Merge pull request #272 from CJackHwang/dev

Remove outdated architecture documentation and improve citation parsing

.env.example

@@ -1,5 +1,8 @@
 # DS2API runtime
+# Runtime listen port inside the app/container
 PORT=5001
+# Docker Compose host port (compose only; container still listens on PORT)
+DS2API_HOST_PORT=6011
 LOG_LEVEL=INFO
 
 # Admin authentication

.github/workflows/quality-gates.yml

@@ -19,7 +19,7 @@ jobs:
       - name: Setup Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.24.x"
+          go-version: "1.26.x"
 
       - name: Setup Node
         uses: actions/setup-node@v4
@@ -28,6 +28,16 @@ jobs:
           cache: "npm"
           cache-dependency-path: webui/package-lock.json
 
+      - name: Setup golangci-lint
+        uses: golangci/golangci-lint-action@v8
+        with:
+          version: v2.11.4
+          install-mode: binary
+          verify: true
+
+      - name: Go Format & Lint Gates
+        run: ./scripts/lint.sh
+
       - name: Refactor Line Gate
         run: ./tests/scripts/check-refactor-line-gate.sh
 

.github/workflows/release-artifacts.yml

@@ -27,7 +27,7 @@ jobs:
       - name: Setup Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.24.x"
+          go-version: "1.26.x"
 
       - name: Setup Node
         uses: actions/setup-node@v4
@@ -79,7 +79,7 @@ jobs:
             CGO_ENABLED=0 GOOS="${GOOS}" GOARCH="${GOARCH}" \
               go build -trimpath -ldflags="-s -w -X ds2api/internal/version.BuildVersion=${BUILD_VERSION}" -o "${STAGE}/${BIN}" ./cmd/ds2api
 
-            cp config.example.json .env.example internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm LICENSE README.MD README.en.md "${STAGE}/"
+            cp config.example.json .env.example LICENSE README.MD README.en.md "${STAGE}/"
             cp -R static/admin "${STAGE}/static/admin"
 
             if [ "${GOOS}" = "windows" ]; then

.gitignore

@@ -9,6 +9,7 @@ config.json
 *.swo
 *~
 .DS_Store
+opencode.json
 
 # Logs
 *.log
@@ -58,3 +59,6 @@ Thumbs.db
 # Claude Code
 .claude/
 CLAUDE.local.md
+
+# Local tool bootstrap cache
+.tmp/

.golangci.yml

@@ -0,0 +1,73 @@
+version: "2"
+
+run:
+  tests: true
+
+linters:
+  default: standard
+  enable:
+    - errcheck
+    - govet
+    - ineffassign
+    - staticcheck
+    - unused
+  settings:
+    dupl:
+      threshold: 100
+    goconst:
+      min-len: 2
+      min-occurrences: 2
+    gocritic:
+      enabled-tags:
+        - diagnostic
+        - experimental
+        - opinionated
+        - performance
+        - style
+      disabled-checks:
+        - wrapperFunc
+        - rangeValCopy
+        - hugeParam
+    gocyclo:
+      min-complexity: 15
+    lll:
+      line-length: 140
+    misspell:
+      locale: US
+    nakedret:
+      max-func-lines: 30
+    prealloc:
+      simple: true
+      range-loops: true
+      for-loops: false
+  exclusions:
+    generated: lax
+    rules:
+      - path: (.+)\.go$
+        text: "ST1000: at least one file in a package should have a package comment"
+    paths:
+      - third_party$
+      - builtin$
+      - examples$
+      - vendor$
+      - webui/node_modules$
+
+issues:
+  max-issues-per-linter: 0
+  max-same-issues: 0
+
+formatters:
+  enable:
+    - gofmt
+  settings:
+    goimports:
+      local-prefixes:
+        - ds2api
+  exclusions:
+    generated: lax
+    paths:
+      - third_party$
+      - builtin$
+      - examples$
+      - vendor$
+      - webui/node_modules$

AGENTS.md

@@ -0,0 +1,23 @@
+# AGENTS.md
+
+These rules apply to all agent-made changes in this repository.
+
+## PR Gate
+
+- Before opening or updating a PR, run the same local gates as `.github/workflows/quality-gates.yml`.
+- Required commands:
+  - `./scripts/lint.sh`
+  - `./tests/scripts/check-refactor-line-gate.sh`
+  - `./tests/scripts/run-unit-all.sh`
+  - `npm run build --prefix webui`
+
+## Go Lint Rules
+
+- Run `gofmt -w` on every changed Go file before commit or push.
+- Do not ignore error returns from I/O-style cleanup calls such as `Close`, `Flush`, `Sync`, or similar methods.
+- If a cleanup error cannot be returned, log it explicitly.
+
+## Change Scope
+
+- Keep changes additive and tightly scoped to the requested feature or bugfix.
+- Do not mix unrelated refactors into feature PRs unless they are required to make the change pass gates.

API.en.md

@@ -4,6 +4,8 @@ Language: [中文](API.md) | [English](API.en.md)
 
 This document describes the actual behavior of the current Go codebase.
 
+Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Deployment](docs/DEPLOY.en.md) / [Testing](docs/TESTING.md)
+
 ---
 
 ## Table of Contents
@@ -35,7 +37,7 @@ This document describes the actual behavior of the current Go codebase.
 
 - OpenAI / Claude / Gemini protocols are now mounted on one shared `chi` router tree assembled in `internal/server/router.go`.
 - Adapter responsibilities are streamlined to: **request normalization → DeepSeek invocation → protocol-shaped rendering**, reducing legacy split-logic paths.
-- Tool-calling semantics are aligned between Go and Node runtime: structured parsing first (JSON/XML/invoke/markup), plus stream-time anti-leak filtering.
+- Tool-calling semantics are aligned between Go and Node runtime: parsing is now centered on XML/Markup-family tool syntax (`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml variants), plus stream-time anti-leak filtering.
 - `Admin API` separates static config from runtime policy: `/admin/config*` for configuration state, `/admin/settings*` for runtime behavior.
 
 ---
@@ -52,8 +54,7 @@ cp config.example.json config.json
 Use it per deployment mode:
 
 - Local run: read `config.json` directly
-- Docker / Vercel: generate Base64 from `config.json`, then set `DS2API_CONFIG_JSON`
-- Compatibility note: `DS2API_CONFIG_JSON` may also contain raw JSON directly; `CONFIG_JSON` is the legacy fallback variable
+- Docker / Vercel: generate Base64 from `config.json`, then set `DS2API_CONFIG_JSON`, or paste raw JSON directly
 
 ```bash
 DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
@@ -107,6 +108,7 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 | POST | `/v1/responses` | Business | OpenAI Responses API (stream/non-stream) |
 | GET | `/v1/responses/{response_id}` | Business | Query stored response (in-memory TTL) |
 | POST | `/v1/embeddings` | Business | OpenAI Embeddings API |
+| POST | `/v1/files` | Business | OpenAI Files upload (multipart/form-data) |
 | GET | `/anthropic/v1/models` | None | Claude model list |
 | POST | `/anthropic/v1/messages` | Business | Claude messages |
 | POST | `/anthropic/v1/messages/count_tokens` | Business | Claude token counting |
@@ -130,15 +132,24 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 | GET | `/admin/config/export` | Admin | Export full config (`config`/`json`/`base64`) |
 | POST | `/admin/keys` | Admin | Add API key |
 | DELETE | `/admin/keys/{key}` | Admin | Delete API key |
+| GET | `/admin/proxies` | Admin | List proxies |
+| POST | `/admin/proxies` | Admin | Add proxy |
+| PUT | `/admin/proxies/{proxyID}` | Admin | Update proxy (empty password keeps old secret) |
+| DELETE | `/admin/proxies/{proxyID}` | Admin | Delete proxy (auto-unbind referenced accounts) |
+| POST | `/admin/proxies/test` | Admin | Test proxy connectivity |
 | GET | `/admin/accounts` | Admin | Paginated account list |
 | POST | `/admin/accounts` | Admin | Add account |
 | DELETE | `/admin/accounts/{identifier}` | Admin | Delete account |
+| PUT | `/admin/accounts/{identifier}/proxy` | Admin | Bind/unbind proxy for an account |
 | GET | `/admin/queue/status` | Admin | Account queue status |
 | POST | `/admin/accounts/test` | Admin | Test one account |
 | POST | `/admin/accounts/test-all` | Admin | Test all accounts |
 | POST | `/admin/accounts/sessions/delete-all` | Admin | Delete all sessions for one account |
 | POST | `/admin/import` | Admin | Batch import keys/accounts |
 | POST | `/admin/test` | Admin | Test API through service |
+| POST | `/admin/dev/raw-samples/capture` | Admin | Fire one request and persist it as a raw sample |
+| GET | `/admin/dev/raw-samples/query` | Admin | Search current in-memory capture chains by prompt keyword |
+| POST | `/admin/dev/raw-samples/save` | Admin | Persist a selected in-memory capture chain as a raw sample |
 | POST | `/admin/vercel/sync` | Admin | Sync config to Vercel |
 | GET | `/admin/vercel/status` | Admin | Vercel sync status |
 | POST | `/admin/vercel/status` | Admin | Vercel sync status / draft compare |
@@ -169,7 +180,7 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 
 ### `GET /v1/models`
 
-No auth required. Returns supported models.
+No auth required. Returns the currently supported DeepSeek native model list.
 
 **Response**:
 
@@ -180,11 +191,21 @@ No auth required. Returns supported models.
     {"id": "deepseek-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
 
+> Note: `/v1/models` returns normalized DeepSeek native model IDs. Common aliases are accepted only as request input and are not expanded as separate items in this endpoint.
+
 ### Model Alias Resolution
 
 For `chat` / `responses` / `embeddings`, DS2API follows a wide-input/strict-output policy:
@@ -194,6 +215,13 @@ For `chat` / `responses` / `embeddings`, DS2API follows a wide-input/strict-outp
 3. If still unmatched, fall back by known family heuristics (`o*`, `gpt-*`, `claude-*`, etc.).
 4. If still unmatched, return `invalid_request_error`.
 
+Current built-in default aliases (excerpt):
+
+- OpenAI: `gpt-4o`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-5`, `gpt-5-mini`, `gpt-5-codex`
+- OpenAI reasoning: `o1`, `o1-mini`, `o3`, `o3-mini`
+- Claude: `claude-sonnet-4-5`, `claude-haiku-4-5`, `claude-opus-4-6` (plus compatibility aliases `claude-3-5-sonnet` / `claude-3-5-haiku` / `claude-3-opus`)
+- Gemini: `gemini-2.5-pro`, `gemini-2.5-flash`
+
 ### `POST /v1/chat/completions`
 
 **Headers**:
@@ -207,7 +235,7 @@ Content-Type: application/json
 
 | Field | Type | Required | Notes |
 | --- | --- | --- | --- |
-| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-4o`, `gpt-5-codex`, `o3`, `claude-sonnet-4-5`, etc.) |
+| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-5`, `gpt-5-mini`, `gpt-5-codex`, `o3`, `claude-opus-4-6`, `gemini-2.5-pro`, `gemini-2.5-flash`, etc.) |
 | `messages` | array | ✅ | OpenAI-style messages |
 | `stream` | boolean | ❌ | Default `false` |
 | `tools` | array | ❌ | Function calling schema |
@@ -265,6 +293,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` models emit `delta.reasoning_content`
 - Text emits `delta.content`
 - Last chunk includes `finish_reason` and `usage`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 #### Tool Calls
 
@@ -297,7 +326,12 @@ When `tools` is present, DS2API performs anti-leak handling:
 }
 ```
 
-**Stream**: Once high-confidence toolcall features are matched, DS2API emits `delta.tool_calls` immediately (without waiting for full JSON closure), then keeps sending argument deltas; confirmed raw tool JSON is never forwarded as `delta.content`.
+**Stream**: Once high-confidence toolcall features are matched, DS2API emits `delta.tool_calls` immediately (without waiting for full argument closure), then keeps sending argument deltas; confirmed tool-call fragments are not forwarded as `delta.content`.
+
+Additional notes:
+
+- The parser currently follows XML/Markup-family tool payloads (`<tool_call>`, `<function_call>`, `<invoke>`, `tool_use`, antml variants). Standalone JSON `tool_calls` payloads are not treated as executable tool calls by default.
+- `tool_calls` shown inside fenced markdown code blocks (for example, ```json ... ```) are treated as examples, not executable calls.
 
 ---
 
@@ -356,7 +390,8 @@ data: [DONE]
 ```
 
 If `tool_choice=required` is violated in stream mode, DS2API emits `response.failed` then `[DONE]` (no `response.completed`).
-Unknown tool names (outside declared `tools`) are rejected and will not be emitted as valid tool calls.
+
+> Current behavior: the parser tries to extract structured tool calls and does not enforce a hard allow-list reject; your tool executor should still validate against a whitelist before executing.
 
 ### `GET /v1/responses/{response_id}`
 
@@ -375,11 +410,27 @@ Business auth required. Returns OpenAI-compatible embeddings shape.
 
 > Requires `embeddings.provider`. Current supported values: `mock` / `deterministic` / `builtin`. If missing/unsupported, returns standard error shape with HTTP 501.
 
+### `POST /v1/files`
+
+Business auth required. OpenAI Files-compatible upload endpoint; currently only `multipart/form-data` is supported.
+
+| Field | Type | Required | Notes |
+| --- | --- | --- | --- |
+| `file` | file | ✅ | Binary payload |
+| `purpose` | string | ❌ | Forwarded purpose field |
+
+Constraints and behavior:
+
+- `Content-Type` must be `multipart/form-data` (otherwise `400`).
+- Total request size limit is `100 MiB` (over-limit returns `413`).
+- Success returns an OpenAI `file` object (`id/object/bytes/filename/purpose/status`, etc.) and includes `account_id` for source-account tracing.
+
 ---
 
 ## Claude-Compatible API
 
 Besides `/anthropic/v1/*`, DS2API also supports shortcut paths: `/v1/messages`, `/messages`, `/v1/messages/count_tokens`, `/messages/count_tokens`.
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `GET /anthropic/v1/models`
 
@@ -401,7 +452,7 @@ No auth required.
 }
 ```
 
-> Note: the example is partial; the real response includes historical Claude 1.x/2.x/3.x/4.x IDs and common aliases.
+> Note: the example is partial; besides the current primary aliases, the real response also includes Claude 4.x snapshots plus historical 3.x / 2.x / 1.x IDs and common aliases.
 
 ### `POST /anthropic/v1/messages`
 
@@ -514,6 +565,7 @@ Supported paths:
 - `/v1/models/{model}:streamGenerateContent` (compat path)
 
 Authentication is the same as other business routes (`Authorization: Bearer <token>` or `x-api-key`).
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -532,6 +584,7 @@ Returns SSE (`text/event-stream`), each chunk as `data: <json>`:
 - regular text: incremental text chunks
 - `tools` mode: buffered and emitted as `functionCall` at finalize phase
 - final chunk: includes `finishReason: "STOP"` and `usageMetadata`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 ---
 
@@ -642,8 +695,9 @@ Reads runtime settings and status, including:
 - `success`
 - `admin` (`has_password_hash`, `jwt_expire_hours`, `jwt_valid_after_unix`, `default_password_warning`)
 - `runtime` (`account_max_inflight`, `account_max_queue`, `global_max_inflight`, `token_refresh_interval_hours`)
+- `compat` (`wide_input_strict_output`, `strip_reference_markers`)
 - `responses` / `embeddings`
-- `auto_delete` (`sessions`)
+- `auto_delete` (`mode`: `none` / `single` / `all`; legacy `sessions=true` is still treated as `all`)
 - `claude_mapping` / `model_aliases`
 - `env_backed`, `needs_vercel_sync`
 - `toolcall` policy is fixed to `feature_match + high` and is no longer returned or editable via settings
@@ -654,9 +708,10 @@ Hot-updates runtime settings. Supported fields:
 
 - `admin.jwt_expire_hours`
 - `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
+- `compat.wide_input_strict_output` / `compat.strip_reference_markers`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
-- `auto_delete.sessions`
+- `auto_delete.mode`
 - `claude_mapping`
 - `model_aliases`
 - `toolcall` policy is fixed and is no longer writable through settings
@@ -684,6 +739,8 @@ The request can send config directly, or wrapped as `{"config": {...}, "mode":"m
 Query params `?mode=merge` / `?mode=replace` are also supported.
 Import accepts `keys`, `accounts`, `claude_mapping` / `claude_model_mapping`, `model_aliases`, `admin`, `runtime`, `responses`, `embeddings`, and `auto_delete`; legacy `toolcall` fields are ignored.
 
+> `compat` fields are managed via `/admin/settings` or the config file; this import endpoint does not update `compat`.
+
 ### `GET /admin/config/export`
 
 Exports full config in three forms: `config`, `json`, and `base64`.
@@ -700,6 +757,26 @@ Exports full config in three forms: `config`, `json`, and `base64`.
 
 **Response**: `{"success": true, "total_keys": 2}`
 
+### `GET /admin/proxies`
+
+Lists proxy configs (password is never returned; use `has_password` as a marker).
+
+### `POST /admin/proxies`
+
+Adds a proxy. Request accepts `id` (optional; auto-generated when omitted), `name`, `type` (`http` / `socks5`), `host`, `port`, `username`, `password`.
+
+### `PUT /admin/proxies/{proxyID}`
+
+Updates a proxy. If `password` is an empty string, the existing secret is preserved.
+
+### `DELETE /admin/proxies/{proxyID}`
+
+Deletes a proxy and automatically clears `proxy_id` on all accounts that reference it.
+
+### `POST /admin/proxies/test`
+
+Tests proxy connectivity: provide `proxy_id` to test a saved proxy; omit it to run a one-off test using proxy fields in the request body.
+
 ### `GET /admin/accounts`
 
 **Query params**:
@@ -707,7 +784,7 @@ Exports full config in three forms: `config`, `json`, and `base64`.
 | Param | Default | Range |
 | --- | --- | --- |
 | `page` | `1` | ≥ 1 |
-| `page_size` | `10` | 1–100 |
+| `page_size` | `10` | 1–5000 |
 | `q` | empty | Filter by identifier / email / mobile |
 
 **Response**:
@@ -748,6 +825,14 @@ Returned items also include `test_status`, usually `ok` or `failed`.
 
 **Response**: `{"success": true, "total_accounts": 5}`
 
+### `PUT /admin/accounts/{identifier}/proxy`
+
+Updates proxy binding for a specific account.
+
+- Request body: `{"proxy_id":"..."}`.
+- Use empty `proxy_id` to unbind proxy.
+- `identifier` supports email / mobile / token-only synthetic id.
+
 ### `GET /admin/queue/status`
 
 ```json
@@ -758,17 +843,25 @@ Returned items also include `test_status`, usually `ok` or `failed`.
   "available_accounts": ["a@example.com"],
   "in_use_accounts": ["b@example.com"],
   "max_inflight_per_account": 2,
-  "recommended_concurrency": 8
+  "global_max_inflight": 8,
+  "recommended_concurrency": 8,
+  "waiting": 0,
+  "max_queue_size": 8
 }
 ```
 
 | Field | Description |
 | --- | --- |
-| `available` | Currently available accounts |
-| `in_use` | Currently in-use accounts |
+| `available` | Accounts that still have spare inflight capacity |
+| `in_use` | Number of occupied in-flight slots |
 | `total` | Total accounts |
+| `available_accounts` | List of account IDs with remaining inflight capacity |
+| `in_use_accounts` | List of account IDs currently in use |
 | `max_inflight_per_account` | Per-account inflight limit |
+| `global_max_inflight` | Global inflight limit |
 | `recommended_concurrency` | Suggested concurrency (`total × max_inflight_per_account`) |
+| `waiting` | Number of queued requests currently waiting |
+| `max_queue_size` | Waiting queue limit |
 
 ### `POST /admin/accounts/test`
 
@@ -871,6 +964,74 @@ Test API availability through the service itself.
 }
 ```
 
+### `POST /admin/dev/raw-samples/capture`
+
+Internally issues one `/v1/chat/completions` request through the service, then persists the request metadata and raw upstream SSE into `tests/raw_stream_samples/<sample-id>/`.
+
+Common request fields:
+
+| Field | Required | Default | Notes |
+| --- | --- | --- | --- |
+| `message` | No | `你好` | Convenience single-turn user message |
+| `messages` | No | Auto-derived from `message` | OpenAI-style message array |
+| `model` | No | `deepseek-chat` | Target model |
+| `stream` | No | `true` | Recommended to keep streaming enabled so raw SSE is recorded |
+| `api_key` | No | First configured key | Business API key to use |
+| `sample_id` | No | Auto-generated | Sample directory name |
+
+On success, the response headers include:
+
+- `X-Ds2-Sample-Id`
+- `X-Ds2-Sample-Dir`
+- `X-Ds2-Sample-Meta`
+- `X-Ds2-Sample-Upstream`
+
+If the request itself succeeds but the process did not record a new upstream capture, the endpoint returns:
+
+```json
+{"detail":"no upstream capture was recorded"}
+```
+
+### `GET /admin/dev/raw-samples/query`
+
+Searches the current process's in-memory capture entries and groups `completion + continue` rounds by `chat_session_id`.
+
+**Query parameters**:
+
+| Param | Default | Notes |
+| --- | --- | --- |
+| `q` | empty | Fuzzy match against request/response text |
+| `limit` | `20` | Max number of chains returned |
+
+**Response fields** include:
+
+- `items[].chain_key`
+- `items[].capture_ids`
+- `items[].round_count`
+- `items[].initial_label`
+- `items[].request_preview`
+- `items[].response_preview`
+
+### `POST /admin/dev/raw-samples/save`
+
+Persists one selected in-memory capture chain into `tests/raw_stream_samples/<sample-id>/`.
+
+Any one of these selectors is accepted:
+
+```json
+{"chain_key":"session:xxxx","sample_id":"tmp-from-memory"}
+```
+
+```json
+{"capture_id":"cap_xxx","sample_id":"tmp-from-memory"}
+```
+
+```json
+{"query":"Guangzhou weather","sample_id":"tmp-from-memory"}
+```
+
+The success payload includes `sample_id`, `dir`, `meta_path`, and `upstream_path`.
+
 ### `POST /admin/vercel/sync`
 
 | Field | Required | Notes |

API.md

@@ -4,6 +4,8 @@
 
 本文档描述当前 Go 代码库的实际 API 行为。
 
+文档导航：[总览](README.MD) / [架构说明](docs/ARCHITECTURE.md) / [部署指南](docs/DEPLOY.md) / [测试指南](docs/TESTING.md)
+
 ---
 
 ## 目录
@@ -35,7 +37,7 @@
 
 - OpenAI / Claude / Gemini 三套协议已统一挂在同一 `chi` 路由树上，由 `internal/server/router.go` 负责装配。
 - 适配器层职责收敛为：**请求归一化 → DeepSeek 调用 → 协议形态渲染**，减少历史版本中“同能力多处实现”的分叉。
-- Tool Calling 的解析策略在 Go 与 Node Runtime 间保持一致：优先结构化解析（JSON/XML/invoke/markup），并在流式场景执行防泄漏筛分。
+- Tool Calling 的解析策略在 Go 与 Node Runtime 间保持一致：当前以 XML/Markup 家族解析为主（含 `<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml 变体），并在流式场景执行防泄漏筛分。
 - `Admin API` 将配置与运行时策略分开：`/admin/config*` 管静态配置，`/admin/settings*` 管运行时行为。
 
 ---
@@ -52,8 +54,7 @@ cp config.example.json config.json
 按部署方式使用：
 
 - 本地运行：直接读取 `config.json`
-- Docker / Vercel：从 `config.json` 生成 Base64，填入 `DS2API_CONFIG_JSON`
-- 兼容写法：`DS2API_CONFIG_JSON` 也可直接填原始 JSON；`CONFIG_JSON` 是旧版兼容回退变量
+- Docker / Vercel：从 `config.json` 生成 Base64，填入 `DS2API_CONFIG_JSON`，也可以直接填原始 JSON
 
 ```bash
 DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
@@ -107,6 +108,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | POST | `/v1/responses` | 业务 | OpenAI Responses 接口（流式/非流式） |
 | GET | `/v1/responses/{response_id}` | 业务 | 查询已生成 response（内存 TTL） |
 | POST | `/v1/embeddings` | 业务 | OpenAI Embeddings 接口 |
+| POST | `/v1/files` | 业务 | OpenAI Files 上传（multipart/form-data） |
 | GET | `/anthropic/v1/models` | 无 | Claude 模型列表 |
 | POST | `/anthropic/v1/messages` | 业务 | Claude 消息接口 |
 | POST | `/anthropic/v1/messages/count_tokens` | 业务 | Claude token 计数 |
@@ -130,15 +132,24 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | GET | `/admin/config/export` | Admin | 导出完整配置（含 `config`/`json`/`base64`） |
 | POST | `/admin/keys` | Admin | 添加 API key |
 | DELETE | `/admin/keys/{key}` | Admin | 删除 API key |
+| GET | `/admin/proxies` | Admin | 代理列表 |
+| POST | `/admin/proxies` | Admin | 添加代理 |
+| PUT | `/admin/proxies/{proxyID}` | Admin | 更新代理（留空 password 表示保留原密码） |
+| DELETE | `/admin/proxies/{proxyID}` | Admin | 删除代理（自动解绑引用该代理的账号） |
+| POST | `/admin/proxies/test` | Admin | 测试代理连通性 |
 | GET | `/admin/accounts` | Admin | 分页账号列表 |
 | POST | `/admin/accounts` | Admin | 添加账号 |
 | DELETE | `/admin/accounts/{identifier}` | Admin | 删除账号 |
+| PUT | `/admin/accounts/{identifier}/proxy` | Admin | 为账号绑定/解绑代理 |
 | GET | `/admin/queue/status` | Admin | 账号队列状态 |
 | POST | `/admin/accounts/test` | Admin | 测试单个账号 |
 | POST | `/admin/accounts/test-all` | Admin | 测试全部账号 |
 | POST | `/admin/accounts/sessions/delete-all` | Admin | 删除某账号的全部会话 |
 | POST | `/admin/import` | Admin | 批量导入 keys/accounts |
 | POST | `/admin/test` | Admin | 测试当前 API 可用性 |
+| POST | `/admin/dev/raw-samples/capture` | Admin | 直接发起一次请求并保存为 raw sample |
+| GET | `/admin/dev/raw-samples/query` | Admin | 按问题关键词查询当前内存抓包链 |
+| POST | `/admin/dev/raw-samples/save` | Admin | 把命中的内存抓包链保存为 raw sample |
 | POST | `/admin/vercel/sync` | Admin | 同步配置到 Vercel |
 | GET | `/admin/vercel/status` | Admin | Vercel 同步状态 |
 | POST | `/admin/vercel/status` | Admin | Vercel 同步状态 / 草稿对比 |
@@ -169,7 +180,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 
 ### `GET /v1/models`
 
-无需鉴权。返回当前支持的模型列表。
+无需鉴权。返回当前支持的 DeepSeek 原生模型列表。
 
 **响应示例**：
 
@@ -180,11 +191,21 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
     {"id": "deepseek-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
 
+> 说明：`/v1/models` 返回的是规范化后的 DeepSeek 原生模型 ID；常见 alias 仅用于请求入参解析，不会在该接口中单独展开返回。
+
 ### 模型 alias 解析策略
 
 对 `chat` / `responses` / `embeddings` 的 `model` 字段采用“宽进严出”：
@@ -194,6 +215,13 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 3. 未命中时按模型家族规则回退（如 `o*`、`gpt-*`、`claude-*`）。
 4. 仍未命中则返回 `invalid_request_error`。
 
+当前内置默认 alias（节选）：
+
+- OpenAI：`gpt-4o`、`gpt-4.1`、`gpt-4.1-mini`、`gpt-4.1-nano`、`gpt-5`、`gpt-5-mini`、`gpt-5-codex`
+- OpenAI Reasoning：`o1`、`o1-mini`、`o3`、`o3-mini`
+- Claude：`claude-sonnet-4-5`、`claude-haiku-4-5`、`claude-opus-4-6`（及 `claude-3-5-sonnet` / `claude-3-5-haiku` / `claude-3-opus` 兼容别名）
+- Gemini：`gemini-2.5-pro`、`gemini-2.5-flash`
+
 ### `POST /v1/chat/completions`
 
 **请求头**：
@@ -207,7 +235,7 @@ Content-Type: application/json
 
 | 字段 | 类型 | 必填 | 说明 |
 | --- | --- | --- | --- |
-| `model` | string | ✅ | 支持 DeepSeek 原生模型 + 常见 alias（如 `gpt-4o`、`gpt-5-codex`、`o3`、`claude-sonnet-4-5`） |
+| `model` | string | ✅ | 支持 DeepSeek 原生模型 + 常见 alias（如 `gpt-5`、`gpt-5-mini`、`gpt-5-codex`、`o3`、`claude-opus-4-6`、`gemini-2.5-pro`、`gemini-2.5-flash` 等） |
 | `messages` | array | ✅ | OpenAI 风格消息数组 |
 | `stream` | boolean | ❌ | 默认 `false` |
 | `tools` | array | ❌ | Function Calling 定义 |
@@ -265,6 +293,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` 模型输出 `delta.reasoning_content`
 - 普通文本输出 `delta.content`
 - 最后一段包含 `finish_reason` 和 `usage`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 #### Tool Calls
 
@@ -297,12 +326,12 @@ data: [DONE]
 }
 ```
 
-**流式**：命中高置信特征后立即输出 `delta.tool_calls`（不等待完整 JSON 闭合），并持续发送 arguments 增量；已确认的 toolcall 原始 JSON 不会回流到 `delta.content`。
+**流式**：命中高置信特征后立即输出 `delta.tool_calls`（不等待完整工具参数闭合），并持续发送 arguments 增量；已确认的工具调用片段不会回流到 `delta.content`。
 
 补充说明：
 
 - **非代码块上下文**下，工具负载即使与普通文本混合，也会按特征识别并产出可执行 tool call（前后普通文本仍可透传）。
-- 解析器以 XML/Markup 为最高优先级，并兼容 JSON、ANTML、text-kv 等格式输入；最终按客户端协议转译为对应 tool call 结构（OpenAI/Claude/Gemini）。
+- 解析器当前走 XML/Markup 家族（包含 `<tool_call>`、`<function_call>`、`<invoke>`、`tool_use`、antml 风格）；纯 JSON `tool_calls` 片段默认不会直接作为可执行调用解析。
 - Markdown fenced code block（例如 ```json ... ```）中的 `tool_calls` 仅视为示例文本，不会被执行。
 
 ---
@@ -382,11 +411,27 @@ data: [DONE]
 
 > 需配置 `embeddings.provider`。当前支持：`mock` / `deterministic` / `builtin`。未配置或不支持时返回标准错误结构（HTTP 501）。
 
+### `POST /v1/files`
+
+需要业务鉴权。兼容 OpenAI Files 上传接口，当前仅支持 `multipart/form-data`。
+
+| 字段 | 类型 | 必填 | 说明 |
+| --- | --- | --- | --- |
+| `file` | file | ✅ | 上传文件二进制 |
+| `purpose` | string | ❌ | 透传到上游用途字段 |
+
+约束与行为：
+
+- 请求必须为 `multipart/form-data`，否则返回 `400`。
+- 请求体总大小上限 `100 MiB`（超限返回 `413`）。
+- 成功返回 OpenAI `file` 对象（`id/object/bytes/filename/purpose/status` 等字段），并附带 `account_id` 便于定位来源账号。
+
 ---
 
 ## Claude 兼容接口
 
 除标准路径 `/anthropic/v1/*` 外，还支持快捷路径 `/v1/messages`、`/messages`、`/v1/messages/count_tokens`、`/messages/count_tokens`。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `GET /anthropic/v1/models`
 
@@ -408,7 +453,7 @@ data: [DONE]
 }
 ```
 
-> 说明：示例仅展示部分模型；实际返回包含 Claude 1.x/2.x/3.x/4.x 历史模型 ID 与常见别名。
+> 说明：示例仅展示部分模型；实际返回除当前主别名外，还包含 Claude 4.x snapshots，以及 3.x / 2.x / 1.x 历史模型 ID 与常见别名。
 
 ### `POST /anthropic/v1/messages`
 
@@ -509,8 +554,6 @@ data: {"type":"message_stop"}
 }
 ```
 
-返回项还会包含 `test_status`，当前值通常为 `ok` 或 `failed`。
-
 ---
 
 ## Gemini 兼容接口
@@ -523,6 +566,7 @@ data: {"type":"message_stop"}
 - `/v1/models/{model}:streamGenerateContent`（兼容路径）
 
 鉴权方式同业务接口（`Authorization: Bearer <token>` 或 `x-api-key`）。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -541,6 +585,7 @@ data: {"type":"message_stop"}
 - 常规文本：持续返回增量文本 chunk
 - `tools` 场景：会缓冲并在结束时输出 `functionCall` 结构
 - 结束 chunk：包含 `finishReason: "STOP"` 与 `usageMetadata`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 ---
 
@@ -651,8 +696,9 @@ data: {"type":"message_stop"}
 - `success`
 - `admin`（`has_password_hash`、`jwt_expire_hours`、`jwt_valid_after_unix`、`default_password_warning`）
 - `runtime`（`account_max_inflight`、`account_max_queue`、`global_max_inflight`、`token_refresh_interval_hours`）
+- `compat`（`wide_input_strict_output`、`strip_reference_markers`）
 - `responses` / `embeddings`
-- `auto_delete`（`sessions`）
+- `auto_delete`（`mode`：`none` / `single` / `all`；旧配置 `sessions=true` 仍按 `all` 处理）
 - `claude_mapping` / `model_aliases`
 - `env_backed`、`needs_vercel_sync`
 - `toolcall` 策略已固定为 `feature_match + high`，不再通过 settings 返回或修改
@@ -663,9 +709,10 @@ data: {"type":"message_stop"}
 
 - `admin.jwt_expire_hours`
 - `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
+- `compat.wide_input_strict_output` / `compat.strip_reference_markers`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
-- `auto_delete.sessions`
+- `auto_delete.mode`
 - `claude_mapping`
 - `model_aliases`
 - `toolcall` 策略已固定，不再作为可写入字段
@@ -693,6 +740,8 @@ data: {"type":"message_stop"}
 也支持在查询参数里传 `?mode=merge` / `?mode=replace`。
 导入时会接受 `keys`、`accounts`、`claude_mapping` / `claude_model_mapping`、`model_aliases`、`admin`、`runtime`、`responses`、`embeddings`、`auto_delete` 等字段；`toolcall` 相关字段会被忽略。
 
+> `compat` 相关字段请通过 `/admin/settings` 或配置文件管理；该导入接口不会更新 `compat`。
+
 ### `GET /admin/config/export`
 
 导出完整配置，返回 `config`、`json`、`base64` 三种格式。
@@ -709,6 +758,26 @@ data: {"type":"message_stop"}
 
 **响应**：`{"success": true, "total_keys": 2}`
 
+### `GET /admin/proxies`
+
+列出代理配置（密码不回传，仅返回 `has_password` 标记）。
+
+### `POST /admin/proxies`
+
+新增代理。请求体支持 `id`（可选，未传则自动生成）、`name`、`type`（`http` / `socks5`）、`host`、`port`、`username`、`password`。
+
+### `PUT /admin/proxies/{proxyID}`
+
+更新指定代理。若请求中 `password` 为空字符串，则保留原密码。
+
+### `DELETE /admin/proxies/{proxyID}`
+
+删除代理，并自动清空所有引用该代理账号的 `proxy_id`。
+
+### `POST /admin/proxies/test`
+
+测试代理连通性：传 `proxy_id` 时测试已保存代理；不传时按请求体代理字段做临时连通性测试。
+
 ### `GET /admin/accounts`
 
 **查询参数**：
@@ -716,7 +785,7 @@ data: {"type":"message_stop"}
 | 参数 | 默认 | 范围 |
 | --- | --- | --- |
 | `page` | `1` | ≥ 1 |
-| `page_size` | `10` | 1–100 |
+| `page_size` | `10` | 1–5000 |
 | `q` | 空 | 按 identifier / email / mobile 过滤 |
 
 **响应**：
@@ -755,6 +824,14 @@ data: {"type":"message_stop"}
 
 **响应**：`{"success": true, "total_accounts": 5}`
 
+### `PUT /admin/accounts/{identifier}/proxy`
+
+更新指定账号绑定代理。
+
+- 请求体：`{"proxy_id":"..."}`；
+- `proxy_id` 传空字符串时表示解绑代理；
+- `identifier` 支持 email / mobile / token-only 合成标识。
+
 ### `GET /admin/queue/status`
 
 ```json
@@ -765,17 +842,25 @@ data: {"type":"message_stop"}
   "available_accounts": ["a@example.com"],
   "in_use_accounts": ["b@example.com"],
   "max_inflight_per_account": 2,
-  "recommended_concurrency": 8
+  "global_max_inflight": 8,
+  "recommended_concurrency": 8,
+  "waiting": 0,
+  "max_queue_size": 8
 }
 ```
 
 | 字段 | 说明 |
 | --- | --- |
-| `available` | 当前可用账号数 |
-| `in_use` | 当前使用中的账号数 |
+| `available` | 仍有剩余并发槽位的账号数 |
+| `in_use` | 当前已占用的 in-flight 槽位数 |
 | `total` | 总账号数 |
+| `available_accounts` | 仍有剩余并发槽位的账号 ID 列表 |
+| `in_use_accounts` | 当前处于使用中的账号 ID 列表 |
 | `max_inflight_per_account` | 每账号并发上限 |
+| `global_max_inflight` | 全局并发上限 |
 | `recommended_concurrency` | 建议并发值（`total × max_inflight_per_account`） |
+| `waiting` | 当前等待中的请求数 |
+| `max_queue_size` | 等待队列上限 |
 
 ### `POST /admin/accounts/test`
 
@@ -877,6 +962,74 @@ data: {"type":"message_stop"}
 }
 ```
 
+### `POST /admin/dev/raw-samples/capture`
+
+直接通过服务自身发起一次 `/v1/chat/completions` 请求，并把请求元信息和上游原始 SSE 保存到 `tests/raw_stream_samples/<sample-id>/`。
+
+常用请求字段：
+
+| 字段 | 必填 | 默认值 | 说明 |
+| --- | --- | --- | --- |
+| `message` | 否 | `你好` | 便捷单轮用户消息 |
+| `messages` | 否 | 自动由 `message` 生成 | OpenAI 风格消息数组 |
+| `model` | 否 | `deepseek-chat` | 目标模型 |
+| `stream` | 否 | `true` | 建议保留流式，以记录原始 SSE |
+| `api_key` | 否 | 配置中第一个 key | 调用业务接口使用的 key |
+| `sample_id` | 否 | 自动生成 | 样本目录名 |
+
+成功时会在响应头里附带：
+
+- `X-Ds2-Sample-Id`
+- `X-Ds2-Sample-Dir`
+- `X-Ds2-Sample-Meta`
+- `X-Ds2-Sample-Upstream`
+
+如果请求本身成功，但当前进程没有记录到新的上游抓包，会返回：
+
+```json
+{"detail":"no upstream capture was recorded"}
+```
+
+### `GET /admin/dev/raw-samples/query`
+
+按关键词查询当前进程内存里的抓包记录，并按 `chat_session_id` 归并 `completion + continue` 链。
+
+**查询参数**：
+
+| 参数 | 默认值 | 说明 |
+| --- | --- | --- |
+| `q` | 空 | 按请求体/响应体关键词模糊匹配 |
+| `limit` | `20` | 返回链条数上限 |
+
+**响应字段**包含：
+
+- `items[].chain_key`
+- `items[].capture_ids`
+- `items[].round_count`
+- `items[].initial_label`
+- `items[].request_preview`
+- `items[].response_preview`
+
+### `POST /admin/dev/raw-samples/save`
+
+把当前内存中的某条抓包链落盘为 `tests/raw_stream_samples/<sample-id>/`。
+
+支持以下任一种选中方式：
+
+```json
+{"chain_key":"session:xxxx","sample_id":"tmp-from-memory"}
+```
+
+```json
+{"capture_id":"cap_xxx","sample_id":"tmp-from-memory"}
+```
+
+```json
+{"query":"广州天气","sample_id":"tmp-from-memory"}
+```
+
+成功响应会返回 `sample_id`、`dir`、`meta_path`、`upstream_path`。
+
 ### `POST /admin/vercel/sync`
 
 | 字段 | 必填 | 说明 |

Dockerfile

@@ -34,7 +34,7 @@ CMD ["/usr/local/bin/ds2api"]
 
 FROM runtime-base AS runtime-from-source
 COPY --from=go-builder /out/ds2api /usr/local/bin/ds2api
-COPY --from=go-builder /app/internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
+
 COPY --from=go-builder /app/config.example.json /app/config.example.json
 COPY --from=webui-builder /app/static/admin /app/static/admin
 
@@ -53,13 +53,13 @@ RUN set -eux; \
     test -n "${PKG_DIR}"; \
     mkdir -p /out/static; \
     cp "${PKG_DIR}/ds2api" /out/ds2api; \
-    cp "${PKG_DIR}/sha3_wasm_bg.7b9ca65ddd.wasm" /out/sha3_wasm_bg.7b9ca65ddd.wasm; \
+
     cp "${PKG_DIR}/config.example.json" /out/config.example.json; \
     cp -R "${PKG_DIR}/static/admin" /out/static/admin
 
 FROM runtime-base AS runtime-from-dist
 COPY --from=dist-extract /out/ds2api /usr/local/bin/ds2api
-COPY --from=dist-extract /out/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
+
 COPY --from=dist-extract /out/config.example.json /app/config.example.json
 COPY --from=dist-extract /out/static/admin /app/static/admin
 

README.MD

@@ -16,6 +16,10 @@
 
 将 DeepSeek Web 对话能力转换为 OpenAI、Claude 与 Gemini 兼容 API。后端为 **Go 全量实现**，前端为 React WebUI 管理台（源码在 `webui/`，部署时自动构建到 `static/admin`）。
 
+文档入口：[文档导航](docs/README.md) / [架构说明](docs/ARCHITECTURE.md) / [接口文档](API.md)
+
+【感谢Linux.do社区及GitHub社区各位开发者对项目的支持与贡献】
+
 > **重要免责声明**
 >
 > 本仓库仅供学习、研究、个人实验和内部验证使用，不提供任何形式的商业授权、适用性保证或结果保证。
@@ -24,7 +28,7 @@
 >
 > 请勿将本项目用于违反服务条款、协议、法律法规或平台规则的场景。商业使用前请自行确认 `LICENSE`、相关协议以及你是否获得了作者的书面许可。
 
-## 架构概览
+## 架构概览（摘要）
 
 ```mermaid
 flowchart LR
@@ -48,7 +52,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(并发槽位 + 等待队列)"]
             DSClient["DeepSeek Client\n(Session / Auth / HTTP)"]
-            Pow["PoW WASM\n(wazero 预加载)"]
+            Pow["PoW 实现\n(纯 Go 毫秒级)"]
             Tool["Tool Sieve\n(Go/Node 语义对齐)"]
         end
     end
@@ -72,32 +76,24 @@ flowchart LR
     Bridge --> Client
 ```
 
+详细架构拆分与目录职责见 [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)。
+
 - **后端**：Go（`cmd/ds2api/`、`api/`、`internal/`），不依赖 Python 运行时
 - **前端**：React 管理台（`webui/`），运行时托管静态构建产物
 - **部署**：本地运行、Docker、Vercel Serverless、Linux systemd
 
-### 3.0 底层架构调整（相较旧版本）
-
-- **统一路由内核**：所有协议入口统一汇聚到 `internal/server/router.go`，并在同一路由树中注册 OpenAI / Claude / Gemini / Admin / WebUI 路由，避免多入口行为漂移。
-- **统一执行链路**：Claude / Gemini 入口先经 `internal/translatorcliproxy` 做协议转换，再进入 `openai.ChatCompletions` 统一处理工具调用与流式语义，最后再转换回原协议响应。
-- **适配器分层更清晰**：`internal/adapter/{claude,gemini}` 负责入口/出口协议封装，`internal/adapter/openai` 负责核心执行，DeepSeek 侧调用只保留在 OpenAI 内核中。
-- **Tool Calling 双运行时对齐**：Go 侧（`internal/util`）与 Vercel Node 侧（`internal/js/helpers/stream-tool-sieve`）保持一致的解析/防泄漏语义，覆盖 JSON / XML / invoke / text-kv 多风格输入。
-- **配置与运行时设置解耦**：静态配置（`config`）与运行时策略（`settings`）通过 Admin API 分离管理，支持热更新和密码轮换失效旧 JWT。
-- **流式能力升级**：`/v1/responses` 与 `/v1/chat/completions` 共享更一致的工具调用增量输出策略，降低不同 SDK 下的行为差异。
-- **可观测与可运维增强**：`/healthz`、`/readyz`、`/admin/version`、`/admin/dev/captures` 形成排障闭环，便于发布后验证。
-
 ## 核心能力
 
 | 能力 | 说明 |
 | --- | --- |
-| OpenAI 兼容 | `GET /v1/models`、`GET /v1/models/{id}`、`POST /v1/chat/completions`、`POST /v1/responses`、`GET /v1/responses/{response_id}`、`POST /v1/embeddings` |
+| OpenAI 兼容 | `GET /v1/models`、`GET /v1/models/{id}`、`POST /v1/chat/completions`、`POST /v1/responses`、`GET /v1/responses/{response_id}`、`POST /v1/embeddings`、`POST /v1/files` |
 | Claude 兼容 | `GET /anthropic/v1/models`、`POST /anthropic/v1/messages`、`POST /anthropic/v1/messages/count_tokens`（及快捷路径 `/v1/messages`、`/messages`） |
 | Gemini 兼容 | `POST /v1beta/models/{model}:generateContent`、`POST /v1beta/models/{model}:streamGenerateContent`（及 `/v1/models/{model}:*` 路径） |
 | 多账号轮询 | 自动 token 刷新、邮箱/手机号双登录方式 |
 | 并发队列控制 | 每账号 in-flight 上限 + 等待队列，动态计算建议并发值 |
-| DeepSeek PoW | WASM 计算（`wazero`），无需外部 Node.js 依赖 |
+| DeepSeek PoW | 纯 Go 高性能实现（DeepSeekHashV1），毫秒级响应 |
 | Tool Calling | 防泄漏处理：非代码块高置信特征识别、`delta.tool_calls` 早发、结构化增量输出 |
-| Admin API | 配置管理、运行时设置热更新、账号测试 / 批量测试、会话清理、导入导出、Vercel 同步、版本检查 |
+| Admin API | 配置管理、运行时设置热更新、代理管理、账号测试 / 批量测试、会话清理、导入导出、Vercel 同步、版本检查 |
 | WebUI 管理台 | `/admin` 单页应用（中英文双语、深色模式） |
 | 运维探针 | `GET /healthz`（存活）、`GET /readyz`（就绪） |
 
@@ -111,37 +107,45 @@ flowchart LR
 | P0 | Anthropic SDK（messages） | ✅ |
 | P0 | Google Gemini SDK（generateContent） | ✅ |
 | P1 | LangChain / LlamaIndex / OpenWebUI（OpenAI 兼容接入） | ✅ |
-| P2 | MCP 独立桥接层 | 规划中 |
 
 ## 模型支持
 
-### OpenAI 接口
+### OpenAI 接口（`GET /v1/models`）
 
-| 模型 | thinking | search |
-| --- | --- | --- |
-| `deepseek-chat` | ❌ | ❌ |
-| `deepseek-reasoner` | ✅ | ❌ |
-| `deepseek-chat-search` | ❌ | ✅ |
-| `deepseek-reasoner-search` | ✅ | ✅ |
+| 模型类型 | 模型 ID | thinking | search |
+| --- | --- | --- | --- |
+| default | `deepseek-chat` | ❌ | ❌ |
+| default | `deepseek-reasoner` | ✅ | ❌ |
+| default | `deepseek-chat-search` | ❌ | ✅ |
+| default | `deepseek-reasoner-search` | ✅ | ✅ |
+| expert | `deepseek-expert-chat` | ❌ | ❌ |
+| expert | `deepseek-expert-reasoner` | ✅ | ❌ |
+| expert | `deepseek-expert-chat-search` | ❌ | ✅ |
+| expert | `deepseek-expert-reasoner-search` | ✅ | ✅ |
+| vision | `deepseek-vision-chat` | ❌ | ❌ |
+| vision | `deepseek-vision-reasoner` | ✅ | ❌ |
+| vision | `deepseek-vision-chat-search` | ❌ | ✅ |
+| vision | `deepseek-vision-reasoner-search` | ✅ | ✅ |
 
-### Claude 接口
+除原生模型外，也支持常见 alias 输入（如 `gpt-5`、`gpt-5-mini`、`gpt-5-codex`、`gpt-4.1`、`o3`、`claude-opus-4-6`、`claude-sonnet-4-5`、`gemini-2.5-pro`、`gemini-2.5-flash` 等），但 `/v1/models` 返回的是规范化后的 DeepSeek 原生模型 ID。
 
-| 模型 | 默认映射 |
+### Claude 接口（`GET /anthropic/v1/models`）
+
+| 当前常用模型 | 默认映射 |
 | --- | --- |
 | `claude-sonnet-4-5` | `deepseek-chat` |
 | `claude-haiku-4-5`（兼容 `claude-3-5-haiku-latest`） | `deepseek-chat` |
 | `claude-opus-4-6` | `deepseek-reasoner` |
 
 可通过配置中的 `claude_mapping` 或 `claude_model_mapping` 覆盖映射关系。
-另外，`/anthropic/v1/models` 现已包含 Claude 1.x/2.x/3.x/4.x 历史模型 ID 与常见别名，便于旧客户端直接兼容。
-
+`/anthropic/v1/models` 除上述当前主别名外，还会返回 Claude 4.x snapshots，以及 3.x / 2.x / 1.x 历史模型 ID 与常见 alias，便于旧客户端直接兼容。
 
 #### Claude Code 接入避坑（实测）
 
 - `ANTHROPIC_BASE_URL` 推荐直接指向 DS2API 根地址（例如 `http://127.0.0.1:5001`），Claude Code 会请求 `/v1/messages?beta=true`。
 - `ANTHROPIC_API_KEY` 需要与 `config.json` 中 `keys` 一致；建议同时保留常规 key 与 `sk-ant-*` 形态 key，兼容不同客户端校验习惯。
 - 若系统设置了代理，建议对 DS2API 地址配置 `NO_PROXY=127.0.0.1,localhost,<你的主机IP>`，避免本地回环请求被代理拦截。
-- 如遇“工具调用输出成文本、未执行”问题，请升级到包含 Claude 工具调用多格式解析（JSON/XML/ANTML/invoke）的版本。
+- 如遇“工具调用输出成文本、未执行”问题，请优先检查模型输出是否为受支持的 XML/Markup 工具块（例如 `<tool_call>` / `<function_call>` / `<invoke>` / `tool_use`），而不是纯 JSON `tool_calls` 片段。
 
 ### Gemini 接口
 
@@ -149,6 +153,15 @@ Gemini 适配器将模型名通过 `model_aliases` 或内置规则映射到 Deep
 
 ## 快速开始
 
+### 部署方式优先级建议
+
+推荐按以下顺序选择部署方式：
+
+1. **下载 Release 构建包运行**：最省事，产物已编译完成，最适合大多数用户。
+2. **Docker / GHCR 镜像部署**：适合需要容器化、编排或云环境部署。
+3. **Vercel 部署**：适合已有 Vercel 环境且接受其平台约束的场景。
+4. **本地源码运行 / 自行编译**：适合开发、调试或需要自行修改代码的场景。
+
 ### 通用第一步（所有部署方式）
 
 把 `config.json` 作为唯一配置源（推荐做法）：
@@ -160,30 +173,21 @@ cp config.example.json config.json
 
 后续部署建议：
 - 本地运行：直接读取 `config.json`
-- Docker / Vercel：由 `config.json` 生成 `DS2API_CONFIG_JSON`（Base64）注入环境变量
-- 兼容写法：`DS2API_CONFIG_JSON` 也可以直接写原始 JSON；`CONFIG_JSON` 是旧版回退变量
+- Docker / Vercel：由 `config.json` 生成 `DS2API_CONFIG_JSON`（Base64）注入环境变量，也可以直接写原始 JSON
 
-### 方式一：本地运行
+### 方式一：下载 Release 构建包
 
-**前置要求**：Go 1.26+，Node.js 20+（仅在需要构建 WebUI 时）
+每次发布 Release 时，GitHub Actions 会自动构建多平台二进制包：
 
 ```bash
-# 1. 克隆仓库
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
-
-# 2. 配置
+# 下载对应平台的压缩包后
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 cp config.example.json config.json
-# 编辑 config.json，填入你的 DeepSeek 账号信息和 API key
-
-# 3. 启动
-go run ./cmd/ds2api
+# 编辑 config.json
+./ds2api
 ```
 
-默认监听地址：`http://localhost:5001`
-
-> **WebUI 自动构建**：本地首次启动时，若 `static/admin` 不存在，会自动尝试执行 `npm ci`（仅在缺少依赖时）和 `npm run build -- --outDir static/admin --emptyOutDir`（需要本机有 Node.js）。你也可以手动构建：`./scripts/build-webui.sh`
-
 ### 方式二：Docker 运行
 
 ```bash
@@ -191,7 +195,7 @@ go run ./cmd/ds2api
 cp .env.example .env
 cp config.example.json config.json
 
-# 2. 编辑 .env（至少设置 DS2API_ADMIN_KEY）
+# 2. 编辑 .env（至少设置 DS2API_ADMIN_KEY；如需修改宿主机端口，可额外设置 DS2API_HOST_PORT）
 #    DS2API_ADMIN_KEY=请替换为强密码
 
 # 3. 启动
@@ -201,7 +205,7 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
-默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请调整 `ports` 配置。
+默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请设置 `DS2API_HOST_PORT=5001`（或者手动调整 `ports` 配置）。
 
 更新镜像：`docker-compose up -d --build`
 
@@ -237,35 +241,28 @@ base64 < config.json | tr -d '\n'
 
 详细部署说明请参阅 [部署指南](docs/DEPLOY.md)。
 
-### 方式四：下载 Release 构建包
+### 方式四：本地源码运行
 
-每次发布 Release 时，GitHub Actions 会自动构建多平台二进制包：
+**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）
 
 ```bash
-# 下载对应平台的压缩包后
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 1. 克隆仓库
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
+
+# 2. 配置
 cp config.example.json config.json
-# 编辑 config.json
-./ds2api
+# 编辑 config.json，填入你的 DeepSeek 账号信息和 API key
+
+# 3. 启动
+go run ./cmd/ds2api
 ```
 
-### 方式五：OpenCode CLI 接入
+默认本地访问地址：`http://127.0.0.1:5001`
 
-1. 复制示例配置：
+服务实际绑定：`0.0.0.0:5001`，因此同一局域网设备通常也可以通过你的内网 IP 访问。
 
-```bash
-cp opencode.json.example opencode.json
-```
-
-2. 编辑 `opencode.json`：
-- 将 `baseURL` 改为你的 DS2API 地址（例如 `https://your-domain.com/v1`）
-- 将 `apiKey` 改为你的 DS2API key（对应 `config.keys`）
-
-3. 在项目目录启动 OpenCode CLI（按你的安装方式运行 `opencode`）。
-
-> 建议优先使用 OpenAI 兼容路径（`/v1/*`），即示例里的 `@ai-sdk/openai-compatible` provider。
-> 若客户端支持 `wire_api`，可分别测试 `responses` 与 `chat`，DS2API 两条链路都兼容。
+> **WebUI 自动构建**：本地首次启动时，若 `static/admin` 不存在，会自动尝试执行 `npm ci`（仅在缺少依赖时）和 `npm run build -- --outDir static/admin --emptyOutDir`（需要本机有 Node.js）。你也可以手动构建：`./scripts/build-webui.sh`
 
 ## 配置说明
 
@@ -286,11 +283,16 @@ cp opencode.json.example opencode.json
   ],
   "model_aliases": {
     "gpt-4o": "deepseek-chat",
+    "gpt-5": "deepseek-chat",
+    "gpt-5-mini": "deepseek-chat",
     "gpt-5-codex": "deepseek-reasoner",
-    "o3": "deepseek-reasoner"
+    "o3": "deepseek-reasoner",
+    "claude-opus-4-6": "deepseek-reasoner",
+    "gemini-2.5-flash": "deepseek-chat"
   },
   "compat": {
-    "wide_input_strict_output": true
+    "wide_input_strict_output": true,
+    "strip_reference_markers": true
   },
   "responses": {
     "store_ttl_seconds": 900
@@ -312,7 +314,7 @@ cp opencode.json.example opencode.json
     "token_refresh_interval_hours": 6
   },
   "auto_delete": {
-    "sessions": false
+    "mode": "none"
   }
 }
 ```
@@ -322,13 +324,14 @@ cp opencode.json.example opencode.json
 - `token`：配置文件中即使填写也会在加载时被清空（不会从 `config.json` 读取 token）；实际 token 仅在运行时内存中维护并自动刷新
 - `model_aliases`：常见模型名（如 GPT/Codex/Claude）到 DeepSeek 模型的映射
 - `compat.wide_input_strict_output`：建议保持 `true`（当前实现默认宽进严出）
-- `toolcall`：策略已固定为特征匹配 + 高置信早发，不再作为可配置项
+- `compat.strip_reference_markers`：建议保持 `true`，用于清理可见输出中的引用/标记
+- `toolcall`：旧字段，当前实现已固定为特征匹配 + 高置信早发；即使保留在配置里也会被忽略
 - `responses.store_ttl_seconds`：`/v1/responses/{id}` 的内存缓存 TTL
 - `embeddings.provider`：embedding 提供方（当前内置 `deterministic/mock/builtin`）
 - `claude_mapping`：字典中 `fast`/`slow` 后缀映射到对应 DeepSeek 模型（兼容读取 `claude_model_mapping`）
 - `admin`：管理后台设置（JWT 过期时间、密码哈希等），可通过 Admin Settings API 热更新
 - `runtime`：运行时参数（并发限制、队列大小、托管账号 token 刷新间隔），可通过 Admin Settings API 热更新；`account_max_queue=0`/`global_max_inflight=0` 表示按推荐值自动计算，`token_refresh_interval_hours=6` 为默认强制重登间隔
-- `auto_delete.sessions`：是否在请求结束后自动清理 DeepSeek 会话（默认 `false`，可在 Settings 热更新）
+- `auto_delete.mode`：请求结束后如何清理 DeepSeek 远端聊天记录，支持 `none`（默认，不删除）、`single`（仅删除当前会话）、`all`（清空全部会话）；旧配置里的 `auto_delete.sessions=true` 仍会被视为 `all`
 
 ### 环境变量
 
@@ -341,31 +344,23 @@ cp opencode.json.example opencode.json
 | `DS2API_JWT_EXPIRE_HOURS` | Admin JWT 过期小时数 | `24` |
 | `DS2API_CONFIG_PATH` | 配置文件路径 | `config.json` |
 | `DS2API_CONFIG_JSON` | 直接注入配置（JSON 或 Base64） | — |
-| `CONFIG_JSON` | 旧版兼容配置注入 | — |
 | `DS2API_ENV_WRITEBACK` | 环境变量模式下自动写回配置文件并切换文件模式（`1/true/yes/on`） | 关闭 |
-| `DS2API_WASM_PATH` | PoW WASM 文件路径 | 自动查找 |
 | `DS2API_STATIC_ADMIN_DIR` | 管理台静态文件目录 | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | 启动时自动构建 WebUI | 本地开启，Vercel 关闭 |
 | `DS2API_DEV_PACKET_CAPTURE` | 本地开发抓包开关（记录最近会话请求/响应体） | 本地非 Vercel 默认开启 |
-| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | 本地抓包保留条数（超出自动淘汰） | `5` |
-| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | 单条响应体最大记录字节数 | `2097152` |
+| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | 本地抓包保留条数（超出自动淘汰） | `20` |
+| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | 单条响应体最大记录字节数 | `5242880` |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | 每账号最大并发 in-flight 请求数 | `2` |
-| `DS2API_ACCOUNT_CONCURRENCY` | 同上（兼容旧名） | — |
 | `DS2API_ACCOUNT_MAX_QUEUE` | 等待队列上限 | `recommended_concurrency` |
-| `DS2API_ACCOUNT_QUEUE_SIZE` | 同上（兼容旧名） | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | 全局最大 in-flight 请求数 | `recommended_concurrency` |
-| `DS2API_MAX_INFLIGHT` | 同上（兼容旧名） | — |
 | `DS2API_VERCEL_INTERNAL_SECRET` | Vercel 混合流式内部鉴权密钥 | 回退用 `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | 流式 lease 过期秒数 | `900` |
-| `DS2API_DEV_PACKET_CAPTURE` | 本地开发抓包开关（记录最近会话请求/响应体） | 本地非 Vercel 默认开启 |
-| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | 本地抓包保留条数（超出自动淘汰） | `5` |
-| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | 单条响应体最大记录字节数 | `2097152` |
 | `VERCEL_TOKEN` | Vercel 同步 token | — |
 | `VERCEL_PROJECT_ID` | Vercel 项目 ID | — |
 | `VERCEL_TEAM_ID` | Vercel 团队 ID | — |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | Vercel 部署保护绕过密钥（内部 Node→Go 调用） | — |
 
-> 提示：当检测到 `DS2API_CONFIG_JSON/CONFIG_JSON` 时，管理台会显示当前模式风险与自动持久化状态（含 `DS2API_CONFIG_PATH` 路径与模式切换说明）。
+> 提示：当检测到 `DS2API_CONFIG_JSON` 时，管理台会显示当前模式风险与自动持久化状态（含 `DS2API_CONFIG_PATH` 路径与模式切换说明）。
 
 ## 鉴权模式
 
@@ -397,7 +392,7 @@ Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `
 当请求中带 `tools` 时，DS2API 会做防泄漏处理与结构化转译：
 
 1. 只在**非代码块上下文**启用执行型 toolcall 识别（代码块示例默认不触发）
-2. 解析层以 XML/Markup 为最高优先级，同时兼容 JSON / ANTML / invoke / text-kv，并统一归一到内部工具调用结构
+2. 解析层当前以 XML/Markup 家族为准（`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml 变体）；纯 JSON `tool_calls` 片段默认不作为可执行调用解析
 3. `responses` 流式严格使用官方 item 生命周期事件（`response.output_item.*`、`response.content_part.*`、`response.function_call_arguments.*`）
 4. `responses` 支持并执行 `tool_choice`（`auto`/`none`/`required`/强制函数）；`required` 违规时非流式返回 `422`，流式返回 `response.failed`
 5. 客户端请求哪种协议，就按该协议返回工具调用（OpenAI/Claude/Gemini 各自原生结构）；模型侧优先约束输出规范 XML，再由兼容层转译
@@ -408,13 +403,13 @@ Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `
 
 ## 本地开发抓包工具
 
-用于定位「responses 思考流/工具调用」等问题。开启后会自动记录最近 N 条 DeepSeek 对话上游请求体与响应体（默认 5 条，超出自动淘汰）。
+用于定位「responses 思考流/工具调用」等问题。开启后会自动记录最近 N 条 DeepSeek 对话上游请求体与响应体（默认 20 条，超出自动淘汰；单条响应体默认最多记录 5 MB）。
 
 启用示例：
 
 ```bash
 DS2API_DEV_PACKET_CAPTURE=true \
-DS2API_DEV_PACKET_CAPTURE_LIMIT=5 \
+DS2API_DEV_PACKET_CAPTURE_LIMIT=20 \
 go run ./cmd/ds2api
 ```
 
@@ -422,6 +417,8 @@ go run ./cmd/ds2api
 
 - `GET /admin/dev/captures`：查看抓包列表（最新在前）
 - `DELETE /admin/dev/captures`：清空抓包
+- `GET /admin/dev/raw-samples/query?q=关键词&limit=20`：按问题关键词查询当前内存抓包，并按 `chat_session_id` 归并 `completion + continue` 链
+- `POST /admin/dev/raw-samples/save`：把命中的某条抓包链保存为 `tests/raw_stream_samples/<sample-id>/` 回放样本
 
 返回字段包含：
 
@@ -429,67 +426,10 @@ go run ./cmd/ds2api
 - `response_body`：上游返回的原始流式内容拼接文本
 - `response_truncated`：是否触发单条大小截断
 
-## 项目结构
+保存接口支持用 `query`、`chain_key` 或 `capture_id` 选中目标。例如：
 
-```text
-ds2api/
-├── app/                     # 统一 HTTP Handler 组装层（供本地与 Serverless 复用）
-├── cmd/
-│   ├── ds2api/              # 本地 / 容器启动入口
-│   └── ds2api-tests/        # 端到端测试集入口
-├── api/
-│   ├── index.go             # Vercel Serverless Go 入口
-│   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # 账号池与并发队列
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI 兼容适配器（含 Tool Call 解析、Vercel 流式 prepare/release）
-│   │   ├── claude/          # Claude 兼容适配器
-│   │   └── gemini/          # Gemini 兼容适配器（generateContent / streamGenerateContent）
-│   ├── admin/               # Admin API handlers（含 Settings 热更新）
-│   ├── auth/                # 鉴权与 JWT
-│   ├── claudeconv/          # Claude 消息格式转换
-│   ├── compat/              # Go 版本兼容与回归测试辅助
-│   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek API 客户端、PoW WASM
-│   ├── js/                  # Node 运行时流式处理与兼容逻辑
-│   ├── devcapture/          # 开发抓包模块
-│   ├── format/              # 输出格式化
-│   ├── prompt/              # Prompt 构建
-│   ├── server/              # HTTP 路由与中间件（chi router）
-│   ├── sse/                 # SSE 解析工具
-│   ├── stream/              # 统一流式消费引擎
-│   ├── testsuite/           # 端到端测试框架与用例编排
-│   ├── translatorcliproxy/  # CLIProxy 桥接与流写入组件
-│   ├── util/                # 通用工具函数
-│   ├── version/             # 版本解析 / 比较与 tag 规范化
-│   └── webui/               # WebUI 静态文件托管与自动构建
-├── webui/                   # React WebUI 源码（Vite + Tailwind）
-│   └── src/
-│       ├── app/             # 路由、鉴权、配置状态管理
-│       ├── features/        # 业务功能模块（account/settings/vercel/apiTester）
-│       ├── components/      # 登录/落地页等通用组件
-│       └── locales/         # 中英文语言包（zh.json / en.json）
-├── scripts/
-│   └── build-webui.sh       # WebUI 手动构建脚本
-├── tests/
-│   ├── compat/              # 兼容性测试夹具与期望输出
-│   ├── node/                # Node 侧单元测试（chat-stream / tool-sieve）
-│   └── scripts/             # 统一测试脚本入口（unit/e2e）
-├── docs/                    # 部署 / 贡献 / 测试等辅助文档
-├── static/admin/            # WebUI 构建产物（不提交到 Git）
-├── .github/
-│   ├── workflows/           # GitHub Actions（质量门禁 + Release 自动构建）
-│   ├── ISSUE_TEMPLATE/      # Issue 模板
-│   └── PULL_REQUEST_TEMPLATE.md
-├── config.example.json      # 配置文件示例
-├── .env.example             # 环境变量示例
-├── Dockerfile               # 多阶段构建（WebUI + Go）
-├── docker-compose.yml       # 生产环境 Docker Compose
-├── docker-compose.dev.yml   # 开发环境 Docker Compose
-├── vercel.json              # Vercel 路由与构建配置
-└── go.mod / go.sum          # Go 模块依赖
+```json
+{"query":"广州天气","sample_id":"gz-weather-from-memory"}
 ```
 
 ## 文档索引
@@ -538,7 +478,7 @@ npm ci --prefix webui && npm run build --prefix webui
 go test ./...
 
 # 运行 tool calls 相关测试（调试工具调用问题）
-go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/toolcall/
 
 # 运行端到端测试
 ./tests/scripts/run-live.sh

README.en.md

@@ -16,6 +16,8 @@ Language: [中文](README.MD) | [English](README.en.md)
 
 DS2API converts DeepSeek Web chat capability into OpenAI-compatible, Claude-compatible, and Gemini-compatible APIs. The backend is a **pure Go implementation**, with a React WebUI admin panel (source in `webui/`, build output auto-generated to `static/admin` during deployment).
 
+Documentation entry: [Docs Index](docs/README.md) / [Architecture](docs/ARCHITECTURE.en.md) / [API Reference](API.en.md)
+
 > **Important Disclaimer**
 >
 > This repository is provided for learning, research, personal experimentation, and internal validation only. It does not grant any commercial authorization and comes with no warranty of fitness, stability, or results.
@@ -24,7 +26,7 @@ DS2API converts DeepSeek Web chat capability into OpenAI-compatible, Claude-comp
 >
 > Do not use this project in ways that violate service terms, agreements, laws, or platform rules. Before any commercial use, review the `LICENSE`, the relevant terms, and confirm that you have the author's written permission.
 
-## Architecture Overview
+## Architecture Overview (Summary)
 
 ```mermaid
 flowchart LR
@@ -48,7 +50,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(in-flight slots + wait queue)"]
             DSClient["DeepSeek Client\n(session / auth / HTTP)"]
-            Pow["PoW WASM\n(wazero preload)"]
+            Pow["PoW Solver\n(Pure Go ms-level)"]
             Tool["Tool Sieve\n(Go/Node semantic parity)"]
         end
     end
@@ -72,32 +74,24 @@ flowchart LR
     Bridge --> Client
 ```
 
+For the full module-by-module architecture and directory responsibilities, see [docs/ARCHITECTURE.en.md](docs/ARCHITECTURE.en.md).
+
 - **Backend**: Go (`cmd/ds2api/`, `api/`, `internal/`), no Python runtime
 - **Frontend**: React admin panel (`webui/`), served as static build at runtime
 - **Deployment**: local run, Docker, Vercel serverless, Linux systemd
 
-### 3.0 Architecture Changes (vs older releases)
-
-- **Unified routing core**: all protocol entries are now centralized through `internal/server/router.go`, with OpenAI / Claude / Gemini / Admin / WebUI routes registered in one tree to avoid multi-entry drift.
-- **Unified execution chain**: Claude/Gemini entries are translated by `internal/translatorcliproxy`, then executed through `openai.ChatCompletions` for shared tool-calling and stream semantics, then translated back to the client protocol.
-- **Cleaner adapter boundaries**: `internal/adapter/{claude,gemini}` handles protocol wrappers, while `internal/adapter/openai` remains the execution core; upstream DeepSeek calls are retained only in the OpenAI core.
-- **Tool-calling parity across runtimes**: Go (`internal/util`) and Vercel Node (`internal/js/helpers/stream-tool-sieve`) follow aligned parsing/anti-leak semantics across JSON / XML / invoke / text-kv inputs.
-- **Config/runtime separation**: static config (`config`) and runtime policy (`settings`) are managed independently via Admin APIs, enabling hot updates and password rotation with JWT invalidation.
-- **Streaming behavior upgrade**: `/v1/responses` and `/v1/chat/completions` now share a more consistent incremental tool-call emission strategy across SDK ecosystems.
-- **Improved operability**: `/healthz`, `/readyz`, `/admin/version`, and `/admin/dev/captures` form a tighter post-deploy diagnostics loop.
-
 ## Key Capabilities
 
 | Capability | Details |
 | --- | --- |
-| OpenAI compatible | `GET /v1/models`, `GET /v1/models/{id}`, `POST /v1/chat/completions`, `POST /v1/responses`, `GET /v1/responses/{response_id}`, `POST /v1/embeddings` |
+| OpenAI compatible | `GET /v1/models`, `GET /v1/models/{id}`, `POST /v1/chat/completions`, `POST /v1/responses`, `GET /v1/responses/{response_id}`, `POST /v1/embeddings`, `POST /v1/files` |
 | Claude compatible | `GET /anthropic/v1/models`, `POST /anthropic/v1/messages`, `POST /anthropic/v1/messages/count_tokens` (plus shortcut paths `/v1/messages`, `/messages`) |
 | Gemini compatible | `POST /v1beta/models/{model}:generateContent`, `POST /v1beta/models/{model}:streamGenerateContent` (plus `/v1/models/{model}:*` paths) |
 | Multi-account rotation | Auto token refresh, email/mobile dual login |
 | Concurrency control | Per-account in-flight limit + waiting queue, dynamic recommended concurrency |
-| DeepSeek PoW | WASM solving via `wazero`, no external Node.js dependency |
+| DeepSeek PoW | Pure Go high-performance solver (DeepSeekHashV1), ms-level response |
 | Tool Calling | Anti-leak handling: non-code-block feature match, early `delta.tool_calls`, structured incremental output |
-| Admin API | Config management, runtime settings hot-reload, account testing/batch test, session cleanup, import/export, Vercel sync, version check |
+| Admin API | Config management, runtime settings hot-reload, proxy management, account testing/batch test, session cleanup, import/export, Vercel sync, version check |
 | WebUI Admin Panel | SPA at `/admin` (bilingual Chinese/English, dark mode) |
 | Health Probes | `GET /healthz` (liveness), `GET /readyz` (readiness) |
 
@@ -111,37 +105,45 @@ flowchart LR
 | P0 | Anthropic SDK (messages) | ✅ |
 | P0 | Google Gemini SDK (generateContent) | ✅ |
 | P1 | LangChain / LlamaIndex / OpenWebUI (OpenAI-compatible integration) | ✅ |
-| P2 | MCP standalone bridge | Planned |
 
 ## Model Support
 
-### OpenAI Endpoint
+### OpenAI Endpoint (`GET /v1/models`)
 
-| Model | thinking | search |
-| --- | --- | --- |
-| `deepseek-chat` | ❌ | ❌ |
-| `deepseek-reasoner` | ✅ | ❌ |
-| `deepseek-chat-search` | ❌ | ✅ |
-| `deepseek-reasoner-search` | ✅ | ✅ |
+| Family | Model ID | thinking | search |
+| --- | --- | --- | --- |
+| default | `deepseek-chat` | ❌ | ❌ |
+| default | `deepseek-reasoner` | ✅ | ❌ |
+| default | `deepseek-chat-search` | ❌ | ✅ |
+| default | `deepseek-reasoner-search` | ✅ | ✅ |
+| expert | `deepseek-expert-chat` | ❌ | ❌ |
+| expert | `deepseek-expert-reasoner` | ✅ | ❌ |
+| expert | `deepseek-expert-chat-search` | ❌ | ✅ |
+| expert | `deepseek-expert-reasoner-search` | ✅ | ✅ |
+| vision | `deepseek-vision-chat` | ❌ | ❌ |
+| vision | `deepseek-vision-reasoner` | ✅ | ❌ |
+| vision | `deepseek-vision-chat-search` | ❌ | ✅ |
+| vision | `deepseek-vision-reasoner-search` | ✅ | ✅ |
 
-### Claude Endpoint
+Besides native IDs, DS2API also accepts common aliases as input (for example `gpt-5`, `gpt-5-mini`, `gpt-5-codex`, `gpt-4.1`, `o3`, `claude-opus-4-6`, `claude-sonnet-4-5`, `gemini-2.5-pro`, `gemini-2.5-flash`), but `/v1/models` returns normalized DeepSeek native model IDs.
 
-| Model | Default Mapping |
+### Claude Endpoint (`GET /anthropic/v1/models`)
+
+| Current common model | Default Mapping |
 | --- | --- |
 | `claude-sonnet-4-5` | `deepseek-chat` |
 | `claude-haiku-4-5` (compatible with `claude-3-5-haiku-latest`) | `deepseek-chat` |
 | `claude-opus-4-6` | `deepseek-reasoner` |
 
 Override mapping via `claude_mapping` or `claude_model_mapping` in config.
-In addition, `/anthropic/v1/models` now includes historical Claude 1.x/2.x/3.x/4.x IDs and common aliases for legacy client compatibility.
-
+Besides the current primary aliases above, `/anthropic/v1/models` also returns Claude 4.x snapshots plus historical 3.x / 2.x / 1.x IDs and common aliases for legacy client compatibility.
 
 #### Claude Code integration pitfalls (validated)
 
 - Set `ANTHROPIC_BASE_URL` to the DS2API root URL (for example `http://127.0.0.1:5001`). Claude Code sends requests to `/v1/messages?beta=true`.
 - `ANTHROPIC_API_KEY` must match an entry in `keys` from `config.json`. Keeping both a regular key and an `sk-ant-*` style key improves client compatibility.
 - If your environment has proxy variables, set `NO_PROXY=127.0.0.1,localhost,<your_host_ip>` for DS2API to avoid proxy interception of local traffic.
-- If tool calls are rendered as plain text and not executed, upgrade to a build that includes multi-format Claude tool-call parsing (JSON/XML/ANTML/invoke).
+- If tool calls are rendered as plain text and not executed, first verify the model output uses supported XML/Markup tool blocks (`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use`) rather than standalone JSON `tool_calls`.
 
 ### Gemini Endpoint
 
@@ -149,6 +151,15 @@ The Gemini adapter maps model names to DeepSeek native models via `model_aliases
 
 ## Quick Start
 
+### Recommended deployment priority
+
+Recommended order when choosing a deployment method:
+
+1. **Download and run release binaries**: the easiest path for most users because the artifacts are already built.
+2. **Docker / GHCR image deployment**: suitable for containerized, orchestrated, or cloud environments.
+3. **Vercel deployment**: suitable if you already use Vercel and accept its platform constraints.
+4. **Run from source / build locally**: suitable for development, debugging, or when you need to modify the code yourself.
+
 ### Universal First Step (all deployment modes)
 
 Use `config.json` as the single source of truth (recommended):
@@ -160,48 +171,39 @@ cp config.example.json config.json
 
 Recommended per deployment mode:
 - Local run: read `config.json` directly
-- Docker / Vercel: generate Base64 from `config.json` and inject as `DS2API_CONFIG_JSON`
-- Compatibility note: `DS2API_CONFIG_JSON` may also contain raw JSON directly; `CONFIG_JSON` is the legacy fallback variable
+- Docker / Vercel: generate Base64 from `config.json` and inject as `DS2API_CONFIG_JSON`, or paste raw JSON directly
 
-### Option 1: Local Run
+### Option 1: Download Release Binaries
 
-**Prerequisites**: Go 1.26+, Node.js 20+ (only if building WebUI locally)
+GitHub Actions automatically builds multi-platform archives on each Release:
 
 ```bash
-# 1. Clone
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
-
-# 2. Configure
+# After downloading the archive for your platform
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 cp config.example.json config.json
-# Edit config.json with your DeepSeek account info and API keys
-
-# 3. Start
-go run ./cmd/ds2api
+# Edit config.json
+./ds2api
 ```
 
-Default URL: `http://localhost:5001`
-
-> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm ci` (only when dependencies are missing) and `npm run build -- --outDir static/admin --emptyOutDir` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
-
-### Option 2: Docker
+### Option 2: Docker / GHCR
 
 ```bash
-# 1. Prepare env file and config file
+# Pull prebuilt image
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
+# Or run a pinned version
+# docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+
+# Prepare env file and config file
 cp .env.example .env
 cp config.example.json config.json
 
-# 2. Edit .env (at least set DS2API_ADMIN_KEY)
-#    DS2API_ADMIN_KEY=replace-with-a-strong-secret
-
-# 3. Start
+# Start with compose
 docker-compose up -d
-
-# 4. View logs
-docker-compose logs -f
 ```
 
-The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, adjust the `ports` mapping.
+The default `docker-compose.yml` uses `ghcr.io/cjackhwang/ds2api:latest` and maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
 
 Rebuild after updates: `docker-compose up -d --build`
 
@@ -237,35 +239,28 @@ base64 < config.json | tr -d '\n'
 
 For detailed deployment instructions, see the [Deployment Guide](docs/DEPLOY.en.md).
 
-### Option 4: Download Release Binaries
+### Option 4: Local Run
 
-GitHub Actions automatically builds multi-platform archives on each Release:
+**Prerequisites**: Go 1.26+, Node.js `20.19+` or `22.12+` (only if building WebUI locally)
 
 ```bash
-# After downloading the archive for your platform
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 1. Clone
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
+
+# 2. Configure
 cp config.example.json config.json
-# Edit config.json
-./ds2api
+# Edit config.json with your DeepSeek account info and API keys
+
+# 3. Start
+go run ./cmd/ds2api
 ```
 
-### Option 5: OpenCode CLI
+Default local URL: `http://127.0.0.1:5001`
 
-1. Copy the example config:
+The server actually binds to `0.0.0.0:5001`, so devices on the same LAN can usually reach it through your private IP as well.
 
-```bash
-cp opencode.json.example opencode.json
-```
-
-2. Edit `opencode.json`:
-- Set `baseURL` to your DS2API endpoint (for example, `https://your-domain.com/v1`)
-- Set `apiKey` to your DS2API key (from `config.keys`)
-
-3. Start OpenCode CLI in the project directory (run `opencode` using your installed method).
-
-> Recommended: use the OpenAI-compatible path (`/v1/*`) via `@ai-sdk/openai-compatible` as shown in the example.
-> If your client supports `wire_api`, test both `responses` and `chat`; DS2API supports both paths.
+> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm ci` (only when dependencies are missing) and `npm run build -- --outDir static/admin --emptyOutDir` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
 
 ## Configuration
 
@@ -286,11 +281,16 @@ cp opencode.json.example opencode.json
   ],
   "model_aliases": {
     "gpt-4o": "deepseek-chat",
+    "gpt-5": "deepseek-chat",
+    "gpt-5-mini": "deepseek-chat",
     "gpt-5-codex": "deepseek-reasoner",
-    "o3": "deepseek-reasoner"
+    "o3": "deepseek-reasoner",
+    "claude-opus-4-6": "deepseek-reasoner",
+    "gemini-2.5-flash": "deepseek-chat"
   },
   "compat": {
-    "wide_input_strict_output": true
+    "wide_input_strict_output": true,
+    "strip_reference_markers": true
   },
   "responses": {
     "store_ttl_seconds": 900
@@ -312,7 +312,7 @@ cp opencode.json.example opencode.json
     "token_refresh_interval_hours": 6
   },
   "auto_delete": {
-    "sessions": false
+    "mode": "none"
   }
 }
 ```
@@ -322,13 +322,14 @@ cp opencode.json.example opencode.json
 - `token`: Even if set in `config.json`, it is cleared during load (DS2API does not read persisted tokens from config); runtime tokens are maintained/refreshed in memory only
 - `model_aliases`: Map common model names (GPT/Codex/Claude) to DeepSeek models
 - `compat.wide_input_strict_output`: Keep `true` (current default policy)
-- `toolcall`: Fixed to feature matching + high-confidence early emit, no longer configurable
+- `compat.strip_reference_markers`: Keep `true`; it strips reference markers from visible output
+- `toolcall`: Legacy field; the current behavior is fixed to feature matching + high-confidence early emit, and any config value is ignored
 - `responses.store_ttl_seconds`: In-memory TTL for `/v1/responses/{id}`
 - `embeddings.provider`: Embeddings provider (`deterministic/mock/builtin` built-in)
 - `claude_mapping`: Maps `fast`/`slow` suffixes to corresponding DeepSeek models (still compatible with `claude_model_mapping`)
 - `admin`: Admin panel settings (JWT expiry, password hash, etc.), hot-reloadable via Admin Settings API
 - `runtime`: Runtime parameters (concurrency limits, queue sizes, managed token refresh interval), hot-reloadable via Admin Settings API; `account_max_queue=0`/`global_max_inflight=0` means auto-calculate from recommended values, `token_refresh_interval_hours=6` is the default forced re-login interval
-- `auto_delete.sessions`: Whether to auto-delete DeepSeek sessions after request completion (default `false`, hot-reloadable via Settings)
+- `auto_delete.mode`: How to clean up DeepSeek remote chat records after each request completes. Supported values: `none` (default, no deletion), `single` (delete only the current session), `all` (delete all sessions); legacy `auto_delete.sessions=true` is still treated as `all`
 
 ### Environment Variables
 
@@ -341,28 +342,23 @@ cp opencode.json.example opencode.json
 | `DS2API_JWT_EXPIRE_HOURS` | Admin JWT TTL in hours | `24` |
 | `DS2API_CONFIG_PATH` | Config file path | `config.json` |
 | `DS2API_CONFIG_JSON` | Inline config (JSON or Base64) | — |
-| `CONFIG_JSON` | Legacy compatibility config input | — |
 | `DS2API_ENV_WRITEBACK` | Auto-write env-backed config to file and transition to file mode (`1/true/yes/on`) | Disabled |
-| `DS2API_WASM_PATH` | PoW WASM file path | Auto-detect |
 | `DS2API_STATIC_ADMIN_DIR` | Admin static assets dir | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | Auto-build WebUI on startup | Enabled locally, disabled on Vercel |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | Max in-flight requests per account | `2` |
-| `DS2API_ACCOUNT_CONCURRENCY` | Alias (legacy compat) | — |
 | `DS2API_ACCOUNT_MAX_QUEUE` | Waiting queue limit | `recommended_concurrency` |
-| `DS2API_ACCOUNT_QUEUE_SIZE` | Alias (legacy compat) | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | Global max in-flight requests | `recommended_concurrency` |
-| `DS2API_MAX_INFLIGHT` | Alias (legacy compat) | — |
 | `DS2API_VERCEL_INTERNAL_SECRET` | Vercel hybrid streaming internal auth | Falls back to `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | Stream lease TTL seconds | `900` |
 | `DS2API_DEV_PACKET_CAPTURE` | Local dev packet capture switch (record recent request/response bodies) | Enabled by default on non-Vercel local runtime |
-| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | Number of captured sessions to retain (auto-evict overflow) | `5` |
-| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | Max recorded bytes per captured response body | `2097152` |
+| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | Number of captured sessions to retain (auto-evict overflow) | `20` |
+| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | Max recorded bytes per captured response body | `5242880` |
 | `VERCEL_TOKEN` | Vercel sync token | — |
 | `VERCEL_PROJECT_ID` | Vercel project ID | — |
 | `VERCEL_TEAM_ID` | Vercel team ID | — |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | Vercel deployment protection bypass for internal Node→Go calls | — |
 
-> Note: when `DS2API_CONFIG_JSON/CONFIG_JSON` is detected, the Admin UI shows mode risk and auto-persistence status (including `DS2API_CONFIG_PATH` and mode-transition hints).
+> Note: when `DS2API_CONFIG_JSON` is detected, the Admin UI shows mode risk and auto-persistence status (including `DS2API_CONFIG_PATH` and mode-transition hints).
 
 ## Authentication Modes
 
@@ -394,21 +390,22 @@ Queue limit = DS2API_ACCOUNT_MAX_QUEUE (default = recommended concurrency)
 When `tools` is present in the request, DS2API performs anti-leak handling:
 
 1. Toolcall feature matching is enabled only in **non-code-block context** (fenced examples are ignored)
-   - In non-code-block context, tool JSON may still be recognized even when mixed with normal prose; surrounding prose can remain as text output.
-2. `responses` streaming strictly uses official item lifecycle events (`response.output_item.*`, `response.content_part.*`, `response.function_call_arguments.*`)
-3. Tool names not declared in the `tools` schema are strictly rejected and will not be emitted as valid tool calls
+2. The parser currently targets XML/Markup-family tool syntax (`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml variants); standalone JSON `tool_calls` payloads are not treated as executable calls by default
+3. `responses` streaming strictly uses official item lifecycle events (`response.output_item.*`, `response.content_part.*`, `response.function_call_arguments.*`)
 4. `responses` supports and enforces `tool_choice` (`auto`/`none`/`required`/forced function); `required` violations return `422` for non-stream and `response.failed` for stream
-5. Valid tool call events are only emitted after passing policy validation, preventing invalid tool names from entering the client execution chain
+5. The output protocol follows the client request (OpenAI / Claude / Gemini native shapes); model-side prompting can prefer XML, and the compatibility layer handles the protocol-specific translation
+
+> Note: the current parser still prioritizes “parse successfully whenever possible”; hard allow-list rejection for undeclared tool names is not enabled yet.
 
 ## Local Dev Packet Capture
 
-This is for debugging issues such as Responses reasoning streaming and tool-call handoff. When enabled, DS2API stores the latest N DeepSeek conversation payload pairs (request body + upstream response body), defaulting to 5 entries with auto-eviction.
+This is for debugging issues such as Responses reasoning streaming and tool-call handoff. When enabled, DS2API stores the latest N DeepSeek conversation payload pairs (request body + upstream response body), defaulting to 20 entries with auto-eviction; each response body is capped at 5 MB by default.
 
 Enable example:
 
 ```bash
 DS2API_DEV_PACKET_CAPTURE=true \
-DS2API_DEV_PACKET_CAPTURE_LIMIT=5 \
+DS2API_DEV_PACKET_CAPTURE_LIMIT=20 \
 go run ./cmd/ds2api
 ```
 
@@ -416,6 +413,8 @@ Inspect/clear (Admin JWT required):
 
 - `GET /admin/dev/captures`: list captured items (newest first)
 - `DELETE /admin/dev/captures`: clear captured items
+- `GET /admin/dev/raw-samples/query?q=keyword&limit=20`: search current in-memory captures by prompt keyword and group `completion + continue` by `chat_session_id`
+- `POST /admin/dev/raw-samples/save`: persist a selected capture chain as `tests/raw_stream_samples/<sample-id>/`
 
 Response fields include:
 
@@ -423,67 +422,10 @@ Response fields include:
 - `response_body`: concatenated raw upstream stream body text
 - `response_truncated`: whether body-size truncation happened
 
-## Project Structure
+The save endpoint can target a chain by `query`, `chain_key`, or `capture_id`. Example:
 
-```text
-ds2api/
-├── app/                     # Unified HTTP handler assembly (shared by local + serverless)
-├── cmd/
-│   ├── ds2api/              # Local / container entrypoint
-│   └── ds2api-tests/        # End-to-end testsuite entrypoint
-├── api/
-│   ├── index.go             # Vercel Serverless Go entry
-│   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # Account pool and concurrency queue
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI adapter (incl. tool call parsing, Vercel stream prepare/release)
-│   │   ├── claude/          # Claude adapter
-│   │   └── gemini/          # Gemini adapter (generateContent / streamGenerateContent)
-│   ├── admin/               # Admin API handlers (incl. Settings hot-reload)
-│   ├── auth/                # Auth and JWT
-│   ├── claudeconv/          # Claude message format conversion
-│   ├── compat/              # Go-version compatibility and regression helpers
-│   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek API client, PoW WASM
-│   ├── js/                  # Node runtime stream/compat logic
-│   ├── devcapture/          # Dev packet capture module
-│   ├── format/              # Output formatting
-│   ├── prompt/              # Prompt construction
-│   ├── server/              # HTTP routing and middleware (chi router)
-│   ├── sse/                 # SSE parsing utilities
-│   ├── stream/              # Unified stream consumption engine
-│   ├── testsuite/           # End-to-end testsuite framework and case orchestration
-│   ├── translatorcliproxy/  # CLIProxy bridge and stream writer components
-│   ├── util/                # Common utilities
-│   ├── version/             # Version parsing/comparison and tag normalization
-│   └── webui/               # WebUI static file serving and auto-build
-├── webui/                   # React WebUI source (Vite + Tailwind)
-│   └── src/
-│       ├── app/             # Routing, auth, config state
-│       ├── features/        # Feature modules (account/settings/vercel/apiTester)
-│       ├── components/      # Shared UI pieces (login/landing, etc.)
-│       └── locales/         # Language packs (zh.json / en.json)
-├── scripts/
-│   └── build-webui.sh       # Manual WebUI build script
-├── tests/
-│   ├── compat/              # Compatibility fixtures and expected outputs
-│   ├── node/                # Node-side unit tests (chat-stream / tool-sieve)
-│   └── scripts/             # Unified test script entrypoints (unit/e2e)
-├── docs/                    # Deployment / contributing / testing docs
-├── static/admin/            # WebUI build output (not committed to Git)
-├── .github/
-│   ├── workflows/           # GitHub Actions (quality gates + release automation)
-│   ├── ISSUE_TEMPLATE/      # Issue templates
-│   └── PULL_REQUEST_TEMPLATE.md
-├── config.example.json      # Config file template
-├── .env.example             # Environment variable template
-├── Dockerfile               # Multi-stage build (WebUI + Go)
-├── docker-compose.yml       # Production Docker Compose
-├── docker-compose.dev.yml   # Development Docker Compose
-├── vercel.json              # Vercel routing and build config
-└── go.mod / go.sum          # Go module dependencies
+```json
+{"query":"Guangzhou weather","sample_id":"gz-weather-from-memory"}
 ```
 
 ## Documentation Index

VERSION

@@ -1 +1 @@
-3.0.0
+3.5.1

app/handler.go

@@ -3,9 +3,17 @@ package app
 import (
 	"net/http"
 
+	"ds2api/internal/config"
 	"ds2api/internal/server"
 )
 
 func NewHandler() http.Handler {
-	return server.NewApp().Router
+	app, err := server.NewApp()
+	if err != nil {
+		config.Logger.Error("[app] init failed", "error", err)
+		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			server.WriteUnhandledError(w, err)
+		})
+	}
+	return app.Router
 }

cmd/ds2api-tests/main.go

@@ -30,8 +30,8 @@ func main() {
 	opts.Timeout = time.Duration(timeoutSeconds) * time.Second
 
 	if err := testsuite.Run(context.Background(), opts); err != nil {
-		fmt.Fprintln(os.Stderr, err.Error())
+		_, _ = fmt.Fprintln(os.Stderr, err.Error())
 		os.Exit(1)
 	}
-	fmt.Fprintln(os.Stdout, "testsuite completed successfully")
+	_, _ = fmt.Fprintln(os.Stdout, "testsuite completed successfully")
 }

cmd/ds2api/main.go

@@ -18,9 +18,17 @@ import (
 )
 
 func main() {
+	if err := config.LoadDotEnv(); err != nil {
+		config.Logger.Warn("[dotenv] load failed", "error", err)
+	}
+	config.RefreshLogger()
 	webui.EnsureBuiltOnStartup()
 	_ = auth.AdminKey()
-	app := server.NewApp()
+	app, err := server.NewApp()
+	if err != nil {
+		config.Logger.Error("server initialization failed", "error", err)
+		os.Exit(1)
+	}
 	port := strings.TrimSpace(os.Getenv("PORT"))
 	if port == "" {
 		port = "5001"

config.example.json

@@ -28,11 +28,8 @@
     "o3": "deepseek-reasoner"
   },
   "compat": {
-    "wide_input_strict_output": true
-  },
-  "toolcall": {
-    "mode": "feature_match",
-    "early_emit_confidence": "high"
+    "wide_input_strict_output": true,
+    "strip_reference_markers": true
   },
   "responses": {
     "store_ttl_seconds": 900
@@ -50,9 +47,10 @@
   "runtime": {
     "account_max_inflight": 2,
     "account_max_queue": 0,
-    "global_max_inflight": 0
+    "global_max_inflight": 0,
+    "token_refresh_interval_hours": 6
   },
   "auto_delete": {
-    "sessions": false
+    "mode": "none"
   }
 }

docker-compose.dev.yml

@@ -16,7 +16,8 @@ services:
     container_name: ds2api-dev
     command: ["go", "run", "./cmd/ds2api"]
     ports:
-      - "${PORT:-5001}:${PORT:-5001}"
+      # Host port is configurable via DS2API_HOST_PORT; container port stays fixed at 5001.
+      - "${DS2API_HOST_PORT:-6011}:5001"
     env_file:
       - .env
     environment:

docker-compose.yml

@@ -1,14 +1,16 @@
 services:
   ds2api:
     image: ghcr.io/cjackhwang/ds2api:latest
-    container_name: ds2api
-    restart: always
-    ports:
-      - "6011:5001"
-    volumes:
-      - ./config.json:/app/config.json    # 配置文件
-      - ./.env:/app/.env                  # 环境变量
-    environment:
-      - TZ=Asia/Shanghai
-      - LOG_LEVEL=INFO
-      - DS2API_ADMIN_KEY=${DS2API_ADMIN_KEY:-ds2api}
+    container_name: ds2api
+    restart: always
+    env_file:
+      - .env
+    ports:
+      # Host port is configurable via DS2API_HOST_PORT; container port stays fixed at 5001.
+      - "${DS2API_HOST_PORT:-6011}:5001"
+    volumes:
+      - ./config.json:/app/config.json    # 配置文件
+    environment:
+      - TZ=Asia/Shanghai
+      - LOG_LEVEL=INFO
+      - DS2API_ADMIN_KEY=${DS2API_ADMIN_KEY:-ds2api}

docs/ARCHITECTURE.en.md

@@ -0,0 +1,136 @@
+# DS2API Architecture & Project Layout
+
+Language: [中文](ARCHITECTURE.md) | [English](ARCHITECTURE.en.md)
+
+> This file is the single architecture source for directory layout, module boundaries, and execution flow.
+
+## 1. Top-level Layout (expanded)
+
+> Notes: this is the **fully expanded** project directory list (excluding metadata/dependency dirs such as `.git/` and `webui/node_modules/`), with each folder annotated by purpose.
+
+```text
+ds2api/
+├── .github/                              # GitHub collaboration and CI config
+│   ├── ISSUE_TEMPLATE/                   # Issue templates
+│   └── workflows/                        # GitHub Actions workflows
+├── api/                                  # Serverless entrypoints (Vercel Go/Node)
+├── app/                                  # Application-level handler assembly
+├── cmd/                                  # Executable entrypoints
+│   ├── ds2api/                           # Main service bootstrap
+│   └── ds2api-tests/                     # E2E testsuite CLI bootstrap
+├── docs/                                 # Project documentation
+├── internal/                             # Core implementation (non-public packages)
+│   ├── account/                          # Account pool, inflight slots, waiting queue
+│   ├── adapter/                          # Multi-protocol adapters
+│   │   ├── claude/                       # Claude protocol adapter
+│   │   ├── gemini/                       # Gemini protocol adapter
+│   │   └── openai/                       # OpenAI adapter and shared execution core
+│   ├── admin/                            # Admin API (config/accounts/ops)
+│   ├── auth/                             # Auth/JWT/credential resolution
+│   ├── claudeconv/                       # Claude message conversion helpers
+│   ├── compat/                           # Compatibility and regression helpers
+│   ├── config/                           # Config loading/validation/hot reload
+│   ├── deepseek/                         # DeepSeek upstream client capabilities
+│   │   └── transport/                    # DeepSeek transport details
+│   ├── devcapture/                       # Dev capture and troubleshooting
+│   ├── format/                           # Response formatting layer
+│   │   ├── claude/                       # Claude output formatting
+│   │   └── openai/                       # OpenAI output formatting
+│   ├── js/                               # Node runtime related logic
+│   │   ├── chat-stream/                  # Node streaming bridge
+│   │   ├── helpers/                      # JS helper modules
+│   │   │   └── stream-tool-sieve/        # JS implementation of tool sieve
+│   │   └── shared/                       # Shared semantics between Go/Node
+│   ├── prompt/                           # Prompt composition
+│   ├── rawsample/                        # Raw sample read/write and management
+│   ├── server/                           # Router and middleware assembly
+│   ├── sse/                              # SSE parsing utilities
+│   ├── stream/                           # Unified stream consumption engine
+│   ├── testsuite/                        # Testsuite execution framework
+│   ├── textclean/                        # Text cleanup
+│   ├── toolcall/                         # Tool-call parsing and repair
+│   ├── translatorcliproxy/               # Cross-protocol translation bridge
+│   ├── util/                             # Shared utility helpers
+│   ├── version/                          # Version query/compare
+│   └── webui/                            # WebUI static hosting logic
+├── plans/                                # Stage plans and manual QA records
+├── pow/                                  # PoW standalone implementation + benchmarks
+├── scripts/                              # Build/release helper scripts
+├── tests/                                # Test assets and scripts
+│   ├── compat/                           # Compatibility fixtures + expected outputs
+│   │   ├── expected/                     # Expected output samples
+│   │   └── fixtures/                     # Fixture inputs
+│   │       ├── sse_chunks/               # SSE chunk fixtures
+│   │       └── toolcalls/                # Tool-call fixtures
+│   ├── node/                             # Node unit tests
+│   ├── raw_stream_samples/               # Upstream raw SSE samples
+│   │   ├── content-filter-trigger-20260405-jwt3/          # Content-filter terminal sample
+│   │   ├── continue-thinking-snapshot-replay-20260405/    # Continue-thinking sample
+│   │   ├── guangzhou-weather-reasoner-search-20260404/    # Search/reference sample
+│   │   ├── markdown-format-example-20260405/              # Markdown sample
+│   │   └── markdown-format-example-20260405-spacefix/     # Space-fix sample
+│   ├── scripts/                          # Test entry scripts
+│   └── tools/                            # Testing helper tools
+└── webui/                                # React admin console source
+    ├── public/                           # Static assets
+    └── src/                              # Frontend source code
+        ├── app/                          # Routing/state scaffolding
+        ├── components/                   # Shared UI components
+        ├── features/                     # Feature modules
+        │   ├── account/                  # Account management page
+        │   ├── apiTester/                # API tester page
+        │   ├── settings/                 # Settings page
+        │   └── vercel/                   # Vercel sync page
+        ├── layout/                       # Layout components
+        ├── locales/                      # i18n strings
+        └── utils/                        # Frontend utilities
+```
+
+## 2. Primary Request Flow
+
+```mermaid
+flowchart LR
+    C[Client/SDK] --> R[internal/server/router.go]
+    R --> OA[OpenAI Adapter]
+    R --> CA[Claude Adapter]
+    R --> GA[Gemini Adapter]
+    R --> AD[Admin API]
+
+    CA --> BR[translatorcliproxy]
+    GA --> BR
+    BR --> CORE[internal/adapter/openai ChatCompletions]
+    OA --> CORE
+
+    CORE --> AUTH[internal/auth + config key/account resolver]
+    CORE --> POOL[internal/account queue + concurrency]
+    CORE --> TOOL[internal/toolcall parser + sieve]
+    CORE --> DS[internal/deepseek client]
+    DS --> U[DeepSeek upstream]
+```
+
+## 3. Responsibilities in `internal/`
+
+- `internal/server`: router tree + middlewares (health, protocol routes, Admin/WebUI).
+- `internal/adapter/openai`: shared execution core (chat/responses/embeddings + tool semantics).
+- `internal/adapter/{claude,gemini}`: protocol wrappers only (no duplicated upstream execution).
+- `internal/translatorcliproxy`: structure translation between Claude/Gemini and OpenAI.
+- `internal/deepseek`: upstream request/session/PoW/SSE handling.
+- `internal/stream` + `internal/sse`: stream parsing and incremental assembly.
+- `internal/toolcall`: XML/Markup-family tool-call parsing + anti-leak sieve (`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml variants).
+- `internal/admin`: config/accounts/vercel sync/version/dev-capture endpoints.
+- `internal/config`: config loading/validation + runtime settings hot-reload.
+- `internal/account`: managed account pool, inflight slots, waiting queue.
+
+## 4. WebUI Runtime Relation
+
+- `webui/` stores frontend source (Vite + React).
+- Runtime serves static output from `static/admin`.
+- On first local startup, if `static/admin` is missing, DS2API may auto-build it (Node.js required).
+
+## 5. Documentation Split Strategy
+
+- Onboarding & quick start: `README.MD` / `README.en.md`
+- Architecture & layout: `docs/ARCHITECTURE*.md` (this file)
+- API contracts: `API.md` / `API.en.md`
+- Deployment/testing/contributing: `docs/DEPLOY*`, `docs/TESTING.md`, `docs/CONTRIBUTING*`
+- Deep topics: `docs/toolcall-semantics.md`, `docs/DeepSeekSSE行为结构说明-2026-04-05.md`

docs/ARCHITECTURE.md

@@ -0,0 +1,136 @@
+# DS2API 架构与项目结构说明
+
+语言 / Language: [中文](ARCHITECTURE.md) | [English](ARCHITECTURE.en.md)
+
+> 本文档用于集中维护“代码目录结构 + 模块边界 + 主链路调用关系”。
+
+## 1. 顶层目录结构（展开）
+
+> 说明：以下为仓库内业务相关目录的**完整展开**（排除 `.git/` 与 `webui/node_modules/` 这类依赖/元数据目录），并标注每个文件夹作用。
+
+```text
+ds2api/
+├── .github/                              # GitHub 协作与 CI 配置
+│   ├── ISSUE_TEMPLATE/                   # Issue 模板
+│   └── workflows/                        # GitHub Actions 工作流
+├── api/                                  # Serverless 入口（Vercel Go/Node）
+├── app/                                  # 应用级 handler 装配层
+├── cmd/                                  # 可执行程序入口
+│   ├── ds2api/                           # 主服务启动入口
+│   └── ds2api-tests/                     # E2E 测试集 CLI 入口
+├── docs/                                 # 项目文档目录
+├── internal/                             # 核心业务实现（不对外暴露）
+│   ├── account/                          # 账号池、并发槽位、等待队列
+│   ├── adapter/                          # 多协议适配层
+│   │   ├── claude/                       # Claude 协议适配
+│   │   ├── gemini/                       # Gemini 协议适配
+│   │   └── openai/                       # OpenAI 协议与统一执行核心
+│   ├── admin/                            # Admin API（配置/账号/运维）
+│   ├── auth/                             # 鉴权/JWT/凭证解析
+│   ├── claudeconv/                       # Claude 消息格式转换工具
+│   ├── compat/                           # 兼容性辅助与回归支持
+│   ├── config/                           # 配置加载、校验、热更新
+│   ├── deepseek/                         # DeepSeek 上游客户端能力
+│   │   └── transport/                    # DeepSeek 传输层细节
+│   ├── devcapture/                       # 开发抓包与调试采集
+│   ├── format/                           # 响应格式化层
+│   │   ├── claude/                       # Claude 输出格式化
+│   │   └── openai/                       # OpenAI 输出格式化
+│   ├── js/                               # Node Runtime 相关逻辑
+│   │   ├── chat-stream/                  # Node 流式输出桥接
+│   │   ├── helpers/                      # JS 辅助函数
+│   │   │   └── stream-tool-sieve/        # Tool sieve JS 实现
+│   │   └── shared/                       # Go/Node 共用语义片段
+│   ├── prompt/                           # Prompt 组装
+│   ├── rawsample/                        # raw sample 读写与管理
+│   ├── server/                           # 路由与中间件装配
+│   ├── sse/                              # SSE 解析工具
+│   ├── stream/                           # 统一流式消费引擎
+│   ├── testsuite/                        # 测试集执行框架
+│   ├── textclean/                        # 文本清洗
+│   ├── toolcall/                         # 工具调用解析与修复
+│   ├── translatorcliproxy/               # 多协议互转桥
+│   ├── util/                             # 通用工具函数
+│   ├── version/                          # 版本查询/比较
+│   └── webui/                            # WebUI 静态托管相关逻辑
+├── plans/                                # 阶段计划与人工验收记录
+├── pow/                                  # PoW 独立实现与基准
+├── scripts/                              # 构建/发布/辅助脚本
+├── tests/                                # 测试资源与脚本
+│   ├── compat/                           # 兼容性夹具与期望输出
+│   │   ├── expected/                     # 预期结果样本
+│   │   └── fixtures/                     # 测试输入夹具
+│   │       ├── sse_chunks/               # SSE chunk 夹具
+│   │       └── toolcalls/                # toolcall 夹具
+│   ├── node/                             # Node 单元测试
+│   ├── raw_stream_samples/               # 上游原始 SSE 样本
+│   │   ├── content-filter-trigger-20260405-jwt3/          # 风控终态样本
+│   │   ├── continue-thinking-snapshot-replay-20260405/    # continue 样本
+│   │   ├── guangzhou-weather-reasoner-search-20260404/    # 搜索+引用样本
+│   │   ├── markdown-format-example-20260405/              # Markdown 样本
+│   │   └── markdown-format-example-20260405-spacefix/     # 空格修复样本
+│   ├── scripts/                          # 测试脚本入口
+│   └── tools/                            # 测试辅助工具
+└── webui/                                # React 管理台源码
+    ├── public/                           # 静态资源
+    └── src/                              # 前端源码
+        ├── app/                          # 路由/状态框架
+        ├── components/                   # 共享组件
+        ├── features/                     # 功能模块
+        │   ├── account/                  # 账号管理页面
+        │   ├── apiTester/                # API 测试页面
+        │   ├── settings/                 # 设置页面
+        │   └── vercel/                   # Vercel 同步页面
+        ├── layout/                       # 布局组件
+        ├── locales/                      # 国际化文案
+        └── utils/                        # 前端工具函数
+```
+
+## 2. 请求主链路
+
+```mermaid
+flowchart LR
+    C[Client/SDK] --> R[internal/server/router.go]
+    R --> OA[OpenAI Adapter]
+    R --> CA[Claude Adapter]
+    R --> GA[Gemini Adapter]
+    R --> AD[Admin API]
+
+    CA --> BR[translatorcliproxy]
+    GA --> BR
+    BR --> CORE[internal/adapter/openai ChatCompletions]
+    OA --> CORE
+
+    CORE --> AUTH[internal/auth + config key/account resolver]
+    CORE --> POOL[internal/account queue + concurrency]
+    CORE --> TOOL[internal/toolcall parser + sieve]
+    CORE --> DS[internal/deepseek client]
+    DS --> U[DeepSeek upstream]
+```
+
+## 3. internal/ 子模块职责
+
+- `internal/server`：路由树和中间件挂载（健康检查、协议入口、Admin/WebUI）。
+- `internal/adapter/openai`：统一执行内核（chat/responses/embeddings 与 tool calling 语义）。
+- `internal/adapter/{claude,gemini}`：协议输入输出适配，不重复实现上游调用逻辑。
+- `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转。
+- `internal/deepseek`：上游请求、会话、PoW、SSE 消费。
+- `internal/stream` + `internal/sse`：流式解析与增量处理。
+- `internal/toolcall`：以 XML/Markup 家族为核心的工具调用解析与防泄漏筛分（`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml 变体）。
+- `internal/admin`：配置管理、账号管理、Vercel 同步、版本检查、开发抓包。
+- `internal/config`：配置加载、校验、运行时 settings 热更新。
+- `internal/account`：托管账号池、并发槽位、等待队列。
+
+## 4. WebUI 与运行时关系
+
+- `webui/` 是前端源码（Vite + React）。
+- 运行时托管目录是 `static/admin`（构建产物）。
+- 本地首次启动若 `static/admin` 缺失，会尝试自动构建（依赖 Node.js）。
+
+## 5. 文档拆分策略
+
+- 总览与快速开始：`README.MD` / `README.en.md`
+- 架构与目录：`docs/ARCHITECTURE*.md`（本文件）
+- 接口协议：`API.md` / `API.en.md`
+- 部署、测试、贡献：`docs/DEPLOY*`、`docs/TESTING.md`、`docs/CONTRIBUTING*`
+- 专题：`docs/toolcall-semantics.md`、`docs/DeepSeekSSE行为结构说明-2026-04-05.md`

docs/CONTRIBUTING.en.md

@@ -9,7 +9,7 @@ Thanks for your interest in contributing to DS2API!
 ### Prerequisites
 
 - Go 1.26+
-- Node.js 20+ (for WebUI development)
+- Node.js `20.19+` or `22.12+` (for WebUI development)
 - npm (bundled with Node.js)
 
 ### Backend Development
@@ -25,7 +25,8 @@ cp config.example.json config.json
 
 # 3. Run backend
 go run ./cmd/ds2api
-# Default: http://localhost:5001
+# Local access: http://127.0.0.1:5001
+# Actual bind: 0.0.0.0:5001, so LAN access is available via your private IP
 ```
 
 ### Frontend Development (WebUI)
@@ -40,6 +41,7 @@ npm install
 # 3. Start dev server (hot reload)
 npm run dev
 # Default: http://localhost:5173, auto-proxies API to backend
+# host: 0.0.0.0 is not configured, so LAN access is not enabled by default
 ```
 
 WebUI tech stack:
@@ -57,7 +59,7 @@ docker-compose -f docker-compose.dev.yml up
 
 | Language | Standards |
 | --- | --- |
-| **Go** | Run `gofmt` and ensure `go test ./...` passes before committing |
+| **Go** | Run `./scripts/lint.sh` (gofmt + golangci-lint) and ensure `go test ./...` passes before committing |
 | **JavaScript/React** | Follow existing project style (functional components) |
 | **Commit messages** | Use semantic prefixes: `feat:`, `fix:`, `docs:`, `refactor:`, `style:`, `perf:`, `chore:` |
 
@@ -92,58 +94,12 @@ Manually build WebUI to `static/admin/`:
 
 ## Project Structure
 
-```text
-ds2api/
-├── app/                     # Shared HTTP handler assembly (local + serverless)
-├── cmd/
-│   ├── ds2api/              # Local/container entrypoint
-│   └── ds2api-tests/        # End-to-end testsuite entrypoint
-├── api/
-│   ├── index.go             # Vercel Serverless Go entry
-│   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # Account pool and concurrency queue
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI adapter
-│   │   ├── claude/          # Claude adapter
-│   │   └── gemini/          # Gemini adapter
-│   ├── admin/               # Admin API handlers
-│   ├── auth/                # Auth and JWT
-│   ├── claudeconv/          # Claude message conversion
-│   ├── compat/              # Go-version compatibility and regression helpers
-│   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek client, PoW WASM
-│   ├── js/                  # Node runtime stream/compat logic
-│   ├── devcapture/          # Dev packet capture
-│   ├── format/              # Output formatting
-│   ├── prompt/              # Prompt building
-│   ├── server/              # HTTP routing (chi router)
-│   ├── sse/                 # SSE parsing utilities
-│   ├── stream/              # Unified stream consumption engine
-│   ├── testsuite/           # Testsuite framework and scenario orchestration
-│   ├── translatorcliproxy/  # CLIProxy bridge and stream writer
-│   ├── util/                # Common utilities
-│   ├── version/             # Version parsing and comparison
-│   └── webui/               # WebUI static hosting
-├── webui/                   # React WebUI source
-│   └── src/
-│       ├── app/             # Routing, auth, config state
-│       ├── features/        # Feature modules
-│       ├── components/      # Shared components
-│       └── locales/         # Language packs
-├── scripts/                 # Build and test scripts
-├── tests/
-│   ├── compat/              # Compatibility fixtures and expected outputs
-│   ├── node/                # Node-side unit tests
-│   └── scripts/             # Test script entrypoints (unit/e2e)
-├── plans/                   # Plans, gates, and manual smoke-test records
-├── static/admin/            # WebUI build output (not committed)
-├── Dockerfile               # Multi-stage build
-├── docker-compose.yml       # Production
-├── docker-compose.dev.yml   # Development
-└── vercel.json              # Vercel config
-```
+To avoid documentation drift, directory layout and module responsibilities were moved to:
+
+- [docs/ARCHITECTURE.en.md](./ARCHITECTURE.en.md)
+- [docs/README.md](./README.md)
+
+Before contributing, review the architecture doc sections for request flow and `internal/` module boundaries.
 
 ## Reporting Issues
 

docs/CONTRIBUTING.md

@@ -9,7 +9,7 @@
 ### 前置要求
 
 - Go 1.26+
-- Node.js 20+（WebUI 开发时）
+- Node.js `20.19+` 或 `22.12+`（WebUI 开发时）
 - npm（随 Node.js 提供）
 
 ### 后端开发
@@ -25,7 +25,8 @@ cp config.example.json config.json
 
 # 3. 启动后端
 go run ./cmd/ds2api
-# 默认监听 http://localhost:5001
+# 本地访问 http://127.0.0.1:5001
+# 实际绑定 0.0.0.0:5001，可通过局域网 IP 访问
 ```
 
 ### 前端开发（WebUI）
@@ -40,6 +41,7 @@ npm install
 # 3. 启动开发服务器（热更新）
 npm run dev
 # 默认监听 http://localhost:5173，自动代理 API 到后端
+# 当前未配置 host: 0.0.0.0，因此默认不对局域网开放
 ```
 
 WebUI 技术栈：
@@ -57,7 +59,7 @@ docker-compose -f docker-compose.dev.yml up
 
 | 语言 | 规范 |
 | --- | --- |
-| **Go** | 提交前运行 `gofmt`，确保 `go test ./...` 通过 |
+| **Go** | 提交前运行 `./scripts/lint.sh`（包含 gofmt+golangci-lint）并确保 `go test ./...` 通过 |
 | **JavaScript/React** | 保持现有代码风格（函数组件） |
 | **提交信息** | 使用语义化前缀：`feat:`、`fix:`、`docs:`、`refactor:`、`style:`、`perf:`、`chore:` |
 
@@ -92,58 +94,12 @@ docker-compose -f docker-compose.dev.yml up
 
 ## 项目结构
 
-```text
-ds2api/
-├── app/                     # 统一 HTTP Handler 装配（本地 + Serverless）
-├── cmd/
-│   ├── ds2api/              # 本地/容器启动入口
-│   └── ds2api-tests/        # 端到端测试集入口
-├── api/
-│   ├── index.go             # Vercel Serverless Go 入口
-│   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # 账号池与并发队列
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI 兼容适配器
-│   │   ├── claude/          # Claude 兼容适配器
-│   │   └── gemini/          # Gemini 兼容适配器
-│   ├── admin/               # Admin API handlers
-│   ├── auth/                # 鉴权与 JWT
-│   ├── claudeconv/          # Claude 消息格式转换
-│   ├── compat/              # Go 版本兼容与回归测试辅助
-│   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek 客户端、PoW WASM
-│   ├── js/                  # Node 运行时流式/兼容逻辑
-│   ├── devcapture/          # 开发抓包
-│   ├── format/              # 输出格式化
-│   ├── prompt/              # Prompt 构建
-│   ├── server/              # HTTP 路由（chi router）
-│   ├── sse/                 # SSE 解析工具
-│   ├── stream/              # 统一流式消费引擎
-│   ├── testsuite/           # 测试集框架与场景编排
-│   ├── translatorcliproxy/  # CLIProxy 桥接与流式写入
-│   ├── util/                # 通用工具
-│   ├── version/             # 版本解析与比较
-│   └── webui/               # WebUI 静态托管
-├── webui/                   # React WebUI 源码
-│   └── src/
-│       ├── app/             # 路由、鉴权、配置状态
-│       ├── features/        # 业务功能模块
-│       ├── components/      # 通用组件
-│       └── locales/         # 语言包
-├── scripts/                 # 构建与测试脚本
-├── tests/
-│   ├── compat/              # 兼容夹具与期望输出
-│   ├── node/                # Node 侧单元测试
-│   └── scripts/             # 测试脚本入口（unit/e2e）
-├── plans/                   # 计划、门禁和手工烟测记录
-├── static/admin/            # WebUI 构建产物（不提交）
-├── Dockerfile               # 多阶段构建
-├── docker-compose.yml       # 生产环境
-├── docker-compose.dev.yml   # 开发环境
-└── vercel.json              # Vercel 配置
-```
+为避免与其他文档重复维护，目录结构与模块职责已迁移到：
+
+- [docs/ARCHITECTURE.md](./ARCHITECTURE.md)
+- [docs/README.md](./README.md)
+
+贡献前建议先阅读架构文档中的“请求主链路”和 `internal/` 模块职责，再定位改动范围。
 
 ## 问题反馈
 

docs/DEPLOY.en.md

@@ -4,15 +4,18 @@ Language: [中文](DEPLOY.md) | [English](DEPLOY.en.md)
 
 This guide covers all deployment methods for the current Go-based codebase.
 
+Doc map: [Index](./README.md) | [Architecture](./ARCHITECTURE.en.md) | [API](../API.en.md) | [Testing](./TESTING.md)
+
 ---
 
 ## Table of Contents
 
+- [Recommended deployment priority](#recommended-deployment-priority)
 - [Prerequisites](#0-prerequisites)
-- [1. Local Run](#1-local-run)
-- [2. Docker Deployment](#2-docker-deployment)
+- [1. Download Release Binaries](#1-download-release-binaries)
+- [2. Docker / GHCR Deployment](#2-docker--ghcr-deployment)
 - [3. Vercel Deployment](#3-vercel-deployment)
-- [4. Download Release Binaries](#4-download-release-binaries)
+- [4. Local Run from Source](#4-local-run-from-source)
 - [5. Reverse Proxy (Nginx)](#5-reverse-proxy-nginx)
 - [6. Linux systemd Service](#6-linux-systemd-service)
 - [7. Post-Deploy Checks](#7-post-deploy-checks)
@@ -20,19 +23,29 @@ This guide covers all deployment methods for the current Go-based codebase.
 
 ---
 
+## Recommended deployment priority
+
+Recommended order when choosing a deployment method:
+
+1. **Download and run release binaries**: the easiest path for most users because the artifacts are already built.
+2. **Docker / GHCR image deployment**: suitable for containerized, orchestrated, or cloud environments.
+3. **Vercel deployment**: suitable if you already use Vercel and accept its platform constraints.
+4. **Run from source / build locally**: suitable for development, debugging, or when you need to modify the code yourself.
+
+---
+
 ## 0. Prerequisites
 
 | Dependency | Minimum Version | Notes |
 | --- | --- | --- |
 | Go | 1.26+ | Build backend |
-| Node.js | 20+ | Only needed to build WebUI locally |
+| Node.js | `20.19+` or `22.12+` | Only needed to build WebUI locally |
 | npm | Bundled with Node.js | Install WebUI dependencies |
 
 Config source (choose one):
 
 - **File**: `config.json` (recommended for local/Docker)
 - **Environment variable**: `DS2API_CONFIG_JSON` (recommended for Vercel; supports raw JSON or Base64)
-- Compatibility note: `CONFIG_JSON` is the legacy fallback variable; `DS2API_CONFIG_JSON` may also contain raw JSON directly
 
 Unified recommendation (best practice):
 
@@ -47,76 +60,67 @@ Use `config.json` as the single source of truth:
 
 ---
 
-## 1. Local Run
+## 1. Download Release Binaries
 
-### 1.1 Basic Steps
+Built-in GitHub Actions workflow: `.github/workflows/release-artifacts.yml`
+
+- **Trigger**: only on Release `published` (no build on normal push)
+- **Outputs**: multi-platform binary archives + `sha256sums.txt`
+- **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
+
+| Platform | Architecture | Format |
+| --- | --- | --- |
+| Linux | amd64, arm64 | `.tar.gz` |
+| macOS | amd64, arm64 | `.tar.gz` |
+| Windows | amd64 | `.zip` |
+
+Each archive includes:
+
+- `ds2api` executable (`ds2api.exe` on Windows)
+- `static/admin/` (built WebUI assets)
+- `config.example.json`, `.env.example`
+- `README.MD`, `README.en.md`, `LICENSE`
+
+### Usage
 
 ```bash
-# Clone
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
+# 1. Download the archive for your platform
+# 2. Extract
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 
-# Copy and edit config
+# 3. Configure
 cp config.example.json config.json
-# Open config.json and fill in:
-#   - keys: your API access keys
-#   - accounts: DeepSeek accounts (email or mobile + password)
+# Edit config.json
 
-# Start
-go run ./cmd/ds2api
-```
-
-Default address: `http://0.0.0.0:5001` (override with `PORT`).
-
-### 1.2 WebUI Build
-
-On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm; when dependencies are missing it runs `npm ci` first, then `npm run build -- --outDir static/admin --emptyOutDir`).
-
-Manual build:
-
-```bash
-./scripts/build-webui.sh
-```
-
-Or step by step:
-
-```bash
-cd webui
-npm install
-npm run build
-# Output goes to static/admin/
-```
-
-Control auto-build via environment variable:
-
-```bash
-# Disable auto-build
-DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
-
-# Force enable auto-build
-DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
-```
-
-### 1.3 Compile to Binary
-
-```bash
-go build -o ds2api ./cmd/ds2api
+# 4. Start
 ./ds2api
 ```
 
+### Maintainer Release Flow
+
+1. Create and publish a GitHub Release (with tag, for example `vX.Y.Z`)
+2. Wait for the `Release Artifacts` workflow to complete
+3. Download the matching archive from Release Assets
+
 ---
 
-## 2. Docker Deployment
+## 2. Docker / GHCR Deployment
 
 ### 2.1 Basic Steps
 
 ```bash
+# Pull prebuilt image
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
 # Copy env template and config file
 cp .env.example .env
 cp config.example.json config.json
 
 # Edit .env and set at least:
 #   DS2API_ADMIN_KEY=your-admin-key
+# Optionally set the host port:
+#   DS2API_HOST_PORT=6011
 
 # Start
 docker-compose up -d
@@ -125,7 +129,13 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
-The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, adjust the `ports` mapping.
+The default `docker-compose.yml` directly uses `ghcr.io/cjackhwang/ds2api:latest` and maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
+
+If you want a pinned version instead of `latest`, you can also pull a specific tag directly:
+
+```bash
+docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+```
 
 ### 2.2 Update
 
@@ -138,7 +148,7 @@ docker-compose up -d --build
 The `Dockerfile` now provides two image paths:
 
 1. **Default local/dev path (`runtime-from-source`)**: a three-stage build (WebUI build + Go build + runtime).
-2. **Release path (`runtime-from-dist`)**: CI first creates `dist/ds2api_<tag>_linux_<arch>.tar.gz`, then Docker directly reuses the binary and `static/admin` assets from those release archives, without running `npm build`/`go build` again.
+2. **Release path (`runtime-from-dist`)**: the release workflow first creates tag-named release archives, then copies the Linux bundles to `dist/docker-input/linux_amd64.tar.gz` / `linux_arm64.tar.gz`; Docker consumes those prepared inputs directly, without rerunning `npm build`/`go build`.
 
 The release path keeps Docker images aligned with release archives and reduces duplicate build work.
 
@@ -198,10 +208,10 @@ Notes:
 2. **Import** the project on Vercel
 3. **Set environment variables** (minimum required: one variable):
 
-   | Variable | Description |
-   | --- | --- |
-   | `DS2API_ADMIN_KEY` | Admin key (required) |
-   | `DS2API_CONFIG_JSON` | Config content, raw JSON or Base64 (optional, recommended) |
+| Variable | Description |
+| --- | --- |
+| `DS2API_ADMIN_KEY` | Admin key (required) |
+| `DS2API_CONFIG_JSON` | Config content, raw JSON or Base64 (optional, recommended) |
 
 4. **Deploy**
 
@@ -244,11 +254,8 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # optional for personal accounts
 | Variable | Description | Default |
 | --- | --- | --- |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | Per-account inflight limit | `2` |
-| `DS2API_ACCOUNT_CONCURRENCY` | Alias (legacy compat) | — |
 | `DS2API_ACCOUNT_MAX_QUEUE` | Waiting queue limit | `recommended_concurrency` |
-| `DS2API_ACCOUNT_QUEUE_SIZE` | Alias (legacy compat) | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | Global inflight limit | `recommended_concurrency` |
-| `DS2API_MAX_INFLIGHT` | Alias (legacy compat) | — |
 | `DS2API_ENV_WRITEBACK` | When `DS2API_CONFIG_JSON` is present, auto-write to `DS2API_CONFIG_PATH` and switch to file-backed mode after success (`1/true/yes/on`) | Disabled |
 | `DS2API_VERCEL_INTERNAL_SECRET` | Hybrid streaming internal auth | Falls back to `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | Stream lease TTL | `900` |
@@ -314,7 +321,7 @@ Error: Command failed: go build -ldflags -s -w -o .../bootstrap ...
 1. Open Vercel Project Settings → Build and Development Settings
 2. **Clear** custom Go Build Flags / Build Command (recommended)
 3. If ldflags must be used, set `-ldflags="-s -w"` (ensure it's one argument)
-4. Verify `go.mod` uses a supported version (currently `go 1.24`)
+4. Verify `go.mod` uses a supported version (currently `go 1.26.0`)
 5. Redeploy (recommended: clear cache)
 
 #### Internal Package Import Error
@@ -350,58 +357,61 @@ If API responses return Vercel HTML `Authentication Required`:
 
 ---
 
-## 4. Download Release Binaries
+## 4. Local Run from Source
 
-Built-in GitHub Actions workflow: `.github/workflows/release-artifacts.yml`
-
-- **Trigger**: only on Release `published` (no build on normal push)
-- **Outputs**: multi-platform binary archives + `sha256sums.txt`
-- **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
-
-| Platform | Architecture | Format |
-| --- | --- | --- |
-| Linux | amd64, arm64 | `.tar.gz` |
-| macOS | amd64, arm64 | `.tar.gz` |
-| Windows | amd64 | `.zip` |
-
-Each archive includes:
-
-- `ds2api` executable (`ds2api.exe` on Windows)
-- `static/admin/` (built WebUI assets)
-- `sha3_wasm_bg.7b9ca65ddd.wasm` (optional; binary has embedded fallback)
-- `config.example.json`, `.env.example`
-- `README.MD`, `README.en.md`, `LICENSE`
-
-### Usage
+### 4.1 Basic Steps
 
 ```bash
-# 1. Download the archive for your platform
-# 2. Extract
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# Clone
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
 
-# 3. Configure
+# Copy and edit config
 cp config.example.json config.json
-# Edit config.json
+# Open config.json and fill in:
+#   - keys: your API access keys
+#   - accounts: DeepSeek accounts (email or mobile + password)
 
-# 4. Start
-./ds2api
+# Start
+go run ./cmd/ds2api
 ```
 
-### Maintainer Release Flow
+Default local access URL: `http://127.0.0.1:5001`; the server actually binds to `0.0.0.0:5001` (override with `PORT`).
 
-1. Create and publish a GitHub Release (with tag, for example `vX.Y.Z`)
-2. Wait for the `Release Artifacts` workflow to complete
-3. Download the matching archive from Release Assets
+### 4.2 WebUI Build
 
-### Pull from GHCR (Optional)
+On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm; when dependencies are missing it runs `npm ci` first, then `npm run build -- --outDir static/admin --emptyOutDir`).
+
+Manual build:
 
 ```bash
-# latest
-docker pull ghcr.io/cjackhwang/ds2api:latest
+./scripts/build-webui.sh
+```
 
-# specific version (example)
-docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+Or step by step:
+
+```bash
+cd webui
+npm install
+npm run build
+# Output goes to static/admin/
+```
+
+Control auto-build via environment variable:
+
+```bash
+# Disable auto-build
+DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
+
+# Force enable auto-build
+DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
+```
+
+### 4.3 Compile to Binary
+
+```bash
+go build -o ds2api ./cmd/ds2api
+./ds2api
 ```
 
 ---
@@ -458,8 +468,6 @@ server {
 # Copy compiled binary and related files to target directory
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# Optional: if you want to use an external WASM file (override the embedded one, from a release package or build output)
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DEPLOY.md

@@ -4,15 +4,18 @@
 
 本指南基于当前 Go 代码库，详细说明各种部署方式。
 
+本页导航：[文档总索引](./README.md)｜[架构说明](./ARCHITECTURE.md)｜[接口文档](../API.md)｜[测试指南](./TESTING.md)
+
 ---
 
 ## 目录
 
+- [部署方式优先级建议](#部署方式优先级建议)
 - [前置要求](#0-前置要求)
-- [一、本地运行](#一本地运行)
-- [二、Docker 部署](#二docker-部署)
+- [一、下载 Release 构建包](#一下载-release-构建包)
+- [二、Docker / GHCR 部署](#二docker--ghcr-部署)
 - [三、Vercel 部署](#三vercel-部署)
-- [四、下载 Release 构建包](#四下载-release-构建包)
+- [四、本地源码运行](#四本地源码运行)
 - [五、反向代理（Nginx）](#五反向代理nginx)
 - [六、Linux systemd 服务化](#六linux-systemd-服务化)
 - [七、部署后检查](#七部署后检查)
@@ -20,19 +23,29 @@
 
 ---
 
+## 部署方式优先级建议
+
+推荐按以下顺序选择部署方式：
+
+1. **下载 Release 构建包运行**：最省事，产物已编译完成，最适合大多数用户。
+2. **Docker / GHCR 镜像部署**：适合需要容器化、编排或云环境部署。
+3. **Vercel 部署**：适合已有 Vercel 环境且接受其平台约束的场景。
+4. **本地源码运行 / 自行编译**：适合开发、调试或需要自行修改代码的场景。
+
+---
+
 ## 0. 前置要求
 
 | 依赖 | 最低版本 | 说明 |
 | --- | --- | --- |
 | Go | 1.26+ | 编译后端 |
-| Node.js | 20+ | 仅在需要本地构建 WebUI 时 |
+| Node.js | `20.19+` 或 `22.12+` | 仅在需要本地构建 WebUI 时 |
 | npm | 随 Node.js 提供 | 安装 WebUI 依赖 |
 
 配置来源（任选其一）：
 
 - **文件方式**：`config.json`（推荐本地/Docker 使用）
-- **环境变量方式**：`DS2API_CONFIG_JSON`（推荐 Vercel 使用，支持 JSON 字符串或 Base64 编码）
-- 兼容写法：`CONFIG_JSON` 是旧版回退变量；`DS2API_CONFIG_JSON` 也可以直接写原始 JSON
+- **环境变量方式**：`DS2API_CONFIG_JSON`（推荐 Vercel 使用，支持 JSON 字符串或 Base64 编码，也可以直接写原始 JSON）
 
 统一建议（最优实践）：
 
@@ -47,76 +60,67 @@ cp config.example.json config.json
 
 ---
 
-## 一、本地运行
+## 一、下载 Release 构建包
 
-### 1.1 基本步骤
+仓库内置 GitHub Actions 工作流：`.github/workflows/release-artifacts.yml`
+
+- **触发条件**：仅在 Release `published` 时触发（普通 push 不会构建）
+- **构建产物**：多平台二进制压缩包 + `sha256sums.txt`
+- **容器镜像发布**：仅发布到 GHCR（`ghcr.io/cjackhwang/ds2api`）
+
+| 平台 | 架构 | 文件格式 |
+| --- | --- | --- |
+| Linux | amd64, arm64 | `.tar.gz` |
+| macOS | amd64, arm64 | `.tar.gz` |
+| Windows | amd64 | `.zip` |
+
+每个压缩包包含：
+
+- `ds2api` 可执行文件（Windows 为 `ds2api.exe`）
+- `static/admin/`（WebUI 构建产物）
+- `config.example.json`、`.env.example`
+- `README.MD`、`README.en.md`、`LICENSE`
+
+### 使用步骤
 
 ```bash
-# 克隆仓库
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
+# 1. 下载对应平台的压缩包
+# 2. 解压
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 
-# 复制并编辑配置
+# 3. 配置
 cp config.example.json config.json
-# 使用你喜欢的编辑器打开 config.json，填入：
-#   - keys: 你的 API 访问密钥
-#   - accounts: DeepSeek 账号（email 或 mobile + password）
+# 编辑 config.json
 
-# 启动服务
-go run ./cmd/ds2api
-```
-
-默认监听 `http://0.0.0.0:5001`，可通过 `PORT` 环境变量覆盖。
-
-### 1.2 WebUI 构建
-
-本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm；缺依赖时会先执行 `npm ci`，再执行 `npm run build -- --outDir static/admin --emptyOutDir`）。
-
-你也可以手动构建：
-
-```bash
-./scripts/build-webui.sh
-```
-
-或手动执行：
-
-```bash
-cd webui
-npm install
-npm run build
-# 产物输出到 static/admin/
-```
-
-通过环境变量控制自动构建行为：
-
-```bash
-# 强制关闭自动构建
-DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
-
-# 强制开启自动构建
-DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
-```
-
-### 1.3 编译为二进制文件
-
-```bash
-go build -o ds2api ./cmd/ds2api
+# 4. 启动
 ./ds2api
 ```
 
+### 维护者发布步骤
+
+1. 在 GitHub 创建并发布 Release（带 tag，如 `vX.Y.Z`）
+2. 等待 Actions 工作流 `Release Artifacts` 完成
+3. 在 Release 的 Assets 下载对应平台压缩包
+
 ---
 
-## 二、Docker 部署
+## 二、Docker / GHCR 部署
 
 ### 2.1 基本步骤
 
 ```bash
+# 拉取预编译镜像
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
 # 复制环境变量模板和配置文件
 cp .env.example .env
 cp config.example.json config.json
 
 # 编辑 .env（请改成你的强密码），至少设置：
 #   DS2API_ADMIN_KEY=your-admin-key
+# 如需修改宿主机端口，可额外设置：
+#   DS2API_HOST_PORT=6011
 
 # 启动
 docker-compose up -d
@@ -125,7 +129,13 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
-默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请调整 `ports` 配置。
+默认 `docker-compose.yml` 直接使用 `ghcr.io/cjackhwang/ds2api:latest`，并把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请设置 `DS2API_HOST_PORT=5001`（或者手动调整 `ports` 配置）。
+
+如需固定版本，也可以直接拉取指定 tag：
+
+```bash
+docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+```
 
 ### 2.2 更新
 
@@ -138,7 +148,7 @@ docker-compose up -d --build
 `Dockerfile` 提供两条构建路径：
 
 1. **本地/开发默认路径（`runtime-from-source`）**：三阶段构建（WebUI 构建 + Go 构建 + 运行阶段）。
-2. **Release 路径（`runtime-from-dist`）**：CI 先生成 `dist/ds2api_<tag>_linux_<arch>.tar.gz`，再由 Docker 直接复用该发布包内的二进制和 `static/admin` 产物组装运行镜像，不再重复执行 `npm build`/`go build`。
+2. **Release 路径（`runtime-from-dist`）**：发布工作流先生成 tag 命名的 Release 压缩包，再把 Linux 产物复制成 `dist/docker-input/linux_amd64.tar.gz` / `linux_arm64.tar.gz`；Docker 构建阶段直接消费这些输入，不再重复执行 `npm build`/`go build`。
 
 Release 路径可确保 Docker 镜像与 release 压缩包使用同一套产物，减少重复构建带来的差异。
 
@@ -198,10 +208,10 @@ healthcheck:
 2. **在 Vercel 上导入项目**
 3. **配置环境变量**（最少只需设置以下一项）：
 
-   | 变量 | 说明 |
-   | --- | --- |
-   | `DS2API_ADMIN_KEY` | 管理密钥（必填） |
-   | `DS2API_CONFIG_JSON` | 配置内容，JSON 字符串或 Base64 编码（可选，建议） |
+| 变量 | 说明 |
+| --- | --- |
+| `DS2API_ADMIN_KEY` | 管理密钥（必填） |
+| `DS2API_CONFIG_JSON` | 配置内容，JSON 字符串或 Base64 编码（可选，建议） |
 
 4. **部署**
 
@@ -244,11 +254,8 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # 个人账号可留空
 | 变量 | 说明 | 默认值 |
 | --- | --- | --- |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | 每账号并发上限 | `2` |
-| `DS2API_ACCOUNT_CONCURRENCY` | 同上（兼容别名） | — |
 | `DS2API_ACCOUNT_MAX_QUEUE` | 等待队列上限 | `recommended_concurrency` |
-| `DS2API_ACCOUNT_QUEUE_SIZE` | 同上（兼容别名） | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | 全局并发上限 | `recommended_concurrency` |
-| `DS2API_MAX_INFLIGHT` | 同上（兼容别名） | — |
 | `DS2API_ENV_WRITEBACK` | 检测到 `DS2API_CONFIG_JSON` 时自动写入 `DS2API_CONFIG_PATH`，并在成功后转为文件模式（`1/true/yes/on`） | 关闭 |
 | `DS2API_VERCEL_INTERNAL_SECRET` | 混合流式内部鉴权 | 回退用 `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | 流式 lease TTL | `900` |
@@ -314,7 +321,7 @@ Error: Command failed: go build -ldflags -s -w -o .../bootstrap ...
 1. 进入 Vercel Project Settings → Build and Development Settings
 2. **清空**自定义 Go Build Flags / Build Command（推荐）
 3. 若必须设置 ldflags，使用 `-ldflags="-s -w"`（保证它是一个参数）
-4. 确认仓库 `go.mod` 为受支持版本（当前为 `go 1.24`）
+4. 确认仓库 `go.mod` 为受支持版本（当前为 `go 1.26.0`）
 5. 重新部署（建议清缓存后 Redeploy）
 
 #### Internal 包导入错误
@@ -350,58 +357,61 @@ No Output Directory named "public" found after the Build completed.
 
 ---
 
-## 四、下载 Release 构建包
+## 四、本地源码运行
 
-仓库内置 GitHub Actions 工作流：`.github/workflows/release-artifacts.yml`
-
-- **触发条件**：仅在 Release `published` 时触发（普通 push 不会构建）
-- **构建产物**：多平台二进制压缩包 + `sha256sums.txt`
-- **容器镜像发布**：仅发布到 GHCR（`ghcr.io/cjackhwang/ds2api`）
-
-| 平台 | 架构 | 文件格式 |
-| --- | --- | --- |
-| Linux | amd64, arm64 | `.tar.gz` |
-| macOS | amd64, arm64 | `.tar.gz` |
-| Windows | amd64 | `.zip` |
-
-每个压缩包包含：
-
-- `ds2api` 可执行文件（Windows 为 `ds2api.exe`）
-- `static/admin/`（WebUI 构建产物）
-- `sha3_wasm_bg.7b9ca65ddd.wasm`（可选；程序内置 embed fallback）
-- `config.example.json`、`.env.example`
-- `README.MD`、`README.en.md`、`LICENSE`
-
-### 使用步骤
+### 4.1 基本步骤
 
 ```bash
-# 1. 下载对应平台的压缩包
-# 2. 解压
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 克隆仓库
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
 
-# 3. 配置
+# 复制并编辑配置
 cp config.example.json config.json
-# 编辑 config.json
+# 使用你喜欢的编辑器打开 config.json，填入：
+#   - keys: 你的 API 访问密钥
+#   - accounts: DeepSeek 账号（email 或 mobile + password）
 
-# 4. 启动
-./ds2api
+# 启动服务
+go run ./cmd/ds2api
 ```
 
-### 维护者发布步骤
+默认本地访问地址是 `http://127.0.0.1:5001`；服务实际绑定 `0.0.0.0:5001`，可通过 `PORT` 环境变量覆盖。
 
-1. 在 GitHub 创建并发布 Release（带 tag，如 `vX.Y.Z`）
-2. 等待 Actions 工作流 `Release Artifacts` 完成
-3. 在 Release 的 Assets 下载对应平台压缩包
+### 4.2 WebUI 构建
 
-### 拉取 GHCR 镜像（可选）
+本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm；缺依赖时会先执行 `npm ci`，再执行 `npm run build -- --outDir static/admin --emptyOutDir`）。
+
+你也可以手动构建：
 
 ```bash
-# latest
-docker pull ghcr.io/cjackhwang/ds2api:latest
+./scripts/build-webui.sh
+```
 
-# 指定版本（示例）
-docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+或手动执行：
+
+```bash
+cd webui
+npm install
+npm run build
+# 产物输出到 static/admin/
+```
+
+通过环境变量控制自动构建行为：
+
+```bash
+# 强制关闭自动构建
+DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
+
+# 强制开启自动构建
+DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
+```
+
+### 4.3 编译为二进制文件
+
+```bash
+go build -o ds2api ./cmd/ds2api
+./ds2api
 ```
 
 ---
@@ -458,8 +468,6 @@ server {
 # 将编译好的二进制文件和相关文件复制到目标目录
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# 可选：若你希望使用外置 WASM 文件（覆盖内置版本，来自 release 包或构建产物）
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DeepSeekSSE流格式字段分析-2026-04-03.md

@@ -1,82 +0,0 @@
-# DeepSeek SSE 流格式字段分析（2026-04-03）
-
-> 日期：2026-04-03（UTC）
-> 
-> 样本：`tests/raw_stream_samples/guangzhou-weather-reasoner-search-20260403/upstream.stream.sse`
-> 
-> 模型：`deepseek-reasoner-search`（搜索 + 思考）
-
-## 1. SSE 事件层结构
-
-原始流由标准 SSE 帧组成，常见形态：
-
-```text
-event: <type>
-data: <json or text>
-
-```
-
-样本中主要 `event` 类型：
-
-- `ready`：流建立后返回请求/响应消息 ID。
-- `update_session`：会话时间戳更新。
-- `finish`：流式阶段结束。
-- （无 `event` 时）默认为 message 事件，`data:` 中承载主要增量数据。
-
-## 2. `data` JSON 常见字段
-
-上游增量主体多为 JSON Patch 风格对象：
-
-- `p`（path）：字段路径，如 `response/fragments/-1/content`。
-- `o`（op，可选）：操作类型，常见 `SET` / `APPEND` / `BATCH`。
-- `v`（value）：值（字符串、布尔、对象、数组都可能）。
-
-示例（语义）：
-
-- `{"p":"response/fragments/-1/content","o":"APPEND","v":"..."}`
-- `{"p":"response/fragments/-16/status","v":"FINISHED"}`
-- `{"p":"response/status","o":"SET","v":"FINISHED"}`
-
-## 3. 搜索+思考场景关键路径
-
-### 3.1 文本内容
-
-- `response/fragments/<idx>/content`
-- `response/content`
-- `response/thinking_content`
-- `response/fragments`（`APPEND` + fragment 数组）
-
-### 3.2 搜索相关
-
-- `response/fragments/<idx>/results`（检索结果数组）
-- `response/search_status`（检索状态，建议跳过展示）
-
-### 3.3 状态相关（重点）
-
-- `response/status = FINISHED`：**最终结束信号**（需要保留用于结束判定）
-- `response/fragments/<idx>/status = FINISHED`：**分片级状态**（高频，建议跳过输出）
-- `response/quasi_status`：过程状态（建议跳过输出）
-
-## 4. 泄露问题根因（FINISHED 重复）
-
-在搜索 + 思考模型中，`response/fragments/<idx>/status` 会出现大量不同 `<idx>`（例如 `-1/-2/-3/-16...`）的 `FINISHED`。
-
-若只过滤固定少量索引（例如仅 `-1/-2/-3`），其他索引的状态会当普通文本透传，导致前端出现：
-
-- `FINISHEDFINISHEDFINISHED...`
-
-## 5. 适配建议（已落地）
-
-1. 跳过所有 `response/fragments/-?\d+/status`。
-2. 继续保留 `response/status=FINISHED` 作为真正结束判定。
-3. 通过独立仿真工具持续回放全部样本，作为回归门禁：
-
-```bash
-./tests/scripts/run-raw-stream-sim.sh
-```
-
-## 6. 后续扩展建议
-
-- 增加不同模型（`deepseek-chat-search` / 非 search / 非 thinking）样本。
-- 增加异常样本（限流、中断、content_filter、空结果）。
-- 为仿真报告加入字段覆盖率统计（路径频次、事件频次、终止路径命中率）。

docs/DeepSeekSSE行为结构说明-2026-04-05.md

@@ -0,0 +1,315 @@
+# DeepSeek SSE 行为结构说明（第三方逆向版）
+
+> 说明：本文基于当前仓库 `tests/raw_stream_samples/` 下全部 `upstream.stream.sse` 原始流样本整理而成，属于第三方逆向观察文档，不是官方协议。
+> 当前 corpus 由 4 份原始流组成，覆盖搜索+引用、风控终态、Markdown 输出和空格敏感输出等行为。
+> 补充：文末还会注明少量“当前实现已确认、但 corpus 尚未完整覆盖”的行为，例如长思考场景下的自动续写状态。
+
+文档导航：[文档总索引](./README.md) / [测试指南](./TESTING.md) / [样本目录说明](../tests/raw_stream_samples/README.md)
+
+## 1. 样本覆盖
+
+下列样本共同构成了本文的观察基础：
+
+| 样本 | 观察重点 |
+| --- | --- |
+| [guangzhou-weather-reasoner-search-20260404](../tests/raw_stream_samples/guangzhou-weather-reasoner-search-20260404/upstream.stream.sse) | 搜索+思考流程，包含 `reference:N` 引用标记与工具片段 |
+| [content-filter-trigger-20260405-jwt3](../tests/raw_stream_samples/content-filter-trigger-20260405-jwt3/upstream.stream.sse) | `CONTENT_FILTER` 终态分支，包含拒答模板与 `ban_regenerate` |
+| [markdown-format-example-20260405](../tests/raw_stream_samples/markdown-format-example-20260405/upstream.stream.sse) | Markdown 输出的早期样本，用于观察 token 级输出形态 |
+| [markdown-format-example-20260405-spacefix](../tests/raw_stream_samples/markdown-format-example-20260405-spacefix/upstream.stream.sse) | Markdown 输出修正样本，用于验证空格 chunk 必须保留 |
+
+当前 corpus 的整体特征是 `message` 帧占绝对多数，控制事件只占很小一部分，但它们决定了流的生命周期和最终状态。
+
+## 2. 总体结构
+
+DeepSeek 的这类输出可以分成两层看：
+
+1. SSE 事件层。
+2. JSON 载荷层。
+
+事件层负责传输边界，载荷层负责业务状态。实现时不要把 HTTP chunk、SSE block 和业务 JSON 混为一体。
+
+最常见的时序可以概括为：
+
+```text
+ready
+update_session
+message(初始化 envelope)
+message(正文 / 片段 / 状态增量)
+message(状态收口)
+finish
+update_session
+title
+close
+```
+
+`finish` 表示生成流结束，但不是唯一的终止信号；真正的语义终态通常还要结合 `response/status`、`quasi_status` 和 `close` 一起判断。
+
+## 3. SSE 事件层
+
+当前 corpus 中观察到的事件类型如下：
+
+| 事件 | 作用 | 处理建议 |
+| --- | --- | --- |
+| `ready` | 传输层就绪，通常携带 `request_message_id`、`response_message_id`、`model_type` | 记录元数据即可，不参与正文拼接 |
+| `update_session` | 会话时间戳或心跳更新 | 当作会话状态帧处理 |
+| `message` | 主体载荷，绝大多数业务信息都在这里 | 必须按顺序解析并保序累积 |
+| `finish` | 生成阶段结束 | 作为流结束标记之一 |
+| `title` | 会话标题生成结果 | 元数据帧，不参与正文拼接 |
+| `close` | 连接关闭信息 | 仅用于收尾与审计 |
+
+说明：
+
+- `message` 是默认事件名，SSE 中没有显式 `event:` 时也应按 `message` 处理。
+- 目前样本里大量 `message` 帧没有独立的业务前缀，不能靠事件名区分正文和控制帧。
+- 可能出现空 payload 的 `message` 帧；它们应被视为 no-op，但不能打乱事件顺序。
+
+## 4. 载荷层形态
+
+`message` 的 `data:` 部分不是单一 schema，而是多种结构混合。当前 corpus 里主要见到以下几种形态：
+
+| 形态 | 典型结构 | 作用 |
+| --- | --- | --- |
+| 初始化 envelope | `{"v":{"response":{...}}}` | 给出会话初始状态、模型状态和片段容器 |
+| 纯文本 token | `{"v":"..."}` | 直接输出可见文本 token |
+| 路径补丁 | `{"p":"...","o":"APPEND|SET|BATCH","v":...}` | 对某个状态路径做增量更新 |
+| 终态 batch | `{"v":[{"p":"status","v":"CONTENT_FILTER"}, ...]}` | 尾部状态收口，常见于风控终态 |
+
+一个简化后的典型样式如下：
+
+```json
+{"v":"输出"}
+{"p":"response/fragments/-1/content","o":"APPEND","v":"..."}
+{"p":"response/fragments","o":"APPEND","v":[...]}
+{"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":211},{"p":"quasi_status","v":"FINISHED"}]}
+{"p":"response/status","o":"SET","v":"FINISHED"}
+```
+
+注意：
+
+- `v` 可能是字符串、对象、数组、布尔值或数字。
+- `o` 当前样本里主要见到 `APPEND`、`SET`、`BATCH`。
+- `v` 为数组时，通常表示一个批量 patch 集合，而不是正文数组。
+
+## 5. 初始化 envelope
+
+每条流开头，常会先出现一个 `message` 帧，内容是完整的 `response` 初始状态。当前 corpus 中，这个 envelope 常见字段包括：
+
+- `message_id`
+- `parent_id`
+- `model`
+- `role`
+- `thinking_enabled`
+- `ban_edit`
+- `ban_regenerate`
+- `status`
+- `incomplete_message`
+- `accumulated_token_usage`
+- `files`
+- `feedback`
+- `inserted_at`
+- `search_enabled`
+- `fragments`
+- `conversation_mode`
+- `has_pending_fragment`
+- `auto_continue`
+- `search_triggered`
+
+这些字段更像会话状态和策略开关，不是正文内容。第三方实现应把它们保留在内部状态树里，而不是直接拼接到最终答案。
+
+## 6. 路径结构
+
+当前 corpus 里观察到的 `p` 路径可以归成几组：
+
+### 6.1 片段级路径
+
+- `response/fragments/-N/content`
+- `response/fragments/-N/status`
+- `response/fragments/-N/results`
+- `response/fragments/-N/elapsed_secs`
+
+这类路径表示某个片段对象的增量更新。`-N` 只是样本中的索引风格，不应被写死成固定数量。
+
+### 6.2 片段容器路径
+
+- `response/fragments`
+- `fragments`
+
+这两类路径通常承载 fragment 数组。前者更像响应树中的分支，后者更像终态批处理里的片段集合。
+
+### 6.3 语义状态路径
+
+- `response/status`
+- `response/has_pending_fragment`
+- `quasi_status`
+- `status`
+- `ban_regenerate`
+
+这类路径决定流是否结束、是否被风控、是否还有待处理片段。它们不应作为正文输出。
+
+尤其是 `response/status` / `status` 这类路径上的字符串值，应被视为控制信号而不是文本 token。当前已确认需要特殊对待的值包括：
+
+- `FINISHED`：正常完成终态，应触发收口。
+- `CONTENT_FILTER`：风控终态，应走拒答/模板分支。
+- `WIP` / `INCOMPLETE` / `AUTO_CONTINUE`：未完成但可继续生成的中间状态，不应直接输出给客户端。
+
+### 6.4 统计与进度路径
+
+- `accumulated_token_usage`
+
+这类路径用于使用量或进度统计，属于元数据。
+
+### 6.5 非命名空间字段
+
+在片段对象内部，还会看到 `content`、`references`、`result`、`queries`、`stage_id` 等字段。它们不一定带 `response/...` 前缀，但仍然是协议语义的一部分。
+
+## 7. fragment 类型
+
+当前 corpus 里已经观察到的 fragment 类型如下：
+
+| 类型 | 作用 | 是否应直接渲染 |
+| --- | --- | --- |
+| `RESPONSE` | 正常回答片段 | 是，属于正文 |
+| `THINK` | 推理或阶段提示 | 通常否，按产品策略决定是否展示 |
+| `TOOL_SEARCH` | 搜索工具调用元数据 | 否 |
+| `TOOL_OPEN` | 打开 / 抽取结果的工具元数据 | 否 |
+| `TIP` | 提示 / 警告类片段，常带 `style: WARNING` | 视产品策略决定，通常作为附注 |
+| `TEMPLATE_RESPONSE` | 风控拒答模板 | 是，但它属于终态 fallback，不是普通正文 |
+
+观察到的典型片段字段：
+
+- `id`
+- `type`
+- `content`
+- `references`
+- `stage_id`
+- `status`
+- `queries`
+- `results`
+- `result`
+- `elapsed_secs`
+- `style`
+- `hide_on_wip`
+
+第三方实现不要把 `fragment.type` 和 `p` 路径混为一谈。`type` 是语义分类，`p` 是状态树位置。
+
+## 8. 终态行为
+
+当前 corpus 里有两条很重要的终态分支。
+
+### 8.1 正常完成
+
+正常回答通常会出现如下收口顺序：
+
+1. `response` 的 `BATCH` 更新 `accumulated_token_usage`。
+2. `response` 的 `BATCH` 或单独 patch 更新 `quasi_status: FINISHED`。
+3. `response/status` 置为 `FINISHED`。
+4. `finish` 事件到来。
+5. 之后可能还有 `update_session`、`title`、`close`。
+
+### 8.2 风控终态
+
+`content-filter-trigger-20260405-jwt3` 展示了另一种终态路径：
+
+1. 先继续输出一段正常正文。
+2. 出现提示类 fragment，例如 `TIP`。
+3. 可能先把 `quasi_status` 提前收口到 `FINISHED`。
+4. 之后出现一个终态 batch，把 `ban_regenerate` 设为 `true`，把 `status` 置为 `CONTENT_FILTER`，并附带 `TEMPLATE_RESPONSE`。
+5. 最后再出现 `finish`，然后是收尾事件。
+
+这个分支说明：
+
+- `finish` 不等于正常结束。
+- `CONTENT_FILTER` 是一个独立终态，不是普通异常。
+- `TEMPLATE_RESPONSE` 不应被当作常规回答流的中间片段，它是终态 fallback。
+
+一个简化的风控尾部可以写成：
+
+```json
+{"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":1269},{"p":"quasi_status","v":"FINISHED"}]}
+{"v":[{"p":"ban_regenerate","v":true},{"p":"status","v":"CONTENT_FILTER"},{"p":"fragments","v":[{"id":38,"type":"TEMPLATE_RESPONSE","content":"..."}]},{"p":"quasi_status","v":"CONTENT_FILTER"}]}
+{"event":"finish"}
+```
+
+### 8.3 自动续写中间态（实现补充）
+
+这部分不是当前 corpus 的直接覆盖项，而是 2026-04-05 在长思考实测中观察到、且已在当前实现中兼容的行为：
+
+1. 上游可能先把 `response/status` 或 envelope 内的 `response.status` 置为 `WIP` / `INCOMPLETE`。
+2. 有时还会伴随 `auto_continue: true`。
+3. 这表示当前轮输出尚未真正结束，客户端或代理层可以继续调用 continue 接口续写同一条回答。
+4. 续写后的内容会承接之前的思考与正文，不应把前一轮状态值泄露成可见文本。
+
+对第三方实现，建议把这一类状态统一当作“可继续的控制信号”：
+
+- 可以据此决定是否继续拉取后续流。
+- 不能把 `INCOMPLETE`、`WIP`、`AUTO_CONTINUE` 直接拼接到最终文本。
+- `finish` 事件本身也不能单独说明回答已完全结束，仍要结合状态字段判断。
+
+## 9. 文本重建规则
+
+如果你的目标是把流重建成最终可见文本，必须遵守下面这些规则：
+
+- 按接收顺序逐个追加 token。
+- 不要对每个 `v` 做 `trim` 或 `TrimSpace`。
+- 不要丢弃只包含空格的 chunk。
+- 不要合并连续空格、换行或 Markdown 符号附近的空白。
+- 不要把 `[reference:N]` 视为协议元数据，它在当前 corpus 里就是正文的一部分。
+- 如果你要屏蔽引用标记，应当把它做成可配置的后处理，而不是在解析阶段硬删。
+- `response/status` / `status` 路径上的状态字符串不应进入正文，即使它们不是终态。
+
+这点对 Markdown、代码块、引用、表格都很关键。样本里已经证明，`#`、`-`、`>`、`|` 这类符号后面的空格必须原样保留，否则渲染结果会变形。
+
+## 10. 推荐实现方式
+
+对第三方开发者，建议把实现拆成三条线：
+
+1. 原始事件线：保留 SSE block 顺序、事件名和完整 JSON 载荷。
+2. 状态树线：维护 `response`、`fragments`、`status`、`quasi_status` 等结构。
+3. 可见文本线：只从明确应渲染的 token / fragment 中拼接最终文本。
+
+一个简单的处理顺序可以是：
+
+```text
+parse SSE block
+  -> 识别 event
+  -> 解析 JSON payload
+  -> 更新状态树
+  -> 识别 status / quasi_status / auto_continue 等控制信号
+  -> 判定是否有可见文本
+  -> 追加到输出缓冲
+  -> 遇到 WIP / INCOMPLETE / AUTO_CONTINUE 时决定是否续写
+  -> 遇到 FINISHED / CONTENT_FILTER / finish 时收口
+```
+
+实现时的兼容原则：
+
+- 未知路径保留，不要报错中断。
+- 未知 fragment.type 保留在日志里。
+- 不要假设所有模型都一定输出 `thinking_content`，当前 corpus 的推理更多是通过 fragment 类型表达。
+- 不要假设 `title` 一定存在，它只是后置元数据。
+
+## 11. 本 corpus 证明了什么
+
+当前样本足以证明以下行为：
+
+- 搜索类模型会把工具调用、结果、引用和正文混在同一条 SSE 流里。
+- 风控不会简单地“没有输出”，而是会在正常生成后切换到 `CONTENT_FILTER` 终态。
+- Markdown 和代码输出对空格非常敏感，空格 chunk 不能吞。
+- `message` 是主体承载层，`ready` / `update_session` / `finish` / `title` / `close` 是控制层。
+- `fragment.type` 是可视化和工具链分层的关键，不应只靠 `p` 路径判断。
+
+结合 2026-04-05 的长思考实测，还可以补充一条当前实现层面的结论：
+
+- 长思考场景下，上游可能先给出 `INCOMPLETE` / `WIP` / `AUTO_CONTINUE` 状态，再通过 continue 链路续写；这些状态值本身不应作为正文输出。
+
+## 12. 适用边界
+
+本文是基于当前 corpus 的逆向说明，不是恒定协议。
+
+- 新模型可能增加新的 `p` 路径。
+- 新版本可能增加新的 fragment.type。
+- `CONTENT_FILTER` 的终态模板内容可能变化。
+- 自动续写相关状态（如 `INCOMPLETE` / `AUTO_CONTINUE`）当前主要来自实测与实现兼容逻辑，后续字段形态仍可能变化。
+- 解析器应当对未知字段、未知路径、未知事件保持容忍。
+
+如果你要把这份说明用于实际开发，建议同时保留原始流样本、回放脚本和回归测试，不要只依赖本文。

docs/README.md

@@ -0,0 +1,53 @@
+# DS2API 文档导航 | Documentation Index
+
+语言 / Language: [中文](README.md) | [English](README.md#english)
+
+## 中文
+
+为减少重复维护，本仓库文档按“入口文档 + 专题文档”拆分。建议从下列顺序阅读：
+
+1. [项目总览（README）](../README.MD)
+2. [架构与目录说明](./ARCHITECTURE.md)
+3. [接口文档（API）](../API.md)
+4. [部署指南](./DEPLOY.md)
+5. [测试指南](./TESTING.md)
+6. [贡献指南](./CONTRIBUTING.md)
+
+### 专题文档
+
+- [Tool Calling 统一语义](./toolcall-semantics.md)
+- [DeepSeek SSE 行为结构说明（逆向观察）](./DeepSeekSSE行为结构说明-2026-04-05.md)
+
+### 文档维护约定
+
+- `README.MD` / `README.en.md`：面向首次接触用户，保留“是什么 + 怎么快速跑起来”。
+- `docs/ARCHITECTURE*.md`：面向开发者，集中维护项目结构、模块职责与调用链。
+- `API*.md`：面向客户端接入者，聚焦接口行为、鉴权和示例。
+- 其他 `docs/*.md`：主题化说明，避免在多个文档重复粘贴同一段内容。
+
+---
+
+## English
+
+To reduce maintenance drift, docs are split into an “entry doc + topical docs” layout.
+
+Recommended reading order:
+
+1. [Project overview (README)](../README.en.md)
+2. [Architecture and project layout](./ARCHITECTURE.en.md)
+3. [API reference](../API.en.md)
+4. [Deployment guide](./DEPLOY.en.md)
+5. [Testing guide](./TESTING.md)
+6. [Contributing guide](./CONTRIBUTING.en.md)
+
+### Topical docs
+
+- [Tool-calling unified semantics](./toolcall-semantics.md)
+- [DeepSeek SSE behavior notes (reverse-engineered)](./DeepSeekSSE行为结构说明-2026-04-05.md)
+
+### Maintenance conventions
+
+- `README.MD` / `README.en.md`: onboarding-oriented (“what + quick start”).
+- `docs/ARCHITECTURE*.md`: developer-oriented source of truth for module boundaries and execution flow.
+- `API*.md`: integration-oriented behavior/contracts.
+- Other `docs/*.md`: focused topics, avoid copy-pasting the same section into multiple files.

docs/TESTING.md

@@ -2,6 +2,8 @@
 
 语言 / Language: 中文 + English（同页）
 
+文档导航： [总览](../README.MD) / [架构说明](./ARCHITECTURE.md) / [部署指南](./DEPLOY.md) / [接口文档](../API.md)
+
 ## 概述 | Overview
 
 DS2API 提供两个层级的测试：
@@ -180,10 +182,10 @@ go test ./...
 
 ```bash
 # 运行 tool calls 相关测试（推荐用于调试 tool call 解析问题）
-go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/toolcall/
 
 # 运行单个测试用例
-go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/util/
+go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/toolcall/
 
 # 运行 format 相关测试
 go test -v ./internal/format/...
@@ -198,13 +200,13 @@ go test -v ./internal/adapter/openai/...
 
 ```bash
 # 1. 运行 tool calls 相关的所有测试
-go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/toolcall/
 
 # 2. 查看测试输出中的详细调试信息
-go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/util/ 2>&1
+go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/toolcall/ 2>&1
 
 # 3. 检查具体测试用例的修复效果
-# 测试用例位于 internal/util/toolcalls_test.go，包含：
+# 测试用例位于 internal/toolcall/toolcalls_test.go，包含：
 # - TestParseToolCallsWithDeepSeekHallucination: DeepSeek 典型幻觉输出
 # - TestRepairLooseJSONWithNestedObjects: 嵌套对象的方括号修复
 # - TestParseToolCallsWithMixedWindowsPaths: Windows 路径处理
@@ -233,9 +235,48 @@ go run ./cmd/ds2api-tests --no-preflight
 ```
 
 说明：
-- 该工具会重放 `tests/raw_stream_samples` 下全部样本，按上游 SSE 顺序做 1:1 仿真解析。
+- 该工具默认重放 `tests/raw_stream_samples/manifest.json` 声明的 canonical 样本，按上游 SSE 顺序做 1:1 仿真解析。
 - 默认校验不出现 `FINISHED` 文本泄露，并要求存在结束信号。
-- 结果会写入 `artifacts/raw-stream-sim/*.json`，可供其他测试脚本或排障流程复用。
+- 默认**不**把 `raw accumulated_token_usage` 与本地解析 token 做强一致校验（当前实现以内容估算为准）；如需强校验可显式加 `--fail-on-token-mismatch`。
+- 每次运行都会把本地派生结果写入 `artifacts/raw-stream-sim/<run-id>/<sample-id>/replay.output.txt`，并输出结构化报告。
+- 如果你有历史基线目录，可以通过 `--baseline-root` 让工具直接做文本对比。
+- 更完整的协议级行为结构说明见 [DeepSeekSSE行为结构说明-2026-04-05.md](./DeepSeekSSE行为结构说明-2026-04-05.md)。
+
+### 对单个样本做回放比对
+
+```bash
+./tests/scripts/compare-raw-stream-sample.sh markdown-format-example-20260405-spacefix
+```
+
+说明：
+- 该脚本会从 raw-only 样本目录读取 `upstream.stream.sse`。
+- 回放结果会写入 `artifacts/raw-stream-sim/<run-id>/<sample-id>/`，便于直接查阅。
+- 如果传入历史基线目录，脚本会自动对比当前回放输出和基线文本。
+
+### 采集永久样本
+
+本地启动服务后，可以直接打：
+
+```bash
+POST /admin/dev/raw-samples/capture
+```
+
+这个接口会把请求元信息和上游原始流写入 `tests/raw_stream_samples/<sample-id>/`，以后可以直接拿来做回放和字段分析。派生输出会在本地回放时再生成，不再落在样本目录里。
+
+### 从内存抓包查询并保存样本
+
+如果问题刚刚在本地复现过，也可以先查当前进程内存里的抓包，再选择性落盘：
+
+```bash
+GET /admin/dev/raw-samples/query?q=广州&limit=10
+POST /admin/dev/raw-samples/save
+{"chain_key":"session:xxxx","sample_id":"tmp-from-memory"}
+```
+
+说明：
+- `query` 会按 `chat_session_id` 把 `completion + continue` 归并成一条链，适合定位接续思考问题。
+- `save` 支持用 `query`、`chain_key` 或 `capture_id` 选中目标。
+- 生成的样本目录仍然是 `tests/raw_stream_samples/<sample-id>/`，可以直接喂给回放脚本。
 
 ### 指定输出目录和超时
 

docs/toolcall-semantics.md

@@ -1,72 +1,74 @@
 # Tool call parsing semantics（Go/Node 统一语义）
 
-本文档描述当前代码中 `ParseToolCallsDetailed` / `parseToolCallsDetailed` 的**实际行为**，用于对齐 Go 与 Node Runtime。
+本文档描述当前代码中工具调用解析链路的**实际行为**（以 `internal/toolcall` 与 `internal/js/helpers/stream-tool-sieve` 为准）。
 
-## 1) 输出结构（当前实现）
+文档导航：[总览](../README.MD) / [架构说明](./ARCHITECTURE.md) / [测试指南](./TESTING.md)
 
-- `calls`：解析得到的工具调用列表（`name` + `input`）。
-- `sawToolCallSyntax`：检测到工具调用语法特征时为 `true`（例如 `tool_calls`、`<tool_call>`、`<function_call>`、`<invoke>`、`function.name:`）。
-- `rejectedByPolicy`：当前实现固定为 `false`（预留字段，尚未启用 allow-list 拒绝）。
+## 1) 当前输出结构
+
+`ParseToolCallsDetailed` / `parseToolCallsDetailed` 返回：
+
+- `calls`：解析出的工具调用列表（`name` + `input`）。
+- `sawToolCallSyntax`：检测到工具调用语法特征时为 `true`。
+- `rejectedByPolicy`：当前实现固定为 `false`（预留字段）。
 - `rejectedToolNames`：当前实现固定为空数组（预留字段）。
 
-> 说明：`filterToolCallsDetailed` 当前仅做结构清洗，不做工具名策略拒绝。
+> 当前 `filterToolCallsDetailed` 仅做结构清洗，不做 allow-list 工具名硬拒绝。
 
-## 2) 解析管线
+## 2) 解析范围（重点）
 
-1. **示例保护**：若判定为 fenced code block 示例上下文，则跳过执行型解析。
-2. **候选片段构建**：从完整文本中构建候选（原文、围绕 `tool_calls` 的 JSON 片段、首尾大括号切片等）。
-3. **按序尝试解析（命中即停）**：
-   - 对“明显 JSON 工具载荷候选”（以 `{`/`[` 开头且包含 `tool_calls`/`\"function\"`）先走 JSON 解析，避免 JSON 字符串内偶发 XML 片段误命中；
-   - 其余候选优先 XML 解析（`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / `antml:function_call` 等）；
-   - JSON 解析（`{"tool_calls": [...]}`、列表、单对象）；
-   - Markup 解析；
-   - Text-KV 回退（如 `function.name:` + `function.arguments:`）。
-4. **兜底**：候选全部失败后，再对全文做 XML / Text-KV 回退。
+当前版本的可执行解析以 **XML/Markup 家族**为主：
 
-## 3) XML 能力边界（当前）
+- `<tool_call>...</tool_call>`
+- `<function_call>...</function_call>`
+- `<invoke ...>...</invoke>`（含自闭合）
+- `<tool_use>...</tool_use>`
+- antml 变体（如 `antml:function_call` / `antml:argument`）
 
-当前已支持输入端的“多 XML/标记风格”解析，包括但不限于：
+并支持在这些标记块内部解析：
 
-- `<tool_call><tool_name>...</tool_name><parameters>...</parameters></tool_call>`
-- `<function_call>tool</function_call><function parameter name="x">...</function parameter>`
-- `<invoke name="tool"><parameter name="x">...</parameter></invoke>`
-- `antml:function_call` / `antml:argument` / `antml:parameters`
-- `tool_use` 家族标签
+- JSON 参数字符串
+- 标签参数（`<parameter name="...">...`）
+- key/value 风格子标签
 
-但**输出端仍统一转换为 OpenAI 兼容 JSON 事件/对象**（`message.tool_calls`、`delta.tool_calls`、`response.function_call_arguments.*`）。
+## 3) 不应再假设的行为
 
-## 4) 关于“是否可以封装成 XML 再喂给模型”
+以下说法在当前实现中已不成立：
 
-结论：**可以做，而且当前解析器已经能兼容 XML 作为输入格式之一**，但代码里并没有 `toolcall.prefer_xml_output` 这个开关。现有可调配置只有：
+1. “纯 JSON `tool_calls` 片段会被直接当作可执行工具调用解析”。
+2. “存在 `toolcall.mode` / `toolcall.early_emit_confidence` 等可配置开关可以改变解析策略”。
 
-- `toolcall.mode`：`feature_match` / `off`
-- `toolcall.early_emit_confidence`：`high` / `low` / `off`
+当前策略在代码中固定为：
 
-推荐思路仍然是“输入兼容层 + 输出按客户端协议渲染”：
+- 特征匹配开启（feature-match on）
+- 高置信度早发开启（early emit on）
+- policy 拒绝字段保留但未启用
 
-1. **Prompt 约束层**：如果你要尝试 XML-first，可以在系统提示词里约束模型输出规范 XML tool block（例如 `<tool_calls><tool_call>...</tool_call></tool_calls>`）。
-2. **解析兼容层**：继续在 parser 中同时接受 JSON / XML / ANTML / invoke / text-kv。
-3. **协议归一层**：无论模型输出什么格式，统一落到内部 `ParsedToolCall`。
-4. **对外渲染层**：根据客户端请求协议渲染（OpenAI / Claude / Gemini 各自格式）。
+## 4) 流式与防泄漏语义
 
-这样可以同时获得：
+在流式链路中（OpenAI / Claude / Gemini 统一内核）：
 
-- 减少模型端 JSON 转义/引号错误；
-- 不破坏现有 SDK / 客户端生态；
-- 逐步灰度（按模型、按租户、按请求开关）。
+- 工具调用片段会被优先提取为结构化增量输出；
+- 已识别的工具调用原始片段不会作为普通文本再次回流；
+- fenced code block 中的示例内容按文本处理，不作为可执行工具调用。
 
-## 5) 落地建议（低风险迭代）
+## 5) 落地建议（按当前实现）
 
-- 继续使用现有的 `toolcall.mode=feature_match` 和 `toolcall.early_emit_confidence=high` 作为默认策略。
-- 如果要试 XML-first，把它放在 prompt 层或上游模板层，不要假设代码里已有专门的 XML 输出开关。
-- 增加观测指标：
-  - `toolcall_parse_source`（json/xml/markup/textkv）；
-  - `toolcall_parse_success_rate`；
-  - `toolcall_malformed_rate`；
-  - `toolcall_repair_rate`。
-- 先在 `responses` 链路灰度，再扩展 `chat.completions`。
+1. Prompt 里优先约束模型输出 XML/Markup 工具块。
+2. 执行器侧继续做工具名白名单与参数 schema 校验（不要依赖 parser 代替安全策略）。
+3. 需要兼容历史“纯 JSON tool_calls”模型输出时，请在上游模板层把输出规范化为 XML/Markup 风格再进入 DS2API。
 
-## 6) 兼容性提醒
+## 6) 回归验证建议
 
-- 上游模型若输出混合文本 + XML，仍可能出现“半结构化”噪声，需要依赖现有 sieve 增量消费策略。
-- XML 不等于安全：仍需做 tool 名、参数 schema、执行权限的服务端校验。
+可直接运行：
+
+```bash
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/toolcall/
+node --test tests/node/stream-tool-sieve.test.js
+```
+
+重点覆盖：
+
+- `<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / antml 变体
+- 参数 JSON 修复与解析
+- 流式增量下的工具调用提取与文本防泄漏

go.mod

@@ -3,23 +3,22 @@ module ds2api
 go 1.26.0
 
 require (
-	github.com/andybalholm/brotli v1.0.6
-	github.com/go-chi/chi/v5 v5.2.3
+	github.com/andybalholm/brotli v1.2.1
+	github.com/go-chi/chi/v5 v5.2.5
 	github.com/google/uuid v1.6.0
 	github.com/refraction-networking/utls v1.8.2
-	github.com/tetratelabs/wazero v1.9.0
+	github.com/router-for-me/CLIProxyAPI/v6 v6.9.14
 )
 
 require (
-	github.com/klauspost/compress v1.17.4 // indirect
-	github.com/router-for-me/CLIProxyAPI/v6 v6.9.8 // indirect
-	github.com/sirupsen/logrus v1.9.3 // indirect
+	github.com/klauspost/compress v1.18.5 // indirect
+	github.com/sirupsen/logrus v1.9.4 // indirect
 	github.com/tidwall/gjson v1.18.0 // indirect
-	github.com/tidwall/match v1.1.1 // indirect
-	github.com/tidwall/pretty v1.2.0 // indirect
+	github.com/tidwall/match v1.2.0 // indirect
+	github.com/tidwall/pretty v1.2.1 // indirect
 	github.com/tidwall/sjson v1.2.5 // indirect
-	golang.org/x/crypto v0.45.0 // indirect
-	golang.org/x/net v0.47.0 // indirect
-	golang.org/x/sys v0.38.0 // indirect
+	golang.org/x/crypto v0.49.0 // indirect
+	golang.org/x/net v0.52.0
+	golang.org/x/sys v0.42.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -1,47 +1,43 @@
-github.com/andybalholm/brotli v1.0.6 h1:Yf9fFpf49Zrxb9NlQaluyE92/+X7UVHlhMNJN2sxfOI=
-github.com/andybalholm/brotli v1.0.6/go.mod h1:fO7iG3H7G2nSZ7m0zPUDn85XEX2GTukHGRSepvi9Eig=
-github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/andybalholm/brotli v1.2.1 h1:R+f5xP285VArJDRgowrfb9DqL18yVK0gKAW/F+eTWro=
+github.com/andybalholm/brotli v1.2.1/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
-github.com/go-chi/chi/v5 v5.2.3 h1:WQIt9uxdsAbgIYgid+BpYc+liqQZGMHRaUwp0JUcvdE=
-github.com/go-chi/chi/v5 v5.2.3/go.mod h1:L2yAIGWB3H+phAw1NxKwWM+7eUH/lU8pOMm5hHcoops=
+github.com/go-chi/chi/v5 v5.2.5 h1:Eg4myHZBjyvJmAFjFvWgrqDTXFyOzjj7YIm3L3mu6Ug=
+github.com/go-chi/chi/v5 v5.2.5/go.mod h1:X7Gx4mteadT3eDOMTsXzmI4/rwUpOwBHLpAfupzFJP0=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
-github.com/klauspost/compress v1.17.4 h1:Ej5ixsIri7BrIjBkRZLTo6ghwrEtHFk7ijlczPW4fZ4=
-github.com/klauspost/compress v1.17.4/go.mod h1:/dCuZOvVtNoHsyb+cuJD3itjs3NbnF6KH9zAO4BDxPM=
+github.com/klauspost/compress v1.18.5 h1:/h1gH5Ce+VWNLSWqPzOVn6XBO+vJbCNGvjoaGBFW2IE=
+github.com/klauspost/compress v1.18.5/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
-github.com/refraction-networking/utls v1.8.1 h1:yNY1kapmQU8JeM1sSw2H2asfTIwWxIkrMJI0pRUOCAo=
-github.com/refraction-networking/utls v1.8.1/go.mod h1:jkSOEkLqn+S/jtpEHPOsVv/4V4EVnelwbMQl4vCWXAM=
 github.com/refraction-networking/utls v1.8.2 h1:j4Q1gJj0xngdeH+Ox/qND11aEfhpgoEvV+S9iJ2IdQo=
 github.com/refraction-networking/utls v1.8.2/go.mod h1:jkSOEkLqn+S/jtpEHPOsVv/4V4EVnelwbMQl4vCWXAM=
-github.com/router-for-me/CLIProxyAPI/v6 v6.9.8 h1:O65R38THenp8E1IK0paQlOfop3Y6UYlfqSdLlepidSY=
-github.com/router-for-me/CLIProxyAPI/v6 v6.9.8/go.mod h1:P1jsIPFXorYGuS2N/3BlZYkpRKi/z7+oR3+1tdG0u4k=
-github.com/sirupsen/logrus v1.9.3 h1:dueUQJ1C2q9oE3F7wvmSGAaVtTmUizReu6fjN8uqzbQ=
-github.com/sirupsen/logrus v1.9.3/go.mod h1:naHLuLoDiP4jHNo9R0sCBMtWGeIprob74mVsIT4qYEQ=
-github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
-github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
-github.com/tetratelabs/wazero v1.9.0 h1:IcZ56OuxrtaEz8UYNRHBrUa9bYeX9oVY93KspZZBf/I=
-github.com/tetratelabs/wazero v1.9.0/go.mod h1:TSbcXCfFP0L2FGkRPxHphadXPjo1T6W+CseNNY7EkjM=
+github.com/router-for-me/CLIProxyAPI/v6 v6.9.14 h1:XItUHrPGE9E5xTeZIPjKGmKqfEs1AZbxl1RPfO5xtrc=
+github.com/router-for-me/CLIProxyAPI/v6 v6.9.14/go.mod h1:P1jsIPFXorYGuS2N/3BlZYkpRKi/z7+oR3+1tdG0u4k=
+github.com/sirupsen/logrus v1.9.4 h1:TsZE7l11zFCLZnZ+teH4Umoq5BhEIfIzfRDZ1Uzql2w=
+github.com/sirupsen/logrus v1.9.4/go.mod h1:ftWc9WdOfJ0a92nsE2jF5u5ZwH8Bv2zdeOC42RjbV2g=
+github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
+github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
 github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY=
 github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
-github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA=
 github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=
-github.com/tidwall/pretty v1.2.0 h1:RWIZEg2iJ8/g6fDDYzMpobmaoGh5OLl4AXtGUGPcqCs=
+github.com/tidwall/match v1.2.0 h1:0pt8FlkOwjN2fPt4bIl4BoNxb98gGHN2ObFEDkrfZnM=
+github.com/tidwall/match v1.2.0/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=
 github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
+github.com/tidwall/pretty v1.2.1 h1:qjsOFOWWQl+N3RsoF5/ssm1pHmJJwhjlSbZ51I6wMl4=
+github.com/tidwall/pretty v1.2.1/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=
 github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28=
-golang.org/x/crypto v0.36.0 h1:AnAEvhDddvBdpY+uR+MyHmuZzzNqXSe/GvuDeob5L34=
-golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
-golang.org/x/crypto v0.45.0 h1:jMBrvKuj23MTlT0bQEOBcAE0mjg8mK9RXFhRH6nyF3Q=
-golang.org/x/crypto v0.45.0/go.mod h1:XTGrrkGJve7CYK7J8PEww4aY7gM3qMCElcJQ8n8JdX4=
-golang.org/x/net v0.47.0 h1:Mx+4dIFzqraBXUugkia1OOvlD6LemFo1ALMHjrXDOhY=
-golang.org/x/net v0.47.0/go.mod h1:/jNxtkgq5yWUGYkaZGqo27cfGZ1c5Nen03aYrrKpVRU=
-golang.org/x/sys v0.0.0-20220715151400-c0bba94af5f8/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.31.0 h1:ioabZlmFYtWhL+TRYpcnNlLwhyxaM9kWTDEmfnprqik=
-golang.org/x/sys v0.31.0/go.mod h1:BJP2sWEmIv4KK5OTEluFJCKSidICx8ciO85XgH3Ak8k=
-golang.org/x/sys v0.38.0 h1:3yZWxaJjBmCWXqhN1qh02AkOnCQ1poK6oF+a7xWL6Gc=
-golang.org/x/sys v0.38.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+github.com/xyproto/randomstring v1.0.5 h1:YtlWPoRdgMu3NZtP45drfy1GKoojuR7hmRcnhZqKjWU=
+github.com/xyproto/randomstring v1.0.5/go.mod h1:rgmS5DeNXLivK7YprL0pY+lTuhNQW3iGxZ18UQApw/E=
+golang.org/x/crypto v0.49.0 h1:+Ng2ULVvLHnJ/ZFEq4KdcDd/cfjrrjjNSXNzxg0Y4U4=
+golang.org/x/crypto v0.49.0/go.mod h1:ErX4dUh2UM+CFYiXZRTcMpEcN8b/1gxEuv3nODoYtCA=
+golang.org/x/net v0.52.0 h1:He/TN1l0e4mmR3QqHMT2Xab3Aj3L9qjbhRm78/6jrW0=
+golang.org/x/net v0.52.0/go.mod h1:R1MAz7uMZxVMualyPXb+VaqGSa3LIaUqk0eEt3w36Sw=
+golang.org/x/sys v0.42.0 h1:omrd2nAlyT5ESRdCLYdm3+fMfNFE/+Rf4bDIQImRJeo=
+golang.org/x/sys v0.42.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

internal/account/pool_edge_test.go

@@ -13,9 +13,7 @@ import (
 
 func TestPoolEmptyNoAccounts(t *testing.T) {
 	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "2")
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
 	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "")
 	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[]}`)
 	pool := NewPool(config.LoadStore())
 	if _, ok := pool.Acquire("", nil); ok {
@@ -165,9 +163,7 @@ func TestPoolAcquireWaitTargetAccount(t *testing.T) {
 
 func TestPoolMaxQueueSizeOverride(t *testing.T) {
 	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "1")
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
 	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "5")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "")
 	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[{"email":"acc1@example.com","token":"t1"}]}`)
 	pool := NewPool(config.LoadStore())
 	status := pool.Status()
@@ -176,19 +172,6 @@ func TestPoolMaxQueueSizeOverride(t *testing.T) {
 	}
 }
 
-func TestPoolQueueSizeAliasEnv(t *testing.T) {
-	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "1")
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
-	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "7")
-	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[{"email":"acc1@example.com","token":"t1"}]}`)
-	pool := NewPool(config.LoadStore())
-	status := pool.Status()
-	if got, ok := status["max_queue_size"].(int); !ok || got != 7 {
-		t.Fatalf("expected max_queue_size=7, got %#v", status["max_queue_size"])
-	}
-}
-
 func TestPoolMultipleAcquireReleaseCycles(t *testing.T) {
 	pool := newSingleAccountPoolForTest(t, "1")
 	for i := 0; i < 10; i++ {

internal/account/pool_limits.go

@@ -29,13 +29,8 @@ func (p *Pool) ApplyRuntimeLimits(maxInflightPerAccount, maxQueueSize, globalMax
 }
 
 func maxInflightFromEnv() int {
-	for _, key := range []string{"DS2API_ACCOUNT_MAX_INFLIGHT", "DS2API_ACCOUNT_CONCURRENCY"} {
-		raw := strings.TrimSpace(os.Getenv(key))
-		if raw == "" {
-			continue
-		}
-		n, err := strconv.Atoi(raw)
-		if err == nil && n > 0 {
+	if raw := strings.TrimSpace(os.Getenv("DS2API_ACCOUNT_MAX_INFLIGHT")); raw != "" {
+		if n, err := strconv.Atoi(raw); err == nil && n > 0 {
 			return n
 		}
 	}
@@ -53,13 +48,8 @@ func defaultRecommendedConcurrency(accountCount, maxInflightPerAccount int) int
 }
 
 func maxQueueFromEnv(defaultSize int) int {
-	for _, key := range []string{"DS2API_ACCOUNT_MAX_QUEUE", "DS2API_ACCOUNT_QUEUE_SIZE"} {
-		raw := strings.TrimSpace(os.Getenv(key))
-		if raw == "" {
-			continue
-		}
-		n, err := strconv.Atoi(raw)
-		if err == nil && n >= 0 {
+	if raw := strings.TrimSpace(os.Getenv("DS2API_ACCOUNT_MAX_QUEUE")); raw != "" {
+		if n, err := strconv.Atoi(raw); err == nil && n >= 0 {
 			return n
 		}
 	}

internal/account/pool_test.go

@@ -12,9 +12,7 @@ import (
 func newPoolForTest(t *testing.T, maxInflight string) *Pool {
 	t.Helper()
 	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", maxInflight)
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
 	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "")
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"keys":["k1"],
 		"accounts":[
@@ -29,9 +27,7 @@ func newPoolForTest(t *testing.T, maxInflight string) *Pool {
 func newSingleAccountPoolForTest(t *testing.T, maxInflight string) *Pool {
 	t.Helper()
 	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", maxInflight)
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
 	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "")
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"keys":["k1"],
 		"accounts":[{"email":"acc1@example.com","token":"token1"}]
@@ -170,9 +166,9 @@ func TestPoolStatusRecommendedConcurrencyRespectsOverride(t *testing.T) {
 	}
 }
 
-func TestPoolAccountConcurrencyAliasEnv(t *testing.T) {
-	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "")
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "4")
+func TestPoolGlobalMaxInflightEnv(t *testing.T) {
+	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "1")
+	t.Setenv("DS2API_GLOBAL_MAX_INFLIGHT", "4")
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"keys":["k1"],
 		"accounts":[
@@ -183,15 +179,15 @@ func TestPoolAccountConcurrencyAliasEnv(t *testing.T) {
 
 	pool := NewPool(config.LoadStore())
 	status := pool.Status()
-	if got, ok := status["max_inflight_per_account"].(int); !ok || got != 4 {
+	if got, ok := status["global_max_inflight"].(int); !ok || got != 4 {
+		t.Fatalf("unexpected global_max_inflight: %#v", status["global_max_inflight"])
+	}
+	if got, ok := status["max_inflight_per_account"].(int); !ok || got != 1 {
 		t.Fatalf("unexpected max_inflight_per_account: %#v", status["max_inflight_per_account"])
 	}
-	if got, ok := status["recommended_concurrency"].(int); !ok || got != 8 {
+	if got, ok := status["recommended_concurrency"].(int); !ok || got != 2 {
 		t.Fatalf("unexpected recommended_concurrency: %#v", status["recommended_concurrency"])
 	}
-	if got, ok := status["max_queue_size"].(int); !ok || got != 8 {
-		t.Fatalf("unexpected max_queue_size: %#v", status["max_queue_size"])
-	}
 }
 
 func TestPoolDropsLegacyTokenOnlyAccountOnLoad(t *testing.T) {
@@ -217,9 +213,7 @@ func TestPoolDropsLegacyTokenOnlyAccountOnLoad(t *testing.T) {
 
 func TestPoolAcquireRotatesIntoTokenlessAccounts(t *testing.T) {
 	t.Setenv("DS2API_ACCOUNT_MAX_INFLIGHT", "1")
-	t.Setenv("DS2API_ACCOUNT_CONCURRENCY", "")
 	t.Setenv("DS2API_ACCOUNT_MAX_QUEUE", "")
-	t.Setenv("DS2API_ACCOUNT_QUEUE_SIZE", "")
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"keys":["k1"],
 		"accounts":[

internal/adapter/claude/deps.go

@@ -22,6 +22,7 @@ type DeepSeekCaller interface {
 
 type ConfigReader interface {
 	ClaudeMapping() map[string]string
+	CompatStripReferenceMarkers() bool
 }
 
 type OpenAIChatRunner interface {

internal/adapter/claude/deps_injection_test.go

@@ -7,6 +7,7 @@ type mockClaudeConfig struct {
 }
 
 func (m mockClaudeConfig) ClaudeMapping() map[string]string { return m.m }
+func (mockClaudeConfig) CompatStripReferenceMarkers() bool  { return true }
 
 func TestNormalizeClaudeRequestUsesConfigInterfaceMapping(t *testing.T) {
 	req := map[string]any{

internal/adapter/claude/handler_messages.go

@@ -64,7 +64,7 @@ func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, store C
 		rec := httptest.NewRecorder()
 		h.OpenAI.ChatCompletions(rec, proxyReq)
 		res := rec.Result()
-		defer res.Body.Close()
+		defer func() { _ = res.Body.Close() }()
 		body, _ := io.ReadAll(res.Body)
 		for k, vv := range res.Header {
 			for _, v := range vv {
@@ -94,7 +94,7 @@ func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, store C
 	rec := httptest.NewRecorder()
 	h.OpenAI.ChatCompletions(rec, proxyReq)
 	res := rec.Result()
-	defer res.Body.Close()
+	defer func() { _ = res.Body.Close() }()
 	body, _ := io.ReadAll(res.Body)
 	if res.StatusCode < 200 || res.StatusCode >= 300 {
 		for k, vv := range res.Header {
@@ -124,7 +124,7 @@ func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, store C
 }
 
 func (h *Handler) handleClaudeStreamRealtime(w http.ResponseWriter, r *http.Request, resp *http.Response, model string, messages []any, thinkingEnabled, searchEnabled bool, toolNames []string) {
-	defer resp.Body.Close()
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeClaudeError(w, http.StatusInternalServerError, string(body))
@@ -149,6 +149,7 @@ func (h *Handler) handleClaudeStreamRealtime(w http.ResponseWriter, r *http.Requ
 		messages,
 		thinkingEnabled,
 		searchEnabled,
+		h.compatStripReferenceMarkers(),
 		toolNames,
 	)
 	streamRuntime.sendMessageStart()

internal/adapter/claude/handler_routes.go

@@ -21,6 +21,13 @@ type Handler struct {
 	OpenAI OpenAIChatRunner
 }
 
+func (h *Handler) compatStripReferenceMarkers() bool {
+	if h == nil || h.Store == nil {
+		return true
+	}
+	return h.Store.CompatStripReferenceMarkers()
+}
+
 var (
 	claudeStreamPingInterval    = time.Duration(deepseek.KeepAliveTimeout) * time.Second
 	claudeStreamIdleTimeout     = time.Duration(deepseek.StreamIdleTimeout) * time.Second

internal/adapter/claude/handler_stream_test.go

@@ -138,77 +138,6 @@ func TestHandleClaudeStreamRealtimeThinkingDelta(t *testing.T) {
 	}
 }
 
-func TestHandleClaudeStreamRealtimeToolSafety(t *testing.T) {
-	h := &Handler{}
-	resp := makeClaudeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\""}`,
-		`data: {"p":"response/content","v":",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
-
-	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "use tool"}}, false, false, []string{"search"})
-
-	frames := parseClaudeFrames(t, rec.Body.String())
-	for _, f := range findClaudeFrames(frames, "content_block_delta") {
-		delta, _ := f.Payload["delta"].(map[string]any)
-		if delta["type"] == "text_delta" && strings.Contains(asString(delta["text"]), `"tool_calls"`) {
-			t.Fatalf("raw tool_calls JSON leaked in text delta: body=%s", rec.Body.String())
-		}
-	}
-
-	foundToolUse := false
-	for _, f := range findClaudeFrames(frames, "content_block_start") {
-		contentBlock, _ := f.Payload["content_block"].(map[string]any)
-		if contentBlock["type"] == "tool_use" {
-			foundToolUse = true
-			break
-		}
-	}
-	if !foundToolUse {
-		t.Fatalf("expected tool_use block in stream, body=%s", rec.Body.String())
-	}
-
-	foundToolUseStop := false
-	for _, f := range findClaudeFrames(frames, "message_delta") {
-		delta, _ := f.Payload["delta"].(map[string]any)
-		if delta["stop_reason"] == "tool_use" {
-			foundToolUseStop = true
-			break
-		}
-	}
-	if !foundToolUseStop {
-		t.Fatalf("expected stop_reason=tool_use, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleClaudeStreamRealtimeToolDetectionFromThinkingFallback(t *testing.T) {
-	h := &Handler{}
-	resp := makeClaudeSSEHTTPResponse(
-		`data: {"p":"response/thinking_content","v":"{\"tool_calls\":[{\"name\":\"search\""}`,
-		`data: {"p":"response/thinking_content","v":",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
-
-	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "use tool"}}, true, false, []string{"search"})
-
-	frames := parseClaudeFrames(t, rec.Body.String())
-	foundToolUse := false
-	for _, f := range findClaudeFrames(frames, "content_block_start") {
-		contentBlock, _ := f.Payload["content_block"].(map[string]any)
-		if contentBlock["type"] == "tool_use" && contentBlock["name"] == "search" {
-			foundToolUse = true
-			break
-		}
-	}
-	if !foundToolUse {
-		t.Fatalf("expected tool_use block from thinking fallback, body=%s", rec.Body.String())
-	}
-}
-
 func TestHandleClaudeStreamRealtimeSkipsThinkingFallbackWhenFinalTextExists(t *testing.T) {
 	h := &Handler{}
 	resp := makeClaudeSSEHTTPResponse(

internal/adapter/claude/handler_util_test.go

@@ -96,7 +96,7 @@ func TestNormalizeClaudeMessagesToolUseToAssistantToolCalls(t *testing.T) {
 	if !containsStr(content, "<tool_calls>") || !containsStr(content, "<tool_name>search_web</tool_name>") {
 		t.Fatalf("expected assistant content to include XML tool call history, got %q", content)
 	}
-	if !containsStr(content, `<parameters>{"query":"latest"}</parameters>`) {
+	if !containsStr(content, "<parameters>\n      <query><![CDATA[latest]]></query>\n    </parameters>") {
 		t.Fatalf("expected assistant content to include serialized parameters, got %q", content)
 	}
 }

internal/adapter/claude/handler_utils.go

@@ -1,12 +1,12 @@
 package claude
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"fmt"
 	"strings"
 
 	"ds2api/internal/prompt"
-	"ds2api/internal/util"
 )
 
 func normalizeClaudeMessages(messages []any) []any {
@@ -98,9 +98,10 @@ func buildClaudeToolPrompt(tools []any) string {
 	}
 	return "You have access to these tools:\n\n" +
 		strings.Join(toolSchemas, "\n\n") + "\n\n" +
-		util.BuildToolCallInstructions(names)
+		toolcall.BuildToolCallInstructions(names)
 }
 
+//nolint:unused // retained for compatibility with pending Claude tool-result prompt flow.
 func formatClaudeToolResultForPrompt(block map[string]any) string {
 	if block == nil {
 		return ""

internal/adapter/claude/handler_utils_sanitize.go

@@ -96,6 +96,7 @@ func looksLikeBase64Payload(v string) bool {
 	return true
 }
 
+//nolint:unused // helper kept for compatibility with upcoming sanitize pipeline.
 func marshalCompactJSON(v any) string {
 	b, err := json.Marshal(v)
 	if err != nil {

internal/adapter/claude/output_clean.go

@@ -0,0 +1,13 @@
+package claude
+
+import textclean "ds2api/internal/textclean"
+
+func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
+	if text == "" {
+		return text
+	}
+	if stripReferenceMarkers {
+		text = textclean.StripReferenceMarkers(text)
+	}
+	return text
+}

internal/adapter/claude/proxy_vercel_test.go

@@ -16,6 +16,8 @@ func (s claudeProxyStoreStub) ClaudeMapping() map[string]string {
 	return s.mapping
 }
 
+func (claudeProxyStoreStub) CompatStripReferenceMarkers() bool { return true }
+
 type openAIProxyStub struct {
 	status int
 	body   string
@@ -32,11 +34,13 @@ func (s openAIProxyStub) ChatCompletions(w http.ResponseWriter, _ *http.Request)
 
 type openAIProxyCaptureStub struct {
 	seenModel string
+	seenReq   map[string]any
 }
 
 func (s *openAIProxyCaptureStub) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 	var req map[string]any
 	_ = json.NewDecoder(r.Body).Decode(&req)
+	s.seenReq = req
 	if m, ok := req["model"].(string); ok {
 		s.seenModel = m
 	}
@@ -82,3 +86,33 @@ func TestClaudeProxyViaOpenAIPreservesClaudeMapping(t *testing.T) {
 		t.Fatalf("expected mapped proxy model deepseek-reasoner, got %q", got)
 	}
 }
+
+func TestClaudeProxyTranslatesInlineImageToOpenAIDataURL(t *testing.T) {
+	openAI := &openAIProxyCaptureStub{}
+	h := &Handler{OpenAI: openAI}
+	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", strings.NewReader(`{"model":"claude-sonnet-4-5","messages":[{"role":"user","content":[{"type":"text","text":"hello"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"QUJDRA=="}}]}],"stream":false}`))
+	rec := httptest.NewRecorder()
+
+	h.Messages(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	messages, _ := openAI.seenReq["messages"].([]any)
+	if len(messages) != 1 {
+		t.Fatalf("expected one translated message, got %#v", openAI.seenReq)
+	}
+	msg, _ := messages[0].(map[string]any)
+	content, _ := msg["content"].([]any)
+	if len(content) != 2 {
+		t.Fatalf("expected translated content blocks, got %#v", msg)
+	}
+	imageBlock, _ := content[1].(map[string]any)
+	if strings.TrimSpace(asString(imageBlock["type"])) != "image_url" {
+		t.Fatalf("expected image_url block, got %#v", imageBlock)
+	}
+	imageURL, _ := imageBlock["image_url"].(map[string]any)
+	if !strings.HasPrefix(strings.TrimSpace(asString(imageURL["url"])), "data:image/png;base64,") {
+		t.Fatalf("expected translated data url, got %#v", imageBlock)
+	}
+}

internal/adapter/claude/standard_request.go

@@ -18,7 +18,7 @@ func normalizeClaudeRequest(store ConfigReader, req map[string]any) (claudeNorma
 	model, _ := req["model"].(string)
 	messagesRaw, _ := req["messages"].([]any)
 	if strings.TrimSpace(model) == "" || len(messagesRaw) == 0 {
-		return claudeNormalizedRequest{}, fmt.Errorf("Request must include 'model' and 'messages'.")
+		return claudeNormalizedRequest{}, fmt.Errorf("request must include 'model' and 'messages'")
 	}
 	if _, ok := req["max_tokens"]; !ok {
 		req["max_tokens"] = 8192
@@ -36,7 +36,7 @@ func normalizeClaudeRequest(store ConfigReader, req map[string]any) (claudeNorma
 		thinkingEnabled = false
 		searchEnabled = false
 	}
-	finalPrompt := deepseek.MessagesPrepare(toMessageMaps(dsPayload["messages"]))
+	finalPrompt := deepseek.MessagesPrepareWithThinking(toMessageMaps(dsPayload["messages"]), thinkingEnabled)
 	toolNames := extractClaudeToolNames(toolsRequested)
 	if len(toolNames) == 0 && len(toolsRequested) > 0 {
 		toolNames = []string{"__any_tool__"}

internal/adapter/claude/stream_runtime_core.go

@@ -19,14 +19,14 @@ type claudeStreamRuntime struct {
 	toolNames []string
 	messages  []any
 
-	thinkingEnabled   bool
-	searchEnabled     bool
-	bufferToolContent bool
+	thinkingEnabled       bool
+	searchEnabled         bool
+	bufferToolContent     bool
+	stripReferenceMarkers bool
 
 	messageID string
 	thinking  strings.Builder
 	text      strings.Builder
-	outputTokens int
 
 	nextBlockIndex     int
 	thinkingBlockOpen  bool
@@ -45,21 +45,23 @@ func newClaudeStreamRuntime(
 	messages []any,
 	thinkingEnabled bool,
 	searchEnabled bool,
+	stripReferenceMarkers bool,
 	toolNames []string,
 ) *claudeStreamRuntime {
 	return &claudeStreamRuntime{
-		w:                  w,
-		rc:                 rc,
-		canFlush:           canFlush,
-		model:              model,
-		messages:           messages,
-		thinkingEnabled:    thinkingEnabled,
-		searchEnabled:      searchEnabled,
-		bufferToolContent:  len(toolNames) > 0,
-		toolNames:          toolNames,
-		messageID:          fmt.Sprintf("msg_%d", time.Now().UnixNano()),
-		thinkingBlockIndex: -1,
-		textBlockIndex:     -1,
+		w:                     w,
+		rc:                    rc,
+		canFlush:              canFlush,
+		model:                 model,
+		messages:              messages,
+		thinkingEnabled:       thinkingEnabled,
+		searchEnabled:         searchEnabled,
+		bufferToolContent:     len(toolNames) > 0,
+		stripReferenceMarkers: stripReferenceMarkers,
+		toolNames:             toolNames,
+		messageID:             fmt.Sprintf("msg_%d", time.Now().UnixNano()),
+		thinkingBlockIndex:    -1,
+		textBlockIndex:        -1,
 	}
 }
 
@@ -67,9 +69,6 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ErrorMessage != "" {
 		s.upstreamErr = parsed.ErrorMessage
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("upstream_error")}
@@ -80,10 +79,11 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 
 	contentSeen := false
 	for _, p := range parsed.Parts {
-		if p.Text == "" {
+		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
+		if cleanedText == "" {
 			continue
 		}
-		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(p.Text) {
+		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(cleanedText) {
 			continue
 		}
 		contentSeen = true
@@ -92,7 +92,11 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 			if !s.thinkingEnabled {
 				continue
 			}
-			s.thinking.WriteString(p.Text)
+			trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
+			if trimmed == "" {
+				continue
+			}
+			s.thinking.WriteString(trimmed)
 			s.closeTextBlock()
 			if !s.thinkingBlockOpen {
 				s.thinkingBlockIndex = s.nextBlockIndex
@@ -112,13 +116,17 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 				"index": s.thinkingBlockIndex,
 				"delta": map[string]any{
 					"type":     "thinking_delta",
-					"thinking": p.Text,
+					"thinking": trimmed,
 				},
 			})
 			continue
 		}
 
-		s.text.WriteString(p.Text)
+		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
+		if trimmed == "" {
+			continue
+		}
+		s.text.WriteString(trimmed)
 		if s.bufferToolContent {
 			if hasUnclosedCodeFence(s.text.String()) {
 				continue
@@ -144,7 +152,7 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 			"index": s.textBlockIndex,
 			"delta": map[string]any{
 				"type": "text_delta",
-				"text": p.Text,
+				"text": trimmed,
 			},
 		})
 	}

internal/adapter/claude/stream_runtime_finalize.go

@@ -1,6 +1,7 @@
 package claude
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"fmt"
 	"time"
@@ -43,12 +44,12 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 	s.closeTextBlock()
 
 	finalThinking := s.thinking.String()
-	finalText := s.text.String()
+	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
 
 	if s.bufferToolContent {
-		detected := util.ParseStandaloneToolCalls(finalText, s.toolNames)
+		detected := toolcall.ParseStandaloneToolCalls(finalText, s.toolNames)
 		if len(detected) == 0 && finalText == "" && finalThinking != "" {
-			detected = util.ParseStandaloneToolCalls(finalThinking, s.toolNames)
+			detected = toolcall.ParseStandaloneToolCalls(finalThinking, s.toolNames)
 		}
 		if len(detected) > 0 {
 			stopReason = "tool_use"
@@ -64,7 +65,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 						"input": map[string]any{},
 					},
 				})
-				
+
 				inputBytes, _ := json.Marshal(tc.Input)
 				s.send("content_block_delta", map[string]any{
 					"type":  "content_block_delta",
@@ -108,9 +109,6 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 	}
 
 	outputTokens := util.EstimateTokens(finalThinking) + util.EstimateTokens(finalText)
-	if s.outputTokens > 0 {
-		outputTokens = s.outputTokens
-	}
 	s.send("message_delta", map[string]any{
 		"type": "message_delta",
 		"delta": map[string]any{

internal/adapter/claude/stream_status_test.go

@@ -28,6 +28,8 @@ func (streamStatusClaudeStoreStub) ClaudeMapping() map[string]string {
 	}
 }
 
+func (streamStatusClaudeStoreStub) CompatStripReferenceMarkers() bool { return true }
+
 func captureClaudeStatusMiddleware(statuses *[]int) func(http.Handler) http.Handler {
 	return func(next http.Handler) http.Handler {
 		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

internal/adapter/gemini/convert_passthrough.go

@@ -5,6 +5,7 @@ import (
 	"strings"
 )
 
+//nolint:unused // compatibility hook for native Gemini request normalization path.
 func collectGeminiPassThrough(req map[string]any) map[string]any {
 	cfg, _ := req["generationConfig"].(map[string]any)
 	if len(cfg) == 0 {

internal/adapter/gemini/convert_request.go

@@ -9,6 +9,7 @@ import (
 	"ds2api/internal/util"
 )
 
+//nolint:unused // kept for native Gemini adapter route compatibility.
 func normalizeGeminiRequest(store ConfigReader, routeModel string, req map[string]any, stream bool) (util.StandardRequest, error) {
 	requestedModel := strings.TrimSpace(routeModel)
 	if requestedModel == "" {
@@ -17,17 +18,17 @@ func normalizeGeminiRequest(store ConfigReader, routeModel string, req map[strin
 
 	resolvedModel, ok := config.ResolveModel(store, requestedModel)
 	if !ok {
-		return util.StandardRequest{}, fmt.Errorf("Model '%s' is not available.", requestedModel)
+		return util.StandardRequest{}, fmt.Errorf("model %q is not available", requestedModel)
 	}
 	thinkingEnabled, searchEnabled, _ := config.GetModelConfig(resolvedModel)
 
 	messagesRaw := geminiMessagesFromRequest(req)
 	if len(messagesRaw) == 0 {
-		return util.StandardRequest{}, fmt.Errorf("Request must include non-empty contents.")
+		return util.StandardRequest{}, fmt.Errorf("request must include non-empty contents")
 	}
 
 	toolsRaw := convertGeminiTools(req["tools"])
-	finalPrompt, toolNames := openai.BuildPromptForAdapter(messagesRaw, toolsRaw, "")
+	finalPrompt, toolNames := openai.BuildPromptForAdapter(messagesRaw, toolsRaw, "", thinkingEnabled)
 	passThrough := collectGeminiPassThrough(req)
 
 	return util.StandardRequest{

internal/adapter/gemini/convert_tools.go

@@ -2,6 +2,7 @@ package gemini
 
 import "strings"
 
+//nolint:unused // kept for native Gemini adapter route compatibility.
 func convertGeminiTools(raw any) []any {
 	tools, _ := raw.([]any)
 	if len(tools) == 0 {

internal/adapter/gemini/deps.go

@@ -22,6 +22,7 @@ type DeepSeekCaller interface {
 
 type ConfigReader interface {
 	ModelAliases() map[string]string
+	CompatStripReferenceMarkers() bool
 }
 
 type OpenAIChatRunner interface {

internal/adapter/gemini/handler_generate.go

@@ -2,6 +2,7 @@ package gemini
 
 import (
 	"bytes"
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -57,7 +58,7 @@ func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, stream
 		rec := httptest.NewRecorder()
 		h.OpenAI.ChatCompletions(rec, proxyReq)
 		res := rec.Result()
-		defer res.Body.Close()
+		defer func() { _ = res.Body.Close() }()
 		body, _ := io.ReadAll(res.Body)
 		for k, vv := range res.Header {
 			for _, v := range vv {
@@ -87,7 +88,7 @@ func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, stream
 	rec := httptest.NewRecorder()
 	h.OpenAI.ChatCompletions(rec, proxyReq)
 	res := rec.Result()
-	defer res.Body.Close()
+	defer func() { _ = res.Body.Close() }()
 	body, _ := io.ReadAll(res.Body)
 	if res.StatusCode < 200 || res.StatusCode >= 300 {
 		for k, vv := range res.Header {
@@ -131,8 +132,9 @@ func writeGeminiErrorFromOpenAI(w http.ResponseWriter, status int, raw []byte) {
 	writeGeminiError(w, status, message)
 }
 
+//nolint:unused // retained for native Gemini non-stream handling path.
 func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *http.Response, model, finalPrompt string, thinkingEnabled bool, toolNames []string) {
-	defer resp.Body.Close()
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeGeminiError(w, resp.StatusCode, strings.TrimSpace(string(body)))
@@ -140,12 +142,20 @@ func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *ht
 	}
 
 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	writeJSON(w, http.StatusOK, buildGeminiGenerateContentResponse(model, finalPrompt, result.Thinking, result.Text, toolNames, result.OutputTokens))
+	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	writeJSON(w, http.StatusOK, buildGeminiGenerateContentResponse(
+		model,
+		finalPrompt,
+		cleanVisibleOutput(result.Thinking, stripReferenceMarkers),
+		cleanVisibleOutput(result.Text, stripReferenceMarkers),
+		toolNames,
+	))
 }
 
-func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string, outputTokens int) map[string]any {
+//nolint:unused // retained for native Gemini non-stream handling path.
+func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string) map[string]any {
 	parts := buildGeminiPartsFromFinal(finalText, finalThinking, toolNames)
-	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText, outputTokens)
+	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText)
 	return map[string]any{
 		"candidates": []map[string]any{
 			{
@@ -162,14 +172,11 @@ func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, final
 	}
 }
 
-func buildGeminiUsage(finalPrompt, finalThinking, finalText string, outputTokens int) map[string]any {
+//nolint:unused // retained for native Gemini non-stream handling path.
+func buildGeminiUsage(finalPrompt, finalThinking, finalText string) map[string]any {
 	promptTokens := util.EstimateTokens(finalPrompt)
 	reasoningTokens := util.EstimateTokens(finalThinking)
 	completionTokens := util.EstimateTokens(finalText)
-	if outputTokens > 0 {
-		completionTokens = outputTokens
-		reasoningTokens = 0
-	}
 	return map[string]any{
 		"promptTokenCount":     promptTokens,
 		"candidatesTokenCount": reasoningTokens + completionTokens,
@@ -177,10 +184,11 @@ func buildGeminiUsage(finalPrompt, finalThinking, finalText string, outputTokens
 	}
 }
 
+//nolint:unused // retained for native Gemini non-stream handling path.
 func buildGeminiPartsFromFinal(finalText, finalThinking string, toolNames []string) []map[string]any {
-	detected := util.ParseToolCalls(finalText, toolNames)
-	if len(detected) == 0 && strings.TrimSpace(finalThinking) != "" {
-		detected = util.ParseToolCalls(finalThinking, toolNames)
+	detected := toolcall.ParseToolCalls(finalText, toolNames)
+	if len(detected) == 0 && finalThinking != "" {
+		detected = toolcall.ParseToolCalls(finalThinking, toolNames)
 	}
 	if len(detected) > 0 {
 		parts := make([]map[string]any, 0, len(detected))
@@ -196,7 +204,7 @@ func buildGeminiPartsFromFinal(finalText, finalThinking string, toolNames []stri
 	}
 
 	text := finalText
-	if strings.TrimSpace(text) == "" {
+	if text == "" {
 		text = finalThinking
 	}
 	return []map[string]any{{"text": text}}

internal/adapter/gemini/handler_routes.go

@@ -17,6 +17,14 @@ type Handler struct {
 	OpenAI OpenAIChatRunner
 }
 
+//nolint:unused // used by native Gemini stream/non-stream runtime helpers.
+func (h *Handler) compatStripReferenceMarkers() bool {
+	if h == nil || h.Store == nil {
+		return true
+	}
+	return h.Store.CompatStripReferenceMarkers()
+}
+
 func RegisterRoutes(r chi.Router, h *Handler) {
 	r.Post("/v1beta/models/{model}:generateContent", h.GenerateContent)
 	r.Post("/v1beta/models/{model}:streamGenerateContent", h.StreamGenerateContent)

internal/adapter/gemini/handler_stream_runtime.go

@@ -12,8 +12,9 @@ import (
 	streamengine "ds2api/internal/stream"
 )
 
+//nolint:unused // retained for native Gemini stream handling path.
 func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Request, resp *http.Response, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string) {
-	defer resp.Body.Close()
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeGeminiError(w, resp.StatusCode, strings.TrimSpace(string(body)))
@@ -27,7 +28,7 @@ func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Req
 
 	rc := http.NewResponseController(w)
 	_, canFlush := w.(http.Flusher)
-	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, toolNames)
+	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames)
 
 	initialType := "text"
 	if thinkingEnabled {
@@ -49,6 +50,7 @@ func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Req
 	})
 }
 
+//nolint:unused // retained for native Gemini stream handling path.
 type geminiStreamRuntime struct {
 	w        http.ResponseWriter
 	rc       *http.ResponseController
@@ -57,16 +59,17 @@ type geminiStreamRuntime struct {
 	model       string
 	finalPrompt string
 
-	thinkingEnabled bool
-	searchEnabled   bool
-	bufferContent   bool
-	toolNames       []string
+	thinkingEnabled       bool
+	searchEnabled         bool
+	bufferContent         bool
+	stripReferenceMarkers bool
+	toolNames             []string
 
 	thinking strings.Builder
 	text     strings.Builder
-	outputTokens int
 }
 
+//nolint:unused // retained for native Gemini stream handling path.
 func newGeminiStreamRuntime(
 	w http.ResponseWriter,
 	rc *http.ResponseController,
@@ -75,21 +78,24 @@ func newGeminiStreamRuntime(
 	finalPrompt string,
 	thinkingEnabled bool,
 	searchEnabled bool,
+	stripReferenceMarkers bool,
 	toolNames []string,
 ) *geminiStreamRuntime {
 	return &geminiStreamRuntime{
-		w:               w,
-		rc:              rc,
-		canFlush:        canFlush,
-		model:           model,
-		finalPrompt:     finalPrompt,
-		thinkingEnabled: thinkingEnabled,
-		searchEnabled:   searchEnabled,
-		bufferContent:   len(toolNames) > 0,
-		toolNames:       toolNames,
+		w:                     w,
+		rc:                    rc,
+		canFlush:              canFlush,
+		model:                 model,
+		finalPrompt:           finalPrompt,
+		thinkingEnabled:       thinkingEnabled,
+		searchEnabled:         searchEnabled,
+		bufferContent:         len(toolNames) > 0,
+		stripReferenceMarkers: stripReferenceMarkers,
+		toolNames:             toolNames,
 	}
 }
 
+//nolint:unused // retained for native Gemini stream handling path.
 func (s *geminiStreamRuntime) sendChunk(payload map[string]any) {
 	b, _ := json.Marshal(payload)
 	_, _ = s.w.Write([]byte("data: "))
@@ -100,33 +106,40 @@ func (s *geminiStreamRuntime) sendChunk(payload map[string]any) {
 	}
 }
 
+//nolint:unused // retained for native Gemini stream handling path.
 func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedDecision {
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" || parsed.Stop {
 		return streamengine.ParsedDecision{Stop: true}
 	}
 
 	contentSeen := false
 	for _, p := range parsed.Parts {
-		if p.Text == "" {
+		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
+		if cleanedText == "" {
 			continue
 		}
-		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(p.Text) {
+		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(cleanedText) {
 			continue
 		}
 		contentSeen = true
 		if p.Type == "thinking" {
 			if s.thinkingEnabled {
-				s.thinking.WriteString(p.Text)
+				trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
+				if trimmed == "" {
+					continue
+				}
+				s.thinking.WriteString(trimmed)
 			}
 			continue
 		}
-		s.text.WriteString(p.Text)
+		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
+		if trimmed == "" {
+			continue
+		}
+		s.text.WriteString(trimmed)
 		if s.bufferContent {
 			continue
 		}
@@ -136,7 +149,7 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 					"index": 0,
 					"content": map[string]any{
 						"role":  "model",
-						"parts": []map[string]any{{"text": p.Text}},
+						"parts": []map[string]any{{"text": trimmed}},
 					},
 				},
 			},
@@ -146,9 +159,10 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	return streamengine.ParsedDecision{ContentSeen: contentSeen}
 }
 
+//nolint:unused // retained for native Gemini stream handling path.
 func (s *geminiStreamRuntime) finalize() {
 	finalThinking := s.thinking.String()
-	finalText := s.text.String()
+	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
 
 	if s.bufferContent {
 		parts := buildGeminiPartsFromFinal(finalText, finalThinking, s.toolNames)
@@ -180,6 +194,6 @@ func (s *geminiStreamRuntime) finalize() {
 			},
 		},
 		"modelVersion":  s.model,
-		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText, s.outputTokens),
+		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText),
 	})
 }

internal/adapter/gemini/handler_test.go

@@ -17,7 +17,8 @@ import (
 
 type testGeminiConfig struct{}
 
-func (testGeminiConfig) ModelAliases() map[string]string { return nil }
+func (testGeminiConfig) ModelAliases() map[string]string   { return nil }
+func (testGeminiConfig) CompatStripReferenceMarkers() bool { return true }
 
 type testGeminiAuth struct {
 	a   *auth.RequestAuth
@@ -41,19 +42,23 @@ func (m testGeminiAuth) Determine(_ *http.Request) (*auth.RequestAuth, error) {
 
 func (testGeminiAuth) Release(_ *auth.RequestAuth) {}
 
+//nolint:unused // reserved test double for native Gemini DS-call path coverage.
 type testGeminiDS struct {
 	resp *http.Response
 	err  error
 }
 
+//nolint:unused // reserved test double for native Gemini DS-call path coverage.
 func (m testGeminiDS) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
 	return "session-id", nil
 }
 
+//nolint:unused // reserved test double for native Gemini DS-call path coverage.
 func (m testGeminiDS) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
 	return "pow", nil
 }
 
+//nolint:unused // reserved test double for native Gemini DS-call path coverage.
 func (m testGeminiDS) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
 	if m.err != nil {
 		return nil, m.err
@@ -62,8 +67,8 @@ func (m testGeminiDS) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ m
 }
 
 type geminiOpenAIErrorStub struct {
-	status int
-	body   string
+	status  int
+	body    string
 	headers map[string]string
 }
 
@@ -77,11 +82,17 @@ func (s geminiOpenAIErrorStub) ChatCompletions(w http.ResponseWriter, _ *http.Re
 }
 
 type geminiOpenAISuccessStub struct {
-	stream bool
-	body   string
+	stream  bool
+	body    string
+	seenReq map[string]any
 }
 
-func (s geminiOpenAISuccessStub) ChatCompletions(w http.ResponseWriter, _ *http.Request) {
+func (s *geminiOpenAISuccessStub) ChatCompletions(w http.ResponseWriter, r *http.Request) {
+	if r != nil {
+		var req map[string]any
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		s.seenReq = req
+	}
 	if s.stream {
 		w.Header().Set("Content-Type", "text/event-stream")
 		w.WriteHeader(http.StatusOK)
@@ -99,6 +110,7 @@ func (s geminiOpenAISuccessStub) ChatCompletions(w http.ResponseWriter, _ *http.
 	_, _ = w.Write([]byte(out))
 }
 
+//nolint:unused // helper retained for native Gemini stream fixture tests.
 func makeGeminiUpstreamResponse(lines ...string) *http.Response {
 	body := strings.Join(lines, "\n")
 	if !strings.HasSuffix(body, "\n") {
@@ -138,7 +150,7 @@ func TestGeminiRoutesRegistered(t *testing.T) {
 func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 	h := &Handler{
 		Store: testGeminiConfig{},
-		OpenAI: geminiOpenAISuccessStub{
+		OpenAI: &geminiOpenAISuccessStub{
 			body: `{"id":"chatcmpl-1","object":"chat.completion","choices":[{"index":0,"message":{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"eval_javascript","arguments":"{\"code\":\"1+1\"}"}}]},"finish_reason":"tool_calls"}]}`,
 		},
 	}
@@ -178,7 +190,7 @@ func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 }
 
 func TestGenerateContentMixedToolSnippetAlsoTriggersFunctionCall(t *testing.T) {
-	h := &Handler{Store: testGeminiConfig{}, OpenAI: geminiOpenAISuccessStub{}}
+	h := &Handler{Store: testGeminiConfig{}, OpenAI: &geminiOpenAISuccessStub{}}
 	r := chi.NewRouter()
 	RegisterRoutes(r, h)
 
@@ -211,7 +223,7 @@ func TestGenerateContentMixedToolSnippetAlsoTriggersFunctionCall(t *testing.T) {
 func TestStreamGenerateContentEmitsSSE(t *testing.T) {
 	h := &Handler{
 		Store:  testGeminiConfig{},
-		OpenAI: geminiOpenAISuccessStub{stream: true},
+		OpenAI: &geminiOpenAISuccessStub{stream: true},
 	}
 	r := chi.NewRouter()
 	RegisterRoutes(r, h)
@@ -245,9 +257,42 @@ func TestStreamGenerateContentEmitsSSE(t *testing.T) {
 	}
 }
 
+func TestGeminiProxyTranslatesInlineImageToOpenAIDataURL(t *testing.T) {
+	openAI := &geminiOpenAISuccessStub{}
+	h := &Handler{Store: testGeminiConfig{}, OpenAI: openAI}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	body := `{"contents":[{"role":"user","parts":[{"text":"hello"},{"inlineData":{"mimeType":"image/png","data":"QUJDRA=="}}]}]}`
+	req := httptest.NewRequest(http.MethodPost, "/v1beta/models/gemini-2.5-pro:generateContent", strings.NewReader(body))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	messages, _ := openAI.seenReq["messages"].([]any)
+	if len(messages) != 1 {
+		t.Fatalf("expected one translated message, got %#v", openAI.seenReq)
+	}
+	msg, _ := messages[0].(map[string]any)
+	content, _ := msg["content"].([]any)
+	if len(content) != 2 {
+		t.Fatalf("expected translated content blocks, got %#v", msg)
+	}
+	imageBlock, _ := content[1].(map[string]any)
+	if strings.TrimSpace(asString(imageBlock["type"])) != "image_url" {
+		t.Fatalf("expected image_url block, got %#v", imageBlock)
+	}
+	imageURL, _ := imageBlock["image_url"].(map[string]any)
+	if !strings.HasPrefix(strings.TrimSpace(asString(imageURL["url"])), "data:image/png;base64,") {
+		t.Fatalf("expected translated data url, got %#v", imageBlock)
+	}
+}
+
 func TestGenerateContentOpenAIProxyErrorUsesGeminiEnvelope(t *testing.T) {
 	h := &Handler{
-		Store:  testGeminiConfig{},
+		Store: testGeminiConfig{},
 		OpenAI: geminiOpenAIErrorStub{
 			status: http.StatusUnauthorized,
 			body:   `{"error":{"message":"invalid api key"}}`,

internal/adapter/gemini/output_clean.go

@@ -0,0 +1,14 @@
+package gemini
+
+import textclean "ds2api/internal/textclean"
+
+//nolint:unused // retained for native Gemini output post-processing path.
+func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
+	if text == "" {
+		return text
+	}
+	if stripReferenceMarkers {
+		text = textclean.StripReferenceMarkers(text)
+	}
+	return text
+}

internal/adapter/openai/chat_stream_runtime.go

@@ -1,6 +1,7 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"net/http"
 	"strings"
@@ -8,7 +9,6 @@ import (
 	openaifmt "ds2api/internal/format/openai"
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
-	"ds2api/internal/util"
 )
 
 type chatStreamRuntime struct {
@@ -22,8 +22,9 @@ type chatStreamRuntime struct {
 	finalPrompt  string
 	toolNames    []string
 
-	thinkingEnabled bool
-	searchEnabled   bool
+	thinkingEnabled       bool
+	searchEnabled         bool
+	stripReferenceMarkers bool
 
 	firstChunkSent       bool
 	bufferToolContent    bool
@@ -36,7 +37,6 @@ type chatStreamRuntime struct {
 	streamToolNames   map[int]string
 	thinking          strings.Builder
 	text              strings.Builder
-	outputTokens      int
 }
 
 func newChatStreamRuntime(
@@ -49,25 +49,27 @@ func newChatStreamRuntime(
 	finalPrompt string,
 	thinkingEnabled bool,
 	searchEnabled bool,
+	stripReferenceMarkers bool,
 	toolNames []string,
 	bufferToolContent bool,
 	emitEarlyToolDeltas bool,
 ) *chatStreamRuntime {
 	return &chatStreamRuntime{
-		w:                   w,
-		rc:                  rc,
-		canFlush:            canFlush,
-		completionID:        completionID,
-		created:             created,
-		model:               model,
-		finalPrompt:         finalPrompt,
-		toolNames:           toolNames,
-		thinkingEnabled:     thinkingEnabled,
-		searchEnabled:       searchEnabled,
-		bufferToolContent:   bufferToolContent,
-		emitEarlyToolDeltas: emitEarlyToolDeltas,
-		streamToolCallIDs:   map[int]string{},
-		streamToolNames:     map[int]string{},
+		w:                     w,
+		rc:                    rc,
+		canFlush:              canFlush,
+		completionID:          completionID,
+		created:               created,
+		model:                 model,
+		finalPrompt:           finalPrompt,
+		toolNames:             toolNames,
+		thinkingEnabled:       thinkingEnabled,
+		searchEnabled:         searchEnabled,
+		stripReferenceMarkers: stripReferenceMarkers,
+		bufferToolContent:     bufferToolContent,
+		emitEarlyToolDeltas:   emitEarlyToolDeltas,
+		streamToolCallIDs:     map[int]string{},
+		streamToolNames:       map[int]string{},
 	}
 }
 
@@ -96,10 +98,23 @@ func (s *chatStreamRuntime) sendDone() {
 	}
 }
 
+func (s *chatStreamRuntime) sendFailedChunk(status int, message, code string) {
+	s.sendChunk(map[string]any{
+		"status_code": status,
+		"error": map[string]any{
+			"message": message,
+			"type":    openAIErrorType(status),
+			"code":    code,
+			"param":   nil,
+		},
+	})
+	s.sendDone()
+}
+
 func (s *chatStreamRuntime) finalize(finishReason string) {
 	finalThinking := s.thinking.String()
-	finalText := sanitizeLeakedOutput(s.text.String())
-	detected := util.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
+	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
+	detected := toolcall.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
 	if len(detected.Calls) > 0 && !s.toolCallsDoneEmitted {
 		finishReason = "tool_calls"
 		delta := map[string]any{
@@ -142,7 +157,7 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 			if evt.Content == "" {
 				continue
 			}
-			cleaned := sanitizeLeakedOutput(evt.Content)
+			cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
 			if cleaned == "" {
 				continue
 			}
@@ -166,13 +181,22 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 	if len(detected.Calls) > 0 || s.toolCallsEmitted {
 		finishReason = "tool_calls"
 	}
-	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
-	if s.outputTokens > 0 {
-		usage["completion_tokens"] = s.outputTokens
-		if prompt, ok := usage["prompt_tokens"].(int); ok {
-			usage["total_tokens"] = prompt + s.outputTokens
+	if len(detected.Calls) == 0 && !s.toolCallsEmitted && strings.TrimSpace(finalText) == "" {
+		status := http.StatusTooManyRequests
+		message := "Upstream model returned empty output."
+		code := "upstream_empty_output"
+		if strings.TrimSpace(finalThinking) != "" {
+			message = "Upstream model returned reasoning without visible output."
 		}
+		if finishReason == "content_filter" {
+			status = http.StatusBadRequest
+			message = "Upstream content filtered the response and returned no output."
+			code = "content_filter"
+		}
+		s.sendFailedChunk(status, message, code)
+		return
 	}
+	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
 		s.completionID,
 		s.created,
@@ -187,10 +211,10 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter {
+		if strings.TrimSpace(s.text.String()) == "" {
+			return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("content_filter")}
+		}
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReasonHandlerRequested}
 	}
 	if parsed.ErrorMessage != "" {
@@ -203,10 +227,11 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 	newChoices := make([]map[string]any, 0, len(parsed.Parts))
 	contentSeen := false
 	for _, p := range parsed.Parts {
-		if s.searchEnabled && sse.IsCitation(p.Text) {
+		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
+		if s.searchEnabled && sse.IsCitation(cleanedText) {
 			continue
 		}
-		if p.Text == "" {
+		if cleanedText == "" {
 			continue
 		}
 		contentSeen = true
@@ -217,21 +242,29 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 		}
 		if p.Type == "thinking" {
 			if s.thinkingEnabled {
-				s.thinking.WriteString(p.Text)
-				delta["reasoning_content"] = p.Text
+				trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
+				if trimmed == "" {
+					continue
+				}
+				s.thinking.WriteString(trimmed)
+				delta["reasoning_content"] = trimmed
 			}
 		} else {
-			s.text.WriteString(p.Text)
+			trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
+			if trimmed == "" {
+				continue
+			}
+			s.text.WriteString(trimmed)
 			if !s.bufferToolContent {
-				delta["content"] = p.Text
+				delta["content"] = trimmed
 			} else {
-				events := processToolSieveChunk(&s.toolSieve, p.Text, s.toolNames)
+				events := processToolSieveChunk(&s.toolSieve, trimmed, s.toolNames)
 				for _, evt := range events {
 					if len(evt.ToolCallDeltas) > 0 {
 						if !s.emitEarlyToolDeltas {
 							continue
 						}
-						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.toolNames, s.streamToolNames)
+						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.streamToolNames)
 						if len(filtered) == 0 {
 							continue
 						}
@@ -264,7 +297,7 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 						continue
 					}
 					if evt.Content != "" {
-						cleaned := sanitizeLeakedOutput(evt.Content)
+						cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
 						if cleaned == "" {
 							continue
 						}

internal/adapter/openai/citation_links.go

@@ -0,0 +1,31 @@
+package openai
+
+import (
+	"fmt"
+	"regexp"
+	"strconv"
+	"strings"
+)
+
+var citationMarkerPattern = regexp.MustCompile(`(?i)\[citation:\s*(\d+)\]`)
+
+func replaceCitationMarkersWithLinks(text string, links map[int]string) string {
+	if strings.TrimSpace(text) == "" || len(links) == 0 {
+		return text
+	}
+	return citationMarkerPattern.ReplaceAllStringFunc(text, func(match string) string {
+		sub := citationMarkerPattern.FindStringSubmatch(match)
+		if len(sub) < 2 {
+			return match
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(sub[1]))
+		if err != nil || idx <= 0 {
+			return match
+		}
+		url := strings.TrimSpace(links[idx])
+		if url == "" {
+			return match
+		}
+		return fmt.Sprintf("[%d](%s)", idx, url)
+	})
+}

internal/adapter/openai/citation_links_test.go

@@ -0,0 +1,28 @@
+package openai
+
+import "testing"
+
+func TestReplaceCitationMarkersWithLinks(t *testing.T) {
+	raw := "这是一条更新[citation:1]，更多信息见[citation:2]。"
+	links := map[int]string{
+		1: "https://example.com/news-1",
+		2: "https://example.com/news-2",
+	}
+
+	got := replaceCitationMarkersWithLinks(raw, links)
+	want := "这是一条更新[1](https://example.com/news-1)，更多信息见[2](https://example.com/news-2)。"
+	if got != want {
+		t.Fatalf("expected %q, got %q", want, got)
+	}
+}
+
+func TestReplaceCitationMarkersWithLinksKeepsUnknownIndex(t *testing.T) {
+	raw := "只有一个来源[citation:1]，未知来源[citation:3]。"
+	links := map[int]string{1: "https://example.com/a"}
+
+	got := replaceCitationMarkersWithLinks(raw, links)
+	want := "只有一个来源[1](https://example.com/a)，未知来源[citation:3]。"
+	if got != want {
+		t.Fatalf("expected %q, got %q", want, got)
+	}
+}

internal/adapter/openai/deps.go

@@ -18,17 +18,21 @@ type AuthResolver interface {
 type DeepSeekCaller interface {
 	CreateSession(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error)
 	GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error)
+	UploadFile(ctx context.Context, a *auth.RequestAuth, req deepseek.UploadFileRequest, maxAttempts int) (*deepseek.UploadFileResult, error)
 	CallCompletion(ctx context.Context, a *auth.RequestAuth, payload map[string]any, powResp string, maxAttempts int) (*http.Response, error)
+	DeleteSessionForToken(ctx context.Context, token string, sessionID string) (*deepseek.DeleteSessionResult, error)
 	DeleteAllSessionsForToken(ctx context.Context, token string) error
 }
 
 type ConfigReader interface {
 	ModelAliases() map[string]string
 	CompatWideInputStrictOutput() bool
+	CompatStripReferenceMarkers() bool
 	ToolcallMode() string
 	ToolcallEarlyEmitConfidence() string
 	ResponsesStoreTTLSeconds() int
 	EmbeddingsProvider() string
+	AutoDeleteMode() string
 	AutoDeleteSessions() bool
 }
 

internal/adapter/openai/deps_injection_test.go

@@ -3,23 +3,31 @@ package openai
 import "testing"
 
 type mockOpenAIConfig struct {
-	aliases      map[string]string
-	wideInput    bool
-	toolMode     string
-	earlyEmit    string
-	responsesTTL int
-	embedProv    string
+	aliases        map[string]string
+	wideInput      bool
+	autoDeleteMode string
+	toolMode       string
+	earlyEmit      string
+	responsesTTL   int
+	embedProv      string
 }
 
 func (m mockOpenAIConfig) ModelAliases() map[string]string { return m.aliases }
 func (m mockOpenAIConfig) CompatWideInputStrictOutput() bool {
 	return m.wideInput
 }
+func (m mockOpenAIConfig) CompatStripReferenceMarkers() bool   { return true }
 func (m mockOpenAIConfig) ToolcallMode() string                { return m.toolMode }
 func (m mockOpenAIConfig) ToolcallEarlyEmitConfidence() string { return m.earlyEmit }
 func (m mockOpenAIConfig) ResponsesStoreTTLSeconds() int       { return m.responsesTTL }
 func (m mockOpenAIConfig) EmbeddingsProvider() string          { return m.embedProv }
-func (m mockOpenAIConfig) AutoDeleteSessions() bool            { return false }
+func (m mockOpenAIConfig) AutoDeleteMode() string {
+	if m.autoDeleteMode == "" {
+		return "none"
+	}
+	return m.autoDeleteMode
+}
+func (m mockOpenAIConfig) AutoDeleteSessions() bool { return false }
 
 func TestNormalizeOpenAIChatRequestWithConfigInterface(t *testing.T) {
 	cfg := mockOpenAIConfig{

internal/adapter/openai/embeddings_handler.go

@@ -26,8 +26,13 @@ func (h *Handler) Embeddings(w http.ResponseWriter, r *http.Request) {
 	}
 	defer h.Auth.Release(a)
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}

internal/adapter/openai/file_inline_upload.go

@@ -0,0 +1,382 @@
+package openai
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/base64"
+	"fmt"
+	"mime"
+	"net/http"
+	"net/url"
+	"path/filepath"
+	"strings"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+const maxInlineFilesPerRequest = 50
+
+type inlineFileUploadError struct {
+	status  int
+	message string
+	err     error
+}
+
+func (e *inlineFileUploadError) Error() string {
+	if e == nil {
+		return ""
+	}
+	if strings.TrimSpace(e.message) != "" {
+		return e.message
+	}
+	if e.err != nil {
+		return e.err.Error()
+	}
+	return "inline file processing failed"
+}
+
+type inlineUploadState struct {
+	ctx          context.Context
+	handler      *Handler
+	auth         *auth.RequestAuth
+	uploadedByID map[string]string
+	uploadCount  int
+}
+
+type inlineDecodedFile struct {
+	Data            []byte
+	ContentType     string
+	Filename        string
+	ReplacementType string
+}
+
+func (h *Handler) preprocessInlineFileInputs(ctx context.Context, a *auth.RequestAuth, req map[string]any) error {
+	if h == nil || h.DS == nil || len(req) == 0 {
+		return nil
+	}
+	state := &inlineUploadState{
+		ctx:          ctx,
+		handler:      h,
+		auth:         a,
+		uploadedByID: map[string]string{},
+	}
+	for _, key := range []string{"messages", "input", "attachments"} {
+		if raw, ok := req[key]; ok {
+			updated, err := state.walk(raw)
+			if err != nil {
+				return err
+			}
+			req[key] = updated
+		}
+	}
+	if refIDs := collectOpenAIRefFileIDs(req); len(refIDs) > 0 {
+		req["ref_file_ids"] = stringsToAnySlice(refIDs)
+	}
+	return nil
+}
+
+func writeOpenAIInlineFileError(w http.ResponseWriter, err error) {
+	inlineErr, ok := err.(*inlineFileUploadError)
+	if !ok || inlineErr == nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "Failed to process file input.")
+		return
+	}
+	status := inlineErr.status
+	if status == 0 {
+		status = http.StatusInternalServerError
+	}
+	message := strings.TrimSpace(inlineErr.message)
+	if message == "" {
+		message = "Failed to process file input."
+	}
+	writeOpenAIError(w, status, message)
+}
+
+func (s *inlineUploadState) walk(raw any) (any, error) {
+	switch x := raw.(type) {
+	case []any:
+		out := make([]any, len(x))
+		for i, item := range x {
+			updated, err := s.walk(item)
+			if err != nil {
+				return nil, err
+			}
+			out[i] = updated
+		}
+		return out, nil
+	case map[string]any:
+		if replacement, replaced, err := s.tryUploadBlock(x); replaced || err != nil {
+			return replacement, err
+		}
+		for _, key := range []string{"messages", "input", "attachments", "content", "files", "items", "data", "source", "file", "image_url"} {
+			if nested, ok := x[key]; ok {
+				updated, err := s.walk(nested)
+				if err != nil {
+					return nil, err
+				}
+				x[key] = updated
+			}
+		}
+		return x, nil
+	default:
+		return raw, nil
+	}
+}
+
+func (s *inlineUploadState) tryUploadBlock(block map[string]any) (map[string]any, bool, error) {
+	decoded, ok, err := decodeOpenAIInlineFileBlock(block)
+	if err != nil {
+		return nil, true, &inlineFileUploadError{status: http.StatusBadRequest, message: err.Error(), err: err}
+	}
+	if !ok {
+		return nil, false, nil
+	}
+	if s.uploadCount >= maxInlineFilesPerRequest {
+		return nil, true, fmt.Errorf("exceeded maximum of %d inline files per request", maxInlineFilesPerRequest)
+	}
+	fileID, err := s.uploadInlineFile(decoded)
+	if err != nil {
+		return nil, true, &inlineFileUploadError{status: http.StatusInternalServerError, message: "Failed to upload inline file.", err: err}
+	}
+	s.uploadCount++
+	replacement := map[string]any{
+		"type":    decoded.ReplacementType,
+		"file_id": fileID,
+	}
+	if decoded.Filename != "" {
+		replacement["filename"] = decoded.Filename
+	}
+	if decoded.ContentType != "" {
+		replacement["mime_type"] = decoded.ContentType
+	}
+	return replacement, true, nil
+}
+
+func (s *inlineUploadState) uploadInlineFile(file inlineDecodedFile) (string, error) {
+	sum := sha256.Sum256(append([]byte(file.ContentType+"\x00"+file.Filename+"\x00"), file.Data...))
+	cacheKey := fmt.Sprintf("%x", sum[:])
+	if fileID, ok := s.uploadedByID[cacheKey]; ok && strings.TrimSpace(fileID) != "" {
+		return fileID, nil
+	}
+	contentType := strings.TrimSpace(file.ContentType)
+	if contentType == "" {
+		contentType = http.DetectContentType(file.Data)
+	}
+	result, err := s.handler.DS.UploadFile(s.ctx, s.auth, deepseek.UploadFileRequest{
+		Filename:    file.Filename,
+		ContentType: contentType,
+		Data:        file.Data,
+	}, 3)
+	if err != nil {
+		return "", err
+	}
+	fileID := strings.TrimSpace(result.ID)
+	if fileID == "" {
+		return "", fmt.Errorf("upload succeeded without file id")
+	}
+	s.uploadedByID[cacheKey] = fileID
+	return fileID, nil
+}
+
+func decodeOpenAIInlineFileBlock(block map[string]any) (inlineDecodedFile, bool, error) {
+	if block == nil {
+		return inlineDecodedFile{}, false, nil
+	}
+	if strings.TrimSpace(asString(block["file_id"])) != "" {
+		return inlineDecodedFile{}, false, nil
+	}
+	if nested, ok := block["file"].(map[string]any); ok {
+		decoded, matched, err := decodeOpenAIInlineFileBlock(nested)
+		if err != nil || !matched {
+			return decoded, matched, err
+		}
+		if decoded.Filename == "" {
+			decoded.Filename = pickInlineFilename(block, decoded.ContentType, defaultInlinePrefix(decoded.ReplacementType))
+		}
+		return decoded, true, nil
+	}
+	blockType := strings.ToLower(strings.TrimSpace(asString(block["type"])))
+	if raw, matched := extractInlineImageDataURL(block); matched {
+		data, contentType, err := decodeInlinePayload(raw, contentTypeFromMap(block))
+		if err != nil {
+			return inlineDecodedFile{}, true, fmt.Errorf("invalid image input")
+		}
+		return inlineDecodedFile{
+			Data:            data,
+			ContentType:     contentType,
+			Filename:        pickInlineFilename(block, contentType, "image"),
+			ReplacementType: "input_image",
+		}, true, nil
+	}
+	if raw, matched := extractInlineFilePayload(block, blockType); matched {
+		data, contentType, err := decodeInlinePayload(raw, contentTypeFromMap(block))
+		if err != nil {
+			return inlineDecodedFile{}, true, fmt.Errorf("invalid file input")
+		}
+		return inlineDecodedFile{
+			Data:            data,
+			ContentType:     contentType,
+			Filename:        pickInlineFilename(block, contentType, defaultInlinePrefix(blockType)),
+			ReplacementType: "input_file",
+		}, true, nil
+	}
+	return inlineDecodedFile{}, false, nil
+}
+
+func extractInlineImageDataURL(block map[string]any) (string, bool) {
+	imageURL := block["image_url"]
+	switch x := imageURL.(type) {
+	case string:
+		if isDataURL(x) {
+			return strings.TrimSpace(x), true
+		}
+	case map[string]any:
+		if raw := strings.TrimSpace(asString(x["url"])); isDataURL(raw) {
+			return raw, true
+		}
+	}
+	if raw := strings.TrimSpace(asString(block["url"])); isDataURL(raw) {
+		return raw, true
+	}
+	return "", false
+}
+
+func extractInlineFilePayload(block map[string]any, blockType string) (string, bool) {
+	for _, value := range []any{block["file_data"], block["base64"], block["data"]} {
+		if raw := strings.TrimSpace(asString(value)); raw != "" {
+			if strings.Contains(blockType, "file") || block["file_data"] != nil || block["filename"] != nil || block["file_name"] != nil || block["name"] != nil {
+				return raw, true
+			}
+		}
+	}
+	return "", false
+}
+
+func decodeInlinePayload(raw string, explicitContentType string) ([]byte, string, error) {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return nil, "", fmt.Errorf("empty payload")
+	}
+	if isDataURL(raw) {
+		return decodeDataURL(raw, explicitContentType)
+	}
+	decoded, err := decodeBase64Flexible(raw)
+	if err != nil {
+		return nil, "", err
+	}
+	contentType := strings.TrimSpace(explicitContentType)
+	if contentType == "" && len(decoded) > 0 {
+		contentType = http.DetectContentType(decoded)
+	}
+	return decoded, contentType, nil
+}
+
+func decodeDataURL(raw string, explicitContentType string) ([]byte, string, error) {
+	raw = strings.TrimSpace(raw)
+	if !isDataURL(raw) {
+		return nil, "", fmt.Errorf("unsupported data url")
+	}
+	header, payload, ok := strings.Cut(raw, ",")
+	if !ok {
+		return nil, "", fmt.Errorf("invalid data url")
+	}
+	meta := strings.TrimSpace(strings.TrimPrefix(header, "data:"))
+	contentType := strings.TrimSpace(explicitContentType)
+	if contentType == "" {
+		contentType = "application/octet-stream"
+		if meta != "" {
+			parts := strings.Split(meta, ";")
+			if len(parts) > 0 && strings.TrimSpace(parts[0]) != "" {
+				contentType = strings.TrimSpace(parts[0])
+			}
+		}
+	}
+	if strings.Contains(strings.ToLower(meta), ";base64") {
+		decoded, err := decodeBase64Flexible(payload)
+		if err != nil {
+			return nil, "", err
+		}
+		return decoded, contentType, nil
+	}
+	decoded, err := url.PathUnescape(payload)
+	if err != nil {
+		return nil, "", err
+	}
+	return []byte(decoded), contentType, nil
+}
+
+func decodeBase64Flexible(raw string) ([]byte, error) {
+	raw = strings.TrimSpace(raw)
+	for _, enc := range []*base64.Encoding{base64.StdEncoding, base64.RawStdEncoding, base64.URLEncoding, base64.RawURLEncoding} {
+		decoded, err := enc.DecodeString(raw)
+		if err == nil {
+			return decoded, nil
+		}
+	}
+	return nil, fmt.Errorf("invalid base64 payload")
+}
+
+func contentTypeFromMap(block map[string]any) string {
+	for _, value := range []any{block["mime_type"], block["mimeType"], block["content_type"], block["contentType"], block["media_type"], block["mediaType"]} {
+		if contentType := strings.TrimSpace(asString(value)); contentType != "" {
+			return contentType
+		}
+	}
+	if imageURL, ok := block["image_url"].(map[string]any); ok {
+		for _, value := range []any{imageURL["mime_type"], imageURL["mimeType"], imageURL["content_type"], imageURL["contentType"]} {
+			if contentType := strings.TrimSpace(asString(value)); contentType != "" {
+				return contentType
+			}
+		}
+	}
+	return ""
+}
+
+func pickInlineFilename(block map[string]any, contentType string, prefix string) string {
+	for _, value := range []any{block["filename"], block["file_name"], block["name"]} {
+		if name := strings.TrimSpace(asString(value)); name != "" {
+			return filepath.Base(name)
+		}
+	}
+	if prefix == "" {
+		prefix = "upload"
+	}
+	ext := ".bin"
+	if parsedType := strings.TrimSpace(contentType); parsedType != "" {
+		if comma := strings.Index(parsedType, ";"); comma >= 0 {
+			parsedType = strings.TrimSpace(parsedType[:comma])
+		}
+		if exts, err := mime.ExtensionsByType(parsedType); err == nil && len(exts) > 0 && strings.TrimSpace(exts[0]) != "" {
+			ext = exts[0]
+		}
+	}
+	return prefix + ext
+}
+
+func defaultInlinePrefix(blockType string) string {
+	blockType = strings.ToLower(strings.TrimSpace(blockType))
+	if strings.Contains(blockType, "image") {
+		return "image"
+	}
+	return "upload"
+}
+
+func isDataURL(raw string) bool {
+	return strings.HasPrefix(strings.ToLower(strings.TrimSpace(raw)), "data:")
+}
+
+func stringsToAnySlice(items []string) []any {
+	out := make([]any, 0, len(items))
+	for _, item := range items {
+		trimmed := strings.TrimSpace(item)
+		if trimmed == "" {
+			continue
+		}
+		out = append(out, trimmed)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}

internal/adapter/openai/file_inline_upload_test.go

@@ -0,0 +1,274 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+type inlineUploadDSStub struct {
+	uploadCalls    []deepseek.UploadFileRequest
+	lastCtx        context.Context
+	completionReq  map[string]any
+	createSession  string
+	uploadErr      error
+	completionResp *http.Response
+}
+
+func (m *inlineUploadDSStub) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	if strings.TrimSpace(m.createSession) == "" {
+		return "session-id", nil
+	}
+	return m.createSession, nil
+}
+
+func (m *inlineUploadDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "pow", nil
+}
+
+func (m *inlineUploadDSStub) UploadFile(ctx context.Context, _ *auth.RequestAuth, req deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	m.lastCtx = ctx
+	m.uploadCalls = append(m.uploadCalls, req)
+	if m.uploadErr != nil {
+		return nil, m.uploadErr
+	}
+	return &deepseek.UploadFileResult{
+		ID:       "file-inline-1",
+		Filename: req.Filename,
+		Bytes:    int64(len(req.Data)),
+		Status:   "uploaded",
+		Purpose:  req.Purpose,
+	}, nil
+}
+
+func (m *inlineUploadDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
+	m.completionReq = payload
+	if m.completionResp != nil {
+		return m.completionResp, nil
+	}
+	return makeOpenAISSEHTTPResponse(
+		`data: {"p":"response/content","v":"ok"}`,
+		`data: [DONE]`,
+	), nil
+}
+
+func (m *inlineUploadDSStub) DeleteSessionForToken(_ context.Context, _ string, _ string) (*deepseek.DeleteSessionResult, error) {
+	return &deepseek.DeleteSessionResult{Success: true}, nil
+}
+
+func (m *inlineUploadDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	return nil
+}
+
+func TestPreprocessInlineFileInputsReplacesDataURLAndCollectsRefFileIDs(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{DS: ds}
+	req := map[string]any{
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{
+						"type":      "image_url",
+						"image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="},
+					},
+				},
+			},
+		},
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	if err := h.preprocessInlineFileInputs(ctx, &auth.RequestAuth{DeepSeekToken: "token"}, req); err != nil {
+		t.Fatalf("preprocess failed: %v", err)
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload, got %d", len(ds.uploadCalls))
+	}
+	if ds.lastCtx != ctx {
+		t.Fatalf("expected upload to use request context")
+	}
+	if ds.uploadCalls[0].ContentType != "image/png" {
+		t.Fatalf("expected image/png, got %q", ds.uploadCalls[0].ContentType)
+	}
+	if ds.uploadCalls[0].Filename != "image.png" {
+		t.Fatalf("expected inferred filename image.png, got %q", ds.uploadCalls[0].Filename)
+	}
+	messages, _ := req["messages"].([]any)
+	first, _ := messages[0].(map[string]any)
+	content, _ := first["content"].([]any)
+	block, _ := content[0].(map[string]any)
+	if block["type"] != "input_image" {
+		t.Fatalf("expected input_image replacement, got %#v", block)
+	}
+	if block["file_id"] != "file-inline-1" {
+		t.Fatalf("expected file-inline-1 replacement id, got %#v", block)
+	}
+	refIDs, _ := req["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected ref_file_ids: %#v", req["ref_file_ids"])
+	}
+}
+
+func TestPreprocessInlineFileInputsDeduplicatesIdenticalPayloads(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{DS: ds}
+	req := map[string]any{
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="}},
+					map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="}},
+				},
+			},
+		},
+	}
+
+	if err := h.preprocessInlineFileInputs(context.Background(), &auth.RequestAuth{DeepSeekToken: "token"}, req); err != nil {
+		t.Fatalf("preprocess failed: %v", err)
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected deduplicated single upload, got %d", len(ds.uploadCalls))
+	}
+	refIDs, _ := req["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected ref_file_ids after dedupe: %#v", req["ref_file_ids"])
+	}
+}
+
+func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	h.ChatCompletions(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	if ds.completionReq == nil {
+		t.Fatal("expected completion payload to be captured")
+	}
+	refIDs, _ := ds.completionReq["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected completion ref_file_ids: %#v", ds.completionReq["ref_file_ids"])
+	}
+}
+
+func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","input":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"input_image","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	refIDs, _ := ds.completionReq["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected completion ref_file_ids: %#v", ds.completionReq["ref_file_ids"])
+	}
+}
+
+func TestChatCompletionsInlineUploadFailureReturnsBadRequest(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,%%%"}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	h.ChatCompletions(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.completionReq != nil {
+		t.Fatalf("did not expect completion call on upload decode error")
+	}
+}
+
+func TestResponsesInlineUploadFailureReturnsInternalServerError(t *testing.T) {
+	ds := &inlineUploadDSStub{uploadErr: errors.New("boom")}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","input":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusInternalServerError {
+		t.Fatalf("expected 500, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.completionReq != nil {
+		t.Fatalf("did not expect completion call after upload failure")
+	}
+}
+
+func TestVercelPrepareUploadsInlineFilesBeforeLeasePayload(t *testing.T) {
+	t.Setenv("VERCEL", "1")
+	t.Setenv("DS2API_VERCEL_INTERNAL_SECRET", "stream-secret")
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions?__stream_prepare=1", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("X-Ds2-Internal-Token", "stream-secret")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	payload, _ := out["payload"].(map[string]any)
+	if payload == nil {
+		t.Fatalf("expected payload in prepare response, got %#v", out)
+	}
+	refIDs, _ := payload["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected payload ref_file_ids: %#v", payload["ref_file_ids"])
+	}
+}

internal/adapter/openai/file_refs.go

@@ -0,0 +1,94 @@
+package openai
+
+import "strings"
+
+func collectOpenAIRefFileIDs(req map[string]any) []string {
+	if len(req) == 0 {
+		return nil
+	}
+	out := make([]string, 0, 4)
+	seen := map[string]struct{}{}
+	for _, key := range []string{
+		"ref_file_ids",
+		"file_ids",
+		"attachments",
+		"messages",
+		"input",
+	} {
+		raw := req[key]
+		if raw == nil {
+			continue
+		}
+		// Skip top-level strings for 'messages' and 'input' as they are likely plain text content,
+		// not file IDs. String file IDs are expected in 'ref_file_ids' or 'file_ids'.
+		if key == "messages" || key == "input" {
+			if _, ok := raw.(string); ok {
+				continue
+			}
+		}
+		appendOpenAIRefFileIDs(&out, seen, raw)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func appendOpenAIRefFileIDs(out *[]string, seen map[string]struct{}, raw any) {
+	switch x := raw.(type) {
+	case string:
+		addOpenAIRefFileID(out, seen, x)
+	case []string:
+		for _, item := range x {
+			addOpenAIRefFileID(out, seen, item)
+		}
+	case []any:
+		for _, item := range x {
+			appendOpenAIRefFileIDs(out, seen, item)
+		}
+	case map[string]any:
+		if fileID := strings.TrimSpace(asString(x["file_id"])); fileID != "" {
+			addOpenAIRefFileID(out, seen, fileID)
+		}
+		if strings.Contains(strings.ToLower(strings.TrimSpace(asString(x["type"]))), "file") {
+			if fileID := strings.TrimSpace(asString(x["id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+		}
+		if fileMap, ok := x["file"].(map[string]any); ok {
+			if fileID := strings.TrimSpace(asString(fileMap["file_id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+			if fileID := strings.TrimSpace(asString(fileMap["id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+		}
+		// Recurse into potential containers. Note: we do NOT recurse into 'content' or 'input'
+		// if they are plain strings (handled by the top-level switch), but they are usually
+		// nested inside the map branch anyway.
+		// To be safe, we only recurse into these known container keys.
+		for _, key := range []string{"ref_file_ids", "file_ids", "attachments", "messages", "input", "content", "files", "items", "data", "source"} {
+			if nested, ok := x[key]; ok {
+				// If it's a message content that is a string, we must NOT treat it as an ID.
+				if key == "content" || key == "input" {
+					if _, ok := nested.(string); ok {
+						continue
+					}
+				}
+				appendOpenAIRefFileIDs(out, seen, nested)
+			}
+		}
+	}
+}
+
+func addOpenAIRefFileID(out *[]string, seen map[string]struct{}, fileID string) {
+	fileID = strings.TrimSpace(fileID)
+	if fileID == "" {
+		return
+	}
+	if _, ok := seen[fileID]; ok {
+		return
+	}
+	seen[fileID] = struct{}{}
+	*out = append(*out, fileID)
+}

internal/adapter/openai/files_route_test.go

@@ -0,0 +1,202 @@
+package openai
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+type managedFilesAuthStub struct{}
+
+func (managedFilesAuthStub) Determine(_ *http.Request) (*auth.RequestAuth, error) {
+	return &auth.RequestAuth{
+		UseConfigToken: true,
+		DeepSeekToken:  "managed-token",
+		CallerID:       "caller:test",
+		AccountID:      "acct-123",
+		TriedAccounts:  map[string]bool{},
+	}, nil
+}
+
+func (managedFilesAuthStub) DetermineCaller(_ *http.Request) (*auth.RequestAuth, error) {
+	return &auth.RequestAuth{
+		UseConfigToken: true,
+		DeepSeekToken:  "managed-token",
+		CallerID:       "caller:test",
+		AccountID:      "acct-123",
+		TriedAccounts:  map[string]bool{},
+	}, nil
+}
+
+func (managedFilesAuthStub) Release(_ *auth.RequestAuth) {}
+
+type filesRouteDSStub struct {
+	lastReq deepseek.UploadFileRequest
+	upload  *deepseek.UploadFileResult
+	err     error
+}
+
+func (m *filesRouteDSStub) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "", nil
+}
+
+func (m *filesRouteDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "", nil
+}
+
+func (m *filesRouteDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, req deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	m.lastReq = req
+	if m.err != nil {
+		return nil, m.err
+	}
+	if m.upload != nil {
+		return m.upload, nil
+	}
+	return &deepseek.UploadFileResult{ID: "file-123", Filename: req.Filename, Bytes: int64(len(req.Data)), Purpose: req.Purpose, Status: "uploaded"}, nil
+}
+
+func (m *filesRouteDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
+	return nil, errors.New("not implemented")
+}
+
+func (m *filesRouteDSStub) DeleteSessionForToken(_ context.Context, _ string, _ string) (*deepseek.DeleteSessionResult, error) {
+	return &deepseek.DeleteSessionResult{Success: true}, nil
+}
+
+func (m *filesRouteDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	return nil
+}
+
+func newMultipartUploadRequest(t *testing.T, purpose string, filename string, data []byte) *http.Request {
+	t.Helper()
+	var body bytes.Buffer
+	writer := multipart.NewWriter(&body)
+	if purpose != "" {
+		if err := writer.WriteField("purpose", purpose); err != nil {
+			t.Fatalf("write purpose failed: %v", err)
+		}
+	}
+	part, err := writer.CreateFormFile("file", filename)
+	if err != nil {
+		t.Fatalf("create form file failed: %v", err)
+	}
+	if _, err := part.Write(data); err != nil {
+		t.Fatalf("write file failed: %v", err)
+	}
+	if err := writer.Close(); err != nil {
+		t.Fatalf("close writer failed: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", &body)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", writer.FormDataContentType())
+	return req
+}
+
+func TestFilesRouteUploadSuccess(t *testing.T) {
+	ds := &filesRouteDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.lastReq.Filename != "notes.txt" {
+		t.Fatalf("expected filename notes.txt, got %q", ds.lastReq.Filename)
+	}
+	if ds.lastReq.Purpose != "assistants" {
+		t.Fatalf("expected purpose assistants, got %q", ds.lastReq.Purpose)
+	}
+	if string(ds.lastReq.Data) != "hello world" {
+		t.Fatalf("unexpected uploaded data: %q", string(ds.lastReq.Data))
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	if out["object"] != "file" {
+		t.Fatalf("expected file object, got %#v", out)
+	}
+	if out["id"] != "file-123" {
+		t.Fatalf("expected file id file-123, got %#v", out["id"])
+	}
+	if out["filename"] != "notes.txt" {
+		t.Fatalf("expected filename notes.txt, got %#v", out["filename"])
+	}
+}
+
+func TestFilesRouteUploadIncludesAccountIDForManagedAccount(t *testing.T) {
+	ds := &filesRouteDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: managedFilesAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	if out["account_id"] != "acct-123" {
+		t.Fatalf("expected account_id acct-123, got %#v", out["account_id"])
+	}
+}
+
+func TestFilesRouteRejectsNonMultipart(t *testing.T) {
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", bytes.NewBufferString(`{"purpose":"assistants"}`))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}
+
+func TestFilesRouteRequiresFileField(t *testing.T) {
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	var body bytes.Buffer
+	writer := multipart.NewWriter(&body)
+	if err := writer.WriteField("purpose", "assistants"); err != nil {
+		t.Fatalf("write field failed: %v", err)
+	}
+	if err := writer.Close(); err != nil {
+		t.Fatalf("close writer failed: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", &body)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", writer.FormDataContentType())
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}

internal/adapter/openai/handler_chat.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"io"
 	"net/http"
+	"strings"
 	"time"
 
 	"ds2api/internal/auth"
@@ -35,39 +36,35 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 		writeOpenAIError(w, status, detail)
 		return
 	}
+	var sessionID string
 	defer func() {
-		// 自动删除会话（同步）
-		// 必须在 Release 之前同步删除，否则：
-		// 1. 异步删除时账号已被 Release
-		// 2. 新请求可能获取到同一账号并开始使用
-		// 3. 异步删除仍在进行，会截断新请求正在使用的会话
-		if h.Store.AutoDeleteSessions() && a.DeepSeekToken != "" {
-			deleteCtx, cancel := context.WithTimeout(r.Context(), 10*time.Second)
-			defer cancel()
-			err := h.DS.DeleteAllSessionsForToken(deleteCtx, a.DeepSeekToken)
-			if err != nil {
-				config.Logger.Warn("[auto_delete_sessions] failed", "account", a.AccountID, "error", err)
-			} else {
-				config.Logger.Debug("[auto_delete_sessions] success", "account", a.AccountID)
-			}
-		}
+		h.autoDeleteRemoteSession(r.Context(), a, sessionID)
 		h.Auth.Release(a)
 	}()
 
 	r = r.WithContext(auth.WithAuth(r.Context(), a))
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	stdReq, err := normalizeOpenAIChatRequest(h.Store, req, requestTraceID(r))
 	if err != nil {
 		writeOpenAIError(w, http.StatusBadRequest, err.Error())
 		return
 	}
 
-	sessionID, err := h.DS.CreateSession(r.Context(), a, 3)
+	sessionID, err = h.DS.CreateSession(r.Context(), a, 3)
 	if err != nil {
 		if a.UseConfigToken {
 			writeOpenAIError(w, http.StatusUnauthorized, "Account token is invalid. Please re-login the account in admin.")
@@ -91,12 +88,45 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 		h.handleStream(w, r, resp, sessionID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames)
 		return
 	}
-	h.handleNonStream(w, r.Context(), resp, sessionID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.ToolNames)
+	h.handleNonStream(w, r.Context(), resp, sessionID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames)
 }
 
-func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled bool, toolNames []string) {
+func (h *Handler) autoDeleteRemoteSession(ctx context.Context, a *auth.RequestAuth, sessionID string) {
+	mode := h.Store.AutoDeleteMode()
+	if mode == "none" || a.DeepSeekToken == "" {
+		return
+	}
+
+	deleteBaseCtx := context.WithoutCancel(ctx)
+	deleteCtx, cancel := context.WithTimeout(deleteBaseCtx, 10*time.Second)
+	defer cancel()
+
+	switch mode {
+	case "single":
+		if sessionID == "" {
+			config.Logger.Warn("[auto_delete_sessions] skipped single-session delete because session_id is empty", "account", a.AccountID)
+			return
+		}
+		_, err := h.DS.DeleteSessionForToken(deleteCtx, a.DeepSeekToken, sessionID)
+		if err != nil {
+			config.Logger.Warn("[auto_delete_sessions] failed", "account", a.AccountID, "mode", mode, "session_id", sessionID, "error", err)
+			return
+		}
+		config.Logger.Debug("[auto_delete_sessions] success", "account", a.AccountID, "mode", mode, "session_id", sessionID)
+	case "all":
+		if err := h.DS.DeleteAllSessionsForToken(deleteCtx, a.DeepSeekToken); err != nil {
+			config.Logger.Warn("[auto_delete_sessions] failed", "account", a.AccountID, "mode", mode, "error", err)
+			return
+		}
+		config.Logger.Debug("[auto_delete_sessions] success", "account", a.AccountID, "mode", mode)
+	default:
+		config.Logger.Warn("[auto_delete_sessions] unknown mode", "account", a.AccountID, "mode", mode)
+	}
+}
+
+func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string) {
 	if resp.StatusCode != http.StatusOK {
-		defer resp.Body.Close()
+		defer func() { _ = resp.Body.Close() }()
 		body, _ := io.ReadAll(resp.Body)
 		writeOpenAIError(w, resp.StatusCode, string(body))
 		return
@@ -104,22 +134,21 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, re
 	_ = ctx
 	result := sse.CollectStream(resp, thinkingEnabled, true)
 
-	finalThinking := result.Thinking
-	finalText := sanitizeLeakedOutput(result.Text)
-	respBody := openaifmt.BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText, toolNames)
-	if result.OutputTokens > 0 {
-		if usage, ok := respBody["usage"].(map[string]any); ok {
-			usage["completion_tokens"] = result.OutputTokens
-			if prompt, ok := usage["prompt_tokens"].(int); ok {
-				usage["total_tokens"] = prompt + result.OutputTokens
-			}
-		}
+	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	finalThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
+	finalText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
+	if searchEnabled {
+		finalText = replaceCitationMarkersWithLinks(finalText, result.CitationLinks)
 	}
+	if writeUpstreamEmptyOutputError(w, finalText, result.ContentFilter) {
+		return
+	}
+	respBody := openaifmt.BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText, toolNames)
 	writeJSON(w, http.StatusOK, respBody)
 }
 
 func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string) {
-	defer resp.Body.Close()
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeOpenAIError(w, resp.StatusCode, string(body))
@@ -138,6 +167,7 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 	created := time.Now().Unix()
 	bufferToolContent := len(toolNames) > 0
 	emitEarlyToolDeltas := h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence()
+	stripReferenceMarkers := h.compatStripReferenceMarkers()
 	initialType := "text"
 	if thinkingEnabled {
 		initialType = "thinking"
@@ -153,6 +183,7 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 		finalPrompt,
 		thinkingEnabled,
 		searchEnabled,
+		stripReferenceMarkers,
 		toolNames,
 		bufferToolContent,
 		emitEarlyToolDeltas,

internal/adapter/openai/handler_chat_auto_delete_test.go

@@ -0,0 +1,143 @@
+package openai
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+type autoDeleteModeDSStub struct {
+	resp          *http.Response
+	singleCalls   int
+	allCalls      int
+	lastSessionID string
+	lastCtxErr    error
+}
+
+func (m *autoDeleteModeDSStub) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "session-id", nil
+}
+
+func (m *autoDeleteModeDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "pow", nil
+}
+
+func (m *autoDeleteModeDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, _ deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	return &deepseek.UploadFileResult{ID: "file-id", Filename: "file.txt", Bytes: 1, Status: "uploaded"}, nil
+}
+
+func (m *autoDeleteModeDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
+	return m.resp, nil
+}
+
+func (m *autoDeleteModeDSStub) DeleteSessionForToken(_ context.Context, _ string, sessionID string) (*deepseek.DeleteSessionResult, error) {
+	m.singleCalls++
+	m.lastSessionID = sessionID
+	return &deepseek.DeleteSessionResult{SessionID: sessionID, Success: true}, nil
+}
+
+func (m *autoDeleteModeDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	m.allCalls++
+	return nil
+}
+
+func (m *autoDeleteModeDSStub) DeleteSessionForTokenCtx(ctx context.Context, _ string, sessionID string) (*deepseek.DeleteSessionResult, error) {
+	m.singleCalls++
+	m.lastSessionID = sessionID
+	m.lastCtxErr = ctx.Err()
+	return &deepseek.DeleteSessionResult{SessionID: sessionID, Success: true}, nil
+}
+
+func TestChatCompletionsAutoDeleteModes(t *testing.T) {
+	tests := []struct {
+		name       string
+		mode       string
+		wantSingle int
+		wantAll    int
+	}{
+		{name: "none", mode: "none"},
+		{name: "single", mode: "single", wantSingle: 1},
+		{name: "all", mode: "all", wantAll: 1},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			ds := &autoDeleteModeDSStub{
+				resp: makeOpenAISSEHTTPResponse(
+					`data: {"p":"response/content","v":"hello"}`,
+					"data: [DONE]",
+				),
+			}
+			h := &Handler{
+				Store: mockOpenAIConfig{
+					wideInput:      true,
+					autoDeleteMode: tc.mode,
+				},
+				Auth: streamStatusAuthStub{},
+				DS:   ds,
+			}
+
+			reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":"hi"}],"stream":false}`
+			req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+			req.Header.Set("Authorization", "Bearer direct-token")
+			req.Header.Set("Content-Type", "application/json")
+			rec := httptest.NewRecorder()
+
+			h.ChatCompletions(rec, req)
+
+			if rec.Code != http.StatusOK {
+				t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+			}
+			if ds.singleCalls != tc.wantSingle {
+				t.Fatalf("single delete calls=%d want=%d", ds.singleCalls, tc.wantSingle)
+			}
+			if ds.allCalls != tc.wantAll {
+				t.Fatalf("all delete calls=%d want=%d", ds.allCalls, tc.wantAll)
+			}
+			if tc.wantSingle > 0 && ds.lastSessionID != "session-id" {
+				t.Fatalf("expected single delete for session-id, got %q", ds.lastSessionID)
+			}
+		})
+	}
+}
+
+type autoDeleteCtxDSStub struct {
+	autoDeleteModeDSStub
+}
+
+func (m *autoDeleteCtxDSStub) DeleteSessionForToken(ctx context.Context, token string, sessionID string) (*deepseek.DeleteSessionResult, error) {
+	return m.DeleteSessionForTokenCtx(ctx, token, sessionID)
+}
+
+func (m *autoDeleteCtxDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	m.allCalls++
+	return nil
+}
+
+func TestAutoDeleteRemoteSessionIgnoresCanceledParentContext(t *testing.T) {
+	ds := &autoDeleteCtxDSStub{}
+	h := &Handler{
+		Store: mockOpenAIConfig{
+			wideInput:      true,
+			autoDeleteMode: "single",
+		},
+		DS: ds,
+	}
+	a := &auth.RequestAuth{DeepSeekToken: "token", AccountID: "acct"}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	h.autoDeleteRemoteSession(ctx, a, "session-id")
+
+	if ds.singleCalls != 1 {
+		t.Fatalf("single delete calls=%d want=1", ds.singleCalls)
+	}
+	if ds.lastCtxErr != nil {
+		t.Fatalf("delete ctx should not inherit cancellation, got %v", ds.lastCtxErr)
+	}
+}

internal/adapter/openai/handler_files.go

@@ -0,0 +1,104 @@
+package openai
+
+import (
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+const openAIUploadMaxMemory = 32 << 20
+
+func (h *Handler) UploadFile(w http.ResponseWriter, r *http.Request) {
+	a, err := h.Auth.Determine(r)
+	if err != nil {
+		status := http.StatusUnauthorized
+		detail := err.Error()
+		if err == auth.ErrNoAccount {
+			status = http.StatusTooManyRequests
+		}
+		writeOpenAIError(w, status, detail)
+		return
+	}
+	defer h.Auth.Release(a)
+	if !strings.HasPrefix(strings.ToLower(strings.TrimSpace(r.Header.Get("Content-Type"))), "multipart/form-data") {
+		writeOpenAIError(w, http.StatusBadRequest, "content-type must be multipart/form-data")
+		return
+	}
+	// Enforce a hard cap on the total request body size to prevent OOM
+	r.Body = http.MaxBytesReader(w, r.Body, openAIUploadMaxSize)
+	if err := r.ParseMultipartForm(openAIUploadMaxMemory); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "file size exceeds limit")
+			return
+		}
+		writeOpenAIError(w, http.StatusBadRequest, "invalid multipart form")
+		return
+	}
+	if r.MultipartForm != nil {
+		defer func() { _ = r.MultipartForm.RemoveAll() }()
+	}
+	r = r.WithContext(auth.WithAuth(r.Context(), a))
+	file, header, err := r.FormFile("file")
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, "file is required")
+		return
+	}
+	defer func() { _ = file.Close() }()
+	data, err := io.ReadAll(file)
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, "failed to read uploaded file")
+		return
+	}
+	contentType := strings.TrimSpace(header.Header.Get("Content-Type"))
+	if contentType == "" && len(data) > 0 {
+		contentType = http.DetectContentType(data)
+	}
+	result, err := h.DS.UploadFile(r.Context(), a, deepseek.UploadFileRequest{
+		Filename:    header.Filename,
+		ContentType: contentType,
+		Purpose:     strings.TrimSpace(r.FormValue("purpose")),
+		Data:        data,
+	}, 3)
+	if err != nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "Failed to upload file.")
+		return
+	}
+	if result != nil && result.AccountID == "" {
+		result.AccountID = a.AccountID
+	}
+	writeJSON(w, http.StatusOK, buildOpenAIFileObject(result))
+}
+
+func buildOpenAIFileObject(result *deepseek.UploadFileResult) map[string]any {
+	if result == nil {
+		obj := map[string]any{
+			"id":             "",
+			"object":         "file",
+			"bytes":          0,
+			"created_at":     time.Now().Unix(),
+			"filename":       "",
+			"purpose":        "",
+			"status":         "uploaded",
+			"status_details": nil,
+		}
+		return obj
+	}
+	obj := map[string]any{
+		"id":             result.ID,
+		"object":         "file",
+		"bytes":          result.Bytes,
+		"created_at":     time.Now().Unix(),
+		"filename":       result.Filename,
+		"purpose":        result.Purpose,
+		"status":         result.Status,
+		"status_details": nil,
+	}
+	if result.AccountID != "" {
+		obj["account_id"] = result.AccountID
+	}
+	return obj
+}

internal/adapter/openai/handler_routes.go

@@ -13,6 +13,13 @@ import (
 	"ds2api/internal/util"
 )
 
+const (
+	// openAIUploadMaxSize limits total multipart request body size (100 MiB).
+	openAIUploadMaxSize = 100 << 20
+	// openAIGeneralMaxSize limits total JSON request body size (100 MiB).
+	openAIGeneralMaxSize = 100 << 20
+)
+
 // writeJSON is a package-internal alias kept to avoid mass-renaming across
 // every call-site in this package.
 var writeJSON = util.WriteJSON
@@ -28,6 +35,13 @@ type Handler struct {
 	responses    *responseStore
 }
 
+func (h *Handler) compatStripReferenceMarkers() bool {
+	if h == nil || h.Store == nil {
+		return true
+	}
+	return h.Store.CompatStripReferenceMarkers()
+}
+
 type streamLease struct {
 	Auth      *auth.RequestAuth
 	ExpiresAt time.Time
@@ -39,6 +53,7 @@ func RegisterRoutes(r chi.Router, h *Handler) {
 	r.Post("/v1/chat/completions", h.ChatCompletions)
 	r.Post("/v1/responses", h.Responses)
 	r.Get("/v1/responses/{response_id}", h.GetResponseByID)
+	r.Post("/v1/files", h.UploadFile)
 	r.Post("/v1/embeddings", h.Embeddings)
 }
 

internal/adapter/openai/handler_toolcall_format.go

@@ -1,6 +1,7 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"fmt"
 	"strings"
@@ -75,7 +76,7 @@ func injectToolPrompt(messages []map[string]any, tools []any, policy util.ToolCh
 
 // buildToolCallInstructions delegates to the shared util implementation.
 func buildToolCallInstructions(toolNames []string) string {
-	return util.BuildToolCallInstructions(toolNames)
+	return toolcall.BuildToolCallInstructions(toolNames)
 }
 
 func formatIncrementalStreamToolCallDeltas(deltas []toolCallDelta, ids map[int]string) []map[string]any {
@@ -112,7 +113,7 @@ func formatIncrementalStreamToolCallDeltas(deltas []toolCallDelta, ids map[int]s
 	return out
 }
 
-func filterIncrementalToolCallDeltasByAllowed(deltas []toolCallDelta, allowedNames []string, seenNames map[int]string) []toolCallDelta {
+func filterIncrementalToolCallDeltasByAllowed(deltas []toolCallDelta, seenNames map[int]string) []toolCallDelta {
 	if len(deltas) == 0 {
 		return nil
 	}
@@ -138,7 +139,7 @@ func filterIncrementalToolCallDeltasByAllowed(deltas []toolCallDelta, allowedNam
 	return out
 }
 
-func formatFinalStreamToolCallsWithStableIDs(calls []util.ParsedToolCall, ids map[int]string) []map[string]any {
+func formatFinalStreamToolCallsWithStableIDs(calls []toolcall.ParsedToolCall, ids map[int]string) []map[string]any {
 	if len(calls) == 0 {
 		return nil
 	}

internal/adapter/openai/handler_toolcall_test.go

@@ -3,7 +3,6 @@ package openai
 import (
 	"context"
 	"encoding/json"
-	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -59,21 +58,6 @@ func parseSSEDataFrames(t *testing.T, body string) ([]map[string]any, bool) {
 	return frames, done
 }
 
-func streamHasRawToolJSONContent(frames []map[string]any) bool {
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			content, _ := delta["content"].(string)
-			if strings.Contains(content, `"tool_calls"`) {
-				return true
-			}
-		}
-	}
-	return false
-}
-
 func streamHasToolCallsDelta(frames []map[string]any) bool {
 	for _, frame := range frames {
 		choices, _ := frame["choices"].([]any)
@@ -101,364 +85,61 @@ func streamFinishReason(frames []map[string]any) string {
 	return ""
 }
 
-func streamToolCallArgumentChunks(frames []map[string]any) []string {
-	out := make([]string, 0, 4)
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			toolCalls, _ := delta["tool_calls"].([]any)
-			for _, tc := range toolCalls {
-				tcm, _ := tc.(map[string]any)
-				fn, _ := tcm["function"].(map[string]any)
-				if args, ok := fn["arguments"].(string); ok && args != "" {
-					out = append(out, args)
-				}
-			}
-		}
-	}
-	return out
-}
-
-func TestHandleNonStreamToolCallInterceptsChatModel(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-
-	h.handleNonStream(rec, context.Background(), resp, "cid1", "deepseek-chat", "prompt", false, []string{"search"})
-	if rec.Code != http.StatusOK {
-		t.Fatalf("unexpected status: %d", rec.Code)
-	}
-
-	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	if len(choices) != 1 {
-		t.Fatalf("unexpected choices: %#v", out["choices"])
-	}
-	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
-	}
-	msg, _ := choice["message"].(map[string]any)
-	if msg["content"] != nil {
-		t.Fatalf("expected content nil, got %#v", msg["content"])
-	}
-	toolCalls, _ := msg["tool_calls"].([]any)
-	if len(toolCalls) != 1 {
-		t.Fatalf("expected 1 tool call, got %#v", msg["tool_calls"])
-	}
-}
-
-func TestHandleNonStreamToolCallInterceptsReasonerModel(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/thinking_content","v":"先想一下"}`,
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-
-	h.handleNonStream(rec, context.Background(), resp, "cid2", "deepseek-reasoner", "prompt", true, []string{"search"})
-	if rec.Code != http.StatusOK {
-		t.Fatalf("unexpected status: %d", rec.Code)
-	}
-
-	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	choice, _ := choices[0].(map[string]any)
-	msg, _ := choice["message"].(map[string]any)
-	if msg["reasoning_content"] != "先想一下" {
-		t.Fatalf("expected reasoning_content, got %#v", msg["reasoning_content"])
-	}
-	if msg["content"] != nil {
-		t.Fatalf("expected content nil, got %#v", msg["content"])
-	}
-	if choice["finish_reason"] != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
-	}
-}
-
-func TestHandleNonStreamUnknownToolIntercepted(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"not_in_schema\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-
-	h.handleNonStream(rec, context.Background(), resp, "cid2b", "deepseek-chat", "prompt", false, []string{"search"})
-	if rec.Code != http.StatusOK {
-		t.Fatalf("unexpected status: %d", rec.Code)
-	}
-
-	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
-	}
-	msg, _ := choice["message"].(map[string]any)
-	toolCalls, _ := msg["tool_calls"].([]any)
-	if len(toolCalls) != 1 {
-		t.Fatalf("expected tool_calls for unknown schema name, got %#v", msg["tool_calls"])
-	}
-}
-
-func TestHandleNonStreamEmbeddedToolCallExamplePromotesToolCall(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"下面是示例："}`,
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: {"p":"response/content","v":"请勿执行。"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-
-	h.handleNonStream(rec, context.Background(), resp, "cid2c", "deepseek-chat", "prompt", false, []string{"search"})
-	if rec.Code != http.StatusOK {
-		t.Fatalf("unexpected status: %d", rec.Code)
-	}
-
-	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
-	}
-	msg, _ := choice["message"].(map[string]any)
-	toolCalls, _ := msg["tool_calls"].([]any)
-	if len(toolCalls) != 1 {
-		t.Fatalf("expected one tool_call field for embedded example: %#v", msg["tool_calls"])
-	}
-	content, _ := msg["content"].(string)
-	if strings.Contains(content, `"tool_calls"`) {
-		t.Fatalf("expected raw tool_calls json stripped from content, got %#v", content)
-	}
-}
-
-func TestHandleNonStreamFencedToolCallExampleDoesNotPromoteToolCall(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		"data: {\"p\":\"response/content\",\"v\":\"```json\\n{\\\"tool_calls\\\":[{\\\"name\\\":\\\"search\\\",\\\"input\\\":{\\\"q\\\":\\\"go\\\"}}]}\\n```\"}",
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-
-	h.handleNonStream(rec, context.Background(), resp, "cid2d", "deepseek-chat", "prompt", false, []string{"search"})
-	if rec.Code != http.StatusOK {
-		t.Fatalf("unexpected status: %d", rec.Code)
-	}
-
-	out := decodeJSONBody(t, rec.Body.String())
-	choices, _ := out["choices"].([]any)
-	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] == "tool_calls" {
-		t.Fatalf("expected fenced example to remain content-only, got finish_reason=%#v", choice["finish_reason"])
-	}
-	msg, _ := choice["message"].(map[string]any)
-	toolCalls, _ := msg["tool_calls"].([]any)
-	if len(toolCalls) != 0 {
-		t.Fatalf("expected no tool_call field for fenced example: %#v", msg["tool_calls"])
-	}
-	content, _ := msg["content"].(string)
-	if !strings.Contains(content, `"tool_calls"`) {
-		t.Fatalf("expected fenced example content preserved, got %q", content)
-	}
-}
-
 // Backward-compatible alias for historical test name used in CI logs.
-func TestHandleNonStreamFencedToolCallExamplePromotesToolCall(t *testing.T) {
-	TestHandleNonStreamFencedToolCallExampleDoesNotPromoteToolCall(t)
-}
-
-func TestHandleStreamToolCallInterceptsWithoutRawContentLeak(t *testing.T) {
+func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\""}`,
-		`data: {"p":"response/content","v":",\"input\":{\"q\":\"go\"}}]}"}`,
+		`data: {"p":"response/content","v":""}`,
 		`data: [DONE]`,
 	)
 	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid3", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	h.handleNonStream(rec, context.Background(), resp, "cid-empty", "deepseek-chat", "prompt", false, false, nil)
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected status 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	foundToolIndex := false
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			toolCalls, _ := delta["tool_calls"].([]any)
-			for _, tc := range toolCalls {
-				tcm, _ := tc.(map[string]any)
-				if _, ok := tcm["index"].(float64); ok {
-					foundToolIndex = true
-				}
-			}
-		}
-	}
-	if !foundToolIndex {
-		t.Fatalf("expected stream tool_calls item with index, body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("raw tool_calls JSON leaked in content delta: %s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
 }
 
-func TestHandleStreamToolCallLargeArgumentsStillIntercepted(t *testing.T) {
+func TestHandleNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
 	h := &Handler{}
-	large := strings.Repeat("a", 9000)
-	payload := fmt.Sprintf(`{"tool_calls":[{"name":"search","input":{"q":"%s"}}]}`, large)
-	splitAt := len(payload) / 2
 	resp := makeSSEHTTPResponse(
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, payload[:splitAt]),
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, payload[splitAt:]),
+		`data: {"code":"content_filter"}`,
 		`data: [DONE]`,
 	)
 	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid3-large", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	h.handleNonStream(rec, context.Background(), resp, "cid-empty-filtered", "deepseek-chat", "prompt", false, false, nil)
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected status 400 for filtered upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("raw tool_calls JSON leaked in content delta: %s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "content_filter" {
+		t.Fatalf("expected code=content_filter, got %#v", out)
 	}
 }
 
-func TestHandleStreamReasonerToolCallInterceptsWithoutRawContentLeak(t *testing.T) {
+func TestHandleNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/thinking_content","v":"思考中"}`,
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
+		`data: {"p":"response/thinking_content","v":"Only thinking"}`,
 		`data: [DONE]`,
 	)
 	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid4", "deepseek-reasoner", "prompt", true, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	h.handleNonStream(rec, context.Background(), resp, "cid-thinking-only", "deepseek-reasoner", "prompt", true, false, nil)
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected status 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	foundToolIndex := false
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			toolCalls, _ := delta["tool_calls"].([]any)
-			for _, tc := range toolCalls {
-				tcm, _ := tc.(map[string]any)
-				if _, ok := tcm["index"].(float64); ok {
-					foundToolIndex = true
-				}
-			}
-		}
-	}
-	if !foundToolIndex {
-		t.Fatalf("expected stream tool_calls item with index, body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("raw tool_calls JSON leaked in content delta: %s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-
-	hasThinkingDelta := false
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if _, ok := delta["reasoning_content"]; ok {
-				hasThinkingDelta = true
-			}
-		}
-	}
-	if !hasThinkingDelta {
-		t.Fatalf("expected reasoning_content delta in reasoner stream: %s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamUnknownToolEmitsToolCall(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"not_in_schema\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid5", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta for unknown schema name, body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("did not expect raw tool_calls json leak for unknown schema name: %s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamUnknownToolNoArgsEmitsToolCall(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"not_in_schema\"}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid5b", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta for unknown schema name (no args), body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("did not expect raw tool_calls json leak for unknown schema name (no args): %s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
 }
 
@@ -500,287 +181,6 @@ func TestHandleStreamToolsPlainTextStreamsBeforeFinish(t *testing.T) {
 	}
 }
 
-func TestHandleStreamToolCallMixedWithPlainTextSegments(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"下面是示例："}`,
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: {"p":"response/content","v":"请勿执行。"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid7", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta in mixed prose stream, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "下面是示例：") || !strings.Contains(got, "请勿执行。") {
-		t.Fatalf("expected pre/post plain text to pass sieve, got=%q", got)
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls for mixed prose, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamToolCallAfterLeadingTextRemainsText(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"我将调用工具。"}`,
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid7b", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "我将调用工具。") {
-		t.Fatalf("expected leading text to keep streaming, got=%q", got)
-	}
-
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamToolCallWithSameChunkTrailingTextRemainsText(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}接下来我会继续说明。"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid7c", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "接下来我会继续说明。") {
-		t.Fatalf("expected trailing plain text to be preserved, got=%q", got)
-	}
-
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamFencedToolCallSnippetPromotesToolCall(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "下面是调用示例：\n```json\n"),
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}\n```\n仅示例，不要执行。"),
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid7f", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta for fenced snippet, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if strings.Contains(strings.ToLower(got), "tool_calls") {
-		t.Fatalf("expected raw fenced tool_calls snippet stripped from content, got=%q", got)
-	}
-	if strings.Contains(strings.ToLower(got), "```json") || strings.Contains(got, "\n```\n") {
-		t.Fatalf("expected consumed fenced tool payload to not leave empty code fence, got=%q", got)
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamStandaloneToolCallAfterClosedFenceKeepsFence(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "先给一个代码示例：\n```text\nhello\n```\n"),
-		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"),
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid7g", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta for standalone payload, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "```") {
-		t.Fatalf("expected closed fence before standalone tool json to be preserved, got=%q", got)
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamToolCallKeyAppearsLateRemainsText(t *testing.T) {
-	h := &Handler{}
-	spaces := strings.Repeat(" ", 200)
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{`+spaces+`"}`,
-		`data: {"p":"response/content","v":"\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}"}`,
-		`data: {"p":"response/content","v":"后置正文C。"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid8", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "后置正文C。") {
-		t.Fatalf("expected stream to continue after tool json convergence, got=%q", got)
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamInvalidToolJSONDoesNotLeakRawObject(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"前置正文D。"}`,
-		`data: {"p":"response/content","v":"{'tool_calls':[{'name':'search','input':{'q':'go'}}]}"}`,
-		`data: {"p":"response/content","v":"后置正文E。"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid9", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta for invalid json, body=%s", rec.Body.String())
-	}
-	content := strings.Builder{}
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			if c, ok := delta["content"].(string); ok {
-				content.WriteString(c)
-			}
-		}
-	}
-	got := content.String()
-	if !strings.Contains(got, "前置正文D。") || !strings.Contains(got, "后置正文E。") {
-		t.Fatalf("expected pre/post plain text to remain, got=%q", content.String())
-	}
-	if !strings.Contains(strings.ToLower(got), "tool_calls") {
-		t.Fatalf("expected invalid embedded tool-like json to pass through as text, got=%q", got)
-	}
-}
-
 func TestHandleStreamIncompleteCapturedToolJSONFlushesAsTextOnFinalize(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
@@ -814,108 +214,3 @@ func TestHandleStreamIncompleteCapturedToolJSONFlushesAsTextOnFinalize(t *testin
 		t.Fatalf("expected incomplete capture to flush as plain text instead of stalling, got=%q", content.String())
 	}
 }
-
-func TestHandleStreamToolCallArgumentsEmitAsSingleCompletedChunk(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go"}`,
-		`data: {"p":"response/content","v":"lang\",\"page\":1}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid11", "deepseek-chat", "prompt", false, false, []string{"search"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-	if streamHasRawToolJSONContent(frames) {
-		t.Fatalf("raw tool_calls JSON leaked in content delta: %s", rec.Body.String())
-	}
-	argChunks := streamToolCallArgumentChunks(frames)
-	if len(argChunks) == 0 {
-		t.Fatalf("expected tool call arguments chunk, got=%v body=%s", argChunks, rec.Body.String())
-	}
-	joined := strings.Join(argChunks, "")
-	if !strings.Contains(joined, `"q":"golang"`) || !strings.Contains(joined, `"page":1`) {
-		t.Fatalf("unexpected merged arguments stream: %q", joined)
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}
-
-func TestHandleStreamMultiToolCallDoesNotMergeNamesOrArguments(t *testing.T) {
-	h := &Handler{}
-	resp := makeSSEHTTPResponse(
-		`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"search_web\",\"input\":{\"query\":\"latest ai news\"}},{"}`,
-		`data: {"p":"response/content","v":"\"name\":\"eval_javascript\",\"input\":{\"code\":\"1+1\"}}]}"}`,
-		`data: [DONE]`,
-	)
-	rec := httptest.NewRecorder()
-	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
-
-	h.handleStream(rec, req, resp, "cid12", "deepseek-chat", "prompt", false, false, []string{"search_web", "eval_javascript"})
-
-	frames, done := parseSSEDataFrames(t, rec.Body.String())
-	if !done {
-		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
-	}
-	if !streamHasToolCallsDelta(frames) {
-		t.Fatalf("expected tool_calls delta, body=%s", rec.Body.String())
-	}
-
-	foundSearch := false
-	foundEval := false
-	foundIndex1 := false
-	toolCallsDeltaLens := make([]int, 0, 2)
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			rawToolCalls, hasToolCalls := delta["tool_calls"]
-			if !hasToolCalls {
-				continue
-			}
-			toolCalls, _ := rawToolCalls.([]any)
-			toolCallsDeltaLens = append(toolCallsDeltaLens, len(toolCalls))
-			for _, tc := range toolCalls {
-				tcm, _ := tc.(map[string]any)
-				if idx, ok := tcm["index"].(float64); ok && int(idx) == 1 {
-					foundIndex1 = true
-				}
-				fn, _ := tcm["function"].(map[string]any)
-				name, _ := fn["name"].(string)
-				switch name {
-				case "search_web":
-					foundSearch = true
-				case "eval_javascript":
-					foundEval = true
-				case "search_webeval_javascript":
-					t.Fatalf("unexpected merged tool name: %s, body=%s", name, rec.Body.String())
-				}
-				if args, ok := fn["arguments"].(string); ok && strings.Contains(args, `}{"`) {
-					t.Fatalf("unexpected concatenated tool arguments: %q, body=%s", args, rec.Body.String())
-				}
-			}
-		}
-	}
-	if !foundSearch || !foundEval {
-		t.Fatalf("expected both tool names in stream deltas, foundSearch=%v foundEval=%v body=%s", foundSearch, foundEval, rec.Body.String())
-	}
-	if len(toolCallsDeltaLens) != 1 || toolCallsDeltaLens[0] != 2 {
-		t.Fatalf("expected exactly one tool_calls delta with two calls, got lens=%v body=%s", toolCallsDeltaLens, rec.Body.String())
-	}
-	if !foundIndex1 {
-		t.Fatalf("expected second tool call index in stream deltas, body=%s", rec.Body.String())
-	}
-	if streamFinishReason(frames) != "tool_calls" {
-		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
-	}
-}

internal/adapter/openai/leaked_output_sanitize.go

@@ -2,16 +2,24 @@ package openai
 
 import (
 	"regexp"
+	"strings"
 )
 
 var emptyJSONFencePattern = regexp.MustCompile("(?is)```json\\s*```")
 var leakedToolCallArrayPattern = regexp.MustCompile(`(?is)\[\{\s*"function"\s*:\s*\{[\s\S]*?\}\s*,\s*"id"\s*:\s*"call[^"]*"\s*,\s*"type"\s*:\s*"function"\s*}\]`)
 var leakedToolResultBlobPattern = regexp.MustCompile(`(?is)<\s*\|\s*tool\s*\|\s*>\s*\{[\s\S]*?"tool_call_id"\s*:\s*"call[^"]*"\s*}`)
 
-// leakedMetaMarkerPattern matches DeepSeek special tokens in BOTH forms:
-//   - ASCII underscore: <｜end_of_sentence｜>
-//   - U+2581 variant:   <｜end▁of▁sentence｜>  (used in some DeepSeek outputs)
-var leakedMetaMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*(?:assistant|tool|end[_▁]of[_▁]sentence|end[_▁]of[_▁]thinking)\s*[｜\|]>`)
+var leakedThinkTagPattern = regexp.MustCompile(`(?is)</?\s*think\s*>`)
+
+// leakedBOSMarkerPattern matches DeepSeek BOS markers in BOTH forms:
+//   - ASCII underscore: <｜begin_of_sentence｜>
+//   - U+2581 variant:   <｜begin▁of▁sentence｜>
+var leakedBOSMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*begin[_▁]of[_▁]sentence\s*[｜\|]>`)
+
+// leakedMetaMarkerPattern matches the remaining DeepSeek special tokens in BOTH forms:
+//   - ASCII underscore: <｜end_of_sentence｜>, <｜end_of_toolresults｜>, <｜end_of_instructions｜>
+//   - U+2581 variant:   <｜end▁of▁sentence｜>, <｜end▁of▁toolresults｜>, <｜end▁of▁instructions｜>
+var leakedMetaMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*(?:assistant|tool|end[_▁]of[_▁]sentence|end[_▁]of[_▁]thinking|end[_▁]of[_▁]toolresults|end[_▁]of[_▁]instructions)\s*[｜\|]>`)
 
 // leakedAgentXMLBlockPatterns catch agent-style XML blocks that leak through
 // when the sieve fails to capture them. These are applied only to complete
@@ -35,11 +43,48 @@ func sanitizeLeakedOutput(text string) string {
 	out := emptyJSONFencePattern.ReplaceAllString(text, "")
 	out = leakedToolCallArrayPattern.ReplaceAllString(out, "")
 	out = leakedToolResultBlobPattern.ReplaceAllString(out, "")
+	out = stripDanglingThinkSuffix(out)
+	out = leakedThinkTagPattern.ReplaceAllString(out, "")
+	out = leakedBOSMarkerPattern.ReplaceAllString(out, "")
 	out = leakedMetaMarkerPattern.ReplaceAllString(out, "")
 	out = sanitizeLeakedAgentXMLBlocks(out)
 	return out
 }
 
+func stripDanglingThinkSuffix(text string) string {
+	matches := leakedThinkTagPattern.FindAllStringIndex(text, -1)
+	if len(matches) == 0 {
+		return text
+	}
+	depth := 0
+	lastOpen := -1
+	for _, loc := range matches {
+		tag := strings.ToLower(text[loc[0]:loc[1]])
+		compact := strings.ReplaceAll(strings.ReplaceAll(strings.TrimSpace(tag), " ", ""), "\t", "")
+		if strings.HasPrefix(compact, "</") {
+			if depth > 0 {
+				depth--
+				if depth == 0 {
+					lastOpen = -1
+				}
+			}
+			continue
+		}
+		if depth == 0 {
+			lastOpen = loc[0]
+		}
+		depth++
+	}
+	if depth == 0 || lastOpen < 0 {
+		return text
+	}
+	prefix := text[:lastOpen]
+	if strings.TrimSpace(prefix) == "" {
+		return ""
+	}
+	return prefix
+}
+
 func sanitizeLeakedAgentXMLBlocks(text string) string {
 	out := text
 	for _, pattern := range leakedAgentXMLBlockPatterns {

internal/adapter/openai/leaked_output_sanitize_test.go

@@ -19,13 +19,29 @@ func TestSanitizeLeakedOutputRemovesLeakedWireToolCallAndResult(t *testing.T) {
 }
 
 func TestSanitizeLeakedOutputRemovesStandaloneMetaMarkers(t *testing.T) {
-	raw := "A<| end_of_sentence |><| Assistant |>B<| end_of_thinking |>C<｜end▁of▁thinking｜>D<｜end▁of▁sentence｜>E"
+	raw := "A<| end_of_sentence |><| Assistant |>B<| end_of_thinking |>C<｜end▁of▁thinking｜>D<｜end▁of▁sentence｜>E<| end_of_toolresults |>F<｜end▁of▁instructions｜>G"
 	got := sanitizeLeakedOutput(raw)
-	if got != "ABCDE" {
+	if got != "ABCDEFG" {
 		t.Fatalf("unexpected sanitize result for meta markers: %q", got)
 	}
 }
 
+func TestSanitizeLeakedOutputRemovesThinkAndBosMarkers(t *testing.T) {
+	raw := "A<think>B</think>C<｜begin▁of▁sentence｜>D<| begin_of_sentence |>E<｜begin_of_sentence｜>F"
+	got := sanitizeLeakedOutput(raw)
+	if got != "ABCDEF" {
+		t.Fatalf("unexpected sanitize result for think/BOS markers: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesDanglingThinkBlock(t *testing.T) {
+	raw := "Answer prefix<think>internal reasoning that never closes"
+	got := sanitizeLeakedOutput(raw)
+	if got != "Answer prefix" {
+		t.Fatalf("unexpected sanitize result for dangling think block: %q", got)
+	}
+}
+
 func TestSanitizeLeakedOutputRemovesAgentXMLLeaks(t *testing.T) {
 	raw := "Done.<attempt_completion><result>Some final answer</result></attempt_completion>"
 	got := sanitizeLeakedOutput(raw)

internal/adapter/openai/models_route_test.go

@@ -22,6 +22,24 @@ func TestGetModelRouteDirectAndAlias(t *testing.T) {
 		}
 	})
 
+	t.Run("direct_expert", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/v1/models/deepseek-expert-chat", nil)
+		rec := httptest.NewRecorder()
+		r.ServeHTTP(rec, req)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+		}
+	})
+
+	t.Run("direct_vision", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/v1/models/deepseek-vision-chat", nil)
+		rec := httptest.NewRecorder()
+		r.ServeHTTP(rec, req)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+		}
+	})
+
 	t.Run("alias", func(t *testing.T) {
 		req := httptest.NewRequest(http.MethodGet, "/v1/models/gpt-4.1", nil)
 		rec := httptest.NewRecorder()

internal/adapter/openai/output_clean.go

@@ -0,0 +1,13 @@
+package openai
+
+import textclean "ds2api/internal/textclean"
+
+func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
+	if text == "" {
+		return text
+	}
+	if stripReferenceMarkers {
+		text = textclean.StripReferenceMarkers(text)
+	}
+	return sanitizeLeakedOutput(text)
+}

internal/adapter/openai/prompt_build.go

@@ -5,22 +5,22 @@ import (
 	"ds2api/internal/util"
 )
 
-func buildOpenAIFinalPrompt(messagesRaw []any, toolsRaw any, traceID string) (string, []string) {
-	return buildOpenAIFinalPromptWithPolicy(messagesRaw, toolsRaw, traceID, util.DefaultToolChoicePolicy())
+func buildOpenAIFinalPrompt(messagesRaw []any, toolsRaw any, traceID string, thinkingEnabled bool) (string, []string) {
+	return buildOpenAIFinalPromptWithPolicy(messagesRaw, toolsRaw, traceID, util.DefaultToolChoicePolicy(), thinkingEnabled)
 }
 
-func buildOpenAIFinalPromptWithPolicy(messagesRaw []any, toolsRaw any, traceID string, toolPolicy util.ToolChoicePolicy) (string, []string) {
+func buildOpenAIFinalPromptWithPolicy(messagesRaw []any, toolsRaw any, traceID string, toolPolicy util.ToolChoicePolicy, thinkingEnabled bool) (string, []string) {
 	messages := normalizeOpenAIMessagesForPrompt(messagesRaw, traceID)
 	toolNames := []string{}
 	if tools, ok := toolsRaw.([]any); ok && len(tools) > 0 {
 		messages, toolNames = injectToolPrompt(messages, tools, toolPolicy)
 	}
-	return deepseek.MessagesPrepare(messages), toolNames
+	return deepseek.MessagesPrepareWithThinking(messages, thinkingEnabled), toolNames
 }
 
 // BuildPromptForAdapter exposes the OpenAI-compatible prompt building flow so
 // other protocol adapters (for example Gemini) can reuse the same tool/history
 // normalization logic and remain behavior-compatible with chat/completions.
-func BuildPromptForAdapter(messagesRaw []any, toolsRaw any, traceID string) (string, []string) {
-	return buildOpenAIFinalPrompt(messagesRaw, toolsRaw, traceID)
+func BuildPromptForAdapter(messagesRaw []any, toolsRaw any, traceID string, thinkingEnabled bool) (string, []string) {
+	return buildOpenAIFinalPrompt(messagesRaw, toolsRaw, traceID, thinkingEnabled)
 }

internal/adapter/openai/prompt_build_test.go

@@ -40,7 +40,7 @@ func TestBuildOpenAIFinalPrompt_HandlerPathIncludesToolRoundtripSemantics(t *tes
 		},
 	}
 
-	finalPrompt, toolNames := buildOpenAIFinalPrompt(messages, tools, "")
+	finalPrompt, toolNames := buildOpenAIFinalPrompt(messages, tools, "", false)
 	if len(toolNames) != 1 || toolNames[0] != "get_weather" {
 		t.Fatalf("unexpected tool names: %#v", toolNames)
 	}
@@ -73,20 +73,31 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 		},
 	}
 
-	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "")
-	if !strings.Contains(finalPrompt, "After receiving a tool result, use it directly.") {
-		t.Fatalf("vercel prepare finalPrompt missing final-answer instruction: %q", finalPrompt)
-	}
-	if !strings.Contains(finalPrompt, "Only call another tool if the result is insufficient.") {
-		t.Fatalf("vercel prepare finalPrompt missing retry guard instruction: %q", finalPrompt)
+	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "", false)
+	if !strings.Contains(finalPrompt, "Remember: The ONLY valid way to use tools is the <tool_calls> XML block at the end of your response.") {
+		t.Fatalf("vercel prepare finalPrompt missing final tool-call anchor instruction: %q", finalPrompt)
 	}
 	if !strings.Contains(finalPrompt, "TOOL CALL FORMAT") {
 		t.Fatalf("vercel prepare finalPrompt missing xml format instruction: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "Do NOT wrap the XML in markdown code fences") {
+	if !strings.Contains(finalPrompt, "Do NOT wrap XML in markdown fences") {
 		t.Fatalf("vercel prepare finalPrompt missing no-fence xml instruction: %q", finalPrompt)
 	}
 	if strings.Contains(finalPrompt, "```json") {
 		t.Fatalf("vercel prepare finalPrompt should not require fenced tool calls: %q", finalPrompt)
 	}
 }
+
+func TestBuildOpenAIFinalPromptWithThinkingAddsContinuationContract(t *testing.T) {
+	messages := []any{
+		map[string]any{"role": "user", "content": "继续回答上一个问题"},
+	}
+
+	finalPrompt, _ := buildOpenAIFinalPrompt(messages, nil, "", true)
+	if !strings.Contains(finalPrompt, "Continue the conversation from the full prior context") {
+		t.Fatalf("expected continuation contract in thinking prompt, got=%q", finalPrompt)
+	}
+	if !strings.Contains(finalPrompt, "final user-facing answer only in reasoning") {
+		t.Fatalf("expected visible-answer contract in thinking prompt, got=%q", finalPrompt)
+	}
+}

internal/adapter/openai/responses_embeddings_test.go

@@ -156,6 +156,33 @@ func TestNormalizeResponsesInputAsMessagesFunctionCallItemPreservesConcatenatedA
 	}
 }
 
+func TestCollectOpenAIRefFileIDs(t *testing.T) {
+	got := collectOpenAIRefFileIDs(map[string]any{
+		"ref_file_ids": []any{"file-top", "file-dup"},
+		"attachments": []any{
+			map[string]any{"file_id": "file-attachment"},
+		},
+		"input": []any{
+			map[string]any{
+				"type": "message",
+				"content": []any{
+					map[string]any{"type": "input_file", "file_id": "file-input"},
+					map[string]any{"type": "input_file", "id": "file-dup"},
+				},
+			},
+		},
+	})
+	want := []string{"file-top", "file-dup", "file-attachment", "file-input"}
+	if len(got) != len(want) {
+		t.Fatalf("expected %d file ids, got %#v", len(want), got)
+	}
+	for i, id := range want {
+		if got[i] != id {
+			t.Fatalf("unexpected file ids at %d: got=%#v want=%#v", i, got, want)
+		}
+	}
+}
+
 func TestExtractEmbeddingInputs(t *testing.T) {
 	got := extractEmbeddingInputs([]any{"a", "b"})
 	if len(got) != 2 || got[0] != "a" || got[1] != "b" {

internal/adapter/openai/responses_handler.go

@@ -1,6 +1,7 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -64,11 +65,20 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	traceID := requestTraceID(r)
 	stdReq, err := normalizeOpenAIResponsesRequest(h.Store, req, traceID)
 	if err != nil {
@@ -102,19 +112,27 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 		h.handleResponsesStream(w, r, resp, owner, responseID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolChoice, traceID)
 		return
 	}
-	h.handleResponsesNonStream(w, resp, owner, responseID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.ToolNames, stdReq.ToolChoice, traceID)
+	h.handleResponsesNonStream(w, resp, owner, responseID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolChoice, traceID)
 }
 
-func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled bool, toolNames []string, toolChoice util.ToolChoicePolicy, traceID string) {
-	defer resp.Body.Close()
+func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolChoice util.ToolChoicePolicy, traceID string) {
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeOpenAIError(w, resp.StatusCode, strings.TrimSpace(string(body)))
 		return
 	}
 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	sanitizedText := sanitizeLeakedOutput(result.Text)
-	textParsed := util.ParseStandaloneToolCallsDetailed(sanitizedText, toolNames)
+	stripReferenceMarkers := h.compatStripReferenceMarkers()
+	sanitizedThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
+	sanitizedText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
+	if searchEnabled {
+		sanitizedText = replaceCitationMarkersWithLinks(sanitizedText, result.CitationLinks)
+	}
+	if writeUpstreamEmptyOutputError(w, sanitizedText, result.ContentFilter) {
+		return
+	}
+	textParsed := toolcall.ParseStandaloneToolCallsDetailed(sanitizedText, toolNames)
 	logResponsesToolPolicyRejection(traceID, toolChoice, textParsed, "text")
 
 	callCount := len(textParsed.Calls)
@@ -123,21 +141,13 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 		return
 	}
 
-	responseObj := openaifmt.BuildResponseObject(responseID, model, finalPrompt, result.Thinking, sanitizedText, toolNames)
-	if result.OutputTokens > 0 {
-		if usage, ok := responseObj["usage"].(map[string]any); ok {
-			usage["output_tokens"] = result.OutputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + result.OutputTokens
-			}
-		}
-	}
+	responseObj := openaifmt.BuildResponseObject(responseID, model, finalPrompt, sanitizedThinking, sanitizedText, toolNames)
 	h.getResponseStore().put(owner, responseID, responseObj)
 	writeJSON(w, http.StatusOK, responseObj)
 }
 
 func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolChoice util.ToolChoicePolicy, traceID string) {
-	defer resp.Body.Close()
+	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		writeOpenAIError(w, resp.StatusCode, strings.TrimSpace(string(body)))
@@ -156,6 +166,7 @@ func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request,
 	}
 	bufferToolContent := len(toolNames) > 0
 	emitEarlyToolDeltas := h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence()
+	stripReferenceMarkers := h.compatStripReferenceMarkers()
 
 	streamRuntime := newResponsesStreamRuntime(
 		w,
@@ -166,6 +177,7 @@ func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request,
 		finalPrompt,
 		thinkingEnabled,
 		searchEnabled,
+		stripReferenceMarkers,
 		toolNames,
 		bufferToolContent,
 		emitEarlyToolDeltas,
@@ -193,7 +205,7 @@ func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request,
 	})
 }
 
-func logResponsesToolPolicyRejection(traceID string, policy util.ToolChoicePolicy, parsed util.ToolCallParseResult, channel string) {
+func logResponsesToolPolicyRejection(traceID string, policy util.ToolChoicePolicy, parsed toolcall.ToolCallParseResult, channel string) {
 	rejected := filteredRejectedToolNamesForLog(parsed.RejectedToolNames)
 	if !parsed.RejectedByPolicy || len(rejected) == 0 {
 		return

internal/adapter/openai/responses_stream_runtime_core.go

@@ -1,6 +1,7 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"net/http"
 	"strings"
 
@@ -23,8 +24,9 @@ type responsesStreamRuntime struct {
 	traceID     string
 	toolChoice  util.ToolChoicePolicy
 
-	thinkingEnabled bool
-	searchEnabled   bool
+	thinkingEnabled       bool
+	searchEnabled         bool
+	stripReferenceMarkers bool
 
 	bufferToolContent    bool
 	emitEarlyToolDeltas  bool
@@ -49,7 +51,6 @@ type responsesStreamRuntime struct {
 	messagePartAdded  bool
 	sequence          int
 	failed            bool
-	outputTokens      int
 
 	persistResponse func(obj map[string]any)
 }
@@ -63,6 +64,7 @@ func newResponsesStreamRuntime(
 	finalPrompt string,
 	thinkingEnabled bool,
 	searchEnabled bool,
+	stripReferenceMarkers bool,
 	toolNames []string,
 	bufferToolContent bool,
 	emitEarlyToolDeltas bool,
@@ -71,40 +73,65 @@ func newResponsesStreamRuntime(
 	persistResponse func(obj map[string]any),
 ) *responsesStreamRuntime {
 	return &responsesStreamRuntime{
-		w:                   w,
-		rc:                  rc,
-		canFlush:            canFlush,
-		responseID:          responseID,
-		model:               model,
-		finalPrompt:         finalPrompt,
-		thinkingEnabled:     thinkingEnabled,
-		searchEnabled:       searchEnabled,
-		toolNames:           toolNames,
-		bufferToolContent:   bufferToolContent,
-		emitEarlyToolDeltas: emitEarlyToolDeltas,
-		streamToolCallIDs:   map[int]string{},
-		functionItemIDs:     map[int]string{},
-		functionOutputIDs:   map[int]int{},
-		functionArgs:        map[int]string{},
-		functionDone:        map[int]bool{},
-		functionAdded:       map[int]bool{},
-		functionNames:       map[int]string{},
-		messageOutputID:     -1,
-		toolChoice:          toolChoice,
-		traceID:             traceID,
-		persistResponse:     persistResponse,
+		w:                     w,
+		rc:                    rc,
+		canFlush:              canFlush,
+		responseID:            responseID,
+		model:                 model,
+		finalPrompt:           finalPrompt,
+		thinkingEnabled:       thinkingEnabled,
+		searchEnabled:         searchEnabled,
+		stripReferenceMarkers: stripReferenceMarkers,
+		toolNames:             toolNames,
+		bufferToolContent:     bufferToolContent,
+		emitEarlyToolDeltas:   emitEarlyToolDeltas,
+		streamToolCallIDs:     map[int]string{},
+		functionItemIDs:       map[int]string{},
+		functionOutputIDs:     map[int]int{},
+		functionArgs:          map[int]string{},
+		functionDone:          map[int]bool{},
+		functionAdded:         map[int]bool{},
+		functionNames:         map[int]string{},
+		messageOutputID:       -1,
+		toolChoice:            toolChoice,
+		traceID:               traceID,
+		persistResponse:       persistResponse,
 	}
 }
 
+func (s *responsesStreamRuntime) failResponse(message, code string) {
+	s.failed = true
+	failedResp := map[string]any{
+		"id":          s.responseID,
+		"type":        "response",
+		"object":      "response",
+		"model":       s.model,
+		"status":      "failed",
+		"output":      []any{},
+		"output_text": "",
+		"error": map[string]any{
+			"message": message,
+			"type":    "invalid_request_error",
+			"code":    code,
+			"param":   nil,
+		},
+	}
+	if s.persistResponse != nil {
+		s.persistResponse(failedResp)
+	}
+	s.sendEvent("response.failed", openaifmt.BuildResponsesFailedPayload(s.responseID, s.model, message, code))
+	s.sendDone()
+}
+
 func (s *responsesStreamRuntime) finalize() {
 	finalThinking := s.thinking.String()
-	finalText := sanitizeLeakedOutput(s.text.String())
+	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
 
 	if s.bufferToolContent {
 		s.processToolStreamEvents(flushToolSieve(&s.sieve, s.toolNames), true)
 	}
 
-	textParsed := util.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
+	textParsed := toolcall.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
 	detected := textParsed.Calls
 	s.logToolPolicyRejections(textParsed)
 
@@ -118,41 +145,21 @@ func (s *responsesStreamRuntime) finalize() {
 	s.closeMessageItem()
 
 	if s.toolChoice.IsRequired() && len(detected) == 0 {
-		s.failed = true
-		message := "tool_choice requires at least one valid tool call."
-		failedResp := map[string]any{
-			"id":          s.responseID,
-			"type":        "response",
-			"object":      "response",
-			"model":       s.model,
-			"status":      "failed",
-			"output":      []any{},
-			"output_text": "",
-			"error": map[string]any{
-				"message": message,
-				"type":    "invalid_request_error",
-				"code":    "tool_choice_violation",
-				"param":   nil,
-			},
+		s.failResponse("tool_choice requires at least one valid tool call.", "tool_choice_violation")
+		return
+	}
+	if len(detected) == 0 && strings.TrimSpace(finalText) == "" {
+		code := "upstream_empty_output"
+		message := "Upstream model returned empty output."
+		if finalThinking != "" {
+			message = "Upstream model returned reasoning without visible output."
 		}
-		if s.persistResponse != nil {
-			s.persistResponse(failedResp)
-		}
-		s.sendEvent("response.failed", openaifmt.BuildResponsesFailedPayload(s.responseID, s.model, message, "tool_choice_violation"))
-		s.sendDone()
+		s.failResponse(message, code)
 		return
 	}
 	s.closeIncompleteFunctionItems()
 
 	obj := s.buildCompletedResponseObject(finalThinking, finalText, detected)
-	if s.outputTokens > 0 {
-		if usage, ok := obj["usage"].(map[string]any); ok {
-			usage["output_tokens"] = s.outputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + s.outputTokens
-			}
-		}
-	}
 	if s.persistResponse != nil {
 		s.persistResponse(obj)
 	}
@@ -160,8 +167,8 @@ func (s *responsesStreamRuntime) finalize() {
 	s.sendDone()
 }
 
-func (s *responsesStreamRuntime) logToolPolicyRejections(textParsed util.ToolCallParseResult) {
-	logRejected := func(parsed util.ToolCallParseResult, channel string) {
+func (s *responsesStreamRuntime) logToolPolicyRejections(textParsed toolcall.ToolCallParseResult) {
+	logRejected := func(parsed toolcall.ToolCallParseResult, channel string) {
 		rejected := filteredRejectedToolNamesForLog(parsed.RejectedToolNames)
 		if !parsed.RejectedByPolicy || len(rejected) == 0 {
 			return
@@ -181,19 +188,17 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" || parsed.Stop {
 		return streamengine.ParsedDecision{Stop: true}
 	}
 
 	contentSeen := false
 	for _, p := range parsed.Parts {
-		if p.Text == "" {
+		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
+		if cleanedText == "" {
 			continue
 		}
-		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(p.Text) {
+		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(cleanedText) {
 			continue
 		}
 		contentSeen = true
@@ -201,21 +206,25 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 			if !s.thinkingEnabled {
 				continue
 			}
-			s.thinking.WriteString(p.Text)
-			s.sendEvent("response.reasoning.delta", openaifmt.BuildResponsesReasoningDeltaPayload(s.responseID, p.Text))
+			trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
+			if trimmed == "" {
+				continue
+			}
+			s.thinking.WriteString(trimmed)
+			s.sendEvent("response.reasoning.delta", openaifmt.BuildResponsesReasoningDeltaPayload(s.responseID, trimmed))
 			continue
 		}
 
-		cleanedText := sanitizeLeakedOutput(p.Text)
-		if cleanedText == "" {
+		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
+		if trimmed == "" {
 			continue
 		}
-		s.text.WriteString(cleanedText)
+		s.text.WriteString(trimmed)
 		if !s.bufferToolContent {
-			s.emitTextDelta(cleanedText)
+			s.emitTextDelta(trimmed)
 			continue
 		}
-		s.processToolStreamEvents(processToolSieveChunk(&s.sieve, cleanedText, s.toolNames), true)
+		s.processToolStreamEvents(processToolSieveChunk(&s.sieve, trimmed, s.toolNames), true)
 	}
 
 	return streamengine.ParsedDecision{ContentSeen: contentSeen}

internal/adapter/openai/responses_stream_runtime_events.go

@@ -48,7 +48,7 @@ func (s *responsesStreamRuntime) processToolStreamEvents(events []toolStreamEven
 			if !s.emitEarlyToolDeltas {
 				continue
 			}
-			filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.toolNames, s.functionNames)
+			filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.functionNames)
 			if len(filtered) == 0 {
 				continue
 			}

internal/adapter/openai/responses_stream_runtime_toolcalls.go

@@ -1,11 +1,11 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"strings"
 
 	openaifmt "ds2api/internal/format/openai"
-	"ds2api/internal/util"
 
 	"github.com/google/uuid"
 )
@@ -69,7 +69,7 @@ func (s *responsesStreamRuntime) ensureMessageContentPartAdded() {
 }
 
 func (s *responsesStreamRuntime) emitTextDelta(content string) {
-	if strings.TrimSpace(content) == "" {
+	if content == "" {
 		return
 	}
 	s.ensureMessageContentPartAdded()
@@ -208,7 +208,7 @@ func (s *responsesStreamRuntime) emitFunctionCallDeltaEvents(deltas []toolCallDe
 	}
 }
 
-func (s *responsesStreamRuntime) emitFunctionCallDoneEvents(calls []util.ParsedToolCall) {
+func (s *responsesStreamRuntime) emitFunctionCallDoneEvents(calls []toolcall.ParsedToolCall) {
 	for idx, tc := range calls {
 		if strings.TrimSpace(tc.Name) == "" {
 			continue

internal/adapter/openai/responses_stream_runtime_toolcalls_finalize.go

@@ -1,12 +1,12 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"encoding/json"
 	"sort"
 	"strings"
 
 	openaifmt "ds2api/internal/format/openai"
-	"ds2api/internal/util"
 )
 
 func (s *responsesStreamRuntime) closeIncompleteFunctionItems() {
@@ -57,7 +57,7 @@ func (s *responsesStreamRuntime) closeIncompleteFunctionItems() {
 	}
 }
 
-func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, finalText string, calls []util.ParsedToolCall) map[string]any {
+func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, finalText string, calls []toolcall.ParsedToolCall) map[string]any {
 	type indexedItem struct {
 		index int
 		item  map[string]any
@@ -83,13 +83,13 @@ func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, fin
 		})
 	} else if len(calls) == 0 {
 		content := make([]map[string]any, 0, 2)
-		if strings.TrimSpace(finalThinking) != "" {
+		if finalThinking != "" {
 			content = append(content, map[string]any{
 				"type": "reasoning",
 				"text": finalThinking,
 			})
 		}
-		if strings.TrimSpace(finalText) != "" {
+		if finalText != "" {
 			content = append(content, map[string]any{
 				"type": "output_text",
 				"text": finalText,
@@ -136,10 +136,10 @@ func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, fin
 	}
 
 	outputText := s.visibleText.String()
-	if strings.TrimSpace(outputText) == "" && len(calls) == 0 {
-		if strings.TrimSpace(finalText) != "" {
+	if outputText == "" && len(calls) == 0 {
+		if finalText != "" {
 			outputText = finalText
-		} else if strings.TrimSpace(finalThinking) != "" {
+		} else if finalThinking != "" {
 			outputText = finalThinking
 		}
 	}

internal/adapter/openai/responses_stream_test.go

@@ -12,149 +12,6 @@ import (
 	"ds2api/internal/util"
 )
 
-func TestHandleResponsesStreamToolCallsHideRawOutputTextInCompleted(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	rawToolJSON := `{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}`
-	streamBody := sseLine(rawToolJSON) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
-
-	completed, ok := extractSSEEventPayload(rec.Body.String(), "response.completed")
-	if !ok {
-		t.Fatalf("expected response.completed event, body=%s", rec.Body.String())
-	}
-	responseObj, _ := completed["response"].(map[string]any)
-	outputText, _ := responseObj["output_text"].(string)
-	if outputText != "" {
-		t.Fatalf("expected empty output_text for tool_calls response, got output_text=%q", outputText)
-	}
-	output, _ := responseObj["output"].([]any)
-	if len(output) == 0 {
-		t.Fatalf("expected structured output entries, got %#v", responseObj["output"])
-	}
-	hasFunctionCall := false
-	hasLegacyWrapper := false
-	for _, item := range output {
-		m, _ := item.(map[string]any)
-		if m == nil {
-			continue
-		}
-		if m["type"] == "function_call" {
-			hasFunctionCall = true
-		}
-		if m["type"] == "tool_calls" {
-			hasLegacyWrapper = true
-		}
-	}
-	if !hasFunctionCall {
-		t.Fatalf("expected function_call item, got %#v", responseObj["output"])
-	}
-	if hasLegacyWrapper {
-		t.Fatalf("did not expect legacy tool_calls wrapper, got %#v", responseObj["output"])
-	}
-	if strings.Contains(outputText, `"tool_calls"`) {
-		t.Fatalf("raw tool_calls JSON leaked in output_text: %q", outputText)
-	}
-}
-
-func TestHandleResponsesStreamUsesOfficialOutputItemEvents(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	streamBody := sseLine(`{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}`) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
-	body := rec.Body.String()
-	if !strings.Contains(body, "event: response.output_item.added") {
-		t.Fatalf("expected response.output_item.added event, body=%s", body)
-	}
-	if !strings.Contains(body, "event: response.output_item.done") {
-		t.Fatalf("expected response.output_item.done event, body=%s", body)
-	}
-	if !strings.Contains(body, "event: response.function_call_arguments.done") {
-		t.Fatalf("expected response.function_call_arguments.done event, body=%s", body)
-	}
-	if strings.Contains(body, "event: response.output_tool_call.delta") || strings.Contains(body, "event: response.output_tool_call.done") {
-		t.Fatalf("legacy response.output_tool_call.* event must not appear, body=%s", body)
-	}
-
-	addedPayloads := extractAllSSEEventPayloads(body, "response.output_item.added")
-	hasFunctionCallAdded := false
-	for _, payload := range addedPayloads {
-		item, _ := payload["item"].(map[string]any)
-		if item == nil || asString(item["type"]) != "function_call" {
-			continue
-		}
-		hasFunctionCallAdded = true
-		if asString(item["arguments"]) != "" {
-			t.Fatalf("expected in-progress function_call.arguments to start empty string, got %#v", item["arguments"])
-		}
-	}
-	if !hasFunctionCallAdded {
-		t.Fatalf("expected function_call output_item.added payload, body=%s", body)
-	}
-
-	donePayload, ok := extractSSEEventPayload(body, "response.function_call_arguments.done")
-	if !ok {
-		t.Fatalf("expected to parse response.function_call_arguments.done payload, body=%s", body)
-	}
-	doneCallID := strings.TrimSpace(asString(donePayload["call_id"]))
-	if doneCallID == "" {
-		t.Fatalf("expected non-empty call_id in done payload, payload=%#v", donePayload)
-	}
-	completed, ok := extractSSEEventPayload(body, "response.completed")
-	if !ok {
-		t.Fatalf("expected response.completed payload, body=%s", body)
-	}
-	responseObj, _ := completed["response"].(map[string]any)
-	output, _ := responseObj["output"].([]any)
-	var completedCallID string
-	for _, item := range output {
-		m, _ := item.(map[string]any)
-		if m == nil || m["type"] != "function_call" {
-			continue
-		}
-		completedCallID = strings.TrimSpace(asString(m["call_id"]))
-		if completedCallID != "" {
-			break
-		}
-	}
-	if completedCallID == "" {
-		t.Fatalf("expected function_call.call_id in completed output, output=%#v", output)
-	}
-	if completedCallID != doneCallID {
-		t.Fatalf("expected completed call_id to match stream done call_id, done=%q completed=%q", doneCallID, completedCallID)
-	}
-}
-
 func TestHandleResponsesStreamDoesNotEmitReasoningTextCompatEvents(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -181,51 +38,6 @@ func TestHandleResponsesStreamDoesNotEmitReasoningTextCompatEvents(t *testing.T)
 	}
 }
 
-func TestHandleResponsesStreamMultiToolCallKeepsNameAndCallIDAligned(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	streamBody := sseLine(`{"tool_calls":[{"name":"search_web","input":{"query":"latest ai news"}},`) +
-		sseLine(`{"name":"eval_javascript","input":{"code":"1+1"}}]}`) +
-		"data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"search_web", "eval_javascript"}, util.DefaultToolChoicePolicy(), "")
-
-	body := rec.Body.String()
-	donePayloads := extractAllSSEEventPayloads(body, "response.function_call_arguments.done")
-	if len(donePayloads) != 2 {
-		t.Fatalf("expected two response.function_call_arguments.done events, got %d body=%s", len(donePayloads), body)
-	}
-	seenNames := map[string]string{}
-	for _, payload := range donePayloads {
-		name := strings.TrimSpace(asString(payload["name"]))
-		callID := strings.TrimSpace(asString(payload["call_id"]))
-		if name != "search_web" && name != "eval_javascript" {
-			t.Fatalf("unexpected tool name in done payload: %#v", payload)
-		}
-		if callID == "" {
-			t.Fatalf("expected non-empty call_id in done payload: %#v", payload)
-		}
-		seenNames[name] = callID
-	}
-	if seenNames["search_web"] == seenNames["eval_javascript"] {
-		t.Fatalf("expected distinct call_id per tool, got %#v", seenNames)
-	}
-}
-
 func TestHandleResponsesStreamEmitsOutputTextDoneBeforeContentPartDone(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -297,123 +109,6 @@ func TestHandleResponsesStreamOutputTextDeltaCarriesItemIndexes(t *testing.T) {
 	}
 }
 
-func TestHandleResponsesStreamThinkingAndMixedToolExampleEmitsFunctionCall(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(path, value string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": path,
-			"v": value,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	streamBody := sseLine("response/thinking_content", "thinking...") +
-		sseLine("response/content", "先读取文件。") +
-		sseLine("response/content", `{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}`) +
-		"data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
-
-	addedPayloads := extractAllSSEEventPayloads(rec.Body.String(), "response.output_item.added")
-	if len(addedPayloads) < 1 {
-		t.Fatalf("expected at least one output_item.added event, got %d body=%s", len(addedPayloads), rec.Body.String())
-	}
-
-	completedPayload, ok := extractSSEEventPayload(rec.Body.String(), "response.completed")
-	if !ok {
-		t.Fatalf("expected response.completed payload, body=%s", rec.Body.String())
-	}
-	responseObj, _ := completedPayload["response"].(map[string]any)
-	output, _ := responseObj["output"].([]any)
-	hasMessage := false
-	hasFunctionCall := false
-	for _, item := range output {
-		m, _ := item.(map[string]any)
-		if m == nil {
-			continue
-		}
-		if asString(m["type"]) == "message" {
-			hasMessage = true
-		}
-		if asString(m["type"]) == "function_call" {
-			hasFunctionCall = true
-		}
-	}
-	if !hasMessage {
-		t.Fatalf("expected message output for mixed prose tool example, output=%#v", output)
-	}
-	if !hasFunctionCall {
-		t.Fatalf("expected function_call output for mixed prose tool example, output=%#v", output)
-	}
-}
-
-func TestHandleResponsesStreamToolChoiceNoneStillAllowsFunctionCall(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	streamBody := sseLine(`{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}`) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-	policy := util.ToolChoicePolicy{Mode: util.ToolChoiceNone}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, nil, policy, "")
-	body := rec.Body.String()
-	if !strings.Contains(body, "event: response.function_call_arguments.done") {
-		t.Fatalf("expected function_call events for tool_choice=none, body=%s", body)
-	}
-}
-
-func TestHandleResponsesStreamMalformedToolJSONFallsBackToText(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	// invalid JSON (NaN) should remain plain text in strict mode.
-	streamBody := sseLine(`{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"},"x":NaN}]}`) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
-	body := rec.Body.String()
-	if strings.Contains(body, "event: response.function_call_arguments.delta") || strings.Contains(body, "event: response.function_call_arguments.done") {
-		t.Fatalf("did not expect function_call events for malformed payload in strict mode, body=%s", body)
-	}
-	if !strings.Contains(body, "event: response.output_text.delta") {
-		t.Fatalf("expected response.output_text.delta for malformed payload, body=%s", body)
-	}
-	if !strings.Contains(body, "event: response.completed") {
-		t.Fatalf("expected response.completed event, body=%s", body)
-	}
-}
-
 func TestHandleResponsesStreamRequiredToolChoiceFailure(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -448,7 +143,7 @@ func TestHandleResponsesStreamRequiredToolChoiceFailure(t *testing.T) {
 	}
 }
 
-func TestHandleResponsesStreamRequiredToolChoiceIgnoresThinkingToolPayload(t *testing.T) {
+func TestHandleResponsesStreamFailsWhenUpstreamHasOnlyThinking(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
 	rec := httptest.NewRecorder()
@@ -461,53 +156,13 @@ func TestHandleResponsesStreamRequiredToolChoiceIgnoresThinkingToolPayload(t *te
 		return "data: " + string(b) + "\n"
 	}
 
-	streamBody := sseLine("response/thinking_content", `{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}`) +
-		sseLine("response/content", "plain text only") +
-		"data: [DONE]\n"
+	streamBody := sseLine("response/thinking_content", "Only thinking") + "data: [DONE]\n"
 	resp := &http.Response{
 		StatusCode: http.StatusOK,
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	policy := util.ToolChoicePolicy{
-		Mode:    util.ToolChoiceRequired,
-		Allowed: map[string]struct{}{"read_file": {}},
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", true, false, []string{"read_file"}, policy, "")
-	body := rec.Body.String()
-	if !strings.Contains(body, "event: response.failed") {
-		t.Fatalf("expected response.failed event for required tool_choice violation, body=%s", body)
-	}
-	if strings.Contains(body, "event: response.completed") {
-		t.Fatalf("did not expect response.completed after failure, body=%s", body)
-	}
-}
-
-func TestHandleResponsesStreamRequiredMalformedToolPayloadFails(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
-	}
-
-	streamBody := sseLine(`{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"},"x":NaN}]}`) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-	policy := util.ToolChoicePolicy{
-		Mode:    util.ToolChoiceRequired,
-		Allowed: map[string]struct{}{"read_file": {}},
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, policy, "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, false, nil, util.DefaultToolChoicePolicy(), "")
 
 	body := rec.Body.String()
 	if !strings.Contains(body, "event: response.failed") {
@@ -516,31 +171,13 @@ func TestHandleResponsesStreamRequiredMalformedToolPayloadFails(t *testing.T) {
 	if strings.Contains(body, "event: response.completed") {
 		t.Fatalf("did not expect response.completed, body=%s", body)
 	}
-}
-
-func TestHandleResponsesStreamAllowsUnknownToolName(t *testing.T) {
-	h := &Handler{}
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
-	rec := httptest.NewRecorder()
-
-	sseLine := func(v string) string {
-		b, _ := json.Marshal(map[string]any{
-			"p": "response/content",
-			"v": v,
-		})
-		return "data: " + string(b) + "\n"
+	payload, ok := extractSSEEventPayload(body, "response.failed")
+	if !ok {
+		t.Fatalf("expected response.failed payload, body=%s", body)
 	}
-
-	streamBody := sseLine(`{"tool_calls":[{"name":"not_in_schema","input":{"q":"go"}}]}`) + "data: [DONE]\n"
-	resp := &http.Response{
-		StatusCode: http.StatusOK,
-		Body:       io.NopCloser(strings.NewReader(streamBody)),
-	}
-
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, util.DefaultToolChoicePolicy(), "")
-	body := rec.Body.String()
-	if !strings.Contains(body, "event: response.function_call_arguments.done") {
-		t.Fatalf("expected function_call events for unknown tool, body=%s", body)
+	errObj, _ := payload["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", payload)
 	}
 }
 
@@ -559,7 +196,7 @@ func TestHandleResponsesNonStreamRequiredToolChoiceViolation(t *testing.T) {
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, []string{"read_file"}, policy, "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, []string{"read_file"}, policy, "")
 	if rec.Code != http.StatusUnprocessableEntity {
 		t.Fatalf("expected 422 for required tool_choice violation, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -586,7 +223,7 @@ func TestHandleResponsesNonStreamRequiredToolChoiceIgnoresThinkingToolPayload(t
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", true, []string{"read_file"}, policy, "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", true, false, []string{"read_file"}, policy, "")
 	if rec.Code != http.StatusUnprocessableEntity {
 		t.Fatalf("expected 422 for required tool_choice violation, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -597,33 +234,69 @@ func TestHandleResponsesNonStreamRequiredToolChoiceIgnoresThinkingToolPayload(t
 	}
 }
 
-func TestHandleResponsesNonStreamToolChoiceNoneStillAllowsFunctionCall(t *testing.T) {
+func TestHandleResponsesNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	rec := httptest.NewRecorder()
 	resp := &http.Response{
 		StatusCode: http.StatusOK,
 		Body: io.NopCloser(strings.NewReader(
-			`data: {"p":"response/content","v":"{\"tool_calls\":[{\"name\":\"read_file\",\"input\":{\"path\":\"README.MD\"}}]}"}` + "\n" +
+			`data: {"p":"response/content","v":""}` + "\n" +
 				`data: [DONE]` + "\n",
 		)),
 	}
-	policy := util.ToolChoicePolicy{Mode: util.ToolChoiceNone}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, nil, policy, "")
-	if rec.Code != http.StatusOK {
-		t.Fatalf("expected 200 for tool_choice=none handling, got %d body=%s", rec.Code, rec.Body.String())
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, nil, util.DefaultToolChoicePolicy(), "")
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
 	out := decodeJSONBody(t, rec.Body.String())
-	output, _ := out["output"].([]any)
-	foundFunctionCall := false
-	for _, item := range output {
-		m, _ := item.(map[string]any)
-		if m != nil && m["type"] == "function_call" {
-			foundFunctionCall = true
-		}
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
-	if !foundFunctionCall {
-		t.Fatalf("expected function_call output item for tool_choice=none, got %#v", output)
+}
+
+func TestHandleResponsesNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
+	h := &Handler{}
+	rec := httptest.NewRecorder()
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body: io.NopCloser(strings.NewReader(
+			`data: {"code":"content_filter"}` + "\n" +
+				`data: [DONE]` + "\n",
+		)),
+	}
+
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, false, nil, util.DefaultToolChoicePolicy(), "")
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400 for filtered empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "content_filter" {
+		t.Fatalf("expected code=content_filter, got %#v", out)
+	}
+}
+
+func TestHandleResponsesNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
+	h := &Handler{}
+	rec := httptest.NewRecorder()
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body: io.NopCloser(strings.NewReader(
+			`data: {"p":"response/thinking_content","v":"Only thinking"}` + "\n" +
+				`data: [DONE]` + "\n",
+		)),
+	}
+
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, false, nil, util.DefaultToolChoicePolicy(), "")
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
 	}
 }
 
@@ -652,30 +325,3 @@ func extractSSEEventPayload(body, targetEvent string) (map[string]any, bool) {
 	}
 	return nil, false
 }
-
-func extractAllSSEEventPayloads(body, targetEvent string) []map[string]any {
-	scanner := bufio.NewScanner(strings.NewReader(body))
-	matched := false
-	out := make([]map[string]any, 0, 2)
-	for scanner.Scan() {
-		line := strings.TrimSpace(scanner.Text())
-		if strings.HasPrefix(line, "event: ") {
-			evt := strings.TrimSpace(strings.TrimPrefix(line, "event: "))
-			matched = evt == targetEvent
-			continue
-		}
-		if !matched || !strings.HasPrefix(line, "data: ") {
-			continue
-		}
-		raw := strings.TrimSpace(strings.TrimPrefix(line, "data: "))
-		if raw == "" || raw == "[DONE]" {
-			continue
-		}
-		var payload map[string]any
-		if err := json.Unmarshal([]byte(raw), &payload); err != nil {
-			continue
-		}
-		out = append(out, payload)
-	}
-	return out
-}

internal/adapter/openai/standard_request.go

@@ -12,11 +12,11 @@ func normalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID
 	model, _ := req["model"].(string)
 	messagesRaw, _ := req["messages"].([]any)
 	if strings.TrimSpace(model) == "" || len(messagesRaw) == 0 {
-		return util.StandardRequest{}, fmt.Errorf("Request must include 'model' and 'messages'.")
+		return util.StandardRequest{}, fmt.Errorf("request must include 'model' and 'messages'")
 	}
 	resolvedModel, ok := config.ResolveModel(store, model)
 	if !ok {
-		return util.StandardRequest{}, fmt.Errorf("Model '%s' is not available.", model)
+		return util.StandardRequest{}, fmt.Errorf("model %q is not available", model)
 	}
 	thinkingEnabled, searchEnabled, _ := config.GetModelConfig(resolvedModel)
 	responseModel := strings.TrimSpace(model)
@@ -24,9 +24,10 @@ func normalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID
 		responseModel = resolvedModel
 	}
 	toolPolicy := util.DefaultToolChoicePolicy()
-	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy)
+	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy, thinkingEnabled)
 	toolNames = ensureToolDetectionEnabled(toolNames, req["tools"])
 	passThrough := collectOpenAIChatPassThrough(req)
+	refFileIDs := collectOpenAIRefFileIDs(req)
 
 	return util.StandardRequest{
 		Surface:        "openai_chat",
@@ -40,6 +41,7 @@ func normalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID
 		Stream:         util.ToBool(req["stream"]),
 		Thinking:       thinkingEnabled,
 		Search:         searchEnabled,
+		RefFileIDs:     refFileIDs,
 		PassThrough:    passThrough,
 	}, nil
 }
@@ -48,11 +50,11 @@ func normalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 	model, _ := req["model"].(string)
 	model = strings.TrimSpace(model)
 	if model == "" {
-		return util.StandardRequest{}, fmt.Errorf("Request must include 'model'.")
+		return util.StandardRequest{}, fmt.Errorf("request must include 'model'")
 	}
 	resolvedModel, ok := config.ResolveModel(store, model)
 	if !ok {
-		return util.StandardRequest{}, fmt.Errorf("Model '%s' is not available.", model)
+		return util.StandardRequest{}, fmt.Errorf("model %q is not available", model)
 	}
 	thinkingEnabled, searchEnabled, _ := config.GetModelConfig(resolvedModel)
 
@@ -68,18 +70,19 @@ func normalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 		messagesRaw = msgs
 	}
 	if len(messagesRaw) == 0 {
-		return util.StandardRequest{}, fmt.Errorf("Request must include 'input' or 'messages'.")
+		return util.StandardRequest{}, fmt.Errorf("request must include 'input' or 'messages'")
 	}
 	toolPolicy, err := parseToolChoicePolicy(req["tool_choice"], req["tools"])
 	if err != nil {
 		return util.StandardRequest{}, err
 	}
-	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy)
+	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy, thinkingEnabled)
 	toolNames = ensureToolDetectionEnabled(toolNames, req["tools"])
 	if !toolPolicy.IsNone() {
 		toolPolicy.Allowed = namesToSet(toolNames)
 	}
 	passThrough := collectOpenAIChatPassThrough(req)
+	refFileIDs := collectOpenAIRefFileIDs(req)
 
 	return util.StandardRequest{
 		Surface:        "openai_responses",
@@ -93,6 +96,7 @@ func normalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 		Stream:         util.ToBool(req["stream"]),
 		Thinking:       thinkingEnabled,
 		Search:         searchEnabled,
+		RefFileIDs:     refFileIDs,
 		PassThrough:    passThrough,
 	}, nil
 }
@@ -152,7 +156,7 @@ func parseToolChoicePolicy(toolChoiceRaw any, toolsRaw any) (util.ToolChoicePoli
 		case "required":
 			policy.Mode = util.ToolChoiceRequired
 		default:
-			return util.ToolChoicePolicy{}, fmt.Errorf("Unsupported tool_choice: %q", v)
+			return util.ToolChoicePolicy{}, fmt.Errorf("unsupported tool_choice: %q", v)
 		}
 	case map[string]any:
 		allowedOverride, hasAllowedOverride, err := parseAllowedToolNames(v["allowed_tools"])
@@ -198,7 +202,7 @@ func parseToolChoicePolicy(toolChoiceRaw any, toolsRaw any) (util.ToolChoicePoli
 			policy.ForcedName = name
 			policy.Allowed = namesToSet([]string{name})
 		default:
-			return util.ToolChoicePolicy{}, fmt.Errorf("Unsupported tool_choice.type: %q", typ)
+			return util.ToolChoicePolicy{}, fmt.Errorf("unsupported tool_choice.type: %q", typ)
 		}
 	default:
 		return util.ToolChoicePolicy{}, fmt.Errorf("tool_choice must be a string or object")
@@ -206,7 +210,7 @@ func parseToolChoicePolicy(toolChoiceRaw any, toolsRaw any) (util.ToolChoicePoli
 
 	if policy.Mode == util.ToolChoiceRequired || policy.Mode == util.ToolChoiceForced {
 		if len(declaredNames) == 0 {
-			return util.ToolChoicePolicy{}, fmt.Errorf("tool_choice=%s requires non-empty tools.", policy.Mode)
+			return util.ToolChoicePolicy{}, fmt.Errorf("tool_choice=%s requires non-empty tools", policy.Mode)
 		}
 	}
 	if policy.Mode == util.ToolChoiceForced {

internal/adapter/openai/standard_request_test.go

@@ -41,6 +41,36 @@ func TestNormalizeOpenAIChatRequest(t *testing.T) {
 	}
 }
 
+func TestNormalizeOpenAIChatRequestCollectsRefFileIDs(t *testing.T) {
+	store := newEmptyStoreForNormalizeTest(t)
+	req := map[string]any{
+		"model": "gpt-5-codex",
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{"type": "input_text", "text": "hello"},
+					map[string]any{"type": "input_file", "file_id": "file-msg"},
+				},
+			},
+		},
+		"attachments": []any{
+			map[string]any{"file_id": "file-attachment"},
+		},
+		"ref_file_ids": []any{"file-top", "file-attachment"},
+	}
+	n, err := normalizeOpenAIChatRequest(store, req, "")
+	if err != nil {
+		t.Fatalf("normalize failed: %v", err)
+	}
+	if len(n.RefFileIDs) != 3 {
+		t.Fatalf("expected 3 distinct file ids, got %#v", n.RefFileIDs)
+	}
+	if n.RefFileIDs[0] != "file-top" || n.RefFileIDs[1] != "file-attachment" || n.RefFileIDs[2] != "file-msg" {
+		t.Fatalf("unexpected file ids: %#v", n.RefFileIDs)
+	}
+}
+
 func TestNormalizeOpenAIResponsesRequestInput(t *testing.T) {
 	store := newEmptyStoreForNormalizeTest(t)
 	req := map[string]any{

internal/adapter/openai/stream_status_test.go

@@ -13,6 +13,7 @@ import (
 	chimw "github.com/go-chi/chi/v5/middleware"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
 )
 
 type streamStatusAuthStub struct{}
@@ -49,10 +50,18 @@ func (m streamStatusDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int
 	return "pow", nil
 }
 
+func (m streamStatusDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, _ deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	return &deepseek.UploadFileResult{ID: "file-id", Filename: "file.txt", Bytes: 1, Status: "uploaded"}, nil
+}
+
 func (m streamStatusDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
 	return m.resp, nil
 }
 
+func (m streamStatusDSStub) DeleteSessionForToken(_ context.Context, _ string, _ string) (*deepseek.DeleteSessionResult, error) {
+	return &deepseek.DeleteSessionResult{Success: true}, nil
+}
+
 func (m streamStatusDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
 	return nil
 }
@@ -137,53 +146,6 @@ func TestResponsesStreamStatusCapturedAs200(t *testing.T) {
 	}
 }
 
-func TestResponsesNonStreamMixedProseToolPayloadHandlerPath(t *testing.T) {
-	statuses := make([]int, 0, 1)
-	content, _ := json.Marshal(map[string]any{
-		"p": "response/content",
-		"v": "我来调用工具\n{\"tool_calls\":[{\"name\":\"read_file\",\"input\":{\"path\":\"README.MD\"}}]}",
-	})
-	h := &Handler{
-		Store: mockOpenAIConfig{wideInput: true},
-		Auth:  streamStatusAuthStub{},
-		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse("data: "+string(content), "data: [DONE]")},
-	}
-	r := chi.NewRouter()
-	r.Use(captureStatusMiddleware(&statuses))
-	RegisterRoutes(r, h)
-
-	reqBody := `{"model":"deepseek-chat","input":"请调用工具","tools":[{"type":"function","function":{"name":"read_file","description":"read","parameters":{"type":"object","properties":{"path":{"type":"string"}}}}}],"stream":false}`
-	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
-	req.Header.Set("Authorization", "Bearer direct-token")
-	req.Header.Set("Content-Type", "application/json")
-	rec := httptest.NewRecorder()
-	r.ServeHTTP(rec, req)
-
-	if rec.Code != http.StatusOK {
-		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
-	}
-	if len(statuses) != 1 || statuses[0] != http.StatusOK {
-		t.Fatalf("expected captured status 200, got %#v", statuses)
-	}
-
-	var out map[string]any
-	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
-		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
-	}
-	outputText, _ := out["output_text"].(string)
-	if outputText != "" {
-		t.Fatalf("expected output_text hidden for mixed prose tool payload, got %q", outputText)
-	}
-	output, _ := out["output"].([]any)
-	if len(output) != 1 {
-		t.Fatalf("expected one output item, got %#v", output)
-	}
-	first, _ := output[0].(map[string]any)
-	if first["type"] != "function_call" {
-		t.Fatalf("expected function_call output item, got %#v", output)
-	}
-}
-
 func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T) {
 	statuses := make([]int, 0, 1)
 	h := &Handler{
@@ -233,3 +195,140 @@ func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T
 		t.Fatalf("expected finish_reason=stop for content-filter upstream stop, got %#v", choice["finish_reason"])
 	}
 }
+
+func TestChatCompletionsStreamEmitsFailureFrameWhenUpstreamOutputEmpty(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse("data: [DONE]")},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":"hi"}],"stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	if len(frames) != 1 {
+		t.Fatalf("expected one failure frame, got %#v body=%s", frames, rec.Body.String())
+	}
+	last := frames[0]
+	statusCode, ok := last["status_code"].(float64)
+	if !ok || int(statusCode) != http.StatusTooManyRequests {
+		t.Fatalf("expected status_code=429, got %#v body=%s", last["status_code"], rec.Body.String())
+	}
+	errObj, _ := last["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", last)
+	}
+}
+
+func TestResponsesStreamUsageIgnoresBatchAccumulatedTokenUsage(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
+			`data: {"p":"response/content","v":"hello"}`,
+			`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":190},{"p":"quasi_status","v":"FINISHED"}]}`,
+		)},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","input":"hi","stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	if len(frames) == 0 {
+		t.Fatalf("expected at least one json frame, body=%s", rec.Body.String())
+	}
+	last := frames[len(frames)-1]
+	resp, _ := last["response"].(map[string]any)
+	if resp == nil {
+		t.Fatalf("expected response payload in final frame, got %#v", last)
+	}
+	usage, _ := resp["usage"].(map[string]any)
+	if usage == nil {
+		t.Fatalf("expected usage in response payload, got %#v", resp)
+	}
+	if got, _ := usage["output_tokens"].(float64); int(got) == 190 {
+		t.Fatalf("expected upstream accumulated token usage to be ignored, got %#v", usage["output_tokens"])
+	}
+}
+
+func TestResponsesNonStreamUsageIgnoresPromptAndOutputTokenUsage(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
+			`data: {"p":"response/content","v":"ok"}`,
+			`data: {"p":"response","o":"BATCH","v":[{"p":"token_usage","v":{"prompt_tokens":11,"completion_tokens":29}},{"p":"quasi_status","v":"FINISHED"}]}`,
+		)},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","input":"hi","stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	usage, _ := out["usage"].(map[string]any)
+	if usage == nil {
+		t.Fatalf("expected usage object, got %#v", out)
+	}
+	input, _ := usage["input_tokens"].(float64)
+	output, _ := usage["output_tokens"].(float64)
+	total, _ := usage["total_tokens"].(float64)
+	if int(output) == 29 {
+		t.Fatalf("expected upstream completion token usage to be ignored, got %#v", usage["output_tokens"])
+	}
+	if int(total) != int(input)+int(output) {
+		t.Fatalf("expected total_tokens=input_tokens+output_tokens, usage=%#v", usage)
+	}
+}

internal/adapter/openai/tool_sieve_core.go

@@ -3,7 +3,7 @@ package openai
 import (
 	"strings"
 
-	"ds2api/internal/util"
+	"ds2api/internal/toolcall"
 )
 
 func processToolSieveChunk(state *toolStreamSieveState, chunk string, toolNames []string) []toolStreamEvent {
@@ -60,7 +60,7 @@ func processToolSieveChunk(state *toolStreamSieveState, chunk string, toolNames
 		if pending == "" {
 			break
 		}
-		start := findToolSegmentStart(pending)
+		start := findToolSegmentStart(state, pending)
 		if start >= 0 {
 			prefix := pending[:start]
 			if prefix != "" {
@@ -74,7 +74,7 @@ func processToolSieveChunk(state *toolStreamSieveState, chunk string, toolNames
 			continue
 		}
 
-		safe, hold := splitSafeContentForToolDetection(pending)
+		safe, hold := splitSafeContentForToolDetection(state, pending)
 		if safe == "" {
 			break
 		}
@@ -114,14 +114,10 @@ func flushToolSieve(state *toolStreamSieveState, toolNames []string) []toolStrea
 		} else {
 			content := state.capture.String()
 			if content != "" {
-				// If the captured text looks like an incomplete XML tool call block,
-				// swallow it to prevent leaking raw XML tags to the client.
-				if hasOpenXMLToolTag(content) {
-					// Drop it silently — incomplete tool call.
-				} else {
-					state.noteText(content)
-					events = append(events, toolStreamEvent{Content: content})
-				}
+				// If capture never resolved into a real tool call, release the
+				// buffered text instead of swallowing it.
+				state.noteText(content)
+				events = append(events, toolStreamEvent{Content: content})
 			}
 		}
 		state.capture.Reset()
@@ -130,109 +126,66 @@ func flushToolSieve(state *toolStreamSieveState, toolNames []string) []toolStrea
 	}
 	if state.pending.Len() > 0 {
 		content := state.pending.String()
-		// Safety: if pending contains XML tool tag fragments (e.g. "tool_calls>"
-		// from a split closing tag), swallow them instead of leaking.
-		if hasOpenXMLToolTag(content) || looksLikeXMLToolTagFragment(content) {
-			// Drop it — likely an incomplete tool call fragment.
-		} else {
-			state.noteText(content)
-			events = append(events, toolStreamEvent{Content: content})
-		}
+		// If pending never resolved into a real tool call, release it as text.
+		state.noteText(content)
+		events = append(events, toolStreamEvent{Content: content})
 		state.pending.Reset()
 	}
 	return events
 }
 
-func splitSafeContentForToolDetection(s string) (safe, hold string) {
+func splitSafeContentForToolDetection(state *toolStreamSieveState, s string) (safe, hold string) {
 	if s == "" {
 		return "", ""
 	}
-	suspiciousStart := findSuspiciousPrefixStart(s)
-	if suspiciousStart < 0 {
-		return s, ""
-	}
-	if suspiciousStart > 0 {
-		return s[:suspiciousStart], s[suspiciousStart:]
-	}
-	// If suspicious content starts at position 0, keep holding until we can
-	// parse a complete tool JSON block or reach stream flush.
-	return "", s
-}
-
-func findSuspiciousPrefixStart(s string) int {
-	start := -1
-	indices := []int{
-		strings.LastIndex(s, "{"),
-		strings.LastIndex(s, "["),
-		strings.LastIndex(s, "```"),
-	}
-	for _, idx := range indices {
-		if idx > start {
-			start = idx
+	if xmlIdx := findPartialXMLToolTagStart(s); xmlIdx >= 0 {
+		if insideCodeFenceWithState(state, s[:xmlIdx]) {
+			return s, ""
 		}
+		if xmlIdx > 0 {
+			return s[:xmlIdx], s[xmlIdx:]
+		}
+		return "", s
 	}
-	// Also check for partial XML tool tag at end of string.
-	if xmlIdx := findPartialXMLToolTagStart(s); xmlIdx >= 0 && xmlIdx > start {
-		start = xmlIdx
-	}
-	return start
+	return s, ""
 }
 
-func findToolSegmentStart(s string) int {
+func findToolSegmentStart(state *toolStreamSieveState, s string) int {
 	if s == "" {
 		return -1
 	}
 	lower := strings.ToLower(s)
-	keywords := []string{"tool_calls", "\"function\"", "function.name:", "\"tool_use\""}
-	bestKeyIdx := -1
-	for _, kw := range keywords {
-		idx := strings.Index(lower, kw)
-		if idx >= 0 && (bestKeyIdx < 0 || idx < bestKeyIdx) {
-			bestKeyIdx = idx
+	offset := 0
+	for {
+		bestKeyIdx := -1
+		matchedTag := ""
+		for _, tag := range xmlToolTagsToDetect {
+			idx := strings.Index(lower[offset:], tag)
+			if idx >= 0 {
+				idx += offset
+				if bestKeyIdx < 0 || idx < bestKeyIdx {
+					bestKeyIdx = idx
+					matchedTag = tag
+				}
+			}
 		}
-	}
-	if fnKeyIdx := findQuotedFunctionCallKeyStart(s); fnKeyIdx >= 0 && (bestKeyIdx < 0 || fnKeyIdx < bestKeyIdx) {
-		bestKeyIdx = fnKeyIdx
-	}
-	// Also detect XML tool call tags.
-	for _, tag := range xmlToolTagsToDetect {
-		idx := strings.Index(lower, tag)
-		if idx >= 0 && (bestKeyIdx < 0 || idx < bestKeyIdx) {
-			bestKeyIdx = idx
+		if bestKeyIdx < 0 {
+			return -1
 		}
-	}
-	if bestKeyIdx < 0 {
-		return -1
-	}
-	// For XML tags, the '<' is itself the segment start.
-	if bestKeyIdx < len(s) && s[bestKeyIdx] == '<' {
-		if fenceStart, ok := openFenceStartBefore(s, bestKeyIdx); ok {
-			return fenceStart
+		if !insideCodeFenceWithState(state, s[:bestKeyIdx]) {
+			return bestKeyIdx
 		}
-		return bestKeyIdx
+		offset = bestKeyIdx + len(matchedTag)
 	}
-	start := strings.LastIndex(s[:bestKeyIdx], "{")
-	if start < 0 {
-		start = bestKeyIdx
-	}
-	// If the keyword matched inside an XML tag (e.g. "tool_calls" in "<tool_calls>"),
-	// back up past the '<' to capture the full tag.
-	if start > 0 && s[start-1] == '<' {
-		start--
-	}
-	if fenceStart, ok := openFenceStartBefore(s, start); ok {
-		return fenceStart
-	}
-	return start
 }
 
-func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix string, calls []util.ParsedToolCall, suffix string, ready bool) {
+func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix string, calls []toolcall.ParsedToolCall, suffix string, ready bool) {
 	captured := state.capture.String()
 	if captured == "" {
 		return "", nil, "", false
 	}
 
-	// Try XML tool call extraction first.
+	// XML tool call extraction only.
 	if xmlPrefix, xmlCalls, xmlSuffix, xmlReady := consumeXMLToolCapture(captured, toolNames); xmlReady {
 		return xmlPrefix, xmlCalls, xmlSuffix, true
 	}
@@ -240,45 +193,5 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	if hasOpenXMLToolTag(captured) {
 		return "", nil, "", false
 	}
-
-	lower := strings.ToLower(captured)
-	keyIdx := -1
-	keywords := []string{"tool_calls", "\"function\"", "function.name:", "\"tool_use\""}
-	for _, kw := range keywords {
-		idx := strings.Index(lower, kw)
-		if idx >= 0 && (keyIdx < 0 || idx < keyIdx) {
-			keyIdx = idx
-		}
-	}
-	if fnKeyIdx := findQuotedFunctionCallKeyStart(captured); fnKeyIdx >= 0 && (keyIdx < 0 || fnKeyIdx < keyIdx) {
-		keyIdx = fnKeyIdx
-	}
-
-	if keyIdx < 0 {
-		return "", nil, "", false
-	}
-	start := strings.LastIndex(captured[:keyIdx], "{")
-	if start < 0 {
-		start = keyIdx
-	}
-	obj, end, ok := extractJSONObjectFrom(captured, start)
-	if !ok {
-		return "", nil, "", false
-	}
-	prefixPart := captured[:start]
-	suffixPart := captured[end:]
-	parsed := util.ParseStandaloneToolCallsDetailed(obj, toolNames)
-	if len(parsed.Calls) == 0 {
-		if parsed.SawToolCallSyntax && parsed.RejectedByPolicy {
-			// Parsed as tool-call payload but rejected by schema/policy:
-			// consume it to avoid leaking raw tool_calls JSON to user content.
-			return prefixPart, nil, suffixPart, true
-		}
-		// If it has obvious keywords but failed to parse even after loose repair,
-		// we still might want to intercept it if it looks like an attempt at tool call.
-		// For now, keep the original logic but rely on loose JSON repair.
-		return captured, nil, "", true
-	}
-	prefixPart, suffixPart = trimWrappingJSONFence(prefixPart, suffixPart)
-	return prefixPart, parsed.Calls, suffixPart, true
+	return "", nil, "", false
 }

internal/adapter/openai/tool_sieve_functioncall.go

@@ -1,100 +0,0 @@
-package openai
-
-import "strings"
-
-func findQuotedFunctionCallKeyStart(s string) int {
-	lower := strings.ToLower(s)
-	quotedIdx := findFunctionCallKeyStart(lower, `"functioncall"`)
-	bareIdx := findFunctionCallKeyStart(lower, "functioncall")
-
-	// Prefer the quoted JSON key whenever we have a structural match.
-	// Bare-key detection is only for loose payloads where the quoted form
-	// is absent.
-	if quotedIdx >= 0 {
-		return quotedIdx
-	}
-	return bareIdx
-}
-
-func findFunctionCallKeyStart(lower, key string) int {
-	for from := 0; from < len(lower); {
-		rel := strings.Index(lower[from:], key)
-		if rel < 0 {
-			return -1
-		}
-		idx := from + rel
-		if isInsideJSONString(lower, idx) {
-			from = idx + 1
-			continue
-		}
-		if !hasJSONObjectContextPrefix(lower[:idx]) {
-			from = idx + 1
-			continue
-		}
-		if !hasJSONKeyBoundary(lower, idx, len(key)) {
-			from = idx + 1
-			continue
-		}
-		j := idx + len(key)
-		for j < len(lower) && (lower[j] == ' ' || lower[j] == '\t' || lower[j] == '\r' || lower[j] == '\n') {
-			j++
-		}
-		if j < len(lower) && lower[j] == ':' {
-			k := j + 1
-			for k < len(lower) && (lower[k] == ' ' || lower[k] == '\t' || lower[k] == '\r' || lower[k] == '\n') {
-				k++
-			}
-			if k < len(lower) && lower[k] != '{' {
-				from = idx + 1
-				continue
-			}
-			return idx
-		}
-		from = idx + 1
-	}
-	return -1
-}
-
-func isInsideJSONString(s string, idx int) bool {
-	inString := false
-	escaped := false
-	for i := 0; i < idx; i++ {
-		c := s[i]
-		if escaped {
-			escaped = false
-			continue
-		}
-		if c == '\\' && inString {
-			escaped = true
-			continue
-		}
-		if c == '"' {
-			inString = !inString
-		}
-	}
-	return inString
-}
-
-func hasJSONObjectContextPrefix(prefix string) bool {
-	return strings.LastIndex(prefix, "{") >= 0
-}
-
-func hasJSONKeyBoundary(s string, idx, keyLen int) bool {
-	if idx > 0 {
-		prev := s[idx-1]
-		if isLowerAlphaNumeric(prev) {
-			return false
-		}
-	}
-	if end := idx + keyLen; end < len(s) {
-		next := s[end]
-		if isLowerAlphaNumeric(next) {
-			return false
-		}
-	}
-	return true
-}
-
-func isLowerAlphaNumeric(b byte) bool {
-	return (b >= 'a' && b <= 'z') || (b >= '0' && b <= '9') || b == '_'
-}

internal/adapter/openai/tool_sieve_functioncall_test.go

@@ -1,23 +0,0 @@
-package openai
-
-import "testing"
-
-func TestFindQuotedFunctionCallKeyStart_PrefersEarlierBareKey(t *testing.T) {
-	input := `{functionCall:{"name":"a","arguments":"{}"},"message":"literal text: \"functionCall\": not a key"}`
-
-	got := findQuotedFunctionCallKeyStart(input)
-	want := 1
-	if got != want {
-		t.Fatalf("findQuotedFunctionCallKeyStart() = %d, want %d", got, want)
-	}
-}
-
-func TestFindQuotedFunctionCallKeyStart_PrefersEarlierQuotedKey(t *testing.T) {
-	input := `{"functionCall":{"name":"a","arguments":"{}"},"note":"functionCall appears in prose"}`
-
-	got := findQuotedFunctionCallKeyStart(input)
-	want := 1
-	if got != want {
-		t.Fatalf("findQuotedFunctionCallKeyStart() = %d, want %d", got, want)
-	}
-}

internal/adapter/openai/tool_sieve_jsonscan.go

@@ -2,48 +2,6 @@ package openai
 
 import "strings"
 
-func extractJSONObjectFrom(text string, start int) (string, int, bool) {
-	if start < 0 || start >= len(text) || text[start] != '{' {
-		return "", 0, false
-	}
-	depth := 0
-	quote := byte(0)
-	escaped := false
-	for i := start; i < len(text); i++ {
-		ch := text[i]
-		if quote != 0 {
-			if escaped {
-				escaped = false
-				continue
-			}
-			if ch == '\\' {
-				escaped = true
-				continue
-			}
-			if ch == quote {
-				quote = 0
-			}
-			continue
-		}
-		if ch == '"' || ch == '\'' {
-			quote = ch
-			continue
-		}
-		if ch == '{' {
-			depth++
-			continue
-		}
-		if ch == '}' {
-			depth--
-			if depth == 0 {
-				end := i + 1
-				return text[start:end], end, true
-			}
-		}
-	}
-	return "", 0, false
-}
-
 func trimWrappingJSONFence(prefix, suffix string) (string, string) {
 	trimmedPrefix := strings.TrimRight(prefix, " \t\r\n")
 	fenceIdx := strings.LastIndex(trimmedPrefix, "```")
@@ -67,18 +25,3 @@ func trimWrappingJSONFence(prefix, suffix string) (string, string) {
 	consumedLeading := len(suffix) - len(trimmedSuffix)
 	return trimmedPrefix[:fenceIdx], suffix[consumedLeading+3:]
 }
-
-func openFenceStartBefore(s string, pos int) (int, bool) {
-	if pos <= 0 || pos > len(s) {
-		return -1, false
-	}
-	segment := s[:pos]
-	lastFence := strings.LastIndex(segment, "```")
-	if lastFence < 0 {
-		return -1, false
-	}
-	if strings.Count(segment, "```")%2 == 1 {
-		return lastFence, true
-	}
-	return -1, false
-}

internal/adapter/openai/tool_sieve_state.go

@@ -1,30 +1,32 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"strings"
-
-	"ds2api/internal/util"
 )
 
 type toolStreamSieveState struct {
-	pending          strings.Builder
-	capture          strings.Builder
-	capturing        bool
-	recentTextTail   string
-	pendingToolRaw   string
-	pendingToolCalls []util.ParsedToolCall
-	disableDeltas    bool
-	toolNameSent     bool
-	toolName         string
-	toolArgsStart    int
-	toolArgsSent     int
-	toolArgsString   bool
-	toolArgsDone     bool
+	pending               strings.Builder
+	capture               strings.Builder
+	capturing             bool
+	codeFenceStack        []int
+	codeFencePendingTicks int
+	codeFenceLineStart    bool
+	recentTextTail        string
+	pendingToolRaw        string
+	pendingToolCalls      []toolcall.ParsedToolCall
+	disableDeltas         bool
+	toolNameSent          bool
+	toolName              string
+	toolArgsStart         int
+	toolArgsSent          int
+	toolArgsString        bool
+	toolArgsDone          bool
 }
 
 type toolStreamEvent struct {
 	Content        string
-	ToolCalls      []util.ParsedToolCall
+	ToolCalls      []toolcall.ParsedToolCall
 	ToolCallDeltas []toolCallDelta
 }
 
@@ -48,9 +50,10 @@ func (s *toolStreamSieveState) resetIncrementalToolState() {
 }
 
 func (s *toolStreamSieveState) noteText(content string) {
-	if strings.TrimSpace(content) == "" {
+	if !hasMeaningfulText(content) {
 		return
 	}
+	updateCodeFenceState(s, content)
 	s.recentTextTail = appendTail(s.recentTextTail, content, toolSieveContextTailLimit)
 }
 
@@ -64,3 +67,107 @@ func appendTail(prev, next string, max int) string {
 	}
 	return combined[len(combined)-max:]
 }
+
+func hasMeaningfulText(text string) bool {
+	return strings.TrimSpace(text) != ""
+}
+
+func insideCodeFenceWithState(state *toolStreamSieveState, text string) bool {
+	if state == nil {
+		return insideCodeFence(text)
+	}
+	simulated := simulateCodeFenceState(
+		state.codeFenceStack,
+		state.codeFencePendingTicks,
+		state.codeFenceLineStart,
+		text,
+	)
+	return len(simulated.stack) > 0
+}
+
+func insideCodeFence(text string) bool {
+	if text == "" {
+		return false
+	}
+	return len(simulateCodeFenceState(nil, 0, true, text).stack) > 0
+}
+
+func updateCodeFenceState(state *toolStreamSieveState, text string) {
+	if state == nil || !hasMeaningfulText(text) {
+		return
+	}
+	next := simulateCodeFenceState(
+		state.codeFenceStack,
+		state.codeFencePendingTicks,
+		state.codeFenceLineStart,
+		text,
+	)
+	state.codeFenceStack = next.stack
+	state.codeFencePendingTicks = next.pendingTicks
+	state.codeFenceLineStart = next.lineStart
+}
+
+type codeFenceSimulation struct {
+	stack        []int
+	pendingTicks int
+	lineStart    bool
+}
+
+func simulateCodeFenceState(stack []int, pendingTicks int, lineStart bool, text string) codeFenceSimulation {
+	chunk := text
+	nextStack := append([]int(nil), stack...)
+	ticks := pendingTicks
+	atLineStart := lineStart
+
+	flushTicks := func() {
+		if ticks > 0 {
+			if atLineStart && ticks >= 3 {
+				applyFenceMarker(&nextStack, ticks)
+			}
+			atLineStart = false
+			ticks = 0
+		}
+	}
+
+	for i := 0; i < len(chunk); i++ {
+		ch := chunk[i]
+		if ch == '`' {
+			ticks++
+			continue
+		}
+		flushTicks()
+		switch ch {
+		case '\n', '\r':
+			atLineStart = true
+		case ' ', '\t':
+			if atLineStart {
+				continue
+			}
+			atLineStart = false
+		default:
+			atLineStart = false
+		}
+	}
+
+	return codeFenceSimulation{
+		stack:        nextStack,
+		pendingTicks: ticks,
+		lineStart:    atLineStart,
+	}
+}
+
+func applyFenceMarker(stack *[]int, ticks int) {
+	if stack == nil || ticks <= 0 {
+		return
+	}
+	if len(*stack) == 0 {
+		*stack = append(*stack, ticks)
+		return
+	}
+	top := (*stack)[len(*stack)-1]
+	if ticks >= top {
+		*stack = (*stack)[:len(*stack)-1]
+		return
+	}
+	*stack = append(*stack, ticks)
+}

internal/adapter/openai/tool_sieve_xml.go

@@ -1,14 +1,14 @@
 package openai
 
 import (
+	"ds2api/internal/toolcall"
 	"regexp"
 	"strings"
-
-	"ds2api/internal/util"
 )
 
 // --- XML tool call support for the streaming sieve ---
 
+//nolint:unused // kept as explicit tag inventory for future XML sieve refinements.
 var xmlToolCallClosingTags = []string{"</tool_calls>", "</tool_call>", "</invoke>", "</function_call>", "</function_calls>", "</tool_use>",
 	// Agent-style XML tags (Roo Code, Cline, etc.)
 	"</attempt_completion>", "</ask_followup_question>", "</new_task>", "</result>"}
@@ -26,14 +26,16 @@ var xmlToolCallTagPairs = []struct{ open, close string }{
 	{"<invoke", "</invoke>"},
 	{"<tool_use", "</tool_use>"},
 	// Agent-style: these are XML "tool call" patterns from coding agents.
-	// They get captured → parsed. If parsing fails, the block is consumed
-	// (swallowed) to prevent raw XML from leaking to the client.
+	// They get captured → parsed. If parsing fails, the raw XML is preserved
+	// so the caller can still see the original text.
 	{"<attempt_completion", "</attempt_completion>"},
 	{"<ask_followup_question", "</ask_followup_question>"},
 	{"<new_task", "</new_task>"},
 }
 
 // xmlToolCallBlockPattern matches a complete XML tool call block (wrapper or standalone).
+//
+//nolint:unused // reserved for future fast-path XML block detection.
 var xmlToolCallBlockPattern = regexp.MustCompile(`(?is)(<tool_calls>\s*(?:.*?)\s*</tool_calls>|<tool_call>\s*(?:.*?)\s*</tool_call>|<invoke\b[^>]*>(?:.*?)</invoke>|<function_calls?\b[^>]*>(?:.*?)</function_calls?>|<tool_use>(?:.*?)</tool_use>|<attempt_completion>(?:.*?)</attempt_completion>|<ask_followup_question>(?:.*?)</ask_followup_question>|<new_task>(?:.*?)</new_task>)`)
 
 // xmlToolTagsToDetect is the set of XML tag prefixes used by findToolSegmentStart.
@@ -43,7 +45,7 @@ var xmlToolTagsToDetect = []string{"<tool_calls>", "<tool_calls\n", "<tool_call>
 	"<attempt_completion>", "<ask_followup_question>", "<new_task>"}
 
 // consumeXMLToolCapture tries to extract complete XML tool call blocks from captured text.
-func consumeXMLToolCapture(captured string, toolNames []string) (prefix string, calls []util.ParsedToolCall, suffix string, ready bool) {
+func consumeXMLToolCapture(captured string, toolNames []string) (prefix string, calls []toolcall.ParsedToolCall, suffix string, ready bool) {
 	lower := strings.ToLower(captured)
 	// Find the FIRST matching open/close pair, preferring wrapper tags.
 	// Tag pairs are ordered longest-first (e.g. <tool_calls before <tool_call)
@@ -66,13 +68,13 @@ func consumeXMLToolCapture(captured string, toolNames []string) (prefix string,
 		xmlBlock := captured[openIdx:closeEnd]
 		prefixPart := captured[:openIdx]
 		suffixPart := captured[closeEnd:]
-		parsed := util.ParseToolCalls(xmlBlock, toolNames)
+		parsed := toolcall.ParseToolCalls(xmlBlock, toolNames)
 		if len(parsed) > 0 {
 			prefixPart, suffixPart = trimWrappingJSONFence(prefixPart, suffixPart)
 			return prefixPart, parsed, suffixPart, true
 		}
-		// Looks like XML tool syntax but failed to parse — consume it to avoid leak.
-		return prefixPart, nil, suffixPart, true
+		// If this block failed to become a tool call, pass it through as text.
+		return prefixPart + xmlBlock, nil, suffixPart, true
 	}
 	return "", nil, "", false
 }
@@ -116,32 +118,3 @@ func findPartialXMLToolTagStart(s string) int {
 	}
 	return -1
 }
-
-// looksLikeXMLToolTagFragment returns true if s looks like a fragment from a
-// split XML tool call tag — for example "tool_calls>" or "/tool_call>\n".
-// These fragments arise when '<' was consumed separately and the tail remains.
-func looksLikeXMLToolTagFragment(s string) bool {
-	trimmed := strings.TrimSpace(s)
-	if trimmed == "" {
-		return false
-	}
-	lower := strings.ToLower(trimmed)
-	// Check for closing tag tails like "tool_calls>" or "/tool_calls>"
-	fragments := []string{
-		"tool_calls>", "tool_call>", "/tool_calls>", "/tool_call>",
-		"function_calls>", "function_call>", "/function_calls>", "/function_call>",
-		"invoke>", "/invoke>", "tool_use>", "/tool_use>",
-		"tool_name>", "/tool_name>", "parameters>", "/parameters>",
-		// Agent-style tag fragments
-		"attempt_completion>", "/attempt_completion>",
-		"ask_followup_question>", "/ask_followup_question>",
-		"new_task>", "/new_task>",
-		"result>", "/result>",
-	}
-	for _, f := range fragments {
-		if strings.Contains(lower, f) {
-			return true
-		}
-	}
-	return false
-}

internal/adapter/openai/tool_sieve_xml_test.go

@@ -78,6 +78,148 @@ func TestProcessToolSieveXMLWithLeadingText(t *testing.T) {
 	}
 }
 
+func TestProcessToolSievePassesThroughNonToolXMLBlock(t *testing.T) {
+	var state toolStreamSieveState
+	chunk := `<tool_call><title>示例 XML</title><body>plain text xml payload</body></tool_call>`
+	events := processToolSieveChunk(&state, chunk, []string{"read_file"})
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		toolCalls += len(evt.ToolCalls)
+	}
+	if toolCalls != 0 {
+		t.Fatalf("expected no tool calls for plain XML payload, got %d events=%#v", toolCalls, events)
+	}
+	if textContent.String() != chunk {
+		t.Fatalf("expected XML payload to pass through unchanged, got %q", textContent.String())
+	}
+}
+
+func TestProcessToolSieveNonToolXMLKeepsSuffixForToolParsing(t *testing.T) {
+	var state toolStreamSieveState
+	chunk := `<tool_call><title>plain xml</title></tool_call><invoke name="read_file"><parameters>{"path":"README.MD"}</parameters></invoke>`
+	events := processToolSieveChunk(&state, chunk, []string{"read_file"})
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		toolCalls += len(evt.ToolCalls)
+	}
+	if !strings.Contains(textContent.String(), `<tool_call><title>plain xml</title></tool_call>`) {
+		t.Fatalf("expected leading non-tool XML to be preserved, got %q", textContent.String())
+	}
+	if strings.Contains(textContent.String(), `<invoke name="read_file">`) {
+		t.Fatalf("expected invoke tool XML to be intercepted, got %q", textContent.String())
+	}
+	if toolCalls != 1 {
+		t.Fatalf("expected exactly one parsed tool call from suffix, got %d events=%#v", toolCalls, events)
+	}
+}
+
+func TestProcessToolSievePassesThroughMalformedExecutableXMLBlock(t *testing.T) {
+	var state toolStreamSieveState
+	chunk := `<tool_call><parameters>{"path":"README.md"}</parameters></tool_call>`
+	events := processToolSieveChunk(&state, chunk, []string{"read_file"})
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if toolCalls != 0 {
+		t.Fatalf("expected malformed executable-looking XML to stay text, got %d events=%#v", toolCalls, events)
+	}
+	if textContent.String() != chunk {
+		t.Fatalf("expected malformed executable-looking XML to pass through unchanged, got %q", textContent.String())
+	}
+}
+
+func TestProcessToolSievePassesThroughFencedXMLToolCallExamples(t *testing.T) {
+	var state toolStreamSieveState
+	input := strings.Join([]string{
+		"Before first example.\n```",
+		"xml\n<tool_call><tool_name>read_file</tool_name><parameters>{\"path\":\"README.md\"}</parameters></tool_call>\n```\n",
+		"Between examples.\n```xml\n",
+		"<tool_call><tool_name>search</tool_name><parameters>{\"q\":\"golang\"}</parameters></tool_call>\n",
+		"```\nAfter examples.",
+	}, "")
+
+	chunks := []string{
+		"Before first example.\n```",
+		"xml\n<tool_call><tool_name>read_file</tool_name><parameters>{\"path\":\"README.md\"}</parameters></tool_call>\n```\n",
+		"Between examples.\n```xml\n",
+		"<tool_call><tool_name>search</tool_name><parameters>{\"q\":\"golang\"}</parameters></tool_call>\n",
+		"```\nAfter examples.",
+	}
+
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file", "search"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"read_file", "search"})...)
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent.WriteString(evt.Content)
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if toolCalls != 0 {
+		t.Fatalf("expected fenced XML examples to stay text, got %d tool calls events=%#v", toolCalls, events)
+	}
+	if textContent.String() != input {
+		t.Fatalf("expected fenced XML examples to pass through unchanged, got %q", textContent.String())
+	}
+}
+
+func TestProcessToolSieveKeepsPartialXMLTagInsideFencedExample(t *testing.T) {
+	var state toolStreamSieveState
+	input := strings.Join([]string{
+		"Example:\n```xml\n<tool_ca",
+		"ll><tool_name>read_file</tool_name><parameters>{\"path\":\"README.md\"}</parameters></tool_call>\n```\n",
+		"Done.",
+	}, "")
+
+	chunks := []string{
+		"Example:\n```xml\n<tool_ca",
+		"ll><tool_name>read_file</tool_name><parameters>{\"path\":\"README.md\"}</parameters></tool_call>\n```\n",
+		"Done.",
+	}
+
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent strings.Builder
+	toolCalls := 0
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent.WriteString(evt.Content)
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if toolCalls != 0 {
+		t.Fatalf("expected partial fenced XML to stay text, got %d tool calls events=%#v", toolCalls, events)
+	}
+	if textContent.String() != input {
+		t.Fatalf("expected partial fenced XML to pass through unchanged, got %q", textContent.String())
+	}
+}
+
 func TestProcessToolSievePartialXMLTagHeldBack(t *testing.T) {
 	var state toolStreamSieveState
 	// Chunk ends with a partial XML tool tag.
@@ -104,15 +246,16 @@ func TestFindToolSegmentStartDetectsXMLToolCalls(t *testing.T) {
 		want  int
 	}{
 		{"tool_calls_tag", "some text <tool_calls>\n", 10},
-		{"gemini_function_call_json", `some text {"functionCall":{"name":"search","args":{"q":"latest"}}}`, 10},
 		{"tool_call_tag", "prefix <tool_call>\n", 7},
 		{"invoke_tag", "text <invoke name=\"foo\">body</invoke>", 5},
+		{"xml_inside_code_fence", "```xml\n<tool_call><tool_name>read_file</tool_name></tool_call>\n```", -1},
 		{"function_call_tag", "<function_call name=\"foo\">body</function_call>", 0},
 		{"no_xml", "just plain text", -1},
+		{"gemini_json_no_detect", `some text {"functionCall":{"name":"search"}}`, -1},
 	}
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
-			got := findToolSegmentStart(tc.input)
+			got := findToolSegmentStart(nil, tc.input)
 			if got != tc.want {
 				t.Fatalf("findToolSegmentStart(%q) = %d, want %d", tc.input, got, tc.want)
 			}
@@ -120,81 +263,6 @@ func TestFindToolSegmentStartDetectsXMLToolCalls(t *testing.T) {
 	}
 }
 
-func TestFindToolSegmentStartIgnoresFunctionCallProse(t *testing.T) {
-	input := "Please explain the functionCall API field and how clients should parse it."
-	if got := findToolSegmentStart(input); got != -1 {
-		t.Fatalf("expected no tool segment start for prose, got %d", got)
-	}
-}
-
-func TestFindToolSegmentStartDetectsQuotedFunctionCallKey(t *testing.T) {
-	input := `prefix {"functionCall": {"name":"search_web","args":{"query":"x"}}}`
-	want := strings.Index(input, "{")
-	if got := findToolSegmentStart(input); got != want {
-		t.Fatalf("expected JSON object start %d, got %d", want, got)
-	}
-}
-
-func TestFindToolSegmentStartDetectsLooseFunctionCallKey(t *testing.T) {
-	input := `prefix {functionCall: {"name":"search_web","args":{"query":"x"}}}`
-	want := strings.Index(input, "{")
-	if got := findToolSegmentStart(input); got != want {
-		t.Fatalf("expected JSON object start %d, got %d", want, got)
-	}
-}
-
-func TestFindToolSegmentStartPrefersQuotedFunctionCallOverEarlierBareProse(t *testing.T) {
-	input := `prefix {note} functionCall: docs hint {"functionCall":{"name":"search_web","args":{"query":"x"}}}`
-	want := strings.Index(input, `{"functionCall"`)
-	if got := findToolSegmentStart(input); got != want {
-		t.Fatalf("expected quoted functionCall JSON start %d, got %d", want, got)
-	}
-}
-
-func TestFindToolSegmentStartIgnoresLooseFunctionCallProse(t *testing.T) {
-	input := "Please explain why functionCall: is used in documentation examples."
-	if got := findToolSegmentStart(input); got != -1 {
-		t.Fatalf("expected no tool segment start for prose, got %d", got)
-	}
-}
-
-func TestProcessToolSieveDoesNotBufferFunctionCallProse(t *testing.T) {
-	var state toolStreamSieveState
-	chunk := "Please explain the functionCall API field and keep streaming this sentence."
-	events := processToolSieveChunk(&state, chunk, []string{"search_web"})
-	var text string
-	for _, evt := range events {
-		text += evt.Content
-		if len(evt.ToolCalls) > 0 {
-			t.Fatalf("expected no tool calls for prose, got %#v", evt.ToolCalls)
-		}
-	}
-	if text != chunk {
-		t.Fatalf("expected prose to pass through immediately, got %q", text)
-	}
-}
-
-func TestProcessToolSieveDetectsGeminiFunctionCallPayload(t *testing.T) {
-	var state toolStreamSieveState
-	events := processToolSieveChunk(&state, `{"functionCall":{"name":"search_web","args":{"query":"latest"}}}`, []string{"search_web"})
-	events = append(events, flushToolSieve(&state, []string{"search_web"})...)
-
-	var textContent string
-	var toolCalls int
-	for _, evt := range events {
-		if evt.Content != "" {
-			textContent += evt.Content
-		}
-		toolCalls += len(evt.ToolCalls)
-	}
-	if toolCalls != 1 {
-		t.Fatalf("expected one tool call from functionCall payload, got events=%#v", events)
-	}
-	if strings.Contains(strings.ToLower(textContent), "functioncall") {
-		t.Fatalf("functionCall json leaked into text content: %q", textContent)
-	}
-}
-
 func TestFindPartialXMLToolTagStart(t *testing.T) {
 	cases := []struct {
 		name  string
@@ -301,8 +369,8 @@ func TestProcessToolSieveTokenByTokenXMLNoLeak(t *testing.T) {
 	}
 }
 
-// Test that flushToolSieve on incomplete XML does NOT leak the raw XML content.
-func TestFlushToolSieveIncompleteXMLDoesNotLeak(t *testing.T) {
+// Test that flushToolSieve on incomplete XML falls back to raw text.
+func TestFlushToolSieveIncompleteXMLFallsBackToText(t *testing.T) {
 	var state toolStreamSieveState
 	// XML block starts but stream ends before completion.
 	chunks := []string{
@@ -324,8 +392,8 @@ func TestFlushToolSieveIncompleteXMLDoesNotLeak(t *testing.T) {
 		}
 	}
 
-	if strings.Contains(textContent, "<tool_call") {
-		t.Fatalf("incomplete XML leaked on flush: %q", textContent)
+	if textContent != strings.Join(chunks, "") {
+		t.Fatalf("expected incomplete XML to fall back to raw text, got %q", textContent)
 	}
 }
 
@@ -362,10 +430,10 @@ func TestOpeningXMLTagNotLeakedAsContent(t *testing.T) {
 	}
 }
 
-func TestProcessToolSieveInterceptsAttemptCompletionLeak(t *testing.T) {
+func TestProcessToolSieveFallsBackToRawAttemptCompletion(t *testing.T) {
 	var state toolStreamSieveState
-	// Simulate an agent outputting attempt_completion XML tag 
-	// which shouldn't leak to text output, even if it fails to parse as a valid tool.
+	// Simulate an agent outputting attempt_completion XML tag.
+	// If it does not parse as a tool call, it should fall back to raw text.
 	chunks := []string{
 		"Done with task.\n",
 		"<attempt_completion>\n",
@@ -389,7 +457,7 @@ func TestProcessToolSieveInterceptsAttemptCompletionLeak(t *testing.T) {
 		t.Fatalf("expected leading text to be emitted, got %q", textContent)
 	}
 
-	if strings.Contains(textContent, "<attempt_completion>") || strings.Contains(textContent, "result>") {
-		t.Fatalf("agent XML tag content leaked to text: %q", textContent)
+	if textContent != strings.Join(chunks, "") {
+		t.Fatalf("expected agent XML to fall back to raw text, got %q", textContent)
 	}
 }

internal/adapter/openai/upstream_empty.go

@@ -0,0 +1,15 @@
+package openai
+
+import "net/http"
+
+func writeUpstreamEmptyOutputError(w http.ResponseWriter, text string, contentFilter bool) bool {
+	if text != "" {
+		return false
+	}
+	if contentFilter {
+		writeOpenAIErrorWithCode(w, http.StatusBadRequest, "Upstream content filtered the response and returned no output.", "content_filter")
+		return true
+	}
+	writeOpenAIErrorWithCode(w, http.StatusTooManyRequests, "Upstream model returned empty output.", "upstream_empty_output")
+	return true
+}

internal/adapter/openai/vercel_stream.go

@@ -52,6 +52,10 @@ func (h *Handler) handleVercelStreamPrepare(w http.ResponseWriter, r *http.Reque
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	if !util.ToBool(req["stream"]) {
 		writeOpenAIError(w, http.StatusBadRequest, "stream must be true")
 		return
@@ -99,10 +103,13 @@ func (h *Handler) handleVercelStreamPrepare(w http.ResponseWriter, r *http.Reque
 		"final_prompt":     stdReq.FinalPrompt,
 		"thinking_enabled": stdReq.Thinking,
 		"search_enabled":   stdReq.Search,
-		"tool_names":       stdReq.ToolNames,
-		"deepseek_token":   a.DeepSeekToken,
-		"pow_header":       powHeader,
-		"payload":          payload,
+		"compat": map[string]any{
+			"strip_reference_markers": h.compatStripReferenceMarkers(),
+		},
+		"tool_names":     stdReq.ToolNames,
+		"deepseek_token": a.DeepSeekToken,
+		"pow_header":     powHeader,
+		"payload":        payload,
 	})
 }
 

internal/admin/deps.go

@@ -32,6 +32,8 @@ type ConfigStore interface {
 	RuntimeAccountMaxQueue(defaultSize int) int
 	RuntimeGlobalMaxInflight(defaultSize int) int
 	RuntimeTokenRefreshIntervalHours() int
+	AutoDeleteMode() string
+	CompatStripReferenceMarkers() bool
 	AutoDeleteSessions() bool
 }
 
@@ -41,6 +43,10 @@ type PoolController interface {
 	ApplyRuntimeLimits(maxInflightPerAccount, maxQueueSize, globalMaxInflight int)
 }
 
+type OpenAIChatCaller interface {
+	ChatCompletions(w http.ResponseWriter, r *http.Request)
+}
+
 type DeepSeekCaller interface {
 	Login(ctx context.Context, acc config.Account) (string, error)
 	CreateSession(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error)

internal/admin/handler.go

@@ -5,9 +5,10 @@ import (
 )
 
 type Handler struct {
-	Store ConfigStore
-	Pool  PoolController
-	DS    DeepSeekCaller
+	Store  ConfigStore
+	Pool   PoolController
+	DS     DeepSeekCaller
+	OpenAI OpenAIChatCaller
 }
 
 func RegisterRoutes(r chi.Router, h *Handler) {
@@ -25,15 +26,24 @@ func RegisterRoutes(r chi.Router, h *Handler) {
 		pr.Get("/config/export", h.configExport)
 		pr.Post("/keys", h.addKey)
 		pr.Delete("/keys/{key}", h.deleteKey)
+		pr.Get("/proxies", h.listProxies)
+		pr.Post("/proxies", h.addProxy)
+		pr.Put("/proxies/{proxyID}", h.updateProxy)
+		pr.Delete("/proxies/{proxyID}", h.deleteProxy)
+		pr.Post("/proxies/test", h.testProxy)
 		pr.Get("/accounts", h.listAccounts)
 		pr.Post("/accounts", h.addAccount)
 		pr.Delete("/accounts/{identifier}", h.deleteAccount)
+		pr.Put("/accounts/{identifier}/proxy", h.updateAccountProxy)
 		pr.Get("/queue/status", h.queueStatus)
 		pr.Post("/accounts/test", h.testSingleAccount)
 		pr.Post("/accounts/test-all", h.testAllAccounts)
 		pr.Post("/accounts/sessions/delete-all", h.deleteAllSessions)
 		pr.Post("/import", h.batchImport)
 		pr.Post("/test", h.testAPI)
+		pr.Post("/dev/raw-samples/capture", h.captureRawSample)
+		pr.Get("/dev/raw-samples/query", h.queryRawSampleCaptures)
+		pr.Post("/dev/raw-samples/save", h.saveRawSampleFromCaptures)
 		pr.Post("/vercel/sync", h.syncVercel)
 		pr.Get("/vercel/status", h.vercelStatus)
 		pr.Post("/vercel/status", h.vercelStatus)