Merge pull request #470 from CJackHwang/dev

feat: DSML/CDATA parsing robustness, tool-call resilience, and completion retry improvements

AGENTS.md

@@ -22,6 +22,13 @@ These rules apply to all agent-made changes in this repository.
 - 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.
 
+## Protocol Adapter Boundary
+
+- Do not let OpenAI Chat, OpenAI Responses, Claude, Gemini, or other interface protocol formatting own shared business behavior.
+- Normalize protocol-specific request shapes into the project standard request/turn model first, run shared business logic in one place, then render back to the target protocol at the boundary.
+- Business logic that must stay globally consistent includes empty-output retry, thinking/reasoning handling, tool-call detection and policy, usage accounting, current-input-file injection, history persistence, file/reference handling, and completion payload assembly.
+- If a behavior must differ by protocol, keep the difference as an explicit adapter/rendering concern and document why it cannot live in the shared normalized path.
+
 ## Documentation Sync
 
 - When business logic or user-visible behavior changes, update the corresponding documentation in the same change.

API.en.md

@@ -18,6 +18,7 @@ Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Depl
 - [OpenAI-Compatible API](#openai-compatible-api)
 - [Claude-Compatible API](#claude-compatible-api)
 - [Gemini-Compatible API](#gemini-compatible-api)
+- [Ollama API](#ollama-api)
 - [Admin API](#admin-api)
 - [Error Payloads](#error-payloads)
 - [cURL Examples](#curl-examples)
@@ -31,7 +32,7 @@ Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Depl
 | Base URL | `http://localhost:5001` or your deployment domain |
 | Default Content-Type | `application/json` |
 | Health probes | `GET /healthz`, `GET /readyz` |
-| CORS | Enabled (uniformly covers `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, and `/admin/*`; echoes the browser `Origin` when present, otherwise `*`; default allow-list includes `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`, and also accepts third-party preflight-requested headers such as `x-stainless-*`; `/v1/chat/completions` on Vercel Node Runtime matches the same behavior; internal-only `X-Ds2-Internal-Token` remains blocked) |
+| CORS | Enabled (uniformly covers `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, `/api/*`, and `/admin/*`; echoes the browser `Origin` when present, otherwise `*`; default allow-list includes `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`, and also accepts third-party preflight-requested headers such as `x-stainless-*`; `/v1/chat/completions` on Vercel Node Runtime matches the same behavior; internal-only `X-Ds2-Internal-Token` remains blocked) |
 
 - All JSON request bodies must be valid UTF-8; malformed byte sequences are rejected on ingress with `400 invalid json`.
 
@@ -39,8 +40,10 @@ Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Depl
 
 - 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: models should output the DSML shell `<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`; DS2API also accepts legacy canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`. DSML is normalized back to XML at the parser entry, so internal parsing remains XML-based, with stream-time anti-leak filtering.
+- Tool-calling semantics are aligned between Go and Node runtime: models should output the fullwidth-separator DSML shell `<｜DSML｜tool_calls>` → `<｜DSML｜invoke name="...">` → `<｜DSML｜parameter name="...">`; DS2API also accepts the halfwidth DSML wrapper `<|DSML|tool_calls>`, DSML wrapper aliases such as `<dsml|tool_calls>`, `<|tool_calls>`, `<｜tool_calls>`, common DSML separator drift such as `<|DSML tool_calls>`, collapsed DSML local names such as `<DSMLtool_calls>`, control-separator drift such as `<DSML␂tool_calls>` / raw STX `\x02`, CJK angle bracket, fullwidth-bang / ideographic-comma separator drift, PascalCase local-name drift, and trailing attribute separator drift such as `<DSM｜parameter name="command"｜>...〈/DSM｜parameter〉`, `<！DSML！invoke name=“Bash”>`, `<、DSML、tool_calls>`, `<DSmartToolCalls>`, or `<DSMLtool_calls※>`, arbitrary protocol prefixes such as `<proto💥tool_calls>`, and legacy canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`. The scanner normalizes fixed local names (`tool_calls` / `invoke` / `parameter`) with non-structural separators before or after them back to XML before parsing, and also tolerates CDATA opener drift such as `<！[CDATA[` / `<、[CDATA[`; only wrapped tool blocks or the narrow missing-opening-wrapper repair path enter the tool path, while bare `<invoke>` does not count as supported syntax. JSON literal parameter bodies are preserved as structured values, explicit empty or whitespace-only parameters are preserved as empty strings, malformed complete wrappers are released as plain text, and loose CDATA is narrowly repaired at final parse/flush when it can preserve a complete outer tool call.
 - `Admin API` separates static config from runtime policy: `/admin/config*` for configuration state, `/admin/settings*` for runtime behavior.
+- When upstream returns a thinking-only response with no visible text, the Go main path for both streaming and non-streaming completions retries once in the same DeepSeek session: it appends the prompt suffix `"Previous reply had no visible output. Please regenerate the visible final answer or tool call now."` and sets `parent_message_id`. If that same-account retry would still end as `429 upstream_empty_output`, managed-account mode switches to the next available account, creates a fresh session, and retries the original payload once before returning 429.
+- Citation/reference marker boundary: streaming output hides upstream `[citation:N]` / `[reference:N]` placeholders by default; non-stream output converts DeepSeek search reference markers into Markdown links.
 
 ---
 
@@ -83,7 +86,7 @@ Two header formats accepted:
 - Token is in `config.keys` → **Managed account mode**: DS2API auto-selects an account via rotation
 - Token is not in `config.keys` → **Direct token mode**: treated as a DeepSeek token directly
 
-**Optional header**: `X-Ds2-Target-Account: <email_or_mobile>` — Pin a specific managed account; if the target account does not exist or the managed-account queue is exhausted, the request returns `429`, and current responses do not include `Retry-After`. If the account exists but login/refresh fails, the request returns the underlying `401` or upstream error.
+**Optional header**: `X-Ds2-Target-Account: <email_or_mobile>` — Pin a specific managed account; if the target account does not exist or the managed-account queue is exhausted, the request returns `429`, and current responses do not include `Retry-After`. If the account exists but login/refresh fails, the request returns the underlying `401` or upstream error. Without a pinned target, managed-account completion requests try one alternate-account fresh retry before returning an empty-output 429; pinned-target requests and requests with no other available account do not switch.
 Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=` as the caller credential source.
 
 ### Admin Endpoints (`/admin/*`)
@@ -123,6 +126,9 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 | POST | `/v1beta/models/{model}:streamGenerateContent` | Business | Gemini stream |
 | POST | `/v1/models/{model}:generateContent` | Business | Gemini non-stream compat path |
 | POST | `/v1/models/{model}:streamGenerateContent` | Business | Gemini stream compat path |
+| GET | `/api/version` | None | Ollama version endpoint |
+| GET | `/api/tags` | None | Ollama model list |
+| POST | `/api/show` | None | Ollama model capability query (returns `id` + `capabilities`) |
 | POST | `/admin/login` | None | Admin login |
 | GET | `/admin/verify` | JWT | Verify admin JWT |
 | GET | `/admin/vercel/config` | Admin | Read preconfigured Vercel creds |
@@ -222,16 +228,18 @@ For `chat` / `responses` / `embeddings`, DS2API follows a wide-input/strict-outp
 
 1. Match DeepSeek native model IDs first.
 2. Then match exact keys in `model_aliases`.
-3. If still unmatched, fall back by known family heuristics (`o*`, `gpt-*`, `claude-*`, etc.).
-4. If still unmatched, return `invalid_request_error`.
+3. If the request name ends with `-nothinking`, resolve the base alias and append the corresponding no-thinking variant.
+4. If still unmatched, return `invalid_request_error`. Unknown model families are not guessed heuristically; add explicit compatibility names through `model_aliases`.
 
 Built-in aliases come from `internal/config/models.go`; `config.model_aliases` can override or add mappings at runtime. Excerpt:
 
 - OpenAI / Codex: `gpt-4o`, `gpt-4.1`, `gpt-5`, `gpt-5.5`, `gpt-5-codex`, `gpt-5.3-codex`, `codex-mini-latest`
 - OpenAI reasoning: `o1`, `o3`, `o3-deep-research`, `o4-mini`
 - Claude: `claude-opus-4-6`, `claude-sonnet-4-6`, `claude-haiku-4-5`, `claude-3-5-sonnet-latest`
-- Gemini: `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-pro-vision`
-- Other compatibility families: `llama-*`, `qwen-*`, `mistral-*`, and `command-*` fall back through family heuristics
+- Gemini: `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-3.1-pro`, `gemini-3-pro`, `gemini-3-flash`, `gemini-3.1-flash-lite`, `gemini-pro-vision`
+- Other exact built-in aliases: `llama-3.1-70b-instruct`, `qwen-max`
+
+Aliases with a `-nothinking` suffix also map to the corresponding forced no-thinking DeepSeek model.
 
 Current vision support resolves only to `deepseek-v4-vision` and does not expose a separate `vision-search` variant.
 
@@ -239,6 +247,8 @@ Retired historical families such as `claude-1.*`, `claude-2.*`, `claude-instant-
 
 ### `POST /v1/chat/completions`
 
+> Path note: besides the canonical `/v1/chat/completions`, DS2API also accepts the root shortcut `/chat/completions`. On Vercel Runtime, `vercel.json` rewrites only the canonical `/v1/chat/completions` path to the Node streaming bridge; the root shortcut stays on the Go primary path. Use `/v1/chat/completions` on Vercel when real-time streaming is required.
+
 **Headers**:
 
 ```http
@@ -250,7 +260,7 @@ Content-Type: application/json
 
 | Field | Type | Required | Notes |
 | --- | --- | --- | --- |
-| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-5.5`, `gpt-5.4-mini`, `gpt-5.3-codex`, `o3`, `claude-opus-4-6`, `gemini-2.5-pro`, `gemini-2.5-flash`, etc.) |
+| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-5.5`, `gpt-5.4-mini`, `gpt-5.3-codex`, `o3`, `claude-opus-4-6`, `gemini-2.5-pro`, `gemini-3.1-pro`, `gemini-3-flash`, etc.); `-nothinking` suffixes force thinking / reasoning off |
 | `messages` | array | ✅ | OpenAI-style messages |
 | `stream` | boolean | ❌ | Default `false` |
 | `tools` | array | ❌ | Function calling schema |
@@ -345,7 +355,8 @@ When `tools` is present, DS2API performs anti-leak handling:
 
 Additional notes:
 
-- The parser treats DSML shell tool blocks (`<|DSML|tool_calls>` / `<|DSML|invoke name="...">` / `<|DSML|parameter name="...">`) and legacy canonical XML tool blocks (`<tool_calls>` / `<invoke name="...">` / `<parameter name="...">`) as executable tool calls. DSML is normalized back to XML at the parser entry; internal parsing remains XML-based. Legacy `<tools>`, `<tool_call>`, `<tool_name>`, `<param>`, `<function_call>`, `tool_use`, antml variants, and standalone JSON `tool_calls` payloads are treated as plain text.
+- The parser treats the recommended DSML shell tool blocks (`<｜DSML｜tool_calls>` / `<｜DSML｜invoke name="...">` / `<｜DSML｜parameter name="...">`), halfwidth DSML shell blocks (`<|DSML|tool_calls>` / `<|DSML|invoke name="...">` / `<|DSML|parameter name="...">`), DSML wrapper aliases (`<dsml|tool_calls>`, `<|tool_calls>`, `<｜tool_calls>`), common DSML separator drift (`<|DSML tool_calls>` / `<|DSML invoke>` / `<|DSML parameter>`), collapsed DSML local names (`<DSMLtool_calls>` / `<DSMLinvoke>` / `<DSMLparameter>`), control-separator drift (`<DSML␂tool_calls>` / raw STX `\x02`), CJK angle bracket, fullwidth-bang / ideographic-comma separator drift, PascalCase local-name drift, and trailing attribute separator drift (`<DSM｜parameter name="command"｜>...〈/DSM｜parameter〉` / `<！DSML！invoke name=“Bash”>` / `<、DSML、tool_calls>` / `<DSmartToolCalls>` / `<DSMLtool_calls※>`), arbitrary protocol prefixes (`<proto💥tool_calls>`), and legacy canonical XML tool blocks (`<tool_calls>` / `<invoke name="...">` / `<parameter name="...">`) as executable tool calls. These shells normalize non-structural separators back to XML first, while internal parsing remains XML-based; CDATA opener drift such as `<！[CDATA[` / `<、[CDATA[` is also normalized for parameter bodies. Legacy `<tools>`, `<tool_call>`, `<tool_name>`, `<param>`, `<function_call>`, `tool_use`, antml variants, and standalone JSON `tool_calls` payloads are treated as plain text; complete but malformed wrappers are also released as plain text.
+- The parser no longer drops tool calls solely because parameter values are empty; explicit empty strings or whitespace-only parameters become empty strings in structured `tool_calls`. Prompting still tells the model not to emit blank parameters, and missing/empty argument rejection belongs in the tool executor or client schema validation.
 - If the final visible response text is empty but the reasoning stream contains an executable tool call, Chat / Responses emits a standard OpenAI `tool_calls` / `function_call` output during finalization. If thinking/reasoning was not enabled by the client, that reasoning text is used only for detection and is not exposed as visible text or `reasoning_content`.
 - `tool_calls` shown inside fenced markdown code blocks (for example, ```json ... ```) are treated as examples, not executable calls.
 
@@ -617,6 +628,20 @@ Returns SSE (`text/event-stream`), each chunk as `data: <json>`:
 
 ---
 
+## Ollama API
+
+- `POST /api/show` request body: `{"model":"<model-id>"}`.
+- Response uses lowercase `id` (not `ID`) and includes `capabilities` for Ollama-style clients and strict schemas.
+
+Example response:
+
+```json
+{
+  "id": "deepseek-v4-flash",
+  "capabilities": ["tools", "thinking"]
+}
+```
+
 ## Admin API
 
 ### `POST /admin/login`
@@ -744,6 +769,7 @@ Reads runtime settings and status, including:
 - `responses` / `embeddings`
 - `auto_delete` (`mode`: `none` / `single` / `all`; legacy `sessions=true` is still treated as `all`)
 - `current_input_file` (`enabled` defaults to `true`, plus `min_chars`)
+- `thinking_injection` (`enabled` defaults to `true`, `prompt`, and `default_prompt`)
 - `model_aliases`
 - `env_backed`, `needs_vercel_sync`
 - `toolcall` policy is fixed to `feature_match + high` and is no longer returned or editable via settings
@@ -758,6 +784,7 @@ Hot-updates runtime settings. Supported fields:
 - `embeddings.provider`
 - `auto_delete.mode`
 - `current_input_file.enabled` / `current_input_file.min_chars`
+- `thinking_injection.enabled` / `thinking_injection.prompt`
 - `model_aliases`
 - `toolcall` policy is fixed and is no longer writable through settings
 
@@ -1238,7 +1265,7 @@ Clients should handle HTTP status code plus `error` / `detail` fields.
 | Code | Meaning |
 | --- | --- |
 | `401` | Authentication failed (invalid key/token, or expired admin JWT) |
-| `429` | Too many requests (exceeded inflight + queue capacity; current responses do not include `Retry-After`) |
+| `429` | Too many requests (exceeded inflight + queue capacity, or upstream thinking-only output with no visible answer; managed-account mode first tries one alternate-account fresh retry; current responses do not include `Retry-After`) |
 | `503` | Model unavailable or upstream error |
 
 ---

API.md

@@ -18,6 +18,7 @@
 - [OpenAI 兼容接口](#openai-兼容接口)
 - [Claude 兼容接口](#claude-兼容接口)
 - [Gemini 兼容接口](#gemini-兼容接口)
+- [Ollama 兼容接口](#ollama-兼容接口)
 - [Admin 接口](#admin-接口)
 - [错误响应格式](#错误响应格式)
 - [cURL 示例](#curl-示例)
@@ -31,7 +32,7 @@
 | Base URL | `http://localhost:5001` 或你的部署域名 |
 | 默认 Content-Type | `application/json` |
 | 健康检查 | `GET /healthz`、`GET /readyz` |
-| CORS | 已启用（统一覆盖 `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/admin/*`；浏览器有 `Origin` 时回显该 Origin，否则为 `*`；默认允许 `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`，并会放行预检里声明的第三方请求头，如 `x-stainless-*`；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同行为；内部专用头 `X-Ds2-Internal-Token` 仍被拦截） |
+| CORS | 已启用（统一覆盖 `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/api/*`、`/admin/*`；浏览器有 `Origin` 时回显该 Origin，否则为 `*`；默认允许 `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`，并会放行预检里声明的第三方请求头，如 `x-stainless-*`；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同行为；内部专用头 `X-Ds2-Internal-Token` 仍被拦截） |
 
 - 所有 JSON 请求体都必须是合法 UTF-8；非法字节序列会在入站阶段被拒绝为 `400 invalid json`。
 
@@ -39,9 +40,9 @@
 
 - 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：推荐模型输出 DSML 外壳 `<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`；兼容层也接受 DSML wrapper 别名 `<dsml|tool_calls>`、`<|tool_calls>`、`<｜tool_calls>`、常见 DSML 分隔符漏写形态（如 `<|DSML tool_calls>`）、`DSML` 与工具标签名黏连的常见 typo（如 `<DSMLtool_calls>`），以及旧式 canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`。实现上采用窄容错结构扫描：只有 `tool_calls` wrapper 或可修复的缺失 opening wrapper 会进入工具路径，裸 `<invoke>` 不计为已支持语法；流式场景继续执行防泄漏筛分。若参数体本身是合法 JSON 字面量（如 `123`、`true`、`null`、数组或对象），会按结构化值输出，不再一律当作字符串；若 CDATA 偶发漏闭合，则会在最终 parse / flush 恢复阶段做窄修复，尽量保住已完整包裹的外层工具调用。
+- Tool Calling 的解析策略在 Go 与 Node Runtime 间保持一致：推荐模型输出全角分隔符 DSML 外壳 `<｜DSML｜tool_calls>` → `<｜DSML｜invoke name="...">` → `<｜DSML｜parameter name="...">`；兼容层也接受半角 DSML wrapper `<|DSML|tool_calls>`、DSML wrapper 别名 `<dsml|tool_calls>`、`<|tool_calls>`、`<｜tool_calls>`、常见 DSML 分隔符漏写形态（如 `<|DSML tool_calls>`）、`DSML` 与工具标签名黏连的常见 typo（如 `<DSMLtool_calls>`）、控制分隔符漂移（如 `<DSML␂tool_calls>` / 原始 STX `\x02`）、CJK 尖括号、全角感叹号、顿号、PascalCase 本地名、弯引号属性值与属性尾部分隔符漂移（如 `<DSM｜parameter name="command"｜>...〈/DSM｜parameter〉` / `<！DSML！invoke name=“Bash”>` / `<、DSML、tool_calls>` / `<DSmartToolCalls>` / `<DSMLtool_calls※>`）、任意协议前缀壳（如 `<proto💥tool_calls>`），以及旧式 canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`。实现上采用结构扫描：只要固定本地标签名是 `tool_calls` / `invoke` / `parameter`，标签名前或标签名后的非结构性分隔符会在解析入口归一化；CDATA 开头也会容错 `<！[CDATA[` / `<、[CDATA[` 这类分隔符漂移；只有 `tool_calls` wrapper 或可修复的缺失 opening wrapper 会进入工具路径，裸 `<invoke>` 不计为已支持语法；流式场景继续执行防泄漏筛分。若参数体本身是合法 JSON 字面量（如 `123`、`true`、`null`、数组或对象），会按结构化值输出，不再一律当作字符串；显式空字符串和纯空白参数会结构化保留为空字符串，是否拒绝缺参由工具执行侧决定；完整但 malformed 的 wrapper 会作为普通文本释放，不会吞掉或伪造成工具调用；若 CDATA 偶发漏闭合，则会在最终 parse / flush 恢复阶段做窄修复，尽量保住已完整包裹的外层工具调用。
 - `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。
-- 当上游返回 thinking-only 响应（模型输出了推理链但无可见文本）时，非流式补全会自动重试一次：以多轮对话 follow-up 方式追加 prompt 后缀 `"Previous reply had no visible output. Please regenerate the visible final answer or tool call now."` 并设置 `parent_message_id` 在同一 DeepSeek session 内让模型重新输出；重试最大 1 次。
+- 当上游返回 thinking-only 响应（模型输出了推理链但无可见文本）时，Go 主路径的流式与非流式补全都会先自动重试一次：以多轮对话 follow-up 方式追加 prompt 后缀 `"Previous reply had no visible output. Please regenerate the visible final answer or tool call now."` 并设置 `parent_message_id` 在同一 DeepSeek session 内让模型重新输出；同账号重试最大 1 次。若同账号重试后仍即将返回 `429 upstream_empty_output`，托管账号模式会在返回 429 前自动切换到下一个可用账号，新建 session，用原始 payload 再 fresh retry 一次。
 - 引用标记处理边界：流式输出默认隐藏 `[citation:N]` / `[reference:N]` 这类上游内部占位符；非流式输出默认把 DeepSeek 搜索引用标记转换为 Markdown 引用链接。
 
 ---
@@ -85,7 +86,7 @@ Vercel 一键部署可先只填 `DS2API_ADMIN_KEY`，部署后在 `/admin` 导
 - token 在 `config.keys` 中 → **托管账号模式**，自动轮询选择账号
 - token 不在 `config.keys` 中 → **直通 token 模式**，直接作为 DeepSeek token 使用
 
-**可选请求头**：`X-Ds2-Target-Account: <email_or_mobile>` — 指定使用某个托管账号；如果目标账号不存在，或管理账号队列已耗尽，相关业务请求会返回 `429`，当前不会附带 `Retry-After` 头。若账号存在但登录/刷新失败，则返回对应的 `401` 或上游错误。
+**可选请求头**：`X-Ds2-Target-Account: <email_or_mobile>` — 指定使用某个托管账号；如果目标账号不存在，或管理账号队列已耗尽，相关业务请求会返回 `429`，当前不会附带 `Retry-After` 头。若账号存在但登录/刷新失败，则返回对应的 `401` 或上游错误。未指定目标账号时，托管账号模式的 completion 空输出 429 会先尝试切到另一个可用账号 fresh retry 一次；指定目标账号或无其他可用账号时不会切号。
 Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=` 作为凭据来源。
 
 ### Admin 接口（`/admin/*`）
@@ -125,6 +126,9 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | POST | `/v1beta/models/{model}:streamGenerateContent` | 业务 | Gemini 流式 |
 | POST | `/v1/models/{model}:generateContent` | 业务 | Gemini 非流式兼容路径 |
 | POST | `/v1/models/{model}:streamGenerateContent` | 业务 | Gemini 流式兼容路径 |
+| GET | `/api/version` | 无 | Ollama 版本接口 |
+| GET | `/api/tags` | 无 | Ollama 模型列表 |
+| POST | `/api/show` | 无 | Ollama 单模型能力查询（返回 `id` 与 `capabilities`） |
 | POST | `/admin/login` | 无 | 管理登录 |
 | GET | `/admin/verify` | JWT | 校验管理 JWT |
 | GET | `/admin/vercel/config` | Admin | 读取 Vercel 预配置 |
@@ -168,12 +172,12 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | GET | `/admin/chat-history/{id}` | Admin | 查看单条服务器端对话记录 |
 | DELETE | `/admin/chat-history/{id}` | Admin | 删除单条服务器端对话记录 |
 | PUT | `/admin/chat-history/settings` | Admin | 更新对话记录保留条数 |
-
-服务器端记录本质上是 DeepSeek 上游响应归档：OpenAI Chat、OpenAI Responses、Claude Messages、Gemini GenerateContent 等直连 DeepSeek 的生成接口，在收到上游响应后会于各协议回译/裁剪前写入记录；列表按请求创建时间倒序展示，流式请求会在生成过程中持续刷新状态与详情。WebUI「API 测试」发出的请求也会进入该记录。
 | GET | `/admin/version` | Admin | 查询当前版本与最新 Release |
 
 OpenAI `/v1/*` 仍是规范路径。对于只配置 DS2API 根地址的客户端，同一套 OpenAI handler 也通过根路径快捷路由暴露：`/models`、`/models/{id}`、`/chat/completions`、`/responses`、`/responses/{response_id}`、`/embeddings`、`/files`、`/files/{file_id}`。
 
+服务器端记录本质上是 DeepSeek 上游响应归档：OpenAI Chat、OpenAI Responses、Claude Messages、Gemini GenerateContent 等直连 DeepSeek 的生成接口，在收到上游响应后会于各协议回译/裁剪前写入记录；列表按请求创建时间倒序展示，流式请求会在生成过程中持续刷新状态与详情。WebUI「API 测试」发出的请求也会进入该记录。
+
 ---
 
 ## 健康检查
@@ -227,16 +231,15 @@ OpenAI `/v1/*` 仍是规范路径。对于只配置 DS2API 根地址的客户端
 1. 先匹配 DeepSeek 原生模型。
 2. 再匹配 `model_aliases` 精确映射。
 3. 如果请求名以 `-nothinking` 结尾，则在最终解析出的规范模型上追加对应的无思考变体。
-4. 未命中时按模型家族规则回退（如 `o*`、`gpt-*`、`claude-*`）。
-5. 仍未命中则返回 `invalid_request_error`。
+4. 仍未命中则返回 `invalid_request_error`。当前不会按未知模型家族做启发式兜底；需要新增兼容名时请通过 `model_aliases` 明确配置。
 
 当前内置默认 alias 来自 `internal/config/models.go`，`config.model_aliases` 会在运行时覆盖或补充同名映射。节选：
 
 - OpenAI / Codex：`gpt-4o`、`gpt-4.1`、`gpt-5`、`gpt-5.5`、`gpt-5-codex`、`gpt-5.3-codex`、`codex-mini-latest`
 - OpenAI reasoning：`o1`、`o3`、`o3-deep-research`、`o4-mini`
 - Claude：`claude-opus-4-6`、`claude-sonnet-4-6`、`claude-haiku-4-5`、`claude-3-5-sonnet-latest`
-- Gemini：`gemini-2.5-pro`、`gemini-2.5-flash`、`gemini-pro-vision`
-- 其他兼容族：`llama-*`、`qwen-*`、`mistral-*`、`command-*` 会按家族启发式回退
+- Gemini：`gemini-2.5-pro`、`gemini-2.5-flash`、`gemini-3.1-pro`、`gemini-3-pro`、`gemini-3-flash`、`gemini-3.1-flash-lite`、`gemini-pro-vision`
+- 其他内置精确 alias：`llama-3.1-70b-instruct`、`qwen-max`
 
 上述 alias 若在请求名后追加 `-nothinking` 后缀，也会映射到对应的强制关闭 thinking 版本。
 当前视觉能力仅对应 `deepseek-v4-vision` / `deepseek-v4-vision-nothinking`，不会解析出独立的 `vision-search` 变体。
@@ -245,6 +248,8 @@ OpenAI `/v1/*` 仍是规范路径。对于只配置 DS2API 根地址的客户端
 
 ### `POST /v1/chat/completions`
 
+> 路径说明：除规范路径 `/v1/chat/completions` 外，也支持根路径快捷别名 `/chat/completions`。在 Vercel Runtime 上，`vercel.json` 仅把规范路径 `/v1/chat/completions` 重写到 Node 流式桥接；根路径快捷别名仍走 Go 主链路。因此 Vercel 上需要实时流式时请使用 `/v1/chat/completions`。
+
 **请求头**：
 
 ```http
@@ -256,7 +261,7 @@ 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-5.5`、`gpt-5.4-mini`、`gpt-5.3-codex`、`o3`、`claude-opus-4-6`、`claude-sonnet-4-6`、`gemini-2.5-pro`、`gemini-2.5-flash` 等）；若模型名带 `-nothinking` 后缀，则强制关闭 thinking / reasoning |
+| `model` | string | ✅ | 支持 DeepSeek 原生模型 + 常见 alias（如 `gpt-5.5`、`gpt-5.4-mini`、`gpt-5.3-codex`、`o3`、`claude-opus-4-6`、`claude-sonnet-4-6`、`gemini-2.5-pro`、`gemini-3.1-pro`、`gemini-3-flash` 等）；若模型名带 `-nothinking` 后缀，则强制关闭 thinking / reasoning |
 | `messages` | array | ✅ | OpenAI 风格消息数组 |
 | `stream` | boolean | ❌ | 默认 `false` |
 | `tools` | array | ❌ | Function Calling 定义 |
@@ -352,7 +357,8 @@ data: [DONE]
 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）。
-- 解析器当前把 DSML 外壳（`<|DSML|tool_calls>` / `<|DSML|invoke name="...">` / `<|DSML|parameter name="...">`）、DSML wrapper 别名（`<dsml|tool_calls>`、`<|tool_calls>`、`<｜tool_calls>`）、常见 DSML 分隔符漏写形态（如 `<|DSML tool_calls>` / `<|DSML invoke>` / `<|DSML parameter>`）、`DSML` 与工具标签名黏连的常见 typo（如 `<DSMLtool_calls>` / `<DSMLinvoke>` / `<DSMLparameter>`）和旧式 canonical XML 工具块（`<tool_calls>` / `<invoke name="...">` / `<parameter name="...">`）作为可执行调用解析；DSML 会先归一化回 XML，内部仍以 XML 解析语义为准。旧式 `<tools>`、`<tool_call>`、`<tool_name>`、`<param>`、`<function_call>`、`tool_use`、antml 风格与纯 JSON `tool_calls` 片段默认都会按普通文本处理。
+- 解析器当前把推荐 DSML 外壳（`<｜DSML｜tool_calls>` / `<｜DSML｜invoke name="...">` / `<｜DSML｜parameter name="...">`）、半角 DSML 外壳（`<|DSML|tool_calls>` / `<|DSML|invoke name="...">` / `<|DSML|parameter name="...">`）、DSML wrapper 别名（`<dsml|tool_calls>`、`<|tool_calls>`、`<｜tool_calls>`）、常见 DSML 分隔符漏写形态（如 `<|DSML tool_calls>` / `<|DSML invoke>` / `<|DSML parameter>`）、`DSML` 与工具标签名黏连的常见 typo（如 `<DSMLtool_calls>` / `<DSMLinvoke>` / `<DSMLparameter>`）、控制分隔符漂移（如 `<DSML␂tool_calls>` / 原始 STX `\x02`）、CJK 尖括号、全角感叹号、顿号、PascalCase 本地名、弯引号属性值与属性尾部分隔符漂移（如 `<DSM｜parameter name="command"｜>...〈/DSM｜parameter〉` / `<！DSML！invoke name=“Bash”>` / `<、DSML、tool_calls>` / `<DSmartToolCalls>` / `<DSMLtool_calls※>`）、任意协议前缀壳（如 `<proto💥tool_calls>`）和旧式 canonical XML 工具块（`<tool_calls>` / `<invoke name="...">` / `<parameter name="...">`）作为可执行调用解析；这些非结构性分隔符壳会先归一化回 XML，内部仍以 XML 解析语义为准，CDATA 开头也会容错 `<！[CDATA[` / `<、[CDATA[`。旧式 `<tools>`、`<tool_call>`、`<tool_name>`、`<param>`、`<function_call>`、`tool_use`、antml 风格与纯 JSON `tool_calls` 片段默认都会按普通文本处理；完整但 malformed 的 wrapper 同样会作为普通文本释放。
+- 解析层不会因为参数值为空而丢弃工具调用；显式空字符串或纯空白参数会按空字符串进入结构化 `tool_calls`。Prompt 会要求模型不要主动输出空参数，缺参/空命令的拒绝应由工具执行侧或客户端 schema 校验负责。
 - 当最终可见正文为空但思维链里包含可执行工具调用时，Chat / Responses 会在收尾阶段补发标准 OpenAI `tool_calls` / `function_call` 输出；如果客户端未开启 thinking / reasoning，该思维链只用于检测，不会作为可见正文或 `reasoning_content` 暴露。
 - Markdown fenced code block（例如 ```json ... ```）中的 `tool_calls` 仅视为示例文本，不会被执行。
 
@@ -628,6 +634,20 @@ data: {"type":"message_stop"}
 
 ---
 
+## Ollama 兼容接口
+
+- `POST /api/show` 请求体：`{"model":"<model-id>"}`。
+- 响应字段使用小写 `id`（不是 `ID`），并返回 `capabilities` 数组，便于与 Ollama 风格客户端/严格 schema 对齐。
+
+示例响应：
+
+```json
+{
+  "id": "deepseek-v4-flash",
+  "capabilities": ["tools", "thinking"]
+}
+```
+
 ## Admin 接口
 
 ### `POST /admin/login`
@@ -755,6 +775,7 @@ data: {"type":"message_stop"}
 - `responses` / `embeddings`
 - `auto_delete`（`mode`：`none` / `single` / `all`；旧配置 `sessions=true` 仍按 `all` 处理）
 - `current_input_file`（`enabled` 默认返回 `true`、`min_chars`）
+- `thinking_injection`（`enabled` 默认返回 `true`、`prompt`、`default_prompt`）
 - `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
@@ -769,6 +790,7 @@ data: {"type":"message_stop"}
 - `embeddings.provider`
 - `auto_delete.mode`
 - `current_input_file.enabled` / `current_input_file.min_chars`
+- `thinking_injection.enabled` / `thinking_injection.prompt`
 - `model_aliases`
 - `toolcall` 策略已固定，不再作为可写入字段
 
@@ -1251,7 +1273,7 @@ Gemini 路由使用 Google 风格错误结构：
 | 状态码 | 说明 |
 | --- | --- |
 | `401` | 鉴权失败（key/token 无效，或 Admin JWT 过期） |
-| `429` | 请求过多（超出并发上限 + 等待队列；当前不附带 `Retry-After` 头） |
+| `429` | 请求过多（超出并发上限 + 等待队列，或上游账号 thinking-only 后仍无可见输出；托管账号模式会先尝试一次切号 fresh retry；当前不附带 `Retry-After` 头） |
 | `503` | 模型不可用或上游服务异常 |
 
 ---

README.MD

@@ -134,7 +134,8 @@ flowchart LR
 | 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`、`GET /v1/files/{file_id}` |
 | 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） |
-| 统一 CORS 兼容 | `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/admin/*` 统一走同一套 CORS 策略；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同放行规则，尽量减少第三方预检请求头限制 |
+| Ollama 兼容 | `GET /api/version`、`GET /api/tags`、`POST /api/show` |
+| 统一 CORS 兼容 | `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/api/*`、`/admin/*` 统一走同一套 CORS 策略；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同放行规则，尽量减少第三方预检请求头限制 |
 | 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 | 纯 Go 高性能实现（DeepSeekHashV1），毫秒级响应 |
@@ -195,11 +196,11 @@ OpenAI `/v1/*` 仍是推荐的规范路径；同时支持 `/models`、`/chat/com
 - `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，请优先检查模型输出是否为推荐的 DSML 工具块：`<|DSML|tool_calls><|DSML|invoke name="..."><|DSML|parameter name="...">...`。兼容层也接受旧式 canonical XML：`<tool_calls><invoke name="..."><parameter name="...">...`；旧式 `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`、`<function_call>`、`tool_use` 或纯 JSON `tool_calls` 片段不会执行。
+- 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推荐的全角分隔符 DSML 工具块：`<｜DSML｜tool_calls><｜DSML｜invoke name="..."><｜DSML｜parameter name="...">...`。兼容层也接受半角 DSML 与旧式 canonical XML：`<tool_calls><invoke name="..."><parameter name="...">...`；旧式 `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`、`<function_call>`、`tool_use` 或纯 JSON `tool_calls` 片段不会执行，会作为普通文本处理。
 
 ### Gemini 接口
 
-Gemini 适配器将模型名通过 `model_aliases` 或内置规则映射到 DeepSeek 原生模型，支持 `generateContent` 和 `streamGenerateContent` 两种调用方式，并完整支持 Tool Calling（`functionDeclarations` → `functionCall` 输出）。若 Gemini 模型名带 `-nothinking` 后缀，例如 `gemini-2.5-pro-nothinking`，会映射到对应的强制关闭思考模型。
+Gemini 适配器将模型名通过 `model_aliases` 或内置精确 alias 映射到 DeepSeek 原生模型（覆盖 `gemini-2.5-*`、`gemini-3*`、`gemini-pro-vision` 等常见名称），支持 `generateContent` 和 `streamGenerateContent` 两种调用方式，并完整支持 Tool Calling（`functionDeclarations` → `functionCall` 输出）。若 Gemini 模型名带 `-nothinking` 后缀，例如 `gemini-2.5-pro-nothinking`，会映射到对应的强制关闭思考模型。
 
 ## 快速开始
 
@@ -295,13 +296,13 @@ cp config.example.json config.json
 base64 < config.json | tr -d '\n'
 ```
 
-> **流式说明**：`/v1/chat/completions` 在 Vercel 上默认走 `api/chat-stream.js`（Node Runtime）以保证实时 SSE。鉴权、账号选择、会话/PoW 准备仍由 Go 内部 prepare 接口完成；流式响应（含 `tools`）在 Node 侧执行与 Go 对齐的输出组装与防泄漏处理。虽然这里只有 OpenAI chat 流式走 Node，但 CORS 放行策略仍与 Go 主路由保持一致，统一覆盖第三方客户端预检场景。
+> **流式说明**：OpenAI Chat 流式在 Vercel 上会由 `api/chat-stream.js`（Node Runtime）承接，但 `vercel.json` 只把规范路径 `/v1/chat/completions` 重写到 Node；根路径快捷别名 `/chat/completions` 仍走 Go 主链路。鉴权、账号选择、会话/PoW 准备仍由 Go 内部 prepare 接口完成；流式响应（含 `tools`）在 Node 侧执行与 Go 对齐的输出组装与防泄漏处理。Vercel 上需要实时流式时请使用 `/v1/chat/completions`。
 
 详细部署说明请参阅 [部署指南](docs/DEPLOY.md)。
 
 ### 方式四：本地源码运行
 
-**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）；同时确保 `npm` 可用，建议 `npm 10+`
+**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时；CI / Docker 构建使用 Node 24）；同时确保 `npm` 可用，建议 `npm 10+`
 
 ```bash
 # 1. 克隆仓库
@@ -320,7 +321,7 @@ go run ./cmd/ds2api
 
 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。
 
-> **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 和 npm）。你也可以手动构建：`./scripts/build-webui.sh`
+> **WebUI 自动构建**：本地首次启动时，若 WebUI 静态目录不存在，会自动尝试执行 `npm ci --prefix webui`（仅在缺少依赖时）和 `npm run build --prefix webui -- --outDir static/admin --emptyOutDir`（需要本机有 Node.js 和 npm；静态目录可用 `DS2API_STATIC_ADMIN_DIR` 覆盖）。你也可以手动构建：`./scripts/build-webui.sh`
 
 ## 配置说明
 
@@ -350,6 +351,7 @@ go run ./cmd/ds2api
 
 可选请求头 `X-Ds2-Target-Account`：指定使用某个托管账号（值为 email 或 mobile）。
 如果指定账号不存在，或者当前管理账号队列已满，请求会返回 `429`；当前 `429` 不附带 `Retry-After` 头。若账号存在但登录/刷新失败，则返回对应的鉴权错误。
+未指定目标账号时，如果 completion 因上游 thinking-only 空输出在同账号补偿重试后仍将返回 `429 upstream_empty_output`，托管账号模式会自动切到下一个可用账号，新建 session，并用原始 payload 再 fresh retry 一次。
 Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `?key=` / `?api_key=` 作为调用方凭据。
 
 ## 并发模型
@@ -363,6 +365,7 @@ Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `
 
 - 当 in-flight 槽位满时，请求进入等待队列，**不会立即 429**
 - 超出总承载上限后才返回 `429 Too Many Requests`，当前响应不附带 `Retry-After`
+- completion 空输出类 429 会先做同账号补偿重试；托管账号模式还会在最终返回 429 前切到另一个可用账号 fresh retry 一次
 - `GET /admin/queue/status` 返回实时并发状态
 
 ## Tool Call 适配
@@ -370,12 +373,13 @@ 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. 解析层当前把 DSML 外壳视为推荐可执行调用：`<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`；兼容旧式 canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`。DSML 只是外壳别名，内部仍以 XML 解析语义为准；旧式 `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`、`<function_call>`、`tool_use` / antml 变体与纯 JSON `tool_calls` 片段都会按普通文本处理
+2. 解析层当前把全角分隔符 DSML 外壳视为推荐可执行调用：`<｜DSML｜tool_calls>` → `<｜DSML｜invoke name="...">` → `<｜DSML｜parameter name="...">`；兼容半角 DSML、旧式 canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`，以及若干 DSML 前缀/分隔符漂移。DSML 只是外壳别名，内部仍以 XML 解析语义为准；旧式 `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`、`<function_call>`、`tool_use` / antml 变体与纯 JSON `tool_calls` 片段都会按普通文本处理，完整但 malformed 的 wrapper 也会作为普通文本释放
 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
 
 > 说明：当前版本 parser 层以”尽量解析成功”为优先，所有格式合法的 XML 工具调用都会通过，不做工具名 allow-list 过滤。
+> 解析层会保留显式空字符串或纯空白参数；Prompt 会要求模型不要主动输出空参数，缺参/空命令的拒绝应由工具执行侧或客户端 schema 校验负责。
 >
 > 想评估”把工具调用封装成 XML 再输入模型”的方案，可参考：`docs/toolcall-semantics.md`。
 

README.en.md

@@ -131,7 +131,8 @@ For the full module-by-module architecture and directory responsibilities, see [
 | 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`, `GET /v1/files/{file_id}` |
 | 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) |
-| Unified CORS compatibility | `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, and `/admin/*` share one CORS policy; on Vercel, the Node Runtime for `/v1/chat/completions` mirrors the same relaxed preflight behavior for third-party clients |
+| Ollama compatible | `GET /api/version`, `GET /api/tags`, `POST /api/show` |
+| Unified CORS compatibility | `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, `/api/*`, and `/admin/*` share one CORS policy; on Vercel, the Node Runtime for `/v1/chat/completions` mirrors the same relaxed preflight behavior for third-party clients |
 | Multi-account rotation | Auto token refresh, email/mobile dual login |
 | Concurrency control | Per-account in-flight limit + waiting queue, dynamic recommended concurrency |
 | DeepSeek PoW | Pure Go high-performance solver (DeepSeekHashV1), ms-level response |
@@ -184,11 +185,11 @@ Besides the primary aliases above, `/anthropic/v1/models` also returns Claude 4.
 - 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, first verify the model output uses the recommended DSML block: `<|DSML|tool_calls><|DSML|invoke name="..."><|DSML|parameter name="...">...`. DS2API also accepts legacy canonical XML: `<tool_calls><invoke name="..."><parameter name="...">...`; legacy `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`, `<function_call>`, `tool_use`, or standalone JSON `tool_calls` are not executed.
+- If tool calls are rendered as plain text and not executed, first verify the model output uses the recommended fullwidth-separator DSML block: `<｜DSML｜tool_calls><｜DSML｜invoke name="..."><｜DSML｜parameter name="...">...`. DS2API also accepts halfwidth DSML and legacy canonical XML: `<tool_calls><invoke name="..."><parameter name="...">...`; legacy `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`, `<function_call>`, `tool_use`, or standalone JSON `tool_calls` are not executed and stay plain text.
 
 ### Gemini Endpoint
 
-The Gemini adapter maps model names to DeepSeek native models via `model_aliases` or built-in heuristics, supporting both `generateContent` and `streamGenerateContent` call patterns with full Tool Calling support (`functionDeclarations` → `functionCall` output).
+The Gemini adapter maps model names to DeepSeek native models via `model_aliases` or exact built-in aliases (covering common `gemini-2.5-*`, `gemini-3*`, and `gemini-pro-vision` names), supporting both `generateContent` and `streamGenerateContent` call patterns with full Tool Calling support (`functionDeclarations` → `functionCall` output). If the Gemini model name has a `-nothinking` suffix, such as `gemini-2.5-pro-nothinking`, it maps to the corresponding forced no-thinking model.
 
 ## Quick Start
 
@@ -283,13 +284,13 @@ Recommended: convert `config.json` to Base64 locally, then paste into `DS2API_CO
 base64 < config.json | tr -d '\n'
 ```
 
-> **Streaming note**: `/v1/chat/completions` on Vercel is routed to `api/chat-stream.js` (Node Runtime) for real-time SSE. Auth, account selection, and session/PoW preparation are still handled by the Go internal prepare endpoint; streaming output (including `tools`) is assembled on Node with Go-aligned anti-leak handling. This is the only interface family currently routed through Node, and its CORS allow behavior is kept aligned with the Go router so third-party preflight handling stays unified.
+> **Streaming note**: OpenAI Chat streaming on Vercel is routed to `api/chat-stream.js` (Node Runtime), but `vercel.json` rewrites only the canonical `/v1/chat/completions` path to Node; the root shortcut `/chat/completions` stays on the Go main path. Auth, account selection, and session/PoW preparation are still handled by the Go internal prepare endpoint; streaming output (including `tools`) is assembled on Node with Go-aligned anti-leak handling. Use `/v1/chat/completions` on Vercel when real-time streaming is required.
 
 For detailed deployment instructions, see the [Deployment Guide](docs/DEPLOY.en.md).
 
 ### Option 4: Local Run
 
-**Prerequisites**: Go 1.26+, Node.js `20.19+` or `22.12+` (only if building WebUI locally)
+**Prerequisites**: Go 1.26+, Node.js `20.19+` or `22.12+` (only if building WebUI locally; CI / Docker builds use Node 24), and npm available; npm 10+ is recommended
 
 ```bash
 # 1. Clone
@@ -308,7 +309,7 @@ Default local URL: `http://127.0.0.1:5001`
 
 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.
 
-> **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`
+> **WebUI auto-build**: On first local startup, if the WebUI static directory is missing, DS2API auto-runs `npm ci --prefix webui` (only when dependencies are missing) and `npm run build --prefix webui -- --outDir static/admin --emptyOutDir` (requires Node.js; `DS2API_STATIC_ADMIN_DIR` can override the static directory). You can also build manually: `./scripts/build-webui.sh`
 
 ## Configuration
 
@@ -336,6 +337,7 @@ For business endpoints (`/v1/*`, `/anthropic/*`, Gemini routes), DS2API supports
 | **Direct token** | If the token is not in `config.keys`, DS2API treats it as a DeepSeek token directly |
 
 Optional header `X-Ds2-Target-Account`: Pin a specific managed account (value is email or mobile).
+When no target account is pinned, if a completion would end as `429 upstream_empty_output` after the same-account empty-output retry, managed-account mode switches to the next available account, creates a fresh session, and retries the original payload once.
 Gemini routes also accept `x-goog-api-key`, or `?key=` / `?api_key=` when no auth header is present.
 
 ## Concurrency Model
@@ -348,7 +350,8 @@ Queue limit = DS2API_ACCOUNT_MAX_QUEUE (default = recommended concurrency)
 ```
 
 - When inflight slots are full, requests enter a waiting queue — **no immediate 429**
-- 429 is returned only when total load exceeds inflight + queue capacity
+- 429 is returned only when total load exceeds inflight + queue capacity; current responses do not include `Retry-After`
+- Completion empty-output 429s first get the same-account compensation retry; managed-account mode also tries one alternate-account fresh retry before returning the final 429
 - `GET /admin/queue/status` returns real-time concurrency state
 
 ## Tool Call Adaptation
@@ -356,12 +359,13 @@ 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)
-2. The parser now treats the DSML shell as the recommended executable tool-calling syntax: `<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`; it also accepts legacy canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`. DSML is a shell alias and internal parsing remains XML-based; legacy `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`, `<function_call>`, `tool_use`, antml variants, and standalone JSON `tool_calls` payloads are treated as plain text
+2. The parser treats the fullwidth-separator DSML shell as the recommended executable tool-calling syntax: `<｜DSML｜tool_calls>` → `<｜DSML｜invoke name="...">` → `<｜DSML｜parameter name="...">`; it also accepts halfwidth DSML, legacy canonical XML `<tool_calls>` → `<invoke name="...">` → `<parameter name="...">`, plus common DSML prefix/separator drift. DSML is a shell alias and internal parsing remains XML-based; legacy `<tools>` / `<tool_call>` / `<tool_name>` / `<param>`, `<function_call>`, `tool_use`, antml variants, and standalone JSON `tool_calls` payloads are treated as plain text, and complete but malformed wrappers are released as plain text too
 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. 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.
+> Explicit empty strings or whitespace-only parameters are preserved by the parser; prompting tells the model not to emit blank parameters, and missing/empty argument rejection belongs in the tool executor or client schema validation.
 
 ## Local Dev Packet Capture
 

VERSION

@@ -1 +1 @@
-4.4.3
+4.5.0

docs/ARCHITECTURE.en.md

@@ -27,7 +27,7 @@ ds2api/
 │   ├── claudeconv/                       # Claude message conversion helpers
 │   ├── compat/                           # Compatibility and regression helpers
 │   ├── assistantturn/                    # Upstream output to canonical assistant turn / stream event semantics
-│   ├── completionruntime/                # Shared Go DeepSeek completion startup, non-stream collection, and retry
+│   ├── completionruntime/                # Shared Go DeepSeek completion startup, collection, empty-output/account-switch retry
 │   ├── config/                           # Config loading/validation/hot reload
 │   ├── deepseek/                         # DeepSeek upstream client/protocol/transport
 │   │   ├── client/                       # Login/session/completion/upload/delete calls
@@ -41,6 +41,7 @@ ds2api/
 │   │   ├── admin/                        # Admin API root assembly and resource packages
 │   │   ├── claude/                       # Claude HTTP protocol adapter
 │   │   ├── gemini/                       # Gemini HTTP protocol adapter
+│   │   ├── ollama/                       # Ollama-compatible model/capability query endpoints
 │   │   ├── openai/                       # OpenAI HTTP surface
 │   │   │   ├── chat/                     # Chat Completions execution entrypoint
 │   │   │   ├── responses/                # Responses API and response store
@@ -57,6 +58,7 @@ ds2api/
 │   ├── prompt/                           # Prompt composition
 │   ├── promptcompat/                     # API request -> DeepSeek web-chat plain-text compatibility
 │   ├── rawsample/                        # Raw sample read/write and management
+│   ├── responsehistory/                  # DeepSeek upstream response archive and session snapshots
 │   ├── server/                           # Router and middleware assembly
 │   │   └── data/                         # Router/runtime helper data
 │   ├── sse/                              # SSE parsing utilities
@@ -188,10 +190,11 @@ flowchart LR
 - `internal/server`: router tree + middlewares (health, protocol routes, Admin/WebUI).
 - `internal/httpapi/openai/*`: OpenAI HTTP surface split into chat, responses, files, embeddings, history, and shared packages; chat/responses share the promptcompat, stream, and toolcall semantics.
 - `internal/httpapi/{claude,gemini}`: protocol adapters that normalize into the same prompt compatibility semantics; normal direct paths must share DeepSeek session/PoW/completion execution through `completionruntime`, while `translatorcliproxy` is reserved for Vercel prepare/release, missing-backend fallback, and regression tests.
+- `internal/httpapi/ollama`: Ollama-compatible model list and capability query endpoints.
 - `internal/httpapi/requestbody`: shared HTTP body reading, JSON pre-validation, and UTF-8 error helpers across protocol adapters.
 - `internal/promptcompat`: compatibility core for turning OpenAI/Claude/Gemini requests into DeepSeek web-chat plain-text context.
 - `internal/assistantturn`: Go output-side canonical semantics, converting DeepSeek SSE collection results and stream finalization state into assistant turns and centralizing thinking, tool call, citation, usage, stop/error behavior.
-- `internal/completionruntime`: shared Go completion execution helpers for DeepSeek session/PoW/call startup, non-stream collection, and empty-output retry; streaming paths use it to start upstream requests, continue to use `internal/stream` for real-time consumption, and use `assistantturn` during finalization.
+- `internal/completionruntime`: shared Go completion execution helpers for DeepSeek session/PoW/call startup, non-stream collection, empty-output retry, and one managed-account fresh retry before a final 429; streaming paths use it to start upstream requests, continue to use `internal/stream` for real-time consumption, and use `assistantturn` during finalization.
 - `internal/translatorcliproxy`: bridge compatibility layer for Claude/Gemini and OpenAI shape translation; it is not the main business protocol conversion center.
 - `internal/deepseek/{client,protocol,transport}`: upstream requests, sessions, PoW adaptation, protocol constants, and transport details.
 - `internal/js/chat-stream` + `api/chat-stream.js`: Vercel Node streaming bridge; Go prepare/release owns auth, account lease, and completion payload assembly, while Node relays real-time SSE with Go-aligned finalization and tool sieve semantics.
@@ -199,6 +202,7 @@ flowchart LR
 - `internal/toolcall` + `internal/toolstream`: DSML shell compatibility plus canonical XML tool-call parsing and anti-leak sieve; DSML is normalized back to XML at the entrypoint, and internal parsing remains XML-based.
 - `internal/httpapi/admin/*`: Admin API root assembly plus auth/accounts/config/settings/proxies/rawsamples/vercel/history/devcapture/version resource packages.
 - `internal/chathistory`: server-side conversation history persistence, pagination, detail lookup, and retention policy.
+- `internal/responsehistory`: DeepSeek upstream response archive, saving assistant text, thinking, raw tool-call fragments, and streaming detail before protocol rendering/trimming.
 - `internal/config`: config loading/validation + runtime settings hot-reload.
 - `internal/account`: managed account pool, inflight slots, waiting queue.
 - `internal/textclean`: text cleanup helpers, e.g. stripping `[reference: N]` markers.

docs/ARCHITECTURE.md

@@ -27,7 +27,7 @@ ds2api/
 │   ├── claudeconv/                       # Claude 消息格式转换工具
 │   ├── compat/                           # 兼容性辅助与回归支持
 │   ├── assistantturn/                    # 上游输出到统一 assistant turn / stream event 的语义层
-│   ├── completionruntime/                # Go 主路径共享 DeepSeek completion 启动、非流式收集与 retry
+│   ├── completionruntime/                # Go 主路径共享 DeepSeek completion 启动、收集、空输出/切号 retry
 │   ├── config/                           # 配置加载、校验、热更新
 │   ├── deepseek/                         # DeepSeek 上游 client/protocol/transport
 │   │   ├── client/                       # 登录、会话、completion、上传/删除等上游调用
@@ -41,6 +41,7 @@ ds2api/
 │   │   ├── admin/                        # Admin API 根装配与资源子包
 │   │   ├── claude/                       # Claude HTTP 协议适配
 │   │   ├── gemini/                       # Gemini HTTP 协议适配
+│   │   ├── ollama/                       # Ollama 兼容模型/能力查询接口
 │   │   ├── openai/                       # OpenAI HTTP surface
 │   │   │   ├── chat/                     # Chat Completions 执行入口
 │   │   │   ├── responses/                # Responses API 与 response store
@@ -57,6 +58,7 @@ ds2api/
 │   ├── prompt/                           # Prompt 组装
 │   ├── promptcompat/                     # API 请求到 DeepSeek 网页纯文本上下文兼容层
 │   ├── rawsample/                        # raw sample 读写与管理
+│   ├── responsehistory/                  # DeepSeek 上游响应归档与会话快照
 │   ├── server/                           # 路由与中间件装配
 │   │   └── data/                         # 路由/运行时辅助数据
 │   ├── sse/                              # SSE 解析工具
@@ -188,10 +190,11 @@ flowchart LR
 - `internal/server`：路由树和中间件挂载（健康检查、协议入口、Admin/WebUI）。
 - `internal/httpapi/openai/*`：OpenAI HTTP surface，按 chat、responses、files、embeddings、history、shared 拆分；chat/responses 共享 promptcompat、stream、toolcall 等核心语义。
 - `internal/httpapi/{claude,gemini}`：协议输入输出适配，归一到同一套 prompt compatibility 语义；正常直连路径必须通过 `completionruntime` 共享 DeepSeek session/PoW/completion 调用，`translatorcliproxy` 仅保留给 Vercel prepare/release、后端缺失 fallback 和回归测试。
+- `internal/httpapi/ollama`：Ollama 兼容的模型列表与能力查询入口。
 - `internal/httpapi/requestbody`：跨协议复用的请求体读取、JSON 解码前置校验与 UTF-8 错误处理辅助。
 - `internal/promptcompat`：OpenAI/Claude/Gemini 请求到 DeepSeek 网页纯文本上下文的兼容内核。
 - `internal/assistantturn`：Go 输出侧统一语义层，把 DeepSeek SSE 收集结果和流式收尾状态归一成 assistant turn，集中处理 thinking、tool call、citation、usage、stop/error 语义。
-- `internal/completionruntime`：Go surface 共享的 completion 执行辅助，负责 DeepSeek session/PoW/call 启动、非流式 collect 和 empty-output retry；流式路径复用它启动上游请求，继续用 `internal/stream` 做实时消费，并在最终收尾阶段接入 `assistantturn`。
+- `internal/completionruntime`：Go surface 共享的 completion 执行辅助，负责 DeepSeek session/PoW/call 启动、非流式 collect、empty-output retry，以及托管账号在最终 429 前的一次切号 fresh retry；流式路径复用它启动上游请求，继续用 `internal/stream` 做实时消费，并在最终收尾阶段接入 `assistantturn`。
 - `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转的桥接兼容层，不作为主业务协议转换中心。
 - `internal/deepseek/{client,protocol,transport}`：上游请求、会话、PoW 适配、协议常量与传输层。
 - `internal/js/chat-stream` + `api/chat-stream.js`：Vercel Node 流式桥；Go prepare/release 管理鉴权、账号租约和 completion payload，Node 侧负责实时 SSE 转发并保持 Go 对齐的终结态和 tool sieve 语义。
@@ -199,6 +202,7 @@ flowchart LR
 - `internal/toolcall` + `internal/toolstream`：DSML 外壳兼容与 canonical XML 工具调用解析、防泄漏筛分；DSML 会在入口归一化回 XML，内部仍按 XML 语义解析。
 - `internal/httpapi/admin/*`：Admin API 根装配与 auth/accounts/config/settings/proxies/rawsamples/vercel/history/devcapture/version 等资源子包。
 - `internal/chathistory`：服务器端对话记录持久化、分页、单条详情和保留策略。
+- `internal/responsehistory`：DeepSeek 上游响应归档，会在协议回译/裁剪前保存 assistant text、thinking、tool-call 原始片段和流式详情。
 - `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。
 - `internal/textclean`：文本清洗，移除 `[reference: N]` 标记等噪声。

docs/CONTRIBUTING.en.md

@@ -9,8 +9,8 @@ Thanks for your interest in contributing to DS2API!
 ### Prerequisites
 
 - Go 1.26+
-- Node.js `20.19+` or `22.12+` (for WebUI development)
-- npm (bundled with Node.js)
+- Node.js `20.19+` or `22.12+` (for WebUI development; CI / Docker builds use Node 24)
+- npm (bundled with Node.js; 10+ recommended)
 
 ### Backend Development
 

docs/CONTRIBUTING.md

@@ -9,8 +9,8 @@
 ### 前置要求
 
 - Go 1.26+
-- Node.js `20.19+` 或 `22.12+`（WebUI 开发时）
-- npm（随 Node.js 提供）
+- Node.js `20.19+` 或 `22.12+`（WebUI 开发时；CI / Docker 构建使用 Node 24）
+- npm（随 Node.js 提供，建议 10+）
 
 ### 后端开发
 

docs/DEPLOY.en.md

@@ -39,8 +39,8 @@ Recommended order when choosing a deployment method:
 | Dependency | Minimum Version | Notes |
 | --- | --- | --- |
 | Go | 1.26+ | Build backend |
-| Node.js | `20.19+` or `22.12+` | Only needed to build WebUI locally |
-| npm | Bundled with Node.js | Install WebUI dependencies |
+| Node.js | `20.19+` or `22.12+` (CI / Docker builds use Node 24) | Only needed to build WebUI locally |
+| npm | Bundled with Node.js; 10+ recommended | Install WebUI dependencies |
 
 Config source (choose one):
 
@@ -299,6 +299,8 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # optional for personal accounts
 | `DS2API_VERCEL_INTERNAL_SECRET` | Hybrid streaming internal auth | Falls back to `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | Stream lease TTL | `900` |
 | `DS2API_RAW_STREAM_SAMPLE_ROOT` | Raw stream sample root for saving/reading samples | `tests/raw_stream_samples` |
+| `DS2API_STATIC_ADMIN_DIR` | WebUI static asset directory | `static/admin` |
+| `DS2API_AUTO_BUILD_WEBUI` | Whether local startup auto-builds missing WebUI assets (`1/true/yes/on` or `0/false/no/off`) | Enabled outside Vercel |
 | `VERCEL_TOKEN` | Vercel sync token | — |
 | `VERCEL_PROJECT_ID` | Vercel project ID | — |
 | `VERCEL_TEAM_ID` | Vercel team ID | — |
@@ -321,7 +323,7 @@ Request ──────┐
 ```
 
 - **Go entry**: `api/index.go` (Serverless Go)
-- **Stream entry**: `api/chat-stream.js` (Node Runtime for real-time SSE)
+- **Stream entry**: `api/chat-stream.js` (Node Runtime for real-time SSE; `vercel.json` rewrites only the canonical `/v1/chat/completions` path here, while the root shortcut `/chat/completions` stays on the Go entry)
 - **Routing**: `vercel.json`
 - **Build command**: `npm ci --prefix webui && npm run build --prefix webui` (automatic)
 
@@ -438,7 +440,7 @@ Default local access URL: `http://127.0.0.1:5001`; the server actually binds to
 
 ### 4.2 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`).
+On first local startup, if the WebUI static directory is missing, DS2API automatically attempts to build it (requires Node.js/npm; when dependencies are missing it runs `npm ci --prefix webui`, then `npm run build --prefix webui -- --outDir <static-dir> --emptyOutDir`). The default static directory is `static/admin/`, and `DS2API_STATIC_ADMIN_DIR` can override it.
 
 Manual build:
 

docs/DEPLOY.md

@@ -39,8 +39,8 @@
 | 依赖 | 最低版本 | 说明 |
 | --- | --- | --- |
 | Go | 1.26+ | 编译后端 |
-| 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 |
+| Node.js | `20.19+` 或 `22.12+`（CI / Docker 构建使用 Node 24） | 仅在需要本地构建 WebUI 时 |
+| npm | 随 Node.js 提供，建议 10+ | 安装 WebUI 依赖 |
 
 配置来源（任选其一）：
 
@@ -299,6 +299,8 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # 个人账号可留空
 | `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` |
 | `DS2API_RAW_STREAM_SAMPLE_ROOT` | raw stream 样本保存/读取根目录 | `tests/raw_stream_samples` |
+| `DS2API_STATIC_ADMIN_DIR` | WebUI 静态资源目录 | `static/admin` |
+| `DS2API_AUTO_BUILD_WEBUI` | 本地启动时是否自动构建缺失的 WebUI（`1/true/yes/on` 或 `0/false/no/off`） | 非 Vercel 默认开启 |
 | `VERCEL_TOKEN` | Vercel 同步 token | — |
 | `VERCEL_PROJECT_ID` | Vercel 项目 ID | — |
 | `VERCEL_TEAM_ID` | Vercel 团队 ID | — |
@@ -331,7 +333,7 @@ api/index.go  api/chat-stream.js
 ```
 
 - **入口文件**：`api/index.go`（Serverless Go）
-- **流式入口**：`api/chat-stream.js`（Node Runtime，保证实时 SSE）
+- **流式入口**：`api/chat-stream.js`（Node Runtime，保证实时 SSE；`vercel.json` 仅把规范路径 `/v1/chat/completions` 重写到这里，根路径快捷别名 `/chat/completions` 仍走 Go 入口）
 - **路由重写**：`vercel.json`
 - **构建命令**：`npm ci --prefix webui && npm run build --prefix webui`（自动执行）
 
@@ -448,7 +450,7 @@ go run ./cmd/ds2api
 
 ### 4.2 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`）。
+本地首次启动时，若 WebUI 静态目录不存在，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 --prefix webui`，再执行 `npm run build --prefix webui -- --outDir <静态目录> --emptyOutDir`）。默认静态目录为 `static/admin/`，可用 `DS2API_STATIC_ADMIN_DIR` 覆盖。
 
 你也可以手动构建：
 

docs/DEVELOPMENT.md

@@ -81,7 +81,7 @@ Tool call 问题优先跑：
 
 ```bash
 go test -v ./internal/toolcall ./internal/toolstream -count=1
-node --test tests/node/stream-tool-sieve.test.js tests/node/chat-stream.test.js
+./tests/scripts/run-unit-node.sh
 ```
 
 ## 5. 测试选择

docs/TESTING.md

@@ -75,7 +75,7 @@ npm run build --prefix webui
 1. **Preflight 检查**：
    - `go test ./... -count=1`（单元测试）
    - `./tests/scripts/check-node-split-syntax.sh`（Node 拆分模块语法门禁）
-   - `node --test tests/node/stream-tool-sieve.test.js tests/node/chat-stream.test.js tests/node/js_compat_test.js`
+   - `node --test --test-concurrency=1 tests/node/stream-tool-sieve.test.js tests/node/chat-stream.test.js tests/node/chat-history-utils.test.js tests/node/js_compat_test.js`
    - `npm run build --prefix webui`（WebUI 构建检查）
 
 2. **隔离启动**：复制 `config.json` 到临时目录，启动独立服务进程
@@ -203,10 +203,10 @@ 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/toolcall/
+go test -v -run 'TestParseToolCalls|TestProcessToolSieve|TestRepair' ./internal/toolcall ./internal/toolstream
 
 # 运行单个测试用例
-go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/toolcall/
+go test -v -run TestParseToolCallsAllowsAllEmptyParameterPayload ./internal/toolcall
 
 # 运行 format 相关测试
 go test -v ./internal/format/...
@@ -221,23 +221,23 @@ go test -v ./internal/httpapi/openai/...
 
 ```bash
 # 1. 运行 tool calls 相关的所有测试
-go test -v -run 'TestParseToolCalls|TestRepair' ./internal/toolcall/
+go test -v -run 'TestParseToolCalls|TestProcessToolSieve|TestRepair' ./internal/toolcall ./internal/toolstream
 
 # 2. 查看测试输出中的详细调试信息
-go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/toolcall/ 2>&1
+go test -v -run TestProcessToolSieveReleasesMalformedExecutableXMLBlock ./internal/toolstream 2>&1
 
 # 3. 检查具体测试用例的修复效果
-# 测试用例位于 internal/toolcall/toolcalls_test.go，包含：
-# - TestParseToolCallsWithDeepSeekHallucination: DeepSeek 典型幻觉输出
+# 重点测试位于 internal/toolcall/toolcalls_test.go 与 internal/toolstream/tool_sieve_xml_test.go，包含：
+# - TestParseToolCallsAllowsAllEmptyParameterPayload: 空参数结构化保留
+# - TestProcessToolSieveReleasesMalformedExecutableXMLBlock: malformed XML wrapper 释放为文本
 # - TestRepairLooseJSONWithNestedObjects: 嵌套对象的方括号修复
-# - TestParseToolCallsWithMixedWindowsPaths: Windows 路径处理
 ```
 
 ### 运行 Node.js 测试
 
 ```bash
 # 运行 Node 测试
-node --test tests/node/stream-tool-sieve.test.js
+node --test --test-concurrency=1 tests/node/stream-tool-sieve.test.js tests/node/chat-stream.test.js tests/node/chat-history-utils.test.js tests/node/js_compat_test.js
 
 # 或使用脚本
 ./tests/scripts/run-unit-node.sh

docs/prompt-compatibility.md

@@ -111,8 +111,8 @@ DS2API 当前的核心思路，不是把客户端传来的 `messages`、`tools`
 - OpenAI Chat / Responses 原生走统一 OpenAI 标准化与 DeepSeek payload 组装；Claude / Gemini 会尽量复用 OpenAI prompt/tool 语义，其中 Gemini 直接复用 `promptcompat.BuildOpenAIPromptForAdapter`。Go 主服务新增 `completionruntime` 启动层，统一执行 DeepSeek session/PoW/call；输出侧新增 `assistantturn` 语义层：非流式 OpenAI Chat / Responses / Claude / Gemini 会把 DeepSeek SSE 收集结果先归一成同一份 assistant turn，再分别渲染成各协议原生外形；流式 OpenAI Chat / Responses / Claude / Gemini 继续保持各协议实时 SSE framing，但最终收尾的 tool fallback、schema 归一、usage、empty-output / content-filter 错误语义同样由 `assistantturn` 判定。Claude / Gemini 的常规 Go 主路径不再依赖内部 `httptest` 转发到 OpenAI handler；`translatorcliproxy` 仅保留用于 Vercel bridge、后端缺失 fallback 和回归测试，不作为主业务协议转换中心。
 - Vercel Node 流式路径本轮不迁移，仍使用现有 Node bridge / stream-tool-sieve 实现；后续若变更 Node 流式语义，需要按 `assistantturn` 的 Go canonical 输出语义同步对齐。
 - 客户端传入的 thinking / reasoning 开关会被归一到下游 `thinking_enabled`。Gemini `generationConfig.thinkingConfig.thinkingBudget` 会翻译成同一套 thinking 开关；关闭时即使上游返回 `response/thinking_content`，兼容层也不会把它当作可见正文输出。若最终解析出的模型名带 `-nothinking` 后缀，则会无条件强制关闭 thinking，优先级高于请求体中的 `thinking` / `reasoning` / `reasoning_effort`。未显式关闭时，各 surface 会按解析后的 DeepSeek 模型默认能力开启 thinking，并用各自协议的原生形态暴露：OpenAI Chat 为 `reasoning_content`，OpenAI Responses 为 `response.reasoning.delta` / `reasoning` content，Claude 为 `thinking` block / `thinking_delta`，Gemini 为 `thought: true` part。
-- 对 OpenAI Chat / Responses 的非流式收尾，如果最终可见正文为空，兼容层会优先尝试把思维链中的独立 DSML / XML 工具块当作真实工具调用解析出来。流式链路也会在收尾阶段做同样的 fallback 检测，但不会因为思维链内容去中途拦截或改写流式输出；真正的工具识别始终基于原始上游文本，而不是基于“已经做过可见输出清洗”的版本，因此即使最终可见层会剥离完整 leaked DSML / XML `tool_calls` wrapper、并抑制全空参数或无效 wrapper 块，也不会影响真实工具调用转成结构化 `tool_calls` / `function_call`。补发结果会作为本轮 assistant 的结构化 `tool_calls` / `function_call` 输出返回，而不是塞进 `content` 文本；如果客户端没有开启 thinking / reasoning，思维链只用于检测，不会作为 `reasoning_content` 或可见正文暴露。只有正文为空且思维链里也没有可执行工具调用时，才继续按空回复错误处理。
-- OpenAI Chat / Responses 的空回复错误处理之前会默认做一次内部补偿重试：第一次上游完整结束后，如果最终可见正文为空、没有解析到工具调用、也没有已经向客户端流式发出工具调用，并且终止原因不是 `content_filter`，兼容层会复用同一个 `chat_session_id`、账号、token 与工具策略，把原始 completion `prompt` 追加固定后缀 `Previous reply had no visible output. Please regenerate the visible final answer or tool call now.` 后重新提交一次。重试遵循 DeepSeek 多轮对话协议：从第一次上游 SSE 流中提取 `response_message_id`，并在重试 payload 中设置 `parent_message_id` 为该值，使重试成为同一会话的后续轮次而非断裂的根消息；同时重新获取一次 PoW（若 PoW 获取失败则回退到原始 PoW）。该重试不会重新标准化消息、不会新建 session、不会切换账号，也不会向流式客户端插入重试标记；第二次 thinking / reasoning 会按正常增量直接接到第一次之后，并继续使用 overlap trim 去重。若第二次仍为空，终端错误码仍保持现有 `upstream_empty_output`；若任一尝试触发空 `content_filter`，不做补偿重试并保持 `content_filter` 错误。JS Vercel 运行时同样设置 `parent_message_id`，但因无法直接调用 PoW API 而复用原始 PoW。
+- 对 OpenAI Chat / Responses 的非流式收尾，如果最终可见正文为空，兼容层会优先尝试把思维链中的独立 DSML / XML 工具块当作真实工具调用解析出来。流式链路也会在收尾阶段做同样的 fallback 检测，但不会因为思维链内容去中途拦截或改写流式输出；真正的工具识别始终基于原始上游文本，而不是基于“已经做过可见输出清洗”的版本。最终可见层会剥离已经成功解析成工具调用的完整 leaked DSML / XML `tool_calls` wrapper；如果遇到完整 wrapper 但内部形态不符合可执行工具调用语义（例如 `<param>` 这类 malformed XML 工具壳），流式 sieve 会把该块作为普通文本释放，而不是吞掉或伪造成工具调用。补发结果会作为本轮 assistant 的结构化 `tool_calls` / `function_call` 输出返回，而不是塞进 `content` 文本；如果客户端没有开启 thinking / reasoning，思维链只用于检测，不会作为 `reasoning_content` 或可见正文暴露。只有正文为空且思维链里也没有可执行工具调用时，才继续按空回复错误处理。
+- OpenAI Chat / Responses、Claude Messages、Gemini generateContent 的空回复错误处理之前会默认做一次内部补偿重试：第一次上游完整结束后，如果最终可见正文为空、没有解析到工具调用、也没有已经向客户端流式发出工具调用，并且终止原因不是 `content_filter`，兼容层会复用同一个 `chat_session_id`、账号、token 与工具策略，把原始 completion `prompt` 追加固定后缀 `Previous reply had no visible output. Please regenerate the visible final answer or tool call now.` 后重新提交一次。Go 主路径的非流式重试由 `completionruntime.ExecuteNonStreamWithRetry` 统一处理；流式重试由 `completionruntime.ExecuteStreamWithRetry` 统一处理，各协议 runtime 只负责消费/渲染本协议 SSE framing。重试遵循 DeepSeek 多轮对话协议：从第一次上游 SSE 流中提取 `response_message_id`，并在重试 payload 中设置 `parent_message_id` 为该值，使重试成为同一会话的后续轮次而非断裂的根消息；同时重新获取一次 PoW（若 PoW 获取失败则回退到原始 PoW）。该同账号重试不会重新标准化消息、不会新建 session，也不会向流式客户端插入重试标记；第二次 thinking / reasoning 会按正常增量直接接到第一次之后，并继续使用 overlap trim 去重。若同账号补偿重试后即将返回 429 `upstream_empty_output`，并且当前是托管账号模式，Go 主路径会在返回 429 前切换到下一个可用账号，新建 `chat_session_id`，使用原始 completion payload 再做一次 fresh retry；该切号重试不携带空回复 prompt 后缀，也不设置上一账号的 `parent_message_id`。如果没有可切换账号，或切号后的 fresh retry 仍没有可见正文或工具调用，则继续按原错误返回：无任何输出为 503 `upstream_unavailable`，有 reasoning 但没有可见正文或工具调用为 429 `upstream_empty_output`。若任一尝试触发空 `content_filter`，不做补偿重试并保持 `content_filter` 错误。JS Vercel 运行时同样设置 `parent_message_id`，但因无法直接调用 PoW API 而复用原始 PoW；切号 fresh retry 目前由 Go 主路径提供。
 
 - 非流式 OpenAI Chat / Responses、Claude Messages、Gemini generateContent 在最终可见正文渲染阶段，会把 DeepSeek 搜索返回中的 `[citation:N]` / `[reference:N]` 标记替换成对应 Markdown 链接。`citation` 标记按一基序号解析；`reference` 标记只有在同一段正文中出现 `[reference:0]`（允许冒号后有空格）时才按零基序号映射，并且不会影响同段正文里的 `citation` 标记。
 - 流式输出仍默认隐藏 `[citation:N]` / `[reference:N]` 这类上游内部标记，避免分片输出中泄漏尚未完成映射的引用占位符。
@@ -167,14 +167,15 @@ OpenAI Chat / Responses 在标准化后、current input file 之前，会默认
 3. 再附上统一的 DSML tool call 外壳格式约束。
 4. 把这整段内容并入 system prompt。
 
-工具调用正例现在优先示范官方 DSML 风格：`<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`。
-兼容层仍接受旧式纯 `<tool_calls>` wrapper，并会容错若干 DSML 标签变体，包括短横线形式 `<dsml-tool-calls>` / `<dsml-invoke>` / `<dsml-parameter>`；但提示词会优先要求模型输出官方 DSML 标签，并强调不能只输出 closing wrapper 而漏掉 opening tag。需要注意：这是“兼容 DSML 外壳，内部仍以 XML 解析语义为准”，不是原生 DSML 全链路实现；DSML 标签会在解析入口归一化回现有 XML 标签后继续走同一套 parser。
+工具调用正例现在优先示范全角分隔符 DSML 风格：`<｜DSML｜tool_calls>` → `<｜DSML｜invoke name="...">` → `<｜DSML｜parameter name="...">`。
+兼容层仍接受旧式纯 `<tool_calls>` wrapper，并会容错若干 DSML 标签变体，包括短横线形式 `<dsml-tool-calls>` / `<dsml-invoke>` / `<dsml-parameter>`、下划线形式 `<dsml_tool_calls>` / `<dsml_invoke>` / `<dsml_parameter>`，以及其他前缀分隔形态如 `<vendor|tool_calls>` / `<vendor_tool_calls>` / `<vendor - tool_calls>`；标签壳扫描还会把全角 ASCII 漂移归一化，例如 `<ｄＳＭＬ｜tool_calls>` 与全角 `＞` 结束符，也会容错 CJK 尖括号、全角感叹号或顿号分隔符、弯引号属性值、PascalCase 本地名和属性尾部分隔符漂移，例如 `<DSM｜parameter name="command"｜>...〈/DSM｜parameter〉`、`<！DSML！invoke name=“Bash”>`、`<、DSML、tool_calls>`、`<DSmartToolCalls>`、`<DSMLtool_calls※>`。更一般地，Go / Node tag 扫描以固定本地标签名 `tool_calls` / `invoke` / `parameter` 为准，标签名前或标签名后的非结构性协议分隔符都会在解析入口剥离，例如 `<DSML␂tool_calls>`、`<proto💥tool_calls>` 这类控制符或非 ASCII 分隔符漂移也会归一化回现有 XML 标签后继续走同一套 parser；结构性字符如 `<` / `>` / `/` / `=` / 引号、空白和 ASCII 字母数字不会被当作这类分隔符。CDATA 开头也使用同一类扫描式容错，`<![CDATA[` / `<！[CDATA[` / `<、[CDATA[` 都会作为参数原文容器处理。但提示词会优先要求模型输出官方 DSML 标签，并强调不能只输出 closing wrapper 而漏掉 opening tag。需要注意：这是“兼容 DSML 外壳，内部仍以 XML 解析语义为准”，不是原生 DSML 全链路实现。解析器会先截获非代码块中的疑似工具 wrapper，完整解析失败或工具语义无效时再按普通文本放行。
 数组参数使用 `<item>...</item>` 子节点表示；当某个参数体只包含 item 子节点时，Go / Node 解析器会把它还原成数组，避免 `questions` / `options` 这类 schema 中要求 array 的参数被误解析成 `{ "item": ... }` 对象。除此之外，解析器还会回收一些更松散的列表写法，例如 JSON array 字面量或逗号分隔的 JSON 项序列，只要它们足够明确；但 `<item>` 仍然是首选形态。若模型把完整结构化 XML fragment 误包进 CDATA，兼容层会在保护 `content` / `command` 等原文字段的前提下，尝试把非原文字段中的 CDATA XML fragment 还原成 object / array。不过，如果 CDATA 只是单个平面的 XML/HTML 标签，例如 `<b>urgent</b>` 这种行内标记，兼容层会保留原始字符串，不会强行升成 object / array；只有明显表示结构的 CDATA 片段，例如多兄弟节点、嵌套子节点或 `item` 列表，才会触发结构化恢复。对 `command` / `content` 等长文本参数，CDATA 内部的 Markdown fenced DSML / XML 示例会作为原文保护；示例里的 `]]></parameter>` 或 `</tool_calls>` 不会截断外层工具调用，解析器会继续等待围栏外真正的参数 / wrapper 结束标签。
 Go 侧读取 DeepSeek SSE 时不再依赖 `bufio.Scanner` 的固定 2MiB 单行上限；当写文件类工具把很长的 `content` 放在单个 `data:` 行里返回时，非流式收集、流式解析和 auto-continue 透传都会保留完整行，再进入同一套工具解析与序列化流程。
 在 assistant 最终回包阶段，如果某个 tool 参数在声明 schema 中明确是 `string`，兼容层会在把解析后的 `tool_calls` / `function_call` 重新序列化成 OpenAI / Responses / Claude 可见参数前，递归把该路径上的 number / bool / object / array 统一转成字符串；其中 object / array 会压成紧凑 JSON 字符串。这个保护只对 schema 明确声明为 string 的路径生效，不会改写本来就是 `number` / `boolean` / `object` / `array` 的参数。这样可以兼容 DeepSeek 输出了结构化片段、但上游客户端工具 schema 又严格要求字符串参数的场景（例如 `content`、`prompt`、`path`、`taskId` 等）。
 工具 schema 的权威来源始终是**当前请求实际携带的 schema**，而不是同名工具在其他 runtime（Claude Code / OpenCode / Codex 等）里的默认印象。兼容层现在会同时兼容 OpenAI 风格 `function.parameters`、直接工具对象上的 `parameters` / `input_schema`、以及 camelCase 的 `inputSchema` / `schema`，并在最终输出阶段按这份请求内 schema 决定是保留 array/object，还是仅对明确声明为 `string` 的路径做字符串化。该规则同样适用于 Claude 的流式收尾和 Vercel Node 流式 tool-call formatter，避免不同 runtime 因 schema shape 差异而出现同名工具参数类型漂移。
 正例中的工具名只会来自当前请求实际声明的工具；如果当前请求没有足够的已知工具形态，就省略对应的单工具、多工具或嵌套示例，避免把不可用工具名写进 prompt。
 对执行类工具，脚本内容必须进入执行参数本身：`Bash` / `execute_command` 使用 `command`，`exec_command` 使用 `cmd`；不要把脚本示范成 `path` / `content` 文件写入参数。
+工具提示词也会明确要求模型按本次调用实际需要填写参数，禁止输出 placeholder、空字符串或纯空白参数；如果必填参数未知，应先追问用户或正常文字回复，而不是输出空工具壳。对 `Bash` / `execute_command` 这类 shell 工具，命令或脚本必须写入 `command` 参数。解析层仍会把空字符串参数结构化返回；是否拒绝空 `command` 由后续工具执行侧 / 客户端 schema 校验决定。
 如果当前请求声明了 `Read` / `read_file` 这类读取工具，兼容层会额外注入一条 read-tool cache guard：当读取结果只表示“文件未变更 / 已在历史中 / 请引用先前上下文 / 没有正文内容”时，模型必须把它视为内容不可用，不能反复调用同一个无正文读取；应改为请求完整正文读取能力，或向用户说明需要重新提供文件内容。这个约束只缓解客户端缓存返回空内容导致的死循环，DS2API 不会也无法凭空恢复客户端本地文件正文。
 
 OpenAI 路径实现：
@@ -205,18 +206,24 @@ assistant 的 reasoning 会变成一个显式标签块：
 
 然后再接可见回答正文。
 
+对最终返回给客户端的 assistant 轮次，reasoning 不会因为本轮输出了工具调用而被丢弃。OpenAI Chat 会在同一个 assistant message 上同时返回 `reasoning_content` 和 `tool_calls`；OpenAI Responses 会先返回一个包含 `reasoning` content 的 assistant message item，再返回后续 `function_call` item；Claude / Gemini 也会在各自原生 thinking / thought 结构后继续返回 tool_use / functionCall。
+
+对进入后续 prompt / `DS2API_HISTORY.txt` 的历史轮次，兼容层也会把同一轮工具调用前的 reasoning 绑定到 assistant tool call 历史上。OpenAI Chat 原生 `reasoning_content + tool_calls` 会直接保留；OpenAI Responses 若以 `reasoning` message item 后接 `function_call` item 的形式回放历史，会在归一化时合并为同一个 assistant 历史块；Claude 的 `thinking` block 会绑定到后续 `tool_use`；Gemini 的 `thought: true` part 会绑定到后续 `functionCall`。最终 prompt 中的顺序固定为 `[reasoning_content]...[/reasoning_content]`，再接 DSML tool call 外壳。
+
 ### 7.2 历史 tool_calls 保留方式
 
 assistant 历史 `tool_calls` 不会保留成 OpenAI 原生 JSON，而会转成 prompt 可见的 DSML 外壳：
 
 ```xml
-<|DSML|tool_calls>
-  <|DSML|invoke name="read_file">
-    <|DSML|parameter name="path"><![CDATA[src/main.go]]></|DSML|parameter>
-  </|DSML|invoke>
-</|DSML|tool_calls>
+<｜DSML｜tool_calls>
+  <｜DSML｜invoke name="read_file">
+    <｜DSML｜parameter name="path"><![CDATA[src/main.go]]></｜DSML｜parameter>
+  </｜DSML｜invoke>
+</｜DSML｜tool_calls>
 ```
 
+如果客户端历史里没有结构化 `tool_calls` 字段、却把一个可独立解析的 assistant 工具块放进了普通 `content`，兼容层会在写入后续 prompt 前先按工具调用解析它，再重渲染为规范 DSML 历史外壳。这样可以避免一次 malformed 工具块未被结构化保存后，作为普通 assistant 文本回灌，继续污染后续模型的 few-shot 工具格式。
+
 解析层同时兼容旧式纯 XML 形态：`<tool_calls>` / `<invoke>` / `<parameter>`。两者都会先归一到现有 XML 解析语义；其他旧格式都会作为普通文本保留，不会作为可执行调用语法。
 例外是 parser 会对一个非常窄的模型失误做修复：如果 assistant 输出了 `<invoke ...>` ... `</tool_calls>`（或 DSML 对应标签），但漏掉最前面的 opening wrapper，解析阶段会补回 wrapper 后再尝试识别。
 
@@ -419,7 +426,8 @@ Prior conversation history and tool progress.
 如果改的是 tool call 相关兼容语义，还应同时检查：
 
 - `go test ./internal/toolcall/...`
-- `node --test tests/node/stream-tool-sieve.test.js`
+- `go test ./internal/toolstream/...`
+- `./tests/scripts/run-unit-node.sh`
 
 ## 14. 文档同步约定
 

docs/toolcall-semantics.md

@@ -6,14 +6,14 @@
 
 ## 1) 当前可执行格式
 
-当前版本推荐模型输出 DSML 外壳：
+当前版本推荐模型输出全角分隔符 DSML 外壳：
 
 ```xml
-<|DSML|tool_calls>
-  <|DSML|invoke name="read_file">
-    <|DSML|parameter name="path"><![CDATA[README.MD]]></|DSML|parameter>
-  </|DSML|invoke>
-</|DSML|tool_calls>
+<｜DSML｜tool_calls>
+  <｜DSML｜invoke name="read_file">
+    <｜DSML｜parameter name="path"><![CDATA[README.MD]]></｜DSML｜parameter>
+  </｜DSML｜invoke>
+</｜DSML｜tool_calls>
 ```
 
 兼容层仍接受旧式 canonical XML：
@@ -30,17 +30,17 @@
 
 约束：
 
-- 必须有 `<|DSML|tool_calls>...</|DSML|tool_calls>` 或 `<tool_calls>...</tool_calls>` wrapper
-- 每个调用必须在 `<|DSML|invoke name="...">...</|DSML|invoke>` 或 `<invoke name="...">...</invoke>` 内
+- 必须有 `<｜DSML｜tool_calls>...</｜DSML｜tool_calls>` 或 `<tool_calls>...</tool_calls>` wrapper
+- 每个调用必须在 `<｜DSML｜invoke name="...">...</｜DSML｜invoke>` 或 `<invoke name="...">...</invoke>` 内
 - 工具名必须放在 `invoke` 的 `name` 属性
-- 参数必须使用 `<|DSML|parameter name="...">...</|DSML|parameter>` 或 `<parameter name="...">...</parameter>`
+- 参数必须使用 `<｜DSML｜parameter name="...">...</｜DSML｜parameter>` 或 `<parameter name="...">...</parameter>`
 - 同一个工具块内不要混用 DSML 标签和旧 XML 工具标签；混搭会被视为非法工具块
 
 兼容修复：
 
 - 如果模型漏掉 opening wrapper，但后面仍输出了一个或多个 invoke 并以 closing wrapper 收尾，Go 解析链路会在解析前补回缺失的 opening wrapper。
-- Go / Node 解析层不再枚举每一种 DSML typo。它会把工具标签名前的 `DSML`、管道符 `|` / `｜`、空白、重复 leading `<` 视为可容忍的协议噪声，然后只匹配固定本地标签名 `tool_calls` / `invoke` / `parameter`。例如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<<DSML|DSML|tool_calls>` 都会归一化；相似但非固定标签名（如 `tool_calls_extra`）仍按普通文本处理。
-- 如果模型在固定工具标签名后多输出一个尾部管道符，例如 `<|DSML|tool_calls|` / `<|DSML|invoke|` / `<|DSML|parameter|`，兼容层会把这个尾部 `|` 当作异常标签终止符并补齐缺失的 `>`；如果后面已经有 `>`，也会消费这个多余 `|` 后再归一化。
+- Go / Node 解析层不再枚举每一种 DSML typo。它以固定本地标签名 `tool_calls` / `invoke` / `parameter` 为准，把标签名前的任意协议前缀壳视为可容忍噪声，并继续兼容管道符 `|` / `｜`、全角感叹号 `！`、顿号 `、`、空白、重复 leading `<`、可视控制符 `␂`、原始 STX `\x02`、非 ASCII 分隔符、CJK 尖括号 `〈` / `〉`、弯引号属性值、PascalCase 本地名等漂移。例如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<DSmartToolCalls>`、`<<DSML|DSML|tool_calls>`、`<DSML␂tool_calls>`、`<proto💥tool_calls>`、`<DSM｜tool_calls>...〈/DSM｜tool_calls〉`、`<！DSML！tool_calls>...<！/DSML！tool_calls>`、`<、DSML、tool_calls>...<、/DSML、tool_calls>` 都会归一化；相似但非固定标签名（如 `tool_calls_extra` / `ToolCallsExtra`）仍按普通文本处理。
+- 如果模型在固定工具标签名后多输出一个非结构性分隔符，例如 `<|DSML|tool_calls|` / `<|DSML|invoke|` / `<|DSML|parameter|` / `<DSMLtool_calls※>`，或在带属性标签的结束符前多输出一个尾部分隔符（如 `<DSM｜parameter name="command"｜>`），兼容层会把这个尾部分隔符当作异常标签终止符并补齐或归一化；如果后面已经有 `>` / `〉`，也会消费这个多余分隔符后再归一化。结构性字符如 `<` / `>` / `/` / `=` / 引号、空白和 ASCII 字母数字不会被当作这类分隔符。
 - 这是一个针对常见模型失误的窄修复，不改变推荐输出格式；prompt 仍要求模型直接输出完整 DSML 外壳。
 - 裸 `<invoke ...>` / `<parameter ...>` 不会被当成“已支持的工具语法”；只有 `tool_calls` wrapper 或可修复的缺失 opening wrapper 才会进入工具调用路径。
 
@@ -54,13 +54,14 @@
 
 在流式链路中（Go / Node 一致）：
 
-- DSML `<|DSML|tool_calls>` wrapper、短横线形式（如 `<dsml-tool-calls>` / `<dsml-invoke>` / `<dsml-parameter>`）、基于固定本地标签名的 DSML 噪声容错形态、尾部管道符形态（如 `<|DSML|tool_calls|`）和 canonical `<tool_calls>` wrapper 都会进入结构化捕获
+- DSML `<｜DSML｜tool_calls>` wrapper、短横线形式（如 `<dsml-tool-calls>` / `<dsml-invoke>` / `<dsml-parameter>`）、基于固定本地标签名的 DSML 噪声容错形态、尾部非结构性分隔符形态（如 `<|DSML|tool_calls|` / `<DSMLtool_calls※>`）和 canonical `<tool_calls>` wrapper 都会进入结构化捕获
 - 如果流里直接从 invoke 开始，但后面补上了 closing wrapper，Go 流式筛分也会按缺失 opening wrapper 的修复路径尝试恢复
 - 已识别成功的工具调用不会再次回流到普通文本
 - 不符合新格式的块不会执行，并继续按原样文本透传
 - fenced code block（反引号 `` ``` `` 和波浪线 `~~~`）中的 XML 示例始终按普通文本处理
 - 支持嵌套围栏（如 4 反引号嵌套 3 反引号）和 CDATA 内围栏保护
 - 对 `command` / `content` 等长文本参数，CDATA 内部如果包含 Markdown fenced DSML / XML 示例，即使示例里出现 `]]></parameter>` / `</tool_calls>` 这类看起来像外层结束标签的片段，也会继续按参数原文保留，直到真正位于围栏外的外层结束标签
+- CDATA 开头也按扫描式识别，除了标准 `<![CDATA[`，还会接受 `<！[CDATA[`、`<、[CDATA[` 这类分隔符漂移，并统一还原为原文字段内容。
 - 如果模型把 `<![CDATA[` 打开后却没有闭合，流式扫描阶段仍会保守地继续缓冲，不会误把 CDATA 里的示例 XML 当成真实工具调用；在最终 parse / flush 恢复阶段，会对这类 loose CDATA 做窄修复，尽量保住外层已完整包裹的真实工具调用
 - 当文本中 mention 了某种标签名（如 `<dsml|tool_calls>` 或 Markdown inline code 里的 `<|DSML|tool_calls>`）而后面紧跟真正工具调用时，sieve 会跳过不可解析的 mention 候选并继续匹配后续真实工具块，不会因 mention 导致工具调用丢失，也不会截断 mention 后的正文
 - Go 侧 SSE 读取不再使用 `bufio.Scanner` 的固定 token 上限；单个 `data:` 行中包含很长的写文件参数时，非流式收集、流式解析与 auto-continue 透传都应保留完整行，再交给 tool parser 处理
@@ -78,11 +79,16 @@
 - `rejectedByPolicy`：当前固定为 `false`
 - `rejectedToolNames`：当前固定为空数组
 
+解析层不会因为参数值为空而丢弃工具调用。若模型输出了显式空字符串或纯空白参数，它们会按空字符串进入结构化 `tool_calls`；是否拒绝缺参或空命令应由后续工具执行侧 / 客户端 schema 校验决定。Prompt 层仍会要求模型不要主动输出空参数。
+
+完整的 DSML / XML wrapper 只有在成功解析出有效 `invoke name`，并且参数节点（如存在）符合 `parameter` 语义后，才会变成结构化工具调用；真正的零参数工具调用仍然有效。如果 wrapper 完整但内部不是可执行工具调用形态（例如使用 `<param>`、缺少有效 `invoke name`、或其他 malformed XML 工具壳），流式 sieve 会把原始 wrapper 作为普通文本释放，不会吞掉内容，也不会生成空的工具调用。
+
 ## 5) 落地建议
 
 1. Prompt 里只示范 DSML 外壳语法。
 2. 上游客户端应直接输出完整 DSML 外壳；DS2API 兼容旧式 canonical XML，并只对“closing tag 在、opening tag 漏掉”的常见失误做窄修复，不会泛化接受其他旧格式。
-3. 不要依赖 parser 做安全控制；执行器侧仍应做工具名和参数校验。
+3. 模型只有在知道本次调用所需参数值时才应输出工具调用；不要输出 placeholder、空字符串或纯空白参数。对 `Bash` / `execute_command`，实际命令必须在 `command` 参数里。
+4. 不要依赖 parser 做安全控制；执行器侧仍应做工具名和参数校验。
 
 ## 6) 回归验证
 
@@ -90,17 +96,18 @@
 
 ```bash
 go test -v -run 'TestParseToolCalls|TestProcessToolSieve' ./internal/toolcall ./internal/toolstream ./internal/httpapi/openai/...
-node --test tests/node/stream-tool-sieve.test.js
+./tests/scripts/run-unit-node.sh
 ```
 
 重点覆盖：
 
-- DSML `<|DSML|tool_calls>` wrapper 正常解析
+- DSML `<｜DSML｜tool_calls>` wrapper 正常解析
 - legacy canonical `<tool_calls>` wrapper 正常解析
-- 固定本地标签名的 DSML 噪声容错形态（如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<<DSML|DSML|tool_calls>`）正常解析
+- 固定本地标签名的 DSML 噪声容错形态（如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<DSmartToolCalls>`、`<<DSML|DSML|tool_calls>`、`<DSM｜tool_calls>...〈/DSM｜tool_calls〉`、`<！DSML！tool_calls>...<！/DSML！tool_calls>`）正常解析
 - 混搭标签（DSML wrapper + canonical inner）归一化后正常解析
 - 波浪线围栏 `~~~` 内的示例不执行
 - 嵌套围栏（4 反引号嵌套 3 反引号）内的示例不执行
 - 文本 mention 标签名后紧跟真正工具调用的场景（含同一 wrapper 变体）
+- 空参数结构化保留，malformed executable-looking XML wrapper 作为文本释放
 - 非兼容内容按普通文本透传
 - 代码块示例不执行

internal/assistantturn/turn.go

@@ -218,7 +218,7 @@ func UpstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int,
 	if strings.TrimSpace(thinking) != "" {
 		return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned reasoning without visible output.", "upstream_empty_output"
 	}
-	return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned empty output.", "upstream_empty_output"
+	return http.StatusServiceUnavailable, "Upstream service is unavailable and returned no output.", "upstream_unavailable"
 }
 
 // ShouldRetryEmptyOutput returns true when the turn produced no visible text

internal/assistantturn/turn_test.go

@@ -1,6 +1,7 @@
 package assistantturn
 
 import (
+	"net/http"
 	"testing"
 
 	"ds2api/internal/promptcompat"
@@ -70,6 +71,13 @@ func TestBuildTurnFromCollectedThinkingOnlyIsEmptyOutput(t *testing.T) {
 	}
 }
 
+func TestBuildTurnFromCollectedPureEmptyOutputIsUpstreamUnavailable(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{}, BuildOptions{})
+	if turn.Error == nil || turn.Error.Status != http.StatusServiceUnavailable || turn.Error.Code != "upstream_unavailable" {
+		t.Fatalf("expected upstream unavailable error, got %#v", turn.Error)
+	}
+}
+
 func TestBuildTurnFromCollectedToolChoiceRequired(t *testing.T) {
 	turn := BuildTurnFromCollected(sse.CollectResult{Text: "hello"}, BuildOptions{
 		ToolChoice: promptcompat.ToolChoicePolicy{Mode: promptcompat.ToolChoiceRequired},

internal/auth/auth_edge_test.go

@@ -241,6 +241,36 @@ func TestSwitchAccountSkipsLoginFailureAndContinues(t *testing.T) {
 	}
 }
 
+func TestSwitchAccountRespectsPinnedTargetAccount(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["managed-key"],
+		"accounts":[
+			{"email":"acc1@test.com","token":"t1"},
+			{"email":"acc2@test.com","token":"t2"}
+		]
+	}`)
+	store := config.LoadStore()
+	pool := account.NewPool(store)
+	r := NewResolver(store, pool, func(_ context.Context, _ config.Account) (string, error) {
+		return "new-token", nil
+	})
+
+	req, _ := http.NewRequest("POST", "/", nil)
+	req.Header.Set("Authorization", "Bearer managed-key")
+	req.Header.Set("X-Ds2-Target-Account", "acc1@test.com")
+	a, err := r.Determine(req)
+	if err != nil {
+		t.Fatalf("determine failed: %v", err)
+	}
+	defer r.Release(a)
+	if r.SwitchAccount(context.Background(), a) {
+		t.Fatal("expected switch to be disabled for pinned target account")
+	}
+	if a.AccountID != "acc1@test.com" {
+		t.Fatalf("expected pinned account to remain selected, got %q", a.AccountID)
+	}
+}
+
 // ─── Release edge cases ─────────────────────────────────────────────
 
 func TestReleaseNilAuth(t *testing.T) {

internal/auth/request.go

@@ -28,6 +28,7 @@ type RequestAuth struct {
 	DeepSeekToken  string
 	CallerID       string
 	AccountID      string
+	TargetAccount  string
 	Account        config.Account
 	TriedAccounts  map[string]bool
 	resolver       *Resolver
@@ -99,6 +100,7 @@ func (r *Resolver) acquireManagedRequestAuth(ctx context.Context, callerID, targ
 			UseConfigToken: true,
 			CallerID:       callerID,
 			AccountID:      acc.Identifier(),
+			TargetAccount:  target,
 			Account:        acc,
 			TriedAccounts:  tried,
 			resolver:       r,
@@ -185,6 +187,9 @@ func (r *Resolver) SwitchAccount(ctx context.Context, a *RequestAuth) bool {
 	if !a.UseConfigToken {
 		return false
 	}
+	if strings.TrimSpace(a.TargetAccount) != "" {
+		return false
+	}
 	if a.TriedAccounts == nil {
 		a.TriedAccounts = map[string]bool{}
 	}
@@ -208,6 +213,13 @@ func (r *Resolver) SwitchAccount(ctx context.Context, a *RequestAuth) bool {
 	}
 }
 
+func (a *RequestAuth) SwitchAccount(ctx context.Context) bool {
+	if a == nil || a.resolver == nil {
+		return false
+	}
+	return a.resolver.SwitchAccount(ctx, a)
+}
+
 func (r *Resolver) Release(a *RequestAuth) {
 	if a == nil || !a.UseConfigToken || a.AccountID == "" {
 		return

internal/completionruntime/nonstream.go

@@ -90,7 +90,11 @@ func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.R
 	if startErr != nil {
 		return NonStreamResult{SessionID: start.SessionID, Payload: start.Payload}, startErr
 	}
-	stdReq = start.Request
+	return ExecuteNonStreamStartedWithRetry(ctx, ds, a, start, opts)
+}
+
+func ExecuteNonStreamStartedWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, start StartResult, opts Options) (NonStreamResult, *assistantturn.OutputError) {
+	stdReq := start.Request
 	maxAttempts := opts.MaxAttempts
 	if maxAttempts <= 0 {
 		maxAttempts = 3
@@ -100,6 +104,7 @@ func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.R
 	pow := start.Pow
 
 	attempts := 0
+	accountSwitchAttempted := false
 	currentResp := start.Response
 	usagePrompt := stdReq.PromptTokenText
 	accumulatedThinking := ""
@@ -108,6 +113,24 @@ func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.R
 	for {
 		turn, outErr := collectAttempt(currentResp, stdReq, usagePrompt, opts)
 		if outErr != nil {
+			if canRetryOnAlternateAccount(ctx, a, outErr, opts.RetryEnabled, &accountSwitchAttempted) {
+				switched, switchErr := startStandardCompletionOnAlternateAccount(ctx, ds, a, stdReq, maxAttempts)
+				if switchErr != nil {
+					return NonStreamResult{SessionID: sessionID, Payload: payload, Attempts: attempts}, switchErr
+				}
+				if switched.Response != nil {
+					config.Logger.Info("[completion_runtime_account_switch_retry] retrying after 429", "surface", stdReq.Surface, "stream", false, "account", a.AccountID)
+					sessionID = switched.SessionID
+					payload = switched.Payload
+					pow = switched.Pow
+					currentResp = switched.Response
+					usagePrompt = stdReq.PromptTokenText
+					accumulatedThinking = ""
+					accumulatedRawThinking = ""
+					accumulatedToolDetectionThinking = ""
+					continue
+				}
+			}
 			return NonStreamResult{SessionID: sessionID, Payload: payload, Attempts: attempts}, outErr
 		}
 		accumulatedThinking += sse.TrimContinuationOverlap(accumulatedThinking, turn.Thinking)
@@ -130,6 +153,24 @@ func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.R
 			retryMax = shared.EmptyOutputRetryMaxAttempts()
 		}
 		if !opts.RetryEnabled || !assistantturn.ShouldRetryEmptyOutput(turn, attempts, retryMax) {
+			if canRetryOnAlternateAccount(ctx, a, turn.Error, opts.RetryEnabled, &accountSwitchAttempted) {
+				switched, switchErr := startStandardCompletionOnAlternateAccount(ctx, ds, a, stdReq, maxAttempts)
+				if switchErr != nil {
+					return NonStreamResult{SessionID: sessionID, Payload: payload, Turn: turn, Attempts: attempts}, switchErr
+				}
+				if switched.Response != nil {
+					config.Logger.Info("[completion_runtime_account_switch_retry] retrying after 429", "surface", stdReq.Surface, "stream", false, "account", a.AccountID)
+					sessionID = switched.SessionID
+					payload = switched.Payload
+					pow = switched.Pow
+					currentResp = switched.Response
+					usagePrompt = stdReq.PromptTokenText
+					accumulatedThinking = ""
+					accumulatedRawThinking = ""
+					accumulatedToolDetectionThinking = ""
+					continue
+				}
+			}
 			return NonStreamResult{SessionID: sessionID, Payload: payload, Turn: turn, Attempts: attempts}, turn.Error
 		}
 
@@ -150,6 +191,37 @@ func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.R
 	}
 }
 
+func canRetryOnAlternateAccount(ctx context.Context, a *auth.RequestAuth, outErr *assistantturn.OutputError, retryEnabled bool, attempted *bool) bool {
+	if outErr == nil || outErr.Status != http.StatusTooManyRequests {
+		return false
+	}
+	if !retryEnabled || attempted == nil || *attempted {
+		return false
+	}
+	if a == nil || !a.UseConfigToken {
+		return false
+	}
+	*attempted = true
+	return a.SwitchAccount(ctx)
+}
+
+func startStandardCompletionOnAlternateAccount(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, stdReq promptcompat.StandardRequest, maxAttempts int) (StartResult, *assistantturn.OutputError) {
+	sessionID, err := ds.CreateSession(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{}, authOutputError(a)
+	}
+	pow, err := ds.GetPow(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID}, &assistantturn.OutputError{Status: http.StatusUnauthorized, Message: "Failed to get PoW (invalid token or unknown error).", Code: "error"}
+	}
+	payload := stdReq.CompletionPayload(sessionID)
+	resp, err := ds.CallCompletion(ctx, a, payload, pow, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID, Payload: payload, Pow: pow}, &assistantturn.OutputError{Status: http.StatusInternalServerError, Message: "Failed to get completion.", Code: "error"}
+	}
+	return StartResult{SessionID: sessionID, Payload: payload, Pow: pow, Response: resp, Request: stdReq}, nil
+}
+
 func collectAttempt(resp *http.Response, stdReq promptcompat.StandardRequest, usagePrompt string, opts Options) (assistantturn.Turn, *assistantturn.OutputError) {
 	defer func() {
 		if err := resp.Body.Close(); err != nil {

internal/completionruntime/nonstream_test.go

@@ -7,15 +7,19 @@ import (
 	"strings"
 	"testing"
 
+	"ds2api/internal/account"
 	"ds2api/internal/auth"
+	"ds2api/internal/config"
 	dsclient "ds2api/internal/deepseek/client"
 	"ds2api/internal/promptcompat"
 )
 
 type fakeDeepSeekCaller struct {
-	responses []*http.Response
-	payloads  []map[string]any
-	uploads   []dsclient.UploadFileRequest
+	responses          []*http.Response
+	payloads           []map[string]any
+	uploads            []dsclient.UploadFileRequest
+	completionAccounts []string
+	sessionByAccount   bool
 }
 
 type currentInputRuntimeConfig struct{}
@@ -23,7 +27,10 @@ type currentInputRuntimeConfig struct{}
 func (currentInputRuntimeConfig) CurrentInputFileEnabled() bool { return true }
 func (currentInputRuntimeConfig) CurrentInputFileMinChars() int { return 0 }
 
-func (f *fakeDeepSeekCaller) CreateSession(context.Context, *auth.RequestAuth, int) (string, error) {
+func (f *fakeDeepSeekCaller) CreateSession(_ context.Context, a *auth.RequestAuth, _ int) (string, error) {
+	if f.sessionByAccount && a != nil && a.AccountID != "" {
+		return "session-" + a.AccountID, nil
+	}
 	return "session-1", nil
 }
 
@@ -36,8 +43,11 @@ func (f *fakeDeepSeekCaller) UploadFile(_ context.Context, _ *auth.RequestAuth,
 	return &dsclient.UploadFileResult{ID: "file-runtime-1"}, nil
 }
 
-func (f *fakeDeepSeekCaller) CallCompletion(_ context.Context, _ *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
+func (f *fakeDeepSeekCaller) CallCompletion(_ context.Context, a *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
 	f.payloads = append(f.payloads, payload)
+	if a != nil {
+		f.completionAccounts = append(f.completionAccounts, a.AccountID)
+	}
 	if len(f.responses) == 0 {
 		return sseHTTPResponse(http.StatusOK, `data: {"p":"response/content","v":"fallback"}`), nil
 	}
@@ -89,9 +99,72 @@ func TestExecuteNonStreamWithRetryBuildsCanonicalTurn(t *testing.T) {
 	}
 }
 
+func TestExecuteNonStreamWithRetrySwitchesManagedAccountBeforeFinal429(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["managed-key"],
+		"accounts":[
+			{"email":"acc1@test.com","password":"pwd"},
+			{"email":"acc2@test.com","password":"pwd"}
+		]
+	}`)
+	store := config.LoadStore()
+	resolver := auth.NewResolver(store, account.NewPool(store), func(_ context.Context, acc config.Account) (string, error) {
+		return "token-" + acc.Identifier(), nil
+	})
+	req, _ := http.NewRequest(http.MethodPost, "/", nil)
+	req.Header.Set("Authorization", "Bearer managed-key")
+	a, err := resolver.Determine(req)
+	if err != nil {
+		t.Fatalf("determine failed: %v", err)
+	}
+	defer resolver.Release(a)
+
+	ds := &fakeDeepSeekCaller{
+		sessionByAccount: true,
+		responses: []*http.Response{
+			sseHTTPResponse(http.StatusOK, `data: {"response_message_id":11,"p":"response/thinking_content","v":"first empty"}`),
+			sseHTTPResponse(http.StatusOK, `data: {"response_message_id":12,"p":"response/thinking_content","v":"retry empty"}`),
+			sseHTTPResponse(http.StatusOK, `data: {"response_message_id":21,"p":"response/content","v":"ok from second account"}`),
+		},
+	}
+	stdReq := promptcompat.StandardRequest{
+		Surface:         "test",
+		ResponseModel:   "deepseek-v4-flash",
+		PromptTokenText: "prompt",
+		FinalPrompt:     "final prompt",
+		Thinking:        true,
+	}
+
+	result, outErr := ExecuteNonStreamWithRetry(context.Background(), ds, a, stdReq, Options{RetryEnabled: true})
+	if outErr != nil {
+		t.Fatalf("unexpected output error after account switch retry: %#v", outErr)
+	}
+	if result.Turn.Text != "ok from second account" {
+		t.Fatalf("text mismatch after switch retry: %q", result.Turn.Text)
+	}
+	if result.SessionID != "session-acc2@test.com" {
+		t.Fatalf("expected switched account session, got %q", result.SessionID)
+	}
+	wantAccounts := []string{"acc1@test.com", "acc1@test.com", "acc2@test.com"}
+	if len(ds.completionAccounts) != len(wantAccounts) {
+		t.Fatalf("completion account count mismatch: got %v want %v", ds.completionAccounts, wantAccounts)
+	}
+	for i, want := range wantAccounts {
+		if ds.completionAccounts[i] != want {
+			t.Fatalf("completion account %d = %q want %q (all=%v)", i, ds.completionAccounts[i], want, ds.completionAccounts)
+		}
+	}
+	if got := ds.payloads[2]["chat_session_id"]; got != "session-acc2@test.com" {
+		t.Fatalf("switched payload session mismatch: %#v", got)
+	}
+	if prompt, _ := ds.payloads[2]["prompt"].(string); strings.Contains(prompt, "Previous reply had no visible output") {
+		t.Fatalf("expected fresh switched-account prompt without empty-output suffix, got %q", prompt)
+	}
+}
+
 func TestExecuteNonStreamWithRetryUsesParentMessageForEmptyRetry(t *testing.T) {
 	ds := &fakeDeepSeekCaller{responses: []*http.Response{
-		sseHTTPResponse(http.StatusOK, `data: {"response_message_id":77,"p":"response/status","v":"FINISHED"}`),
+		sseHTTPResponse(http.StatusOK, `data: {"response_message_id":77,"p":"response/thinking_content","v":"plan"}`),
 		sseHTTPResponse(http.StatusOK, `data: {"response_message_id":78,"p":"response/content","v":"ok"}`),
 	}}
 	stdReq := promptcompat.StandardRequest{

internal/completionruntime/stream_retry.go

@@ -0,0 +1,179 @@
+package completionruntime
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+
+	"ds2api/internal/assistantturn"
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+	"ds2api/internal/httpapi/openai/shared"
+)
+
+type StreamRetryOptions struct {
+	Surface          string
+	Stream           bool
+	RetryEnabled     bool
+	RetryMaxAttempts int
+	MaxAttempts      int
+	UsagePrompt      string
+}
+
+type StreamRetryHooks struct {
+	ConsumeAttempt  func(resp *http.Response, allowDeferEmpty bool) (terminalWritten bool, retryable bool)
+	Finalize        func(attempts int)
+	ParentMessageID func() int
+	OnRetry         func(attempts int)
+	OnRetryPrompt   func(prompt string)
+	OnRetryFailure  func(status int, message, code string)
+	OnAccountSwitch func(sessionID string)
+	OnTerminal      func(attempts int)
+}
+
+func ExecuteStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, initialResp *http.Response, payload map[string]any, pow string, opts StreamRetryOptions, hooks StreamRetryHooks) {
+	if hooks.ConsumeAttempt == nil {
+		return
+	}
+	surface := strings.TrimSpace(opts.Surface)
+	if surface == "" {
+		surface = "completion"
+	}
+	maxAttempts := opts.MaxAttempts
+	if maxAttempts <= 0 {
+		maxAttempts = 3
+	}
+	retryMax := opts.RetryMaxAttempts
+	if retryMax <= 0 {
+		retryMax = shared.EmptyOutputRetryMaxAttempts()
+	}
+
+	attempts := 0
+	accountSwitchAttempted := false
+	currentResp := initialResp
+	currentPayload := clonePayload(payload)
+	for {
+		allowAccountSwitch := opts.RetryEnabled && attempts >= retryMax && !accountSwitchAttempted && a != nil && a.UseConfigToken
+		terminalWritten, retryable := hooks.ConsumeAttempt(currentResp, opts.RetryEnabled && (attempts < retryMax || allowAccountSwitch))
+		if terminalWritten {
+			if hooks.OnTerminal != nil {
+				hooks.OnTerminal(attempts)
+			}
+			return
+		}
+		if !retryable || !opts.RetryEnabled {
+			if hooks.Finalize != nil {
+				hooks.Finalize(attempts)
+			}
+			return
+		}
+
+		if attempts >= retryMax {
+			if canRetryOnAlternateAccount(ctx, a, &assistantturn.OutputError{Status: http.StatusTooManyRequests}, opts.RetryEnabled, &accountSwitchAttempted) {
+				switched, switchErr := startPayloadCompletionOnAlternateAccount(ctx, ds, a, payload, maxAttempts)
+				if switchErr != nil {
+					if hooks.OnRetryFailure != nil {
+						hooks.OnRetryFailure(switchErr.Status, switchErr.Message, switchErr.Code)
+					}
+					return
+				}
+				if switched.Response != nil {
+					config.Logger.Info("[completion_runtime_account_switch_retry] retrying after 429", "surface", surface, "stream", opts.Stream, "account", a.AccountID)
+					currentResp = switched.Response
+					currentPayload = switched.Payload
+					pow = switched.Pow
+					if hooks.OnAccountSwitch != nil {
+						hooks.OnAccountSwitch(switched.SessionID)
+					}
+					if hooks.OnRetryPrompt != nil {
+						hooks.OnRetryPrompt(opts.UsagePrompt)
+					}
+					continue
+				}
+			}
+			if hooks.Finalize != nil {
+				hooks.Finalize(attempts)
+			}
+			return
+		}
+
+		attempts++
+		parentMessageID := 0
+		if hooks.ParentMessageID != nil {
+			parentMessageID = hooks.ParentMessageID()
+		}
+		config.Logger.Info("[completion_runtime_empty_retry] attempting synthetic retry", "surface", surface, "stream", opts.Stream, "retry_attempt", attempts, "parent_message_id", parentMessageID)
+		retryPow, powErr := ds.GetPow(ctx, a, maxAttempts)
+		if powErr != nil {
+			config.Logger.Warn("[completion_runtime_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", surface, "stream", opts.Stream, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		nextResp, err := ds.CallCompletion(ctx, a, shared.ClonePayloadForEmptyOutputRetry(currentPayload, parentMessageID), retryPow, maxAttempts)
+		if err != nil {
+			if hooks.OnRetryFailure != nil {
+				hooks.OnRetryFailure(http.StatusInternalServerError, "Failed to get completion.", "error")
+			}
+			config.Logger.Warn("[completion_runtime_empty_retry] retry request failed", "surface", surface, "stream", opts.Stream, "retry_attempt", attempts, "error", err)
+			return
+		}
+		if nextResp.StatusCode != http.StatusOK {
+			body, readErr := io.ReadAll(nextResp.Body)
+			if readErr != nil {
+				config.Logger.Warn("[completion_runtime_empty_retry] retry error body read failed", "surface", surface, "stream", opts.Stream, "retry_attempt", attempts, "error", readErr)
+			}
+			closeRetryBody(surface, nextResp.Body)
+			msg := strings.TrimSpace(string(body))
+			if msg == "" {
+				msg = http.StatusText(nextResp.StatusCode)
+			}
+			if hooks.OnRetryFailure != nil {
+				hooks.OnRetryFailure(nextResp.StatusCode, msg, "error")
+			}
+			return
+		}
+		if hooks.OnRetry != nil {
+			hooks.OnRetry(attempts)
+		}
+		if hooks.OnRetryPrompt != nil {
+			hooks.OnRetryPrompt(shared.UsagePromptWithEmptyOutputRetry(opts.UsagePrompt, attempts))
+		}
+		currentResp = nextResp
+	}
+}
+
+func startPayloadCompletionOnAlternateAccount(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, payload map[string]any, maxAttempts int) (StartResult, *assistantturn.OutputError) {
+	sessionID, err := ds.CreateSession(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{}, authOutputError(a)
+	}
+	pow, err := ds.GetPow(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID}, &assistantturn.OutputError{Status: http.StatusUnauthorized, Message: "Failed to get PoW (invalid token or unknown error).", Code: "error"}
+	}
+	nextPayload := clonePayload(payload)
+	nextPayload["chat_session_id"] = sessionID
+	delete(nextPayload, "parent_message_id")
+	resp, err := ds.CallCompletion(ctx, a, nextPayload, pow, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID, Payload: nextPayload, Pow: pow}, &assistantturn.OutputError{Status: http.StatusInternalServerError, Message: "Failed to get completion.", Code: "error"}
+	}
+	return StartResult{SessionID: sessionID, Payload: nextPayload, Pow: pow, Response: resp}, nil
+}
+
+func clonePayload(payload map[string]any) map[string]any {
+	clone := make(map[string]any, len(payload))
+	for k, v := range payload {
+		clone[k] = v
+	}
+	return clone
+}
+
+func closeRetryBody(surface string, body io.Closer) {
+	if body == nil {
+		return
+	}
+	if err := body.Close(); err != nil {
+		config.Logger.Warn("[completion_runtime_empty_retry] retry response body close failed", "surface", surface, "error", err)
+	}
+}

internal/completionruntime/stream_retry_test.go

@@ -0,0 +1,150 @@
+package completionruntime
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"ds2api/internal/account"
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+	"ds2api/internal/httpapi/openai/shared"
+)
+
+func TestExecuteStreamWithRetryUsesSharedRetryPayloadAndUsagePrompt(t *testing.T) {
+	ds := &fakeDeepSeekCaller{responses: []*http.Response{
+		sseHTTPResponse(http.StatusOK, `data: {"p":"response/content","v":"ok"}`),
+	}}
+	initial := sseHTTPResponse(http.StatusOK, `data: {"response_message_id":77,"p":"response/thinking_content","v":"plan"}`)
+	payload := map[string]any{"prompt": "original prompt"}
+	attemptsSeen := 0
+	retryPrompt := ""
+
+	ExecuteStreamWithRetry(context.Background(), ds, &auth.RequestAuth{}, initial, payload, "pow", StreamRetryOptions{
+		Surface:      "test.stream",
+		Stream:       true,
+		RetryEnabled: true,
+		UsagePrompt:  "original prompt",
+	}, StreamRetryHooks{
+		ConsumeAttempt: func(resp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			defer func() {
+				if err := resp.Body.Close(); err != nil {
+					t.Fatalf("close failed: %v", err)
+				}
+			}()
+			_, _ = io.ReadAll(resp.Body)
+			attemptsSeen++
+			return attemptsSeen == 2, attemptsSeen == 1 && allowDeferEmpty
+		},
+		ParentMessageID: func() int {
+			return 77
+		},
+		OnRetryPrompt: func(prompt string) {
+			retryPrompt = prompt
+		},
+	})
+
+	if attemptsSeen != 2 {
+		t.Fatalf("expected two stream attempts, got %d", attemptsSeen)
+	}
+	if len(ds.payloads) != 1 {
+		t.Fatalf("expected one retry completion call, got %d", len(ds.payloads))
+	}
+	if got := ds.payloads[0]["parent_message_id"]; got != 77 {
+		t.Fatalf("retry parent_message_id mismatch: %#v", got)
+	}
+	if prompt, _ := ds.payloads[0]["prompt"].(string); !strings.Contains(prompt, shared.EmptyOutputRetrySuffix) {
+		t.Fatalf("expected retry suffix in payload prompt, got %q", prompt)
+	}
+	if !strings.Contains(retryPrompt, shared.EmptyOutputRetrySuffix) {
+		t.Fatalf("expected retry suffix in usage prompt, got %q", retryPrompt)
+	}
+}
+
+func TestExecuteStreamWithRetrySwitchesManagedAccountBeforeFinal429(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["managed-key"],
+		"accounts":[
+			{"email":"acc1@test.com","password":"pwd"},
+			{"email":"acc2@test.com","password":"pwd"}
+		]
+	}`)
+	store := config.LoadStore()
+	resolver := auth.NewResolver(store, account.NewPool(store), func(_ context.Context, acc config.Account) (string, error) {
+		return "token-" + acc.Identifier(), nil
+	})
+	req, _ := http.NewRequest(http.MethodPost, "/", nil)
+	req.Header.Set("Authorization", "Bearer managed-key")
+	a, err := resolver.Determine(req)
+	if err != nil {
+		t.Fatalf("determine failed: %v", err)
+	}
+	defer resolver.Release(a)
+
+	ds := &fakeDeepSeekCaller{
+		sessionByAccount: true,
+		responses: []*http.Response{
+			sseHTTPResponse(http.StatusOK, `data: {"response_message_id":12,"p":"response/thinking_content","v":"retry empty"}`),
+			sseHTTPResponse(http.StatusOK, `data: {"response_message_id":21,"p":"response/content","v":"ok from second account"}`),
+		},
+	}
+	initial := sseHTTPResponse(http.StatusOK, `data: {"response_message_id":11,"p":"response/thinking_content","v":"first empty"}`)
+	payload := map[string]any{"prompt": "original prompt", "chat_session_id": "session-acc1@test.com"}
+	attemptsSeen := 0
+	switchedSession := ""
+
+	ExecuteStreamWithRetry(context.Background(), ds, a, initial, payload, "pow", StreamRetryOptions{
+		Surface:          "test.stream",
+		Stream:           true,
+		RetryEnabled:     true,
+		RetryMaxAttempts: 1,
+		UsagePrompt:      "original prompt",
+	}, StreamRetryHooks{
+		ConsumeAttempt: func(resp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			defer func() {
+				if err := resp.Body.Close(); err != nil {
+					t.Fatalf("close failed: %v", err)
+				}
+			}()
+			body, _ := io.ReadAll(resp.Body)
+			attemptsSeen++
+			if strings.Contains(string(body), "ok from second account") {
+				return true, false
+			}
+			if !allowDeferEmpty {
+				t.Fatalf("expected empty attempt %d to be deferred before final 429", attemptsSeen)
+			}
+			return false, true
+		},
+		ParentMessageID: func() int {
+			return 11 + attemptsSeen
+		},
+		OnAccountSwitch: func(sessionID string) {
+			switchedSession = sessionID
+		},
+	})
+
+	if attemptsSeen != 3 {
+		t.Fatalf("expected three stream attempts, got %d", attemptsSeen)
+	}
+	if switchedSession != "session-acc2@test.com" {
+		t.Fatalf("expected switched session id, got %q", switchedSession)
+	}
+	wantAccounts := []string{"acc1@test.com", "acc2@test.com"}
+	if len(ds.completionAccounts) != len(wantAccounts) {
+		t.Fatalf("completion accounts mismatch: got %v want %v", ds.completionAccounts, wantAccounts)
+	}
+	for i, want := range wantAccounts {
+		if ds.completionAccounts[i] != want {
+			t.Fatalf("completion account %d = %q want %q (all=%v)", i, ds.completionAccounts[i], want, ds.completionAccounts)
+		}
+	}
+	if got := ds.payloads[1]["chat_session_id"]; got != "session-acc2@test.com" {
+		t.Fatalf("switched payload session mismatch: %#v", got)
+	}
+	if prompt, _ := ds.payloads[1]["prompt"].(string); strings.Contains(prompt, shared.EmptyOutputRetrySuffix) {
+		t.Fatalf("expected switched-account prompt without empty-output suffix, got %q", prompt)
+	}
+}

internal/config/models.go

@@ -1,6 +1,9 @@
 package config
 
-import "strings"
+import (
+	"strings"
+	"time"
+)
 
 type ModelInfo struct {
 	ID         string `json:"id"`
@@ -9,6 +12,16 @@ type ModelInfo struct {
 	OwnedBy    string `json:"owned_by"`
 	Permission []any  `json:"permission,omitempty"`
 }
+type OllamaModelInfo struct {
+	Name       string `json:"name"`
+	Model      string `json:"model"`
+	Size       int64  `json:"size"`
+	ModifiedAt string `json:"modified_at"`
+}
+type OllamaCapabilitiesModelInfo struct {
+	ID           string   `json:"id"`
+	Capabilities []string `json:"capabilities"`
+}
 
 type ModelAliasReader interface {
 	ModelAliases() map[string]string
@@ -24,8 +37,21 @@ var deepSeekBaseModels = []ModelInfo{
 	{ID: "deepseek-v4-vision", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 }
 
-var DeepSeekModels = appendNoThinkingVariants(deepSeekBaseModels)
+var OllamaCapabilitiesModels = []OllamaCapabilitiesModelInfo{
+	{ID: "deepseek-v4-flash", Capabilities: []string{"tools", "thinking"}},
+	{ID: "deepseek-v4-pro", Capabilities: []string{"tools", "thinking"}},
+	{ID: "deepseek-v4-flash-search", Capabilities: []string{"tools", "thinking"}},
+	{ID: "deepseek-v4-pro-search", Capabilities: []string{"tools", "thinking"}},
+	{ID: "deepseek-v4-vision", Capabilities: []string{"tools", "thinking", "vision"}},
+	{ID: "deepseek-v4-flash-nothinking", Capabilities: []string{"tools"}},
+	{ID: "deepseek-v4-pro-nothinking", Capabilities: []string{"tools"}},
+	{ID: "deepseek-v4-flash-search-nothinking", Capabilities: []string{"tools"}},
+	{ID: "deepseek-v4-pro-search-nothinking", Capabilities: []string{"tools"}},
+	{ID: "deepseek-v4-vision-nothinking", Capabilities: []string{"tools", "vision"}},
+}
 
+var DeepSeekModels = appendNoThinkingVariants(deepSeekBaseModels)
+var OllamaModels = mapToOllamaModels(DeepSeekModels)
 var claudeBaseModels = []ModelInfo{
 	// Current aliases
 	{ID: "claude-opus-4-6", Object: "model", Created: 1715635200, OwnedBy: "anthropic"},
@@ -247,6 +273,23 @@ func OpenAIModelByID(store ModelAliasReader, id string) (ModelInfo, bool) {
 	return ModelInfo{}, false
 }
 
+func OllamaModelsResponse() map[string]any {
+	return map[string]any{"models": OllamaModels}
+}
+
+func OllamaModelByID(store ModelAliasReader, id string) (OllamaCapabilitiesModelInfo, bool) {
+	canonical, ok := ResolveModel(store, id)
+	if !ok {
+		return OllamaCapabilitiesModelInfo{}, false
+	}
+	for _, model := range OllamaCapabilitiesModels {
+		if model.ID == canonical {
+			return model, true
+		}
+	}
+	return OllamaCapabilitiesModelInfo{}, false
+}
+
 func ClaudeModelsResponse() map[string]any {
 	resp := map[string]any{"object": "list", "data": ClaudeModels}
 	if len(ClaudeModels) > 0 {
@@ -270,6 +313,23 @@ func appendNoThinkingVariants(models []ModelInfo) []ModelInfo {
 	}
 	return out
 }
+func mapToOllamaModels(models []ModelInfo) []OllamaModelInfo {
+	out := make([]OllamaModelInfo, 0, len(models))
+	for _, model := range models {
+		var modifiedAt string
+		if model.Created > 0 {
+			modifiedAt = time.Unix(model.Created, 0).Format(time.RFC3339)
+		}
+		ollamaModel := OllamaModelInfo{
+			Name:       model.ID,
+			Model:      model.ID,
+			Size:       0,
+			ModifiedAt: modifiedAt,
+		}
+		out = append(out, ollamaModel)
+	}
+	return out
+}
 
 func splitNoThinkingModel(model string) (string, bool) {
 	model = lower(strings.TrimSpace(model))

internal/config/paths.go

@@ -58,6 +58,11 @@ func RawStreamSampleRoot() string {
 }
 
 func ChatHistoryPath() string {
+	// On Vercel, /var/task is read-only at runtime. If no explicit path is set,
+	// default to /tmp/chat_history.json (the only writable directory).
+	if IsVercel() && strings.TrimSpace(os.Getenv("DS2API_CHAT_HISTORY_PATH")) == "" {
+		return "/tmp/chat_history.json"
+	}
 	return ResolvePath("DS2API_CHAT_HISTORY_PATH", "data/chat_history.json")
 }
 

internal/format/openai/render_responses.go

@@ -21,6 +21,18 @@ func BuildResponseObjectWithToolCalls(responseID, model, finalPrompt, finalThink
 	output := make([]any, 0, 2)
 	if len(detected) > 0 {
 		exposedOutputText = ""
+		if strings.TrimSpace(finalThinking) != "" {
+			output = append(output, map[string]any{
+				"type":   "message",
+				"id":     "msg_" + strings.ReplaceAll(uuid.NewString(), "-", ""),
+				"role":   "assistant",
+				"status": "completed",
+				"content": []any{map[string]any{
+					"type": "reasoning",
+					"text": finalThinking,
+				}},
+			})
+		}
 		output = append(output, toResponsesFunctionCallItems(detected, toolsRaw)...)
 	} else {
 		content := make([]any, 0, 2)

internal/format/openai/render_test.go

@@ -85,12 +85,24 @@ func TestBuildResponseObjectPromotesToolCallFromThinkingWhenTextEmpty(t *testing
 	)
 
 	output, _ := obj["output"].([]any)
-	if len(output) != 1 {
-		t.Fatalf("expected one output item, got %#v", obj["output"])
+	if len(output) != 2 {
+		t.Fatalf("expected reasoning message plus function_call output, got %#v", obj["output"])
 	}
 	first, _ := output[0].(map[string]any)
-	if first["type"] != "function_call" {
-		t.Fatalf("expected function_call output, got %#v", first["type"])
+	if first["type"] != "message" {
+		t.Fatalf("expected reasoning message output first, got %#v", first["type"])
+	}
+	content, _ := first["content"].([]any)
+	if len(content) != 1 {
+		t.Fatalf("expected reasoning content, got %#v", first["content"])
+	}
+	block0, _ := content[0].(map[string]any)
+	if block0["type"] != "reasoning" {
+		t.Fatalf("expected reasoning block, got %#v", block0["type"])
+	}
+	second, _ := output[1].(map[string]any)
+	if second["type"] != "function_call" {
+		t.Fatalf("expected function_call output, got %#v", second["type"])
 	}
 }
 

internal/httpapi/claude/handler_messages.go

@@ -145,7 +145,7 @@ func (h *Handler) handleClaudeDirectStream(w http.ResponseWriter, r *http.Reques
 		return
 	}
 	streamReq := start.Request
-	h.handleClaudeStreamRealtime(w, r, start.Response, streamReq.ResponseModel, streamReq.Messages, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, historySession)
+	h.handleClaudeStreamRealtimeWithRetry(w, r, a, start.Response, start.Payload, start.Pow, streamReq.ResponseModel, streamReq.Messages, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, streamReq.PromptTokenText, historySession)
 }
 
 func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, store ConfigReader) bool {
@@ -360,3 +360,112 @@ func (h *Handler) handleClaudeStreamRealtime(w http.ResponseWriter, r *http.Requ
 		OnFinalize: streamRuntime.onFinalize,
 	})
 }
+
+func (h *Handler) handleClaudeStreamRealtimeWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, model string, messages []any, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, promptTokenText string, historySession *responsehistory.Session) {
+	if resp.StatusCode != http.StatusOK {
+		defer func() { _ = resp.Body.Close() }()
+		body, _ := io.ReadAll(resp.Body)
+		if historySession != nil {
+			historySession.Error(resp.StatusCode, strings.TrimSpace(string(body)), "error", "", "")
+		}
+		writeClaudeError(w, http.StatusInternalServerError, string(body))
+		return
+	}
+
+	w.Header().Set("Content-Type", "text/event-stream")
+	w.Header().Set("Cache-Control", "no-cache, no-transform")
+	w.Header().Set("Connection", "keep-alive")
+	w.Header().Set("X-Accel-Buffering", "no")
+	rc := http.NewResponseController(w)
+	_, canFlush := w.(http.Flusher)
+	if !canFlush {
+		config.Logger.Warn("[claude_stream] response writer does not support flush; streaming may be buffered")
+	}
+
+	streamRuntime := newClaudeStreamRuntime(
+		w,
+		rc,
+		canFlush,
+		model,
+		messages,
+		thinkingEnabled,
+		searchEnabled,
+		stripReferenceMarkersEnabled(),
+		toolNames,
+		toolsRaw,
+		promptTokenText,
+		historySession,
+	)
+	streamRuntime.sendMessageStart()
+
+	completionruntime.ExecuteStreamWithRetry(r.Context(), h.DS, a, resp, payload, pow, completionruntime.StreamRetryOptions{
+		Surface:      "claude.messages",
+		Stream:       true,
+		RetryEnabled: true,
+		MaxAttempts:  3,
+		UsagePrompt:  promptTokenText,
+	}, completionruntime.StreamRetryHooks{
+		ConsumeAttempt: func(currentResp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			return h.consumeClaudeStreamAttempt(r, currentResp, streamRuntime, thinkingEnabled, allowDeferEmpty)
+		},
+		Finalize: func(_ int) {
+			streamRuntime.finalize("end_turn", false)
+		},
+		ParentMessageID: func() int {
+			return streamRuntime.responseMessageID
+		},
+		OnRetryPrompt: func(prompt string) {
+			streamRuntime.promptTokenText = prompt
+		},
+		OnRetryFailure: func(status int, message, code string) {
+			streamRuntime.sendErrorWithCode(status, strings.TrimSpace(message), code)
+		},
+	})
+}
+
+func (h *Handler) consumeClaudeStreamAttempt(r *http.Request, resp *http.Response, streamRuntime *claudeStreamRuntime, thinkingEnabled bool, allowDeferEmpty bool) (bool, bool) {
+	defer func() { _ = resp.Body.Close() }()
+	initialType := "text"
+	if thinkingEnabled {
+		initialType = "thinking"
+	}
+	finalReason := streamengine.StopReason("")
+	var scannerErr error
+	streamengine.ConsumeSSE(streamengine.ConsumeConfig{
+		Context:             r.Context(),
+		Body:                resp.Body,
+		ThinkingEnabled:     thinkingEnabled,
+		InitialType:         initialType,
+		KeepAliveInterval:   claudeStreamPingInterval,
+		IdleTimeout:         claudeStreamIdleTimeout,
+		MaxKeepAliveNoInput: claudeStreamMaxKeepaliveCnt,
+	}, streamengine.ConsumeHooks{
+		OnKeepAlive: func() {
+			streamRuntime.sendPing()
+		},
+		OnParsed: streamRuntime.onParsed,
+		OnFinalize: func(reason streamengine.StopReason, err error) {
+			finalReason = reason
+			scannerErr = err
+		},
+	})
+	if string(finalReason) == "upstream_error" {
+		if streamRuntime.history != nil {
+			streamRuntime.history.Error(500, streamRuntime.upstreamErr, "upstream_error", responsehistory.ThinkingForArchive(streamRuntime.rawThinking.String(), streamRuntime.toolDetectionThinking.String(), streamRuntime.thinking.String()), responsehistory.TextForArchive(streamRuntime.rawText.String(), streamRuntime.text.String()))
+		}
+		streamRuntime.sendError(streamRuntime.upstreamErr)
+		return true, false
+	}
+	if scannerErr != nil {
+		if streamRuntime.history != nil {
+			streamRuntime.history.Error(500, scannerErr.Error(), "error", responsehistory.ThinkingForArchive(streamRuntime.rawThinking.String(), streamRuntime.toolDetectionThinking.String(), streamRuntime.thinking.String()), responsehistory.TextForArchive(streamRuntime.rawText.String(), streamRuntime.text.String()))
+		}
+		streamRuntime.sendError(scannerErr.Error())
+		return true, false
+	}
+	terminalWritten := streamRuntime.finalize("end_turn", allowDeferEmpty)
+	if terminalWritten {
+		return true, false
+	}
+	return false, true
+}

internal/httpapi/claude/handler_util_test.go

@@ -93,14 +93,51 @@ func TestNormalizeClaudeMessagesToolUseToAssistantToolCalls(t *testing.T) {
 		t.Fatalf("expected call id preserved, got %#v", call)
 	}
 	content, _ := m["content"].(string)
-	if !containsStr(content, "<|DSML|tool_calls>") || !containsStr(content, `<|DSML|invoke name="search_web">`) {
+	if !containsStr(content, "<｜DSML｜tool_calls>") || !containsStr(content, `<｜DSML｜invoke name="search_web">`) {
 		t.Fatalf("expected assistant content to include DSML tool call history, got %q", content)
 	}
-	if !containsStr(content, `<|DSML|parameter name="query"><![CDATA[latest]]></|DSML|parameter>`) {
+	if !containsStr(content, `<｜DSML｜parameter name="query"><![CDATA[latest]]></｜DSML｜parameter>`) {
 		t.Fatalf("expected assistant content to include serialized parameters, got %q", content)
 	}
 }
 
+func TestNormalizeClaudeMessagesPreservesThinkingOnToolUseHistory(t *testing.T) {
+	msgs := []any{
+		map[string]any{
+			"role": "assistant",
+			"content": []any{
+				map[string]any{"type": "thinking", "thinking": "need live search before answering"},
+				map[string]any{
+					"type":  "tool_use",
+					"id":    "call_1",
+					"name":  "search_web",
+					"input": map[string]any{"query": "latest"},
+				},
+			},
+		},
+	}
+
+	got := normalizeClaudeMessages(msgs)
+	if len(got) != 1 {
+		t.Fatalf("expected one normalized tool-call message, got %#v", got)
+	}
+	m := got[0].(map[string]any)
+	if m["reasoning_content"] != "need live search before answering" {
+		t.Fatalf("expected thinking preserved as reasoning_content, got %#v", m)
+	}
+	tc, _ := m["tool_calls"].([]any)
+	if len(tc) != 1 {
+		t.Fatalf("expected one tool call, got %#v", m["tool_calls"])
+	}
+	prompt := buildClaudePromptTokenText(got, true)
+	if !containsStr(prompt, "[reasoning_content]\nneed live search before answering\n[/reasoning_content]") {
+		t.Fatalf("expected thinking in prompt history, got %q", prompt)
+	}
+	if !containsStr(prompt, `<｜DSML｜invoke name="search_web">`) {
+		t.Fatalf("expected tool call in prompt history, got %q", prompt)
+	}
+}
+
 func TestNormalizeClaudeMessagesDoesNotPromoteUserToolUse(t *testing.T) {
 	msgs := []any{
 		map[string]any{
@@ -292,7 +329,7 @@ func TestBuildClaudeToolPromptSingleTool(t *testing.T) {
 	if !containsStr(prompt, "Search the web") {
 		t.Fatalf("expected description in prompt")
 	}
-	if !containsStr(prompt, "<|DSML|tool_calls>") {
+	if !containsStr(prompt, "<｜DSML｜tool_calls>") {
 		t.Fatalf("expected DSML tool_calls format in prompt")
 	}
 	if !containsStr(prompt, "TOOL CALL FORMAT") {

internal/httpapi/claude/handler_utils.go

@@ -25,14 +25,21 @@ func normalizeClaudeMessages(messages []any) []any {
 		switch content := msg["content"].(type) {
 		case []any:
 			textParts := make([]string, 0, len(content))
+			pendingThinking := ""
 			flushText := func() {
 				if len(textParts) == 0 {
 					return
 				}
-				out = append(out, map[string]any{
+				message := map[string]any{
 					"role":    role,
 					"content": strings.Join(textParts, "\n"),
-				})
+				}
+				if role == "assistant" && strings.TrimSpace(pendingThinking) != "" {
+					message["reasoning_content"] = pendingThinking
+					message["content"] = prependClaudeReasoningForPrompt(pendingThinking, safeStringValue(message["content"]))
+					pendingThinking = ""
+				}
+				out = append(out, message)
 				textParts = textParts[:0]
 			}
 			for _, block := range content {
@@ -46,10 +53,29 @@ func normalizeClaudeMessages(messages []any) []any {
 					if t, ok := b["text"].(string); ok {
 						textParts = append(textParts, t)
 					}
+				case "thinking":
+					if role == "assistant" {
+						if thinking := extractClaudeThinkingBlockText(b); thinking != "" {
+							if pendingThinking == "" {
+								pendingThinking = thinking
+							} else {
+								pendingThinking += "\n" + thinking
+							}
+						}
+						continue
+					}
+					if raw := strings.TrimSpace(formatClaudeUnknownBlockForPrompt(b)); raw != "" {
+						textParts = append(textParts, raw)
+					}
 				case "tool_use":
 					if role == "assistant" {
 						flushText()
 						if toolMsg := normalizeClaudeToolUseToAssistant(b, state); toolMsg != nil {
+							if strings.TrimSpace(pendingThinking) != "" {
+								toolMsg["reasoning_content"] = pendingThinking
+								toolMsg["content"] = prependClaudeReasoningForPrompt(pendingThinking, safeStringValue(toolMsg["content"]))
+								pendingThinking = ""
+							}
 							out = append(out, toolMsg)
 						}
 						continue
@@ -69,6 +95,13 @@ func normalizeClaudeMessages(messages []any) []any {
 				}
 			}
 			flushText()
+			if role == "assistant" && strings.TrimSpace(pendingThinking) != "" {
+				out = append(out, map[string]any{
+					"role":              "assistant",
+					"reasoning_content": pendingThinking,
+					"content":           formatClaudeReasoningForPrompt(pendingThinking),
+				})
+			}
 		default:
 			copied := cloneMap(msg)
 			out = append(out, copied)
@@ -77,6 +110,39 @@ func normalizeClaudeMessages(messages []any) []any {
 	return out
 }
 
+func prependClaudeReasoningForPrompt(reasoning, content string) string {
+	reasoning = strings.TrimSpace(reasoning)
+	content = strings.TrimSpace(content)
+	if reasoning == "" {
+		return content
+	}
+	block := formatClaudeReasoningForPrompt(reasoning)
+	if content == "" {
+		return block
+	}
+	return block + "\n\n" + content
+}
+
+func formatClaudeReasoningForPrompt(reasoning string) string {
+	reasoning = strings.TrimSpace(reasoning)
+	if reasoning == "" {
+		return ""
+	}
+	return "[reasoning_content]\n" + reasoning + "\n[/reasoning_content]"
+}
+
+func extractClaudeThinkingBlockText(block map[string]any) string {
+	if block == nil {
+		return ""
+	}
+	for _, key := range []string{"thinking", "text", "content"} {
+		if text := strings.TrimSpace(safeStringValue(block[key])); text != "" {
+			return text
+		}
+	}
+	return ""
+}
+
 func buildClaudeToolPrompt(tools []any) string {
 	toolSchemas := make([]string, 0, len(tools))
 	names := make([]string, 0, len(tools))

internal/httpapi/claude/stream_runtime_core.go

@@ -29,9 +29,10 @@ type claudeStreamRuntime struct {
 	bufferToolContent     bool
 	stripReferenceMarkers bool
 
-	messageID string
-	thinking  strings.Builder
-	text      strings.Builder
+	messageID         string
+	thinking          strings.Builder
+	text              strings.Builder
+	responseMessageID int
 
 	sieve                 toolstream.State
 	rawText               strings.Builder
@@ -92,6 +93,9 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 		s.upstreamErr = parsed.ErrorMessage
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("upstream_error")}
 	}
+	if parsed.ResponseMessageID > 0 {
+		s.responseMessageID = parsed.ResponseMessageID
+	}
 	if parsed.Stop {
 		return streamengine.ParsedDecision{Stop: true}
 	}

internal/httpapi/claude/stream_runtime_emit.go

@@ -22,16 +22,27 @@ func (s *claudeStreamRuntime) send(event string, v any) {
 }
 
 func (s *claudeStreamRuntime) sendError(message string) {
+	s.sendErrorWithCode(500, message, "internal_error")
+}
+
+func (s *claudeStreamRuntime) sendErrorWithCode(status int, message, code string) {
 	msg := strings.TrimSpace(message)
 	if msg == "" {
 		msg = "upstream stream error"
 	}
+	if code == "" {
+		code = "internal_error"
+	}
+	errType := "api_error"
+	if status == 429 {
+		errType = "rate_limit_error"
+	}
 	s.send("error", map[string]any{
 		"type": "error",
 		"error": map[string]any{
-			"type":    "api_error",
+			"type":    errType,
 			"message": msg,
-			"code":    "internal_error",
+			"code":    code,
 			"param":   nil,
 		},
 	})

internal/httpapi/claude/stream_runtime_finalize.go

@@ -63,13 +63,10 @@ func (s *claudeStreamRuntime) sendToolUseBlock(idx int, tc toolcall.ParsedToolCa
 	})
 }
 
-func (s *claudeStreamRuntime) finalize(stopReason string) {
+func (s *claudeStreamRuntime) finalize(stopReason string, deferEmptyOutput bool) bool {
 	if s.ended {
-		return
+		return true
 	}
-	s.ended = true
-
-	s.closeThinkingBlock()
 
 	if s.bufferToolContent {
 		for _, evt := range toolstream.Flush(&s.sieve, s.toolNames) {
@@ -123,6 +120,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 		RawThinking:           s.rawThinking.String(),
 		VisibleThinking:       s.thinking.String(),
 		DetectionThinking:     s.toolDetectionThinking.String(),
+		ResponseMessageID:     s.responseMessageID,
 		AlreadyEmittedCalls:   s.toolCallsDetected,
 		AlreadyEmittedToolRaw: s.toolCallsDetected,
 	}, assistantturn.BuildOptions{
@@ -137,6 +135,22 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{
 		AlreadyEmittedToolCalls: s.toolCallsDetected,
 	})
+	if outcome.ShouldFail {
+		if deferEmptyOutput {
+			return false
+		}
+		s.ended = true
+		s.closeThinkingBlock()
+		s.closeTextBlock()
+		if s.history != nil {
+			s.history.Error(outcome.Error.Status, outcome.Error.Message, outcome.Error.Code, responsehistory.ThinkingForArchive(turn.RawThinking, turn.DetectionThinking, turn.Thinking), responsehistory.TextForArchive(turn.RawText, turn.Text))
+		}
+		s.sendErrorWithCode(outcome.Error.Status, outcome.Error.Message, outcome.Error.Code)
+		return true
+	}
+
+	s.ended = true
+	s.closeThinkingBlock()
 
 	if s.bufferToolContent && !s.toolCallsDetected {
 		if len(turn.ToolCalls) > 0 {
@@ -197,6 +211,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 		},
 	})
 	s.send("message_stop", map[string]any{"type": "message_stop"})
+	return true
 }
 
 func (s *claudeStreamRuntime) onFinalize(reason streamengine.StopReason, scannerErr error) {
@@ -214,5 +229,5 @@ func (s *claudeStreamRuntime) onFinalize(reason streamengine.StopReason, scanner
 		s.sendError(scannerErr.Error())
 		return
 	}
-	s.finalize("end_turn")
+	s.finalize("end_turn", false)
 }

internal/httpapi/gemini/convert_messages.go

@@ -44,14 +44,20 @@ func geminiMessagesFromRequest(req map[string]any) []any {
 		}
 
 		textParts := make([]string, 0, len(parts))
+		pendingThinking := ""
 		flushText := func() {
 			if len(textParts) == 0 {
 				return
 			}
-			out = append(out, map[string]any{
+			msg := map[string]any{
 				"role":    role,
 				"content": strings.Join(textParts, "\n"),
-			})
+			}
+			if role == "assistant" && strings.TrimSpace(pendingThinking) != "" {
+				msg["reasoning_content"] = pendingThinking
+				pendingThinking = ""
+			}
+			out = append(out, msg)
 			textParts = textParts[:0]
 		}
 
@@ -61,6 +67,14 @@ func geminiMessagesFromRequest(req map[string]any) []any {
 				continue
 			}
 			if text := strings.TrimSpace(asString(part["text"])); text != "" {
+				if role == "assistant" && isGeminiThoughtPart(part) {
+					if pendingThinking == "" {
+						pendingThinking = text
+					} else {
+						pendingThinking += "\n" + text
+					}
+					continue
+				}
 				textParts = append(textParts, text)
 				continue
 			}
@@ -75,7 +89,7 @@ func geminiMessagesFromRequest(req map[string]any) []any {
 						}
 					}
 					lastToolCallIDByName[strings.ToLower(name)] = callID
-					out = append(out, map[string]any{
+					msg := map[string]any{
 						"role": "assistant",
 						"tool_calls": []any{
 							map[string]any{
@@ -87,7 +101,12 @@ func geminiMessagesFromRequest(req map[string]any) []any {
 								},
 							},
 						},
-					})
+					}
+					if strings.TrimSpace(pendingThinking) != "" {
+						msg["reasoning_content"] = pendingThinking
+						pendingThinking = ""
+					}
+					out = append(out, msg)
 				}
 				continue
 			}
@@ -132,10 +151,29 @@ func geminiMessagesFromRequest(req map[string]any) []any {
 			}
 		}
 		flushText()
+		if role == "assistant" && strings.TrimSpace(pendingThinking) != "" {
+			out = append(out, map[string]any{
+				"role":              "assistant",
+				"reasoning_content": pendingThinking,
+			})
+		}
 	}
 	return out
 }
 
+func isGeminiThoughtPart(part map[string]any) bool {
+	if part == nil {
+		return false
+	}
+	if v, ok := part["thought"].(bool); ok {
+		return v
+	}
+	if v, ok := part["thoughtSignature"].(string); ok && strings.TrimSpace(v) != "" {
+		return true
+	}
+	return false
+}
+
 func normalizeGeminiSystemInstruction(raw any) string {
 	switch v := raw.(type) {
 	case string:

internal/httpapi/gemini/convert_messages_test.go

@@ -1,6 +1,7 @@
 package gemini
 
 import (
+	"ds2api/internal/promptcompat"
 	"strings"
 	"testing"
 )
@@ -53,6 +54,46 @@ func TestGeminiMessagesFromRequestPreservesFunctionRoundtrip(t *testing.T) {
 	}
 }
 
+func TestGeminiMessagesFromRequestPreservesThoughtOnFunctionCallHistory(t *testing.T) {
+	req := map[string]any{
+		"contents": []any{
+			map[string]any{
+				"role": "model",
+				"parts": []any{
+					map[string]any{"text": "need current state before answering", "thought": true},
+					map[string]any{
+						"functionCall": map[string]any{
+							"id":   "call_g1",
+							"name": "search_web",
+							"args": map[string]any{"query": "ai"},
+						},
+					},
+				},
+			},
+		},
+	}
+
+	got := geminiMessagesFromRequest(req)
+	if len(got) != 1 {
+		t.Fatalf("expected one normalized message, got %#v", got)
+	}
+	assistant, _ := got[0].(map[string]any)
+	if assistant["reasoning_content"] != "need current state before answering" {
+		t.Fatalf("expected thought preserved as reasoning_content, got %#v", assistant)
+	}
+	tc, _ := assistant["tool_calls"].([]any)
+	if len(tc) != 1 {
+		t.Fatalf("expected one tool call, got %#v", assistant["tool_calls"])
+	}
+	prompt, _ := promptcompat.BuildOpenAIPromptForAdapter(got, nil, "", true)
+	if !strings.Contains(prompt, "[reasoning_content]\nneed current state before answering\n[/reasoning_content]") {
+		t.Fatalf("expected thought in prompt history, got %q", prompt)
+	}
+	if !strings.Contains(prompt, `<｜DSML｜invoke name="search_web">`) {
+		t.Fatalf("expected tool call in prompt history, got %q", prompt)
+	}
+}
+
 func TestGeminiMessagesFromRequestPreservesUnknownPartAsRawJSONText(t *testing.T) {
 	req := map[string]any{
 		"contents": []any{

internal/httpapi/gemini/handler_generate.go

@@ -137,7 +137,7 @@ func (h *Handler) handleGeminiDirectStream(w http.ResponseWriter, r *http.Reques
 		return
 	}
 	streamReq := start.Request
-	h.handleStreamGenerateContent(w, r, start.Response, streamReq.ResponseModel, streamReq.PromptTokenText, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, historySession)
+	h.handleStreamGenerateContentWithRetry(w, r, a, start.Response, start.Payload, start.Pow, streamReq.ResponseModel, streamReq.PromptTokenText, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, historySession)
 }
 
 func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, stream bool) bool {

internal/httpapi/gemini/handler_stream_runtime.go

@@ -1,6 +1,7 @@
 package gemini
 
 import (
+	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -8,6 +9,8 @@ import (
 	"time"
 
 	"ds2api/internal/assistantturn"
+	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	"ds2api/internal/responsehistory"
 	"ds2api/internal/sse"
@@ -54,7 +57,7 @@ func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Req
 	}, streamengine.ConsumeHooks{
 		OnParsed: runtime.onParsed,
 		OnFinalize: func(_ streamengine.StopReason, _ error) {
-			runtime.finalize()
+			runtime.finalize(false)
 		},
 	})
 }
@@ -78,9 +81,83 @@ type geminiStreamRuntime struct {
 	accumulator       *assistantturn.Accumulator
 	contentFilter     bool
 	responseMessageID int
+	finalErrorStatus  int
+	finalErrorMessage string
+	finalErrorCode    string
 	history           *responsehistory.Session
 }
 
+func (h *Handler) handleStreamGenerateContentWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *responsehistory.Session) {
+	if resp.StatusCode != http.StatusOK {
+		defer func() { _ = resp.Body.Close() }()
+		body, _ := io.ReadAll(resp.Body)
+		if historySession != nil {
+			historySession.Error(resp.StatusCode, strings.TrimSpace(string(body)), "error", "", "")
+		}
+		writeGeminiError(w, resp.StatusCode, strings.TrimSpace(string(body)))
+		return
+	}
+
+	w.Header().Set("Content-Type", "text/event-stream")
+	w.Header().Set("Cache-Control", "no-cache, no-transform")
+	w.Header().Set("Connection", "keep-alive")
+	w.Header().Set("X-Accel-Buffering", "no")
+
+	rc := http.NewResponseController(w)
+	_, canFlush := w.(http.Flusher)
+	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, stripReferenceMarkersEnabled(), toolNames, toolsRaw, historySession)
+
+	completionruntime.ExecuteStreamWithRetry(r.Context(), h.DS, a, resp, payload, pow, completionruntime.StreamRetryOptions{
+		Surface:      "gemini.generate_content",
+		Stream:       true,
+		RetryEnabled: true,
+		MaxAttempts:  3,
+		UsagePrompt:  finalPrompt,
+	}, completionruntime.StreamRetryHooks{
+		ConsumeAttempt: func(currentResp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			return h.consumeGeminiStreamAttempt(r.Context(), currentResp, runtime, thinkingEnabled, allowDeferEmpty)
+		},
+		Finalize: func(_ int) {
+			runtime.finalize(false)
+		},
+		ParentMessageID: func() int {
+			return runtime.responseMessageID
+		},
+		OnRetryPrompt: func(prompt string) {
+			runtime.finalPrompt = prompt
+		},
+		OnRetryFailure: func(status int, message, _ string) {
+			runtime.sendErrorChunk(status, strings.TrimSpace(message))
+		},
+	})
+}
+
+func (h *Handler) consumeGeminiStreamAttempt(ctx context.Context, resp *http.Response, runtime *geminiStreamRuntime, thinkingEnabled bool, allowDeferEmpty bool) (bool, bool) {
+	defer func() { _ = resp.Body.Close() }()
+	initialType := "text"
+	if thinkingEnabled {
+		initialType = "thinking"
+	}
+	streamengine.ConsumeSSE(streamengine.ConsumeConfig{
+		Context:             ctx,
+		Body:                resp.Body,
+		ThinkingEnabled:     thinkingEnabled,
+		InitialType:         initialType,
+		KeepAliveInterval:   time.Duration(dsprotocol.KeepAliveTimeout) * time.Second,
+		IdleTimeout:         time.Duration(dsprotocol.StreamIdleTimeout) * time.Second,
+		MaxKeepAliveNoInput: dsprotocol.MaxKeepaliveCount,
+	}, streamengine.ConsumeHooks{
+		OnParsed: runtime.onParsed,
+		OnFinalize: func(_ streamengine.StopReason, _ error) {
+		},
+	})
+	terminalWritten := runtime.finalize(allowDeferEmpty)
+	if terminalWritten {
+		return true, false
+	}
+	return false, true
+}
+
 //nolint:unused // retained for native Gemini stream handling path.
 func newGeminiStreamRuntime(
 	w http.ResponseWriter,
@@ -127,6 +204,35 @@ func (s *geminiStreamRuntime) sendChunk(payload map[string]any) {
 	}
 }
 
+func (s *geminiStreamRuntime) sendErrorChunk(status int, message string) {
+	msg := strings.TrimSpace(message)
+	if msg == "" {
+		msg = http.StatusText(status)
+	}
+	errorStatus := "INVALID_ARGUMENT"
+	switch status {
+	case http.StatusUnauthorized:
+		errorStatus = "UNAUTHENTICATED"
+	case http.StatusForbidden:
+		errorStatus = "PERMISSION_DENIED"
+	case http.StatusTooManyRequests:
+		errorStatus = "RESOURCE_EXHAUSTED"
+	case http.StatusNotFound:
+		errorStatus = "NOT_FOUND"
+	default:
+		if status >= 500 {
+			errorStatus = "INTERNAL"
+		}
+	}
+	s.sendChunk(map[string]any{
+		"error": map[string]any{
+			"code":    status,
+			"message": msg,
+			"status":  errorStatus,
+		},
+	})
+}
+
 //nolint:unused // retained for native Gemini stream handling path.
 func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedDecision {
 	if !parsed.Parsed {
@@ -192,7 +298,7 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 }
 
 //nolint:unused // retained for native Gemini stream handling path.
-func (s *geminiStreamRuntime) finalize() {
+func (s *geminiStreamRuntime) finalize(deferEmptyOutput bool) bool {
 	rawText, text, rawThinking, thinking, detectionThinking := s.accumulator.Snapshot()
 	turn := assistantturn.BuildTurnFromStreamSnapshot(assistantturn.StreamSnapshot{
 		RawText:           rawText,
@@ -211,6 +317,19 @@ func (s *geminiStreamRuntime) finalize() {
 		ToolsRaw:              s.toolsRaw,
 	})
 	outcome := assistantturn.FinalizeTurn(turn, assistantturn.FinalizeOptions{})
+	if outcome.ShouldFail {
+		if deferEmptyOutput {
+			s.finalErrorStatus = outcome.Error.Status
+			s.finalErrorMessage = outcome.Error.Message
+			s.finalErrorCode = outcome.Error.Code
+			return false
+		}
+		if s.history != nil {
+			s.history.Error(outcome.Error.Status, outcome.Error.Message, outcome.Error.Code, responsehistory.ThinkingForArchive(turn.RawThinking, turn.DetectionThinking, turn.Thinking), responsehistory.TextForArchive(turn.RawText, turn.Text))
+		}
+		s.sendErrorChunk(outcome.Error.Status, outcome.Error.Message)
+		return true
+	}
 	if s.history != nil {
 		s.history.Success(
 			http.StatusOK,
@@ -257,4 +376,5 @@ func (s *geminiStreamRuntime) finalize() {
 			"totalTokenCount":      outcome.Usage.TotalTokens,
 		},
 	})
+	return true
 }

internal/httpapi/ollama/handler_routes.go

@@ -0,0 +1,58 @@
+package ollama
+
+import (
+	"ds2api/internal/config"
+	"ds2api/internal/util"
+	"encoding/json"
+	"github.com/go-chi/chi/v5"
+	"log/slog"
+	"net/http"
+)
+
+var WriteJSON = util.WriteJSON
+
+type ConfigReader interface {
+	ModelAliases() map[string]string
+}
+
+type Handler struct {
+	Store ConfigReader
+}
+
+type OllamaModelRequest struct {
+	Model string `json:"model"`
+}
+
+func RegisterRoutes(r chi.Router, h *Handler) {
+	r.Get("/api/version", h.GetVersion)
+	r.Get("/api/tags", h.ListOllamaModels)
+	r.Post("/api/show", h.GetOllamaModel)
+}
+
+func (h *Handler) GetVersion(w http.ResponseWriter, r *http.Request) {
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	_, _ = w.Write([]byte(`{"version":"0.23.1"}`))
+}
+func (h *Handler) ListOllamaModels(w http.ResponseWriter, r *http.Request) {
+	WriteJSON(w, http.StatusOK, config.OllamaModelsResponse())
+}
+func (h *Handler) GetOllamaModel(w http.ResponseWriter, r *http.Request) {
+	var payload OllamaModelRequest
+	if err := json.NewDecoder(r.Body).Decode(&payload); err != nil {
+		http.Error(w, "Invalid JSON body: "+err.Error(), http.StatusBadRequest)
+		return
+	}
+	defer func() {
+		if err := r.Body.Close(); err != nil {
+			slog.Warn("[ollama] failed to close request body", "error", err)
+		}
+	}()
+	modelID := payload.Model
+	model, ok := config.OllamaModelByID(h.Store, modelID)
+	if !ok {
+		http.Error(w, "Model not found.", http.StatusNotFound)
+		return
+	}
+	WriteJSON(w, http.StatusOK, model)
+}

internal/httpapi/ollama/handler_routes_test.go

@@ -0,0 +1,127 @@
+package ollama
+
+import (
+	"encoding/json"
+	"github.com/go-chi/chi/v5"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+type ollamaTestSurface struct {
+	Store   ConfigReader
+	handler *Handler
+}
+
+func (h *ollamaTestSurface) apiHandler() *Handler {
+	if h.handler == nil {
+		h.handler = &Handler{Store: h.Store}
+	}
+	return h.handler
+}
+
+func registerOllamaTestRoutes(r chi.Router, h *ollamaTestSurface) {
+	r.Get("/api/version", h.apiHandler().GetVersion)
+	r.Get("/api/tags", h.apiHandler().ListOllamaModels)
+	r.Post("/api/show", h.apiHandler().GetOllamaModel)
+}
+
+func TestGetOllamaVersionRoute(t *testing.T) {
+	h := &ollamaTestSurface{}
+	r := chi.NewRouter()
+	registerOllamaTestRoutes(r, h)
+	req := httptest.NewRequest(http.MethodGet, "/api/version", 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())
+	}
+}
+
+func TestGetOllamaModelsRoute(t *testing.T) {
+	h := &ollamaTestSurface{}
+	r := chi.NewRouter()
+	registerOllamaTestRoutes(r, h)
+	req := httptest.NewRequest(http.MethodGet, "/api/tags", 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())
+	}
+}
+
+func TestGetOllamaModelRoute(t *testing.T) {
+	h := &ollamaTestSurface{}
+	r := chi.NewRouter()
+	registerOllamaTestRoutes(r, h)
+
+	t.Run("direct", func(t *testing.T) {
+		body := `{"model":"deepseek-v4-flash"}`
+		req := httptest.NewRequest(http.MethodPost, "/api/show", strings.NewReader(body))
+		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())
+		}
+		var payload map[string]any
+		if err := json.Unmarshal(rec.Body.Bytes(), &payload); err != nil {
+			t.Fatalf("expected valid json body, got err=%v body=%s", err, rec.Body.String())
+		}
+		if _, ok := payload["id"]; !ok {
+			t.Fatalf("expected response has lowercase id field, body=%s", rec.Body.String())
+		}
+		if _, ok := payload["ID"]; ok {
+			t.Fatalf("expected response does not expose uppercase ID field, body=%s", rec.Body.String())
+		}
+	})
+
+	t.Run("direct_nothinking", func(t *testing.T) {
+		body := `{"model":"deepseek-v4-flash-nothinking"}`
+		req := httptest.NewRequest(http.MethodPost, "/api/show", strings.NewReader(body))
+		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())
+		}
+	})
+
+	t.Run("direct_expert", func(t *testing.T) {
+		body := `{"model":"deepseek-v4-pro"}`
+		req := httptest.NewRequest(http.MethodPost, "/api/show", strings.NewReader(body))
+		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())
+		}
+	})
+
+	t.Run("direct_vision", func(t *testing.T) {
+		body := `{"model":"deepseek-v4-vision"}`
+		req := httptest.NewRequest(http.MethodPost, "/api/show", strings.NewReader(body))
+		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())
+		}
+	})
+}
+
+func TestGetOllamaModelRouteNotFound(t *testing.T) {
+	h := &ollamaTestSurface{}
+	r := chi.NewRouter()
+	registerOllamaTestRoutes(r, h)
+
+	body := `{"model":"not-exists"}`
+	req := httptest.NewRequest(http.MethodPost, "/api/show", strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+	if rec.Code != http.StatusNotFound {
+		t.Fatalf("expected 404, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}

internal/httpapi/openai/chat/empty_retry_runtime.go

@@ -4,11 +4,11 @@ import (
 	"context"
 	"io"
 	"net/http"
-	"strings"
 	"time"
 
 	"ds2api/internal/assistantturn"
 	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	openaifmt "ds2api/internal/format/openai"
@@ -17,191 +17,94 @@ import (
 	streamengine "ds2api/internal/stream"
 )
 
-type chatNonStreamResult struct {
-	rawThinking           string
-	rawText               string
-	thinking              string
-	toolDetectionThinking string
-	text                  string
-	contentFilter         bool
-	detectedCalls         int
-	body                  map[string]any
-	finishReason          string
-	responseMessageID     int
-	outputError           *assistantturn.OutputError
-}
-
-func (r chatNonStreamResult) historyText() string {
-	return historyTextForArchive(r.rawText, r.text)
-}
-
-func (r chatNonStreamResult) historyThinking() string {
-	return historyThinkingForArchive(r.rawThinking, r.toolDetectionThinking, r.thinking)
-}
-
 func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
-	attempts := 0
-	currentResp := resp
-	usagePrompt := finalPrompt
-	accumulatedThinking := ""
-	accumulatedRawThinking := ""
-	accumulatedToolDetectionThinking := ""
-	for {
-		result, ok := h.collectChatNonStreamAttempt(w, currentResp, completionID, model, usagePrompt, thinkingEnabled, searchEnabled, toolNames, toolsRaw)
-		if !ok {
-			return
-		}
-		accumulatedThinking += sse.TrimContinuationOverlap(accumulatedThinking, result.thinking)
-		accumulatedRawThinking += sse.TrimContinuationOverlap(accumulatedRawThinking, result.rawThinking)
-		accumulatedToolDetectionThinking += sse.TrimContinuationOverlap(accumulatedToolDetectionThinking, result.toolDetectionThinking)
-		result.thinking = accumulatedThinking
-		result.rawThinking = accumulatedRawThinking
-		result.toolDetectionThinking = accumulatedToolDetectionThinking
-		detected := detectAssistantToolCalls(result.rawText, result.text, result.rawThinking, result.toolDetectionThinking, toolNames)
-		result.detectedCalls = len(detected.Calls)
-		result.body = openaifmt.BuildChatCompletionWithToolCalls(completionID, model, usagePrompt, result.thinking, result.text, detected.Calls, toolsRaw)
-		addRefFileTokensToUsage(result.body, refFileTokens)
-		result.finishReason = chatFinishReason(result.body)
-		if !shouldRetryChatNonStream(result, attempts) {
-			h.finishChatNonStreamResult(w, result, attempts, usagePrompt, refFileTokens, historySession)
-			return
-		}
-
-		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "parent_message_id", result.responseMessageID)
-		retryPow, powErr := h.DS.GetPow(ctx, a, 3)
-		if powErr != nil {
-			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "error", powErr)
-			retryPow = pow
-		}
-		retryPayload := clonePayloadForEmptyOutputRetry(payload, result.responseMessageID)
-		nextResp, err := h.DS.CallCompletion(ctx, a, retryPayload, retryPow, 3)
-		if err != nil {
-			if historySession != nil {
-				historySession.error(http.StatusInternalServerError, "Failed to get completion.", "error", result.historyThinking(), result.historyText())
-			}
-			writeOpenAIError(w, http.StatusInternalServerError, "Failed to get completion.")
-			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "error", err)
-			return
-		}
-		usagePrompt = usagePromptWithEmptyOutputRetry(usagePrompt, attempts)
-		currentResp = nextResp
-	}
-}
-
-func (h *Handler) collectChatNonStreamAttempt(w http.ResponseWriter, resp *http.Response, completionID, model, usagePrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any) (chatNonStreamResult, bool) {
 	if resp.StatusCode != http.StatusOK {
 		defer func() { _ = resp.Body.Close() }()
 		body, _ := io.ReadAll(resp.Body)
-		writeOpenAIError(w, resp.StatusCode, string(body))
-		return chatNonStreamResult{}, false
-	}
-	result := sse.CollectStream(resp, thinkingEnabled, true)
-	turn := assistantturn.BuildTurnFromCollected(result, assistantturn.BuildOptions{
-		Model:         model,
-		Prompt:        usagePrompt,
-		SearchEnabled: searchEnabled,
-		ToolNames:     toolNames,
-		ToolsRaw:      toolsRaw,
-	})
-	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, usagePrompt, turn.Thinking, turn.Text, turn.ToolCalls, toolsRaw)
-	return chatNonStreamResult{
-		rawThinking:           result.Thinking,
-		rawText:               result.Text,
-		thinking:              turn.Thinking,
-		toolDetectionThinking: result.ToolDetectionThinking,
-		text:                  turn.Text,
-		contentFilter:         result.ContentFilter,
-		detectedCalls:         len(turn.ToolCalls),
-		body:                  respBody,
-		finishReason:          chatFinishReason(respBody),
-		responseMessageID:     result.ResponseMessageID,
-		outputError:           turn.Error,
-	}, true
-}
-
-func (h *Handler) finishChatNonStreamResult(w http.ResponseWriter, result chatNonStreamResult, attempts int, usagePrompt string, refFileTokens int, historySession *chatHistorySession) {
-	if result.detectedCalls == 0 && strings.TrimSpace(result.text) == "" {
-		status, message, code := upstreamEmptyOutputDetail(result.contentFilter, result.text, result.thinking)
-		if result.outputError != nil {
-			status, message, code = result.outputError.Status, result.outputError.Message, result.outputError.Code
-		}
 		if historySession != nil {
-			historySession.error(status, message, code, result.historyThinking(), result.historyText())
+			historySession.error(resp.StatusCode, string(body), "error", "", "")
 		}
-		writeOpenAIErrorWithCode(w, status, message, code)
-		config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "chat.completions", "stream", false, "retry_attempts", attempts, "success_source", "none", "content_filter", result.contentFilter)
+		writeOpenAIError(w, resp.StatusCode, string(body))
 		return
 	}
-	if historySession != nil {
-		historySession.success(http.StatusOK, result.historyThinking(), result.historyText(), result.finishReason, openaifmt.BuildChatUsageForModel("", usagePrompt, result.thinking, result.text, refFileTokens))
+	stdReq := promptcompat.StandardRequest{
+		Surface:         "chat.completions",
+		ResponseModel:   model,
+		PromptTokenText: finalPrompt,
+		FinalPrompt:     finalPrompt,
+		RefFileTokens:   refFileTokens,
+		Thinking:        thinkingEnabled,
+		Search:          searchEnabled,
+		ToolNames:       toolNames,
+		ToolsRaw:        toolsRaw,
+		ToolChoice:      promptcompat.DefaultToolChoicePolicy(),
 	}
-	writeJSON(w, http.StatusOK, result.body)
-	source := "first_attempt"
-	if attempts > 0 {
-		source = "synthetic_retry"
-	}
-	config.Logger.Info("[openai_empty_retry] completed", "surface", "chat.completions", "stream", false, "retry_attempts", attempts, "success_source", source)
-}
-
-func chatFinishReason(respBody map[string]any) string {
-	if choices, ok := respBody["choices"].([]map[string]any); ok && len(choices) > 0 {
-		if fr, _ := choices[0]["finish_reason"].(string); strings.TrimSpace(fr) != "" {
-			return fr
+	retryEnabled := h != nil && h.DS != nil && emptyOutputRetryEnabled()
+	result, outErr := completionruntime.ExecuteNonStreamStartedWithRetry(ctx, h.DS, a, completionruntime.StartResult{
+		SessionID: completionID,
+		Payload:   payload,
+		Pow:       pow,
+		Response:  resp,
+		Request:   stdReq,
+	}, completionruntime.Options{
+		RetryEnabled:     retryEnabled,
+		RetryMaxAttempts: emptyOutputRetryMaxAttempts(),
+	})
+	if outErr != nil {
+		if historySession != nil {
+			historySession.error(outErr.Status, outErr.Message, outErr.Code, historyThinkingForArchive(result.Turn.RawThinking, result.Turn.DetectionThinking, result.Turn.Thinking), historyTextForArchive(result.Turn.RawText, result.Turn.Text))
 		}
+		writeOpenAIErrorWithCode(w, outErr.Status, outErr.Message, outErr.Code)
+		return
 	}
-	return "stop"
+	respBody := openaifmt.BuildChatCompletionWithToolCalls(result.SessionID, model, result.Turn.Prompt, result.Turn.Thinking, result.Turn.Text, result.Turn.ToolCalls, toolsRaw)
+	respBody["usage"] = assistantturn.OpenAIChatUsage(result.Turn)
+	outcome := assistantturn.FinalizeTurn(result.Turn, assistantturn.FinalizeOptions{})
+	if historySession != nil {
+		historySession.success(http.StatusOK, historyThinkingForArchive(result.Turn.RawThinking, result.Turn.DetectionThinking, result.Turn.Thinking), historyTextForArchive(result.Turn.RawText, result.Turn.Text), outcome.FinishReason, assistantturn.OpenAIChatUsage(result.Turn))
+	}
+	writeJSON(w, http.StatusOK, respBody)
 }
 
-func shouldRetryChatNonStream(result chatNonStreamResult, attempts int) bool {
-	return emptyOutputRetryEnabled() &&
-		attempts < emptyOutputRetryMaxAttempts() &&
-		!result.contentFilter &&
-		result.detectedCalls == 0 &&
-		strings.TrimSpace(result.text) == ""
-}
-
-func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, historySession *chatHistorySession) {
+func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID string, sessionIDRef *string, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, historySession *chatHistorySession) {
 	streamRuntime, initialType, ok := h.prepareChatStreamRuntime(w, resp, completionID, model, finalPrompt, refFileTokens, thinkingEnabled, searchEnabled, toolNames, toolsRaw, toolChoice, historySession)
 	if !ok {
 		return
 	}
-	attempts := 0
-	currentResp := resp
-	for {
-		terminalWritten, retryable := h.consumeChatStreamAttempt(r, currentResp, streamRuntime, initialType, thinkingEnabled, historySession, attempts < emptyOutputRetryMaxAttempts())
-		if terminalWritten {
-			logChatStreamTerminal(streamRuntime, attempts)
-			return
-		}
-		if !retryable || !emptyOutputRetryEnabled() || attempts >= emptyOutputRetryMaxAttempts() {
+	completionruntime.ExecuteStreamWithRetry(r.Context(), h.DS, a, resp, payload, pow, completionruntime.StreamRetryOptions{
+		Surface:          "chat.completions",
+		Stream:           true,
+		RetryEnabled:     emptyOutputRetryEnabled(),
+		RetryMaxAttempts: emptyOutputRetryMaxAttempts(),
+		MaxAttempts:      3,
+		UsagePrompt:      finalPrompt,
+	}, completionruntime.StreamRetryHooks{
+		ConsumeAttempt: func(currentResp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			return h.consumeChatStreamAttempt(r, currentResp, streamRuntime, initialType, thinkingEnabled, historySession, allowDeferEmpty)
+		},
+		Finalize: func(attempts int) {
 			streamRuntime.finalize("stop", false)
 			recordChatStreamHistory(streamRuntime, historySession)
 			config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "chat.completions", "stream", true, "retry_attempts", attempts, "success_source", "none")
-			return
-		}
-		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "parent_message_id", streamRuntime.responseMessageID)
-		retryPow, powErr := h.DS.GetPow(r.Context(), a, 3)
-		if powErr != nil {
-			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "error", powErr)
-			retryPow = pow
-		}
-		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadForEmptyOutputRetry(payload, streamRuntime.responseMessageID), retryPow, 3)
-		if err != nil {
-			failChatStreamRetry(streamRuntime, historySession, http.StatusInternalServerError, "Failed to get completion.", "error")
-			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "chat.completions", "stream", true, "retry_attempt", attempts, "error", err)
-			return
-		}
-		if nextResp.StatusCode != http.StatusOK {
-			defer func() { _ = nextResp.Body.Close() }()
-			body, _ := io.ReadAll(nextResp.Body)
-			failChatStreamRetry(streamRuntime, historySession, nextResp.StatusCode, string(body), "error")
-			return
-		}
-		streamRuntime.finalPrompt = usagePromptWithEmptyOutputRetry(finalPrompt, attempts)
-		currentResp = nextResp
-	}
+		},
+		ParentMessageID: func() int {
+			return streamRuntime.responseMessageID
+		},
+		OnRetryPrompt: func(prompt string) {
+			streamRuntime.finalPrompt = prompt
+		},
+		OnRetryFailure: func(status int, message, code string) {
+			failChatStreamRetry(streamRuntime, historySession, status, message, code)
+		},
+		OnAccountSwitch: func(sessionID string) {
+			if sessionIDRef != nil {
+				*sessionIDRef = sessionID
+			}
+		},
+		OnTerminal: func(attempts int) {
+			logChatStreamTerminal(streamRuntime, attempts)
+		},
+	})
 }
 
 func (h *Handler) prepareChatStreamRuntime(w http.ResponseWriter, resp *http.Response, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, historySession *chatHistorySession) (*chatStreamRuntime, string, bool) {

internal/httpapi/openai/chat/handler.go

@@ -106,10 +106,6 @@ func cleanVisibleOutput(text string, stripReferenceMarkers bool) string {
 	return shared.CleanVisibleOutput(text, stripReferenceMarkers)
 }
 
-func upstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
-	return shared.UpstreamEmptyOutputDetail(contentFilter, text, thinking)
-}
-
 func emptyOutputRetryEnabled() bool {
 	return shared.EmptyOutputRetryEnabled()
 }
@@ -118,14 +114,6 @@ func emptyOutputRetryMaxAttempts() int {
 	return shared.EmptyOutputRetryMaxAttempts()
 }
 
-func clonePayloadForEmptyOutputRetry(payload map[string]any, parentMessageID int) map[string]any {
-	return shared.ClonePayloadForEmptyOutputRetry(payload, parentMessageID)
-}
-
-func usagePromptWithEmptyOutputRetry(originalPrompt string, retryAttempts int) string {
-	return shared.UsagePromptWithEmptyOutputRetry(originalPrompt, retryAttempts)
-}
-
 func formatIncrementalStreamToolCallDeltas(deltas []toolstream.ToolCallDelta, ids map[int]string) []map[string]any {
 	return shared.FormatIncrementalStreamToolCallDeltas(deltas, ids)
 }
@@ -137,7 +125,3 @@ func filterIncrementalToolCallDeltasByAllowed(deltas []toolstream.ToolCallDelta,
 func formatFinalStreamToolCallsWithStableIDs(calls []toolcall.ParsedToolCall, ids map[int]string, toolsRaw any) []map[string]any {
 	return shared.FormatFinalStreamToolCallsWithStableIDs(calls, ids, toolsRaw)
 }
-
-func detectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
-	return shared.DetectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking, toolNames)
-}

internal/httpapi/openai/chat/handler_chat.go

@@ -114,7 +114,7 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 	}
 	streamReq := start.Request
 	refFileTokens := streamReq.RefFileTokens
-	h.handleStreamWithRetry(w, r, a, start.Response, start.Payload, start.Pow, sessionID, streamReq.ResponseModel, streamReq.PromptTokenText, refFileTokens, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, streamReq.ToolChoice, historySession)
+	h.handleStreamWithRetry(w, r, a, start.Response, start.Payload, start.Pow, sessionID, &sessionID, streamReq.ResponseModel, streamReq.PromptTokenText, refFileTokens, streamReq.Thinking, streamReq.Search, streamReq.ToolNames, streamReq.ToolsRaw, streamReq.ToolChoice, historySession)
 }
 
 func (h *Handler) autoDeleteRemoteSession(ctx context.Context, a *auth.RequestAuth, sessionID string) {

internal/httpapi/openai/chat/handler_toolcall_test.go

@@ -85,8 +85,7 @@ func streamFinishReason(frames []map[string]any) string {
 	return ""
 }
 
-// Backward-compatible alias for historical test name used in CI logs.
-func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
+func TestHandleNonStreamSingleAttemptReturns503WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		`data: {"p":"response/content","v":""}`,
@@ -95,17 +94,17 @@ func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	rec := httptest.NewRecorder()
 
 	h.handleNonStream(rec, resp, "cid-empty", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, 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 rec.Code != http.StatusServiceUnavailable {
+		t.Fatalf("expected status 503 for 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"]) != "upstream_empty_output" {
-		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
+	if asString(errObj["code"]) != "upstream_unavailable" {
+		t.Fatalf("expected code=upstream_unavailable, got %#v", out)
 	}
 }
 
-func TestHandleNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
+func TestHandleNonStreamSingleAttemptReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		`data: {"code":"content_filter"}`,
@@ -124,7 +123,7 @@ func TestHandleNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutp
 	}
 }
 
-func TestHandleNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
+func TestHandleNonStreamSingleAttemptReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		`data: {"p":"response/thinking_content","v":"Only thinking"}`,

internal/httpapi/openai/chat/ref_file_tokens.go

@@ -1,26 +0,0 @@
-package chat
-
-// addRefFileTokensToUsage adds inline-uploaded file token estimates to an existing
-// usage map inside a response object. This keeps the token accounting aware of file
-// content that the upstream model processes but that is not part of the prompt text.
-func addRefFileTokensToUsage(obj map[string]any, refFileTokens int) {
-	if refFileTokens <= 0 || obj == nil {
-		return
-	}
-	usage, ok := obj["usage"].(map[string]any)
-	if !ok || usage == nil {
-		return
-	}
-	for _, key := range []string{"input_tokens", "prompt_tokens"} {
-		if v, ok := usage[key]; ok {
-			if n, ok := v.(int); ok {
-				usage[key] = n + refFileTokens
-			}
-		}
-	}
-	if v, ok := usage["total_tokens"]; ok {
-		if n, ok := v.(int); ok {
-			usage["total_tokens"] = n + refFileTokens
-		}
-	}
-}

internal/httpapi/openai/history_split_test.go

@@ -84,7 +84,7 @@ func TestBuildOpenAICurrentInputContextTranscriptUsesNumberedHistorySections(t *
 		"latest user turn",
 		"[reasoning_content]",
 		"hidden reasoning",
-		"<|DSML|tool_calls>",
+		"<｜DSML｜tool_calls>",
 	} {
 		if !strings.Contains(transcript, want) {
 			t.Fatalf("expected transcript to contain %q, got %q", want, transcript)

internal/httpapi/openai/responses/empty_retry_runtime.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	"ds2api/internal/promptcompat"
@@ -19,41 +20,34 @@ func (h *Handler) handleResponsesStreamWithRetry(w http.ResponseWriter, r *http.
 	if !ok {
 		return
 	}
-	attempts := 0
-	currentResp := resp
-	for {
-		terminalWritten, retryable := h.consumeResponsesStreamAttempt(r, currentResp, streamRuntime, initialType, thinkingEnabled, attempts < emptyOutputRetryMaxAttempts())
-		if terminalWritten {
-			logResponsesStreamTerminal(streamRuntime, attempts)
-			return
-		}
-		if !retryable || !emptyOutputRetryEnabled() || attempts >= emptyOutputRetryMaxAttempts() {
+	completionruntime.ExecuteStreamWithRetry(r.Context(), h.DS, a, resp, payload, pow, completionruntime.StreamRetryOptions{
+		Surface:          "responses",
+		Stream:           true,
+		RetryEnabled:     emptyOutputRetryEnabled(),
+		RetryMaxAttempts: emptyOutputRetryMaxAttempts(),
+		MaxAttempts:      3,
+		UsagePrompt:      finalPrompt,
+	}, completionruntime.StreamRetryHooks{
+		ConsumeAttempt: func(currentResp *http.Response, allowDeferEmpty bool) (bool, bool) {
+			return h.consumeResponsesStreamAttempt(r, currentResp, streamRuntime, initialType, thinkingEnabled, allowDeferEmpty)
+		},
+		Finalize: func(attempts int) {
 			streamRuntime.finalize("stop", false)
 			config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "responses", "stream", true, "retry_attempts", attempts, "success_source", "none", "error_code", streamRuntime.finalErrorCode)
-			return
-		}
-		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", true, "retry_attempt", attempts, "parent_message_id", streamRuntime.responseMessageID)
-		retryPow, powErr := h.DS.GetPow(r.Context(), a, 3)
-		if powErr != nil {
-			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "responses", "stream", true, "retry_attempt", attempts, "error", powErr)
-			retryPow = pow
-		}
-		nextResp, err := h.DS.CallCompletion(r.Context(), a, clonePayloadForEmptyOutputRetry(payload, streamRuntime.responseMessageID), retryPow, 3)
-		if err != nil {
-			streamRuntime.failResponse(http.StatusInternalServerError, "Failed to get completion.", "error")
-			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "responses", "stream", true, "retry_attempt", attempts, "error", err)
-			return
-		}
-		if nextResp.StatusCode != http.StatusOK {
-			defer func() { _ = nextResp.Body.Close() }()
-			body, _ := io.ReadAll(nextResp.Body)
-			streamRuntime.failResponse(nextResp.StatusCode, strings.TrimSpace(string(body)), "error")
-			return
-		}
-		streamRuntime.finalPrompt = usagePromptWithEmptyOutputRetry(finalPrompt, attempts)
-		currentResp = nextResp
-	}
+		},
+		ParentMessageID: func() int {
+			return streamRuntime.responseMessageID
+		},
+		OnRetryPrompt: func(prompt string) {
+			streamRuntime.finalPrompt = prompt
+		},
+		OnRetryFailure: func(status int, message, code string) {
+			streamRuntime.failResponse(status, strings.TrimSpace(message), code)
+		},
+		OnTerminal: func(attempts int) {
+			logResponsesStreamTerminal(streamRuntime, attempts)
+		},
+	})
 }
 
 func (h *Handler) prepareResponsesStreamRuntime(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string, historySession *responsehistory.Session) (*responsesStreamRuntime, string, bool) {

internal/httpapi/openai/responses/handler.go

@@ -103,14 +103,6 @@ func emptyOutputRetryMaxAttempts() int {
 	return shared.EmptyOutputRetryMaxAttempts()
 }
 
-func clonePayloadForEmptyOutputRetry(payload map[string]any, parentMessageID int) map[string]any {
-	return shared.ClonePayloadForEmptyOutputRetry(payload, parentMessageID)
-}
-
-func usagePromptWithEmptyOutputRetry(originalPrompt string, retryAttempts int) string {
-	return shared.UsagePromptWithEmptyOutputRetry(originalPrompt, retryAttempts)
-}
-
 func filterIncrementalToolCallDeltasByAllowed(deltas []toolstream.ToolCallDelta, seenNames map[int]string) []toolstream.ToolCallDelta {
 	return shared.FilterIncrementalToolCallDeltasByAllowed(deltas, seenNames)
 }

internal/httpapi/openai/responses/responses_stream_runtime_toolcalls_finalize.go

@@ -81,6 +81,22 @@ func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, fin
 				},
 			},
 		})
+	} else if len(calls) > 0 && strings.TrimSpace(finalThinking) != "" {
+		indexed = append(indexed, indexedItem{
+			index: s.ensureMessageOutputIndex(),
+			item: map[string]any{
+				"id":     s.ensureMessageItemID(),
+				"type":   "message",
+				"role":   "assistant",
+				"status": "completed",
+				"content": []map[string]any{
+					{
+						"type": "reasoning",
+						"text": finalThinking,
+					},
+				},
+			},
+		})
 	} else if len(calls) == 0 {
 		content := make([]map[string]any, 0, 2)
 		if finalThinking != "" {

internal/httpapi/openai/responses/responses_stream_test.go

@@ -397,7 +397,7 @@ func TestHandleResponsesNonStreamRequiredToolChoiceIgnoresThinkingToolPayloadWhe
 	}
 }
 
-func TestHandleResponsesNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
+func TestHandleResponsesNonStreamSingleAttemptReturns503WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	rec := httptest.NewRecorder()
 	resp := &http.Response{
@@ -409,17 +409,17 @@ func TestHandleResponsesNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T)
 	}
 
 	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
-	if rec.Code != http.StatusTooManyRequests {
-		t.Fatalf("expected 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	if rec.Code != http.StatusServiceUnavailable {
+		t.Fatalf("expected 503 for 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"]) != "upstream_empty_output" {
-		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
+	if asString(errObj["code"]) != "upstream_unavailable" {
+		t.Fatalf("expected code=upstream_unavailable, got %#v", out)
 	}
 }
 
-func TestHandleResponsesNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
+func TestHandleResponsesNonStreamSingleAttemptReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutput(t *testing.T) {
 	h := &Handler{}
 	rec := httptest.NewRecorder()
 	resp := &http.Response{
@@ -441,7 +441,7 @@ func TestHandleResponsesNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWi
 	}
 }
 
-func TestHandleResponsesNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
+func TestHandleResponsesNonStreamSingleAttemptReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
 	h := &Handler{}
 	rec := httptest.NewRecorder()
 	resp := &http.Response{

internal/httpapi/openai/shared/upstream_empty.go

@@ -17,7 +17,7 @@ func UpstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int,
 	if thinking != "" {
 		return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned reasoning without visible output.", "upstream_empty_output"
 	}
-	return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned empty output.", "upstream_empty_output"
+	return http.StatusServiceUnavailable, "Upstream service is unavailable and returned no output.", "upstream_unavailable"
 }
 
 func WriteUpstreamEmptyOutputError(w http.ResponseWriter, text, thinking string, contentFilter bool) bool {

internal/httpapi/openai/stream_status_test.go

@@ -274,12 +274,12 @@ func TestChatCompletionsStreamEmitsFailureFrameWhenUpstreamOutputEmpty(t *testin
 	}
 	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())
+	if !ok || int(statusCode) != http.StatusServiceUnavailable {
+		t.Fatalf("expected status_code=503, 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)
+	if asString(errObj["code"]) != "upstream_unavailable" {
+		t.Fatalf("expected code=upstream_unavailable, got %#v", last)
 	}
 }
 
@@ -345,7 +345,7 @@ func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 
 func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"response_message_id":99}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":99,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -496,7 +496,7 @@ func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 
 func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"response_message_id":88}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":88,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -537,8 +537,15 @@ func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if len(content) == 0 {
 		t.Fatalf("expected content entries, got %#v", item)
 	}
-	textEntry, _ := content[0].(map[string]any)
-	if asString(textEntry["type"]) != "output_text" || asString(textEntry["text"]) != "visible" {
+	var textEntry map[string]any
+	for _, entry := range content {
+		obj, _ := entry.(map[string]any)
+		if asString(obj["type"]) == "output_text" {
+			textEntry = obj
+			break
+		}
+	}
+	if asString(textEntry["text"]) != "visible" {
 		t.Fatalf("expected visible text entry, got %#v", content)
 	}
 }

internal/js/chat-stream/cors.js

@@ -19,13 +19,15 @@ const BLOCKED_CORS_REQUEST_HEADERS = new Set([
 function setCorsHeaders(res, req) {
   const origin = asString(readHeader(req, 'origin'));
   res.setHeader('Access-Control-Allow-Origin', origin || '*');
+  if (origin) {
+    addVaryHeader(res, 'Origin');
+  }
   res.setHeader('Access-Control-Allow-Methods', 'GET, POST, OPTIONS, PUT, DELETE');
   res.setHeader('Access-Control-Max-Age', '600');
   res.setHeader(
     'Access-Control-Allow-Headers',
     buildCORSAllowHeaders(req),
   );
-  addVaryHeader(res, 'Origin');
   addVaryHeader(res, 'Access-Control-Request-Headers');
   if (asString(readHeader(req, 'access-control-request-private-network')).toLowerCase() === 'true') {
     res.setHeader('Access-Control-Allow-Private-Network', 'true');

internal/js/chat-stream/index.js

@@ -88,7 +88,7 @@ function isVercelRuntime() {
 
 function isNodeStreamSupportedPath(rawURL) {
   const path = extractPathname(rawURL);
-  return path === '/v1/chat/completions';
+  return path === '/v1/chat/completions' || path === '/chat/completions';
 }
 
 function extractPathname(rawURL) {

internal/js/chat-stream/vercel_stream_impl.js

@@ -641,9 +641,9 @@ function upstreamEmptyOutputDetail(contentFilter, _text, thinking) {
     };
   }
   return {
-    status: 429,
-    message: 'Upstream account hit a rate limit and returned empty output.',
-    code: 'upstream_empty_output',
+    status: 503,
+    message: 'Upstream service is unavailable and returned no output.',
+    code: 'upstream_unavailable',
   };
 }
 

internal/js/helpers/stream-tool-sieve/parse.js

@@ -113,9 +113,10 @@ function filterToolCallsDetailed(parsed, toolNames) {
     if (!tc || !tc.name) {
       continue;
     }
+    const input = tc.input && typeof tc.input === 'object' && !Array.isArray(tc.input) ? tc.input : {};
     calls.push({
       name: tc.name,
-      input: tc.input && typeof tc.input === 'object' && !Array.isArray(tc.input) ? tc.input : {},
+      input,
     });
   }
   return { calls, rejectedToolNames: [] };

internal/js/helpers/stream-tool-sieve/parse_payload.js

@@ -1,10 +1,11 @@
 'use strict';
 
-const CDATA_PATTERN = /^<!\[CDATA\[([\s\S]*?)]]>$/i;
+const CDATA_PATTERN = /^(?:<|〈)(?:!|！)\[CDATA\[([\s\S]*?)]](?:>|＞|〉)$/i;
 const XML_ATTR_PATTERN = /\b([a-z0-9_:-]+)\s*=\s*("([^"]*)"|'([^']*)')/gi;
 const TOOL_MARKUP_NAMES = [
   { raw: 'tool_calls', canonical: 'tool_calls' },
   { raw: 'tool-calls', canonical: 'tool_calls', dsmlOnly: true },
+  { raw: 'toolcalls', canonical: 'tool_calls', dsmlOnly: true },
   { raw: 'invoke', canonical: 'invoke' },
   { raw: 'parameter', canonical: 'parameter' },
 ];
@@ -87,7 +88,8 @@ function isFenceCloseLine(trimmed, fenceChar, fenceLen) {
 }
 
 function cdataStartsBeforeFence(line) {
-  const cdataIdx = line.toLowerCase().indexOf('<![cdata[');
+  const cdataOpen = findNextCDATAOpen(line, 0);
+  const cdataIdx = cdataOpen.ok ? cdataOpen.start : -1;
   if (cdataIdx < 0) return false;
   const fenceIdx = Math.min(
     line.indexOf('```') >= 0 ? line.indexOf('```') : Infinity,
@@ -102,15 +104,16 @@ function updateCDATAStateLine(inCDATA, line) {
   let state = inCDATA;
   while (pos < lower.length) {
     if (state) {
-      const end = lower.indexOf(']]>', pos);
+      const cdataEnd = findCDATAEnd(lower, pos);
+      const end = cdataEnd.index;
       if (end < 0) return true;
-      pos = end + ']]>'.length;
+      pos = end + cdataEnd.len;
       state = false;
       continue;
     }
-    const start = lower.indexOf('<![cdata[', pos);
-    if (start < 0) return false;
-    pos = start + '<![cdata['.length;
+    const start = findNextCDATAOpen(line, pos);
+    if (!start.ok) return false;
+    pos = start.bodyStart;
     state = true;
   }
   return state;
@@ -252,8 +255,9 @@ function replaceDSMLToolMarkupOutsideIgnored(text) {
     const tag = scanToolMarkupTagAt(raw, i);
     if (tag) {
       if (tag.dsmlLike) {
-        out += `<${tag.closing ? '/' : ''}${tag.name}${raw.slice(tag.nameEnd, tag.end + 1)}`;
-        if (raw[tag.end] !== '>') {
+        const tail = normalizeToolMarkupTagTailForXML(raw.slice(tag.nameEnd, tag.end + 1));
+        out += `<${tag.closing ? '/' : ''}${tag.name}${tail}`;
+        if (!tail.endsWith('>')) {
           out += '>';
         }
       } else {
@@ -408,12 +412,14 @@ function findMatchingXmlEndTagOutsideCDATA(text, tag, from) {
 }
 
 function skipXmlIgnoredSection(lower, i) {
-  if (lower.startsWith('<![cdata[', i)) {
-    const end = lower.indexOf(']]>', i + '<![cdata['.length);
+  const cdataOpen = matchCDATAOpenAt(lower, i);
+  if (cdataOpen.ok) {
+    const cdataEnd = findCDATAEnd(lower, cdataOpen.bodyStart);
+    const end = cdataEnd.index;
     if (end < 0) {
       return { advanced: false, blocked: true, next: i };
     }
-    return { advanced: true, blocked: false, next: end + ']]>'.length };
+    return { advanced: true, blocked: false, next: end + cdataEnd.len };
   }
   if (lower.startsWith('<!--', i)) {
     const end = lower.indexOf('-->', i + '<!--'.length);
@@ -425,44 +431,112 @@ function skipXmlIgnoredSection(lower, i) {
   return { advanced: false, blocked: false, next: i };
 }
 
+function findNextCDATAOpen(text, from) {
+  const raw = toStringSafe(text);
+  for (let i = Math.max(0, from || 0); i < raw.length; i += 1) {
+    if (normalizeFullwidthASCIIChar(raw[i]) !== '<') {
+      continue;
+    }
+    const open = matchCDATAOpenAt(raw, i);
+    if (open.ok) {
+      return { ok: true, start: i, bodyStart: open.bodyStart };
+    }
+  }
+  return { ok: false, start: -1, bodyStart: -1 };
+}
+
+function matchCDATAOpenAt(text, start) {
+  const raw = toStringSafe(text);
+  if (start < 0 || start >= raw.length || normalizeFullwidthASCIIChar(raw[start]) !== '<') {
+    return { ok: false, bodyStart: start };
+  }
+  let i = start + 1;
+  for (let skipped = 0; skipped <= 4 && i < raw.length; skipped += 1) {
+    const matched = matchNormalizedASCII(raw, i, '[cdata[');
+    if (matched.ok) {
+      return { ok: true, bodyStart: i + matched.len };
+    }
+    if (!isCDATAOpenSeparator(raw[i])) {
+      break;
+    }
+    i += 1;
+  }
+  return { ok: false, bodyStart: start };
+}
+
+function isCDATAOpenSeparator(ch) {
+  return isToolMarkupSeparator(ch);
+}
+
+function findCDATAEnd(text, from) {
+  const ascii = text.indexOf(']]>', from);
+  const fullwidth = text.indexOf(']]＞', from);
+  const cjk = text.indexOf(']]〉', from);
+  if (ascii < 0 && fullwidth < 0 && cjk < 0) {
+    return { index: -1, len: 0 };
+  }
+  let best = { index: -1, len: 0 };
+  for (const candidate of [
+    { index: ascii, len: ']]>'.length },
+    { index: fullwidth, len: ']]＞'.length },
+    { index: cjk, len: ']]〉'.length },
+  ]) {
+    if (candidate.index >= 0 && (best.index < 0 || candidate.index < best.index)) {
+      best = candidate;
+    }
+  }
+  return best;
+}
+
 function scanToolMarkupTagAt(text, start) {
   const raw = toStringSafe(text);
-  if (!raw || start < 0 || start >= raw.length || raw[start] !== '<') {
+  if (!raw || start < 0 || start >= raw.length || normalizeFullwidthASCIIChar(raw[start]) !== '<') {
     return null;
   }
   const lower = raw.toLowerCase();
   let i = start + 1;
-  while (i < raw.length && raw[i] === '<') {
+  while (i < raw.length && normalizeFullwidthASCIIChar(raw[i]) === '<') {
     i += 1;
   }
-  const closing = raw[i] === '/';
+  let closing = raw[i] === '/';
   if (closing) {
     i += 1;
   }
   const prefix = consumeToolMarkupNamePrefix(raw, lower, i);
+  const prefixStart = i;
   i = prefix.next;
-  const dsmlLike = prefix.dsmlLike;
-  const { name, len } = matchToolMarkupName(lower, i, dsmlLike);
+  let dsmlLike = prefix.dsmlLike;
+  let { name, len } = matchToolMarkupName(raw, i, dsmlLike);
   if (!name) {
-    return null;
+    const fallback = matchToolMarkupNameAfterArbitraryPrefix(raw, prefixStart);
+    if (!fallback.ok) {
+      return null;
+    }
+    if (!closing && toolMarkupPrefixContainsSlash(raw.slice(prefixStart, fallback.start))) {
+      closing = true;
+    }
+    name = fallback.name;
+    i = fallback.start;
+    len = fallback.len;
+    dsmlLike = true;
   }
   const originalNameEnd = i + len;
   let nameEnd = originalNameEnd;
-  while (nameEnd < raw.length && isToolMarkupPipe(raw[nameEnd])) {
+  while (nameEnd < raw.length && isToolMarkupSeparator(raw[nameEnd])) {
     nameEnd += 1;
   }
-  const hasTrailingPipe = nameEnd > originalNameEnd;
+  const hasTrailingSeparator = nameEnd > originalNameEnd;
   if (!hasXmlTagBoundary(raw, nameEnd)) {
     return null;
   }
   let end = findXmlTagEnd(raw, nameEnd);
   if (end < 0) {
-    if (!hasTrailingPipe) {
+    if (!hasTrailingSeparator) {
       return null;
     }
     end = nameEnd - 1;
   }
-  if (hasTrailingPipe) {
+  if (hasTrailingSeparator) {
     const nextLT = raw.indexOf('<', nameEnd);
     if (nextLT >= 0 && end >= nextLT) {
       end = nameEnd - 1;
@@ -541,7 +615,7 @@ function findPartialToolMarkupStart(text) {
   }
   const start = includeDuplicateLeadingLessThan(raw, lastLT);
   const tail = raw.slice(start);
-  if (tail.includes('>')) {
+  if (tail.includes('>') || tail.includes('＞')) {
     return -1;
   }
   return isPartialToolMarkupTagPrefix(tail) ? start : -1;
@@ -555,8 +629,15 @@ function includeDuplicateLeadingLessThan(text, idx) {
   return out;
 }
 
-function isToolMarkupPipe(ch) {
-  return ch === '|' || ch === '｜';
+function isToolMarkupSeparator(ch) {
+  const normalized = normalizeFullwidthASCIIChar(ch || '');
+  if (!normalized || ['<', '>', '/', '=', '"', "'", '['].includes(normalized)) {
+    return false;
+  }
+  if ([' ', '\t', '\n', '\r'].includes(normalized)) {
+    return false;
+  }
+  return !/^[A-Za-z0-9]$/.test(normalized);
 }
 
 function isPartialToolMarkupTagPrefix(text) {
@@ -579,10 +660,13 @@ function isPartialToolMarkupTagPrefix(text) {
     if (i === raw.length) {
       return true;
     }
-    if (hasToolMarkupNamePrefix(lower.slice(i))) {
+    if (hasToolMarkupNamePrefix(raw, i)) {
       return true;
     }
-    if ('dsml'.startsWith(lower.slice(i))) {
+    if (normalizedASCIITailAt(raw, i).startsWith('dsml') || 'dsml'.startsWith(normalizedASCIITailAt(raw, i))) {
+      return true;
+    }
+    if (hasPartialToolMarkupNameAfterArbitraryPrefix(raw, i)) {
       return true;
     }
     const next = consumeToolMarkupNamePrefixOnce(raw, lower, i);
@@ -607,39 +691,174 @@ function consumeToolMarkupNamePrefix(raw, lower, idx) {
   }
 }
 
-function consumeToolMarkupNamePrefixOnce(raw, lower, idx) {
-  if (idx < raw.length && isToolMarkupPipe(raw[idx])) {
-    return { next: idx + 1, ok: true };
-  }
-  if (idx < raw.length && [' ', '\t', '\r', '\n'].includes(raw[idx])) {
-    return { next: idx + 1, ok: true };
-  }
-  if (lower.startsWith('dsml', idx)) {
-    let next = idx + 'dsml'.length;
-    if (next < raw.length && raw[next] === '-') {
-      next += 1;
+function matchToolMarkupNameAfterArbitraryPrefix(raw, start) {
+  for (let idx = start; idx < raw.length;) {
+    if (isToolMarkupTagTerminator(raw, idx)) {
+      return { ok: false };
     }
-    return { next, ok: true };
+    for (const name of TOOL_MARKUP_NAMES) {
+      const matched = matchNormalizedASCII(raw, idx, name.raw);
+      if (!matched.ok) continue;
+      if (!toolMarkupPrefixAllowsLocalNameAt(raw, start, idx)) continue;
+      return { ok: true, name: name.canonical, start: idx, len: matched.len };
+    }
+    idx += 1;
   }
-  return { next: idx, ok: false };
+  return { ok: false };
 }
 
-function hasToolMarkupNamePrefix(lowerTail) {
-  for (const name of TOOL_MARKUP_NAMES) {
-    if (lowerTail.startsWith(name.raw) || name.raw.startsWith(lowerTail)) {
+function hasPartialToolMarkupNameAfterArbitraryPrefix(raw, start) {
+  for (let idx = start; idx < raw.length;) {
+    if (isToolMarkupTagTerminator(raw, idx)) {
+      return false;
+    }
+    if (toolMarkupPrefixAllowsLocalNameAt(raw, start, idx) && hasToolMarkupNamePrefix(raw, idx)) {
+      return true;
+    }
+    if (toolMarkupPrefixAllowsLocalNameAt(raw, start, idx) && hasDSMLNamePrefixOrPartial(raw, idx)) {
+      return true;
+    }
+    idx += 1;
+  }
+  return toolMarkupPrefixAllowsLocalName(raw.slice(start));
+}
+
+function hasDSMLNamePrefixOrPartial(raw, start) {
+  const tail = normalizedASCIITailAt(raw, start);
+  return tail.startsWith('dsml') || 'dsml'.startsWith(tail);
+}
+
+function toolMarkupPrefixAllowsLocalName(prefix) {
+  if (!prefix) {
+    return false;
+  }
+  if (normalizedASCIITailAt(prefix, 0).includes('dsml')) {
+    return true;
+  }
+  if (/[="'"]/.test(prefix)) {
+    return false;
+  }
+  const previous = normalizeFullwidthASCIIChar(prefix[prefix.length - 1] || '');
+  return !/^[A-Za-z0-9]$/.test(previous);
+}
+
+function toolMarkupPrefixAllowsLocalNameAt(raw, start, localStart) {
+  if (start < 0 || localStart <= start || localStart > raw.length) {
+    return false;
+  }
+  const prefix = raw.slice(start, localStart);
+  if (toolMarkupPrefixAllowsLocalName(prefix)) {
+    return true;
+  }
+  if (/[="'"]/.test(prefix)) {
+    return false;
+  }
+  const previous = normalizeFullwidthASCIIChar(prefix[prefix.length - 1] || '');
+  const next = normalizeFullwidthASCIIChar(raw[localStart] || '');
+  return /^[A-Za-z0-9]$/.test(previous) && /^[A-Z]$/.test(next);
+}
+
+function toolMarkupPrefixContainsSlash(prefix) {
+  for (const ch of toStringSafe(prefix)) {
+    if (normalizeFullwidthASCIIChar(ch) === '/') {
       return true;
     }
   }
   return false;
 }
 
-function matchToolMarkupName(lower, start, dsmlLike) {
+function isToolMarkupTagTerminator(raw, idx) {
+  return raw[idx] === '>' || normalizeFullwidthASCIIChar(raw[idx] || '') === '>';
+}
+
+function consumeToolMarkupNamePrefixOnce(raw, lower, idx) {
+  if (idx < raw.length && isToolMarkupSeparator(raw[idx])) {
+    return { next: idx + 1, ok: true };
+  }
+  if (idx < raw.length && [' ', '\t', '\r', '\n'].includes(raw[idx])) {
+    return { next: idx + 1, ok: true };
+  }
+  const dsml = matchNormalizedASCII(raw, idx, 'dsml');
+  if (dsml.ok) {
+    let next = idx + dsml.len;
+    const sep = normalizeFullwidthASCIIChar(raw[next] || '');
+    if (next < raw.length && (sep === '-' || sep === '_')) {
+      next += 1;
+    }
+    return { next, ok: true };
+  }
+  const arbitrary = consumeArbitraryToolMarkupNamePrefix(raw, lower, idx);
+  if (arbitrary.ok) {
+    return arbitrary;
+  }
+  return { next: idx, ok: false };
+}
+
+function consumeArbitraryToolMarkupNamePrefix(raw, lower, idx) {
+  const first = consumeToolMarkupPrefixSegment(raw, idx);
+  if (!first.ok) {
+    return { next: idx, ok: false };
+  }
+  let j = first.next;
+  while (j < raw.length) {
+    const segment = consumeToolMarkupPrefixSegment(raw, j);
+    if (!segment.ok) break;
+    j = segment.next;
+  }
+  let k = j;
+  while (k < raw.length && [' ', '\t', '\r', '\n'].includes(raw[k])) {
+    k += 1;
+  }
+  let next = k;
+  let ok = false;
+  if (next < raw.length && isToolMarkupSeparator(raw[next])) {
+    next += 1;
+    ok = true;
+  } else if (next < raw.length && ['_', '-'].includes(normalizeFullwidthASCIIChar(raw[next]))) {
+    next += 1;
+    ok = true;
+  }
+  if (!ok) {
+    return { next: idx, ok: false };
+  }
+  while (next < raw.length && [' ', '\t', '\r', '\n'].includes(raw[next])) {
+    next += 1;
+  }
+  if (!hasToolMarkupNamePrefix(raw, next)) {
+    return { next: idx, ok: false };
+  }
+  return { next, ok: true };
+}
+
+function consumeToolMarkupPrefixSegment(raw, idx) {
+  if (idx < 0 || idx >= raw.length) {
+    return { next: idx, ok: false };
+  }
+  const ch = normalizeFullwidthASCIIChar(raw[idx]);
+  if (/^[A-Za-z0-9]$/.test(ch)) {
+    return { next: idx + 1, ok: true };
+  }
+  return { next: idx, ok: false };
+}
+
+function hasToolMarkupNamePrefix(raw, start) {
+  const tail = normalizedASCIITailAt(raw, start);
+  for (const name of TOOL_MARKUP_NAMES) {
+    if (tail.startsWith(name.raw) || name.raw.startsWith(tail)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+function matchToolMarkupName(raw, start, dsmlLike) {
   for (const name of TOOL_MARKUP_NAMES) {
     if (name.dsmlOnly && !dsmlLike) {
       continue;
     }
-    if (lower.startsWith(name.raw, start)) {
-      return { name: name.canonical, len: name.raw.length };
+    const matched = matchNormalizedASCII(raw, start, name.raw);
+    if (matched.ok) {
+      return { name: name.canonical, len: matched.len };
     }
   }
   return { name: '', len: 0 };
@@ -649,17 +868,18 @@ function findXmlTagEnd(text, from) {
   let quote = '';
   for (let i = Math.max(0, from || 0); i < text.length; i += 1) {
     const ch = text[i];
+    const normalized = normalizeFullwidthASCIIChar(ch);
     if (quote) {
-      if (ch === quote) {
+      if (normalized === quote) {
         quote = '';
       }
       continue;
     }
-    if (ch === '"' || ch === "'") {
-      quote = ch;
+    if (normalized === '"' || normalized === "'") {
+      quote = normalized;
       continue;
     }
-    if (ch === '>') {
+    if (normalized === '>') {
       return i;
     }
   }
@@ -670,13 +890,96 @@ function hasXmlTagBoundary(text, idx) {
   if (idx >= text.length) {
     return true;
   }
-  return [' ', '\t', '\n', '\r', '>', '/'].includes(text[idx]);
+  return [' ', '\t', '\n', '\r', '>', '/'].includes(text[idx])
+    || normalizeFullwidthASCIIChar(text[idx]) === '>';
 }
 
 function isSelfClosingXmlTag(startTag) {
   return toStringSafe(startTag).trim().endsWith('/');
 }
 
+function normalizeFullwidthASCIIChar(ch) {
+  if (!ch) {
+    return ch;
+  }
+  if (ch === '〈') {
+    return '<';
+  }
+  if (ch === '〉') {
+    return '>';
+  }
+  if (ch === '“' || ch === '”') {
+    return '"';
+  }
+  if (ch === '‘' || ch === '’') {
+    return "'";
+  }
+  const code = ch.charCodeAt(0);
+  if (code >= 0xff01 && code <= 0xff5e) {
+    return String.fromCharCode(code - 0xfee0);
+  }
+  return ch;
+}
+
+function normalizedASCIITailAt(raw, start) {
+  let out = '';
+  for (let i = Math.max(0, start || 0); i < raw.length; i += 1) {
+    const ch = normalizeFullwidthASCIIChar(raw[i]).toLowerCase();
+    if (ch.charCodeAt(0) > 0x7f) {
+      break;
+    }
+    out += ch;
+  }
+  return out;
+}
+
+function matchNormalizedASCII(raw, start, expected) {
+  let idx = start;
+  for (let j = 0; j < expected.length; j += 1) {
+    if (idx >= raw.length) {
+      return { ok: false, len: 0 };
+    }
+    const ch = normalizeFullwidthASCIIChar(raw[idx]).toLowerCase();
+    if (ch !== expected[j].toLowerCase()) {
+      return { ok: false, len: 0 };
+    }
+    idx += 1;
+  }
+  return { ok: true, len: idx - start };
+}
+
+function normalizeToolMarkupTagTailForXML(tail) {
+  let out = '';
+  const raw = typeof tail === 'string' ? tail : String(tail || '');
+  let quote = '';
+  for (let i = 0; i < raw.length; i += 1) {
+    const ch = raw[i];
+    const normalized = normalizeFullwidthASCIIChar(ch);
+    if (quote) {
+      out += normalized;
+      if (normalized === quote) {
+        quote = '';
+      }
+    } else if (normalized === '"' || normalized === "'") {
+      quote = normalized;
+      out += normalized;
+    } else if (normalized === '|' || normalized === '!') {
+      let j = i + 1;
+      while (j < raw.length && [' ', '\t', '\r', '\n'].includes(raw[j])) {
+        j += 1;
+      }
+      if (normalizeFullwidthASCIIChar(raw[j] || '') !== '>') {
+        out += normalized;
+      }
+    } else if (['>', '/', '='].includes(normalized)) {
+      out += normalized;
+    } else {
+      out += ch;
+    }
+  }
+  return out;
+}
+
 function parseMarkupInput(raw) {
   const s = toStringSafe(raw).trim();
   if (!s) {
@@ -1017,16 +1320,34 @@ function unescapeHtml(safe) {
 
 function extractStandaloneCDATA(inner) {
   const s = toStringSafe(inner).trim();
-  const cdataMatch = s.match(CDATA_PATTERN);
-  if (cdataMatch && cdataMatch[1] !== undefined) {
-    return { ok: true, value: cdataMatch[1] };
-  }
-  if (s.toLowerCase().startsWith('<![cdata[')) {
-    return { ok: true, value: s.slice('<![CDATA['.length) };
+  const open = matchCDATAOpenAt(s, 0);
+  if (open.ok) {
+    const close = findStandaloneCDATAEnd(s, open.bodyStart);
+    if (close.index < 0) {
+      return { ok: true, value: s.slice(open.bodyStart) };
+    }
+    return { ok: true, value: s.slice(open.bodyStart, close.index) };
   }
   return { ok: false, value: '' };
 }
 
+function findStandaloneCDATAEnd(text, from) {
+  const raw = toStringSafe(text);
+  let best = { index: -1, len: 0 };
+  for (let searchFrom = Math.max(0, from || 0); searchFrom < raw.length;) {
+    const close = findCDATAEnd(raw, searchFrom);
+    if (close.index < 0) {
+      break;
+    }
+    const closeEnd = close.index + close.len;
+    if (!raw.slice(closeEnd).trim()) {
+      best = close;
+    }
+    searchFrom = closeEnd;
+  }
+  return best;
+}
+
 function parseJSONLiteralValue(raw) {
   const s = toStringSafe(raw).trim();
   if (!s) {

internal/js/helpers/stream-tool-sieve/sieve-xml.js

@@ -1,5 +1,5 @@
 'use strict';
-const { parseToolCalls } = require('./parse');
+const { parseToolCallsDetailed } = require('./parse');
 const {
   findToolMarkupTagOutsideIgnored,
   findMatchingToolMarkupClose,
@@ -27,19 +27,30 @@ function consumeXMLToolCapture(captured, toolNames, trimWrappingJSONFence) {
     const xmlBlock = captured.slice(openTag.start, closeTag.end + 1);
     const prefixPart = captured.slice(0, openTag.start);
     const suffixPart = captured.slice(closeTag.end + 1);
-    const parsed = parseToolCalls(xmlBlock, toolNames);
-    if (Array.isArray(parsed) && parsed.length > 0) {
+    const parsed = parseToolCallsDetailed(xmlBlock, toolNames);
+    if (Array.isArray(parsed.calls) && parsed.calls.length > 0) {
       const trimmedFence = trimWrappingJSONFence(prefixPart, suffixPart);
       if (!best || openTag.start < best.start) {
         best = {
           start: openTag.start,
           prefix: trimmedFence.prefix,
-          calls: parsed,
+          calls: parsed.calls,
           suffix: trimmedFence.suffix,
         };
       }
       break;
     }
+    if (parsed.sawToolCallSyntax) {
+      if (!rejected || openTag.start < rejected.start) {
+        rejected = {
+          start: openTag.start,
+          prefix: prefixPart + xmlBlock,
+          suffix: suffixPart,
+        };
+      }
+      searchFrom = openTag.end + 1;
+      continue;
+    }
     if (!rejected || openTag.start < rejected.start) {
       rejected = {
         start: openTag.start,
@@ -69,16 +80,19 @@ function consumeXMLToolCapture(captured, toolNames, trimWrappingJSONFence) {
         const xmlBlock = '<tool_calls>' + captured.slice(invokeTag.start, closeTag.end + 1);
         const prefixPart = captured.slice(0, invokeTag.start);
         const suffixPart = captured.slice(closeTag.end + 1);
-        const parsed = parseToolCalls(xmlBlock, toolNames);
-        if (Array.isArray(parsed) && parsed.length > 0) {
+        const parsed = parseToolCallsDetailed(xmlBlock, toolNames);
+        if (Array.isArray(parsed.calls) && parsed.calls.length > 0) {
           const trimmedFence = trimWrappingJSONFence(prefixPart, suffixPart);
           return {
             ready: true,
             prefix: trimmedFence.prefix,
-            calls: parsed,
+            calls: parsed.calls,
             suffix: trimmedFence.suffix,
           };
         }
+        if (parsed.sawToolCallSyntax) {
+          return { ready: true, prefix: prefixPart + captured.slice(invokeTag.start, closeTag.end + 1), calls: [], suffix: suffixPart };
+        }
         return { ready: true, prefix: prefixPart + captured.slice(invokeTag.start, closeTag.end + 1), calls: [], suffix: suffixPart };
       }
     }

internal/prompt/tool_calls.go

@@ -16,6 +16,15 @@ var promptXMLTextEscaper = strings.NewReplacer(
 
 var promptXMLNamePattern = regexp.MustCompile(`^[A-Za-z_][A-Za-z0-9_.:-]*$`)
 
+const (
+	promptDSMLToolCallsOpen  = "<｜DSML｜tool_calls>"
+	promptDSMLToolCallsClose = "</｜DSML｜tool_calls>"
+	promptDSMLInvokeOpen     = "<｜DSML｜invoke"
+	promptDSMLInvokeClose    = "</｜DSML｜invoke>"
+	promptDSMLParameterOpen  = "<｜DSML｜parameter"
+	promptDSMLParameterClose = "</｜DSML｜parameter>"
+)
+
 // FormatToolCallsForPrompt renders a tool_calls slice into the prompt-visible
 // invoke/parameter history block used across adapters.
 func FormatToolCallsForPrompt(raw any) string {
@@ -38,7 +47,7 @@ func FormatToolCallsForPrompt(raw any) string {
 	if len(blocks) == 0 {
 		return ""
 	}
-	return "<|DSML|tool_calls>\n" + strings.Join(blocks, "\n") + "\n</|DSML|tool_calls>"
+	return promptDSMLToolCallsOpen + "\n" + strings.Join(blocks, "\n") + "\n" + promptDSMLToolCallsClose
 }
 
 // StringifyToolCallArguments normalizes tool arguments into a compact string
@@ -94,12 +103,12 @@ func formatToolCallForPrompt(call map[string]any) string {
 
 	parameters := formatToolCallParametersForPrompt(argsRaw)
 	if parameters == "" {
-		return `  <|DSML|invoke name="` + escapeXMLAttribute(name) + `"></|DSML|invoke>`
+		return `  ` + promptDSMLInvokeOpen + ` name="` + escapeXMLAttribute(name) + `">` + promptDSMLInvokeClose
 	}
 
-	return "  <|DSML|invoke name=\"" + escapeXMLAttribute(name) + "\">\n" +
+	return "  " + promptDSMLInvokeOpen + " name=\"" + escapeXMLAttribute(name) + "\">\n" +
 		parameters + "\n" +
-		"  </|DSML|invoke>"
+		"  " + promptDSMLInvokeClose
 }
 
 func formatToolCallParametersForPrompt(raw any) string {
@@ -113,7 +122,7 @@ func formatToolCallParametersForPrompt(raw any) string {
 	if strings.TrimSpace(fallback) == "" {
 		return ""
 	}
-	return "    <|DSML|parameter name=\"content\">" + renderPromptXMLText(fallback) + "</|DSML|parameter>"
+	return "    " + promptDSMLParameterOpen + " name=\"content\">" + renderPromptXMLText(fallback) + promptDSMLParameterClose
 }
 
 func renderPromptToolParameters(value any, indent string) (string, bool) {
@@ -149,9 +158,9 @@ func renderPromptToolParameters(value any, indent string) (string, bool) {
 		}
 		return strings.Join(lines, "\n"), true
 	case string:
-		return indent + `<|DSML|parameter name="content">` + renderPromptXMLText(v) + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="content">` + renderPromptXMLText(v) + promptDSMLParameterClose, true
 	default:
-		return indent + `<|DSML|parameter name="value">` + renderPromptXMLText(fmt.Sprint(v)) + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="value">` + renderPromptXMLText(fmt.Sprint(v)) + promptDSMLParameterClose, true
 	}
 }
 
@@ -162,29 +171,29 @@ func renderPromptParameterNode(name string, value any, indent string) (string, b
 	}
 	switch v := value.(type) {
 	case nil:
-		return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + `"></|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + `">` + promptDSMLParameterClose, true
 	case map[string]any:
 		body, ok := renderPromptToolXMLBody(v, indent+"  ")
 		if !ok {
 			return "", false
 		}
 		if strings.TrimSpace(body) == "" {
-			return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + `"></|DSML|parameter>`, true
+			return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + `">` + promptDSMLParameterClose, true
 		}
-		return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + "\">\n" + body + "\n" + indent + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + "\">\n" + body + "\n" + indent + promptDSMLParameterClose, true
 	case []any:
 		body, ok := renderPromptToolXMLArray(v, indent+"  ")
 		if !ok {
 			return "", false
 		}
 		if strings.TrimSpace(body) == "" {
-			return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + `"></|DSML|parameter>`, true
+			return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + `">` + promptDSMLParameterClose, true
 		}
-		return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + "\">\n" + body + "\n" + indent + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + "\">\n" + body + "\n" + indent + promptDSMLParameterClose, true
 	case string:
-		return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + `">` + renderPromptXMLText(v) + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + `">` + renderPromptXMLText(v) + promptDSMLParameterClose, true
 	default:
-		return indent + `<|DSML|parameter name="` + escapeXMLAttribute(trimmedName) + `">` + renderPromptXMLText(fmt.Sprint(v)) + `</|DSML|parameter>`, true
+		return indent + promptDSMLParameterOpen + ` name="` + escapeXMLAttribute(trimmedName) + `">` + renderPromptXMLText(fmt.Sprint(v)) + promptDSMLParameterClose, true
 	}
 }
 

internal/prompt/tool_calls_test.go

@@ -22,7 +22,7 @@ func TestFormatToolCallsForPromptDSML(t *testing.T) {
 	if got == "" {
 		t.Fatal("expected non-empty formatted tool calls")
 	}
-	if got != "<|DSML|tool_calls>\n  <|DSML|invoke name=\"search_web\">\n    <|DSML|parameter name=\"query\"><![CDATA[latest]]></|DSML|parameter>\n  </|DSML|invoke>\n</|DSML|tool_calls>" {
+	if got != "<｜DSML｜tool_calls>\n  <｜DSML｜invoke name=\"search_web\">\n    <｜DSML｜parameter name=\"query\"><![CDATA[latest]]></｜DSML｜parameter>\n  </｜DSML｜invoke>\n</｜DSML｜tool_calls>" {
 		t.Fatalf("unexpected formatted tool call DSML: %q", got)
 	}
 }
@@ -34,7 +34,7 @@ func TestFormatToolCallsForPromptEscapesXMLEntities(t *testing.T) {
 			"arguments": `{"q":"a < b && c > d"}`,
 		},
 	})
-	want := "<|DSML|tool_calls>\n  <|DSML|invoke name=\"search&lt;&amp;&gt;\">\n    <|DSML|parameter name=\"q\"><![CDATA[a < b && c > d]]></|DSML|parameter>\n  </|DSML|invoke>\n</|DSML|tool_calls>"
+	want := "<｜DSML｜tool_calls>\n  <｜DSML｜invoke name=\"search&lt;&amp;&gt;\">\n    <｜DSML｜parameter name=\"q\"><![CDATA[a < b && c > d]]></｜DSML｜parameter>\n  </｜DSML｜invoke>\n</｜DSML｜tool_calls>"
 	if got != want {
 		t.Fatalf("unexpected escaped tool call XML: %q", got)
 	}
@@ -50,7 +50,7 @@ func TestFormatToolCallsForPromptUsesCDATAForMultilineContent(t *testing.T) {
 			},
 		},
 	})
-	want := "<|DSML|tool_calls>\n  <|DSML|invoke name=\"write_file\">\n    <|DSML|parameter name=\"content\"><![CDATA[#!/bin/bash\nprintf \"hello\"\n]]></|DSML|parameter>\n    <|DSML|parameter name=\"path\"><![CDATA[script.sh]]></|DSML|parameter>\n  </|DSML|invoke>\n</|DSML|tool_calls>"
+	want := "<｜DSML｜tool_calls>\n  <｜DSML｜invoke name=\"write_file\">\n    <｜DSML｜parameter name=\"content\"><![CDATA[#!/bin/bash\nprintf \"hello\"\n]]></｜DSML｜parameter>\n    <｜DSML｜parameter name=\"path\"><![CDATA[script.sh]]></｜DSML｜parameter>\n  </｜DSML｜invoke>\n</｜DSML｜tool_calls>"
 	if got != want {
 		t.Fatalf("unexpected multiline cdata tool call XML: %q", got)
 	}

internal/promptcompat/message_normalize.go

@@ -4,6 +4,7 @@ import (
 	"strings"
 
 	"ds2api/internal/prompt"
+	"ds2api/internal/toolcall"
 )
 
 const assistantReasoningLabel = "reasoning_content"
@@ -62,6 +63,9 @@ func buildAssistantContentForPrompt(msg map[string]any) string {
 		reasoning = strings.TrimSpace(extractOpenAIReasoningContentFromMessage(msg["content"]))
 	}
 	toolHistory := prompt.FormatToolCallsForPrompt(msg["tool_calls"])
+	if toolHistory == "" {
+		content = normalizeAssistantToolMarkupContentForPrompt(content)
+	}
 	parts := make([]string, 0, 3)
 	if reasoning != "" {
 		parts = append(parts, formatPromptLabeledBlock(assistantReasoningLabel, reasoning))
@@ -82,6 +86,40 @@ func buildAssistantContentForPrompt(msg map[string]any) string {
 	}
 }
 
+func normalizeAssistantToolMarkupContentForPrompt(content string) string {
+	trimmed := strings.TrimSpace(content)
+	if trimmed == "" || !isStandaloneAssistantToolMarkupBlock(trimmed) {
+		return content
+	}
+	parsed := toolcall.ParseStandaloneToolCallsDetailed(trimmed, nil)
+	if len(parsed.Calls) == 0 {
+		return content
+	}
+	raw := make([]any, 0, len(parsed.Calls))
+	for _, call := range parsed.Calls {
+		raw = append(raw, map[string]any{
+			"name":  call.Name,
+			"input": call.Input,
+		})
+	}
+	if formatted := prompt.FormatToolCallsForPrompt(raw); formatted != "" {
+		return formatted
+	}
+	return content
+}
+
+func isStandaloneAssistantToolMarkupBlock(trimmed string) bool {
+	tag, ok := toolcall.FindToolMarkupTagOutsideIgnored(trimmed, 0)
+	if !ok || tag.Start != 0 || tag.Closing || tag.Name != "tool_calls" {
+		return false
+	}
+	closeTag, ok := toolcall.FindMatchingToolMarkupClose(trimmed, tag)
+	if !ok {
+		return false
+	}
+	return strings.TrimSpace(trimmed[closeTag.End+1:]) == ""
+}
+
 func normalizeOpenAIReasoningContentForPrompt(v any) string {
 	switch x := v.(type) {
 	case string:

internal/promptcompat/message_normalize_test.go

@@ -38,10 +38,10 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsAndToolResult(t *tes
 		t.Fatalf("expected 4 normalized messages with assistant tool history preserved, got %d", len(normalized))
 	}
 	assistantContent, _ := normalized[2]["content"].(string)
-	if !strings.Contains(assistantContent, "<|DSML|tool_calls>") {
+	if !strings.Contains(assistantContent, "<｜DSML｜tool_calls>") {
 		t.Fatalf("assistant tool history should be preserved in DSML form, got %q", assistantContent)
 	}
-	if !strings.Contains(assistantContent, `<|DSML|invoke name="get_weather">`) {
+	if !strings.Contains(assistantContent, `<｜DSML｜invoke name="get_weather">`) {
 		t.Fatalf("expected tool name in preserved history, got %q", assistantContent)
 	}
 	if !strings.Contains(normalized[3]["content"].(string), `"temp":18`) {
@@ -49,7 +49,7 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsAndToolResult(t *tes
 	}
 
 	prompt := util.MessagesPrepare(normalized)
-	if !strings.Contains(prompt, "<|DSML|tool_calls>") {
+	if !strings.Contains(prompt, "<｜DSML｜tool_calls>") {
 		t.Fatalf("expected preserved assistant tool history in prompt: %q", prompt)
 	}
 }
@@ -177,10 +177,10 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantMultipleToolCallsRemainSepara
 		t.Fatalf("expected assistant tool_call-only message preserved, got %#v", normalized)
 	}
 	content, _ := normalized[0]["content"].(string)
-	if strings.Count(content, "<|DSML|invoke name=") != 2 {
+	if strings.Count(content, "<｜DSML｜invoke name=") != 2 {
 		t.Fatalf("expected two preserved tool call blocks, got %q", content)
 	}
-	if !strings.Contains(content, `<|DSML|invoke name="search_web">`) || !strings.Contains(content, `<|DSML|invoke name="eval_javascript">`) {
+	if !strings.Contains(content, `<｜DSML｜invoke name="search_web">`) || !strings.Contains(content, `<｜DSML｜invoke name="eval_javascript">`) {
 		t.Fatalf("expected both tool names in preserved history, got %q", content)
 	}
 }
@@ -258,11 +258,47 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantNilContentDoesNotInjectNullLi
 	if strings.Contains(content, "null") {
 		t.Fatalf("expected no null literal injection, got %q", content)
 	}
-	if !strings.Contains(content, "<|DSML|tool_calls>") {
+	if !strings.Contains(content, "<｜DSML｜tool_calls>") {
 		t.Fatalf("expected assistant tool history in normalized content, got %q", content)
 	}
 }
 
+func TestNormalizeOpenAIMessagesForPrompt_CanonicalizesStandaloneAssistantToolMarkupContent(t *testing.T) {
+	raw := []any{
+		map[string]any{
+			"role": "assistant",
+			"content": `<！DSML！tool_calls>
+  <！DSML！invoke name=“Bash”>
+  <！DSML！parameter name=“command”><！[CDATA[lsof -i :4321 -t]]><！/DSML！parameter>
+  <！DSML！parameter name=“description”><！[CDATA[Verify port 4321 is free]]><！/DSML！parameter>
+  <！/DSML！invoke>
+  <！/DSML！tool_calls>`,
+		},
+	}
+
+	normalized := NormalizeOpenAIMessagesForPrompt(raw, "")
+	if len(normalized) != 1 {
+		t.Fatalf("expected one normalized assistant message, got %#v", normalized)
+	}
+	content, _ := normalized[0]["content"].(string)
+	for _, want := range []string{
+		"<｜DSML｜tool_calls>",
+		`<｜DSML｜invoke name="Bash">`,
+		`<｜DSML｜parameter name="command"><![CDATA[lsof -i :4321 -t]]></｜DSML｜parameter>`,
+		`<｜DSML｜parameter name="description"><![CDATA[Verify port 4321 is free]]></｜DSML｜parameter>`,
+		"</｜DSML｜tool_calls>",
+	} {
+		if !strings.Contains(content, want) {
+			t.Fatalf("expected canonicalized assistant tool markup to contain %q, got %q", want, content)
+		}
+	}
+	for _, bad := range []string{"<！DSML", "！tool_calls", "“", "”"} {
+		if strings.Contains(content, bad) {
+			t.Fatalf("expected malformed assistant tool markup to be removed from prompt history, found %q in %q", bad, content)
+		}
+	}
+}
+
 func TestNormalizeOpenAIMessagesForPrompt_DeveloperRoleMapsToSystem(t *testing.T) {
 	raw := []any{
 		map[string]any{"role": "developer", "content": "必须先走工具调用"},

internal/promptcompat/prompt_build_test.go

@@ -47,10 +47,10 @@ func TestBuildOpenAIFinalPrompt_HandlerPathIncludesToolRoundtripSemantics(t *tes
 	if !strings.Contains(finalPrompt, `"condition":"sunny"`) {
 		t.Fatalf("handler finalPrompt should preserve tool output content: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "<|DSML|tool_calls>") {
+	if !strings.Contains(finalPrompt, "<｜DSML｜tool_calls>") {
 		t.Fatalf("handler finalPrompt should preserve assistant tool history: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, `<|DSML|invoke name="get_weather">`) {
+	if !strings.Contains(finalPrompt, `<｜DSML｜invoke name="get_weather">`) {
 		t.Fatalf("handler finalPrompt should include tool name history: %q", finalPrompt)
 	}
 }
@@ -74,7 +74,7 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 	}
 
 	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "", false)
-	if !strings.Contains(finalPrompt, "Remember: The ONLY valid way to use tools is the <|DSML|tool_calls>...</|DSML|tool_calls> block at the end of your response.") {
+	if !strings.Contains(finalPrompt, "Remember: The ONLY valid way to use tools is the <｜DSML｜tool_calls>...</｜DSML｜tool_calls> 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") {

internal/promptcompat/responses_input_items_test.go

@@ -1,6 +1,9 @@
 package promptcompat
 
-import "testing"
+import (
+	"strings"
+	"testing"
+)
 
 func TestNormalizeResponsesInputItemPreservesAssistantReasoningContent(t *testing.T) {
 	item := map[string]any{
@@ -48,3 +51,44 @@ func TestNormalizeResponsesInputItemAssistantMessageWithReasoningBlocks(t *testi
 		t.Fatalf("expected content blocks preserved, got %#v", got["content"])
 	}
 }
+
+func TestNormalizeResponsesInputArrayMergesReasoningMessageIntoFunctionCallHistory(t *testing.T) {
+	input := []any{
+		map[string]any{
+			"type": "message",
+			"role": "assistant",
+			"content": []any{
+				map[string]any{"type": "reasoning", "text": "need fresh docs before answering"},
+			},
+		},
+		map[string]any{
+			"type":      "function_call",
+			"call_id":   "call_search",
+			"name":      "search_web",
+			"arguments": `{"query":"docs"}`,
+		},
+	}
+
+	got := NormalizeResponsesInputAsMessages(input)
+	if len(got) != 1 {
+		t.Fatalf("expected reasoning and function_call merged into one assistant message, got %#v", got)
+	}
+	msg, _ := got[0].(map[string]any)
+	if msg["role"] != "assistant" {
+		t.Fatalf("expected assistant message, got %#v", msg)
+	}
+	if msg["reasoning_content"] != "need fresh docs before answering" {
+		t.Fatalf("expected reasoning_content on tool-call message, got %#v", msg)
+	}
+	toolCalls, _ := msg["tool_calls"].([]any)
+	if len(toolCalls) != 1 {
+		t.Fatalf("expected one tool call, got %#v", msg["tool_calls"])
+	}
+	history := BuildOpenAIHistoryTranscript(got)
+	if !strings.Contains(history, "[reasoning_content]\nneed fresh docs before answering\n[/reasoning_content]") {
+		t.Fatalf("expected reasoning in history transcript, got %q", history)
+	}
+	if !strings.Contains(history, `<｜DSML｜invoke name="search_web">`) {
+		t.Fatalf("expected tool call in history transcript, got %q", history)
+	}
+}

internal/promptcompat/responses_input_normalize.go

@@ -61,19 +61,52 @@ func normalizeResponsesInputArray(items []any) []any {
 	out := make([]any, 0, len(items))
 	callNameByID := map[string]string{}
 	fallbackParts := make([]string, 0, len(items))
+	pendingAssistantReasoning := ""
 	flushFallback := func() {
 		if len(fallbackParts) == 0 {
 			return
 		}
+		if pendingAssistantReasoning != "" {
+			out = append(out, map[string]any{"role": "assistant", "reasoning_content": pendingAssistantReasoning})
+			pendingAssistantReasoning = ""
+		}
 		out = append(out, map[string]any{"role": "user", "content": strings.Join(fallbackParts, "\n")})
 		fallbackParts = fallbackParts[:0]
 	}
+	flushPendingReasoning := func() {
+		if pendingAssistantReasoning == "" {
+			return
+		}
+		out = append(out, map[string]any{"role": "assistant", "reasoning_content": pendingAssistantReasoning})
+		pendingAssistantReasoning = ""
+	}
 
 	for _, item := range items {
 		switch x := item.(type) {
 		case map[string]any:
 			if msg := normalizeResponsesInputItemWithState(x, callNameByID); msg != nil {
+				if reasoning := assistantReasoningOnlyContent(msg); reasoning != "" {
+					if pendingAssistantReasoning == "" {
+						pendingAssistantReasoning = reasoning
+					} else {
+						pendingAssistantReasoning += "\n" + reasoning
+					}
+					continue
+				}
+				if isAssistantToolCallMessage(msg) && pendingAssistantReasoning != "" {
+					if strings.TrimSpace(normalizeOpenAIReasoningContentForPrompt(msg["reasoning_content"])) == "" {
+						msg["reasoning_content"] = pendingAssistantReasoning
+					}
+					pendingAssistantReasoning = ""
+				} else {
+					flushPendingReasoning()
+				}
 				flushFallback()
+				if isAssistantToolCallMessage(msg) && len(out) > 0 {
+					if merged := mergeResponsesAssistantToolCalls(out[len(out)-1], msg); merged {
+						continue
+					}
+				}
 				out = append(out, msg)
 				continue
 			}
@@ -86,9 +119,55 @@ func normalizeResponsesInputArray(items []any) []any {
 			}
 		}
 	}
+	flushPendingReasoning()
 	flushFallback()
 	if len(out) == 0 {
 		return nil
 	}
 	return out
 }
+
+func assistantReasoningOnlyContent(msg map[string]any) string {
+	if !isAssistantMessage(msg) || isAssistantToolCallMessage(msg) {
+		return ""
+	}
+	if _, hasContent := msg["content"]; hasContent {
+		normalizedContent := strings.TrimSpace(NormalizeOpenAIContentForPrompt(msg["content"]))
+		reasoningFromContent := strings.TrimSpace(extractOpenAIReasoningContentFromMessage(msg["content"]))
+		if normalizedContent != "" && normalizedContent != reasoningFromContent {
+			return ""
+		}
+		if reasoningFromContent != "" {
+			return reasoningFromContent
+		}
+	}
+	return strings.TrimSpace(normalizeOpenAIReasoningContentForPrompt(msg["reasoning_content"]))
+}
+
+func isAssistantMessage(msg map[string]any) bool {
+	return strings.EqualFold(strings.TrimSpace(asString(msg["role"])), "assistant")
+}
+
+func isAssistantToolCallMessage(msg map[string]any) bool {
+	if !isAssistantMessage(msg) {
+		return false
+	}
+	toolCalls, ok := msg["tool_calls"].([]any)
+	return ok && len(toolCalls) > 0
+}
+
+func mergeResponsesAssistantToolCalls(prev any, next map[string]any) bool {
+	prevMsg, ok := prev.(map[string]any)
+	if !ok || !isAssistantToolCallMessage(prevMsg) || !isAssistantToolCallMessage(next) {
+		return false
+	}
+	prevCalls, _ := prevMsg["tool_calls"].([]any)
+	nextCalls, _ := next["tool_calls"].([]any)
+	prevMsg["tool_calls"] = append(prevCalls, nextCalls...)
+	if strings.TrimSpace(normalizeOpenAIReasoningContentForPrompt(prevMsg["reasoning_content"])) == "" {
+		if reasoning := strings.TrimSpace(normalizeOpenAIReasoningContentForPrompt(next["reasoning_content"])); reasoning != "" {
+			prevMsg["reasoning_content"] = reasoning
+		}
+	}
+	return true
+}

internal/server/router.go

@@ -22,6 +22,7 @@ import (
 	"ds2api/internal/httpapi/admin"
 	"ds2api/internal/httpapi/claude"
 	"ds2api/internal/httpapi/gemini"
+	"ds2api/internal/httpapi/ollama"
 	"ds2api/internal/httpapi/openai/chat"
 	"ds2api/internal/httpapi/openai/embeddings"
 	"ds2api/internal/httpapi/openai/files"
@@ -68,6 +69,7 @@ func NewApp() (*App, error) {
 	claudeHandler := &claude.Handler{Store: store, Auth: resolver, DS: dsClient, OpenAI: chatHandler, ChatHistory: chatHistoryStore}
 	geminiHandler := &gemini.Handler{Store: store, Auth: resolver, DS: dsClient, OpenAI: chatHandler, ChatHistory: chatHistoryStore}
 	adminHandler := &admin.Handler{Store: store, Pool: pool, DS: dsClient, OpenAI: chatHandler, ChatHistory: chatHistoryStore}
+	ollamaHandler := &ollama.Handler{Store: store}
 	webuiHandler := webui.NewHandler()
 
 	r := chi.NewRouter()
@@ -112,6 +114,7 @@ func NewApp() (*App, error) {
 	r.Post("/embeddings", embeddingsHandler.Embeddings)
 	claude.RegisterRoutes(r, claudeHandler)
 	gemini.RegisterRoutes(r, geminiHandler)
+	ollama.RegisterRoutes(r, ollamaHandler)
 	r.Route("/admin", func(ar chi.Router) {
 		admin.RegisterRoutes(ar, adminHandler)
 	})

internal/toolcall/tool_prompt.go

@@ -11,46 +11,55 @@ import "strings"
 func BuildToolCallInstructions(toolNames []string) string {
 	return `TOOL CALL FORMAT — FOLLOW EXACTLY:
 
-<|DSML|tool_calls>
-  <|DSML|invoke name="TOOL_NAME_HERE">
-    <|DSML|parameter name="PARAMETER_NAME"><![CDATA[PARAMETER_VALUE]]></|DSML|parameter>
-  </|DSML|invoke>
-</|DSML|tool_calls>
+<｜DSML｜tool_calls>
+  <｜DSML｜invoke name="TOOL_NAME_HERE">
+    <｜DSML｜parameter name="PARAMETER_NAME"><![CDATA[PARAMETER_VALUE]]></｜DSML｜parameter>
+  </｜DSML｜invoke>
+</｜DSML｜tool_calls>
 
 RULES:
-1) Use the <|DSML|tool_calls> wrapper format.
-2) Put one or more <|DSML|invoke> entries under a single <|DSML|tool_calls> root.
-3) Put the tool name in the invoke name attribute: <|DSML|invoke name="TOOL_NAME">.
+1) Use the <｜DSML｜tool_calls> wrapper format.
+2) Put one or more <｜DSML｜invoke> entries under a single <｜DSML｜tool_calls> root.
+3) Put the tool name in the invoke name attribute: <｜DSML｜invoke name="TOOL_NAME">.
+3a) Tag punctuation alphabet: ASCII < > / = " plus the fullwidth vertical bar ｜.
 4) All string values must use <![CDATA[...]]>, even short ones. This includes code, scripts, file contents, prompts, paths, names, and queries.
-5) Every top-level argument must be a <|DSML|parameter name="ARG_NAME">...</|DSML|parameter> node.
+5) Every top-level argument must be a <｜DSML｜parameter name="ARG_NAME">...</｜DSML｜parameter> node.
 6) Objects use nested XML elements inside the parameter body. Arrays may repeat <item> children.
 7) Numbers, booleans, and null stay plain text.
 8) Use only the parameter names in the tool schema. Do not invent fields.
-9) Do NOT wrap XML in markdown fences. Do NOT output explanations, role markers, or internal monologue.
-10) If you call a tool, the first non-whitespace characters of that tool block must be exactly <|DSML|tool_calls>.
-11) Never omit the opening <|DSML|tool_calls> tag, even if you already plan to close with </|DSML|tool_calls>.
-12) Compatibility note: the runtime also accepts the legacy XML tags <tool_calls> / <invoke> / <parameter>, but prefer the DSML-prefixed form above.
+9) Fill parameters with the actual values required for this call. Do not emit placeholder, blank, or whitespace-only parameters.
+10) If a required parameter value is unknown, ask the user or answer normally instead of outputting an empty tool call.
+11) For shell tools such as Bash / execute_command, the command/script must be inside the command parameter. Never call them with an empty command.
+12) Do NOT wrap XML in markdown fences. Do NOT output explanations, role markers, or internal monologue.
+13) If you call a tool, the first non-whitespace characters of that tool block must be exactly <｜DSML｜tool_calls>.
+14) Never omit the opening <｜DSML｜tool_calls> tag, even if you already plan to close with </｜DSML｜tool_calls>.
+15) Compatibility note: the runtime also accepts the legacy XML tags <tool_calls> / <invoke> / <parameter>, but prefer the DSML-prefixed form above.
 
 PARAMETER SHAPES:
-- string => <|DSML|parameter name="x"><![CDATA[value]]></|DSML|parameter>
-- object => <|DSML|parameter name="x"><field>...</field></|DSML|parameter>
-- array => <|DSML|parameter name="x"><item>...</item><item>...</item></|DSML|parameter>
-- number/bool/null => <|DSML|parameter name="x">plain_text</|DSML|parameter>
+- string => <｜DSML｜parameter name="x"><![CDATA[value]]></｜DSML｜parameter>
+- object => <｜DSML｜parameter name="x"><field>...</field></｜DSML｜parameter>
+- array => <｜DSML｜parameter name="x"><item>...</item><item>...</item></｜DSML｜parameter>
+- number/bool/null => <｜DSML｜parameter name="x">plain_text</｜DSML｜parameter>
 
 【WRONG — Do NOT do these】:
 
 Wrong 1 — mixed text after XML:
-  <|DSML|tool_calls>...</|DSML|tool_calls> I hope this helps.
+  <｜DSML｜tool_calls>...</｜DSML｜tool_calls> I hope this helps.
 Wrong 2 — Markdown code fences:
   ` + "```xml" + `
-  <|DSML|tool_calls>...</|DSML|tool_calls>
+  <｜DSML｜tool_calls>...</｜DSML｜tool_calls>
   ` + "```" + `
 Wrong 3 — missing opening wrapper:
-  <|DSML|invoke name="TOOL_NAME">...</|DSML|invoke>
-  </|DSML|tool_calls>
-
-Remember: The ONLY valid way to use tools is the <|DSML|tool_calls>...</|DSML|tool_calls> block at the end of your response.
+  <｜DSML｜invoke name="TOOL_NAME">...</｜DSML｜invoke>
+  </｜DSML｜tool_calls>
+Wrong 4 — empty parameters:
+  <｜DSML｜tool_calls>
+    <｜DSML｜invoke name="Bash">
+      <｜DSML｜parameter name="command"></｜DSML｜parameter>
+    </｜DSML｜invoke>
+  </｜DSML｜tool_calls>
 
+Remember: The ONLY valid way to use tools is the <｜DSML｜tool_calls>...</｜DSML｜tool_calls> block at the end of your response.
 ` + buildCorrectToolExamples(toolNames)
 }
 
@@ -141,21 +150,21 @@ func firstScriptExample(names []string) (promptToolExample, bool) {
 
 func renderToolExampleBlock(calls []promptToolExample) string {
 	var b strings.Builder
-	b.WriteString("<|DSML|tool_calls>\n")
+	b.WriteString("<｜DSML｜tool_calls>\n")
 	for _, call := range calls {
-		b.WriteString(`  <|DSML|invoke name="`)
+		b.WriteString(`  <｜DSML｜invoke name="`)
 		b.WriteString(call.name)
 		b.WriteString(`">` + "\n")
 		b.WriteString(indentPromptParameters(call.params, "    "))
-		b.WriteString("\n  </|DSML|invoke>\n")
+		b.WriteString("\n  </｜DSML｜invoke>\n")
 	}
-	b.WriteString("</|DSML|tool_calls>")
+	b.WriteString("</｜DSML｜tool_calls>")
 	return b.String()
 }
 
 func indentPromptParameters(body, indent string) string {
 	if strings.TrimSpace(body) == "" {
-		return indent + `<|DSML|parameter name="content"></|DSML|parameter>`
+		return indent + `<｜DSML｜parameter name="content"></｜DSML｜parameter>`
 	}
 	lines := strings.Split(body, "\n")
 	for i, line := range lines {
@@ -169,7 +178,7 @@ func indentPromptParameters(body, indent string) string {
 }
 
 func wrapParameter(name, inner string) string {
-	return `<|DSML|parameter name="` + name + `">` + inner + `</|DSML|parameter>`
+	return `<｜DSML｜parameter name="` + name + `">` + inner + `</｜DSML｜parameter>`
 }
 
 func exampleBasicParams(name string) (string, bool) {
@@ -195,7 +204,7 @@ func exampleBasicParams(name string) (string, bool) {
 	case "Edit":
 		return wrapParameter("file_path", promptCDATA("README.md")) + "\n" + wrapParameter("old_string", promptCDATA("foo")) + "\n" + wrapParameter("new_string", promptCDATA("bar")), true
 	case "MultiEdit":
-		return wrapParameter("file_path", promptCDATA("README.md")) + "\n" + `<|DSML|parameter name="edits"><item><old_string>` + promptCDATA("foo") + `</old_string><new_string>` + promptCDATA("bar") + `</new_string></item></|DSML|parameter>`, true
+		return wrapParameter("file_path", promptCDATA("README.md")) + "\n" + `<｜DSML｜parameter name="edits"><item><old_string>` + promptCDATA("foo") + `</old_string><new_string>` + promptCDATA("bar") + `</new_string></item></｜DSML｜parameter>`, true
 	}
 	return "", false
 }
@@ -203,11 +212,11 @@ func exampleBasicParams(name string) (string, bool) {
 func exampleNestedParams(name string) (string, bool) {
 	switch strings.TrimSpace(name) {
 	case "MultiEdit":
-		return wrapParameter("file_path", promptCDATA("README.md")) + "\n" + `<|DSML|parameter name="edits"><item><old_string>` + promptCDATA("foo") + `</old_string><new_string>` + promptCDATA("bar") + `</new_string></item></|DSML|parameter>`, true
+		return wrapParameter("file_path", promptCDATA("README.md")) + "\n" + `<｜DSML｜parameter name="edits"><item><old_string>` + promptCDATA("foo") + `</old_string><new_string>` + promptCDATA("bar") + `</new_string></item></｜DSML｜parameter>`, true
 	case "Task":
 		return wrapParameter("description", promptCDATA("Investigate flaky tests")) + "\n" + wrapParameter("prompt", promptCDATA("Run targeted tests and summarize failures")), true
 	case "ask_followup_question":
-		return wrapParameter("question", promptCDATA("Which approach do you prefer?")) + "\n" + `<|DSML|parameter name="follow_up"><item><text>` + promptCDATA("Option A") + `</text></item><item><text>` + promptCDATA("Option B") + `</text></item></|DSML|parameter>`, true
+		return wrapParameter("question", promptCDATA("Which approach do you prefer?")) + "\n" + `<｜DSML｜parameter name="follow_up"><item><text>` + promptCDATA("Option A") + `</text></item><item><text>` + promptCDATA("Option B") + `</text></item></｜DSML｜parameter>`, true
 	}
 	return "", false
 }

internal/toolcall/tool_prompt_test.go

@@ -7,20 +7,20 @@ import (
 
 func TestBuildToolCallInstructions_ExecCommandUsesCmdExample(t *testing.T) {
 	out := BuildToolCallInstructions([]string{"exec_command"})
-	if !strings.Contains(out, `<|DSML|invoke name="exec_command">`) {
+	if !strings.Contains(out, `<｜DSML｜invoke name="exec_command">`) {
 		t.Fatalf("expected exec_command in examples, got: %s", out)
 	}
-	if !strings.Contains(out, `<|DSML|parameter name="cmd"><![CDATA[pwd]]></|DSML|parameter>`) {
+	if !strings.Contains(out, `<｜DSML｜parameter name="cmd"><![CDATA[pwd]]></｜DSML｜parameter>`) {
 		t.Fatalf("expected cmd parameter example for exec_command, got: %s", out)
 	}
 }
 
 func TestBuildToolCallInstructions_ExecuteCommandUsesCommandExample(t *testing.T) {
 	out := BuildToolCallInstructions([]string{"execute_command"})
-	if !strings.Contains(out, `<|DSML|invoke name="execute_command">`) {
+	if !strings.Contains(out, `<｜DSML｜invoke name="execute_command">`) {
 		t.Fatalf("expected execute_command in examples, got: %s", out)
 	}
-	if !strings.Contains(out, `<|DSML|parameter name="command"><![CDATA[pwd]]></|DSML|parameter>`) {
+	if !strings.Contains(out, `<｜DSML｜parameter name="command"><![CDATA[pwd]]></｜DSML｜parameter>`) {
 		t.Fatalf("expected command parameter example for execute_command, got: %s", out)
 	}
 }
@@ -34,20 +34,20 @@ func TestBuildToolCallInstructions_BashUsesCommandAndDescriptionExamples(t *test
 
 	sawDescription := false
 	for _, block := range blocks {
-		if !strings.Contains(block, `<|DSML|parameter name="command">`) {
+		if !strings.Contains(block, `<｜DSML｜parameter name="command">`) {
 			t.Fatalf("expected every Bash example to use command parameter, got: %s", block)
 		}
-		if strings.Contains(block, `<|DSML|parameter name="path">`) || strings.Contains(block, `<|DSML|parameter name="content">`) {
+		if strings.Contains(block, `<｜DSML｜parameter name="path">`) || strings.Contains(block, `<｜DSML｜parameter name="content">`) {
 			t.Fatalf("expected Bash examples not to use file write parameters, got: %s", block)
 		}
-		if strings.Contains(block, `<|DSML|parameter name="description">`) {
+		if strings.Contains(block, `<｜DSML｜parameter name="description">`) {
 			sawDescription = true
 		}
 	}
 	if !sawDescription {
 		t.Fatalf("expected Bash long-script example to include description, got: %s", out)
 	}
-	if strings.Contains(out, `<|DSML|invoke name="Read">`) {
+	if strings.Contains(out, `<｜DSML｜invoke name="Read">`) {
 		t.Fatalf("expected examples to avoid unavailable hard-coded Read tool, got: %s", out)
 	}
 }
@@ -60,10 +60,10 @@ func TestBuildToolCallInstructions_ExecuteCommandLongScriptUsesCommand(t *testin
 	}
 
 	for _, block := range blocks {
-		if !strings.Contains(block, `<|DSML|parameter name="command">`) {
+		if !strings.Contains(block, `<｜DSML｜parameter name="command">`) {
 			t.Fatalf("expected execute_command examples to use command parameter, got: %s", block)
 		}
-		if strings.Contains(block, `<|DSML|parameter name="path">`) || strings.Contains(block, `<|DSML|parameter name="content">`) {
+		if strings.Contains(block, `<｜DSML｜parameter name="path">`) || strings.Contains(block, `<｜DSML｜parameter name="content">`) {
 			t.Fatalf("expected execute_command examples not to use file write parameters, got: %s", block)
 		}
 	}
@@ -80,10 +80,10 @@ func TestBuildToolCallInstructions_ExecCommandLongScriptUsesCmd(t *testing.T) {
 	}
 
 	for _, block := range blocks {
-		if !strings.Contains(block, `<|DSML|parameter name="cmd">`) {
+		if !strings.Contains(block, `<｜DSML｜parameter name="cmd">`) {
 			t.Fatalf("expected exec_command examples to use cmd parameter, got: %s", block)
 		}
-		if strings.Contains(block, `<|DSML|parameter name="command">`) || strings.Contains(block, `<|DSML|parameter name="path">`) || strings.Contains(block, `<|DSML|parameter name="content">`) {
+		if strings.Contains(block, `<｜DSML｜parameter name="command">`) || strings.Contains(block, `<｜DSML｜parameter name="path">`) || strings.Contains(block, `<｜DSML｜parameter name="content">`) {
 			t.Fatalf("expected exec_command examples not to use command or file write parameters, got: %s", block)
 		}
 	}
@@ -100,10 +100,10 @@ func TestBuildToolCallInstructions_WriteUsesFilePathAndContent(t *testing.T) {
 	}
 
 	for _, block := range blocks {
-		if !strings.Contains(block, `<|DSML|parameter name="file_path">`) || !strings.Contains(block, `<|DSML|parameter name="content">`) {
+		if !strings.Contains(block, `<｜DSML｜parameter name="file_path">`) || !strings.Contains(block, `<｜DSML｜parameter name="content">`) {
 			t.Fatalf("expected Write examples to use file_path and content, got: %s", block)
 		}
-		if strings.Contains(block, `<|DSML|parameter name="path">`) {
+		if strings.Contains(block, `<｜DSML｜parameter name="path">`) {
 			t.Fatalf("expected Write examples not to use path, got: %s", block)
 		}
 	}
@@ -111,7 +111,7 @@ func TestBuildToolCallInstructions_WriteUsesFilePathAndContent(t *testing.T) {
 
 func TestBuildToolCallInstructions_AnchorsMissingOpeningWrapperFailureMode(t *testing.T) {
 	out := BuildToolCallInstructions([]string{"read_file"})
-	if !strings.Contains(out, "Never omit the opening <|DSML|tool_calls> tag") {
+	if !strings.Contains(out, "Never omit the opening <｜DSML｜tool_calls> tag") {
 		t.Fatalf("expected explicit missing-opening-tag warning, got: %s", out)
 	}
 	if !strings.Contains(out, "Wrong 3 — missing opening wrapper") {
@@ -119,8 +119,35 @@ func TestBuildToolCallInstructions_AnchorsMissingOpeningWrapperFailureMode(t *te
 	}
 }
 
+func TestBuildToolCallInstructions_RejectsEmptyParametersInPrompt(t *testing.T) {
+	out := BuildToolCallInstructions([]string{"Bash"})
+	for _, want := range []string{
+		"Do not emit placeholder, blank, or whitespace-only parameters.",
+		"If a required parameter value is unknown, ask the user or answer normally instead of outputting an empty tool call.",
+		"Never call them with an empty command.",
+		"Wrong 4 — empty parameters",
+	} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("expected empty-parameter instruction %q, got: %s", want, out)
+		}
+	}
+}
+
+func TestBuildToolCallInstructions_UsesPositiveTagPunctuationAlphabet(t *testing.T) {
+	out := BuildToolCallInstructions([]string{"Bash"})
+	want := `Tag punctuation alphabet: ASCII < > / = " plus the fullwidth vertical bar ｜.`
+	if !strings.Contains(out, want) {
+		t.Fatalf("expected positive tag punctuation alphabet %q, got: %s", want, out)
+	}
+	for _, bad := range []string{"lookalike", "substitute", "！", "〈", "〉", "“", "”", "、"} {
+		if strings.Contains(out, bad) {
+			t.Fatalf("tool prompt should not include negative punctuation examples %q, got: %s", bad, out)
+		}
+	}
+}
+
 func findInvokeBlocks(text, name string) []string {
-	open := `<|DSML|invoke name="` + name + `">`
+	open := `<｜DSML｜invoke name="` + name + `">`
 	remaining := text
 	blocks := []string{}
 	for {
@@ -129,11 +156,11 @@ func findInvokeBlocks(text, name string) []string {
 			return blocks
 		}
 		remaining = remaining[start:]
-		end := strings.Index(remaining, `</|DSML|invoke>`)
+		end := strings.Index(remaining, `</｜DSML｜invoke>`)
 		if end < 0 {
 			return blocks
 		}
-		end += len(`</|DSML|invoke>`)
+		end += len(`</｜DSML｜invoke>`)
 		blocks = append(blocks, remaining[:end])
 		remaining = remaining[end:]
 	}

internal/toolcall/toolcalls_dsml.go

@@ -1,6 +1,9 @@
 package toolcall
 
-import "strings"
+import (
+	"strings"
+	"unicode/utf8"
+)
 
 func normalizeDSMLToolCallMarkup(text string) (string, bool) {
 	if text == "" {
@@ -17,11 +20,10 @@ func rewriteDSMLToolMarkupOutsideIgnored(text string) string {
 	if text == "" {
 		return ""
 	}
-	lower := strings.ToLower(text)
 	var b strings.Builder
 	b.Grow(len(text))
 	for i := 0; i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			b.WriteString(text[i:])
 			break
@@ -43,8 +45,9 @@ func rewriteDSMLToolMarkupOutsideIgnored(text string) string {
 				b.WriteByte('/')
 			}
 			b.WriteString(tag.Name)
-			b.WriteString(text[tag.NameEnd : tag.End+1])
-			if text[tag.End] != '>' {
+			tail := normalizeToolMarkupTagTailForXML(text[tag.NameEnd : tag.End+1])
+			b.WriteString(tail)
+			if !strings.HasSuffix(tail, ">") {
 				b.WriteByte('>')
 			}
 			i = tag.End + 1
@@ -55,3 +58,57 @@ func rewriteDSMLToolMarkupOutsideIgnored(text string) string {
 	}
 	return b.String()
 }
+
+func normalizeToolMarkupTagTailForXML(tail string) string {
+	if tail == "" {
+		return ""
+	}
+	var b strings.Builder
+	b.Grow(len(tail))
+	quote := rune(0)
+	for i := 0; i < len(tail); {
+		r, size := utf8.DecodeRuneInString(tail[i:])
+		if r == utf8.RuneError && size == 1 {
+			b.WriteByte(tail[i])
+			i++
+			continue
+		}
+		ch := normalizeFullwidthASCII(r)
+		if quote != 0 {
+			b.WriteRune(ch)
+			if ch == quote {
+				quote = 0
+			}
+			i += size
+			continue
+		}
+		switch ch {
+		case '"', '\'':
+			quote = ch
+			b.WriteRune(ch)
+		case '|', '!':
+			j := i + size
+			for j < len(tail) {
+				next, nextSize := utf8.DecodeRuneInString(tail[j:])
+				if nextSize <= 0 {
+					break
+				}
+				if next == ' ' || next == '\t' || next == '\r' || next == '\n' {
+					j += nextSize
+					continue
+				}
+				break
+			}
+			next, _ := normalizedASCIIAt(tail, j)
+			if next != '>' {
+				b.WriteRune(ch)
+			}
+		case '>', '/', '=':
+			b.WriteRune(ch)
+		default:
+			b.WriteString(tail[i : i+size])
+		}
+		i += size
+	}
+	return b.String()
+}

internal/toolcall/toolcalls_markup.go

@@ -9,9 +9,6 @@ import (
 
 var toolCallMarkupKVPattern = regexp.MustCompile(`(?is)<(?:[a-z0-9_:-]+:)?([a-z0-9_\-.]+)\b[^>]*>(.*?)</(?:[a-z0-9_:-]+:)?([a-z0-9_\-.]+)>`)
 
-// cdataPattern matches a standalone CDATA section.
-var cdataPattern = regexp.MustCompile(`(?is)^<!\[CDATA\[(.*?)]]>$`)
-
 func parseMarkupKVObject(text string) map[string]any {
 	matches := toolCallMarkupKVPattern.FindAllStringSubmatch(strings.TrimSpace(text), -1)
 	if len(matches) == 0 {
@@ -108,15 +105,32 @@ func extractRawTagValue(inner string) string {
 
 func extractStandaloneCDATA(inner string) (string, bool) {
 	trimmed := strings.TrimSpace(inner)
-	if cdataMatches := cdataPattern.FindStringSubmatch(trimmed); len(cdataMatches) >= 2 {
-		return cdataMatches[1], true
-	}
-	if strings.HasPrefix(strings.ToLower(trimmed), "<![cdata[") {
-		return trimmed[len("<![CDATA["):], true
+	if bodyStart, ok := matchToolCDATAOpenAt(trimmed, 0); ok {
+		end := findStandaloneCDATAEnd(trimmed, bodyStart)
+		if end < 0 {
+			return trimmed[bodyStart:], true
+		}
+		return trimmed[bodyStart:end], true
 	}
 	return "", false
 }
 
+func findStandaloneCDATAEnd(text string, from int) int {
+	end := -1
+	for searchFrom := from; searchFrom < len(text); {
+		next := indexToolCDATAClose(text, searchFrom)
+		if next < 0 {
+			break
+		}
+		closeEnd := next + toolCDATACloseLenAt(text, next)
+		if strings.TrimSpace(text[closeEnd:]) == "" {
+			end = next
+		}
+		searchFrom = closeEnd
+	}
+	return end
+}
+
 func parseJSONLiteralValue(raw string) (any, bool) {
 	trimmed := strings.TrimSpace(raw)
 	if trimmed == "" {
@@ -145,7 +159,6 @@ func SanitizeLooseCDATA(text string) string {
 		return ""
 	}
 
-	lower := strings.ToLower(text)
 	const openMarker = "<![cdata["
 	const closeMarker = "]]>"
 
@@ -154,17 +167,16 @@ func SanitizeLooseCDATA(text string) string {
 	changed := false
 	pos := 0
 	for pos < len(text) {
-		startRel := strings.Index(lower[pos:], openMarker)
-		if startRel < 0 {
+		start := indexASCIIFold(text, pos, openMarker)
+		if start < 0 {
 			b.WriteString(text[pos:])
 			break
 		}
-		start := pos + startRel
 		contentStart := start + len(openMarker)
 		b.WriteString(text[pos:start])
 
-		if endRel := strings.Index(lower[contentStart:], closeMarker); endRel >= 0 {
-			end := contentStart + endRel + len(closeMarker)
+		if endRel := indexASCIIFold(text, contentStart, closeMarker); endRel >= 0 {
+			end := endRel + len(closeMarker)
 			b.WriteString(text[start:end])
 			pos = end
 			continue

internal/toolcall/toolcalls_parse.go

@@ -92,45 +92,11 @@ func filterToolCallsDetailed(parsed []ParsedToolCall) ([]ParsedToolCall, []strin
 		if tc.Input == nil {
 			tc.Input = map[string]any{}
 		}
-		if len(tc.Input) > 0 && !toolCallInputHasMeaningfulValue(tc.Input) {
-			continue
-		}
 		out = append(out, tc)
 	}
 	return out, nil
 }
 
-func toolCallInputHasMeaningfulValue(v any) bool {
-	switch x := v.(type) {
-	case nil:
-		return false
-	case string:
-		return strings.TrimSpace(x) != ""
-	case map[string]any:
-		if len(x) == 0 {
-			return false
-		}
-		for _, child := range x {
-			if toolCallInputHasMeaningfulValue(child) {
-				return true
-			}
-		}
-		return false
-	case []any:
-		if len(x) == 0 {
-			return false
-		}
-		for _, child := range x {
-			if toolCallInputHasMeaningfulValue(child) {
-				return true
-			}
-		}
-		return false
-	default:
-		return true
-	}
-}
-
 func looksLikeToolCallSyntax(text string) bool {
 	hasDSML, hasCanonical := ContainsToolCallWrapperSyntaxOutsideIgnored(text)
 	return hasDSML || hasCanonical
@@ -212,17 +178,16 @@ func firstFenceMarkerIndex(line string) int {
 }
 
 func updateCDATAStateForStrip(inCDATA bool, cdataFenceMarker, line string) (bool, string) {
-	lower := strings.ToLower(line)
 	pos := 0
 	state := inCDATA
 	fenceMarker := cdataFenceMarker
 	lineForFence := line
 	if !state {
-		start := strings.Index(lower[pos:], "<![cdata[")
+		start := indexASCIIFold(line, pos, "<![cdata[")
 		if start < 0 {
 			return false, ""
 		}
-		pos += start + len("<![cdata[")
+		pos = start + len("<![cdata[")
 		state = true
 		lineForFence = line[pos:]
 	}
@@ -239,24 +204,23 @@ func updateCDATAStateForStrip(inCDATA bool, cdataFenceMarker, line string) (bool
 		fenceMarker = ""
 	}
 
-	for pos < len(lower) {
-		end := strings.Index(lower[pos:], "]]>")
-		if end < 0 {
+	for pos < len(line) {
+		endPos := indexASCIIFold(line, pos, "]]>")
+		if endPos < 0 {
 			return true, fenceMarker
 		}
-		endPos := pos + end
 		pos = endPos + len("]]>")
 		if fenceMarker != "" {
 			continue
 		}
-		if cdataEndLooksStructural(lower, pos) || strings.TrimSpace(lower[pos:]) == "" {
+		if cdataEndLooksStructural(line, pos) || strings.TrimSpace(line[pos:]) == "" {
 			state = false
-			for pos < len(lower) {
-				start := strings.Index(lower[pos:], "<![cdata[")
+			for pos < len(line) {
+				start := indexASCIIFold(line, pos, "<![cdata[")
 				if start < 0 {
 					return false, ""
 				}
-				pos += start + len("<![cdata[")
+				pos = start + len("<![cdata[")
 				state = true
 				trimmedTail := strings.TrimLeft(line[pos:], " \t")
 				if marker, ok := parseFenceOpen(trimmedTail); ok {

internal/toolcall/toolcalls_parse_markup.go

@@ -6,6 +6,7 @@ import (
 	"html"
 	"regexp"
 	"strings"
+	"unicode/utf8"
 )
 
 var xmlAttrPattern = regexp.MustCompile(`(?is)\b([a-z0-9_:-]+)\s*=\s*("([^"]*)"|'([^']*)')`)
@@ -141,10 +142,9 @@ func findXMLElementBlocks(text, tag string) []xmlElementBlock {
 }
 
 func findXMLStartTagOutsideCDATA(text, tag string, from int) (start, bodyStart int, attrs string, ok bool) {
-	lower := strings.ToLower(text)
 	target := "<" + strings.ToLower(tag)
 	for i := maxInt(from, 0); i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			return -1, -1, "", false
 		}
@@ -152,7 +152,7 @@ func findXMLStartTagOutsideCDATA(text, tag string, from int) (start, bodyStart i
 			i = next
 			continue
 		}
-		if strings.HasPrefix(lower[i:], target) && hasXMLTagBoundary(text, i+len(target)) {
+		if hasASCIIPrefixFoldAt(text, i, target) && hasXMLTagBoundary(text, i+len(target)) {
 			end := findXMLTagEnd(text, i+len(target))
 			if end < 0 {
 				return -1, -1, "", false
@@ -165,12 +165,11 @@ func findXMLStartTagOutsideCDATA(text, tag string, from int) (start, bodyStart i
 }
 
 func findMatchingXMLEndTagOutsideCDATA(text, tag string, from int) (closeStart, closeEnd int, ok bool) {
-	lower := strings.ToLower(text)
 	openTarget := "<" + strings.ToLower(tag)
 	closeTarget := "</" + strings.ToLower(tag)
 	depth := 1
 	for i := maxInt(from, 0); i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			return -1, -1, false
 		}
@@ -178,7 +177,7 @@ func findMatchingXMLEndTagOutsideCDATA(text, tag string, from int) (closeStart,
 			i = next
 			continue
 		}
-		if strings.HasPrefix(lower[i:], closeTarget) && hasXMLTagBoundary(text, i+len(closeTarget)) {
+		if hasASCIIPrefixFoldAt(text, i, closeTarget) && hasXMLTagBoundary(text, i+len(closeTarget)) {
 			end := findXMLTagEnd(text, i+len(closeTarget))
 			if end < 0 {
 				return -1, -1, false
@@ -190,7 +189,7 @@ func findMatchingXMLEndTagOutsideCDATA(text, tag string, from int) (closeStart,
 			i = end + 1
 			continue
 		}
-		if strings.HasPrefix(lower[i:], openTarget) && hasXMLTagBoundary(text, i+len(openTarget)) {
+		if hasASCIIPrefixFoldAt(text, i, openTarget) && hasXMLTagBoundary(text, i+len(openTarget)) {
 			end := findXMLTagEnd(text, i+len(openTarget))
 			if end < 0 {
 				return -1, -1, false
@@ -206,16 +205,20 @@ func findMatchingXMLEndTagOutsideCDATA(text, tag string, from int) (closeStart,
 	return -1, -1, false
 }
 
-func skipXMLIgnoredSection(text, lower string, i int) (next int, advanced bool, blocked bool) {
-	switch {
-	case strings.HasPrefix(lower[i:], "<![cdata["):
-		end := findToolCDATAEnd(text, lower, i+len("<![cdata["))
+func skipXMLIgnoredSection(text string, i int) (next int, advanced bool, blocked bool) {
+	if i < 0 || i >= len(text) {
+		return i, false, false
+	}
+	if bodyStart, ok := matchToolCDATAOpenAt(text, i); ok {
+		end := findToolCDATAEnd(text, bodyStart)
 		if end < 0 {
 			return 0, false, true
 		}
-		return end + len("]]>"), true, false
-	case strings.HasPrefix(lower[i:], "<!--"):
-		end := strings.Index(lower[i+len("<!--"):], "-->")
+		return end + toolCDATACloseLenAt(text, end), true, false
+	}
+	switch {
+	case strings.HasPrefix(text[i:], "<!--"):
+		end := strings.Index(text[i+len("<!--"):], "-->")
 		if end < 0 {
 			return 0, false, true
 		}
@@ -225,43 +228,141 @@ func skipXMLIgnoredSection(text, lower string, i int) (next int, advanced bool,
 	}
 }
 
-func findToolCDATAEnd(text, lower string, from int) int {
-	if from < 0 || from > len(text) {
-		return -1
+func matchToolCDATAOpenAt(text string, start int) (int, bool) {
+	i, ok := consumeToolMarkupLessThan(text, start)
+	if !ok {
+		return start, false
 	}
-	const closeMarker = "]]>"
-	firstNonFenceEnd := -1
-	for searchFrom := from; searchFrom < len(text); {
-		rel := strings.Index(lower[searchFrom:], closeMarker)
-		if rel < 0 {
+	for skipped := 0; skipped <= 4 && i < len(text); skipped++ {
+		if cdataLen, ok := matchASCIIPrefixFoldAt(text, i, "[cdata["); ok {
+			return i + cdataLen, true
+		}
+		r, size := utf8.DecodeRuneInString(text[i:])
+		if size <= 0 || !isToolCDATAOpenSeparator(r) {
 			break
 		}
-		end := searchFrom + rel
-		searchFrom = end + len(closeMarker)
+		i += size
+	}
+	return start, false
+}
+
+func isToolCDATAOpenSeparator(r rune) bool {
+	return isToolMarkupSeparator(r)
+}
+
+func hasASCIIPrefixFoldAt(text string, start int, prefix string) bool {
+	_, ok := matchASCIIPrefixFoldAt(text, start, prefix)
+	return ok
+}
+
+func matchASCIIPrefixFoldAt(text string, start int, prefix string) (int, bool) {
+	if start < 0 || start >= len(text) && prefix != "" {
+		return 0, false
+	}
+	idx := start
+	for j := 0; j < len(prefix); j++ {
+		if idx >= len(text) {
+			return 0, false
+		}
+		ch, size := normalizedASCIIAt(text, idx)
+		if size <= 0 || asciiLower(ch) != asciiLower(prefix[j]) {
+			return 0, false
+		}
+		idx += size
+	}
+	return idx - start, true
+}
+
+func asciiLower(b byte) byte {
+	if b >= 'A' && b <= 'Z' {
+		return b + ('a' - 'A')
+	}
+	return b
+}
+
+// indexASCIIFold returns the absolute byte position in s where substr (ASCII-only) is
+// found case-insensitively, scanning forward from start. Returns -1 if not found.
+// Unlike strings.Index on a lowercased copy, this does not allocate or risk byte-length
+// mismatch when non-ASCII runes change width under case folding.
+func indexASCIIFold(s string, start int, substr string) int {
+	if start < 0 || len(s)-start < len(substr) {
+		return -1
+	}
+	end := len(s) - len(substr) + 1
+	for i := start; i < end; i++ {
+		if hasASCIIPrefixFoldAt(s, i, substr) {
+			return i
+		}
+	}
+	return -1
+}
+
+func findToolCDATAEnd(text string, from int) int {
+	if from < 0 || from >= len(text) {
+		return -1
+	}
+	firstNonFenceEnd := -1
+	for searchFrom := from; searchFrom < len(text); {
+		end := indexToolCDATAClose(text, searchFrom)
+		if end < 0 {
+			break
+		}
+		closeLen := toolCDATACloseLenAt(text, end)
+		searchFrom = end + closeLen
 		if cdataOffsetIsInsideMarkdownFence(text[from:end]) {
 			continue
 		}
+		if cdataEndLooksStructural(text, searchFrom) {
+			return end
+		}
 		if firstNonFenceEnd < 0 {
 			firstNonFenceEnd = end
 		}
-		if cdataEndLooksStructural(lower, searchFrom) {
-			return end
-		}
 	}
 	return firstNonFenceEnd
 }
 
-func cdataEndLooksStructural(lower string, after int) bool {
-	for after < len(lower) {
-		switch lower[after] {
-		case ' ', '\t', '\r', '\n':
-			after++
-			continue
-		default:
-		}
-		break
+func indexToolCDATAClose(text string, from int) int {
+	if from < 0 {
+		from = 0
 	}
-	return strings.HasPrefix(lower[after:], "</")
+	asciiIdx := strings.Index(text[from:], "]]>")
+	fullIdx := strings.Index(text[from:], "]]＞")
+	cjkIdx := strings.Index(text[from:], "]]〉")
+	if asciiIdx < 0 && fullIdx < 0 && cjkIdx < 0 {
+		return -1
+	}
+	best := -1
+	for _, idx := range []int{asciiIdx, fullIdx, cjkIdx} {
+		if idx >= 0 && (best < 0 || idx < best) {
+			best = idx
+		}
+	}
+	return from + best
+}
+
+func toolCDATACloseLenAt(text string, idx int) int {
+	if strings.HasPrefix(text[idx:], "]]〉") {
+		return len("]]〉")
+	}
+	if strings.HasPrefix(text[idx:], "]]＞") {
+		return len("]]＞")
+	}
+	return len("]]>")
+}
+
+func cdataEndLooksStructural(text string, after int) bool {
+	for after < len(text) {
+		switch {
+		case text[after] == ' ' || text[after] == '\t' || text[after] == '\r' || text[after] == '\n':
+			after++
+		case after+1 < len(text) && text[after] == '<' && text[after+1] == '/':
+			return true
+		default:
+			return false
+		}
+	}
+	return false
 }
 
 func cdataOffsetIsInsideMarkdownFence(fragment string) bool {
@@ -289,22 +390,29 @@ func cdataOffsetIsInsideMarkdownFence(fragment string) bool {
 }
 
 func findXMLTagEnd(text string, from int) int {
-	quote := byte(0)
-	for i := maxInt(from, 0); i < len(text); i++ {
-		ch := text[i]
+	quote := rune(0)
+	for i := maxInt(from, 0); i < len(text); {
+		r, size := utf8.DecodeRuneInString(text[i:])
+		if r == utf8.RuneError && size == 0 {
+			break
+		}
+		ch := normalizeFullwidthASCII(r)
 		if quote != 0 {
 			if ch == quote {
 				quote = 0
 			}
+			i += size
 			continue
 		}
 		if ch == '"' || ch == '\'' {
 			quote = ch
+			i += size
 			continue
 		}
 		if ch == '>' {
-			return i
+			return i + size - 1
 		}
+		i += size
 	}
 	return -1
 }
@@ -317,7 +425,8 @@ func hasXMLTagBoundary(text string, idx int) bool {
 	case ' ', '\t', '\n', '\r', '>', '/':
 		return true
 	default:
-		return false
+		r, _ := utf8.DecodeRuneInString(text[idx:])
+		return normalizeFullwidthASCII(r) == '>'
 	}
 }
 

internal/toolcall/toolcalls_scan.go

@@ -1,6 +1,9 @@
 package toolcall
 
-import "strings"
+import (
+	"strings"
+	"unicode/utf8"
+)
 
 type toolMarkupNameAlias struct {
 	raw       string
@@ -11,6 +14,7 @@ type toolMarkupNameAlias struct {
 var toolMarkupNames = []toolMarkupNameAlias{
 	{raw: "tool_calls", canonical: "tool_calls"},
 	{raw: "tool-calls", canonical: "tool_calls", dsmlOnly: true},
+	{raw: "toolcalls", canonical: "tool_calls", dsmlOnly: true},
 	{raw: "invoke", canonical: "invoke"},
 	{raw: "parameter", canonical: "parameter"},
 }
@@ -28,9 +32,8 @@ type ToolMarkupTag struct {
 }
 
 func ContainsToolMarkupSyntaxOutsideIgnored(text string) (hasDSML, hasCanonical bool) {
-	lower := strings.ToLower(text)
 	for i := 0; i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			return hasDSML, hasCanonical
 		}
@@ -56,9 +59,8 @@ func ContainsToolMarkupSyntaxOutsideIgnored(text string) (hasDSML, hasCanonical
 }
 
 func ContainsToolCallWrapperSyntaxOutsideIgnored(text string) (hasDSML, hasCanonical bool) {
-	lower := strings.ToLower(text)
 	for i := 0; i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			return hasDSML, hasCanonical
 		}
@@ -88,9 +90,8 @@ func ContainsToolCallWrapperSyntaxOutsideIgnored(text string) (hasDSML, hasCanon
 }
 
 func FindToolMarkupTagOutsideIgnored(text string, start int) (ToolMarkupTag, bool) {
-	lower := strings.ToLower(text)
 	for i := maxInt(start, 0); i < len(text); {
-		next, advanced, blocked := skipXMLIgnoredSection(text, lower, i)
+		next, advanced, blocked := skipXMLIgnoredSection(text, i)
 		if blocked {
 			return ToolMarkupTag{}, false
 		}
@@ -107,7 +108,7 @@ func FindToolMarkupTagOutsideIgnored(text string, start int) (ToolMarkupTag, boo
 }
 
 func FindMatchingToolMarkupClose(text string, open ToolMarkupTag) (ToolMarkupTag, bool) {
-	if text == "" || open.Name == "" || open.Closing {
+	if text == "" || open.Name == "" || open.Closing || open.End >= len(text) {
 		return ToolMarkupTag{}, false
 	}
 	depth := 1
@@ -134,41 +135,56 @@ func FindMatchingToolMarkupClose(text string, open ToolMarkupTag) (ToolMarkupTag
 }
 
 func scanToolMarkupTagAt(text string, start int) (ToolMarkupTag, bool) {
-	if start < 0 || start >= len(text) || text[start] != '<' {
+	next, ok := consumeToolMarkupLessThan(text, start)
+	if !ok {
 		return ToolMarkupTag{}, false
 	}
-	lower := strings.ToLower(text)
-	i := start + 1
-	for i < len(text) && text[i] == '<' {
-		i++
+	i := next
+	for {
+		next, ok := consumeToolMarkupLessThan(text, i)
+		if !ok {
+			break
+		}
+		i = next
 	}
 	closing := false
 	if i < len(text) && text[i] == '/' {
 		closing = true
 		i++
 	}
-	i, dsmlLike := consumeToolMarkupNamePrefix(lower, text, i)
-	name, nameLen := matchToolMarkupName(lower, i, dsmlLike)
+	prefixStart := i
+	i, dsmlLike := consumeToolMarkupNamePrefix(text, i)
+	name, nameLen := matchToolMarkupName(text, i, dsmlLike)
 	if nameLen == 0 {
-		return ToolMarkupTag{}, false
+		fallbackName, fallbackStart, fallbackLen, ok := matchToolMarkupNameAfterArbitraryPrefix(text, prefixStart)
+		if !ok {
+			return ToolMarkupTag{}, false
+		}
+		if !closing && toolMarkupPrefixContainsSlash(text[prefixStart:fallbackStart]) {
+			closing = true
+		}
+		name = fallbackName
+		i = fallbackStart
+		nameLen = fallbackLen
+		dsmlLike = true
 	}
 	nameEnd := i + nameLen
-	nameEndBeforePipes := nameEnd
-	for next, ok := consumeToolMarkupPipe(text, nameEnd); ok; next, ok = consumeToolMarkupPipe(text, nameEnd) {
+	nameEndBeforeSeparators := nameEnd
+	for next, ok := consumeToolMarkupSeparator(text, nameEnd); ok; next, ok = consumeToolMarkupSeparator(text, nameEnd) {
 		nameEnd = next
 	}
-	hasTrailingPipe := nameEnd > nameEndBeforePipes
+	hasTrailingSeparator := nameEnd > nameEndBeforeSeparators
 	if !hasToolMarkupBoundary(text, nameEnd) {
 		return ToolMarkupTag{}, false
 	}
 	end := findXMLTagEnd(text, nameEnd)
 	if end < 0 {
-		if !hasTrailingPipe {
+		if !hasTrailingSeparator {
 			return ToolMarkupTag{}, false
 		}
 		end = nameEnd - 1
 	}
-	if hasTrailingPipe {
+	if hasTrailingSeparator {
 		if nextLT := strings.IndexByte(text[nameEnd:], '<'); nextLT >= 0 && end >= nameEnd+nextLT {
 			end = nameEnd - 1
 		}
@@ -188,10 +204,9 @@ func scanToolMarkupTagAt(text string, start int) (ToolMarkupTag, bool) {
 }
 
 func IsPartialToolMarkupTagPrefix(text string) bool {
-	if text == "" || text[0] != '<' || strings.Contains(text, ">") {
+	if text == "" || text[0] != '<' || strings.Contains(text, ">") || strings.Contains(text, "＞") {
 		return false
 	}
-	lower := strings.ToLower(text)
 	i := 1
 	for i < len(text) && text[i] == '<' {
 		i++
@@ -206,13 +221,16 @@ func IsPartialToolMarkupTagPrefix(text string) bool {
 		if i == len(text) {
 			return true
 		}
-		if hasToolMarkupNamePrefix(lower[i:]) {
+		if hasToolMarkupNamePrefix(text, i) {
 			return true
 		}
-		if strings.HasPrefix("dsml", lower[i:]) {
+		if hasASCIIPartialPrefixFoldAt(text, i, "dsml") {
 			return true
 		}
-		next, ok := consumeToolMarkupNamePrefixOnce(lower, text, i)
+		if hasPartialToolMarkupNameAfterArbitraryPrefix(text, i) {
+			return true
+		}
+		next, ok := consumeToolMarkupNamePrefixOnce(text, i)
 		if !ok {
 			return false
 		}
@@ -221,10 +239,10 @@ func IsPartialToolMarkupTagPrefix(text string) bool {
 	return false
 }
 
-func consumeToolMarkupNamePrefix(lower, text string, idx int) (int, bool) {
+func consumeToolMarkupNamePrefix(text string, idx int) (int, bool) {
 	dsmlLike := false
 	for {
-		next, ok := consumeToolMarkupNamePrefixOnce(lower, text, idx)
+		next, ok := consumeToolMarkupNamePrefixOnce(text, idx)
 		if !ok {
 			return idx, dsmlLike
 		}
@@ -233,55 +251,262 @@ func consumeToolMarkupNamePrefix(lower, text string, idx int) (int, bool) {
 	}
 }
 
-func consumeToolMarkupNamePrefixOnce(lower, text string, idx int) (int, bool) {
-	if next, ok := consumeToolMarkupPipe(text, idx); ok {
+func consumeToolMarkupNamePrefixOnce(text string, idx int) (int, bool) {
+	if next, ok := consumeToolMarkupSeparator(text, idx); ok {
 		return next, true
 	}
 	if idx < len(text) && (text[idx] == ' ' || text[idx] == '\t' || text[idx] == '\r' || text[idx] == '\n') {
 		return idx + 1, true
 	}
-	if strings.HasPrefix(lower[idx:], "dsml") {
-		next := idx + len("dsml")
-		if next < len(text) && (text[next] == '-' || text[next] == '_') {
-			next++
+	if hasASCIIPrefixFoldAt(text, idx, "dsml") {
+		dsmlLen, _ := matchASCIIPrefixFoldAt(text, idx, "dsml")
+		next := idx + dsmlLen
+		if sep, size := normalizedASCIIAt(text, next); sep == '-' || sep == '_' {
+			next += size
 		}
 		return next, true
 	}
+	if next, ok := consumeArbitraryToolMarkupNamePrefix(text, idx); ok {
+		return next, true
+	}
 	return idx, false
 }
 
-func hasToolMarkupNamePrefix(lowerTail string) bool {
+func consumeArbitraryToolMarkupNamePrefix(text string, idx int) (int, bool) {
+	nextSegment, ok := consumeToolMarkupPrefixSegment(text, idx)
+	if !ok {
+		return idx, false
+	}
+	j := nextSegment
+	for {
+		nextSegment, ok = consumeToolMarkupPrefixSegment(text, j)
+		if !ok {
+			break
+		}
+		j = nextSegment
+	}
+	k := j
+	for k < len(text) && (text[k] == ' ' || text[k] == '\t' || text[k] == '\r' || text[k] == '\n') {
+		k++
+	}
+	next, ok := consumeToolMarkupSeparator(text, k)
+	if !ok {
+		if sep, size := normalizedASCIIAt(text, k); sep == '_' || sep == '-' {
+			next = k + size
+			ok = true
+		}
+	}
+	if !ok {
+		return idx, false
+	}
+	for next < len(text) && (text[next] == ' ' || text[next] == '\t' || text[next] == '\r' || text[next] == '\n') {
+		next++
+	}
+	if !hasToolMarkupNamePrefix(text, next) {
+		return idx, false
+	}
+	return next, true
+}
+
+func consumeToolMarkupPrefixSegment(text string, idx int) (int, bool) {
+	ch, size := normalizedASCIIAt(text, idx)
+	if size <= 0 {
+		return idx, false
+	}
+	if (ch >= 'a' && ch <= 'z') || (ch >= 'A' && ch <= 'Z') || (ch >= '0' && ch <= '9') {
+		return idx + size, true
+	}
+	return idx, false
+}
+
+func hasASCIIPartialPrefixFoldAt(text string, start int, prefix string) bool {
+	if start < 0 || start >= len(text) {
+		return false
+	}
+	idx := start
+	matched := 0
+	for matched < len(prefix) && idx < len(text) {
+		ch, size := normalizedASCIIAt(text, idx)
+		if size <= 0 || asciiLower(ch) != asciiLower(prefix[matched]) {
+			return false
+		}
+		idx += size
+		matched++
+	}
+	return matched > 0 && matched < len(prefix) && idx == len(text)
+}
+
+func hasToolMarkupNamePrefix(text string, start int) bool {
 	for _, name := range toolMarkupNames {
-		if strings.HasPrefix(lowerTail, name.raw) || strings.HasPrefix(name.raw, lowerTail) {
+		if hasASCIIPrefixFoldAt(text, start, name.raw) {
+			return true
+		}
+		if hasASCIIPartialPrefixFoldAt(text, start, name.raw) {
 			return true
 		}
 	}
 	return false
 }
 
-func matchToolMarkupName(lower string, start int, dsmlLike bool) (string, int) {
+func matchToolMarkupName(text string, start int, dsmlLike bool) (string, int) {
 	for _, name := range toolMarkupNames {
 		if name.dsmlOnly && !dsmlLike {
 			continue
 		}
-		if strings.HasPrefix(lower[start:], name.raw) {
-			return name.canonical, len(name.raw)
+		if nameLen, ok := matchASCIIPrefixFoldAt(text, start, name.raw); ok {
+			return name.canonical, nameLen
 		}
 	}
 	return "", 0
 }
 
-func consumeToolMarkupPipe(text string, idx int) (int, bool) {
+func matchToolMarkupNameAfterArbitraryPrefix(text string, start int) (string, int, int, bool) {
+	for idx := start; idx < len(text); {
+		if isToolMarkupTagTerminator(text, idx) {
+			return "", 0, 0, false
+		}
+		for _, name := range toolMarkupNames {
+			nameLen, ok := matchASCIIPrefixFoldAt(text, idx, name.raw)
+			if !ok {
+				continue
+			}
+			if !toolMarkupPrefixAllowsLocalNameAt(text, start, idx) {
+				continue
+			}
+			return name.canonical, idx, nameLen, true
+		}
+		_, size := utf8.DecodeRuneInString(text[idx:])
+		if size <= 0 {
+			size = 1
+		}
+		idx += size
+	}
+	return "", 0, 0, false
+}
+
+func hasPartialToolMarkupNameAfterArbitraryPrefix(text string, start int) bool {
+	for idx := start; idx < len(text); {
+		if isToolMarkupTagTerminator(text, idx) {
+			return false
+		}
+		if toolMarkupPrefixAllowsLocalNameAt(text, start, idx) && hasToolMarkupNamePrefix(text, idx) {
+			return true
+		}
+		if toolMarkupPrefixAllowsLocalNameAt(text, start, idx) && hasDSMLNamePrefixOrPartial(text, idx) {
+			return true
+		}
+		_, size := utf8.DecodeRuneInString(text[idx:])
+		if size <= 0 {
+			size = 1
+		}
+		idx += size
+	}
+	return toolMarkupPrefixAllowsLocalName(text[start:])
+}
+
+func toolMarkupPrefixAllowsLocalNameAt(text string, start, localStart int) bool {
+	if start < 0 || localStart <= start || localStart > len(text) {
+		return false
+	}
+	prefix := text[start:localStart]
+	if toolMarkupPrefixAllowsLocalName(prefix) {
+		return true
+	}
+	if strings.ContainsAny(prefix, "=\"'") {
+		return false
+	}
+	prev, prevSize := utf8.DecodeLastRuneInString(prefix)
+	next, _ := utf8.DecodeRuneInString(text[localStart:])
+	if prevSize <= 0 || next == utf8.RuneError {
+		return false
+	}
+	return isASCIIAlphaNumeric(normalizeFullwidthASCII(prev)) && isASCIIUpper(normalizeFullwidthASCII(next))
+}
+
+func hasDSMLNamePrefixOrPartial(text string, start int) bool {
+	return hasASCIIPrefixFoldAt(text, start, "dsml") || hasASCIIPartialPrefixFoldAt(text, start, "dsml")
+}
+
+func toolMarkupPrefixAllowsLocalName(prefix string) bool {
+	if prefix == "" {
+		return false
+	}
+	if strings.Contains(normalizedASCIILowerString(prefix), "dsml") {
+		return true
+	}
+	if strings.ContainsAny(prefix, "=\"'") {
+		return false
+	}
+	r, _ := utf8.DecodeLastRuneInString(prefix)
+	r = normalizeFullwidthASCII(r)
+	return (r < 'a' || r > 'z') && (r < 'A' || r > 'Z') && (r < '0' || r > '9')
+}
+
+func normalizedASCIILowerString(text string) string {
+	var b strings.Builder
+	b.Grow(len(text))
+	for _, r := range text {
+		r = normalizeFullwidthASCII(r)
+		if r >= 'A' && r <= 'Z' {
+			r += 'a' - 'A'
+		}
+		if r <= 0x7f {
+			b.WriteRune(r)
+		}
+	}
+	return b.String()
+}
+
+func isASCIIAlphaNumeric(r rune) bool {
+	return (r >= 'a' && r <= 'z') || (r >= 'A' && r <= 'Z') || (r >= '0' && r <= '9')
+}
+
+func isASCIIUpper(r rune) bool {
+	return r >= 'A' && r <= 'Z'
+}
+
+func isToolMarkupTagTerminator(text string, idx int) bool {
+	if idx >= len(text) {
+		return false
+	}
+	if text[idx] == '>' {
+		return true
+	}
+	r, _ := utf8.DecodeRuneInString(text[idx:])
+	return normalizeFullwidthASCII(r) == '>'
+}
+
+func consumeToolMarkupSeparator(text string, idx int) (int, bool) {
 	if idx >= len(text) {
 		return idx, false
 	}
-	if text[idx] == '|' {
-		return idx + 1, true
+	r, size := utf8.DecodeRuneInString(text[idx:])
+	if size <= 0 || !isToolMarkupSeparator(r) {
+		return idx, false
 	}
-	if strings.HasPrefix(text[idx:], "｜") {
-		return idx + len("｜"), true
+	return idx + size, true
+}
+
+func isToolMarkupSeparator(r rune) bool {
+	ch := normalizeFullwidthASCII(r)
+	if ch == 0 || ch == '<' || ch == '>' || ch == '/' || ch == '=' || ch == '"' || ch == '\'' || ch == '[' {
+		return false
 	}
-	return idx, false
+	if ch == ' ' || ch == '\t' || ch == '\n' || ch == '\r' {
+		return false
+	}
+	if (ch >= 'a' && ch <= 'z') || (ch >= 'A' && ch <= 'Z') || (ch >= '0' && ch <= '9') {
+		return false
+	}
+	return true
+}
+
+func consumeToolMarkupLessThan(text string, idx int) (int, bool) {
+	ch, size := normalizedASCIIAt(text, idx)
+	if size <= 0 || ch != '<' {
+		return idx, false
+	}
+	return idx + size, true
 }
 
 func hasToolMarkupBoundary(text string, idx int) bool {
@@ -292,6 +517,48 @@ func hasToolMarkupBoundary(text string, idx int) bool {
 	case ' ', '\t', '\n', '\r', '>', '/':
 		return true
 	default:
-		return false
+		r, _ := utf8.DecodeRuneInString(text[idx:])
+		return normalizeFullwidthASCII(r) == '>'
 	}
 }
+
+func normalizedASCIIAt(text string, idx int) (byte, int) {
+	if idx < 0 || idx >= len(text) {
+		return 0, 0
+	}
+	r, size := utf8.DecodeRuneInString(text[idx:])
+	if r == utf8.RuneError && size == 0 {
+		return 0, 0
+	}
+	normalized := normalizeFullwidthASCII(r)
+	if normalized > 0x7f {
+		return 0, 0
+	}
+	return byte(normalized), size
+}
+
+func normalizeFullwidthASCII(r rune) rune {
+	switch r {
+	case '〈':
+		return '<'
+	case '〉':
+		return '>'
+	case '“', '”':
+		return '"'
+	case '‘', '’':
+		return '\''
+	}
+	if r >= '！' && r <= '～' {
+		return r - 0xFEE0
+	}
+	return r
+}
+
+func toolMarkupPrefixContainsSlash(prefix string) bool {
+	for _, r := range prefix {
+		if normalizeFullwidthASCII(r) == '/' {
+			return true
+		}
+	}
+	return false
+}

internal/toolcall/toolcalls_test.go

@@ -72,6 +72,159 @@ EOF
 	}
 }
 
+func TestParseToolCallsSupportsUnderscoredDSMLShell(t *testing.T) {
+	text := `<dsml_tool_calls>
+<dsml_invoke name="search_web">
+<dsml_parameter name="query"><![CDATA[2026年5月 热点事件]]></dsml_parameter>
+<dsml_parameter name="topic"><![CDATA[news]]></dsml_parameter>
+</dsml_invoke>
+<dsml_invoke name="eval_javascript">
+<dsml_parameter name="code"><![CDATA[1 + 1]]></dsml_parameter>
+</dsml_invoke>
+</dsml_tool_calls>`
+	calls := ParseToolCalls(text, []string{"search_web", "eval_javascript"})
+	if len(calls) != 2 {
+		t.Fatalf("expected two underscored DSML calls, got %#v", calls)
+	}
+	if calls[0].Name != "search_web" || calls[0].Input["query"] != "2026年5月 热点事件" || calls[0].Input["topic"] != "news" {
+		t.Fatalf("unexpected first underscored DSML call: %#v", calls[0])
+	}
+	if calls[1].Name != "eval_javascript" || calls[1].Input["code"] != "1 + 1" {
+		t.Fatalf("unexpected second underscored DSML call: %#v", calls[1])
+	}
+}
+
+func TestParseToolCallsSupportsArbitraryPrefixedToolMarkup(t *testing.T) {
+	cases := []string{
+		`<abc|tool_calls><abc|invoke name="Read"><abc|parameter name="file_path">README.md</abc|parameter></abc|invoke></abc|tool_calls>`,
+		`<vendor_tool_calls><vendor_invoke name="Read"><vendor_parameter name="file_path">README.md</vendor_parameter></vendor_invoke></vendor_tool_calls>`,
+		`<agent - tool_calls><agent - invoke name="Read"><agent - parameter name="file_path">README.md</agent - parameter></agent - invoke></agent - tool_calls>`,
+	}
+	for _, text := range cases {
+		calls := ParseToolCalls(text, []string{"Read"})
+		if len(calls) != 1 {
+			t.Fatalf("expected one arbitrary-prefixed tool call for %q, got %#v", text, calls)
+		}
+		if calls[0].Name != "Read" || calls[0].Input["file_path"] != "README.md" {
+			t.Fatalf("unexpected arbitrary-prefixed parse result: %#v", calls[0])
+		}
+	}
+}
+
+func TestParseToolCallsSupportsCamelPrefixedToolMarkup(t *testing.T) {
+	text := `<DSmartToolCalls><DSmartInvoke name="Bash"><DSmartParameter name="command"><![CDATA[git push]]></DSmartParameter><DSmartParameter name="description"><![CDATA[Push dev branch to origin]]></DSmartParameter></DSmartInvoke></DSmartToolCalls>`
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected one camel-prefixed tool call, got %#v", calls)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "git push" || calls[0].Input["description"] != "Push dev branch to origin" {
+		t.Fatalf("unexpected camel-prefixed tool call: %#v", calls[0])
+	}
+}
+
+func TestParseToolCallsRejectsCamelPrefixedToolMarkupLookalike(t *testing.T) {
+	text := `<DSmartToolCallsExtra><DSmartInvoke name="Bash"><DSmartParameter name="command">git push</DSmartParameter></DSmartInvoke></DSmartToolCallsExtra>`
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 0 {
+		t.Fatalf("expected camel-prefixed lookalike to be ignored, got %#v", calls)
+	}
+}
+
+func TestParseToolCallsSupportsFullwidthDSMLShell(t *testing.T) {
+	text := `<ｄＳＭＬ｜tool_calls>
+  <ｄＳＭＬ｜invoke name="Read">
+    <ｄＳＭＬ｜parameter name="file_path"＞<![CDATA[/Users/aq/Desktop/myproject/Personal_Blog/README.md]]＞</ｄＳＭＬ｜parameter>
+  </ｄＳＭＬ｜invoke>
+  <ｄＳＭＬ｜invoke name="Read">
+    <ｄＳＭＬ｜parameter name="file_path"＞<![CDATA[/Users/aq/Desktop/myproject/Personal_Blog/index.html]]＞</ｄＳＭＬ｜parameter>
+  </ｄＳＭＬ｜invoke>
+</ｄＳＭＬ｜tool_calls>`
+	calls := ParseToolCalls(text, []string{"Read"})
+	if len(calls) != 2 {
+		t.Fatalf("expected two fullwidth DSML calls, got %#v", calls)
+	}
+	if calls[0].Name != "Read" || calls[0].Input["file_path"] != "/Users/aq/Desktop/myproject/Personal_Blog/README.md" {
+		t.Fatalf("unexpected first fullwidth DSML call: %#v", calls[0])
+	}
+	if calls[1].Name != "Read" || calls[1].Input["file_path"] != "/Users/aq/Desktop/myproject/Personal_Blog/index.html" {
+		t.Fatalf("unexpected second fullwidth DSML call: %#v", calls[1])
+	}
+}
+
+func TestParseToolCallsSupportsCJKAngleDSMDrift(t *testing.T) {
+	text := `<DSM｜tool_calls>
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Show commits on local dev not on origin/dev]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git log --oneline origin/dev..dev]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Show commits on origin/dev not on local dev]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git log --oneline dev..origin/dev]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Check tracking branch status]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git status -b --short]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+〈/DSM｜tool_calls〉`
+
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 3 {
+		t.Fatalf("expected three CJK-angle DSM drift calls, got %#v", calls)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "git log --oneline origin/dev..dev" {
+		t.Fatalf("unexpected first CJK-angle DSM drift call: %#v", calls[0])
+	}
+	if calls[1].Name != "Bash" || calls[1].Input["description"] != "Show commits on origin/dev not on local dev" {
+		t.Fatalf("unexpected second CJK-angle DSM drift call: %#v", calls[1])
+	}
+	if calls[2].Name != "Bash" || calls[2].Input["command"] != "git status -b --short" {
+		t.Fatalf("unexpected third CJK-angle DSM drift call: %#v", calls[2])
+	}
+}
+
+func TestParseToolCallsSupportsFullwidthBangDSMLDrift(t *testing.T) {
+	text := `<！DSML！tool_calls>
+  <！DSML！invoke name=“Bash”>
+  <！DSML！parameter name=“command”><！[CDATA[lsof -i :4321 -t]]><！/DSML！parameter>
+  <！DSML！parameter name=“description”><！[CDATA[Verify port 4321 is free]]><！/DSML！parameter>
+  <！/DSML！invoke>
+  <！/DSML！tool_calls>`
+
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected one fullwidth-bang DSML drift call, got %#v", calls)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "lsof -i :4321 -t" || calls[0].Input["description"] != "Verify port 4321 is free" {
+		t.Fatalf("unexpected fullwidth-bang DSML drift call: %#v", calls[0])
+	}
+}
+
+func TestParseToolCallsSupportsIdeographicCommaDSMLDrift(t *testing.T) {
+	text := `<、DSML、tool_calls>
+  <、DSML、invoke name="Bash">
+    <、DSML、parameter name="command"><、[CDATA[git commit -m "$(cat <<'EOF'
+feat: expand fullwidth bang separator and curly quote tolerance in DSML tool parsing
+
+Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
+EOF
+)"]]><、/DSML、parameter>
+    <、DSML、parameter name="description"><、[CDATA[Create commit with staged changes]]><、/DSML、parameter>
+  <、/DSML、invoke>
+<、/DSML、tool_calls>`
+
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected one ideographic-comma DSML drift call, got %#v", calls)
+	}
+	command, _ := calls[0].Input["command"].(string)
+	if calls[0].Name != "Bash" || !strings.Contains(command, `git commit -m "$(cat <<'EOF'`) || !strings.Contains(command, "Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com") {
+		t.Fatalf("unexpected ideographic-comma DSML drift call: %#v", calls[0])
+	}
+	if calls[0].Input["description"] != "Create commit with staged changes" {
+		t.Fatalf("unexpected ideographic-comma description: %#v", calls[0])
+	}
+}
+
 func TestParseToolCallsIgnoresBareHyphenatedToolCallsLookalike(t *testing.T) {
 	text := `<tool-calls><invoke name="Bash"><parameter name="command">pwd</parameter></invoke></tool-calls>`
 	calls := ParseToolCalls(text, []string{"Bash"})
@@ -104,6 +257,23 @@ func TestParseToolCallsToleratesDSMLTrailingPipeTagTerminator(t *testing.T) {
 	}
 }
 
+func TestParseToolCallsToleratesDSMLTrailingNovelSeparatorTagTerminator(t *testing.T) {
+	text := strings.Join([]string{
+		`<DSMLtool_calls※>`,
+		`  <DSMLinvoke name="Bash"※>`,
+		`    <DSMLparameter name="command"※><![CDATA[pwd]]></DSMLparameter※>`,
+		`  </DSMLinvoke※>`,
+		`</DSMLtool_calls※>`,
+	}, "\n")
+	calls := ParseToolCalls(text, []string{"Bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected one trailing-separator DSML call, got %#v", calls)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "pwd" {
+		t.Fatalf("unexpected trailing-separator DSML parse result: %#v", calls[0])
+	}
+}
+
 func TestParseToolCallsToleratesExtraLeadingLessThanBeforeDSML(t *testing.T) {
 	text := `<<|DSML|tool_calls><<|DSML|invoke name="Bash"><<|DSML|parameter name="command"><![CDATA[pwd]]></|DSML|parameter></|DSML|invoke></|DSML|tool_calls>`
 	calls := ParseToolCalls(text, []string{"Bash"})
@@ -516,14 +686,17 @@ func TestParseToolCallsDetailedMarksToolCallsSyntax(t *testing.T) {
 	}
 }
 
-func TestParseToolCallsRejectsAllEmptyParameterPayload(t *testing.T) {
+func TestParseToolCallsAllowsAllEmptyParameterPayload(t *testing.T) {
 	text := `<tool_calls><invoke name="Bash"><parameter name="command"></parameter><parameter name="description">   </parameter><parameter name="timeout"></parameter></invoke></tool_calls>`
 	res := ParseToolCallsDetailed(text, []string{"Bash"})
 	if !res.SawToolCallSyntax {
 		t.Fatalf("expected tool syntax to be detected, got %#v", res)
 	}
-	if len(res.Calls) != 0 {
-		t.Fatalf("expected all-empty payload to be rejected, got %#v", res.Calls)
+	if len(res.Calls) != 1 {
+		t.Fatalf("expected all-empty payload to be parsed, got %#v", res.Calls)
+	}
+	if res.Calls[0].Input["command"] != "" || res.Calls[0].Input["description"] != "" || res.Calls[0].Input["timeout"] != "" {
+		t.Fatalf("expected empty parameters to be preserved, got %#v", res.Calls[0].Input)
 	}
 }
 
@@ -892,3 +1065,139 @@ func TestParseToolCallsSkipsProseMentionOfSameWrapperVariant(t *testing.T) {
 		t.Fatalf("expected command to parse, got %q", got)
 	}
 }
+
+func TestTurkishILowercaseMapping(t *testing.T) {
+	tests := []struct {
+		name     string
+		text     string
+		start    int
+		wantOk   bool
+		wantName string
+	}{
+		{"turkish_i_at_name_start", "İ<tool>", 0, false, ""},
+		{"turkish_i_at_name_end", "<toolİ>", 0, false, ""},
+		{"turkish_i_before_tag", "İ<tool>", 0, false, ""},
+		{"normal_tool_calls", "<tool_calls>", 0, true, "tool_calls"},
+		{"normal_invoke", "<invoke name=\"test\">", 0, true, "invoke"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, ok := FindToolMarkupTagOutsideIgnored(tt.text, tt.start)
+			if ok != tt.wantOk {
+				t.Errorf("FindToolMarkupTagOutsideIgnored(%q, %d) ok = %v, want %v", tt.text, tt.start, ok, tt.wantOk)
+				return
+			}
+			if ok && got.Name != tt.wantName {
+				t.Errorf("FindToolMarkupTagOutsideIgnored(%q, %d) name = %q, want %q", tt.text, tt.start, got.Name, tt.wantName)
+			}
+		})
+	}
+}
+
+func TestSkipXMLIgnoredSectionBoundaryConditions(t *testing.T) {
+	text := "hello"
+
+	tests := []struct {
+		name     string
+		i        int
+		wantNext int
+		wantAdv  bool
+		wantBlk  bool
+	}{
+		{"valid_index", 2, 2, false, false},
+		{"at_end_equal_len", 5, 5, false, false},
+		{"beyond_end", 6, 6, false, false},
+		{"negative", -1, -1, false, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			next, adv, blk := skipXMLIgnoredSection(text, tt.i)
+			if next != tt.wantNext || adv != tt.wantAdv || blk != tt.wantBlk {
+				t.Errorf("skipXMLIgnoredSection(%q, %d) = (%d, %v, %v), want (%d, %v, %v)",
+					text, tt.i, next, adv, blk, tt.wantNext, tt.wantAdv, tt.wantBlk)
+			}
+		})
+	}
+}
+
+func TestSkipXMLIgnoredSectionCommentWithUnicodeKeepsByteOffset(t *testing.T) {
+	text := "<!-- İ -->x<tool_calls>"
+
+	next, adv, blk := skipXMLIgnoredSection(text, 0)
+	if blk || !adv {
+		t.Fatalf("skipXMLIgnoredSection() = (%d, %v, %v), want advanced unblocked comment", next, adv, blk)
+	}
+	if want := len("<!-- İ -->"); next != want {
+		t.Fatalf("skipXMLIgnoredSection() next = %d, want %d", next, want)
+	}
+}
+
+func TestSkipXMLIgnoredSectionMatchesCDATAWithoutAllocatingTail(t *testing.T) {
+	text := "<![cDaTa[<tool_calls>]]><tool_calls>"
+
+	next, adv, blk := skipXMLIgnoredSection(text, 0)
+	if blk || !adv {
+		t.Fatalf("skipXMLIgnoredSection() = (%d, %v, %v), want advanced unblocked CDATA", next, adv, blk)
+	}
+	if want := len("<![cDaTa[<tool_calls>]]>"); next != want {
+		t.Fatalf("skipXMLIgnoredSection() next = %d, want %d", next, want)
+	}
+
+	tag, ok := FindToolMarkupTagOutsideIgnored(text, 0)
+	if !ok {
+		t.Fatal("expected tool tag after skipped CDATA")
+	}
+	if tag.Start != next {
+		t.Fatalf("FindToolMarkupTagOutsideIgnored() start = %d, want %d", tag.Start, next)
+	}
+}
+
+func TestFindToolCDATAEndBoundaryConditions(t *testing.T) {
+	text := "<![CDATA[hello]]>"
+
+	tests := []struct {
+		name       string
+		from       int
+		wantResult int
+	}{
+		{"valid", 12, 14},
+		{"at_end", 17, -1},
+		{"beyond_end", 18, -1},
+		{"negative", -1, -1},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := findToolCDATAEnd(text, tt.from)
+			if got != tt.wantResult {
+				t.Errorf("findToolCDATAEnd(%q, %d) = %d, want %d",
+					text, tt.from, got, tt.wantResult)
+			}
+		})
+	}
+}
+
+func TestFindMatchingToolMarkupCloseBoundaryConditions(t *testing.T) {
+	tests := []struct {
+		name   string
+		text   string
+		open   ToolMarkupTag
+		wantOk bool
+	}{
+		{"empty_text", "", ToolMarkupTag{Name: "tool_calls", End: 0}, false},
+		{"open_end_beyond_text", "hello", ToolMarkupTag{Name: "tool_calls", End: 100}, false},
+		{"open_end_equals_len", "hello", ToolMarkupTag{Name: "tool_calls", End: 5}, false},
+		{"valid_simple", "<tool_calls></tool_calls>", ToolMarkupTag{Name: "tool_calls", End: 11}, true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			_, ok := FindMatchingToolMarkupClose(tt.text, tt.open)
+			if ok != tt.wantOk {
+				t.Errorf("FindMatchingToolMarkupClose(%q, %+v) ok = %v, want %v", tt.text, tt.open, ok, tt.wantOk)
+			}
+		})
+	}
+}

internal/toolstream/tool_sieve_xml.go

@@ -54,7 +54,7 @@ func consumeXMLToolCapture(captured string, toolNames []string) (prefix string,
 		}
 		if parsed.SawToolCallSyntax {
 			if rejected == nil || tag.Start < rejected.start {
-				rejected = &rejectedBlock{start: tag.Start, prefix: prefixPart, suffix: suffixPart}
+				rejected = &rejectedBlock{start: tag.Start, prefix: prefixPart + xmlBlock, suffix: suffixPart}
 			}
 			searchFrom = tag.End + 1
 			continue
@@ -88,7 +88,7 @@ func consumeXMLToolCapture(captured string, toolNames []string) (prefix string,
 					return prefixPart, parsed.Calls, suffixPart, true
 				}
 				if parsed.SawToolCallSyntax {
-					return prefixPart, nil, suffixPart, true
+					return prefixPart + captured[invokeTag.Start:closeTag.End+1], nil, suffixPart, true
 				}
 				return prefixPart + captured[invokeTag.Start:closeTag.End+1], nil, suffixPart, true
 			}

internal/toolstream/tool_sieve_xml_test.go

@@ -1,6 +1,7 @@
 package toolstream
 
 import (
+	"ds2api/internal/toolcall"
 	"strings"
 	"testing"
 )
@@ -104,6 +105,99 @@ func TestProcessToolSieveInterceptsDSMLTrailingPipeToolCallWithoutLeak(t *testin
 	}
 }
 
+func TestProcessToolSieveInterceptsDSMLControlSeparatorWithoutLeak(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		sep  string
+	}{
+		{name: "control_picture", sep: "␂"},
+		{name: "raw_stx", sep: "\x02"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			sep := tc.sep
+			var state State
+			chunks := []string{
+				"<DSML" + sep + "tool",
+				"_calls>\n",
+				`  <DSML` + sep + `invoke name="Read">` + "\n",
+				`    <DSML` + sep + `parameter name="file_path"><![CDATA[/tmp/input.txt]]></DSML` + sep + `parameter>` + "\n",
+				"  </DSML" + sep + "invoke>\n",
+				"</DSML" + sep + "tool_calls>",
+			}
+			var events []Event
+			for _, c := range chunks {
+				events = append(events, ProcessChunk(&state, c, []string{"Read"})...)
+			}
+			events = append(events, Flush(&state, []string{"Read"})...)
+
+			var textContent strings.Builder
+			var calls []any
+			for _, evt := range events {
+				textContent.WriteString(evt.Content)
+				for _, call := range evt.ToolCalls {
+					calls = append(calls, call)
+				}
+			}
+			if text := textContent.String(); strings.Contains(strings.ToLower(text), "dsml") || strings.Contains(text, "Read") || strings.Contains(text, sep) {
+				t.Fatalf("control-separator DSML tool call leaked to text: %q events=%#v", text, events)
+			}
+			if len(calls) != 1 {
+				t.Fatalf("expected one control-separator DSML tool call, got %d events=%#v", len(calls), events)
+			}
+		})
+	}
+}
+
+func TestProcessToolSieveInterceptsArbitraryPrefixedToolTagsWithoutLeak(t *testing.T) {
+	var state State
+	chunks := []string{
+		"<proto💥tool",
+		"_calls>\n",
+		`  <proto💥invoke name="Read">` + "\n",
+		`    <proto💥parameter name="file_path"><![CDATA[/tmp/input.txt]]></proto💥parameter>` + "\n",
+		"  </proto💥invoke>\n",
+		"</proto💥tool_calls>",
+	}
+	var events []Event
+	for _, c := range chunks {
+		events = append(events, ProcessChunk(&state, c, []string{"Read"})...)
+	}
+	events = append(events, Flush(&state, []string{"Read"})...)
+
+	var textContent strings.Builder
+	var calls []any
+	for _, evt := range events {
+		textContent.WriteString(evt.Content)
+		for _, call := range evt.ToolCalls {
+			calls = append(calls, call)
+		}
+	}
+	if text := textContent.String(); strings.Contains(text, "proto") || strings.Contains(text, "Read") || strings.Contains(text, "💥") {
+		t.Fatalf("arbitrary-prefixed tool call leaked to text: %q events=%#v", text, events)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected one arbitrary-prefixed tool call, got %d events=%#v", len(calls), events)
+	}
+}
+
+func TestProcessToolSieveEmitsEmptyDSMLControlSeparatorBlockWithoutLeak(t *testing.T) {
+	sep := "␂"
+	chunks := []string{
+		"<DSML" + sep + "tool_calls>\n",
+		`  <DSML` + sep + `invoke name="Read">` + "\n",
+		`    <DSML` + sep + `parameter name="file_path"></DSML` + sep + `parameter>` + "\n",
+		"  </DSML" + sep + "invoke>\n",
+		"</DSML" + sep + "tool_calls>",
+	}
+	calls := collectToolCallsForChunks(t, chunks, []string{"Read"})
+	if len(calls) != 1 {
+		t.Fatalf("expected empty control-separator block to produce one call, got %#v", calls)
+	}
+	if calls[0].Name != "Read" || calls[0].Input["file_path"] != "" {
+		t.Fatalf("expected empty file_path parameter to be preserved, got %#v", calls)
+	}
+}
+
 func TestProcessToolSieveInterceptsExtraLeadingLessThanDSMLToolCallWithoutLeak(t *testing.T) {
 	var state State
 	chunks := []string{
@@ -490,7 +584,7 @@ func TestProcessToolSieveNonToolXMLKeepsSuffixForToolParsing(t *testing.T) {
 	}
 }
 
-func TestProcessToolSieveSuppressesMalformedExecutableXMLBlock(t *testing.T) {
+func TestProcessToolSieveReleasesMalformedExecutableXMLBlock(t *testing.T) {
 	var state State
 	chunk := `<tool_calls><invoke name="read_file"><param>{"path":"README.md"}</param></invoke></tool_calls>`
 	events := ProcessChunk(&state, chunk, []string{"read_file"})
@@ -506,13 +600,12 @@ func TestProcessToolSieveSuppressesMalformedExecutableXMLBlock(t *testing.T) {
 	if toolCalls != 0 {
 		t.Fatalf("expected malformed executable-looking XML not to become a tool call, got %d events=%#v", toolCalls, events)
 	}
-	if textContent.Len() != 0 {
-		t.Fatalf("expected malformed executable-looking XML to be suppressed, got %q", textContent.String())
+	if textContent.String() != chunk {
+		t.Fatalf("expected malformed executable-looking XML to be released as text, got %q", textContent.String())
 	}
 }
 
-func TestProcessToolSieveSuppressesAllEmptyDSMLToolBlock(t *testing.T) {
-	var state State
+func TestProcessToolSieveEmitsAllEmptyDSMLToolBlock(t *testing.T) {
 	chunk := strings.Join([]string{
 		`<|DSML|tool_calls>`,
 		`<|DSML|invoke name="Bash">`,
@@ -522,22 +615,69 @@ func TestProcessToolSieveSuppressesAllEmptyDSMLToolBlock(t *testing.T) {
 		`</|DSML|invoke>`,
 		`</|DSML|tool_calls>`,
 	}, "\n")
-	events := ProcessChunk(&state, chunk, []string{"Bash"})
-	events = append(events, Flush(&state, []string{"Bash"})...)
+	calls := collectToolCallsForChunks(t, []string{chunk}, []string{"Bash"})
+	if len(calls) != 1 {
+		t.Fatalf("expected all-empty DSML block to produce one tool call, got %#v", calls)
+	}
+	if calls[0].Input["command"] != "" || calls[0].Input["description"] != "" || calls[0].Input["timeout"] != "" {
+		t.Fatalf("expected empty parameters to be preserved, got %#v", calls[0].Input)
+	}
+}
+
+func TestProcessToolSieveEmitsChunkedAllEmptyArbitraryPrefixedToolBlock(t *testing.T) {
+	chunk := strings.Join([]string{
+		`<T｜DSML｜tool_calls>`,
+		`  <T｜DSML｜invoke name="TaskOutput">`,
+		`  <T｜DSML｜parameter name="task_id"></T｜DSML｜parameter>`,
+		`  <T｜DSML｜parameter name="block"></T｜DSML｜parameter>`,
+		`  <T｜DSML｜parameter name="timeout"></T｜DSML｜parameter>`,
+		`  </T｜DSML｜invoke>`,
+		`  </T｜DSML｜tool_calls>`,
+	}, "\n")
+	calls := collectToolCallsForChunks(t, splitEveryNRBytes(chunk, 8), []string{"TaskOutput"})
+	if len(calls) != 1 {
+		t.Fatalf("expected chunked all-empty arbitrary-prefixed block to produce one tool call, got %#v", calls)
+	}
+	if calls[0].Name != "TaskOutput" || calls[0].Input["task_id"] != "" || calls[0].Input["block"] != "" || calls[0].Input["timeout"] != "" {
+		t.Fatalf("expected empty TaskOutput parameters to be preserved, got %#v", calls)
+	}
+}
+
+func collectToolCallsForChunks(t *testing.T, chunks []string, toolNames []string) []toolcall.ParsedToolCall {
+	t.Helper()
+	var state State
+	var events []Event
+	for _, chunk := range chunks {
+		events = append(events, ProcessChunk(&state, chunk, toolNames)...)
+	}
+	events = append(events, Flush(&state, toolNames)...)
 
 	var textContent strings.Builder
-	toolCalls := 0
+	var calls []toolcall.ParsedToolCall
 	for _, evt := range events {
 		textContent.WriteString(evt.Content)
-		toolCalls += len(evt.ToolCalls)
-	}
-
-	if toolCalls != 0 {
-		t.Fatalf("expected all-empty DSML block not to produce tool calls, got %d events=%#v", toolCalls, events)
+		calls = append(calls, evt.ToolCalls...)
 	}
 	if textContent.Len() != 0 {
-		t.Fatalf("expected all-empty DSML block not to leak as text, got %q", textContent.String())
+		t.Fatalf("expected tool block not to leak as text, got %q", textContent.String())
 	}
+	return calls
+}
+
+func splitEveryNRBytes(s string, n int) []string {
+	if n <= 0 {
+		return []string{s}
+	}
+	out := make([]string, 0, len(s)/n+1)
+	for len(s) > 0 {
+		if len(s) <= n {
+			out = append(out, s)
+			break
+		}
+		out = append(out, s[:n])
+		s = s[n:]
+	}
+	return out
 }
 
 func TestProcessToolSievePassesThroughFencedXMLToolCallExamples(t *testing.T) {
@@ -671,6 +811,8 @@ func TestFindPartialXMLToolTagStart(t *testing.T) {
 		{"partial_tool_calls", "Hello <tool_ca", 6},
 		{"partial_dsml_trailing_pipe", "Hello <|DSML|tool_calls|", 6},
 		{"partial_dsml_extra_leading_less_than", "Hello <<|DSML|tool_calls", 6},
+		{"partial_arbitrary_prefix_before_dsml", "Hello <T｜DS", 6},
+		{"partial_arbitrary_prefix_after_dsml_pipe", "Hello <T｜DSML｜", 6},
 		{"partial_invoke", "Hello <inv", 6},
 		{"bare_tool_call_not_held", "Hello <tool_name", -1},
 		{"partial_lt_only", "Text <", 5},
@@ -1086,3 +1228,110 @@ func TestProcessToolSieveDSMLBarePrefixVariantDoesNotLeak(t *testing.T) {
 		t.Fatalf("expected one tool call from DSML bare prefix variant, got %d events=%#v", toolCalls, events)
 	}
 }
+
+func TestProcessToolSieveCJKAngleDSMDriftDoesNotLeak(t *testing.T) {
+	var state State
+	chunks := []string{
+		"<DSM｜tool_calls>\n",
+		"<DSM｜invoke name=\"Bash\">\n",
+		"<DSM｜parameter name=\"description\"｜>〈![CDATA[Check tracking branch status]]〉〈/DSM｜parameter〉\n",
+		"<DSM｜parameter name=\"command\"｜>〈![CDATA[git status -b --short]]〉〈/DSM｜parameter〉\n",
+		"〈/DSM｜invoke〉\n",
+		"〈/DSM｜tool_calls〉",
+	}
+	var events []Event
+	for _, c := range chunks {
+		events = append(events, ProcessChunk(&state, c, []string{"Bash"})...)
+	}
+	events = append(events, Flush(&state, []string{"Bash"})...)
+
+	var textContent string
+	var calls []toolcall.ParsedToolCall
+	for _, evt := range events {
+		textContent += evt.Content
+		calls = append(calls, evt.ToolCalls...)
+	}
+
+	if strings.Contains(textContent, "DSM") || strings.Contains(textContent, "git status") {
+		t.Fatalf("CJK-angle DSM drift leaked to text: %q events=%#v", textContent, events)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected one CJK-angle DSM drift tool call, got %d events=%#v", len(calls), events)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "git status -b --short" {
+		t.Fatalf("unexpected CJK-angle DSM drift call: %#v", calls[0])
+	}
+}
+
+func TestProcessToolSieveFullwidthBangDSMLDriftDoesNotLeak(t *testing.T) {
+	var state State
+	chunks := []string{
+		"<！DSML！tool_calls>\n",
+		"  <！DSML！invoke name=“Bash”>\n",
+		"  <！DSML！parameter name=“command”><！[CDATA[lsof -i :4321 -t]]><！/DSML！parameter>\n",
+		"  <！DSML！parameter name=“description”><！[CDATA[Verify port 4321 is free]]><！/DSML！parameter>\n",
+		"  <！/DSML！invoke>\n",
+		"  <！/DSML！tool_calls>",
+	}
+	var events []Event
+	for _, c := range chunks {
+		events = append(events, ProcessChunk(&state, c, []string{"Bash"})...)
+	}
+	events = append(events, Flush(&state, []string{"Bash"})...)
+
+	var textContent string
+	var calls []toolcall.ParsedToolCall
+	for _, evt := range events {
+		textContent += evt.Content
+		calls = append(calls, evt.ToolCalls...)
+	}
+
+	if strings.Contains(textContent, "DSML") || strings.Contains(textContent, "lsof") {
+		t.Fatalf("fullwidth-bang DSML drift leaked to text: %q events=%#v", textContent, events)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected one fullwidth-bang DSML drift tool call, got %d events=%#v", len(calls), events)
+	}
+	if calls[0].Name != "Bash" || calls[0].Input["command"] != "lsof -i :4321 -t" {
+		t.Fatalf("unexpected fullwidth-bang DSML drift call: %#v", calls[0])
+	}
+}
+
+func TestProcessToolSieveIdeographicCommaDSMLDriftDoesNotLeak(t *testing.T) {
+	var state State
+	chunks := []string{
+		"<、DSML、tool_calls>\n",
+		"  <、DSML、invoke name=\"Bash\">\n",
+		"    <、DSML、parameter name=\"command\"><、[CDATA[git commit -m \"$(cat <<'EOF'\n",
+		"feat: expand fullwidth bang separator and curly quote tolerance in DSML tool parsing\n\n",
+		"Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com\n",
+		"EOF\n",
+		")\"]]><、/DSML、parameter>\n",
+		"    <、DSML、parameter name=\"description\"><、[CDATA[Create commit with staged changes]]><、/DSML、parameter>\n",
+		"  <、/DSML、invoke>\n",
+		"<、/DSML、tool_calls>",
+	}
+	var events []Event
+	for _, c := range chunks {
+		events = append(events, ProcessChunk(&state, c, []string{"Bash"})...)
+	}
+	events = append(events, Flush(&state, []string{"Bash"})...)
+
+	var textContent string
+	var calls []toolcall.ParsedToolCall
+	for _, evt := range events {
+		textContent += evt.Content
+		calls = append(calls, evt.ToolCalls...)
+	}
+
+	if strings.Contains(textContent, "DSML") || strings.Contains(textContent, "git commit") {
+		t.Fatalf("ideographic-comma DSML drift leaked to text: %q events=%#v", textContent, events)
+	}
+	if len(calls) != 1 {
+		t.Fatalf("expected one ideographic-comma DSML drift tool call, got %d events=%#v", len(calls), events)
+	}
+	command, _ := calls[0].Input["command"].(string)
+	if calls[0].Name != "Bash" || !strings.Contains(command, "git commit -m") {
+		t.Fatalf("unexpected ideographic-comma DSML drift call: %#v", calls[0])
+	}
+}

tests/node/chat-stream.test.js

@@ -187,9 +187,9 @@ test('vercel stream emits Go-parity empty-output failure on DONE', async () => {
   const { frames } = await runMockVercelStream(['data: [DONE]\n\n']);
   assert.equal(frames.length, 2);
   const failed = JSON.parse(frames[0]);
-  assert.equal(failed.status_code, 429);
-  assert.equal(failed.error.type, 'rate_limit_error');
-  assert.equal(failed.error.code, 'upstream_empty_output');
+  assert.equal(failed.status_code, 503);
+  assert.equal(failed.error.type, 'service_unavailable_error');
+  assert.equal(failed.error.code, 'upstream_unavailable');
   assert.equal(frames[1], '[DONE]');
 });
 
@@ -209,6 +209,21 @@ test('vercel stream retries empty output once and keeps one terminal frame', asy
   assert.match(completionBodies[1].prompt, /Previous reply had no visible output\. Please regenerate the visible final answer or tool call now\.$/);
 });
 
+test('vercel stream retries thinking-only output once', async () => {
+  const { frames, fetchURLs, fetchBodies } = await runMockVercelStreamSequence([
+    ['data: {"response_message_id":42,"p":"response/thinking_content","v":"plan"}\n\n', 'data: [DONE]\n\n'],
+    ['data: {"p":"response/content","v":"visible"}\n\n', 'data: [DONE]\n\n'],
+  ], { thinking_enabled: true });
+  const parsed = frames.filter((frame) => frame !== '[DONE]').map((frame) => JSON.parse(frame));
+  const completionBodies = fetchBodies.filter((body) => Object.hasOwn(body, 'prompt'));
+  assert.equal(fetchURLs.filter((url) => url === 'https://chat.deepseek.com/api/v0/chat/completion').length, 2);
+  assert.equal(frames.filter((frame) => frame === '[DONE]').length, 1);
+  assert.equal(completionBodies[1].parent_message_id, 42);
+  assert.equal(parsed[0].choices[0].delta.reasoning_content, 'plan');
+  assert.equal(parsed[1].choices[0].delta.content, 'visible');
+  assert.equal(parsed[2].choices[0].finish_reason, 'stop');
+});
+
 test('vercel stream coalesces many small content deltas while keeping one choice', async () => {
   const lines = Array.from({ length: 100 }, () => `data: ${JSON.stringify({ p: 'response/content', v: '字' })}\n\n`);
   lines.push('data: [DONE]\n\n');
@@ -758,9 +773,11 @@ test('shouldSkipPath skips dynamic response/fragments/*/status paths only', () =
   assert.equal(shouldSkipPath('response/status'), false);
 });
 
-test('node stream path guard only allows /v1/chat/completions', () => {
+test('node stream path guard allows OpenAI v1 and root alias chat completions paths', () => {
   assert.equal(isNodeStreamSupportedPath('/v1/chat/completions'), true);
   assert.equal(isNodeStreamSupportedPath('/v1/chat/completions?x=1'), true);
+  assert.equal(isNodeStreamSupportedPath('/chat/completions'), true);
+  assert.equal(isNodeStreamSupportedPath('/chat/completions?x=1'), true);
   assert.equal(isNodeStreamSupportedPath('/v1beta/models/gemini-2.5-flash:streamGenerateContent'), false);
   assert.equal(isNodeStreamSupportedPath('/anthropic/v1/messages'), false);
 });
@@ -768,6 +785,7 @@ test('node stream path guard only allows /v1/chat/completions', () => {
 test('extractPathname strips query only', () => {
   assert.equal(extractPathname('/v1/chat/completions?stream=true'), '/v1/chat/completions');
   assert.equal(extractPathname('/v1beta/models/gemini-2.5-flash:streamGenerateContent?key=1'), '/v1beta/models/gemini-2.5-flash:streamGenerateContent');
+  assert.equal(extractPathname('/chat/completions?stream=true'), '/chat/completions');
 });
 
 test('setCorsHeaders reflects requested third-party headers and blocks internal-only headers', () => {

tests/node/stream-tool-sieve.test.js

@@ -80,6 +80,169 @@ EOF
   assert.equal(calls[0].input.command.includes('Co-Authored-By: Claude Opus 4.7'), true);
 });
 
+test('parseToolCalls parses underscored DSML shell (Vercel parity)', () => {
+  const payload = `<dsml_tool_calls>
+<dsml_invoke name="search_web">
+<dsml_parameter name="query"><![CDATA[2026年5月 热点事件]]></dsml_parameter>
+<dsml_parameter name="topic"><![CDATA[news]]></dsml_parameter>
+</dsml_invoke>
+<dsml_invoke name="eval_javascript">
+<dsml_parameter name="code"><![CDATA[1 + 1]]></dsml_parameter>
+</dsml_invoke>
+</dsml_tool_calls>`;
+  const calls = parseToolCalls(payload, ['search_web', 'eval_javascript']);
+  assert.equal(calls.length, 2);
+  assert.equal(calls[0].name, 'search_web');
+  assert.deepEqual(calls[0].input, { query: '2026年5月 热点事件', topic: 'news' });
+  assert.equal(calls[1].name, 'eval_javascript');
+  assert.deepEqual(calls[1].input, { code: '1 + 1' });
+});
+
+test('parseToolCalls parses arbitrary-prefixed tool markup shells', () => {
+  const samples = [
+    '<abc|tool_calls><abc|invoke name="Read"><abc|parameter name="file_path">README.md</abc|parameter></abc|invoke></abc|tool_calls>',
+    '<vendor_tool_calls><vendor_invoke name="Read"><vendor_parameter name="file_path">README.md</vendor_parameter></vendor_invoke></vendor_tool_calls>',
+    '<agent - tool_calls><agent - invoke name="Read"><agent - parameter name="file_path">README.md</agent - parameter></agent - invoke></agent - tool_calls>',
+  ];
+  for (const payload of samples) {
+    const calls = parseToolCalls(payload, ['Read']);
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].name, 'Read');
+    assert.deepEqual(calls[0].input, { file_path: 'README.md' });
+  }
+});
+
+test('parseToolCalls parses camel-prefixed tool markup shell', () => {
+  const payload = '<DSmartToolCalls><DSmartInvoke name="Bash"><DSmartParameter name="command"><![CDATA[git push]]></DSmartParameter><DSmartParameter name="description"><![CDATA[Push dev branch to origin]]></DSmartParameter></DSmartInvoke></DSmartToolCalls>';
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'Bash');
+  assert.deepEqual(calls[0].input, {
+    command: 'git push',
+    description: 'Push dev branch to origin',
+  });
+});
+
+test('parseToolCalls ignores camel-prefixed tool markup lookalike', () => {
+  const payload = '<DSmartToolCallsExtra><DSmartInvoke name="Bash"><DSmartParameter name="command">git push</DSmartParameter></DSmartInvoke></DSmartToolCallsExtra>';
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 0);
+});
+
+test('parseToolCalls parses fullwidth DSML shell drift', () => {
+  const payload = `<ｄＳＭＬ｜tool_calls>
+  <ｄＳＭＬ｜invoke name="Read">
+    <ｄＳＭＬ｜parameter name="file_path"＞<![CDATA[/Users/aq/Desktop/myproject/Personal_Blog/README.md]]＞</ｄＳＭＬ｜parameter>
+  </ｄＳＭＬ｜invoke>
+  <ｄＳＭＬ｜invoke name="Read">
+    <ｄＳＭＬ｜parameter name="file_path"＞<![CDATA[/Users/aq/Desktop/myproject/Personal_Blog/index.html]]＞</ｄＳＭＬ｜parameter>
+  </ｄＳＭＬ｜invoke>
+</ｄＳＭＬ｜tool_calls>`;
+  const calls = parseToolCalls(payload, ['Read']);
+  assert.equal(calls.length, 2);
+  assert.equal(calls[0].name, 'Read');
+  assert.deepEqual(calls[0].input, { file_path: '/Users/aq/Desktop/myproject/Personal_Blog/README.md' });
+  assert.equal(calls[1].name, 'Read');
+  assert.deepEqual(calls[1].input, { file_path: '/Users/aq/Desktop/myproject/Personal_Blog/index.html' });
+});
+
+test('parseToolCalls parses CJK-angle DSM drift', () => {
+  const payload = `<DSM｜tool_calls>
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Show commits on local dev not on origin/dev]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git log --oneline origin/dev..dev]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Show commits on origin/dev not on local dev]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git log --oneline dev..origin/dev]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+<DSM｜invoke name="Bash">
+<DSM｜parameter name="description"｜>〈![CDATA[Check tracking branch status]]〉〈/DSM｜parameter〉
+<DSM｜parameter name="command"｜>〈![CDATA[git status -b --short]]〉〈/DSM｜parameter〉
+〈/DSM｜invoke〉
+〈/DSM｜tool_calls〉`;
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 3);
+  assert.equal(calls[0].name, 'Bash');
+  assert.equal(calls[0].input.command, 'git log --oneline origin/dev..dev');
+  assert.equal(calls[1].input.description, 'Show commits on origin/dev not on local dev');
+  assert.equal(calls[2].input.command, 'git status -b --short');
+});
+
+test('parseToolCalls parses fullwidth-bang DSML drift', () => {
+  const payload = `<！DSML！tool_calls>
+  <！DSML！invoke name=“Bash”>
+  <！DSML！parameter name=“command”><！[CDATA[lsof -i :4321 -t]]><！/DSML！parameter>
+  <！DSML！parameter name=“description”><！[CDATA[Verify port 4321 is free]]><！/DSML！parameter>
+  <！/DSML！invoke>
+  <！/DSML！tool_calls>`;
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'Bash');
+  assert.equal(calls[0].input.command, 'lsof -i :4321 -t');
+  assert.equal(calls[0].input.description, 'Verify port 4321 is free');
+});
+
+test('parseToolCalls parses ideographic-comma DSML drift', () => {
+  const payload = `<、DSML、tool_calls>
+  <、DSML、invoke name="Bash">
+    <、DSML、parameter name="command"><、[CDATA[git commit -m "$(cat <<'EOF'
+feat: expand fullwidth bang separator and curly quote tolerance in DSML tool parsing
+
+Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
+EOF
+)"]]><、/DSML、parameter>
+    <、DSML、parameter name="description"><、[CDATA[Create commit with staged changes]]><、/DSML、parameter>
+  <、/DSML、invoke>
+<、/DSML、tool_calls>`;
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'Bash');
+  assert.equal(calls[0].input.command.includes('git commit -m "$(cat <<\'EOF\''), true);
+  assert.equal(calls[0].input.command.includes('Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com'), true);
+  assert.equal(calls[0].input.description, 'Create commit with staged changes');
+});
+
+test('parseToolCalls parses DSML control separator drift', () => {
+  for (const sep of ['␂', '\x02']) {
+    const payload = `<DSML${sep}tool_calls>
+  <DSML${sep}invoke name="Read">
+    <DSML${sep}parameter name="file_path"><![CDATA[/tmp/input.txt]]></DSML${sep}parameter>
+  </DSML${sep}invoke>
+</DSML${sep}tool_calls>`;
+    const calls = parseToolCalls(payload, ['Read']);
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].name, 'Read');
+    assert.deepEqual(calls[0].input, { file_path: '/tmp/input.txt' });
+  }
+});
+
+test('parseToolCalls parses arbitrary-prefixed tool tags', () => {
+  const payload = `<proto💥tool_calls>
+  <proto💥invoke name="Read">
+    <proto💥parameter name="file_path"><![CDATA[/tmp/input.txt]]></proto💥parameter>
+  </proto💥invoke>
+</proto💥tool_calls>`;
+  const calls = parseToolCalls(payload, ['Read']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'Read');
+  assert.deepEqual(calls[0].input, { file_path: '/tmp/input.txt' });
+});
+
+test('parseToolCalls allows all-empty parameter payloads', () => {
+  const payload = `<T｜DSML｜tool_calls>
+  <T｜DSML｜invoke name="TaskOutput">
+    <T｜DSML｜parameter name="task_id"></T｜DSML｜parameter>
+    <T｜DSML｜parameter name="block"></T｜DSML｜parameter>
+    <T｜DSML｜parameter name="timeout"></T｜DSML｜parameter>
+  </T｜DSML｜invoke>
+</T｜DSML｜tool_calls>`;
+  const calls = parseToolCalls(payload, ['TaskOutput']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'TaskOutput');
+  assert.deepEqual(calls[0].input, { task_id: '', block: '', timeout: '' });
+});
+
 test('parseToolCalls ignores bare hyphenated tool_calls lookalike', () => {
   const payload = '<tool-calls><invoke name="Bash"><parameter name="command">pwd</parameter></invoke></tool-calls>';
   const calls = parseToolCalls(payload, ['Bash']);
@@ -101,6 +264,20 @@ test('parseToolCalls tolerates DSML trailing pipe tag terminator', () => {
   assert.deepEqual(calls[0].input, { command: 'find "/home" -type d', timeout: 10 });
 });
 
+test('parseToolCalls tolerates DSML trailing novel separator tag terminator', () => {
+  const payload = [
+    '<DSMLtool_calls※>',
+    '  <DSMLinvoke name="Bash"※>',
+    '    <DSMLparameter name="command"※><![CDATA[pwd]]></DSMLparameter※>',
+    '  </DSMLinvoke※>',
+    '</DSMLtool_calls※>',
+  ].join('\n');
+  const calls = parseToolCalls(payload, ['Bash']);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'Bash');
+  assert.deepEqual(calls[0].input, { command: 'pwd' });
+});
+
 test('parseToolCalls tolerates extra leading less-than before DSML tags', () => {
   const payload = [
     '<<|DSML|tool_calls>',
@@ -396,6 +573,116 @@ test('sieve emits tool_calls for DSML trailing pipe tag terminator', () => {
   assert.equal(text.toLowerCase().includes('dsml'), false);
 });
 
+test('sieve emits tool_calls for DSML control separator drift', () => {
+  for (const sep of ['␂', '\x02']) {
+    const events = runSieve([
+      `<DSML${sep}tool`,
+      '_calls>\n',
+      `<DSML${sep}invoke name="Read">\n`,
+      `<DSML${sep}parameter name="file_path"><![CDATA[/tmp/input.txt]]></DSML${sep}parameter>\n`,
+      `</DSML${sep}invoke>\n`,
+      `</DSML${sep}tool_calls>`,
+    ], ['Read']);
+    const finalCalls = events.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+    assert.equal(finalCalls.length, 1);
+    assert.equal(finalCalls[0].name, 'Read');
+    assert.equal(finalCalls[0].input.file_path, '/tmp/input.txt');
+    const text = collectText(events);
+    assert.equal(text.toLowerCase().includes('dsml'), false);
+    assert.equal(text.includes(sep), false);
+  }
+});
+
+test('sieve emits tool_calls for arbitrary-prefixed tool tags', () => {
+  const events = runSieve([
+    '<proto💥tool',
+    '_calls>\n',
+    '<proto💥invoke name="Read">\n',
+    '<proto💥parameter name="file_path"><![CDATA[/tmp/input.txt]]></proto💥parameter>\n',
+    '</proto💥invoke>\n',
+    '</proto💥tool_calls>',
+  ], ['Read']);
+  const finalCalls = events.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'Read');
+  assert.equal(finalCalls[0].input.file_path, '/tmp/input.txt');
+  const text = collectText(events);
+  assert.equal(text.includes('proto'), false);
+  assert.equal(text.includes('💥'), false);
+});
+
+test('sieve emits tool_calls for CJK-angle DSM drift', () => {
+  const events = runSieve([
+    '<DSM｜tool_calls>\n',
+    '<DSM｜invoke name="Bash">\n',
+    '<DSM｜parameter name="description"｜>〈![CDATA[Check tracking branch status]]〉〈/DSM｜parameter〉\n',
+    '<DSM｜parameter name="command"｜>〈![CDATA[git status -b --short]]〉〈/DSM｜parameter〉\n',
+    '〈/DSM｜invoke〉\n',
+    '〈/DSM｜tool_calls〉',
+  ], ['Bash']);
+  const finalCalls = events.flatMap((evt) => (evt.type === 'tool_calls' ? evt.calls : []));
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'Bash');
+  assert.equal(finalCalls[0].input.command, 'git status -b --short');
+  assert.equal(collectText(events), '');
+});
+
+test('sieve emits tool_calls for fullwidth-bang DSML drift', () => {
+  const events = runSieve([
+    '<！DSML！tool_calls>\n',
+    '  <！DSML！invoke name=“Bash”>\n',
+    '  <！DSML！parameter name=“command”><！[CDATA[lsof -i :4321 -t]]><！/DSML！parameter>\n',
+    '  <！DSML！parameter name=“description”><！[CDATA[Verify port 4321 is free]]><！/DSML！parameter>\n',
+    '  <！/DSML！invoke>\n',
+    '  <！/DSML！tool_calls>',
+  ], ['Bash']);
+  const finalCalls = events.flatMap((evt) => (evt.type === 'tool_calls' ? evt.calls : []));
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'Bash');
+  assert.equal(finalCalls[0].input.command, 'lsof -i :4321 -t');
+  assert.equal(collectText(events), '');
+});
+
+test('sieve emits tool_calls for ideographic-comma DSML drift', () => {
+  const events = runSieve([
+    '<、DSML、tool_calls>\n',
+    '  <、DSML、invoke name="Bash">\n',
+    "    <、DSML、parameter name=\"command\"><、[CDATA[git commit -m \"$(cat <<'EOF'\n",
+    'feat: expand fullwidth bang separator and curly quote tolerance in DSML tool parsing\n\n',
+    'Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com\n',
+    'EOF\n',
+    ')"]]><、/DSML、parameter>\n',
+    '    <、DSML、parameter name="description"><、[CDATA[Create commit with staged changes]]><、/DSML、parameter>\n',
+    '  <、/DSML、invoke>\n',
+    '<、/DSML、tool_calls>',
+  ], ['Bash']);
+  const finalCalls = events.flatMap((evt) => (evt.type === 'tool_calls' ? evt.calls : []));
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'Bash');
+  assert.equal(finalCalls[0].input.command.includes('git commit -m'), true);
+  assert.equal(collectText(events), '');
+});
+
+test('sieve emits all-empty arbitrary-prefixed tool tags without leaking text', () => {
+  const payload = [
+    '<T｜DSML｜tool_calls>\n',
+    '  <T｜DSML｜invoke name="TaskOutput">\n',
+    '    <T｜DSML｜parameter name="task_id"></T｜DSML｜parameter>\n',
+    '    <T｜DSML｜parameter name="block"></T｜DSML｜parameter>\n',
+    '    <T｜DSML｜parameter name="timeout"></T｜DSML｜parameter>\n',
+    '  </T｜DSML｜invoke>\n',
+    '</T｜DSML｜tool_calls>',
+  ].join('');
+  for (const chunks of [[payload], payload.match(/.{1,8}/gs)]) {
+    const events = runSieve(chunks, ['TaskOutput']);
+    const finalCalls = events.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+    assert.equal(finalCalls.length, 1);
+    assert.equal(finalCalls[0].name, 'TaskOutput');
+    assert.deepEqual(finalCalls[0].input, { task_id: '', block: '', timeout: '' });
+    assert.equal(collectText(events), '');
+  }
+});
+
 test('sieve emits tool_calls for extra leading less-than DSML tags without leaking prefix', () => {
   const events = runSieve([
     '<<|DSML|tool_calls>\n',
@@ -734,7 +1021,7 @@ test('sieve keeps embedded invalid tool-like json as normal text to avoid stream
   assert.equal(leakedText.toLowerCase().includes('tool_calls'), true);
 });
 
-test('sieve passes malformed executable-looking XML through as text', () => {
+test('sieve releases malformed executable-looking XML wrappers as text', () => {
   const chunk = '<tool_calls><invoke name="read_file"><param>{"path":"README.MD"}</param></invoke></tool_calls>';
   const events = runSieve([chunk], ['read_file']);
   const leakedText = collectText(events);