PR #460 introduced fullwidth pipe characters (｜) in DSML tool call formatting
to improve parsing robustness, but models exposed to these fullwidth pipes in
system prompts exhibit significantly higher rates of tool output hallucinations.
Reverting to halfwidth pipes (|) drastically reduces tokenizer/perplexity-driven
hallucinations while retaining the existing confusable-hardening in the parser.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Detect camelCase→PascalCase boundaries between arbitrary prefixes and fixed
local names (tool_calls/invoke/parameter), so that fused forms like
<DSmartToolCalls> are recognized without explicit separator characters.
Also add the underscore-free alias "toolcalls" as a valid DSML local name.
Includes lookalike rejection tests to ensure near-matches like
<DSmartToolCallsExtra> are not falsely accepted.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace the hardcoded isToolMarkupPipe (matching |, ｜, ␂, \x02, !) and
isToolCDATAOpenSeparator (exclusion-based) with a single isToolMarkupSeparator
that treats any Unicode punctuation outside structural characters as a valid
DSML separator. This eliminates the need for a per-character allowlist — novel
separators like ※ are automatically supported without code changes. Also
removes the unused cdataPattern regexp and updates docs to use "non-structural
separator" terminology.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Stream: strip both and [reference:N] markers to prevent
leaking partial link metadata during incremental output
- Non-stream: convert citation/reference markers to Markdown links for
Claude Messages, Gemini generateContent, and OpenAI Chat/Responses
- Remove StripReferenceMarkers option from call sites; behavior is now
determined automatically by stream vs non-stream context
- Extend JS runtime stripReferenceMarkersText() to also match [citation:N]
- Add tests for streaming marker stripping and non-stream link conversion
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add compatibility for <dsml-tool-calls>/<dsml-invoke>/<dsml-parameter>
tag forms alongside the canonical pipe-prefixed DSML shell. Hyphenated
forms only activate when a DSML prefix is detected, preventing false
matches on bare XML lookalikes. Go and Node parsers aligned, with tests
covering here-doc CDATA, streaming sieve, and negative lookalike cases.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Replace bufio.Scanner with bufio.NewReaderSize + ReadBytes('\n') across all
SSE read paths to preserve long single-line data (e.g. write_file content).
Add quasi_status and auto_continue handling as direct path-based patches in
both Go continue observer and Node vercel_stream_impl, mirroring existing
batch-patch logic. Add 2MiB+ line throughput tests at every SSE layer.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Replace hardcoded DSML typo variant lists in Go/Node tool call parsers with
generalized prefix consumption that tolerates repeated leading <, repeated DSML
prefix noise, and trailing pipe terminators. Split tiktoken-dependent token
counting into a build-tagged file for non-cgo platform compatibility. Add /data
directory to Dockerfile for bind-mount permissions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Previously retry/continue requests reused the initial PoW header and
lacked parent_message_id, causing them to land as disconnected root
messages in the DeepSeek session instead of proper follow-up turns.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Auto-retry Chat/Responses streams once when upstream output is empty but not content-filtered, reusing session/token/PoW and appending a regeneration suffix to the prompt
- Wire DeepSeek continue API into Vercel streams for multi-round thinking output exhaustion
- Defer empty-output errors in stream finalizers to enable synthetic retry; only surface failure when the retry budget is exhausted
- Track content_filter stops to avoid retry on filtered outputs
- Add comprehensive tests for stream/non-stream retry, Responses retry, and content_filter no-retry
- Update prompt-compatibility.md documentation
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Add shared ToolMarkupTag scanner (toolcalls_scan.go) replacing hardcoded alias tables
- Support DSML collapsed tag names (<DSMLtool_calls>, <DSMLinvoke>, <DSMLparameter>)
- Parse JSON literal values from parameter bodies (123→number, true→bool, null)
- Recover unclosed CDATA in final parse/flush via SanitizeLooseCDATA
- Align Go and Node implementations (scanToolMarkupTagAt, findMatchingToolMarkupClose)
- Reject bare <invoke> as unsupported syntax, only tool_calls wrapper triggers tool path
- Update API.md and toolcall-semantics.md documentation
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Guard current_input_file.enabled / thinking_injection.{enabled,prompt} with hasNestedSettingsKey so partial updates don't overwrite omitted fields
- Expand DSML alias support to tolerate space-separated tags (e.g. <|dsml invoke>) alongside pipe-separated forms
- Sync Go sieve, Node sieve, toolcall parser, and tests for all new DSML variants
- Update API.md and toolcall-semantics.md with expanded alias coverage
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add support for DSML wrapper aliases (<dsml|tool_calls>, <|tool_calls>,
<｜tool_calls>) alongside canonical XML. Normalize mixed DSML/canonical
tags instead of rejecting them. Add tilde fence (~~~) support, fix
nested fence and unclosed fence handling, support CDATA-protected fence
content, and skip prose mentions when scanning for real tool blocks.
Mirror all changes between Go and Node.js runtimes.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- replace legacy DeepSeek ids with the new deepseek-v4 model family\n- move thinking control to request parameters and preserve assistant reasoning content\n- switch history split to IGNORE transcript injection and map upload auth failures to 401\n- update admin defaults, API docs, samples, and tests for the new model scheme
