Ignore tool_call payloads inside fenced code blocks and chat envelopes; stream-aware code-fence tracking

[CJackHwang]

Motivation

• Prevent accidental parsing/execution of tool_calls that appear only inside fenced code examples or inside chat transcript envelopes.
• Improve streaming detection of fenced code blocks so tool detection does not misfire when backticks are split across chunks.
• Support more robust detection for nested/qualified markup tool syntax and extend the context tail window for streaming state.

Description

• Add heuristics to skip tool-call parsing when the only occurrences are inside fenced code blocks by implementing stripFencedCodeBlocks / shouldSkipToolCallParsingForCodeFenceExample in both JS and Go parsers.
• Detect and ignore chat-transcript-style envelopes that contain a tool_calls field using isLikelyChatMessageEnvelope / isLikelyChatMessageEnvelope in both languages.
• Make sieve code stream-aware of fenced code state by increasing TOOL_SIEVE_CONTEXT_TAIL_LIMIT from 256 to 4096 and adding incremental fence tracking (codeFenceStack, codeFencePendingTicks, codeFenceLineStart), plus insideCodeFenceWithState, updateCodeFenceState, and related simulation helpers.
• Update parsing flow to consult the stream-aware fence state in findToolSegmentStart and consumeToolCapture, and add markup-syntax checks to avoid false positives.
• Mirror the JS changes in Go utility functions (toolcalls_candidates.go, toolcalls_parse.go) and add unit tests and expected fixtures updates to cover ignoring fenced JSON blocks, chat envelope cases, chunk-split fences, nested fences, and legacy behavior adjustments.

Testing

• Updated and ran unit tests for the Go package (go test ./...) and Node tests (npm test / test suite used in repository), including new tests for fenced JSON ignoring, transcript envelope ignoring, chunk-split fence handling, and nested fences, and all tests passed.
• Updated compatibility fixtures (tests/compat/expected) to reflect the new behavior for fenced examples and validated against the test expectations.
• Existing parsing and sieve tests were exercised and confirmed to still detect legitimate tool_calls outside of ignored contexts.

---

Codex Task

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ds2api	Ready	Preview, Comment	Mar 22, 2026 8:37am

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b108a7915a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".