Do not promote fenced code examples to tool calls and centralize tool-keyword detection

[CJackHwang]

Motivation

• Prevent accidental promotion of fenced code block examples into actionable tool_calls/tool_use events so example content remains as text.
• Reduce duplicated keyword logic by centralizing tool-segment keywords and earliest-index lookup for more robust parsing.

Description

• Add internal/js/helpers/stream-tool-sieve/tool-keywords.js with TOOL_SEGMENT_KEYWORDS and earliestKeywordIndex to centralize keyword detection.
• Update JS parser (parse.js) and sieve (sieve.js) to use TOOL_SEGMENT_KEYWORDS and a small TOOL_MARKUP_PREFIXES list and simplify fence-checking logic so fenced examples are preserved as content-only when appropriate.
• Adjust Go parsing helpers (internal/util/toolcalls_candidates.go) to stop promoting fenced code examples by simplifying shouldSkipToolCallParsingForCodeFenceExample behavior.
• Update tests across adapters and format packages (handler_stream_test.go, handler_toolcall_test.go, render_test.go) to expect fenced JSON examples to remain content-only and add backward-compatible alias test functions for historical test names.

Testing

• Ran the Go unit tests for the modified packages via go test ./... and verified the updated tests in internal/adapter/claude, internal/adapter/openai, and internal/format/openai passed.
• Exercised the updated stream-tool-sieve logic through the modified unit tests which confirmed fenced examples are no longer promoted to tool calls.

---

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 8:33am

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.