Relax tool-name allow-listing and improve tool-call detection/parsing across adapters and sieve

[CJackHwang]

Motivation

• Ensure streaming tool-detection remains active even when client tools schemas are malformed, empty, or when tool_choice=none, so tool payloads are still intercepted and emitted.
• Stop canonicalizing or rejecting unknown tool names during parsing so original tool names/casing are preserved and unknown tools can be emitted as executable tool calls.
• Improve robustness of the stream sieve and XML/markup parsers to detect more tool-call forms (including Claude <tool_use> variants and array payloads) and to detect tool calls when mixed with leading prose.

Description

• Enabled a sentinel tool name ("__any_tool__") when tools are present but provide no valid names by adding ensureToolDetectionEnabled and equivalent logic in the JS policy resolver.
• Relaxed incremental filtering in filterIncrementalToolCallDeltasByAllowed to preserve seen names and avoid blocking when allowedNames is empty, and adjusted formatting logic accordingly.
• Stopped strict allow-list enforcement in parser layers and changed parsing to preserve original tool name casing and to keep unknown tool names; removed rejection paths from ParseToolCallsDetailed/filterToolCallsDetailed and updated JS helpers to match.
• Extended the tool-call candidate extraction to include top-level arrays and look for "function" JSON keyword, and adjusted sieve keyword lists to catch more payload shapes.
• Added support for additional <tool_use> XML/markup styles in parseXMLToolCalls and helpers to extract tool_name/function_name and parameter bodies.
• Adjusted the Claude stream runtime to buffer and finalize tool detection using the standalone parsing routines (and added a test TestHandleClaudeStreamRealtimeDetectsToolUseWithLeadingProse).
• Updated many tests and test expectations to reflect the new permissive behavior and added new unit tests for the new XML/markup styles and stream cases.

Testing

• Ran Go unit tests with go test ./... covering internal/adapter/claude, internal/adapter/openai, and internal/util, and they all passed.
• Ran the JS unit tests with the project's Node test harness (npm test / yarn test) for the stream-tool-sieve helpers and they passed.
• Updated and added multiple tests (e.g. TestHandleClaudeStreamRealtimeDetectsToolUseWithLeadingProse, several ParseToolCalls cases, and stream-tool-sieve node tests) to assert the new permissive parsing and they succeed.

---

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 1:27pm

[chatgpt-codex-connector]

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