Preserve code fences around standalone tool JSON and add marker-output guards

[CJackHwang]

Motivation

• Avoid consuming legitimate closed code fences when a standalone tool JSON payload follows and prevent leaking tool-history markers into assistant output.

Description

• Implemented fence-aware parsing by adding trimWrappingJSONFence and openFenceStartBefore to internal/adapter/openai/tool_sieve_core.go and the corresponding trimWrappingJSONFence in internal/js/helpers/stream-tool-sieve/sieve.js to only strip opening fences that truly wrap standalone tool JSON.
• Updated findToolSegmentStart and consumeToolCapture to use the new fence helpers and to preserve trailing prose/closing fences when appropriate.
• Added guard instructions to prompts to prevent models from emitting [TOOL_CALL_HISTORY] or [TOOL_RESULT_HISTORY] markers in internal/adapter/claude/handler_utils.go and internal/adapter/openai/handler_toolcall_format.go and updated related tests to expect the new instructions.
• Added and adjusted tests to verify behavior, including TestHandleStreamStandaloneToolCallAfterClosedFenceKeepsFence in Go and the sieve preserves closed fence before standalone tool payload test in Node, plus minor updates to existing prompt/build tests.

Testing

• Ran Go unit tests for the modified packages including internal/adapter/openai and internal/adapter/claude, and the new/updated tests (e.g. TestHandleStreamStandaloneToolCallAfterClosedFenceKeepsFence) passed.
• Ran Node unit tests for the stream tool sieve (tests/node/stream-tool-sieve.test.js), including the new fence-preservation test, and they passed.
• Stream handling assertions confirming finish_reason=tool_calls and no leakage of raw tool_calls or stray code fences were validated and succeeded.

---

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 21, 2026 9:39am