Sanitize leaked tool-call wire format in assistant text

[CJackHwang]

Motivation

• Long toolcall conversations can surface intermediate wire-format fragments (tool-call arrays, result blobs, and meta markers) into visible assistant text when relay/client layers mis-handle or reserialize tool events.
• Once leaked into visible text these fragments can be treated as context and then amplified by subsequent model turns.
• Add server-side defensive sanitization to remove these artifacts before they contaminate downstream conversation state.

Description

• Add three regex detectors leakedToolCallArrayPattern, leakedToolResultBlobPattern, and leakedMetaMarkerPattern in internal/adapter/openai/tool_history_sanitize.go and apply them in sanitizeLeakedToolHistory to strip wire-format arrays, <| Tool |>… result blobs, and <| … |> meta markers.
• Preserve existing cleanup behavior for [TOOL_CALL_HISTORY]/[TOOL_RESULT_HISTORY] blocks and empty JSON fences (emptyJSONFencePattern).
• Add regression tests in internal/adapter/openai/tool_history_sanitize_test.go that exercise leaked wire-format tool-call + result sequences and standalone meta-marker leaks.

Testing

• Ran go test ./internal/adapter/openai -run SanitizeLeakedToolHistory -count=1 and the test run completed successfully.
• The new tests in tool_history_sanitize_test.go passed, confirming the sanitizer removes the targeted leak patterns.

---

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 2:36pm

[chatgpt-codex-connector]

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