[diegosouzapw]

Closes #4385

Problem

Routing a Claude-protocol conversation (/v1/messages) to an OpenAI-compatible provider (command-code, custom openai-compatible) returned a 502 from the upstream:

Messages with role 'tool' must be a response to a preceding message with 'tool_calls'

claudeToOpenAIRequest emits a role:"tool" message for every Claude tool_result block, but never drops an orphan one — a tool_result whose tool_use_id has no matching assistant.tool_calls (e.g. when history truncation / compression removed the assistant turn that issued the call but kept the result). OpenAI-compatible upstreams reject those sequences. This is why downgrading to 3.8.25 "worked" — it depends on the exact history shape the agent sends.

Fix

After fixMissingToolResponses, filter out role:"tool" messages whose tool_call_id matches no assistant.tool_calls id. This mirrors the filter already applied on the Responses→Chat path (openai-responses.ts, #2893). Placeholders inserted by fixMissingToolResponses (which DO match a tool_call) are kept; only genuine orphans are removed.

Validation (TDD)

New tests/unit/translator-claude-to-openai-orphan-tool-4385.test.ts — fails on the unfixed code (orphan survives), passes after the fix (3/3): drops an orphan, preserves a paired tool_result, and keeps valid while dropping orphan in the same turn.

One existing test (...converts tool_result blocks into tool messages...) built an unpaired tool_result (technically an orphan that produces the 502); it now pairs it with the assistant tool_call so it still exercises the content-extraction mechanics on a valid sequence — assertions on "20C"/"Thanks" are unchanged. Full translator suites green; typecheck:core + lint clean.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

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