fix(responses): insert synthetic tool messages for missing function_call_output items

[ThinkPeace]

Summary

When converting Responses API input items to Chat Completions messages, function_call items without matching function_call_output items produce assistant messages with tool_calls but no tool responses. Downstream providers reject such incomplete tool call cycles.

Fix

Add fixMissingToolCallOutputs — a post-processing step in convertInputToMessages that:

1. Collects all tool_call_ids that already have responses
2. Inserts synthetic empty role: "tool" messages for any tool call lacking a response

Changes

• llm/transformer/openai/responses/inbound.go: Add fixMissingToolCallOutputs, called at end of convertInputToMessages
• llm/transformer/openai/responses/inbound_test.go: Add 9 test cases

All existing tests in transformer/openai/responses/ continue to pass.

Closes #2049

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR repairs incomplete tool-call cycles before forwarding requests. The main changes are:

• Inserts empty tool responses for unanswered function calls.
• Matches responses within each assistant turn so reused IDs remain independent.
• Adds missing function-tool definitions during request processing.
• Covers complete, partial, missing, terminal, and repeated-ID cases.

Confidence Score: 5/5

This looks safe to merge.

• Reused tool-call IDs are now matched independently for each assistant turn.
• A second middleware application does not add duplicate synthetic responses.
• No blocking issue was found in the updated code.

Important Files Changed

Filename	Overview
llm/transformer/openai/responses/inbound.go	Adds turn-local matching and synthetic responses for unanswered tool calls.
llm/pipeline/cc/fix_tool_calls.go	Adds middleware for missing tool responses and function definitions.
internal/server/orchestrator/orchestrator.go	Registers the new request middleware in the orchestrator.
llm/transformer/openai/responses/inbound_test.go	Adds coverage for the repaired tool-call matching behavior.

_{Reviews (3): Last reviewed commit: "fix(pipeline): universal fix for missing..." | Re-trigger Greptile}

[ThinkPeace]

Good catch! Fixed in 459675e.

The updated algorithm does per-assistant-message independent checking instead of using a global responded map:

1. For each assistant message, collect its tool_call_ids into a missing set
2. Look ahead at immediately-following role: "tool" messages (stop at first non-tool), removing covered IDs from missing
3. Insert synthetic tool messages for any remaining uncovered IDs

This means [assistant(x), tool(x), assistant(x), user] correctly gets a synthetic tool(x) inserted after the second assistant turn — the first turn's tool(x) is before the second assistant, not after it, so it does not cover the second call.

Added a dedicated test case: "same ID across two assistant turns — each checked independently" that validates this exact scenario.

[ThinkPeace]

Updated with the full fix. The original approach only covered the /v1/responses path. Testing revealed the /v1/chat/completions path also needs the fix.

Changes since last review

1. Exported FixMissingToolCallOutputs so it can be called from other packages
2. Added llm/pipeline/cc/fix_tool_calls.go — a pipeline middleware that runs on ALL request paths (both /v1/responses and /v1/chat/completions)
3. Registered middleware in orchestrator as a global middleware
4. Added tool definition fix: the middleware now also ensures every tool name referenced in conversation history has a matching entry in the tools array (prevents tool call id X is not found errors when Codex-specific tools like exec_command appear in history but not in the client's tool definitions)
5. 8 test cases covering the exported function

Middleware does two things on every request:

1. Calls FixMissingToolCallOutputs to insert synthetic tool messages for any assistant tool_calls lacking responses
2. Scans conversation history for tool names not in request.Tools and adds minimal dummy definitions

[looplj]

这是客户端问题吧