fix(server): replay function_call items when translating Responses input

[Carleon]

Problem

Agentic clients that use /v1/responses never make progress: the model re-issues the same tool call on every turn and the conversation loops forever.

I hit this with Codex CLI 0.150.1, which is a realistic trigger because Codex removed wire_api = "chat" support, so it can only speak the Responses format. A one-step task (cat a.txt) executed the same command ~100 times in five minutes without finishing.

Root cause

In responsesInputToMessages (internal/server/responses.go), function_call_output is translated into a role: "tool" message, but function_call is skipped:

case "function_call", "reasoning", "item_reference":
    // Locally unexecutable or non-replayable items are skipped
    // (the reference does the same — the upstream cannot run
    // them).
    continue

The result is an orphaned tool reply: a tool_call_id with no preceding assistant tool_calls entry. The translated conversation reaching upstream looks like

user:  What is the weather in Berlin?
tool:  {"temp_c": 19, "sky": "clear"}     <- answers nothing

Chat backends drop a tool message that answers no call, so the model never sees the result and asks again.

The existing comment's reasoning is about execution, but replaying a function_call does not ask the upstream to run anything. It only records that the call happened, which is exactly what makes the following tool message well-formed.

Reproduction

Same conversation, both wire formats, against v1.6.0. Chat is correct, Responses is not.

/v1/chat/completions:

{"messages":[
  {"role":"user","content":"What is the weather in Berlin?"},
  {"role":"assistant","tool_calls":[{"id":"call_w1","type":"function","function":{"name":"get_weather","arguments":"{\"city\":\"Berlin\"}"}}]},
  {"role":"tool","tool_call_id":"call_w1","content":"{\"temp_c\": 19, \"sky\": \"clear\"}"}
]}

→ "The weather in Berlin is currently 19°C with clear skies" ✅

/v1/responses:

{"input":[
  {"type":"message","role":"user","content":[{"type":"input_text","text":"What is the weather in Berlin?"}]},
  {"type":"function_call","call_id":"call_w1","name":"get_weather","arguments":"{\"city\":\"Berlin\"}"},
  {"type":"function_call_output","call_id":"call_w1","output":"{\"temp_c\": 19, \"sky\": \"clear\"}"}
]}

→ returns the identical function_call for get_weather again ❌

Fix

Emit the assistant tool_calls message rather than skipping it, so each call is replayed immediately before the tool message that answers it.

reasoning and item_reference keep their existing skip behaviour, since they carry no state the upstream can use. A function_call missing call_id or name is also still skipped rather than emitted malformed, and an absent arguments becomes "{}" because the Chat format requires a JSON string there.

Tests

New internal/server/responses_function_call_replay_test.go covers:

• the call/output pair surviving translation in the right order (the regression)
• omitted arguments defaulting to "{}"
• function_call missing call_id or name, plus reasoning and item_reference, still being skipped
• parallel tool calls each being replayed before their own result

Verified red/green: with the fix reverted, three of the four fail with the orphaned-tool-reply output shown above. go build ./... and go test ./... are clean across all 21 packages.

Verification against a live upstream

Built and deployed against real traffic. After the fix, the same Codex task completes in a single tool call, and a multi-step task (count lines in one file, read a number from another, sum them) also completes correctly in one pass.

🤖 Generated with Claude Code

[trefeon]

Reviewed and merged manually as c4ac39d (this PR could not merge directly: it was based on e0f134c, before the repo split moved all Go code under backend/).

Verdict: approved. The diff matches the report exactly — function_call replayed as an assistant tool_calls message immediately before its function_call_output -> tool translation, reasoning/item_reference stay skipped, malformed calls are dropped rather than emitted, absent arguments default to "{}", and ordering is preserved for parallel calls. All four regression tests pass as-is.

The port also fixed two stale doc comments that still claimed function_call is non-replayable/skipped, so the documented translation guarantees now match the wire behavior.

Verified end-to-end: hermetic Go suite, golangci-lint (0 issues), and the -race CI job all pass on main. Thank you for the report and reproduction — this unblocked Codex-style agentic clients on /v1/responses.