codex: treat an empty successful completion after a tool_result tail as a valid end_turn

[dmitryanchikov]

Summary

When a conversation ends assistant: tool_use → user: tool_result and the tool call itself
delivered the user-visible answer (e.g. a messaging/reply tool), gpt-5.x models legitimately
end the follow-up turn with zero output items — a successful terminal event with nothing
to add. The empty-completion protection from #70/#71 treats this as a failure: it retries
(bounded 1+10), burns ~3 minutes, then surfaces 503 Codex completed without producing output — and the Anthropic-side client (Claude Code) retries the whole request on top,
producing multi-minute hangs on trivial turns.

For a Telegram channel bot whose reply tool IS the answer, this tail shape occurs on
essentially every turn: ~68% of its real turns failed this way over 2h (2026-08-25,
websocket transport, main @ 0185409).

Deterministic repro

{"model":"gpt-5.5","max_tokens":1024,"stream":false,
 "system":"You are a channel agent. When a user pings you, reply via the reply tool. The tool call IS your answer to the user.",
 "tools":[{"name":"reply","description":"Send a reply message to the chat. This delivers your answer to the user.","input_schema":{"type":"object","properties":{"text":{"type":"string"}},"required":["text"]}}],
 "messages":[
  {"role":"user","content":"ping"},
  {"role":"assistant","content":[{"type":"tool_use","id":"toolu_01","name":"reply","input":{"text":"pong"}}]},
  {"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_01","content":"Message sent."}]}]}

Before: 503 after 175s, empty. (Adding a system line "always end with a short text
line" makes the identical request return 200 in 1.8s — the model is fine; the empties are
semantic, not transient.)

Change

Both empty-completion sites (buffered and live-stream) now consult the translated request's
tail: a function_call_output tail means the empty completion is the turn's real end, and it
translates into a valid empty end_turn instead of entering the retry loop. Empty
completions without a tool-result tail keep the full #70/#71 retry-then-503 behavior.

After (verified in production, both transports)

• non-streaming: 200 in ~1.7s, {"content":[],"stop_reason":"end_turn","role":"assistant",...}
• streaming: 200 in ~1.6s, well-formed SSE (message_start → content_block_start/stop →
  message_delta → message_stop) — accepted by Claude Code.

Full cargo test green. Happy to add a regression test in the #71 harness style if you can
point me at the preferred fixture shape.

[dmitryanchikov]

Force-pushed a v2 after internal review flagged the original as too broad — the empty-acceptance is now doubly gated:

1. Predicate narrowed: runs on the original Anthropic body and requires a tool_result without is_error — an error result demands a model reaction, so a deterministic empty there stays a loud failure.
2. One retry as the glitch discriminator: a qualifying tail alone isn't proof (an internal tool result has the same shape and can meet a transient upstream glitch, which is what fix(codex): retry terminal-only live completions #70/fix(codex): retry empty buffered completions #71 exist for). The first empty still retries once; only the second consecutive empty is accepted as a valid empty end_turn. Both transports share one accept_empty_completion().

Added tests: tail predicate (success/error/non-tool/ordering/empty), the accept truth table, and a live-path behavioral pair (flag off → Retry with the empty-completion detail; flag on → delivered stream ending in message_stop). Verified in production: the repro returns 200 in ~4.5s (one discriminator roundtrip included), and an is_error-tail control produced an actual model reaction (thinking + a retry tool call) rather than a silent empty.

[pierzchalski]

I reproduced this with Claude Code 2.1.278 and a proxy built from 55bf0b5 (v0.1.35 + 6 commits), using CCP_CODEX_TRANSPORT=http. Claude Code’s Agent-tool subagents deliver their report with a SubagentHandback tool call and are instructed to “call it and then stop,” so the request after that tool result is exactly this tail and the model legitimately returns nothing. In one session, 15 handbacks succeeded across 7 gpt-5.6-sol and gpt-5.6-terra subagents; 8 then stalled for 19–56 minutes, with 5 ending in 503 Codex completed without producing output and 3 interrupted by the user first, while the other 7 happened to emit a follow-up text turn and completed normally. Claude Code showed each stalled subagent as still running, then reported “Agent terminated early due to an API error” despite the report already being delivered, and each stall re-sent the full subagent context up to 10 times. f7ee7e8 cherry-picks cleanly onto v0.1.41 (2b639129); we are deploying it locally.