fix(codex): rescue reasoning-only turns that die with 'remained incomplete after 3 continuation attempts'

[IpastorSan]

Problem

On the xAI /v1/responses surface (xai-oauth), grok-4.x sometimes ends a turn with only reasoning items — no message output item, no tool calls. When that happens, the turn deterministically dies with:

Codex response remained incomplete after 3 continuation attempts

Two compounding causes:

1. The answer can land in the reasoning channel. The model occasionally emits its final answer inside the reasoning item, delimited by grok's internal <response> tag. The answer exists, but _normalize_codex_response classifies the turn as reasoning-only → finish_reason="incomplete".

2. Continuations are a no-op. grok's reasoning items on this surface carry no encrypted_content, so the interim assistant message (plain-text reasoning only, no content, no replayable items) emits nothing in _chat_messages_to_responses_input. Every continuation request is therefore byte-identical to the one that just came back incomplete — the model repeats the reasoning-only response until the retry budget is exhausted.

Observed live with grok-4.20 on xai-oauth (2026-07): the failure pattern shows up as triplets of finish_reason="incomplete" assistant messages with populated reasoning and empty content, across multiple unrelated sessions in ordinary usage. This is a sibling of the grok-composer web_search incomplete-loop fix already documented in transports/codex.py.

Related issues

• Codex hidden-reasoning-only incomplete turns can poison gateway sessions #51628 — same user-facing failure ("remained incomplete after 3 continuation attempts") from hidden-reasoning-only turns on the Codex ChatGPT backend. The continuation-nudge half of this PR is provider-agnostic and fires exactly when the interim turn has no replayable payload, which covers the mechanism described there (encrypted-reasoning-only turns keep their existing replay path and are untouched).
• [Bug]: status=incomplete (incomplete_details=None) on an OpenAI-compatible /v1/responses endpoint fans one turn into ~60 near-identical messages — codex_responses incomplete-continuation dedup/cap insufficient #52711 — the opposite failure mode of the same continuation branch (near-identical incompletes fanning out). Not addressed here, but the nudge does not worsen it: interims with drifting encrypted items are replayable, so the nudge never fires for them.

Fix

• _normalize_codex_response (scoped to issuer_kind == "xai_responses"): when a turn has no visible content and no tool calls but the reasoning text contains grok's <response> delimiter, promote the delimited tail to assistant content (the untagged prefix stays as thinking text). The turn completes normally on the first attempt.
• Codex-incomplete continuation path (conversation_loop.py): when the interim message has nothing the input converter will replay (no content, no encrypted reasoning items, no message items), append a user-role nudge so the retry actually differs and explicitly asks for the final answer / pending tool call. Mirrors the existing _get_continuation_prompt pattern used for length truncation, and is guarded against duplicate insertion. Turns whose interim is replayable (visible content, encrypted reasoning) are untouched.

Tests

• 4 new tests in tests/agent/test_codex_responses_adapter.py: salvage with/without closing tag, xAI-only scoping, no-marker case stays incomplete.
• 2 new tests in tests/run_agent/test_run_agent_codex_responses.py: nudge is appended exactly once after an unreplayable reasoning-only interim (and the turn then completes); nudge does NOT fire for a replayable interim.
• scripts/run_tests.sh tests/agent/test_codex_responses_adapter.py tests/run_agent/test_run_agent_codex_responses.py (hermetic CI wrapper): 114 passed, 0 failed.
• scripts/check-windows-footguns.py --diff: clean (change is pure string/dict handling, no OS-specific code).

Platforms tested

• Linux (Arch, Python 3.11 via uv). No file I/O, process, or terminal handling touched, so no platform-specific surface beyond the test suite.

🤖 Generated with Claude Code

[teknium1]

Merged via PR #64768 — your commit was cherry-picked onto current main with your authorship preserved in git log (rebase merge). We added a small follow-up: an alternation guard on the continuation nudge so an interim turn too empty to append can never produce a user→user / tool→user sequence on strict providers. Excellent PR — the two-cause analysis (answer in the reasoning channel + byte-identical retries from unreplayable interims) and the live grok-4.20 observations made this a straightforward salvage.