fix: resample empty provider responses like Grok Build to prevent premature session completion

[Blankeos]

Sessions on xAI/Grok would silently finish after reasoning-only or no-visible-content responses because response.completed lacks end_turn and phase metadata. The stream treated these as finished turns and stopped the agent loop.

This change adds a resample loop in stream_with_tools that detects empty completions (no text, no tool calls) and retries the same request, matching Grok Build's AttemptOutcome::Empty behavior. Doom loop triggers are now parsed from ResponseCompleted payloads and used to inject recovery reminders when reasoning-only samples hit tail repetition.

• Extract doom_loop_triggers_from and summarize_response_output_types in openai.rs
• Log response.completed with status, end_turn, incomplete_reason, output_types, doom_loop
• Add doom_loop_triggers field to ChunkType::ResponseCompleted
• Add retryable patterns for "empty response", "reasoning_only", "no_visible_content"
• Add x-grok-exact-repetition-check header to xAI build requests
• Add tests: resamples_reasoning_only_completed_response_like_grok_build, reasoning_only_with_thinking_doom_loop_resamples_with_reminder

[chatgpt-codex-connector]

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

[cloudflare-workers-and-pages]

Deploying crabcode with    Cloudflare Pages

Latest commit:	ed6d8e5
Status:	✅  Deploy successful!
Preview URL:	https://7679fe77.crabcode.pages.dev
Branch Preview URL:	https://fix-grok-4-6-emptyresponselo.crabcode.pages.dev

View logs

[Blankeos]

PR Review (updated)

fix/grok-4.6-emptyresponseloop → main

PR: #42

Merge confidence: 4.5/5 (was 3.5/5)

---

Suggested GitHub PR body

Grok 4.6 (xAI Responses / cli-chat-proxy) was silently ending the session after a reasoning-only response.completed — no text, no tool calls, end_turn=None, no phase metadata. The finish gate treated that as StopReason::Finish.

This matches Grok Build: classify those payloads as empty (reasoning_only / no_visible_content), resample the same request (rollback streamed reasoning), and fail after the retry budget instead of pretending the turn completed. Doom-loop tail_repetition@thinking on an empty sample injects the recovery reminder. Preamble text with no tools still finishes. Hosted provider-tool steps and content-filter empties are not resampled.

Also: parse doom_loop_check off response.completed, log output item types, send x-grok-exact-repetition-check: 64.

---

What changed

File	Role
src/aisdk/response.rs	Nested 'resample loop; empty terminal → rollback + retry; doom-loop reminder on empty+confident trigger; hosted tools / content-filter skipped
src/aisdk/chunk.rs	ResponseCompleted.doom_loop_triggers
src/aisdk/providers/openai.rs	Parse/log completed status, output_types, doom_loop_check
src/aisdk/retry.rs	empty response / reasoning_only / no_visible_content are retryable
src/llm/xai_build.rs	x-grok-exact-repetition-check: 64
_plans/PREMATURE_COMPLETE_BUG.md	Incident notes

1 commit: 3780c94

---

Checks run (read-only)

Check	Result
git diff --check	clean
cargo fmt	clean
just check-aisdk-boundary	ok
cargo test --bin crabcode aisdk::response::tests	42/42
cargo test --bin crabcode aisdk::providers::openai::tests	45/45 (prior run)
cargo test --bin crabcode aisdk::retry::tests	8/8 (prior run)
cargo test --bin crabcode llm::xai_build::tests	11/11 (prior run)
Live grok-4.6 continue smoke	no longer terminates; that session still flaky

New tests that passed: resamples_reasoning_only_completed_response_like_grok_build, resamples_no_visible_content_completed_response, empty_response_exhaustion_fails_instead_of_finish, content_filter_empty_does_not_resample, hosted_tool_only_completed_response_is_not_empty, reasoning_only_with_thinking_doom_loop_resamples_with_reminder. Empty-resample tests use start_paused so they do not wait real backoff.

---

Regressions?

Covered path is fixed. Reasoning-only / no-visible-content response.completed no longer Finishes. Preamble text still finishes. Hosted-tool-only and Chat Completions content-filter empties do not resample. Exhaustion is StopReason::Error, not a silent finish.

Intentional behavior change: persistent empty completions fail after retries instead of idle-complete. Better than silent stop.

Remaining risks (not blockers):

1. Global in stream_with_tools. Empty resample applies to every provider, not just xAI Responses. Likely correct (Grok Build treats empty as transport-agnostic), but a legitimate empty stop from another provider now retries then errors.
2. Responses content-filter. Skip is FinishReason::ContentFilter on ChunkType::End. response.completed never sets that, so a content-filtered empty Responses payload would still resample.
3. Doom recovery exhausted ≠ accept. After 2 recoveries the tool-call doom path disarms. On the empty path, exhausted recoveries fall through to empty retry, then Failed.
4. Backoff is real. Empty samples share PROVIDER_STEP_MAX_RETRIES (10) and 2s exponential sleep. Matches Grok Build EmptyResponse retryability. A sticky empty can stall ~3.5 minutes then fail. Live smoke already showed this is preferable to silent finish.

---

Migrations?

No.

---

Merge checklist

• Open the GitHub PR (#42)
• Tests: no_visible_content, exhaustion → Error, content-filter skip, hosted ProviderToolCall skip
• Assert x-grok-exact-repetition-check=64
• Keep network backoff (matches Grok Build)
• Live grok-4.6 continue smoke — no longer terminates
• Optional: skip empty resample when Responses incomplete_reason / status is content-filter
• Optional: after DOOM_LOOP_MAX_RECOVERIES, accept the empty sample instead of failing

---

Confidence bumps

If you do this	New score
Responses-path content-filter skip (incomplete_reason / completed status)	4.7/5
+ doom-exhausted empty sample is accepted (Grok Build disarm) instead of Failed	4.9/5

A 5/5 needs those two edges or a documented “won’t fix” on them. The original silent-finish bug is covered.