feat(cli): post partial results when the process is signalled

[MattJColes]

What

On SIGINT/SIGTERM, lgtmaybe stops dispatching new model calls and posts whatever the review has already produced, with a notice naming the interruption — instead of dying with nothing on the PR.

Why

Two routine ways a review currently posts nothing at all — no findings, no failure comment:

1. The job exceeds its timeout-minutes. max_review_seconds defaults to 3600s, so a workflow with timeout-minutes: 25 gets killed long before the soft deadline can fire and post partial findings. The workaround has been hand-tuning max_review_seconds under the job timeout — fragile config discipline every user has to rediscover (our own .github/workflows/lgtmaybe.yml carries a comment explaining the layering).
2. concurrency: cancel-in-progress: true cancels an in-flight review when a new push lands.

In both cases the runner signals the process before killing it. We ignored the signal.

How

Reuses the deadline mechanism rather than building a parallel one. The engine already degrades an over-running review to partial-results-with-a-notice when max_review_seconds passes. A signal now sets the same state:

• engine.request_interrupt() sets a process-global flag that _skip_reason checks alongside the deadline and the token budget, so queued lens calls are skipped at execution time exactly as they are past the deadline. In-flight calls finish and their findings post.
• Everything downstream is shared and unchanged: the "N of M review calls failed" notice, the hidden incomplete marker (so the notice also posts as its own PR comment), the skipped-reflection notice, and the ReviewIncompleteError path when every call was skipped — an interruption can never become a silent 👍 LGTM.
• Only the reason wording differs (review interrupted (termination signal) vs the deadline's), so nobody goes hunting for a max_review_seconds they never hit.

Handler installation (cli.graceful_interrupt), against the constraints:

• Installed by the CLI entrypoint only, held open for the subcommand via the click context — importing lgtmaybe as a library never touches a host application's handlers (guarded by a fresh-interpreter test).
• signal.signal() only works on the main thread; off it (and on a platform missing a signal) the context manager degrades to a no-op rather than raising — the engine runs a ThreadPoolExecutor, so this matters.
• The previous handler is restored on the way out and as the handler fires, so a second signal still kills the process if the wind-down itself hangs. lgtmaybe never becomes unkillable, and it never permanently steals the signal.
• Windows-safe: signals are looked up with getattr and signal.signal failures are tolerated, and the tests parametrise over whichever signals the platform actually has.

Tests

TDD, red first (both files failed to import before the implementation existed).

tests/engine/test_interrupt.py

• test_flag_starts_clear_and_round_trips
• test_queued_calls_are_skipped_and_partial_findings_still_post
• test_notice_names_the_interruption_not_the_deadline
• test_reflection_is_skipped_with_an_honest_notice
• test_every_call_skipped_still_fails_loud

tests/cli/test_signal.py

• test_installs_and_restores_the_previous_handler (per available signal)
• test_first_signal_requests_the_wind_down (per available signal — also asserts the previous handler is already back)
• test_off_the_main_thread_is_a_no_op
• test_main_wraps_the_command_in_the_wind_down
• test_importing_the_library_installs_no_handler

Also

• Spec: new cli.graceful-interrupt requirement + anchor in openspec/specs/cli-and-local/ (the review-pipeline section it would otherwise extend is already at the 40-line cap). uv run pytest tests/specs -q and openspec validate --specs green.
• Docs: action.yml's max_review_seconds description, the Action input table, docs/explanation/architecture.md, ARCHITECTURE.md, CLAUDE.md, regenerated docs/llms*.txt.
• .github/workflows/lgtmaybe.yml: the timeout-layering comment now names the wind-down as the backstop. Timeout layering is still the calm path — the signal handler can only skip queued calls, so a job cut short mid-call is still racing the runner's grace period. No workflow if: conditions touched.

Gate: ruff check / ruff format --check / mypy / pytest -q (1839 passed, 3 skipped) all green.

---

🤖 Generated with Claude Code

---

Generated by Claude Code

[lgtmaybe]

⚠️ 3 of 4 review calls failed (ProviderTruncated: response hit the 32768-token max_tokens ceiling (34012 reasoning) before finishing — raise max_tokens, or lower max_input_tokens so each call covers less); results may be incomplete.

⏱️ 1 batch was too big for one call (timed out, or ran past the max_tokens ceiling) and was reviewed in smaller pieces instead. Consider a lower max_input_tokens, a higher max_tokens, or a faster model.

2 findings · provider openrouter · model ~deepseek/deepseek-v4-flash-latest · lgtmaybe 1.12.2

[lgtmaybe]

⚠️ 2 of 4 review calls failed (ProviderTruncated: response hit the 32768-token max_tokens ceiling (33849 reasoning) before finishing — raise max_tokens, or lower max_input_tokens so each call covers less); results may be incomplete.

⏱️ 1 batch was too big for one call (timed out, or ran past the max_tokens ceiling) and was reviewed in smaller pieces instead. Consider a lower max_input_tokens, a higher max_tokens, or a faster model.

💬 2 earlier lgtmaybe conversations are still unresolved on this PR — this run's count covers what it reviewed now, not those.

0 findings · provider openrouter · model ~deepseek/deepseek-v4-flash-latest · lgtmaybe 1.12.2

Incremental review of the changes since 8403343 — earlier findings stay open until fixed.

[lgtmaybe]

⚠️ 2 of 4 review calls failed (ProviderTruncated: response hit the 32768-token max_tokens ceiling (33849 reasoning) before finishing — raise max_tokens, or lower max_input_tokens so each call covers less); results may be incomplete.

⏱️ 1 batch was too big for one call (timed out, or ran past the max_tokens ceiling) and was reviewed in smaller pieces instead. Consider a lower max_input_tokens, a higher max_tokens, or a faster model.

💬 2 earlier lgtmaybe conversations are still unresolved on this PR — this run's count covers what it reviewed now, not those.

0 findings · provider openrouter · model ~deepseek/deepseek-v4-flash-latest · lgtmaybe 1.12.2

Incremental review of the changes since 8403343 — earlier findings stay open until fixed.