mattrobinsonsre/terrapod v0.30.8

on GitHub

A bundle of correctness fixes around AI-summary rendering, log streaming through the BFF, the PR-comment dispatcher, and — most importantly — the run lifecycle when a user cancels an in-flight apply.

Highlights

• Cancel-while-applying now honours reality — cancel_run on an applying run no longer transitions straight to canceled. A new canceling intermediate state lets the reconciler pick the terminal status from observable Job outcome: if a state-version was uploaded, the run is applied (apply landed; bookkeeping must follow reality); if the Job was killed cleanly with no state, the run is canceled and the workspace's state_diverged flag is set so the operator sees the drift-warning signal. Artifact uploads are deliberately never rejected by tracking status — refusing a state upload from a Job that mutated infra would be strictly worse than the zombie symptom.
• Plan output no longer truncates at the tail — upload_plan_log / upload_apply_log now publish a log_updated SSE event after writing storage, so the UI's log-fetcher comes back for the bytes the runner's EXIT trap landed (typically the PLAN_HAS_CHANGES=... line and post-plan OPA output). The PR #423 fix in v0.30.7 was the necessary half — kept the poll open by holding off ETX on the Redis path — but the UI only re-fetches on event, and there was no event when the final log landed.
• No more duplicate PR comments on fast status flips — _find_or_create_comment now serialises across replicas via a Redis SETNX mutex keyed on (workspace_id, pr_number). Without it, a queued → planning flip within the same scheduler tick produces two dedup-distinct triggers; both dispatchers found an empty cache + empty marker search and POSTed a brand-new comment, leaving a duplicate alongside the one being edited.
• AI plan summary risk_factors no longer drop out — the schema enforces presence, but neither schema nor constrained-decoding can express "non-empty when risk_level isn't low". The model occasionally returned risk_level: "medium" with risk_factors: [], leaving the operator with a yellow pill and no enumerated reasons. The prompt rule is now a hard rule (not a schema-comment hint), and a parse-time observability log fires when the violation slips through anyway.

Status

Beta — production-deployable; the cancel-zombie change touches the run state machine, so re-test cancel paths against your live deployment before promoting.

Full Changelog: v0.30.7...v0.30.8