devintern
← All releases

v2.13.0

September 24, 2026

Actioned-ticket release: the worker now records every ticket it opens a PR for and keeps it out of the sweep until the ticket genuinely changes, so label-based trackers stop re-implementing the same open issue in a loop, with per-repo GitHub credentials for fleet events and steadier PR watching when GitHub returns 404.

Added

  • Actioned markers stop re-implemented tickets (DEV-125): once a pull request is created, the worker records the ticket as actioned in .devintern-code/queue.db and skips it on later sweeps even when it still matches [defaults].task_query — the loop a label-based tracker (GitHub, GitLab) would otherwise fall into. This works for every polled tracker. The marker captures the summary, description, status/state, and labels and re-arms when any of them changes (edit, re-open, or add/remove a label); it is captured after the worker’s own comment and status transition so those never re-trigger a run. If the project configures prStatus in .devintern-code/settings.json the worker also moves the ticket to that status/label first, and a missing label, missing permission, or transient API error never loses track of the ticket or fails the PR run — the marker is recorded anyway with a warning suggesting a prStatus. Skips are logged as ⏭️ skipping KEY (already actioned; no change since the PR). The marker gates polling, dashboard retries, and relayed changes, and uses a pinned DEVINTERN_ACTIONED_SOURCE key so the subprocess recorder and the workspace gate can never derive different keys from a stale .env TASK_TRACKER or a leaked team env

Fixed

  • Actioned marker is race- and failure-hardened: a relayed task.changed caused by the worker’s own post-PR label/comment is no longer re-implemented, an unreadable tracker record is persisted as an unverified marker that keeps suppressing and refreshes to verified on the first successful read, the gate caches the last-verified ticket updated stamp (persisted so it survives a restart) to avoid re-reading the tracker per ticket on every poll, and a state-DB or disk-I/O failure during the transition warns instead of being misreported as a PR-creation failure
  • Open PRs stay watched when GitHub returns 404: the review poller, CI watcher, and PR reconciler now treat a GitHub 404 as “still open, access warning” rather than “gone”, because GitHub returns 404 for an inaccessible PR that is still open; only a definite closed/merged state drops a PR from the watch list
  • Fleet events use per-repo GitHub credentials: GitHub review, CI, mention, and relay requests now resolve the matching repo’s GITHUB_TOKEN layers (with a fixed team’s env_file/[teams.env] able to override), falling back to the workspace token only when no repo or team override exists