Skip to content

0.113.0

Choose a tag to compare

@github-actions github-actions released this 12 Oct 19:17
· 674 commits to main since this release
chore(release): bump to v0.113.0

- fix(epics): stop double-hydrate clobbering sessions on inbound auto-approve
- fix(plugin): sync session-manager-dev marketplace version to 0.2.1
- docs(validation): errors-logs-hardening plan review record
- feat(telemetry): app-wide first-run telemetry notice banner
- feat(telemetry): log flush failures and breaker trip to app log
- feat(watchdog): alert once when relaunch capped, throttle alive lines, prune old logs
- feat(atomicFs): sweep stale tmp files of dead writers at boot
- feat(scheduler): run-log retention on by default (14d, keep newest 3 per slug)
- fix(diagnostics): raise main loop stall warn threshold to 1000 ms
- perf(scheduler): throttle retired flat PRD sweep to once per hour
- fix(webremote): treat invalid_token ticket failure as terminal, expose repair-needed
- fix(scheduler): carry verifier verdict and reason on needs_review transitions
- fix(config): validate config:watch per path and drop sst-* tmp cwds
- fix(supervisor): pin haiku probe, size budget, record budget/error exits distinctly
- fix(credentials): make CLI fallback the sole refresh avenue, drop private OAuth grant
- fix(scheduler): share dep-satisfied predicate between dispatch and post-run verify
- docs(validation): record feedback-autostart plan validation
- feat(feedback): wire inbound auto-approve at boot, workspace hydrate, and approval bar
- docs(domain-model): cross-project feedback auto-starts only on verified human request
- feat(feedback): expose autoStart/humanRequestQuote via feedback_open_session MCP, catalog, and skill
- feat(feedback): add renderer inbound auto-approve primitive and preload channel
- feat(feedback): verify human quote to auto-start cross-project sessions
- feat(chat): thread prompt origin from send() through dispatchSend to chat.run
- feat(chat): record chat:run prompts to human ledger only when origin is human
- feat(feedback): add in-memory human prompt ledger
- docs(validation): validate false rate-limit pause plan (1857-1860)
- docs(scheduler): clarify utilizationThreshold vs rate_limit pause
- fix(scheduler): only use billing reset for rate-limit pause when at/above threshold
- fix(scheduler): take rate-limit reset from the rejected event's window
- fix(scheduler): detect rate limit only on genuine rejection events
- docs(manual): documentsAppVersion 0.112.0 (no content change)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>