v0.3.0

Highlights

Added

• Pooled standalone-task pods — runs within a workflow now share pods, scaling out to workflows.maxPodInstances replicas each hosting up to workflows.maxAgentsPerPod concurrent runs (mirrors repo pod scaling). Runs track assigned pods via workflow_runs.pod_id with last_pod_id for retry affinity, and pool selection follows preferred → least-loaded → scale-up → overflow. Fixes a leak where a burst of triggers would spawn one pod per run even though only a few ran at once.

Changed

• Reconciliation control plane is now authoritative — the K8s-style reconciler (shadow mode in 0.2.0) now owns PR-driven transitions, auto-merge, complete-on-merge, fail-on-close, auto-resume, review launch, stall detection, pod-death detection, and control intent (cancel/retry/resume/restart) for both Repo Tasks and Standalone Tasks.
• Shared auth banner, state badge, and metadata card across task pages for a consistent UX.

Fixed

• Reconciler: clear stale finishedAt when retrying a standalone run.
• Reconciler: use unique jobIds for executor enqueues to prevent BullMQ dedup collisions.
• Agent adapters: include cache_read and cache_creation tokens in input totals (#457).
• API: trigger auth banner when the usage endpoint detects an expired OAuth token (#455).
• API: detect Claude auth failures mid-run in standalone task runs and override nominally-successful exit codes.

Docs

• Document the unified reconciler and the Repo vs Standalone Task model.

Full changelog: v0.2.0...v0.3.0