batch-work

Bucket: workflow · Invocation: manual · /batch-work (direct Claude), /agent-os:batch-work (Claude plugin), or $batch-work (Codex)

Plans and runs several decision-ready units through isolated workers and aggregate verification. Each worker is simply deliver-work starting in an isolated workspace.

Batch-work consumes an existing set of implementation-ready, dependency-mapped issues. It does not discover or decompose the product shape, and several issues do not invoke it automatically. The developer explicitly chooses when the issue graph should run as an integrated batch.

The request decides whether the workflow plans, executes, or does both. A planning-only request creates the manifest and stops. An execution request continues without a second approval checkpoint.

Each task has a stable key, outcome, scope, dependencies, checks, and definition hash. Independent frontier tasks run in isolated branches and worktrees. Workers return concise results; the coordinator owns the manifest and integration branch.

After each integration the coordinator reruns relevant checks. After all tasks integrate, fresh aggregate checks decide whether the complete candidate is ready. Worker-local success is never a substitute for integrated behavior.

An executed batch is material. The coordinator applies the deliver-work review gate to the integrated candidate before delivery, resolves supported findings, and reruns affected aggregate checks.

Definition drift invalidates an affected result. Merge, deploy, destructive cleanup, and external effects follow the request or project policy.

See Batch manifests.