release-sudowork

Use this skill only for Workcell record structure and file protocol. It does not decide whether work should be delegated or how code should be implemented.

Workcell root is ~/.local/share/workcell unless WORKCELL_ROOT overrides it. Records are outside Git worktrees and are analysis records, not cross-device execution state.

The canonical Markdown templates are in this skill's templates/ directory. Follow them rather than recreating formats from memory.

CLI

The session start hook reports the CLI's absolute path and the record root. Invoke it as python "<reported path>", or as workcell when a shim is on PATH.

workcell task create <slug> creates one task directory and prints its absolute path.

workcell task revise <task-dir> archives the current TASK as history/task-rev-NNN.md, increments revision, resets status to preparing, and preserves the body for main to edit.

workcell run create <task-dir> creates the next runs/run-NNN.md and matching artifacts/run-NNN/, then prints the RUN absolute path.

The CLI does not create questions, choose statuses, integrate commits, push, rebase, squash, or analyze telemetry.

TASK ownership

Main owns TASK. Executor reads it but does not edit it. The current TASK is authoritative. Goal changes substantially create a new task; Scope, Constraints, Success criteria, or Parent context changes with the same Goal use revision.

RUN ownership

One fresh executor owns one RUN. The executor updates it during its work. After that executor finishes, the RUN is immutable. A later executor gets a new RUN.

Artifacts

Large logs, measurements, JSON, images, and other detailed evidence belong under artifacts/run-NNN/. Refer to them from RUN by relative path. Do not store hidden chain-of-thought.

Questions

Questions are one Markdown file, so main creates them directly from templates/QUESTION.md rather than through the CLI. Number them with the next unused q-NNN-<slug>.md in the task's date directory questions/.

Use blocking only when the related work cannot continue without human-only input. Use advisory when main can state a provisional action and continue. Human answers arrive through normal chat; main records the answer and marks the file answered.