auto-skill-loop

name	auto-skill-loop
description	Explicit-only. Run a named autonomous mode, repeatedly selecting, implementing, and shipping coherent work while preserving anything blocked.

auto-skill-loop

Act as the orchestrator. Delegate all selected work; do not implement it in the orchestrator context.

Inputs

Require:

• a mode matching references/modes/<mode>.md
• standing authorization to merge work that passes every gate below
• planner, implementer, code_reviewer, and rote_worker dispatch targets
• one or more plan_reviewers targets when the mode always plans; otherwise they are optional until the selected work needs planning
• any additional targets required by the selected mode

Read the selected mode file and references/merge-authorization.md. Validate the mode and all required targets before any side effect. Pass each target unchanged through /agent-dispatch; do not infer or substitute targets.

A mode may add instructions to any stage of the core loop. Every section is optional; an omitted section uses the core behavior unchanged.

Context and isolation

When compacting, retain the caller's loop invocation and this skill text verbatim in addition to what you would normally retain.

Write every fresh-dispatch prompt as a self-contained task. State what the subagent must do, the authoritative files or links it should inspect, and any required output format. Do not include internal role names or a narrative of the parent agent's workflow. For follow-up dispatches, pass artifact and report paths, changed files, and unresolved finding IDs.

Initialize

Dispatch planner once to initialize progress, applying any mode-specific initialization instructions. Its output must record what was examined and what remains eligible.

Loop

1. Select. Prompt a fresh planner with the selected mode, current progress, latest reconciliation result, and plan_reviewers. Require it to select as much eligible work as can be implemented and reviewed coherently together, update progress, and run /plan-work when substantive design is needed. Include any mode-specific Execute instructions in the plan. Terminate only when it proves a complete pass is exhausted under any mode-specific criteria. If planning is required, do not proceed without a non-empty plan_reviewers list and an implementation-ready result.

2. Execute. Prepare or reuse a branch for the selected work. Dispatch implementer with any mode-specific Execute instructions and an implementation-ready plan, or the selected work directly when no plan was needed. Then dispatch code_reviewer with /review-uncommitted-code, return material findings to implementer, and rerun affected checks.

3. Prepare the PR. Dispatch rote_worker to run /ship-pr. Continue only when it returns a merge-authorization request for an exact PR and head.

4. Authorize. Follow references/merge-authorization.md for the exact PR and head. On changes-required, dispatch implementer fresh with the findings, rerun affected checks, and return to PR preparation. Preserve a blocked PR and continue independent work. Continue to merge only on authorize.

5. Merge. Resume the same rote_worker with authorization for that exact PR and head, derived from the standing loop authorization. Continue to reconciliation after the shipping dispatch returns.

6. Reconcile. Reconcile the actual merged, open, or preserved PR with its source, applying the mode's Reconcile section when present. Then return to selection. With stop_after=one-pr, stop after this reconciliation.

Termination

Continue whenever evidence supports a safe, in-scope next step. Preserve blocked work and keep selecting independent work. Ask the user only after a complete pass finds no independent work and further progress requires authority only the user can provide or a consequential choice the available evidence cannot resolve.

Report why the loop ended, the smallest remaining user questions, and every preserved branch or PR.