Auto Skill Loop

auto-skill-loop

Inputs

Required:

• a mode matching references/modes/<mode>.md
• merge_authorization: standing authorization to merge under the gate below
• operator, planner, one or more plan_reviewers, implementer, code_reviewer, pr_worker, and rote_worker dispatch targets

Optional:

• loop_context: additional context for the orchestrator only
• planner_context: additional context included only in step 1
• ship_pr_context: additional context included only in step 2
• operator_context: additional context included only in operator dispatches

Every input must be explicitly named in the skill invocation. Do not infer an unnamed input from unstructured text or from another input.

Rules

• Use /dispatch-agent for every dispatch.
• Act as the orchestrator. Delegate all work.
• Build each dispatch prompt only from its specified prompt template.
• When compacting, retain the original user inputs and this skill verbatim in addition to what you would normally retain.

Acting on the User's Behalf

This loop must run without human intervention. Each iteration is intended to result in a merged PR. For each loop that requires human input, dispatch operator in a fresh session. Use dispatch_continue for multiple invocations within a single loop. The first prompt should include the complete contents of references/human-guidance.md, followed by operator_context if provided, then the item requiring human input with all provided details verbatim.

If the operator determines that real human intervention is required, save the work to an appropriate remote branch for future handling, then check out the primary branch. Continue with another loop iteration. Do not block the loop waiting for human input.

Loop

1. Dispatch planner with skill implement. Use the following as its prompt:

<complete contents of references/modes/<mode>.md>
<planner_context, if provided>

implementer: <implementer>
plan_reviewers: <plan_reviewers>
code_reviewer: <code_reviewer>

Return a self-contained `<implementation_input>` that states the actual task,
request, or spec you implemented and includes paths to plan artifacts if used.
This context will preserve the intended scope during later review.

If planner is unable to find work to complete, repeat the dispatch one more time. If two invocations in a row cannot find work, exit and inform the user.

2. Dispatch rote_worker with skill ship-pr and this prompt:

pr_worker: <pr_worker>
<ship_pr_context, if provided>

Use the following context to preserve intended scope and behavior throughout the PR
workflow:

<implementation_input>

Continue when it returns a merge-authorization request.

3. Dispatch operator in a new session. Use the complete contents of references/merge-authorization.md as its prompt, then append:

<operator_context, if provided>
request: <rote_worker merge-authorization request>

4. If authorized, continue the dispatch session from step 2 with the exact authorization. If not, continue the dispatch to preserve the PR and return to the primary branch without merging. Then return to step 1.