Slice Plan Runner

Slice Plan Runner

Purpose

Run a saved slice plan unattended, one slice at a time, committing each completed slice and stopping on ambiguity, contract failure, failed validation, unsafe scope, or user-judgment needs. Use this skill only when the user explicitly invokes $slice-plan-runner.

This is an execution loop, not planning. Do not invent a plan, broaden scope, infer missing contract fields, or replace acceptance criteria. Status updates are limited to the runtime contract's definition.

Inputs

• Prefer an explicit plan path, slice limit, and run constraints.
• If no plan path is provided, find the active plan from the current conversation, repo plan index, or repo docs. Ask for the path if more than one plan is plausible.
• Use Runner completion scope as the plan-authored maximum. The effective run limit is the strictest limit from the user request, Runner completion scope, Slice/run limits, approvals, runtime contract, local safety policy, and stop conditions. If no explicit limit exists, run all remaining slices only when the user clearly asked for a full unattended run and Runner completion scope allows it; otherwise stop and ask for a limit.
• Treat constraints such as max 3 slices, no network, stop before live data, or approval requirements as hard limits.

Loop Rules

• Read installed _shared/slice-runtime-contract.md before selecting the first slice. If it cannot be read, stop.

• Read installed _shared/slice-execution-contract.md or _shared/slice-planning-contract.md only when starting a fresh plan, validation warns or fails, the plan changed beyond allowed status/Handoff-note updates, or the selected slice touches plan/schema behavior.

• Run strict plan validation, without --allow-draft, before the first slice and after status updates when the plan path ends in .md:

  python path/to/to-slice-plan/scripts/validate_slice_plan.py <plan-path>
  

  Non-Markdown plans are outside the contract unless the user explicitly approved them. Validator warnings do not block by themselves; inspect them for drift and report them.

• Honor the Execution Handoff, user constraints, and runtime contract before each slice.

• Run validation commands sequentially when they may share temp directories, caches, ports, databases, global CLI state, generated outputs, environment variables, or live service state. Parallelize only clearly independent checks.

• Execute slices sequentially in plan order unless the plan explicitly permits a different order.

• Re-check current state and record a fresh dirty baseline before every slice.

• For each selected slice, run the shared Single-Slice Execution Cycle.

• Use continuation mode after a clean committed slice only when the Execution Handoff allows it and every runtime-contract continuation condition holds. Continuation shortens rereading; it never skips strict validation, dirty checks, approvals, Review gate, or commit hygiene.

• Stop after review findings, failed validation, missing Review gate skills, or status update conflicts unless the fix is clearly local to the current or just-completed slice.

• Do not batch slices, skip malformed slices, or continue after a blocker that needs user judgment.

Continuation Check

After each committed slice:

1. Re-read the updated plan, next slice, and Execution Handoff.
2. Continue only when validation passed, validator warnings were inspected and do not indicate drift, the commit succeeded, remaining slices are contract-complete, dirty-work boundaries are intact, Continuation mode and Runner completion scope allow the next pass, and no stricter limit or approval boundary is reached.
3. Stop before network, live-data, credentialed, destructive, or remote-write work without explicit approval.
4. Stop and recommend prefactoring or replanning when repeated awkwardness across adjacent slices shows the plan is fighting the code.

Completion criterion: continuing is justified by the plan, runtime contract, handoff, and current repo state.

Completion Report

End with the shared Executor Completion Report, preceded by the plan path, requested or inferred slice limit, completed slices, encountered skipped slices, and blocked slices. Include any slice the run stopped before starting.

If the run stops early, make the stopping reason the headline. Do not describe a partial run as complete.