Next Slice

Next Slice

Purpose

Execute exactly one slice from a saved plan, validate it, commit it when completed, report any blocker, and stop. Use this skill only when the user explicitly invokes $next-slice.

This is execution, not planning. Do not rewrite the plan, broaden scope, infer missing hard fields, or continue into later slices unless the user asks for replanning or a separate run. Status updates are limited to the runtime contract's definition.

Runner completion scope is ignored by $next-slice; this skill executes exactly one selected pending slice.

Within the selected slice, treat the plan as a boundary, not a script. Adapt implementation details when current code demands it, as long as the Objective, Acceptance criteria, safety and approvals, Validation, Review gate, and Commit boundary remain intact.

Inputs

• Prefer an explicit plan path and slice name or number.
• If no plan path is provided, use the active plan only when it is unambiguous from the current conversation, repo plan index, or repo docs.
• Ask for the plan path when more than one plan is plausible.

Workflow

1. Load the plan and runtime contract.

   • Read installed _shared/slice-runtime-contract.md, the selected plan, its Execution Handoff, local AGENTS.md, and slice-named specs, docs, tests, CLIs, or artifacts.
   • Read installed _shared/slice-execution-contract.md or _shared/slice-planning-contract.md only when this is the first slice for the plan in this conversation, strict validation warns or fails, the plan changed beyond status/Handoff note, or the selected slice touches plan/schema behavior.
   • If the runtime contract cannot be read, stop.
   • Completion criterion: the target slice, execution limits, branch state, and dirty baseline are known.

2. Run strict plan validation.

   • Run strict plan validation, without --allow-draft, 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 as Validation warnings: None or list the warnings inspected.

   • Apply the runtime contract's Selection Rules and Stop Conditions.

   • Stop if the selected slice is missing hard fields, is malformed, or asks the executor to infer intent.

   • Stop if the selected slice has non-None Fog of war; normal implementation uncertainty belongs in Risks/review traps, not Fog of war.

   • Completion criterion: the selected slice is contract-complete before editing.

3. Run the shared Single-Slice Execution Cycle once from the validated selection.

   • Do not start later slices.
   • Use continuation mode only when the Execution Handoff allows it and every runtime-contract continuation condition holds. It may shorten rereading for repeated $next-slice invocations in the same plan/session, but never changes the one-slice execution limit or skips strict validation, dirty checks, approvals, Review gate, or commit hygiene.
   • If repeated awkwardness across adjacent slices suggests the plan is fighting the code, stop after the selected slice or blocker and recommend prefactoring or replanning instead of broadening silently. Examples: needing to change acceptance criteria, move work between slices, add a new architecture decision, or touch unrelated modules not named by the plan.
   • If the Review gate is $review and no fixed point is named or unambiguously derivable from the plan, repo docs, or user request, stop and ask. Do not invent one.
   • Completion criterion: the selected slice completed and committed, or the blocker was reported by the shared cycle.

Completion Report

End with the shared Executor Completion Report, preceded by the plan path and slice selected, completed, or blocked.