| name | execute-plan |
| group | Lifecycle |
| description | Drive an APXM plan to completion without scope creep. Tracks phases with the harness's task tracker, runs focused per-phase verification, refuses to add features beyond the plan, and surfaces blockers immediately. Invoke only after plan produces an approved plan. |
| user-invocable | true |
APXM Execute Plan
Load _shared/apxm-development-rules.md before broad work (it points at
the comment and test rules).
Execute against an approved plan with disciplined progress and no scope
creep. Invoke only after plan and an explicit user approval.
What this skill does
- Restate the plan's phase boundaries at the start. Be explicit
about what each phase produces and where it ends.
- Track phases with the current harness's task tracker.
Mark each
in_progress when starting and completed as soon as
done — do not batch.
- After each phase, run focused verification from the plan's
"Verification per phase" section. Do not run the full workspace
when a per-crate check suffices.
- Refuse scope creep. If something tempting but out-of-scope
appears (a refactor, a comment cleanup, a docs fix), surface it as
a follow-up task and keep going. Do not silently expand the plan.
- Surface blockers immediately. If a verification fails, a
precondition turns out wrong, or the approach hits a dead end:
stop, summarize, and re-invoke
plan rather than working
around the blocker.
- Honor the AIS codegen contract. After any
.td or
TableGen-shim edit, run dekk agents build-dialect then
dekk agents codegen before the next phase begins.
Per-phase rhythm
1. Mark task in_progress.
2. Read the relevant file(s).
3. Make the edit.
4. Run the phase's verification command.
5. If green: mark completed, mark next task in_progress, repeat.
6. If red: stop, summarize, ask for guidance.
Use the smallest correct command:
- Touched a
crates/<x> source? dekk agents test -p <x>.
- Touched the CLI?
dekk agents test-cli.
- Touched Python frontend?
dekk agents test-python-frontend.
- Touched
.td? build-dialect && codegen then the test commands.
- Touched a zoo manifest?
dekk agents vllm zoo-status.
- Touched anything that lint might care about?
Run the relevant focused check, or
dekk agents release check when the
change is release-facing.
Anti-patterns
- Running the full workspace test when one crate changed.
- Marking a task completed when verification failed.
- "Bonus" edits not in the plan ("while I'm in this file I'll also
fix…"). Even good ideas wait for the next plan.
- Continuing past a failed Dekk check without addressing or
acknowledging it.
- Skipping
build-dialect/codegen after .td edits and then being
confused by phantom type errors.
Done condition
All plan phases completed and verified. Then hand off to
simplify.