Skip to content

workflows.md

Latest commit

 

History

History
96 lines (75 loc) · 4.05 KB

workflows.md

File metadata and controls

96 lines (75 loc) · 4.05 KB

Workflows

End-to-end recipes. Each is a short scenario and the ordered skills to run. They all assume the plugin is installed (see Quick Start).

1. Turn a prose spec into a running loop

You have a goal in prose and want it executable.

  1. loopify-spec — compile the spec into a Loop Contract, and ask it to bootstrap a folder: 
    Use loopify-spec on docs/goals/goal.md and bootstrap it into .loopify/loops/001-feature/.
    (loopify-spec hands off to loopify-bootstrap when you say "bootstrap it"; without that phrase it writes only the contract.)
  2. loopify-run — execute until a stop reason: 
    Use loopify-run on .loopify/loops/001-feature/.

You now have a folder with a contract, a trace, and a final report.

2. Run a loop to a verified success

The checks pass — but in Loopify the maker can't call its own work done.

  1. loopify-run drives the loop until the gate is green and manual items are queued, then requests verification (it will not write verdict.md itself).
  2. loopify-review in verifier mode — run it as an independent party (ideally a different model or a fresh read-only context), pointed at the loop folder. It re-runs the checks, grades every evidence-map row, and writes verdict.md ending in approve / reject / cannot-verify.
  3. Validate the claim mechanically: 
    node plugins/loopify/skills/loopify-run/scripts/check-stop-reason.mjs .loopify/loops/001-feature success
    Exit 0 means success is legal: an approving verdict exists, its verifier differs from the maker, and the gate re-runs green.

If the verdict is reject, treat the findings as observed failures, fix, and re-verify — that's the loop, not a setback. See Mandatory independent verification.

3. Rescue a stuck loop

The same failure keeps coming back, or the loop is burning turns.

  1. loopify-debug — classify the stuck pattern and get a recovery plan or a blocked/escalated report: 
    Use loopify-debug on .loopify/loops/001-feature/trace.md and tell me why it isn't converging.
  2. Apply the recovery: narrow the scope, strengthen or isolate the failing check, ask for the human decision, or mark the loop blocked/escalated. The debug skill insists the next diagnostic differs from the failed attempts — no re-running the same patch and hoping.

4. Strengthen weak checks

A requirement is real but the contract proves it with a vibe.

  1. loopify-checks — find the weak/fake/missing evidence and propose checks by class (automated, visual, manual): 
    Use loopify-checks to strengthen .loopify/loops/001-feature/loop-contract.md for the "fast" and "polished" requirements.
  2. It updates the Requirement Evidence Map — turning automated requirements into real commands and naming the visual/manual ones honestly instead of faking them.

5. Add governance to a risky loop

The loop can touch production, money, data, dependencies, or its own harness.

  1. loopify-governance — classify the risk and add budgets, allowed/denied paths, approvals, rollback, and a harness-change contract: 
    Use loopify-governance to add budgets, denied paths, and a rollback plan to .loopify/loops/002-deploy/loop-contract.md.
  2. High-risk actions now require explicit, recorded approval, and any mutation of the loop's own checks/prompts/tools is reviewed as a change to future behavior.

6. Learn from a finished loop

The loop is done; bank what it taught you.

  1. loopify-retro — read the contract, trace, diff, and final report, and extract reusable lessons (including the useful rejected attempts): 
    Use loopify-retro on .loopify/loops/001-feature/ and write its retro.md.
  2. Feed the recommendations back into your next Loop Contract — the missing check it found becomes a check the next loop starts with. That's how one loop's failures become better future loops.