IMPLEMENTAUDIT
On this page
  1. Plain version
  2. Formal vocabulary
  3. Why it matters
  4. Method source surfaces
  5. Plain-first method table
  6. Lean terms as behavior
  7. Method boundary
  8. Default sequence
  9. Authorization defaults
  10. Failure defaults
  11. Scope-creep register
  12. Ledger priority taxonomy
  13. Polish and Harden
  14. Plugin cache warning
  15. Default source surfaces
  16. Compared with a normal agent run
Core model

Operating method

The ordinary meaning behind the Lean vocabulary.

The method exists to make work inspectable.

Plan the work, inspect real files, check before claiming, stop when evidence breaks, reflect on the cause, and close only with proof.
The Lean labels are secondary to those visible behaviors.

Plain version

Plan the work, inspect real files, check before claiming, stop on failure, inspect cause, reflect, and close only with evidence.

Formal vocabulary

PDCA, Gemba, Smoke Before Claim, Andon, Hansei, 5 Whys, and Plan Closure remain as formal labels after the plain meaning.

Why it matters

The method is there to make agent work inspectable, not to make readers learn manufacturing language before they can try it.

Method source surfaces

skills/implementaudit/references/lean-operating-discipline.md and README.md operating method.

Plain-first method table

Plain behaviorFormal label
Inspect real filesGemba
Check before claimingSmoke Before Claim
Stop on abnormal evidenceAndon / Jidoka
Reflect after failureHansei
Find the real blocker5 Whys
Fold durable fixes into standardKaizen

Lean terms as behavior

The method terms are useful only when they map to visible proof surfaces.

  • 5S_CHECK means the run checks cleanliness and artifact placement before claiming closure.
  • Muda/Mura/Muri means waste, unevenness, and overload are treated as inspectable risk, not slogans.
  • Hoshin/Nemawashi means alignment, owner/source, and affected surfaces are named before dispatch.
  • Poka-yoke means validators and templates catch repeatable mistakes before a human reviewer has to.

Method boundary

The method is not a badge.
A term belongs on the page only when it maps to a visible behavior, artifact, or gate.

Default sequence

When the user gives a repo request, IMPLEMENTAUDIT defaults to evidence, bounded scope, and honest closure.

  1. Validate the input or synthesize a bounded work record for governed casual-build intake.
  2. Inspect the repo and applicable instructions before promising edits.
  3. Route existing-repo repair through the repair route (DMAIC) or new-build/replacement work through the design route (DMADV).
  4. Write a run folder for phased work, or bind the compact work record when a separate run folder is not needed.
  5. Run before/after checks, record failures, and stop on abnormal evidence.
  6. Close with final audit or handoff.

Authorization defaults

An implementation run does not imply:

  • commit
  • push
  • tag
  • release
  • publication
  • provenance
  • host install
  • sidecar setup
  • plugin-state mutation

It also does not authorize high-risk maintenance actions:

  • delete data
  • alter credentials or secrets
  • rewrite history
  • commit raw diagnostic outputs
  • hand-edit generated artifacts when a source generator exists
  • claim proof without evidence

If the audit contradicts AGENTS.md, classify the conflict as OWNER DECISION instead of choosing by agent judgment.

Each action needs separate authorization and evidence.

Failure defaults

When evidence fails, the run:

  • stops
  • classifies the abnormality
  • inspects cause
  • proposes a countermeasure
  • reruns the relevant evidence
  • escalates only with new evidence

If no bounded countermeasure remains, it hands off honestly.

Scope-creep register

Adjacent work is not a free add-on.
If the run finds nearby defects, it records them without silently expanding the task.

  • Log the adjacent issue with owner/source and evidence.
  • Promote at most one level only when it is required for safe closure.
  • Do not chain promotions into a new project.
  • Ask for an owner decision or hand off when scope, risk, or authority is unclear.

Ledger priority taxonomy

Default closure order is P0 -> P1 -> P2.
A finding may also become OWNER DECISION, DEFERRED, or OUT OF SCOPE when the evidence says it should not be patched inside the current run.

ClassMeaningDefault action
P0Blocks safe completion, evidence truth, or user-requested acceptance.Fix or hand off before claiming completion.
P1Material correctness, usability, docs, or validation gap.Fix after P0 items unless risk requires split.
P2Lower-risk polish, maintainability, or clarity issue.Fix when in scope and cheap, or defer explicitly.
OWNER DECISIONRequires authorization, policy choice, scope change, or business judgment.Stop and ask or hand off with the smallest decision needed.
DEFERREDValid but intentionally outside the current closure window.Record reason, owner/source, and follow-up surface.
OUT OF SCOPENot part of the supplied audit object or authorized request.Do not patch; explain boundary if relevant.

Polish and Harden

A terminal Polish and Harden phase can be used when the main work is closed but operational quality still needs inspection.
It covers cleanliness, identity hygiene, generated-artifact freshness, proof-boundary wording, and anti-repeat standardization.
It is not permission to add new feature work.

At a release gate, identity hygiene runs public-claim boundary checks with terms supplied at runtime, for example through scripts/check-public-claim-boundaries.sh.
The ledger records PASS/FAIL; disallowed public-claim terms must not be written into tracked files.

Plugin cache warning

Claude plugin cache refresh depends on a version-field change.
Shipped code changes need a plugin version bump before claiming update behavior; this warning does not prove release, install, or provenance state.

Default source surfaces

skills/implementaudit/SKILL.md safety defaults, skills/implementaudit/templates/PROTOCOL.md phase loop, skills/implementaudit/references/routing.md, skills/implementaudit/references/phase-design.md, and README.md priority/closure descriptions.

Compared with a normal agent run

A normal agent run can compress plan, edit, retry, and done into a chat narrative.
IMPLEMENTAUDIT splits those states into files, markers, before/after evidence, failure rows, and final audit closure.