Agent-Skills.md

Agent Skills: Praxis-Audit — did THIS session's own methods FIRE, or was it theater?

>-

UncategorizedID: ekson73/multi-agent-os/praxis-audit

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/praxis-audit

Skill Files

Browse the full folder contents for praxis-audit.

Download Skill

Loading file tree…

skills/praxis-audit/SKILL.md

Skill Metadata

Name
praxis-audit
Description
>-

Praxis-Audit — did THIS session's own methods FIRE, or was it theater?

Thin self-referential session-method audit preset. It re-implements nothing — it composes existing primitives and contributes only what the family lacks: a self-referential subject (the session's OWN enacted tool-use — not the standing governance corpus, not the produced result) plus the wiring that carries that audit through research → council → fix.

praxis (πρᾶξις, "the enacted doing") = exactly the subject; named via the anima engine (naming-authority). Full provenance, sibling-distinctions, --auto-self-* mapping, self-validity proofs, external grounding, §DUED and the dogfood log → reference.md ([C07b] inline-vs-spec: this SKILL.md is the lean operational entry). Authored by Claude Opus 4.8 under operator /enhance directive 2026-07-14; reviewed by @ekson73.

Purpose

Audit whether the methods/tools THIS session actually enacted were sound — each FIRED-WELL (real effect that advanced the goal) vs THEATER (ceremony) vs INCONSISTENT vs MISFIRE (nearby-but-wrong) vs GAP (needed-but-absent) — then research better methods, converge a diverse INDEPENDENT council (verifier>generator), and drive confirmed fixes to done. It fills the seam its siblings leave: corpus-firing-audit → the standing corpus (dormant vs firing); end-of-action-self-audit → the produced result; praxis-audit → the enacted praxis (the how).

When to use

  • End of / mid a substantive session — "were our METHODS sound, or ceremony?" — before trusting its conclusions.
  • A recurring smell that a gate/recon/council/verifier was invoked but hollow (rubber-stamp council, a "validation" that validated nothing, a plan never executed).
  • You want not just a report but research + council-converge + fix of the method-level defects.

When not to use

  • Standing GOVERNANCE corpus (rules + MEMORY) → corpus-firing-audit. Produced RESULT quality → end-of-action-self-audit.
  • Recovering what the session is doing (goal, not method) → goal-recovery / ooda-loop. One adversarial attack on ONE claim → red-team.
  • Single-shot edit / read-only Q&A → answer directly. Destructive ops in FIX → always HITL.

The phases — a RECON pre-step (phase 0) + the core RECAP → FIX sequence

RECON    (phase 0, read-only) OBSERVE before assuming (Skopos / CASC Gate-0). Read any prior
         praxis-audit ledger (idempotency). Capability-detect the enacted-tool-use sources
         (ASH journal · transcript tool-call log · .remember/ · CPT topology · commits); degrade-not-block.
RECAP    enumerate the methods/tools THIS session actually ENACTED → ONE deduplicated, IDed
         praxis-register (M1..Mn). No hallucination: a method appears only with evidence it was
         invoked. Kind-tag (recon-probe · mutation · delegation · research · decision-gate · verify · persist).
AUDIT    per method (retarget the corpus-firing-audit lens onto session praxis; kind-aware, evidence-cited):
         verdict := FIRED-WELL | THEATER | INCONSISTENT | MISFIRE | GAP
         For each THEATER/INCONSISTENT/MISFIRE/GAP → propose the CURE, ranked Eisenhower, preferring
         **sharpen an existing fire-point > add a new passive rule** (a passive rule to fix "methods
         don't fire" would itself be theater).
RESEARCH internal + external for better methods per flagged item (enhance-pipeline EXPAND, run as a
         Task sub-skill carrying its own web allowlist; Skopos recon-first; cite, don't fabricate).
COUNCIL  convergence-engine routes the findings: perspective-trio (breadth) + persona-pipeline
         (INDEPENDENT verify, experts != the auditor) + red-team ("is this 'theater' verdict itself
         theater?") → converge (5-act, reject-log). verifier>generator is the master condition — a
         finding survives only if an independent seat confirms it. Kills false-positive "theater" calls.
FIX      (the only mutating phase) gap-loop drives COUNCIL-confirmed fixes to convergence +
         independently validates + persists (worktree-disciplined; --auto-merge default hold, EKO-66).
         Regenerate the ledger. --dry-run stops after COUNCIL (print the fix-plan; drive nothing).

Idempotency contract

  • Probe before acting — read the prior ledger's generated-at + verdict set first (RECON).
  • Read-only through COUNCIL — only write before FIX is the ledger.
  • Convergence — no new evidence + no confirmed findings ⇒ semantically identical ledger.
  • FIX is gated — mutation only in FIX, via gap-loop's worktree + auto-merge discipline; --dry-run skips it.

Bounds

--max-iterations=6          (FIX loop cap → gap-loop; then park-state + escalate)   depth <= 2   Sentinel HIGH auto-blocks
--principles=[DRY, SSOT, KIS (Keep It Simple; not simplistic), YAGNI, ANTI-OVER-ENG, ANTI-THEATER, CONTINUITY, HAND-OFF, BOY-SCOUT]
--principle-exception="only with documented justification (SDP)"   --meta-rule="BEING > rule" (§0 SER>Regras)

Composition (the wiring — DRY; every phase lands on an existing primitive)

| Phase | Composes (reimplements nothing) | |---|---| | RECON | maos:preflight / skills/pulse + Skopos recon (CASC Gate-0) + prior-ledger read | | RECAP | skills/agentic-session-harness (ASH tool-call meta-trace) + transcript / .remember/ / CPT topology → praxis-register | | AUDIT | skills/corpus-firing-audit lens, retargeted onto session praxis (FIRING/THEATER/STALE → FIRED-WELL/THEATER/INCONSISTENT/MISFIRE/GAP; sharpen>add) | | RESEARCH | skills/enhance-pipeline (EXPAND — internal+external; Task sub-skill with its own web allowlist) | | COUNCIL | skills/convergence-engine (regime-router + master condition) · maos:perspective-trio (breadth) · maos:persona-pipeline (independent verify) · skills/red-team (adversarial) · skills/converge (5-act + reject-log) | | FIX | skills/gap-loop (5-phase: MoE resolve + independent VALIDATE + derived score + PERSIST; worktree-disciplined) · maos:postflight P1-SWEEP | | SUBJECT | this skill's own contribution — the session's OWN enacted praxis (self-referential; the how) |

Operator --auto-self-fix/-heal = the FIX phase (gap-loop); --family-aware = by construction (reads/classifies the family's own tools). NOT new machinery — de-theatered mapping in reference.md.

Override parameters

| Flag | Default | Allowed / Notes | |---|---|---| | "<instructions>" (positional) | empty | extra free-text narrowing the audit focus | | --scope | this.session | this.session | session:<id> | branch | ticket:<id> | | --source | auto | auto (capability-detect) | ash | transcript | remember | commits | topology | | --autonomy-threshold | 0.85 | 0.0-1.0 — FIX score gate (bands SSOT agents/COWORK-AUTONOMY-POLICY.md) | | --max-iterations | 6 | int — FIX loop cap → gap-loop | | --auto-merge | hold | authorized | hold | off — default hold (EKO-66: STAGE-only) | | --auto-merge-reason | (none) | non-empty — required when --auto-merge=authorized (auto-merge-standing G8) | | --dry-run | off | RECON→COUNCIL only; print confirmed findings + fix-plan; drive NO FIX | | --output | text | text | json (run envelope below) |

Defaults are fixed; --source=auto and the COUNCIL seat-roster are computed dynamically at runtime (capability-detect + convergence-engine rotation).

Output contract (--output=json)

{"stage":"RECAP|AUDIT|RESEARCH|COUNCIL|FIX","status":"ok|hitl|error",
 "stop_marker":"STOP-DONE|STOP-HITL|STOP-ERROR|CONTINUE",
 "praxis_register":[{"id":"M1","method":"<...>","kind":"<...>","verdict":"FIRED-WELL|THEATER|INCONSISTENT|MISFIRE|GAP","evidence":"<ref>","cure":"<sharpen|add|none>"}],
 "pulse":{"fired_well":0,"theater":0,"inconsistent":0,"misfire":0,"gap":0},"confirmed_by_council":0,"autonomy_score":0.0}

Exit: 0 STOP-DONE · 1 STOP-ERROR · 2 STOP-HITL (authority gap / dry-run parked). Emit exactly ONE <!--ORCH-STATUS: STOP-DONE|STOP-HITL|STOP-ERROR|CONTINUE --> as the last line of each turn.

Protocol Rules (anti-loop invariants)

  • Read-only through COUNCIL; the ONLY mutating phase is FIX (worktree-disciplined; --dry-run skips it). allowed-tools (Task, Read, Write, Edit, Bash, Grep, Glob) matches sibling conductors gap-loop/ooda-loop; Write/Edit/Bash are exercised only in FIX — the restriction is a protocol contract (per-phase frontmatter tool-gating is not a Claude Code primitive).
  • verifier != generator (persona-pipeline / gap-loop VALIDATE) — a verdict survives only if an INDEPENDENT seat confirms it. Never re-loosened. red-team the audit itself: a finding the adversary dissolves is dropped.
  • RECAP/AUDIT read the session's OWN enacted-tool log (ASH/transcript/.remember/CPT) + public composition manifests — NOT peer skills' prompt-bodies or secrets.
  • --max-iterations (6) caps FIX; same-panel re-run FORBIDDEN. HUMAN_DOMAIN + non-negotiable guardrails (secrets/PII, force-push protected, prod/irreversible, cross-org) are HARD gates → escalate, never self-authorize.

Examples

praxis-audit --scope=this.session --dry-run                 # audit this session's methods; print findings + fix-plan; drive nothing
praxis-audit --scope=this.session                           # audit → research → council → fix (FIX STAGE-only by default, EKO-66)
praxis-audit "focus on the verifier != generator gates"     # narrow the focus
praxis-audit --auto-merge=authorized --auto-merge-reason="method-fixes, green CI"   # reason required when authorized

Validation

tests/validate-plugin.sh checks (generically) that skills/praxis-audit/ contains a SKILL.md and that commands/praxis-audit.md begins with ---. It does not parse YAML or assert a name: match — deeper frontmatter/name validation is by review + the skills/skill-writer/SKILL.md 10-item checklist, not the CI script. No separate protocol file (methodology inherited from the composed primitives — corpus-firing-audit's self-contained precedent).

Related

commands/praxis-audit.md · skills/corpus-firing-audit (the retargeted lens) · skills/{goal-recovery,ooda-loop} (the what) · skills/enhance-pipeline (RESEARCH) · skills/convergence-engine + agents/{perspective-trio,persona-pipeline}.md + skills/{red-team,converge} (COUNCIL) · skills/gap-loop (FIX) · skills/agentic-session-harness (RECAP) · agents/COWORK-AUTONOMY-POLICY.md. Extended rationale, proofs, external grounding, §DUED, dogfood log → reference.md.

License

MIT (matches multi-agent-os repo LICENSE).