athenaeum-audit

name	athenaeum-audit
description	Code-aware agent-stack audit + reconcile in one pass. Use when auditing an existing agent topology, reviewing a fleet setup, or triangulating understanding of a codebase's agent architecture across model families. Triggers — "audit the stack", "agent topology review", "fleet audit", "review agent architecture", "triangulate agent state".

Athenaeum — Audit

When to use what

Skill	Use when
athenaeum-design	Building something new; need to settle branches from scratch
athenaeum-reconcile	Agents already disagree; need structured convergence
athenaeum-audit	Reviewing existing architecture; code-aware triangulation in one pass

Pick a mode

Mode	When to use	Attestation
quick	Verbal read-out, same machine, low stakes	No SHA-256; verbal confidence
formal	Cross-model, cross-machine, high stakes	Full sign-off with SHA-256

Default: quick if unspecified.

4-step workflow

1. Init — scripts/athenaeum init <topic> --mode audit

   • Scaffolds .athenaeum/<topic>/ with audit templates.
   • Prompts for auditor identities and scope.

2. Dump — each auditor scans the repo independently.

   • Write <agent>.claims.yaml with 13-branch coverage.
   • Do NOT read peer claims yet. Context leakage defeats triangulation.
   • See templates/reconcile/claims.yaml.template.

3. Triangulate — compute three buckets (same as reconcile):

   • agreed → merge directly
   • disagreed → grill loop (asker = lower confidence)
   • only-one-knows → grill loop (asker = the one who doesn't know)
   • Round budget: 3. Exhaustion → ESCALATE to unresolved.md.
   • On convergence, both peers write RATIFY: <id> | <statement>.

4. Report — merge writer writes audit-report.md.

   • Findings per branch, confidence levels, source citations, action items.
   • Formal mode: LF-normalize, SHA-256, append to signoff.md.
   • See templates/audit/audit-report.md.template.

Code-anchored confidence rules

• high — quoted file path + line range AND a verifying read
• medium — read it, but artifact may be stale
• low — inferred from adjacent files
• No source found → no claim. Do not invent.

A2A interop

This skill speaks A2A natively. An A2A Task represents this workflow.

Task metadata:

• athenaeum_mode: audit
• status lifecycle: submitted → working → input-required → completed

Artifacts produced:

• audit-report.md — text/markdown
• <agent>.claims.yaml — application/yaml
• state.merged.yaml — application/yaml

Enable A2A for this workflow:

athenaeum init my-topic --mode audit --transport a2a

Invoking via A2A:

{
  "jsonrpc": "2.0",
  "method": "tasks/send",
  "params": {
    "task": {
      "id": "athenaeum-audit-my-topic",
      "sessionId": "<agent-session>",
      "status": "submitted",
      "metadata": {"athenaeum_mode": "audit", "topic": "my-topic"}
    }
  }
}

Sibling skills

• athenaeum-design — design something before auditing it
• athenaeum-reconcile — reconcile divergent state before ratifying
• athenaeum-ratify — fleet-wide attestation after audit

Full 13-branch protocol and confidence rules: see REFERENCE.md §1.