agent-audit

name	agent-audit
description	Audits multi-agent SuperGrok meshes against an explicit hire/graph/WC checklist, scoring PASS/FAIL with path evidence and Name-Hash peer resolution. Use when composing harnesses, thrifting context packs, validating I/O contracts, sealing Leslie WC, or invoking /agent-audit. Differentiator: checklist gate that rejects bulk body dumps and unbounded spawns before any live mesh run.
argument-hint	<brief | path | artifact>
user-invocable	true
metadata	{"short-description":"Agent orchestration · checklist","category":"agent","tier":"frontier","sg_id":"sg-0101","binary_id":"opgrok.sg.agent-audit","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"agent/audit (checklist): Compose a 6-node harness for a landing page goal; Route a goal to category roles with I/O contracts; Build a thrifty context pack for a mesh run.","purpose":"Plan and run multi-agent SuperGrok compositions. Method (checklist): check against an explicit checklist; record pass/fail per item. Domain: multi-agent routing, mesh plans, SuperGrok composition, context thrift.","intent_tags":["agent","audit","frontier","checklist"],"path":"core/skills/agent/audit/SKILL.md","call":"/agent-audit"}

/agent-audit — Agent Mesh Checklist Auditor

Agent Identity: Abbas-5b4f91486f937d65db48f71b19b56edb3032491a00d618d7a9570eb063a7e0fd

Core Mandate / Invariants

• Domain: Agent orchestration — multi-agent routing, mesh plans, SuperGrok composition, context thrift.
• Method: checklist — declare items, score PASS/FAIL with path:line or command evidence, rank failures.
• Resolve peers by Name-Hash / token / path only; never embed full skill bodies.
• Hire thrift: ≤12 specialists; each node has explicit IPO + win criteria.
• Leslie WC must be present or planned before harness is treated complete.
• Evidence over assertion; dry-run before live spawn.
• Stay in domain; escalate mesh-wide orchestration to leslie or /opgrok.

Procedural Workflow

Domain procedure

1. Parse goal → role set, dependency edges, I/O contracts.
2. Resolve SuperGrok tokens via core/registry/named-hashes.json + core/skills/_framework/REGISTRY.json; prefer identity tokens over paths.
3. Draft mesh / harness; mark dry-run vs live; note context budget.

Role method (audit)

1. Declare checklist scoped to this mesh (hire thrift, graph I/O, WC, identity tokens, dry-run path, context cost).
2. Run python3 core/tools/craft_harness.py (or inspect existing harness artifact) and verify node list ≤12 with IPO fields populated.
3. Execute python3 core/tools/run_harness.py --dry-run when a runnable plan exists; treat silence as dry-run success, not API failure.
4. Score every checklist item PASS/FAIL with evidence (path:line, registry key, or command output). Rank FAILs; patch only in-scope defensive fixes.
5. Confirm Leslie WC block exists or is explicitly scheduled before any live spawn.

Domain checklist

• Hire list thrifty (≤12)
• Graph edges + I/O contracts explicit
• WC present or planned
• Identity tokens preferred over body dumps
• Dry-run path known / exercised
• Context pack excludes full skill bodies

Eval dimensions

• Hire thrift
• Graph clarity
• WC readiness
• Context cost

Close

1. Verify: every FAIL has path:line or command evidence. Fix once in scope or escalate to leslie.
2. Emit:

WIN: PASS|FAIL
SG: sg-0101 agent-audit
EVIDENCE:
- ...

Constraints & Gotchas

• Unbounded hires explode context and cost; hard-cap ≤12.
• Missing Leslie WC leaves harness unsealed — never mark complete without it.
• Loading full skill bodies instead of Name-Hash tokens wastes the context budget.
• Graphs without I/O contracts deadlock or thrash under spawn.
• Dry-run silence ≠ API failure; only non-zero exit or explicit error is FAIL.
• Identity drift: registry token mismatch vs on-disk skill → re-resolve before hire.
• Do not use outside Agent orchestration (route via /cat-agent or /opgrok).

Anti-patterns

• Spawning unbounded agents / agent spam
• Skipping Winning Condition
• Embedding all skill bodies in one prompt
• Treating dry-run quiet as broken tooling
• Hiring by free-text name without Name-Hash lookup
• Writing exploits, malware, or undisclosed destructive automation

Definition of Done

• Checklist fully scored; every FAIL has concrete evidence.
• Hire list ≤12, graph I/O clear, WC present/planned, tokens not bodies.
• WIN: PASS with evidence paths/commands; downstream SuperGroks can consume without clarification.
• On residual FAIL: one fix attempt or explicit escalate — no silent ship.

Optional Tool Surface

• python3 core/tools/craft_harness.py
• python3 core/tools/run_harness.py --dry-run
• core/registry/named-hashes.json + IdentityIndex
• core/skills/_framework/REGISTRY.json
• Agent runtime: spawn_subagent, read_file, todo_write
• Binary id: opgrok.sg.agent-audit

References

• core/skills/agent/SKILL.md (category navigator)
• core/tools/domain_enrichment.py (source expertise tables)
• core/skills/_framework/ENHANCEMENT_PROTOCOL.md
• core/skills/_framework/NAVIGATION.md