/agentic-audit

Launch a panel of specialized agents to audit a pull request from multiple perspectives before merge.

Arguments

• $ARGUMENTS - PR number (optional, defaults to current branch's PR), plus optional agent count (default: 5). Examples:
  • 42 (PR #42, 5 agents)
  • 42 7 (PR #42, 7 agents)
  • 3 (PR #3 or 3 agents — resolved by checking if a PR with that number exists)

Instructions

1. Gather PR Context

PR_NUM=${1:-$(gh pr view --json number -q .number)}
AGENT_COUNT=${2:-5}  # min: 3, max: 8
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)

# PR metadata
gh pr view ${PR_NUM} --json title,body,headRefName,baseRefName,files,additions,deletions

# Full diff
gh pr diff ${PR_NUM}

# Changed file list
gh pr view ${PR_NUM} --json files -q '.files[].path'

Build a context bundle for agents containing: PR title, description, branch names, full diff, and the list of changed files.

2. Select Agents

Pick AGENT_COUNT agents from the roster. Always include the core 3. Fill remaining slots from the extended roster based on what the diff touches.

Core (always included)

Extended (pick by relevance to the diff)

3. Launch Agents

Launch all agents in parallel using the Task tool. Each agent receives the full PR context bundle and their persona.

Agent prompt:

You are "{AGENT_NAME}" — {PERSONALITY}

Audit this pull request from the lens of **{LENS}**.

## PR Context
- **Title**: {PR_TITLE}
- **Branch**: {HEAD} → {BASE}
- **Changed files**: {FILE_LIST}

## Diff
{FULL_DIFF}

## PR Description
{PR_BODY}

## Your Review
1. Rate the PR overall 1–5
2. List your top findings (up to 5), each with:
   - Severity: critical / major / minor / nit
   - File and line reference
   - What's wrong and why it matters
   - Suggested fix (concrete, not vague)
3. Note anything done well worth preserving
4. One-paragraph verdict

## Rules
- Review the DIFF, but also read surrounding code in changed files for context.
- Be specific. Reference exact lines. "This might be a problem" is useless.
- Be opinionated. Strong views, loosely held.
- 3/5 = adequate, would merge with minor comments. 5/5 = exemplary. 1/5 = should not merge.

Run agents as foreground Task calls. If more than 4 agents, batch: first 4 in parallel, then the rest.

Delegation tiers (cost discipline — do not strip):

Roles per /tiered-delegation: workhorse = one tier below the session ceiling, mechanical = cheapest available tier. Concrete model names live only in that skill's CUSTOMIZE-marker ladder — never here.

• Panel agents run at the workhorse tier by default. Reading a diff through a persona lens is workhorse work; do not spawn panel agents at the ceiling. The orchestrator's synthesis (step 4) carries the expensive judgment.
• Mechanical prep at the mechanical tier. Building the context bundle (file lists, diff chunking) and any pre-classification of changed files runs at the mechanical tier at low effort.
• Sample-verify instead of up-tiering. If the panel's findings look unreliable, re-check a ~10% sample with a stronger tier rather than re-running the panel up-tier.
• Fan-out budget: ~12 workhorse/ceiling subagents per run. The budget counts workhorse- and ceiling-tier agents (the panel); mechanical-tier prep agents and the ~10% sample-verify checks do not count against it. Exceeding 12 requires an explicit one-line justification in the synthesis; 20 is a hard ceiling that includes every refute/verify agent, and no justification lifts it (≤3 refuters per finding — queue rounds rather than widen). The 8-agent panel maximum sits well inside this budget.

4. Synthesize and Report

After all agents return, output a single synthesis directly to the user (no files written):

## Agentic Audit: PR #{PR_NUM} — {TITLE}

**Agents**: {count} | **Aggregate**: {average}/5

### Agent Panel
| Agent | Rating | Top Finding |
|-------|--------|-------------|
| Skeptic | X/5 | One-line summary |
| Builder | X/5 | One-line summary |
| ... | ... | ... |

### Consensus Findings
Issues where 3+ agents agree. These are high-confidence and should be addressed.

1. **{Finding}** (Skeptic, Builder, Guardian) — {description + file:line}
2. ...

### Contested Points
Where agents disagree. Present both sides, assess who's right.

### Action Items
Prioritized list of recommended changes before merge, ordered by severity.

### Verdict
One paragraph: merge as-is, merge with changes, or rework needed.

5. Offer Next Steps

After presenting the synthesis, ask the user:

Want me to fix any of the findings? I can address them as commits on this branch.

Rating Scale