| name | adversarial-swarm |
| model | sonnet |
| effort | high |
| description | Spawn 6 parallel adversarial sub-agents to audit a worktree or diff range from distinct personas (security, auth, ReDoS, dead-code, API contract, performance), then dedupe and rank findings via coordinator. Use for compressed multi-day audits. Activates on "adversarial swarm", "parallel audit", "swarm audit". |
| tags | ["audit","parallel","adversarial","swarm","security"] |
| triggers | ["adversarial swarm","parallel audit","swarm audit","adversarial-swarm","swarm this branch"] |
Adversarial Audit Swarm
Purpose
Compress multi-day wave audits into a single afternoon by running 6 specialized adversaries in parallel, then deduping findings. Complements audit-wave — use swarm for discovery, audit-wave for execution.
When to Use
- Branch has 1000+ LOC delta vs main
- Time-boxed review window (need findings today)
- Multiple feature areas touched (swarm parallelizes well when diffs are independent)
Don't use for:
- Small PRs (overhead > benefit)
- Single-feature branches (one Explore agent suffices)
- Performance-critical audits needing profiler runs (use chrome-devtools-mcp directly)
Execution
Phase 1: Scope
- Diff range: default
main...HEAD, override via user message
- Capture file list:
git diff <range> --name-only
- Capture LOC delta:
git diff <range> --stat | tail -1
Phase 2: Dispatch Swarm
Use superpowers:dispatching-parallel-agents for the mechanics. Spawn 6 Explore subagents in one message (single Agent tool block with 6 tool_use entries). Dispatch each finder with model: haiku — each persona is a fixed adversarial checklist over the diff, so the cheap model is sufficient; the sonnet coordinator (this skill) does the judgment-heavy dedup and severity ranking in Phase 3. Each gets:
| # | Persona | Focus |
|---|
| 1 | Security/Injection | XSS, SQLi, prompt injection, command injection, DOMParser misuse, unsafe innerHTML |
| 2 | Auth/Authz | Auth bypass, missing token validation, hardcoded secrets, dev-namespace leakage, IMS misuse |
| 3 | ReDoS & Input Validation | Catastrophic regex, unbounded input, missing length caps, URL parsing bugs |
| 4 | Dead Code | Unused exports, unreachable branches, orphaned helpers, commented-out blocks, deprecated configs coexisting with replacements |
| 5 | API Contract Drift | Changed response shapes, removed fields, silently widened types, breaking param reorder |
| 6 | Performance/Memory | Memory leaks, missing cleanup in disconnectedCallback, unbounded arrays, sync work in render, N+1 queries |
Subagent Prompt Template
Each subagent gets this prompt (swap <PERSONA> and <FOCUS>):
You are a <PERSONA> adversary auditing a code diff range for a pre-review pass.
Scope: files in `<DIFF_RANGE>` — get the list with `git diff <DIFF_RANGE> --name-only`.
Focus exclusively on: <FOCUS>
For each finding, output one JSON object per line (JSONL):
{"file": "path/file.js", "line": 123, "severity": "critical|high|medium|low", "title": "one-line", "repro": "one-sentence how to reproduce or why it's broken", "fix_sketch": "one-sentence suggested fix"}
Rules:
- Only report issues you can cite with file:line
- Do NOT propose refactors outside your persona
- Rank severity honestly — critical = ship blocker, low = nice-to-have
- Skip findings already covered by existing tests (grep for the symbol in test/ dirs)
- Return under 30 findings; pick the highest-severity ones if you have more
Phase 3: Coordinator Dedup
After all 6 return:
- Parse each JSONL output into a list
- Dedupe by
(file, line) — if multiple personas flagged the same location, merge titles and keep highest severity
- Sort by severity (critical → low), then by file
- Write to
.claude/plans/<TICKET>-swarm-findings.md with this structure:
# <TICKET> Adversarial Swarm Findings
Generated: <timestamp>
Diff range: <range>
Personas: security, auth, redos, dead-code, api-contract, performance
## Critical (N)
- `file:line` — title (from: persona1, persona2) — repro → fix sketch
## High (N)
- ...
## Medium (N)
...
## Low (N)
...
## Deduped out
- `file:line` — merged under Critical #3 (auth + security agreed)
Phase 4: Handoff to audit-wave
Present the summary to the user with total count per severity and ask:
"Ready to feed these into audit-wave as Wave 1 (critical+high) and Wave 3 (medium+low)?"
If yes, invoke the audit-wave skill and populate the plan doc's Wave 1/3 sections from the findings file.
Token Budget Notes
- 6 parallel Explore agents ≈ 6 × ~3-5k tokens of tool output
- Keep each agent's prompt under 500 tokens
- Cap per-agent findings at 30 to prevent context bloat
- Coordinator reads JSONL (cheap) not full agent transcripts
Related
audit-wave — execution framework for findings produced heresuperpowers:dispatching-parallel-agents — parallel agent mechanics.claude/rules/coding.md — problem-solving priority (DELETE > MODIFY > ADD)
