SMAC — Stochastic Multi-Agent Consensus

A Claude Code skill for deep codebase research via fan-out/fan-in multi-agent consensus with cross-verification.

What it does

/smac dispatches 5+ Sonnet "researcher" agents in parallel, each investigating a different angle of your codebase. Then 5+ Sonnet "verifier" agents circularly fact-check each researcher's findings by re-reading the actual code. Finally, Opus synthesizes everything into a ranked, evidence-backed report.

Every finding carries:

• file:line citations (no speculation)
• Impact × Confidence × Verification score
• CONFIRMED / PARTIAL / DISPUTED verdict from a second agent

Why

Single-thread analysis misses angles. Self-verification has blind spots. SMAC solves both: parallel coverage + circular cross-checking so no agent grades its own work.

Installation

Clone into your Claude Code skills directory:

git clone https://github.com/<user>/smac-skill ~/.claude/skills/smac

Restart Claude Code. Invoke with /smac or /smac "your research question".

Usage

/smac
/smac "what are the riskiest code paths?"
/smac "where are the performance bottlenecks?"

Output:

• Full report → docs/smac-reports/YYYY-MM-DD-{slug}.md
• Run log → docs/smac-reports/_smac.log
• Top 5 findings printed to terminal
• New findings auto-appended to _backlog.md as [NEW] entries (pocketDEV integration)

pocketDEV integration

Before researching, SMAC reads _changelog.md and _backlog.md from the project root and passes that context to every researcher so they do NOT re-propose solved work. After synthesis, CONFIRMED and PARTIAL findings are appended to _backlog.md automatically (duplicates skipped).

Cost

~280–420k tokens per run (10 Sonnet agents + Opus scout + Opus synthesis). Deliberate, high-value tool. Use for release audits, not casual checks.

Credits

Original design by a friend. Hardened fork by Ionuț Roșu: portable paths, operational logging, changelog/backlog integration, auto-append to pocketDEV backlog.

License

MIT