Multi Agent Evolution Framework — Data skill for Claude Code

MAF — Multi-Agent Framework

CLI-agnostic orchestration for code-modifying agents. MAF plans a DAG of agent tasks, dispatches each node to a named role (coder, tester, security, reviewer) with its own tool allowlist and policy bindings, and emits a signed SLSA-style attestation bundle for every run.

The framework wraps any of several backend LLM CLIs — Claude Code, Gemini, Codex, Ollama, or OpenRouter — behind a single adapter interface, then layers:

• Per-role dispatch — each DAG node runs as a specific role with its own system prompt, allowed tools, and policy rules. See docs/ROLES.md.
• Two execution modes per role — cli (opaque single invocation, the default) or in-process (a gated multi-turn agent loop where MAF intercepts every tool call). See In-process execution below.
• Typed processor pipeline — per-turn behavior (prompt assembly, secret redaction, transcript, security gate) is a bundle of hook-indexed Processors with contract-enforced, permitted mutations. See @maf/processors.
• Policy engine — minimatch-based path globs, agentRole predicates, and Cypher memory-pattern queries. Three verdicts: Allow / Deny / Escalate-for-approval.
• Memory — KuzuDB graph for structured run history (tasks, failures, tool invocations) plus an LCM context engine for transcript compression and retrieval.
• Automatic security gate — every coder diff is reviewed by a security agent in parallel with the human review gate. Critical/high findings block the run (on both execution paths, regardless of run outcome).
• Secret redaction — genuine credential formats (API/LLM/cloud keys, tokens, private keys) are stripped from tool output before it reaches the model history, transcript, memory graph, or the signed attestation. The attestation stays byte-faithful except those secret substrings.
• Attestation — every run produces a signed in-toto bundle recording the toolchain,