8 Canonical Agent Roles
Specialized role state machines: planner, researcher, implementer, verifier, etc.
M31A partitions autonomous engineering tasks across 8 specialized agent roles, each with dedicated prompt contracts, capability boundaries, and step budgets.
Canonical Agent Roles
Analyzes objectives, decomposes requirements, and generates acyclic candidate task DAGs.
Explores codebase structure, index symbols, reads docs, and surveys dependencies.
Defines cross-module interfaces, contract types, API boundaries, and architectural patterns.
Authors source code modifications, writes tests, and applies targeted patches.
Performs semantic code reviews, auditing proposed diffs for security and regressions.
Executes automated test suites, type checking, formatting, and produces cryptographic evidence.
Analyzes failed verification runs, compiler diagnostics, and classifies failure modes.
Consolidates verified worktrees, stages changes, generates RFC-compliant commit trailers.
Anti-Fake-Diff Enforcement
The reviewer and verifier roles actively scan proposed code modifications for placeholder shortcuts such as todo!(), unimplemented!(), or truncated comments. Any patch containing fake progress is rejected before task completion.