Skip to main content

Autonomous Execution Lifecycle

Every execution run in M31A transitions through a deterministic state machine managed by src/controller/ and src/state_machine/.

1. Intake & Discovery​

The operator's prompt is parsed into an authoritative ExecutionRun record with a unique UUID v7. The runtime discovers workspace baselines, Git HEAD, and configuration profiles.

2. Planning & DAG Decomposition​

The planner agent breaks down the high-level objective into discrete, typed tasks forming a Directed Acyclic Graph (DAG). Dependencies and critical paths are resolved using the Critical Path Method (CPM).

3. Gated Execution​

Tasks with satisfied dependencies are dispatched to worker agents (implementer, researcher, verifier). Each tool call passes through the non-bypassable Policy Gate.

4. 7-Tier Verification Gate​

Upon task completion, the runtime executes the 7-tier verification hierarchy. The task is only marked Succeeded if all required checks pass with cryptographic evidence.

5. Isolated Attribution & Git Commit​

When the execution run completes, changes are staged, trailers (including M31A-Execution-Run, M31A-Agent-Role, M31A-Verification-Id) are generated, and a clean Git commit is recorded.