12-Stage Autonomy Loop
The 12-stage execution lifecycle, loop detector, and 10-dimensional budget model.
Missions are orchestrated deterministically by the AutonomyController across a 12-stage execution lifecycle.
The 12-Stage Execution Lifecycle
- Intake & Validation: Ingests mission intent, resolves effective profile, validates invariants.
- Context Compilation: Gathers repository AST, workspace baseline commit, security policies.
- Plan Generation: Consults planning model to generate a candidate task DAG.
- DAG Reconciliation: Reconciles candidate plan with existing petgraph TaskGraph.
- Topological Scheduling: Selects ready tasks whose dependencies are satisfied.
- Agent Dispatch: Instantiates specialized agent role state machine.
- Two-Phase Reservation: Pre-allocates tokens, memory, and step quotas before execution.
- Tool Execution & Policy Gate: Evaluates tool proposals through 11-stage policy gate.
- Post-Execution Settlement: Reconciles actual consumed resources and logs telemetry.
- Evidence-Based Verification: Runs cargo test, clippy, and syntax checks.
- Checkpointing: Captures atomic two-phase snapshot committing staged artifacts.
- Completion Gating: Evaluates completion criteria and produces signed report.
The 5 Autonomy Modes
| Mode | Mutations Permitted? | Approval Behavior (ASK) | Primary Use Case |
|---|---|---|---|
| Plan | No (Strictly Read-Only) | Denied immediately | Architecture planning, DAG generation |
| Safe | Requires Approval | Pauses mission; awaits confirmation | Sensitive repos, unfamiliar codebases |
| Assisted | Inside Workspace | Pauses only for external side effects | Standard interactive pairing (default) |
| Autonomous | Inside Workspace | Pauses only on explicit policy veto | Self-directed refactoring and features |
| Unattended | Inside Workspace | Converts ASK to DENY fail-closed | Headless CI/CD evaluation runners |
10-Dimensional Resource Budget Enforcer
M31A clamps execution across 10 hard resource bounds: wall-clock seconds, concurrent agents, agent decision steps, model invocations, token usage, financial cost ($USD), process CPU time, memory bytes (cgroups memory.max), artifact disk storage, and task retries.