Glossary & Core Concepts

Authoritative definitions of core terminology used across the M31A platform.

Agent Role

A specialized LLM prompt and policy profile with constrained capabilities. Examples: planner, implementer, verifier, reviewer, diagnostician, architect.

Autonomy Mode

The operational posture of the runtime controlling approval requirements. M31A implements 5 modes: Safe, Plan, Assisted, Autonomous, and Unattended.

Capability

A typed system interface (e.g. fs.read, process.exec, git.commit) that tools must acquire to perform work.

Checkpoint

A durable, content-addressed snapshot of execution state, workspace files, and artifact ledger entries committed atomically to SQLite.

ExecutionRun

The canonical root orchestrator entity representing an autonomous task execution from intake to completion (consolidated in database migrations 030-032).

FsArtifactStore

Content-addressed filesystem storage located in platform directories (~/.local/share/m31a/artifacts) storing generated diffs, test logs, and review summaries.

Nyquist Gate

The evidence-backed completion gate enforcing that no task or execution run may transition to Succeeded without deterministic verification evidence.

Policy Gate

The Layer 1 security evaluator intercepting all tool calls and classifying them into Allow, Deny, Ask, or Escalate.

Ratatui Cockpit

The high-performance interactive terminal user interface projecting live execution state.

Worktree Isolation

A dedicated Git worktree branch (m31a/run/...) created for an execution run to ensure experiments never corrupt the operator's working branch.