Runs

A run is a single execution instance of a work item — multiple runs can exist per work item through retries, dry runs, and replays. Each run is an immutable record capturing the execution mode, org repo git commit hash, context envelope snapshot, agent runtime identifier, and check results, making it the unit of reproducibility in matic. This section covers the mechanics of how runs are produced, budgeted, persisted, recovered, and re-executed.

Run Modes

The three execution modes — live (side effects enabled), dry-run (side effects suppressed), and replay (deterministic re-execution from a prior record). Learn more

Run Record

The immutable on-disk artifact under runs/<run-index>/ containing record metadata, context envelope snapshot, and produced output — and what makes a closed run reproducible. Learn more

Execute Step

Step 5 of the work lifecycle: agent invocation through a runtime adapter, where generation, tool use, skills application, and helper dispatch occur within the context envelope. Learn more

Context Budget

How the seven-item context envelope is assembled and truncated to fit the agent runtime's token window — which components are never truncated, summarised, or trimmed under pressure. Learn more

Recovery Protocol

Failure handling across agent timeouts and daemon crashes — checkpoint preservation in wip-assets/, reassignment criteria, signal replay windows, and cross-referenced recovery runs. Learn more

Replay

Deterministic re-execution from a run record — restoring shared asset state via the recorded git commit hash, reusing the original context envelope, and creating new cross-referenced run records without overwriting originals. Learn more