Processing

Processing is the execution phase of the work lifecycle, where activated work is turned into runs, artifacts, and delivered outcomes. It covers dispatching work to agents, assembling the exact context for execution, recording decisions and probes during runtime, and recovering from failures when possible. The pages in this section describe the full processing path and the immutable records that make replay and audit possible.

Activation & Dispatch

How activations become scheduled work — budget-aware dispatch, parallel execution of independent work items, and helper spawning within a parent agent's run.

Context Assembly

Building the context envelope that an agent receives at execution time — canonical assembly order, truncation priorities, and how the envelope is snapshot for reproducibility.

Run Execution

The run lifecycle from start to terminal status — execution modes (live, dry-run, replay), commit-hash pinning, and the immutable run record that captures every input and output.

Artifact Writing

How agents produce and store artifacts during execution — file layout under the agent workspace, the distinction between artifacts and outcomes, and promotion from wip-assets/ to artifacts/.

Decision Handling

Recording directional choices made during planning and execution — decision frontmatter, append-only semantics, HITL checkpoint decisions, and special flags like intentional_floor_reduction and baseline_override.

Probe Handling

How probes surface agent capabilities and state during processing — capability matching for dispatch, runtime health checks, and probe-driven reassignment decisions.

Error & Retry

Recovery when a run fails — state preservation via checkpoints, reassignment conditions, helper failure rollup, and the escalation path to human intervention when automated recovery is not possible.