Orchestrator Configuration

orchestrator.yaml holds settings for the per-task pipeline's review phase and its observability.

File: ~/.engineer/config/orchestrator.yamlHot-reload: No — requires a daemon restart.

Both review.lenses and observability.live_activity change behavior. Notification delivery is controlled by the daemon — see notification retry and notification suppression.

Review

The Review phase runs one or more lenses — focused agent passes that each examine the change through a single concern (correctness, security, code quality, architecture) and write findings. Then refine consolidates those findings, fixes what it can directly in the code, and decides whether to ship the change, re-check it, or hand it back to an earlier phase.

Field	Type	Default	Description
review.lenses	string[]	["self-review"]	Which review lenses run before delivery. Each is a focused agent pass that writes findings; refine then consolidates them and fixes in place. Options: "self-review", "security", "code-quality", "architecture".

self-review is the default and is usually enough. The others are opt-in for a change that warrants a dedicated pass — add the lens name to the list. Each lens is a single file under the pipeline's review/ folder, so adding one is a small change.

Observability

Controls how much of a task's inner work The Engineer streams to the dashboard. These knobs are additive and best-effort: turning one off only reduces what is shown — it never changes what the pipeline does, costs, or how long it takes.

Field	Type	Default	Description
observability.live_activity	boolean	true	Stream each agent run's live conversation — assistant text, thinking, tool calls and their results — to the dashboard's Agent Calls tab as it happens, and keep it for re-watching afterward.

When on, an agent run is no longer a black box between prompt and result: the owner watches the agent think and act in near real time, and can replay any past run. Only agents that report supports_activity_streaming emit a feed; others are unaffected and fall back to the post-run transcript. The feed is observation-only — a slow or failing dashboard can never affect a run. Turn it off if you want a quieter trail (the agent_call span, with its prompt/result/transcript blobs, is still recorded either way). See the Live Agent Activity how-to for the mechanics.

Example

review:
  lenses: [self-review]   # add "security", "code-quality", and/or "architecture" for deeper passes

observability:
  live_activity: true     # stream each agent run's live conversation to the dashboard (default: true)