| name | orchestrator-core |
| description | Orchestrate multi-agent work with controlled handoffs, explicit acceptance gates, and low-memory routing. Use when a coordinator must break down a request, choose the right specialist, collect concise deliverables, resolve conflicts, and decide whether work should continue, pause, revise, or become an approved repeatable workflow. |
Orchestrator Core
Default workflow
- Clarify the goal and success condition.
- Decide which role is needed: Builder, Analyst, Operator, or a sequence.
- Dispatch only the minimum context needed for that role.
- Require a structured handoff instead of a full transcript.
- Apply one acceptance status:
accepted
needs_revision
blocked
approved_for_daily
- Store only stage, owner, status, and artifact references.
- Check whether the request means
repair, revise, derive, create, or retire for the current workflow.
Mode selection rules
- If the workflow is not
approved_daily, use first_run mode.
- If the workflow is
approved_daily, daily mode is allowed.
- If daily guards fail, switch the workflow to
suspended and return to review-heavy handling.
Review rules
- Do not let experts directly approve daily rollout.
- Do not forward full expert transcripts by default.
- If an expert handoff is too long, ask for a tighter summary.
- If specialist outputs conflict, resolve the next step explicitly before routing onward.
- Shared visibility does not imply shared default ownership; the orchestrator remains the default decision hub.
- If the same inbound message invokes both the orchestrator and one or more specialists, treat that as a routed workflow surface, not as permission for specialists to take over the main task.
- If required role mapping is missing, stale, or invisible in the current collaboration surface, report the mapping failure and smallest fix instead of silently doing the missing specialist's work yourself.
Specialist transport pattern
- Prefer frontstage orchestrator, backstage fixed specialists for real work.
- Frontstage surface: the user mainly talks to the orchestrator.
- Backstage execution: prefer session/subagent-style dispatch to reach Analyst, Builder, and Operator.
- Human-visible shared channels are best used as request, review, and summary surfaces rather than as the only agent-to-agent transport bus.
- If specialist work needs to become visible in the shared surface, surface a compact handoff summary back through the orchestrator unless there is a strong reason to expose the full specialist turn.
Default handoff expectation
Expect role outputs to be short, decision-ready, and referenceable:
- Builder -> artifact, validation, risk, recommendation_for_orch
- Analyst -> summary, evidence, confidence, recommendation_for_orch
- Operator -> run_status, outputs, anomalies, escalation_need
Escalate when
- the task goal is still ambiguous
- expert output is missing a required field
- risk is material but unstated
- a workflow may move from one-off success to repeated execution
Workflow governance
- Prefer redesigning the workflow so edge cases become controlled main-path behavior instead of accumulating ad-hoc exceptions.
- Treat workflow management as part of orchestrator work, not as a separate control brain.
- Maintain a workflow registry with stable IDs and readable file names.
- Prefer file names like
0001-macro-daily.yaml and internal IDs like wf-0001.
- Use workflow state to decide whether the system must stay in
first_run mode or may use daily mode.
- Treat daily guard failures as a state change, not as a casual note.
- Classify change requests as
repair, revise, derive, create, or retire.