--- name: orchestrate description: 'Coordinate several workers: what idle agents do next, which finished work gets checked first, how to recover a dead one. Use when: managing multiple agents.' practices: [team-topologies, evidence-based-engineering] hexagonal_role: supporting consumes: [accepted-intent, native-work-state, candidate-evidence] produces: [native-handoffs, reconciled-feedback] context_rel: [] skill_api_version: 1 user-invocable: true metadata: graph_root: true tier: execution dependencies: [] capabilities: [coordinate_native_work, recover_assignments, reconcile_feedback] effects: [dispatch_authorized_workers, update_native_handoffs] canonical_status: canonical disposition: keep output_contract: observed native assignments, exact candidate and judgment references, affected-work handoffs and explicit remaining gaps --- # Orchestrate Coordinate caller-authorized work through its existing tracker and runtime; a clear task needs zero mandatory skills. The tracker owns assignments and dependencies, the runtime owns running contexts and bounds, and repository policy owns integration and delivery. Orchestrate decides what each worker does next. [Agent Native](../agent-native/SKILL.md) launches and observes workers; [Navigate](../navigate/SKILL.md) picks the wave on a bead graph toward frozen acceptance and records verdicts. ## Before any dispatch - **Finished is not done.** Exit 0, a pushed branch, a closed tracker item or a worker saying "done" makes a candidate. It still owes its checks, integration and, when one is owed, a fresh judgment. - **Drain before starting.** While candidates wait on checks, repair, integration or an owed judgment, free capacity goes there first: candidates that unblock held dependents, then the oldest. A free slot alone is not a dispatch reason; start new implementation with what is left. - **Judges did not author.** Review or validation of a candidate goes to a context that did not write it. - **Readiness is content.** A prerequisite counts only when its bytes are in the intended checkout. A closed item whose change is missing, stale or unavailable there is not ready: hold its dependents, keep its native status, and report the gap. - **Reconcile before resuming.** Match tracker assignments against observed runtime state, validators included, before dispatching. Never duplicate an assignment because this conversation lacks it. - **Decisions are not build work.** A decision item goes to the resolver it names ([Navigate](../navigate/SKILL.md#foggy-epics)), never to an implementer; with BD, pull build work with `bd ready --exclude-type decision`. ## Recover the actual work Read acceptance, examples, scope, settled choices and open questions from their owner and the existing native handoff, without repeating settled interviews or requiring the full transcript; a missing or contradictory pointer needs source investigation, not a guess. For each task, check who holds it, its workspace and starting content, the write scope it occupies, pending checks and review, the current candidate, who integrates it, and whether each prerequisite's content is present. An empty ready list does not prove completion. ## Choose the next useful dispatch Concurrency follows the observed bottleneck. Reserve capacity for integration, review and repair, and reduce new starts while candidates accumulate. Record only the concrete constraint and next action in the existing native handoff, then reassess when evidence changes. Do not add a capacity ledger or queue. [Agent Native](../agent-native/SKILL.md#before-launch-and-while-running) owns runtime mechanics, disjoint write scopes and isolation; one-shot headless runs go through [Codex Exec](../codex-exec/SKILL.md), [Claude Exec](../claude-exec/SKILL.md) or [AGY Native](../agy-native/SKILL.md). Serialize shared paths, generated companions included. A worktree separates Git edits; it does not enforce restricted-source or model-egress limits ([model dispatch](../agent-native/references/model-dispatch.md)). Dispatch a genuinely fresh implementer for one coherent accepted task, without the coordinator's transcript or unrelated research; a new goal, role label, cleared summary or resumed author context is not fresh. Pass the accepted examples, applicable constraints, exact starting content, usable prerequisites, authorized write/output scope, source pointers, required checks, integration responsibility and real remaining bounds; brevity cannot omit a constraint. A pointer counts only if the worker can open it: a link relative to a skill's own folder, or a file outside the worker's workspace, does not resolve there, so pass a path it can read or paste the rule itself. Record the worker's observed identity at startup through Agent Native's association procedure. [Implement](../implement/SKILL.md) owns the complete change, meaningful checks and direct repair; it is optional guidance for that worker, not a compulsory stage. The handoff returns candidate identity, changed scope, check facts, discoveries and gaps; prompt delivery proves neither engagement nor acceptance. When several workers apply one set of decisions, such as a batch of review findings or one rewrite that must read the same in many files, make each decision once in one context and send workers only what was decided. A worker returns an uncovered case, or a blocker with evidence the decider could not see, instead of improvising or reopening the decision; the decider settles it and never drops it. [Decide once, apply in parallel](references/decide-once-apply-in-parallel.md) covers both shapes. ## Integrate and obtain judgment Name the integration responsibility before launch, and decide then whether the integrated candidate needs a fresh judgment at all. Follow the consumer repository's integration policy, include all changed paths and generated companions, and run affected checks on the actual integrated subject. Acceptance of a leaf does not establish the combined release. Within that policy, combine parallel results one at a time in dependency order and run the affected checks after each, so a failure points at one result; a merge without conflicts does not show that two changes work together. A result that conflicts, or whose own changed paths or prerequisites changed on the target after it started, goes back to its author through a follow-up that authorizes that update ([Agent Native](../agent-native/SKILL.md#launch-and-observe)), or is redone from the current target, and reruns its checks; it is not merged blind. A write outside a worker's declared scope, or collisions that keep recurring, show the scope split cannot be trusted for this change: finish the rest serially. For an ordinary candidate the integrated checks and CI are the gate. Assign one fresh author-distinct judgment through [Validate](../validate/SKILL.md#when-a-fresh-judgment-is-worth-it) only in the costly cases it names: the caller asks, a mistake cannot be cheaply undone after it lands, or no deterministic check covers the changed behavior. Validate owns the single round, what counts as a finding, and when a repair needs only its checks rerun. Preserve every explicitly required review leg; advisory Review, Plan challenge and Council advice cannot stand in for a judgment the caller requested. ## Reconcile feedback and resume Preserve successful and failed evidence in the existing native task or handoff, and update affected unfinished work's native dependencies or handoff within authority. Stop or explicitly re-scope an affected active assignment before it continues on a disproven premise; the revision takes effect only on observable acknowledgment or stopped runtime state. Unaffected work continues unchanged; reconciling unchanged facts creates no new artifact or dispatch. [Plan](../plan/SKILL.md) owns consequential uncertainty and refining the next slice. New evidence may change an approach within the accepted outcome; a different promised outcome or authority choice returns to the caller, never to agent advice. Do not reopen accepted work to fit a revised story. Known failures return to the responsible task for direct repair. A genuine causal stall is an unknown cause, a recurrence, no progress or a wrong objective ([RPI](../rpi/SKILL.md#causal-stall-and-bounds)); it admits at most one authorized bounded fresh helper for that incident within remaining bounds, and an unhelpful answer ends that attempt. Cancellation, refusal or exhausted bounds skip help. Replacement workers, retries, new subjects and compaction never reset those bounds. Use native waits for unchanged pending state instead of repeated analysis or probes. To recover a dead worker, confirm its process stopped, not only its pane; capture its state and preserve its partial work recoverably; move its tracker assignment to the replacement rather than filing a duplicate; and start the fresh worker from that branch as exact starting content while dependents hold. When maintained context could change the next action, use [Memory find/recall](../memory/references/recall.md). A caller-selected external factory keeps its coordinator in control: hand it source intent through its supported door, never run its sessions by hand or mirror its work in a second tracker, and judge the returned content independently; its completion neither authorizes delivery nor establishes acceptance. For Gas City, follow [Using GC](../using-gc/SKILL.md). ## Status block When the caller asks for coordination status, return this shape; answer a quick question in prose. No other report or persisted artifact is required unless an existing consumer needs one. ```text assignments: -> , state as observed (how); last activity | unobserved candidates: -> ; checks ; judgment next: -> , because held: , blocked by handoffs: gaps: ```