--- name: orchestrate-reference description: Reference rules for the Concierge orchestrator that few objectives need - customer routing tables and retired routes, todo rows for dependencies, cancellation, rollover and resume, and learning-store scopes in handoffs. Only the Concierge orchestrator loads it, when its profile or the orchestrate skill points here. --- # Orchestrate reference Load this only for the case at hand and follow the matching section. It grants no tool, access or authority. ## Routing tables If a `copilot-concierge-routing` skill or skills named `copilot-concierge-routing-` are available, load them first and follow their table: - A row whose role is a Concierge role (`scout`, `product-manager`, `worker`, `enterprise-ux`, `react-frontend`, `node-backend`, `graphql`, `reviewer`, `escalation`) sends that role's work to the listed agent instead. - A row with any other role adds a specialist for the work in its "Use when" column. - Skills in a row go into that worker's handoff under `skills to load`, whichever agent does the work. - When rows conflict, `copilot-concierge-routing` wins over `copilot-concierge-routing-`. - Prefer the explicitly routed agent when its exact ID is available. A model pin is a preference, not an eligibility requirement. Preserve applicable skills and role restrictions. - Compatibility: existing role keys remain valid, and an available custom agent named by a row still wins. For a row naming the retired built-in `copilot-concierge:react-frontend`, `copilot-concierge:node-backend` or `copilot-concierge:graphql`, dispatch `copilot-concierge:worker` as a separate delegation and add the corresponding domain skill alongside all row skills. Report this migration and chosen agent. For retired `copilot-concierge:enterprise-ux`, report that the profile and bundled skill were removed; use a suitable available custom route or worker within scope, without claiming the removed specialist ran. If a row names the former `enterprise-ux` skill and no such custom skill is available, report that gap and substitute `plan` for design planning and `react-frontend` only for applicable React work; retain all other row skills. An available user-supplied skill is not a removed built-in. Never rewrite the user's routing files silently. - If a route or preferred model is unavailable, use a suitable available shipped agent or execute directly within scope. Report the substitution and actual model when known. If no permitted capability can complete the work, report the exact blocker. Never bypass host permissions or a user restriction. Report substitutions and reconcile uncertain effects before retrying. Resolve matching routing rows before choosing direct execution: explicit rows remain delegation instructions and cannot be bypassed by doing the work yourself. Unrouted default work is eligible for direct execution. A plain question can be answered directly. A clear ticket needs no interview. Give tightly coupled layers and their tests one execution owner. Do not split only by stack names. Dispatch independent work in parallel after interfaces are clear; serialize work that touches the same files or unsettled contracts. Keep review proportional: simple nonbehavioral edits may use focused checks, while meaningful or risky changes require independent review. ## Todo rows Your profile says when to write the checklist and its `todos` mirror. Also: - Add a `todo_deps` row (`todo_id`, `depends_on`) only for a real ordering between your tasks. When a revised plan drops a task, delete that row and its dependency edges. - On cancellation, set your unfinished rows to `blocked` with `description` "Cancelled". - `ROLLOVER` and `RESUMED` keep the earlier ids: read them first with `SELECT id, title, status FROM todos WHERE id LIKE '/%'` and continue those rows instead of adding duplicates. ## Learning stores in handoffs Carry the user's exact file-write restrictions into every handoff and handover. Set `learning` separately for private memory and repository knowledge. Normal Scout research may curate these stores before implementation approval; wholly read-only work or an explicit user prohibition on all file writes disables both, and a repository-only or memory-only restriction stays scoped to that store. Do not invent a blanket no-write restriction from orientation or “do not implement changes.” Pass the `scout` skill and the same learning scope to a routed Scout or escalation doing Scout work; Sol using the Scout skill follows the same helper and learning scope; other roles have no learning-write exception. Serialize private memory mutations per OS user and knowledge mutations per repository across objectives; the helper's revision checks detect missed coordination. Report every learning write as a changed file. Private memory lives at `~/.config/copilot-concierge/memory.json`; shared knowledge at the assigned repository's `docs/concierge/knowledge.md`. The old `.copilot-concierge/project-notes.md` is only read-only, individually reviewed migration input. Read or maintain stores only through the loaded Scout skill and its bundled helper, directly or through Scout; never directly edit a store or bypass a helper refusal. Retrieve only relevant preference summaries. Apply them to Concierge's interaction and narrowly appropriate task instructions; never put raw memory, user evidence or personal history in repository docs, worker handoffs, commits, tickets or pull requests. Memory never changes approval, security, tools or model pins. A request to inspect/forget private memory is not a request to inspect/delete repo knowledge. Current user instructions and source evidence win. If storage is unavailable, continue the objective session-only and report the gap; do not block ordinary work on optional personalization.