agent-orchestrator
/agent-orchestrator
Planuje i koordynuje autoryzowaną pracę wielu agentów z zależnościami, właścicielami zasobów i sprawdzeniami akceptacji.
- Autor
- ios3kov ↗
- Repozytorium
- codex-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 7
- Skrypty
- 1
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 4
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 1 skrypt w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/validate_plan.py
Co robi i kiedy
Planuje i koordynuje autoryzowaną pracę wielu agentów z zależnościami, właścicielami zasobów i sprawdzeniami akceptacji.
Oryginalny opis (EN)
Plan and coordinate authorized multi-agent work with explicit dependencies, resource ownership, acceptance checks and integration. Use when the user requests delegation, or an authorized engineering workflow identifies independent tasks worth parallel execution and host policy permits it. Small serial tasks remain direct.
Jak działa
- Decide and divide
- Coordinate and integrate
Podgląd SKILL.md
--- name: agent-orchestrator description: Plan and coordinate authorized multi-agent work with explicit dependencies, resource ownership, acceptance checks and integration. Use when the user requests delegation, or an authorized engineering workflow identifies independent tasks worth parallel execution and host policy permits it. Small serial tasks remain direct. --- # Agent Orchestrator Use `production-engineering` for requirements and final acceptance, and `smart-router` before choosing worker models/effort. Resolve companions by name in the current catalog. One coordinator owns integration and the combined-result check. ## Decide and divide Check user authorization, host delegation policy and available tools. Use the host subagent API for subtasks; a user-visible new chat requires the user's specific request. If delegation is unavailable, unauthorized or more costly than direct work, continue serially. This skill installs no runtime or background workers. A team needs at least two useful independent outcomes or one bounded specialist/review task with a concrete benefit. Account for briefing, duplicated context, waiting and integration overhead. Divide by outcomes and real interfaces while preserving the original requirements. Each worker receives objective, raw inputs, acceptance checks, owned paths/resources, dependencies, constraints and side-effect boundaries. Use a flat team with direct children only. Default to at most two active workers and four distinct workers for the task, counting reviewers and replacements. Apply any stricter host or user limit. Set the budget before dispatch and reuse an existing worker when its context still fits. Children return requests for further work to the coordinator. Dispatch ready, compatible tasks together. Treat reads versus another worker's writes, and concurrent writes, as conflicts unless an explicit snapshot/ownership arrangement makes them safe. Isolate worktrees/copies or serialize conflicts, including external records and test services. Only the coordinator releases a dependent task, after accepting its prerequisite against the agreed checks and pinning the artifact revision. A worker finishing or reporting PASS does not release a dependency. For file-plan validation, read [plan-validation.md](references/plan-validation.md) before using the optional helper. It checks declared IDs, cycles and unordered path conflicts; actual filesystem aliases and non-file effects still need review. ## Coordinate and integrate Keep one versioned task record, written by the coordinator. Give workers the relevant brief, source links and decisions rather than the whole conversation by default; use a fresh or selectively seeded context when the host supports it, and pass the exact constraints and source revision needed for the task. Context separation does not isolate files or grant permissions. Ask each worker for a short result with task ID, input and result revision or artifact identity, actual changed paths, and each applicable check with its command or procedure, result and evidence pointer. Include blockers, unrun checks and decisions that affect others. Reconcile these into the shared record and notify affected workers before dependent work continues. Shared files do not automatically synchronize agent context. Track actual agent IDs, attempts, input revisions and pending/running/completed/failed/blocked states. Use bounded waits and independent work while waiting. A missing reply or timeout leaves the task unfinished. Retry a recoverable failure only after checking prior effects; reconcile unknown external effects before replaying a mutation. Apply the delegated-result checks in `independent-reviewer` as coordinator self-review unless a separate review is justified and authorized. Resolve conflicting findings, protect user changes, integrate once and run combined-result checks. For independent review, give the reviewer requirements and candidate artifacts without coaching the verdict. Report who actually reviewed what; coordinator self-review is not independent review. Design inputs reviewed 2026-10-20: [OpenAI subagents](https://learn.chatgpt.com/docs/agent-configuration/subagents), [Ruflo coordination](https://github.com/ruvnet/ruflo/tree/main/plugins/ruflo-swarm), and [Model Picker task ownership](https://github.com/firecrawl/model-picker). Original bounded workflow and helper; no swarm infrastructure, hooks, consensus engine or automatic permanent memory is installed. Scope design input: [session contracts and acceptance loop](https://github.com/adagradschool/scope/tree/bb4c0fb81f9ea77c46f15cbaa9b452214d49d28c).
Pliki w paczce (7)
- LICENSE 1 KB
- LICENSE-SCOPE.md 1 KB
- SKILL.md 5 KB
- agents/openai.yaml 1 KB
- assets/icon.svg 2 KB
- references/plan-validation.md 1 KB
- scripts/validate_plan.py 4 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · ios3kov/codex-skills | 85 | MIT |
| eliyce/paqad-ai | 100 | MIT |
| lingegrace/claude-research-skills | 100 | MIT |
| cravingmaker/skills | 90 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 8fed292b9bb4 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90multi-agent
Koordynuje niezależne podzadania agentów z jasnymi zakresami, własnością i dowodami, gdy delegowanie jest uzasadnione. - 90orchestrator-contract
Ustala zasady delegowania dla agenta nadrzędnego: jakość zleceń, wybór podagentów, przeglądy i odpowiedzialność za wynik. - 90multi-agent-collaboration
Koordynuje wielu agentów przy wdrożeniu lub badaniu: rozdziela odpowiedzialność, porządkuje przekazania i rozstrzyga konflikty na podstawie dowodów. - 100workflow-orchestration
Koordynuje ustrukturyzowane myślenie i równoległe wykonanie przez wielu agentów przy złożonych zadaniach: planowanie pracy równoległej, rozbijanie problemów, koordynacja specjalistów i decyzje architektoniczne. - 100team-orchestrator
Koordynuje zespoły agentów przy równoległym rozwoju: skład, podział zadań, zależności i agregację wyników. - 95relaycrew
Koordynuje równoległą pracę kodową małego zespołu agentów: osobne wątki badań, jasny podział plików i niezależna recenzja.