relaycrew
/relaycrew
Koordynuje równoległą pracę kodową małego zespołu agentów: osobne wątki badań, jasny podział plików i niezależna recenzja.
- Autor
- ScriptNegative ↗
- Repozytorium
- relaycrew ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 10
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/plan.mjs, scripts/session-core.mjs, scripts/session.mjs
Co robi i kiedy
Koordynuje równoległą pracę kodową małego zespołu agentów: osobne wątki badań, jasny podział plików i niezależna recenzja.
Oryginalny opis (EN)
Coordinate parallel coding work, independent investigations, and implementation followed by independent review with a small agent team. Use when a coding task benefits from separate workstreams, explicit file ownership, or focused handoffs; keep small dependent edits solo unless the user requests a team.
Jak działa
- Choose the smallest useful team
- Make ownership explicit
- Relay a brief, not the entire conversation
- Integrate evidence before continuing
Podgląd SKILL.md
--- name: relaycrew description: Coordinate parallel coding work, independent investigations, and implementation followed by independent review with a small agent team. Use when a coding task benefits from separate workstreams, explicit file ownership, or focused handoffs; keep small dependent edits solo unless the user requests a team. license: MIT metadata: author: ScriptNegative version: "0.5.0" --- # RelayCrew Spend context on the work, and add workers only when they have independent work to do. ## Choose the smallest useful team Inspect enough of the project to identify the acceptance checks and shared interfaces. In automatic mode, keep a small fix or a dependent chain in the main agent. Respect an explicit request for solo, team, or review mode. For separable work, retain one coordinator and start with at most two active workers, within the host's actual capacity and the user's budget. A larger queue can run in waves; more workers are not a quality metric. Read [options](references/options.md) when the user requests a mode, budget, review, or checkpoint customization. Use native delegation only when available and permitted by the host and the user's instructions. Otherwise execute the same tasks sequentially and report solo mode. Never invent agents, tool names, or isolation. This skill does not grant permission to publish, deploy, buy services, alter access, or contact other people. Check actual file, editing, execution, and delegation capabilities before promising an outcome. Without source access, request the necessary artifacts or provide a clearly labeled plan. Without execution tools, distinguish a reasoned review from tested behavior. The workflow does not require a terminal; the optional planner does. Read [host notes](references/hosts.md) only when choosing how to delegate in Codex or Claude Code. Do not install an agent runtime to make the skill work. ## Make ownership explicit Before dispatch, give each task an ID, a concrete outcome, dependencies, read paths, write paths, and observable acceptance checks. Declare shared mutable resources such as a database or generated index too. Resolve interface decisions before splitting their consumers. Workers sharing a checkout must not overlap write/write or write/read claims. Two readers can overlap. Treat a directory claim as covering descendants; include lockfiles, generated files, tests, and configuration in ownership. Verify actual isolation before allowing overlapping edits in separate checkouts. Combine changes through the coordinator and check the combined result. For several tasks, optionally run `node <skill-dir>/scripts/plan.mjs <plan.json>`; add `--next` to retrieve only the next wave's handoff capsules. Read [planner format](references/planner.md) only when using it. Record `in_progress` to reserve active workers and their claims. It schedules declared claims; it neither launches agents nor enforces filesystem isolation. For a simple split, keep the plan in conversation instead. ## Relay a brief, not the entire conversation Give each worker this capsule, normally within 2,400 characters. The cap is a briefing target, not a token or spending limit; preserve necessary detail rather than silently truncating it. ```text Task: ID and outcome Context: relevant decisions, paths, and interface facts Read: paths needed to investigate Write: exclusive paths, or none Resources: shared mutable resources, or none Done when: observable checks Return: changed paths; evidence and checks; blockers; follow-up needed ``` Include relevant repository rules and inherited user constraints. Let workers read permitted source files directly. Prefer a fresh context when the host supports it; if history is inherited, disclose that the briefing cap does not bound total context. Do not copy secrets, full logs, or unrelated history into a capsule. Workers must request an ownership change before editing outside their claim. Keep delegation flat unless deeper delegation is explicitly justified and permitted. ## Integrate evidence before continuing Run only ready tasks with compatible ownership. A planned wave is not a runtime guarantee: a worker timeout, failed check, partial edit, or changed interface invalidates dependent work. Inspect the actual diff, update the plan, and resume only the affected tasks after prerequisites are verified. Before transferring ownership, stop the previous worker and confirm that it has stopped writing. An unknown or interrupted worker keeps its reservation; a timeout does not prove it stopped. Preserve useful partial work; do not retry an unchanged failure repeatedly. For interrupted or multi-session work that needs durable attempt tracking, read [session checkpoints](references/sessions.md). The optional helper records a reservation before dispatch, distinguishes a returned report from coordinator acceptance, and rejects stale attempts or changed plan definitions. Pass the current task and attempt IDs to each worker and require both on its return. Inspect the actual diff and reproduce the required checks before accepting. A checkpoint validates bookkeeping, not the truth of reported evidence. For a short run, conversation state or the planner's state arrays are enough. When requirements change, use the reference's reviewed plan migration only after every worker is confirmed stopped and its active attempt is resolved. Compare the old and new plans before applying: retain completed work only after checking that its definition, prerequisites, and actual result remain valid. Changed tasks and affected dependents need fresh work; a matching plan definition does not establish workspace freshness. Ask workers for concise evidence with file locations and actual check results. A worker saying "done" is not proof. Reconcile disagreements against source and reproducible behavior. For an independent review, pass the requirement and resulting diff without the implementer's conclusions; findings must identify a concrete failure or missing check. The coordinator checks the integrated change against the user's original acceptance criteria, then reports the outcome, checks, limitations, and whether delegation was actually used. Report token or cost savings only when measured against a disclosed baseline. The brief cap, worker count, and instruction size do not establish savings.
Pliki w paczce (10)
- LICENSE 1 KB
- SKILL.md 6 KB
- agents/openai.yaml 1 KB
- references/hosts.md 3 KB
- references/options.md 4 KB
- references/planner.md 7 KB
- references/sessions.md 12 KB
- scripts/plan.mjs 15 KB
- scripts/session-core.mjs 16 KB
- scripts/session.mjs 6 KB
Wersje
W bazie nie ma innych skilli o nazwie relaycrew.
Migawka bazy z 13.11.2026 · wersja paczki d25a9e8f00f1 (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
- 100delegate-work
Koordynuje równoległe subagenty, dzieli pracę, izoluje kontekst i unika konfliktów w plikach, a potem sprawdza wyniki. - 90baton
Koordynuje agentów kodujących w jednym projekcie: przydziela zadania, blokady plików, punkty Git, przekazania i niezależny przegląd. - 100team-agents
Uruchamia i koordynuje równoległe zespoły agentów przy zadaniach z co najmniej dwoma niezależnymi strumieniami pracy, od tworzenia do sprzątania. - 85agent-orchestrator
Planuje i koordynuje autoryzowaną pracę wielu agentów z zależnościami, właścicielami zasobów i sprawdzeniami akceptacji. - 100native-agent-swarms
Koordynuje małe zespoły wyspecjalizowanych agentów Codex z ograniczoną równoległością, jawną własnością, natywnym przesyłaniem wiadomości i syntezą opartą na dowodach. - 90coding
Deleguje pracę kodową rolom worker, explorer, advisor i reviewer, a bieżący agent koordynuje i weryfikuje wynik.