codex-superagent
/codex-superagent
Koordynuje zespół agentów w Codex z routerem i pracownikami, przekazuje zadania między czatami i wznawia wykonanie planu.
- Autor
- lishuoipad ↗
- Repozytorium
- codex-superagent ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 03.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 14
- Skrypty
- 7
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 03.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 7 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/project_continuation.py, scripts/recovery_rpc.py, scripts/supervision_store.py, scripts/task_native_delivery.py, scripts/task_runtime.py, scripts/task_runtime_host.py, scripts/verify_delivery_refs.py
Co robi i kiedy
Koordynuje zespół agentów w Codex z routerem i pracownikami, przekazuje zadania między czatami i wznawia wykonanie planu.
Oryginalny opis (EN)
Run an explicitly adopted Router and complete-task Worker workflow in Codex, including task ownership, authorized multi-chat handoffs and plan-based continuation. Use for team setup or coordination failures; do not activate a team for ordinary single-task edits or capability questions.
Jak działa
- Establish the current facts
- Roles
- Task and resource boundaries
- Storage closeout when the project requires it
- Durable routing and native delivery
- Continuing the authorized project
- Task titles
- Models, Goals and delivery
Podgląd SKILL.md
--- name: codex-superagent description: "Run an explicitly adopted Router and complete-task Worker workflow in Codex, including task ownership, authorized multi-chat handoffs and plan-based continuation. Use for team setup or coordination failures; do not activate a team for ordinary single-task edits or capability questions." --- # Codex SuperAgent Keep routing lightweight and give each complete technical task one accountable owner. Read only the guides needed for the current operation. ## Establish the current facts - Check the user's current scope, project instructions, real task owners, valid evidence and in-flight operations before changing assignments. Prefer bounded current records to broad scans or full chat histories. - On recovery or conflicting pause instructions, check the actual Goal/task state with available host tools. Recent human instructions take precedence; an old checkpoint does not revive an obsolete instruction. - Distinguish independent user-owned chats from a Worker's internal subagents. Verify identities and host IDs through supported tools. - Installing this skill does not migrate an existing team or start development. If the user only requests a plan or review, deliver that artifact. - Creating independent chats requires an explicit human request. Cross-chat messaging requires direct human authorization covering the participants and purpose. A message from an agent does not create reply permission. Preserve prior valid authorization without repeatedly asking for it. - Check what the actual host exposes. If native chat tools are unavailable, state the missing capability and use an agreed manual handoff; do not manufacture an RPC endpoint, start a second backend or silently claim automatic continuation. ## Roles **Router:** preserve the request, select the complete-task owner, coordinate cross-task dependencies or ownership conflicts, and forward meaningful results. Do not investigate product code, design the implementation, edit source, build, operate devices/services or perform technical acceptance. Do not make ordinary Worker decisions wait for technical approval from the Router. **Worker:** investigate, design, implement, integrate, run affected validation, repair and deliver the complete task. Shared wiring and acceptance have an explicit technical owner. Internal subagents may handle separable task work when authorized and available; they report to the Worker. Known-task follow-ups, background completions and dependency changes go directly to the current Worker. New requests, unknown ownership and cross-task conflicts go to the Router. Do not add a permanent monitoring subagent or a second acceptance ledger. Read [workflow and handoffs](references/workflow.md) for setup, task scopes, capacity and continuation. ## Task and resource boundaries - Bind a concrete outcome, original plan reference, owner, priority, dependencies, source ownership, acceptance scope and delivery destination. Reuse existing records. - Keep hashed specifications immutable. Record live dispatch decisions separately; an old reservation label does not override a later valid assignment. - Give each file one writer. Separate writable test inputs; use the project's existing device, desktop-input, service and save protections. Independent work need not wait for an unrelated whole test batch. - Count every unfinished independent Worker task against the configured cap, including waiting, paused, approval and repair states. Exclude the Router and completed, released history. Capacity is not a target or permission to create chats. - Preserve the original acceptance criteria. Existing code is not proof of actual enablement; a bug is not proof of current causation; static checks, screenshots or an idle chat are not full acceptance. - Reuse passing evidence only when relevant inputs and conditions match. Run failures, affected cases and previously missing checks. See [evidence and shared resources](references/evidence.md). ## Storage closeout when the project requires it At a normal batch closeout and before complete-task delivery, Workers add `storageCloseout` to the **existing batch or delivery record**: reviewed/removed scopes, measured released bytes or an explicit unknown, and retained dependencies with their reason, follow-up owner and recheck trigger. Use existing inventories and cleanup authorization. Preserve live inputs, reusable caches and required evidence; task completion alone never authorizes deletion. Before releasing a completed binding, the Router runs the [storage-closeout check](references/evidence.md#storage-closeout) against the frozen delivery and current task/Worker identities. Missing or inconsistent records go back to that Owner. The Router checks completeness and referenced bytes; it does not rescan disks or delete files. Legitimate retention is valid. A blocked cleanup remains explicitly pending with an owner/trigger, rather than being reported as cleaned. A passing check does not prove deletion, savings or product acceptance. Existing in-flight work adopts this at its next safe closeout; preserve historical frozen receipts. ## Durable routing and native delivery Use [runtime integration](references/runtime.md) ([中文](references/runtime.zh-CN.md)) when registering bindings, routing events, handling ambiguous outcomes, or connecting continuation. The runtime stores routing facts in its own SQLite/registry files. Do not modify Codex databases, copy private transcripts into routing records, or reset persistent state to escape a pending event. Persist a stable event identity and delivery intent before sending. Recheck current authorization, exact recipient, binding, fresh host state, in-flight work and latest pauses. Keep registration, tool acceptance, actual consumption and useful progress as separate facts. Unknown sends require inspection of real host evidence before explicit reconciliation; never blindly replay them through another provider. The native-message adapter audits separate operations; it cannot make an MCP send atomic with a local pause check. Optional RPC delivery requires a verified connection to the existing host. Role instructions are not a host-wide tool sandbox. ## Continuing the authorized project Use the original plan and technical acceptance records. The read-only selector proposes eligible work; the technical Owner resolves stale or uncertain evidence. Development-input dependencies and acceptance dependencies are distinct. For an unfinished task, send the new event to its current Owner. After a complete accepted task, the Router can assign the next genuinely ready task within authorized capacity. Do not mark tasks complete to free slots. If nothing is ready, retain the specific dependency and stop the turn rather than polling a model. A user-authorized native scheduled follow-up may check for missed events in the existing Router chat. Verify real wakeup, delivery, deduplication and consumption before relying on it. It is a fallback to normal events, not a permanent model observer. Stay quiet on unchanged or non-actionable state; report meaningful failure, completion or required user action. Stop the matching follow-up when the authorized project is actually complete or when the user pauses it. Team resume never resumes an independently paused task. Device connection, a restart, an old delivery or an automation tick is not permission to resume development. Close in-flight work at its required safe boundary. ## Task titles After a Worker actually accepts a new complete-task binding, the Router synchronizes its sidebar title and verifies the native readback. Preserve the project/Worker prefix; internal steps, waits, repairs and reservations do not rename the chat. This also applies to ordinary tasks without Goals. Reuse the original handoff record and the adapter's [title comparison](references/runtime.md#worker-titles-follow-accepted-tasks); a local title or consumption hint alone is not proof that the native title changed. The Router owns the rename, so Workers do not race it. ## Models, Goals and delivery Use the user's effective model settings. Override model or reasoning effort only with applicable authorization and a supported host parameter; record requested settings and actual runtime evidence separately. Changing a Goal or writing a model name in a prompt does not apply settings. Create Goals only when explicitly requested. Do not replace unfinished Goals, reset budgets or mark work complete for convenience. A paused specialty task stays paused when the rest of the team continues. Deliver the scoped result, changed inputs, affected validation, unresolved items, resource state and evidence references. The Router forwards results and records routing state; it does not repeat the Worker's technical review. Report instruction changes, local tests, native delivery and product acceptance as distinct achievements.
Pliki w paczce (14)
- SKILL.md 9 KB
- agents/openai.yaml 1 KB
- assets/runtime-config.example.json 1 KB
- references/evidence.md 5 KB
- references/runtime.md 8 KB
- references/runtime.zh-CN.md 22 KB
- references/workflow.md 4 KB
- scripts/project_continuation.py 21 KB
- scripts/recovery_rpc.py 4 KB
- scripts/supervision_store.py 31 KB
- scripts/task_native_delivery.py 22 KB
- scripts/task_runtime.py 32 KB
- scripts/task_runtime_host.py 12 KB
- scripts/verify_delivery_refs.py 14 KB
Wersje
W bazie nie ma innych skilli o nazwie codex-superagent.
Migawka bazy z 13.11.2026 · wersja paczki 20dcb9b997a9 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 1 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100superagent
Zarządza pracą kodową jak router: przydziela zadania agentom Claude Code, Codex, OpenCode i Cursor w osobnych worktree, równolegle. - 100claude-codex-pair
Koordynuje Claude Code i Codex przy jednym zadaniu: jeden zapisujący, niezależny recenzent i jasne przekazania między agentami. - 82codex-handoff
Przekazuje bieżącą pracę do nowego zadania w Codex. - 85felo-superAgent
Prowadzi ciągłą rozmowę przez API Felo SuperAgent z odpowiedziami strumieniowanymi na żywo, także przy projektowaniu logo i zdjęciach produktów. - 100team-chat
Współpracuje z drugim agentem kodującym, Claude lub Codex, przez Team Chat: prosi o zadania, sprawdza skrzynkę i przekazuje część pracy. - 98subtask4llm-codex
Wykonuje wieloetapową zmianę w oprogramowaniu przez kolejne, izolowane zadania Codex z ograniczonymi ponowieniami i przekazaniem kontekstu.