Przejdź do treści
Biblioteka skilli
AI i agenci · Automatyzacja

codex-agent-relay

/codex-agent-relay

Przekazuje niedokończone zadania wykonawcze innemu agentowi, gdy limit użycia Codex jest niski.

Licencja
MIT
Ostatnia zmiana repozytorium
30.10.2026

Ocena i bezpieczeństwo

90/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść25 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
2
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 30.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Przekazuje niedokończone zadania wykonawcze innemu agentowi, gdy limit użycia Codex jest niski.

Oryginalny opis (EN)

Preserve and hand off unfinished execution tasks to a project-configured Agent when Codex usage quota becomes low. Use when a configured project hook requests relay registration or quota handoff, or the user requests Agent Relay setup or status. Do not delegate casual questions, completed work, Plan mode, or work awaiting user decisions.

Jak działa

  1. Track an opted-in execution task
  2. Hand off at the quota reminder
  3. Failures
Sekcje z nagłówków SKILL.md, w oryginale · 31 linii, 733 słów
Podgląd SKILL.md
---
name: codex-agent-relay
description: Preserve and hand off unfinished execution tasks to a project-configured Agent when Codex usage quota becomes low. Use when a configured project hook requests relay registration or quota handoff, or the user requests Agent Relay setup or status. Do not delegate casual questions, completed work, Plan mode, or work awaiting user decisions.
---

# Codex Agent Relay

This skill works with the bundled `agent_relay` MCP server and reviewed/trusted Hooks. Installing this Markdown alone does not create automatic monitoring. Read [configuration](../../docs/configuration.md) for setup or changing the project's target Agent/model.

For first-time setup or a model change, call `relay_setup` and open its settings page in the right browser panel. Let the user select the target; do not guess a model or substitute one.

## Track an opted-in execution task

- Use the hook-supplied `session_id` and cwd; never invent a Codex session ID. Register through `relay_register` only when the project is enabled and there is authorized execution work. The returned `id` is the task ID.
- Checkpoints include `goal`, `completed`, `remaining`, `decisions`, `files`, `tests`, `processes`, `permissions`, and `acceptance`. All fields except goal/permissions are string arrays. Capture useful verified evidence, not the whole conversation. Do not include credentials or unrelated personal/project context.
- Update `relay_checkpoint` after meaningful milestones and immediately before handoff. Record existing process IDs, remote job IDs, command/tool IDs and observed state so the receiver can inspect them without restarting them.
- Call `relay_finish` when acceptance criteria are met. Use `relay_task_state` for `waiting_user` or `paused` when appropriate. A later authorized continuation can set an undispatched task back to `active`; do not resume a user-paused task autonomously.
- Plan-mode tasks stay paused. Neither quota pressure nor this skill grants permission to execute a plan.

## Hand off at the quota reminder

1. Decide whether the original task is already complete. If so, mark it complete instead of handing off.
2. Finish current file operations and save a fresh checkpoint. If a tool is intentionally running an existing background job, inspect it first and record its identity/state. Only then pass its hook tool ID in `backgroundToolIds`; this removes it from the pending-operation barrier without stopping/restarting it.
3. Tell the user: `额度即将用尽,开始调用 <configured Agent>,模型:<configured model>。` Use the actual project target. This is an external Agent invocation, not a change to Codex's selected model.
4. Call `relay_handoff` with the task ID and concrete remaining instructions. Include the next step, necessary constraints and validation. Do not repeatedly retry uncertain submissions. The relay persists a job ID and returns the same job for duplicate calls.
5. Report what the returned state proves. `preparing` is not acceptance; acceptance requires a durable receipt plus provider/model, sandbox-mode and approval-policy attestation. `failed` means no success claim is justified. Display the returned error concisely. The bundled adapter dispatches for real into the Harness `workspace-write` sandbox; it refuses to start when Windows project-folder ACLs do not let the sandbox provision, when the route or sandbox cannot be attested, or when project approval was revoked. Do not retry, substitute another model, or describe the attempt as a successful handoff. If the receiver asks for a decision, the relay cannot approve on the user's behalf; the job needs your review and may be `waiting_user` or failed.
6. Open the returned `dashboardUrl` in the current Codex window's right browser panel using `open_in_codex` if available (browser target, right placement). Otherwise use the host's built-in browser. If neither capability exists, provide the local link and state that automatic panel opening is unavailable.
7. Stop executing the source task. The external process owns the remaining work. The progress page reads its events directly; do not keep polling with model turns to imitate live progress.

The page continues independently of Codex inference. A run ending and an Agent saying “complete” do not prove acceptance. When asked to review the outcome, inspect the actual files/tests before using `relay_verify` with concrete evidence. `relay_verify` records evidence only; it does not certify a pass. Requests to change Agent/model update project configuration for future jobs only.

## Failures

Unknown quota is not zero. If the quota is already exhausted, retain the checkpoint and explain that no confirmed handoff occurred. Do not fabricate a late Codex instruction or silently submit a stale checkpoint. Missing target models fail explicitly; do not substitute another model or loosen permissions. Handoff also fails closed when project approval was revoked by a `user.json` launch/adapter/scanner edit, when the local secret scanner is missing or reports a finding, when the Windows ACL preflight rejects the project folder, or when the route/sandbox cannot be attested. Explain the specific error; never bypass the scanner, edit ACLs, or approve configuration on the user's behalf. If hooks or MCP tools are unavailable, explain the missing installation capability instead of claiming automatic protection is active.
Pliki w paczce (2)
  • SKILL.md 5 KB
  • agents/openai.yaml 1 KB

Wersje

W bazie nie ma innych skilli o nazwie codex-agent-relay.

Migawka bazy z 13.11.2026 · wersja paczki f3399877e36a (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

  • 100relay-codex
    Pozwala innemu lokalnemu agentowi uzyskać jednorazową odpowiedź z Codex CLI, gdy zachowanie interaktywne jest akceptowalne lub kontrolowane.
  • 81relay-setup
    Konfiguruje Codex Agent Relay przy pierwszej instalacji i ustawieniach agentów, a zmianę modelu odbiorczego zostawia użytkownikowi.
  • 82codex-handoff
    Przekazuje bieżącą pracę do nowego zadania w Codex.
  • 95relay-work-order
    Przekazuje zlecenia innemu agentowi AI przez e-mail według protokołu tagów HANDOFF, ASK, FYI i DONE.
  • 100relay
    Przekazuje stan pracy nowemu agentowi przy około 70 proc. zapełnionego kontekstu, zamiast kompaktować rozmowę.
  • 95agent-relay
    Pozwala Codexowi koordynować wielu agentów przez Agent Relay: wiadomości peer-to-peer, przekazania i wspólny status.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.