AGENT SKILL

orchestrate-work

Coordinate or resume repository delivery with the selected model and the host's native agents.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
orchestrate-work
description
Coordinate or resume repository delivery with the selected model and the host's native agents.

03 SKILL.md 正文

Orchestrate work

Optimize rápido → bien → barato: time to the complete usable result, including corrections. The selected conversation model owns orchestration. Use this as the execution method; reuse settled planning and applicable repo rules instead of restarting a lifecycle or loading a second coordination stack.

Establish the current target

Recover the requested outcome, latest corrections, authorized endpoint, exact working root, branch and current changes. Verify a linked PR against Git; thread titles and sidebar associations can be stale. Preserve existing work. Make done observable: required behavior, relevant validation, material findings resolved, and the authorized delivery endpoint reached. Continue those steps without stopping for approval of an unrequested first-draft checkpoint. For a broad assignment, keep a short requirement → implementation → verification record. Examples emphasized in a follow-up do not silently replace its scope.

Read only the references that decide this task. For visual implementation, open the approved images and design decisions before delegating; a URL nobody can retrieve is not a usable design handoff. Resolve missing assets early while continuing independent work.

Use the host's agents

Choose currently available models and tools within the user's provider, capability and quota constraints. A recommended roster is optional guidance, not a replacement parent or evidence of runtime identity. No particular model, paid plan or OpenAI access is required by this skill. Never silently substitute an excluded provider. Match vision/browser assignments to actual tool access.

Delegate when it shortens the critical path or supplies needed independent judgment. In OpenCode, use native task subagents: general for writable work, explore for a bounded code question, and scout for external research when available. Other hosts use their equivalent native tools. Do not build another agent launcher or launch nested CLIs merely to reproduce these capabilities. If delegation is unavailable, report that and continue useful work directly.

Give each worker a compact packet:

  • Exact root and relevant revision; owned files or surface.
  • Required behavior, existing contracts, and necessary source/asset pointers.
  • Constraints, focused checks, expected output, and any known blocker.
  • Routing and planning are settled; execute this packet, preserve others' work, and report changed files, check results and remaining uncertainty.

Start independent packets together where the host supports it. Keep one writer per surface; the parent owns shared contracts, integration and conflicts. Receive completion events rather than repeatedly polling or requesting status. Reuse a worker for a focused correction when its context remains useful.

Specify outcomes more strongly than implementation guesses. Check consequential provider/API limits before prescribing an approach. A worker should challenge a conflicting plan with concrete evidence, rather than reproduce its false assumptions in code and permissive mocks.

Integrate, verify, finish

Inspect the integrated diff against the whole requested behavior. Choose the smallest adequate existing checks and required repository gates. Add a test only when it catches a named failure and helps correct it. Preserve check exit codes; output filtering must not conceal failure. Repeat checks only for a relevant change, failure or unresolved concern. A mocked service does not prove the real provider accepts the operation.

For UI work, verify the affected interaction with useful data and the intended role. Check retained selection, edit/cancel behavior and permission transitions when the change affects them. Use isolated fixtures; clean up authorized test mutations without overwriting concurrent user changes.

When visual fidelity is part of acceptance, use the approved direction and the available Impeccable/design skills selectively. Have an independent agent with vision critique the rendered result; correct material findings before final evidence. If that capability is missing, state the acceptance gap. A screenshot alone is not a critique, and a text-only model must not claim to have seen it.

Capture final evidence on the validated candidate, after data and layout settle. Use actual mobile dimensions for mobile claims. Stable capture steps can run as a script; they need no model inference per click or frame. Share only curated media and reports, never auth state, secrets or entire session folders.

Finish at the authorized endpoint. Distinguish implemented code, verified behavior, PR, Preview and production. Report the exact candidate, working links, checks and remaining gaps. For a separate review, leave a short handoff with goal, root, revision, diff, verification receipts and known limitations; do not export the whole conversation by default.

04 同仓库更多 Skills