dispatching-agent-work
Routes a user's request to a separate task, subagent, automation, or current agent with an appropriate model role, reasoning effort, and workspace isolation. Use when the user or an authoritative instruction asks to dispatch, delegate, orchestrate, enter dispatch mode, run work in another session, select an execution backend, or schedule work. While dispatch mode is active, proactively send substantive execution to the host's preferred worker boundary: a new user-visible task in Codex Desktop App and a subagent on hosts without durable tasks. Do not invoke merely because ordinary work is complex, long-running, or parallelizable, or when this skill is only being mentioned, reviewed, or edited.
abagames · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add abagames/agentic-gamedev-skills --skill dispatching-agent-work技能说明
name: dispatching-agent-work description: "Routes a user's request to a separate task, subagent, automation, or current agent with an appropriate model role, reasoning effort, and workspace isolation. Use when the user or an authoritative instruction asks to dispatch, delegate, orchestrate, enter dispatch mode, run work in another session, select an execution backend, or schedule work. While dispatch mode is active, proactively send substantive execution to the host's preferred worker boundary: a new user-visible task in Codex Desktop App and a subagent on hosts without durable tasks. Do not invoke merely because ordinary work is complex, long-running, or parallelizable, or when this skill is only being mentioned, reviewed, or edited."
Dispatching Agent Work
Route work without coupling the decision to one vendor's model names or agent APIs. When dispatch mode is active, keep the current conversation as the coordinator and prefer a separate worker for substantive execution. Otherwise prefer the smallest execution unit and lowest sufficient model role that can meet the success criteria reliably.
Workflow
0. Confirm activation and duration
Proceed only when the current request or an authoritative host or repository instruction explicitly asks for dispatch or execution-boundary selection, or when a previously and explicitly established dispatch mode remains active in the conversation. Mentioning, reviewing, or editing this skill does not activate dispatch. Complexity, duration, cost, or possible parallelism alone does not activate it. When the current request provides no activation and no established dispatch mode is active, continue with the current agent under normal host policy.
Distinguish one-request activation from dispatch mode. A request to dispatch or delegate one objective applies only to that objective. A request such as "enter dispatch mode," "act as an orchestrator," or an authoritative instruction establishing that role activates dispatch mode for later objectives in the same conversation. Keep that mode active until the user asks to leave it or an authoritative instruction changes the role. Do not mistake a request to inspect, review, or edit this skill for activation.
In dispatch mode, treat the current agent primarily as a coordinator. Proactively dispatch every substantive objective that has a self-contained handoff and success criterion; the user does not need to repeat "dispatch" on each turn. Keep only clarification, routing, coordination, synthesis, and genuinely trivial one-step actions in the current conversation. Do not create an empty worker merely to await an objective: collect enough task content for a lean handoff first.
Activation permits routing analysis but does not override host policy or broaden authority. Before considering a boundary, remove any backend that the host forbids or that requires authorization not present in the current request.
1. Capture the dispatch contract
Extract only:
- objective and success evidence;
- answer, diagnose, change, monitor, or schedule intent;
- files, systems, and people in scope;
- approval and destructive-action boundaries;
- persistence, follow-up, deadline, and isolation needs;
- any explicitly requested model, backend, or skill.
Do not broaden authority while delegating. Preserve an explicitly requested model or execution boundary when the host supports it.
2. Discover host capabilities
Inspect available tools and policies instead of assuming that the host supports tasks, subagents, worktrees, model overrides, waits, or approval forwarding. Read execution-backends.md when more than one backend is available or a fallback is required.
When the destination is the Codex app and a separate user-visible task may match, read codex-app.md before dispatching.
3. Choose the execution boundary
Classify the instruction before selecting a boundary. Reuse a worker only when the instruction is affirmatively a continuation of that worker's objective and has the same artifact surface, authority, lifecycle, required model role, and required reasoning effort. Assess the required model role and effort fresh for the instruction before comparing them with the worker. If any dimension differs, is unknown, or has not been checked, treat the instruction as a new objective. A shared project, repository, conversation, chronology, or available worker is not continuity.
When dispatch mode is active, first select the host-preferred worker boundary:
- Codex Desktop App: create a new durable user-visible task for each new substantive objective. Prefer the current project and an isolated worktree for repository writes when host policy supports it. Use a new task even when the parent will later synthesize the result; task visibility is the mode's default, not a special case requiring the user to ask again.
- Other hosts with subagents but no durable tasks: create a subagent for each bounded substantive objective and return or synthesize its result in the parent.
- Hosts with neither: perform the work in the current agent and report the capability limit.
- Scheduled or recurring work: use an automation or queued job instead of either default.
Outside dispatch mode, or after classifying a continuation or exception within it, use the first matching boundary:
- Existing worker — the continuation gate above passes and the user did not request a new boundary. Send a concise delta, not the original prompt again.
- Automation or queued job — work must run later or recur.
- Durable task — the user requested a separate user-visible task, needs direct follow-up or approvals, or the work has an independent lifecycle.
- Subagent — a bounded independent subproblem whose result should return to the parent for synthesis.
- Current agent — clarification, routing, synthesis, or one trivial action for which creatin