Agent Delegate

Agent Delegate

Use this skill as the explicit external editful worker/session adapter. It runs Claude Fable/Opus, Codex GPT/GBT models or Fugu profiles, Cursor Composer, Grok, or Kimi K3 as a separate process, captures a durable receipt, and can resume the exact external session. It does not own the caller's decomposition, role choice, context choice, review policy, or final integration.

For ordinary same-host work, use the active host's native child directly when that child can satisfy the role. Choose this external lane when the provider, model/profile, lifecycle, process isolation, automation surface, or receipt provides a concrete benefit worth the extra process and integration cost. Those are recognition examples, not a finite allowlist or an approval gate.

Fresh-resumable is the external default: the worker starts clean from disk and the delegation brief, not from the parent chat, and the parent captures an exact session handle. A resume continues only that worker with a bounded delta.

The child may read files, edit files, run commands, verify its work, and use installed skills when they fit the task.

This is a prompt-engineering skill. It ships no scripts, shims, hook controllers, state machines, parsers, detached monitors, or install-time automation.

When to use

• "Use an external worker with a durable receipt to implement this."
• "Run two cross-provider workers on these independent fixes."
• "From this Codex host, delegate the refactor to Claude and report back."
• "From this Claude host, use Codex to fix the docs drift and run the checks."
• "Use Cursor Agent Composer to implement this slice and report back."
• "Use Grok Build to investigate and patch this workflow."
• "Use Kimi K3 max to implement this slice and keep its exact session handle."
• "Use the exact Fugu Ultra profile and preserve a resumable receipt."
• "Have an external Claude session use $skill-authoring to patch this skill package and preserve its exact handle."
• "Resume the same delegated Claude session and continue with higher effort."
• Another skill has already selected an external foreground worker because a concrete provider, model, lifecycle, isolation, automation, or receipt benefit matters.

When not to use

• The task is ordinary same-host implementation, investigation, or repair and the active host's native child can satisfy the role. Dispatch that child directly under ../_shared/agent-orchestration-policy.md instead of routing native work through this adapter.
• The user wants a clean read, second opinion, consistency check, completion audit, or readability/confusion check with no file edits. Use $fresh-consult.
• The user specifically asks for the existing Codex -p yolo review pattern. Use $codex-review-yolo.
• The user wants two models to iterate on a plan, architecture, design, or concept until they converge. Use $model-consensus.
• The work is an ordered subprocess workflow with manifests, critics, repair loops, or persistent orchestration. Use $stepwise or $arch-epic.
• The task needs a detached/background worker, a separate git worktree, or merge machinery. This skill is foreground and shared-worktree only.
• There is no concrete task, work root, success bar, or write scope for the child being launched.
• The requested runtime CLI is not installed.
• The caller wants to resume "the latest" child without an explicit session id or prior run directory and ambiguity cannot be resolved safely.

Non-Negotiables

• Apply ../_shared/agent-orchestration-policy.md before selecting this lane. Apply $prompt-authoring to the actual populated worker brief before launch and after a material reframe. Connect external transport to a concrete benefit, but do not turn that explanation into a permission checklist or ask for approval merely because the process is external.
• On a Codex host, account for the real cost of launching another Codex process: separate processes can contend on shared Codex SQLite/WAL state and have caused host-wide stalls. This is a tradeoff, not a ban, approval gate, or fixed process-count limit.
• Derive each delegated task and success bar from the user's request, a canonical plan or issue, local doctrine, or verified acceptance evidence. Resolve the work root, allowed write scope, constraints, delegation mode, and exact user-named inputs before launch. Keep the parent's diagnosis, preferred implementation, and selected evidence path advisory unless the authoritative source actually chose them.
• Runtime must be known. Effort must also be known except that GPT-6 Astra defaults an omitted effort to xhigh and Kimi K3 defaults it to max. Model or profile must be known except that a Codex lane with no named model defaults to gpt-6-astra and a Kimi lane defaults to kimi-code/k3. Codex runs GPT/GBT/OpenAI model ids and Fugu profiles, Claude Code runs supported Claude models, Cursor Agent runs composer-2.5-fast, Grok CLI naturally resolves to grok-4.6, and Kimi Code runs kimi-code/k3. Explicit legacy Grok ids remain exact and discovery-gated. If any other required value is missing or ambiguous, ask one consolidated question before invoking.
• Never run GPT/GBT model ids, Fugu profiles, or Claude models through Cursor Agent, Grok, or Kimi. Do not pass Grok or Kimi model ids through another runtime.
• Delegation mode is one of fresh-one-shot, fresh-resumable, or resume. Default to fresh-resumable. Use fresh-one-shot only when the caller explicitly asks for a stateless, ephemeral, no-resume, or throwaway worker. Kimi always persists sessions, so it cannot satisfy a load-bearing stateless/no-persist requirement even when the receipt would be ignored.
• Resume mode requires an explicit session id or a previous run directory with session_id.txt. Refuse missing, empty, UNRECOVERABLE, cross-runtime, or "latest session" resume requests.
• Treat model text as intent, not a fuzzy alias. Preserve exact family and numeric version; never silently substitute a nearby model.
• Run the child hook-suppressed where the runtime supports it, unsandboxed, and in the shared worktree per this repo's convention. Prompt boundaries define the task; the sandbox does not. A clean external session is a context choice, not filesystem, permission, or worktree isolation.
• Fresh one-shot runs may be stateless. Fresh-resumable runs must capture a session handle. Resume runs must use the same runtime as the captured handle: Claude resumes through Claude, Codex resumes through Codex, Cursor Agent resumes through Cursor Agent, Grok resumes through Grok, and Kimi resumes through Kimi with the exact captured id.
• For a single delegation, create one namespaced run directory under /tmp/agent-delegate/ and keep prompt.md, final.txt, events.jsonl, stderr.log, and execution.json there. For fresh-resumable and resume runs, also keep session_id.txt; for resume runs, also keep resume_from.txt.
• For explicit parallel delegations, create one group directory under /tmp/agent-delegate/ and one ordinary child run directory per worker. Do not add a controller, detached monitor, separate worktree, or merge layer. The parent owns the fanout, concurrency budget, scope separation, and final integration.
• Brief the child like a capable colleague walking in cold: include the authoritative outcome and success criteria, source evidence, work root, allowed write scope, real constraints, exact user-named inputs, the parent-owned external-topology boundary, and report contract. Let the child choose the evidence and implementation path inside that contract and report when the brief's premise conflicts with repo truth. For resume prompts, state the new instruction or evidence and what remains unchanged from the original delegation.
• Tell the child to read local instructions such as AGENTS.md before editing covered files.
• In parallel groups, tell each child it is not alone in the codebase, must not revert unfamiliar changes, should make the smallest task-relevant edits, and should report actual conflicts if they happen.
• Tell every worker it may fan out to its own native sub-agents on its host whenever that helps, with no assigned nested scope or budget, and that it must not spawn external agents: a worker started through this skill cannot itself use $agent-delegate or any other external transport to start another agent.
• Do not paste secrets into prompts. If a token is needed, source it into the child environment and tell the child which environment variable to read.
• Do not ask the child to commit, push, open PRs, rewrite history, stash, or revert unrelated work unless the delegated task explicitly requires that exact operation.
• Do not use external continuation controllers as a strategy. This skill is a foreground subprocess path, not an ordered workflow runner.
• If the child changed files or reports a blocker, inspect the repo state before presenting the result as fact. If the child claims a work-product artifact — a screenshot, report, spreadsheet, generated file, export — open the artifact itself and verify its substance matches the claim before relaying it: existence is not evidence, a screenshot showing a blank page, error, or wrong screen is a failed claim, and plausible spreadsheet cells are not checked math. Relay the child's analytical conclusions as verified only when their anchors were checked; otherwise present them explicitly as the child's claims.

First Move

1. Read ../_shared/agent-orchestration-policy.md.
2. Read references/model-and-invocation.md.
3. Read references/delegate-prompt-and-output.md.
4. Inspect the active host's native-child surface. If an ordinary same-host native child can satisfy the role, dispatch it directly instead of using this skill. Make that native context choice explicit: Codex uses fork_turns: "none", a positive bounded count, or "all"; Claude uses a clean named/custom subagent, an explicit full-conversation fork, or a context: fork isolated clean skill context as appropriate. Resume the exact subagent only when the same role continues; use a background agent only when work must outlive the foreground turn, and a team only when direct peer communication genuinely helps. Context inheritance remains separate from permissions and worktree sharing. Otherwise state the concrete benefit of the external lane.
5. Identify the delegated task or parallel delegated tasks, success bar, work root, exact user-named inputs, allowed write scope, constraints, and requested runtime/model/effort from the user's words. Apply $prompt-authoring to the populated brief: preserve source-backed authority, mark useful parent conjecture as advisory, and leave the worker free to reject it from repo evidence.
6. Identify the delegation mode. Use fresh-resumable unless the caller explicitly asks for a stateless one-shot worker or to resume a previous delegate.
7. When an external Codex lane omits its model, apply gpt-6-astra; when that Astra lane also omits effort, apply the preference default xhigh. For Kimi, apply kimi-code/k3 and the model-default max effort when omitted. If another required execution value, write scope, or resume handle is incomplete, ask one question that names exactly what is missing and what it controls.
8. Confirm the selected CLI exists with command -v codex, command -v claude, command -v agent, command -v grok, or command -v kimi.
9. Create the run directory or group directory and write each delegation prompt to its own prompt.md.
10. Invoke each worker with the exact command shape from the invocation reference.

Workflow

1. Confirm the external dispatch. Keep role and decomposition with the caller. Record why this worker is external, whether it is a new clean session or an exact resume, its capabilities, its write scope, and the parent-owned return contract.
2. Shape the delegation. State the source-backed outcome, allowed write scope, observable success bar, real constraints, verified evidence, and exact user-named inputs such as paths, failing commands, repro steps, or docs. Do not assume parent-chat inheritance or promote a parent diagnosis, preferred fix, or selected file list into task authority.
3. Resolve execution. Map the raw model phrase to runtime=<claude|codex|agent|grok|kimi>, model=<runnable id>, and effort=<level-or-encoded-in-model>. Announce the mapping before execution.
4. Select single or parallel. Use one worker by default. Use a parallel group only when the user asks for parallel workers or gives multiple delegated tasks and the parent can own every scope and result.
5. Select continuity. Use fresh-resumable by default. Use fresh-one-shot only for explicit stateless, ephemeral, no-resume, or throwaway delegation. Use resume only with an explicit same-runtime session id or prior run directory. Keep resume on the single-worker path.
6. Run the worker or workers. Use disabled hooks, no sandbox, a shared worktree, namespaced run directories, and live event capture. Fresh one-shot runs start cold; fresh-resumable runs start persistent child sessions; resume runs continue the captured session.
7. Monitor patiently. Normal delegated work often takes 5+ minutes; broad repo edits, verification, xhigh, or max can reasonably take 20-40 minutes. Poll live events.jsonl and stderr.log every few minutes, not every few seconds.
8. Consume the result. Read final.txt, locate the status footer, and inspect events.jsonl/stderr.log when the final output is missing or malformed. For fresh-resumable and resume runs, preserve the session handle for the next explicit resume.
9. Inspect local truth. Check git status and any changed files named by the child or children before reporting upstream. Open any claimed work-product artifact and verify its substance shows the claimed result — view the screenshot, check the spreadsheet's math, read the report — not merely that the file exists.
10. Report upstream. For one worker, lead with status, changed files, verification, blockers, confidence limits, and the run directory. For a parallel group, report one compact child-by-child table plus the parent repo state check and combined outcome.

Output Expectations

• A concise parent-facing report:
  • transport: external
  • concrete external benefit
  • starting context: clean prompt-and-disk context or the existing exact session context
  • continuation: new one-shot, new resumable session, or exact-session resume
  • delegation mode
  • runtime/model/effort used
  • delegated task status, or one status per child for parallel groups
  • changed files or none
  • work products claimed, with the parent's load-and-verify result for each, or none
  • skills the child says it used or none
  • verification run or not run: <reason>
  • blockers or none
  • follow-up needed or none
  • run directory path, or group directory plus child run directories
  • session id, or none only for explicit fresh-one-shot or failed capture
• If the child output is missing or malformed, say that plainly and preserve the run directory for debugging. Do not invent a status.
• If the child is wrong about a changed file, blocker, verification result, artifact's content, or a conclusion's anchors, say so explicitly and cite the evidence that contradicts it.

Reference Map

• references/model-and-invocation.md - exact external runtime/model/effort resolution, command shapes, session capture/resume, and receipts
• references/delegate-prompt-and-output.md - external-worker brief, status footer, parent integration report, and anti-patterns