메뉴
Dispatch a gate-approved GLM-5.2 lane through delegation-glm. Provisional lanes require an explicit --allow-provisional decision. Never use it before the evaluation gate or as a silent fallback.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Dispatch a provisional Kimi K3 lane through delegation-kimi after an explicit decision. Never use it before the evaluation gate or as a silent fallback.
How and when to delegate coding work across models — pick the right lane (cheap execution, senior review, explicit judgement, or dual super-judgement), size the reviewer to the work, and know when a manual gate is required. Use when deciding whether to spawn a subagent, which model/effort to hand a task, how to route a code review, or when to escalate. Also covers reaching Codex from Claude.
Run a task too big for one pass as a lead-executor-advisor loop — plan, fan the work out across cheap parallel workers, verify each result against its own criteria, and have the expensive model judge the plan up front and the ship at the end. Use when the user says "too big for one pass", "fan this out", "orchestrate across a model team", "run the advisor-worker loop", or asks to parallelize research/generation across many subtasks. NOT for single-file edits, work one pass handles, or when you are already inside a Workflow/ultra run (that is the orchestrator already — no double fan-out).
| name | glm-executor |
| description | Dispatch a gate-approved GLM-5.2 lane through delegation-glm. Provisional lanes require an explicit --allow-provisional decision. Never use it before the evaluation gate or as a silent fallback. |
GLM-5.2 is an optional external executor, not a native Claude or Codex model.
Before dispatch, run delegation-glm check --json. The current gate exposes
clerk and scout only in provisional_lanes; use them only after an explicit
decision and pass --allow-provisional. Builder is a blocked candidate.
The small local run is compatibility evidence, not enough for full qualification.
Builder and reviewer are not dispatchable. The installed routing JSON remains
authoritative if a later versioned evaluation changes that set.
delegation-evidence lane builder shows the dated external rows: they provide
context, not a local harness score and not permission to widen the gate.
Write the self-contained worker brief to a file, then run:
delegation-glm run --lane <clerk|scout> --effort auto --allow-provisional \
--backend auto --prompt-file "$brief" --output "$result" --workdir "$repo"
auto and claude-zai both resolve to the isolated Claude→Z.AI backend, which
needs a key: ZAI_API_KEY in the environment, else the 600-mode key the
installer stored. Keep --effort auto; an explicit effort the gate did not pin
is refused (78), and only an --evaluation run may measure a new combination.
Exit 69 means unavailable; exit 78 means the lane or effort did not pass
evaluation. In either case, route deliberately to a documented incumbent rather
than pretending GLM ran. Treat all returned output as unverified and exercise the
deliverable before accepting it.