socketdev delegating-execution
The fleet tier doctrine: Fable plans, a lesser model executes, Fable reviews, the lesser model follows up. This skill is the EXECUTION-tiering complement to grilling-plan (which stress-tests a plan's content interactively before building) and operates at the Workflow-harness tier (agent({model, effort})). The CLI-subprocess layer — codex/opencode/kimi backends and when to prefer each — is documented in .claude/skills/fleet/_shared/multi-agent-backends.md; this skill sits above it on the Anthropic Workflow tier.
When to use
- Non-trivial build/design work: multi-file changes, new enforcement surfaces, cross-repo cascades.
- Work whose plan is a deliverable (
plan-review-nudgeshape: numbered steps, named files + rules). - NOT for one-file fixes — just do them.
- NOT for plan critique — that is
grilling-plan; run it before this skill when the design is unsettled.
The cycle
- Plan — big brain writes the numbered plan + a fenced execution prompt to
.claude/plans/. - Execute — floor model follows the plan verbatim in a
git worktree. - Review — big brain diffs the result against the plan; emits severity +
file:linefindings. - Follow-up — floor model applies each finding, re-runs gates, commits.
Sensitivity routing
Mirror — source of truth: scripts/fleet/lib/delegating-execution/route.mts.
| phase | sensitivity | model | effort |
|---|---|---|---|
| plan | benign | claude-fable-5 |
undefined |
| plan | security | claude-opus-4-8 |
high |
| review | benign | claude-fable-5 |
undefined |
| review | security | claude-opus-4-8 |
high |
| execute | benign | claude-sonnet-4-6 |
medium |
| execute | security | claude-sonnet-4-6 |
medium |
| followup | benign | claude-sonnet-4-6 |
medium |
| followup | security | claude-sonnet-4-6 |
medium |
- Benign infra/docs planning → Fable (apex tier; adaptive-only, no effort knob).
- Security-sensitive planning (vuln analysis, supply-chain, auth/token/secret handling, exploit-adjacent, hardening) → Opus 4.8 DIRECTLY: Fable's classifiers false-positive on benign security work and the refusal→Opus fallback in
spawnAiAgentis not live yet (upstream socket-lib; seefable-fallback). When unsure, saysecurity— the workflow also defaults to it (fail-safe).
How to invoke
Workflow({ name: "delegating-execution", args: { sensitivity: "benign" | "security", task: "<one-paragraph task summary>" } })
sensitivity is optional and defaults to "security".
Output contract
- The plan phase writes
.claude/plans/delegating-<slug>.mdcontaining (a) the numbered plan naming files + rules and (b) a fenced execution prompt the floor model runs verbatim. - Review findings and follow-up receipts are appended to the same doc.
- Plans never land on a committable path (
plan-location-guard).
Backing code
scripts/fleet/lib/delegating-execution/route.mts— the canonical tier table.scripts/fleet/lib/delegating-execution/prompts.mts— phase prompt builders (source of truth; the workflow JS carries a mirror)..claude/workflows/delegating-execution.js— the Workflow runner.