Skill: Model-Pin Gate

Every non-Cursor spawn pins its model explicitly (canon #5); a non-default model such as Fable is used only by explicit per-invocation selection, never inherited. Routing: see /agent-routing § Routing rules (SSOT).

What It Is

model-pin-gate is a Claude Code PreToolUse hook for Task|Agent|Workflow and, with the prepared settings matcher, cmux spawn_agent, dispatch_to_agent, and repoGolem-style new_split launches. It reads the current session's recent assistant model from transcript_path with a bounded tail read. It also honors explicit apex-seat identifiers (orchestrator, orc, skillcreatorLead) when a caller deliberately targets Fable.

The hook enforces Etan's fleet model policy:

The Workflow path is intentionally heuristic and advisory-only. It counts agent( occurrences versus model: occurrences and does not attempt full JavaScript parsing inside a 500ms hook.

Hook Contract

Input is the Claude Code PreToolUse stdin payload:

{
  "hook_event_name": "PreToolUse",
  "transcript_path": "/path/to/transcript.jsonl",
  "tool_name": "Agent",
  "seat_id": "skillcreatorWorker",
  "tool_input": { "model": "opus", "prompt": "..." }
}

Output schema:

• allow: {}
• block: {"decision":"block","reason":"..."}
• advisory: {"systemMessage":"..."}

The hook is local-only: no network, no BrainLayer, no subprocesses, bounded stdin/transcript reads, and fail-open on malformed input or internal errors. Caught failures emit a visible advisory that the check could not run; the action is still allowed and exception details are omitted.

install-snippet.json pins the absolute Node path: $HOME/.nvm/versions/node/v22.22.0/bin/node.

Run It

bun test skills/golem-powers/model-pin-gate/evals/model-pin-gate.test.mjs
python3 skills/golem-powers/model-pin-gate/evals/run_suite.py

Programmatic: import { detectModelPin } from "./src/model-pin-gate.mjs" → detectModelPin(payload) returns { verdict, seatModel, tool, violations, advisories }.

Stated Limits

• Unpinned inheritance detection depends on a recent assistant model field in the transcript. If the model is absent or unreadable and no explicit Fable target is present, the gate allows the call.
• Explicit Fable targets are allowed only when the payload carries an apex seat identifier. Unknown seat identifiers are treated as non-apex for explicit Fable targets.
• Workflow checks are advisory-only by design. They are a cheap reminder to pin each agent() call and avoid brittle parsing in a hot hook path.