Agent Factory Blueprint

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{ "type": "command",
        "command": "npx tsx ~/.claude/hooks/fmt.ts" }]
    }],
    "Stop": [{
      "hooks": [{ "type": "command",
        "command": "npx tsx ~/.claude/hooks/done-gate.ts" }]
    }]
  }
}

import { execSync } from "node:child_process";
import { existsSync } from "node:fs";

const sh = (c: string) => execSync(c, { encoding: "utf8" });
const block = (reason: string) => {
  console.log(JSON.stringify({ decision: "block", reason: reason.slice(-9000) }));
  process.exit(0);
};

if (!sh("git diff --name-only HEAD").trim()) process.exit(0);

for (const c of ["pnpm typecheck", "pnpm lint --max-warnings 0",
  "pnpm knip", "pnpm jscpd:changed", "pnpm test:affected"]) {
  try { sh(c); } catch (e: any) { block(`${c} failed:\n${e.stdout}`); }
}

const sha = sh("git diff HEAD | git hash-object --stdin").trim();
if (!existsSync(`.claude/evidence/${sha}.md`))
  block("Run /simplify, /code-review --fix, /verify (+ design review if UI). " +
        `Write proof to .claude/evidence/${sha}.md`);

Gotchas: exit code 1 does not block (print the JSON block shown). A mistyped hook path silently disables the gate. Claude Code overrides after 8 consecutive blocks unless CLAUDE_CODE_STOP_HOOK_BLOCK_CAP is raised. Hook output caps at 10,000 characters, so trim logs. Starting points to fork rather than write: Every's compound-engineering plugin (/lfg runs plan → build → simplify → review), obra's Superpowers verification skill, OneRedOak's design-review agent.

1. Stop gates in the session: types, lint, knip, jscpd, affected tests, simplify, review, verify.
2. Locked tests written by QA from the PRD before code. Builders mostly cheat by editing tests; read-only is the proven fix.
3. Mutation check on changed lines, so tests must catch planted bugs.
4. Hidden scenarios run by a browser on the PR preview, judged by the other vendor.
5. Security: SAST, secrets scan, new-dependency policy.
6. Verdict in code: AUTO_MERGE, FIX or NEEDS_HUMAN. AI findings are inputs, never the gate.
7. Canary with auto-rollback on error rate or SLO burn.

Run the verdict in shadow for a week before trusting it: it logs what it would do, you compare. Go live at ≥90% agreement.

Faster to start, less control: Claude Code Projects (beta; cloud threads skip repo skills and settings) or Cursor Automations (best triggers; second vendor's harness).

Skip Fable for coding: 2.5x Opus's price, lower agentic score. Fast modes cut latency, not throughput; use them only when you're blocked.