메뉴
Verify phase — evidence-based validation. Detect language, run the full check suite, dispatch the sentinel against the plan's acceptance criteria, and produce a pass/fail scorecard where every PASS cites fresh command output from this session.
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
Choose how to handle a new coding request — native for tiny edits vs a fakoli-claw flow for multi-file, risky, design, or explicit crew/Fakoli work.
Design phase — refine an idea into an approved spec through structured one-question-at-a-time dialogue, propose 2-3 approaches, present the design section by section, write the spec, and hand off to flow:plan. Works headless over any channel.
Execute phase — wave-based crew dispatch on OpenClaw with critic gates and evidence-based verification. Load an intent-driven plan, group tasks into dependency waves, sessions_spawn specialists in parallel per wave (local SGLang tier), run a critic gate after every code wave, then a sentinel evidence gate.
Ship phase — re-run fresh tests, then present exactly four options (merge locally, push+PR, keep as-is, discard) and execute only the explicit choice. Never auto-merge, auto-push, or discard without the typed word discard.
Plan phase — break an approved spec into an intent-driven task list for OpenClaw crew execution. Verify assumptions with a scout sub-agent, map files to owners, write WHAT-not-HOW tasks with acceptance criteria and Depends on, then hand off to flow:execute.
Fast path — skip brainstorm/plan/waves for small tasks under 3 files. Estimate scope, dispatch a single specialist (welder by default), run verification, run a one-cycle critic gate, done. Escalates to flow:execute if scope or fixes exceed the limit.
| name | flow-verify |
| description | Verify phase — evidence-based validation. Detect language, run the full check suite, dispatch the sentinel against the plan's acceptance criteria, and produce a pass/fail scorecard where every PASS cites fresh command output from this session. |
Verification is not an opinion. It is a command you ran, output you read, and a result you can cite. Every PASS cites fresh output from this session; every FAIL cites what the output showed.
Invoked automatically after flow:execute, after flow:quick, or manually.
tsconfig.json/package.json → TypeScript; Cargo.toml → Rust;
pyproject.toml/setup.py → Python. Prefer the most specific. If none, ask.npx tsc --noEmit && bun testruff check . && mypy . && pytestcargo check && cargo test
Capture full output, read exit codes, count errors explicitly.sessions_spawn(agentId="fakoli-sentinel", task=…)
with the acceptance criteria + exact verify command per task from the most recent plan
(ls docs/plans | sort | tail -1; if several for today, ask which). Derive a verify- run id
and inject the absolute status path; sessions_yield. Quick-mode (no plan) → verify modified
files against the user's original task description; run id verify-quick-<YYYYMMDDHHmm>.
If crew absent, run the criteria checks yourself with the same evidence gate.[PASS] … Evidence: <cmd> → <output> or [FAIL] … Evidence: ….
End: Result: N/M criteria PASS — READY TO SHIP or — NOT READY TO SHIP.A non-zero exit, any typecheck error line, any failed test, a criterion with no verifiable output, or a missing plan file outside quick mode.