菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | flow-quick |
| description | 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. |
One agent, one pass, critic gate, done. For changes too small to justify the full pipeline.
Invocation: flow:quick "<task>" (inline task, no spec, no plan, no waves).
flow:brainstorm (offer --force to override). Under 3 files → continue.tsconfig.json/Cargo.toml/pyproject.toml).sessions_spawn: welder by default (code/bug/param), guido for
design/naming, scout for research. Packet = task + scope files + language + "keep it minimal,
don't refactor unrelated code". If crew absent, do it directly with the same scope constraint.npx tsc --noEmit && bun test; Py ruff check . && mypy . && pytest; Rust
cargo check && cargo test. Read full output + exit code.sessions_spawn(agentId="fakoli-critic", …) on the modified files → PASS /
SHOULD FIX / MUST FIX.flow:execute.Quick mode has no spec, no plan, no sentinel — intentionally limited. If a "small" task turns out
to need a design decision, stop and route to flow:brainstorm.
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.
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.