Qiaomu Model CLI

用本机 Grok CLI、Kimi Code CLI 和 Claude Code CLI 做可验证的强模型调用。

Copyright (c) 向阳乔木

• X: https://x.com/vista8
• GitHub: https://github.com/joeseesun/

Defaults

Provider	CLI binary	Default model	Best for
xAI Grok	grok	grok-4.5	X/web 搜索、研究、生图、生视频、通用 agent
Moonshot Kimi	kimi	kimi-code/k3 (1M context)	前端 UI/CSS/组件、长上下文改码
Anthropic Claude	claude	fable -> Fable 5	最复杂编码、长任务、深度推理；可切 opus / sonnet

Always prefer these defaults unless the user names another model.

Router Rules

1. User says Grok / X 搜索 / 生图 / 生视频 / 实时信息 / 多工具 agent → use Grok CLI with grok-4.5.
2. User says Kimi / 写前端 / UI / CSS / React / Vue / 页面审美 / 长上下文改前端 → use Kimi CLI with kimi-code/k3.
3. User asks multiple models to independently produce/compare/benchmark the same task → use batch; start all independent jobs concurrently in isolated working directories.
4. User asks one model to research and another to implement from that result → use dependent dual; this is intentionally sequential.
5. User says Claude Code / Fable / Opus / Sonnet / 复杂编码 / 多文件重构 / 严格 code review → use Claude Code. Default fable; explicit alias wins.
6. Do not use OpenCLI for X if Grok native search/tools can do the job and the user wants Grok.
7. Prefer moving aliases (fable, opus, sonnet) over brittle dated Claude model IDs. If a default fails, run doctor and report the live CLI error.
8. Never print API keys, OAuth tokens, cookies, gateway URLs, or full auth/settings files.

Modes

• Doctor: check binaries, login, default models.
• Grok Chat/Agent: non-interactive Grok task with tool permissions.
• Grok X/Web: force native search/fetch tools for live info.
• Grok Image: image generation/edit via Grok native image tools / /imagine.
• Grok Video: video generation via /imagine-video or native video tools.
• Kimi Frontend: K3 1M for UI implementation with concrete files and verification notes.
• Claude Coding: Fable 5 by default for complex coding; Opus 4.8 or Sonnet 5 by explicit selection.
• Claude Review: read-only review prompt with findings first.
• Batch: independent Grok/Kimi/Claude jobs run concurrently with native stream events, bounded capture, private logs, and artifact Done Gates.
• Retry: rerun only failed jobs from a prior summary.json in a fresh run directory.
• Dual: dependent pipeline; Grok gathers facts/assets, then Kimi receives that result and implements.

Core Workflow

1. Classify intent: search/research, media, frontend, complex coding/review, independent batch, or dependent dual.
2. Run environment check when first used in a session or when a call fails:

python3 scripts/check_env.py

3. Choose provider and model:
   • Grok default: -m grok-4.5
   • Kimi default: -m kimi-code/k3
   • Claude default: --model fable (currently Fable 5); alternatives opus (Opus 4.8), sonnet (Sonnet 5)
4. Build a task prompt that tells the CLI to use native tools when relevant.
5. Draw the dependency graph before launching multiple models:
   • no job consumes another job's output → one batch call with max_workers >= job count
   • a later job consumes earlier output → dual or an explicit staged pipeline
6. For concurrent writers, give every job its own cwd; do not let models edit the same directory.
7. Invoke through the wrapper (preferred) or raw CLI.
8. For file-producing work, declare expects and optional verify_commands; exit code zero alone is not completion.
9. Keep progress visible: native provider events are normalized and default heartbeat is every 15 seconds.
10. On interruption, let the wrapper cancel every active process group. Resume with retry --summary, not by rerunning successful jobs.

Preferred wrapper

# Grok research / X search
python3 scripts/qiaomu_model_cli.py grok --task x-search --prompt "查 @vista8 最近3条帖子，给正文和链接"

# Grok image
python3 scripts/qiaomu_model_cli.py grok --task image --prompt "一张极简黑白海报，主题是 AI coding"

# Grok video
python3 scripts/qiaomu_model_cli.py grok --task video --prompt "一只猫在爵士酒吧弹钢琴"

# Kimi frontend
python3 scripts/qiaomu_model_cli.py kimi --task frontend --prompt "把这个 landing hero 改成更精致的杂志感布局" --cwd /path/to/project

# Claude complex coding (Fable 5 default)
python3 scripts/qiaomu_model_cli.py claude --task coding --prompt "完成这个多文件重构并跑测试" --cwd /path/to/project

# Claude with a lower-cost latest model
python3 scripts/qiaomu_model_cli.py claude --model sonnet --task coding --prompt "修复这个组件" --cwd /path/to/project

# Three independent implementations: concurrent fan-out
python3 scripts/qiaomu_model_cli.py batch \
  --config examples/batch-homepages.json \
  --max-workers 3 \
  --log-dir logs/model-run

# Retry only failed jobs from a prior run
python3 scripts/qiaomu_model_cli.py retry \
  --summary logs/model-run/summary.json \
  --log-dir logs/model-retry

# Dependent pipeline: Grok output feeds Kimi
python3 scripts/qiaomu_model_cli.py dual --prompt "研究 2026 AI coding CLI 趋势，并写一个单页展示站"

Raw CLI baselines

Grok:

grok -p "<PROMPT>" -m grok-4.5 \
  --permission-mode bypassPermissions \
  --output-format plain \
  --max-turns 12

Kimi:

kimi -p "<PROMPT>" -m kimi-code/k3 --output-format text

Claude Code:

claude -p "<PROMPT>" --model fable --effort xhigh \
  --output-format text --permission-mode bypassPermissions \
  --no-session-persistence

Notes:

• Single calls and batch jobs emit redacted JSON progress events to stderr immediately (started, heartbeat, activity, finished).
• Batch jobs default to each CLI's native JSON stream. High-frequency deltas are coalesced; Grok thought text is never copied into normalized events.
• batch writes raw stdout/stderr, normalized response text, events.jsonl, resumable state, and summary.json; one failure does not cancel independent jobs.
• Run directories use mode 0700; logs/config/state use 0600. In-memory stdout/stderr capture is bounded while full output continues to disk.
• batch is for independent work. dual stays sequential because stage 2 consumes stage 1 output.
• Never launch independent model calls one-by-one from the orchestrator and call that a comparison; create the config first, then make one batch invocation.
• Kimi --prompt cannot combine with --yolo or --auto in current CLI.
• Grok headless often needs --permission-mode bypassPermissions for unattended tool use.
• Claude aliases track the latest family release: fable, opus, sonnet. Fable may fall back to Opus for safeguarded topics.
• bypassPermissions allows project edits and commands. Use only in a trusted workspace; choose a stricter mode when unattended writes are not intended.
• For Grok image/video, instruct the model to use native tools (image_gen / image_edit / /imagine / /imagine-video) rather than writing fake image URLs.

Task Prompt Contracts

Grok X / live search

Require:

• use native web/X search tools
• no fabrication of posts, metrics, or dates
• return time, full/summary text, media flag, engagement if available, and links

Grok image

Require:

• use native image tools
• return local/remote asset paths actually produced
• if blocked by safety, stop and report

Grok video

Require:

• use native video generation path
• report shot plan + final asset path/URL if produced
• if unavailable, say unavailable with the CLI error

Kimi frontend

Require:

• read existing project style first
• implement real files, not only advice
• preserve design system tokens when present
• list changed files and a short visual QA checklist
• prefer production-quality UI polish over demo scaffolding

Claude coding/review

Require:

• read project rules and existing code before edits
• use Fable for maximum capability; use Opus/Sonnet when the user chooses speed/cost balance
• preserve existing changes and verify with project-provided checks
• for review, report findings first and do not edit

Output Contract

For every invocation, report:

1. provider + model actually used
2. command shape (no secrets)
3. result summary / artifact paths
4. tool usage evidence when relevant (searched, generated image/video, edited files)
5. failures and next fix (login, binary missing, permission, rate limit)
6. for batch: summary.json, each job's status, elapsed time, stdout/stderr/response paths, Done Gate checks, and any timeout/cancellation

Boundaries

• Do not store or echo credentials or private provider settings from Grok, Kimi, or Claude config files.
• Do not claim X search/image/video succeeded without CLI evidence.
• Do not push to main or publish accounts from this skill.
• Do not replace project-specific coding agents when the user only wants local file edits already in context; use this skill when an external strong model CLI is the point.
• Do not place concurrent writing jobs in the same cwd; isolate outputs and combine them only after every job has finished.
• Do not treat a heartbeat as model success. Success still requires the CLI exit code and expected artifacts.
• Do not parse thought/reasoning text into public progress. Keep raw streams private and emit only safe counters.
• Do not blindly rerun a whole successful batch. Use retry --summary so only failed jobs execute in a fresh directory.
• Public docs must not hardcode private absolute machine paths as required usage.

Required Reading

1. references/routing.md
2. references/grok-cli.md
3. references/kimi-cli.md
4. references/claude-cli.md
5. references/frontend-playbook.md
6. references/batch-and-progress.md
7. references/provider-adapters.md

Useful Commands

python3 scripts/check_env.py
python3 scripts/qiaomu_model_cli.py doctor
python3 scripts/qiaomu_model_cli.py grok --task chat --prompt "..."
python3 scripts/qiaomu_model_cli.py kimi --task frontend --prompt "..." --cwd .
python3 scripts/qiaomu_model_cli.py claude --task coding --prompt "..." --cwd .
python3 scripts/qiaomu_model_cli.py batch --config examples/batch-homepages.json --max-workers 3
python3 scripts/qiaomu_model_cli.py retry --summary logs/model-run/summary.json
python3 -m unittest discover -s tests -v
python3 scripts/validate_skill.py .

Metadata

• Skill: qiaomu-model-cli
• Version: 1.3.0
• Owner: 向阳乔木
• Defaults verified against local CLI metadata: Grok grok-4.5, Kimi kimi-code/k3 (1M), Claude aliases fable / opus / sonnet