菜单
Use when switching the local MLX inference server model on Mac, checking what model is loaded, monitoring downloads, or configuring Cline/Roo Code to use a local model.
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
Mark a logical phase completion in haivemind. More formal than /snapshot — used at step/phase boundaries in multi-phase work. Creates a named, resumable point that /continue can restore from. Triggered by /checkpoint, "checkpoint this", "mark this as a checkpoint", or at phase boundaries in long-running tasks.
Run the Qwen Code CLI as a one-shot bounded executor in a plan/do/verify loop. A frontier model (or you) plans the next bounded step and verifies the result; qwen-do makes a local model do exactly that one step, smoke-test first, then stop. Use when you want a cheap local model to do grunt coding while a smarter orchestrator keeps it on rails.
Scan git commits for PII, credentials, API keys, and passwords before they are pushed. Automatically gates commits via a pre-commit hook. Use when installing PII scanning on a repo, manually scanning a file or diff for credentials, or when a commit is blocked by the hook and remediation is needed. Triggers on "scan for PII", "check for credentials", "install PII guard", "pii hook", or when a commit is blocked by a PII check.
Runs a full CORE-EEAT 80-item content quality audit, scoring content across 8 dimensions with weighted scoring by content type. Produces a detailed report with per-item scores, dimension analysis, and a prioritized action plan.
Deep analysis and debugging with Opus. Systematically investigate problems, identify root causes, and recommend fixes with evidence.
PR review skill. Two modes: /pr [link] = internal thoughts only, what to look for, what to test — NO GitHub action ever. /pr-comment [link] "make it so #1" = draft a GitHub PR comment in <USER_NAME>s voice, present for approval, post only on explicit "go". Passphrase + "go" are both required before any GitHub write. Without both gates, nothing leaves the building.
| name | mlx-switch |
| description | Use when switching the local MLX inference server model on Mac, checking what model is loaded, monitoring downloads, or configuring Cline/Roo Code to use a local model. |
| category | execution |
| runtimes | ["claude"] |
| pii_safe | true |
Local MLX inference server on <MAC_HOST> (:8080). OpenAI-compatible. One model loaded at a time — model field in requests is cosmetic, server always uses the loaded model.
ssh <MAC_HOST> "bash /usr/local/bin/mlx-switch <alias>"
Wait for ready:
until ssh <MAC_HOST> "curl -sf http://localhost:8080/v1/models >/dev/null 2>&1"; do sleep 4; done && ssh <MAC_HOST> "tail -1 /tmp/mlx-server.log"
| Alias | Model ID | Context | Type | Best For |
|---|---|---|---|---|
qwen36big | mlx-community/Qwen3.6-35B-A3B-4bit | 256K | MoE 35B/3B | Default coder (73.4% SWE-bench, ~91 tok/s) |
coder-next | mlx-community/Qwen3-Coder-Next-4bit | 256K | MoE 80B/25B | Code backup |
gemma | mlx-community/gemma-4-26b-a4b-it-4bit | 256K | MoE 26B/4B | Vision/multimodal |
http://<PRIVATE_IP>:8080/v1local)mlx-community/Qwen3-Coder-Next-4bit)gemma supports image/screenshot input<think> blocks — add /no_think to system prompt or use devstral2/coder-next insteadssh <MAC_HOST> "curl -s http://localhost:8080/v1/models | python3 -c 'import json,sys; [print(m[\"id\"]) for m in json.load(sys.stdin)[\"data\"]]'"
# Note: lists ALL cached models, not just active one
# Active model = what's in the plist:
ssh <MAC_HOST> "/usr/libexec/PlistBuddy -c 'Print :ProgramArguments:2' ~/Library/LaunchAgents/com.<ORG_NAME>.mlx-server.plist"
ssh <MAC_HOST> "tail -5 /tmp/mlx-devstral2.log && du -sh ~/.bun/cache/huggingface/hub/models--mlx-community--Devstral-2-123B-Instruct-2512-4bit/ 2>/dev/null"
Cache location: /Users/<USER_NAME>/.bun/cache/huggingface/hub/
HF_TOKEN: stored in ~/.zprofile and plist env on Mac.
alias) echo "mlx-community/Model-ID-4bit" ;; to /usr/local/bin/mlx-switch case statement/usr/local/bin/mlx-downloadssh <MAC_HOST> "bash /usr/local/bin/mlx-switch <alias>"| Problem | Fix |
|---|---|
declare -A error | Mac uses bash 3.2 — script uses case statement, not associative arrays |
| Cache not found | Plist needs HF_HUB_CACHE=/Users/<USER_NAME>/.bun/cache/huggingface/hub |
| Model field ignored | mlx_lm.server loads one model at startup — set Cline model to match active |
<think> blocks breaking tool calls | Use devstral2, or add /no_think to system prompt |
| Architecture not supported | Nemotron (NemotronH_Nano) not supported in mlx_lm 0.31.3 |