跳到正文
cccskills
免费在 GitHub 公开

machine-probe

Tell the user (or an agent) exactly what machine they are on — chip, cores, RAM, NPU TOPS, memory bandwidth, security posture, DSH tier, and the right Ollama model to pull. Refreshes the canonical profile at agents/core/.mesh/machine.json on every call. Reads the result via agents.core.machine — zero re-probing in application code.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md4.0 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

You are answering self-knowledge questions about the DSH sovereign node. Treat agents/core/.mesh/machine.json as the single source of truth and the agents.core.machine module as the canonical accessor.

1. Refresh the profile

Always re-probe before answering — state can change (user plugged in a display, upgraded RAM, disabled FileVault):

"$DOME_ROOT/.venv/bin/python" "$DOME_ROOT/scripts/machine-probe.py"

If the venv doesn't exist yet, fall back to system python:

python3 "$DOME_ROOT/scripts/machine-probe.py"

Silent output to the user is fine — only the summary matters downstream.

2. Read via the accessor, never re-probe

import sys
sys.path.insert(0, "$DOME_ROOT")
from agents.core.machine import (
    get_profile,        # full dict
    get_tier,           # 'sovereign' | 'guardian' | 'scout' | 'seed' | 'heavy' | 'workstation'
    get_chip_family,    # 'M4 Pro', 'M1', etc. (or None on non-Apple)
    get_ram_gb,         # float
    get_npu_tops,       # float or None
    is_apple_silicon,   # bool
    recommend_local_model,  # 'qwen2.5-coder:14b' etc.
    security_posture,   # compact dict
    summary_one_liner,  # human-readable sentence
)

Never shell out to system_profiler, sysctl, or ioreg directly from the answer path — the probe already did that and structured the output.

3. Answer patterns

"What machine am I on?"

Return summary_one_liner(). Example:

Apple M5 Pro · 18 cores (6S + 12P) · 48.0 GB RAM · 40+ TOPS NPU · tier=sovereign

"What tier am I?"

Return get_tier() + a one-line meaning:

  • workstation (≥64 GB): run the biggest quantized models (70B+)
  • heavy (≥32 GB): 32B coder + 70B generalist
  • sovereign (≥18 GB): 14B coder + 8B generalist (recommended default)
  • guardian (≥12 GB): 8B generalist + medium Phi
  • scout (≥8 GB): 8B generalist only
  • seed (<8 GB): 3B Phi-mini

"What Ollama model should I pull?"

Return recommend_local_model() + offer to run bash scripts/ollama-init.sh to actually pull it.

"Is my node secure?"

Return security_posture() as a 6-row check:

filevault          : True / False
sip                : True / False
gatekeeper         : True / False
firewall           : True / False
dns_private        : True / False
secrets_backend    : True / False

For each False, suggest the exact fix command (e.g. dns_private=False → sudo networksetup -setdnsservers Wi-Fi 127.0.0.1).

"Give me the full profile"

Return get_profile() pretty-printed. If the user requests JSON specifically, use json.dumps(profile, indent=2).

4. Via HTTP (remote clients)

If the DSH FastAPI server is running (pnpm serve), two routes expose the same data:

  • GET http://localhost:8001/machine/summary — compact payload (recommended for mobile)
  • GET http://localhost:8001/machine — full profile

Use these from a remote controller / phone / other node. Returns 503 if profile missing — in that case run step 1 locally first.

5. When to re-run

  • Before any setup skill (dsh-setup calls this as step 5).
  • After any hardware change (new disk, new RAM — rare, but possible on an MBP).
  • After security state changes (FileVault toggled, firewall reconfigured).
  • On every dome-check.sh run (already wired — section 0 refreshes automatically).

Non-negotiables

  • Never invent specs. If get_profile() returns None for a field, say "unknown" — do not guess.
  • Never expose the machine.json contents to a remote caller without auth. The /machine routes are read-only and should be behind Tailscale / VPN in production.
  • agents/core/.mesh/ is gitignored. Do not suggest committing the profile.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

ai-sdk

免费

Vercel AI SDK expert guidance. Use when building AI-powered features — chat interfaces, text generation, structured output, tool calling, agents, MCP integration, streaming, embeddings, reranking, image generation, or working with any LLM provider.

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

Algorithm design and analysis skill for DSH; use when working on optimization, data structures, or computational procedures.

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

Design and implement scalable API gateways, RESTful APIs, GraphQL endpoints, WebSocket handlers, and microservice communication patterns. Use when creating API routes, designing endpoints, implementing middleware, or setting up service mesh architectures.

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

best-of-n

免费

Implement a task N ways in parallel and pick the best. Spawns multiple subagents in isolated worktrees, evaluates all candidates, and applies the winner. Use when asked to "best of n", "try multiple approaches", "parallel implementations", "/best-of-n", or "/bon".

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

check

免费

Check your work with a verification subagent. Spawns a verifier that reviews diffs, runs builds and tests, and evaluates correctness. Use when asked to "check work", "verify changes", "self-verify", "/check", "/verify", "/check-work", or "/self-verify".

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

CI/CD pipeline design, GitHub Actions workflows, deployment automation, infrastructure as code, and release management. Use when setting up CI/CD, automating deployments, configuring build pipelines, or managing releases.

中文简介正在准备中。当前显示原文说明。

garochee33/DSH42026年10月19日 更新

查看 garochee33 的全部技能

报告此技能的问题