AGENT SKILL

agent-accuracy-guard

暂无简介

License 许可
MIT
最近更新
2026-09-18
Stars
★ 0
Forks
0

02 资源文件 2

03 SKILL.md 正文

Agent Accuracy Guard

Use this before high-risk actions.

Non-negotiable checks

  1. Authorization

    • If the user asked for a plan first, stop after the plan.
    • Do not run agent swarm, bulk jobs, paid APIs, email, or external sends without explicit approval.
    • "I guess continue" counts as approval only if the user is clearly telling you to continue the same pending action.
  2. Provider/model verification

    • Never infer model availability from a local proxy unless the user explicitly said to use that proxy.
    • Alibaba Coding Plan is direct Alibaba Model Studio/Coding Plan, not Command Code, OpenCode, 9router, or any local proxy.
    • Verify against the authoritative source or current configured provider before stating a model is available.
  3. Delivery method

    • Telegram delivery means use Telegram tools/attachments.
    • Do not invoke Resend/email rules for Telegram delivery.
    • If the user asks for an MD file over Telegram, send the actual .md file attachment, not a summary.
  4. Known bad recommendations

    • Do not recommend Meilisearch on this VPS. It has caused memory leaks / RAM exhaustion / daily crashes.
    • Do not recommend local GPU model services; VPS has no GPU.
    • Do not recommend more OCR stacks unless the user explicitly asks; OCR is already strong enough.
    • Command Code via Hermes: Do not claim Kimi/command-code lacks Telegram, VPS access, or Hermes skills. Delivery = Hermes Gateway. VPS shell = CLI proxy :9877. Skills = read SKILL.md from disk (no native skill_view). Load command-code-proxy before stating capabilities.
  5. Verification before claiming done

    • Before saying a file was sent, verify the tool result includes an attachment success.
    • Before saying a field/file/service was fixed, inspect or test it.
    • If not verified, say exactly what was done and what is unverified.
    • Notary-log cal host: Before saying Settings shows tokens, "Create my account", or Cal UI exists, verify deployed bundle (bun run build + service restart + grep dist or live fetch). Code in repo alone is not proof Joseph sees it.
    • Before running verify-cal-host.mjs, warn it wipes cal DB users/bookings — Joseph's phone token goes stale.
    • Cal "100% complete": split Plan A MVP (paste-link + webhooks on cal host) vs OAuth vs Worker ship vs real Cal booking on user's phone.

5a. Verify before claiming ABSENCE (hard-won, recurring failure)

  • Do NOT say "there is no native import / endpoint / feature / file / capability" based on a shallow first-pass grep or memory of "how these tools usually work." This pattern burned the user badly on 2026-07-25: the assistant twice made confident negative claims ("no native 9router→OmniRoute import exists", "Qwen session cookies are not persisted / can't be extracted") that were both wrong, and the user correctly called it lying.
  • Before claiming absence of a feature/endpoint/file/capability in a tool on this VPS (OmniRoute, 9router, qwen-gate, etc.):
    1. Grep the source tree for the symbol/endpoint name (case-insensitive, both camelCase and snake_case).
    2. Check the actual routes/endpoints the running server exposes (curl localhost:<port>/api/..., read the route files).
    3. Check the actual on-disk state (SQLite tables, browser profiles, .env, config files) — don't answer from memory.
    4. Only THEN say "no, it doesn't exist" — and cite what you checked.
  • If you didn't do steps 1-3, say "I need to verify" and do them. A wrong "no" wastes more time than a verification pass. The user has zero tolerance for confident-but-unverified negative claims — repeated instances get treated as lying, not as honest mistakes.
  1. Version / update claims (Joe)
    • Do not equate "installed ≠ latest release" with "you need to update" for Hermes, mem0, QWEN-gate, or 9router unless Joe asked for upgrades or the install is broken/missing.
    • Hermes at /usr/local/lib/hermes-agent may be a custom build on purpose.
    • If Joe says an item does not need updating, answer NO — do not override with GitHub/PyPI tags.
    • If you already gave a wrong yes/no in the same thread, correct it in one direct sentence.

Command

Run:

python3 /home/workspace/Skills/agent-accuracy-guard/scripts/preflight.py --task "<task>"

It prints a checklist to follow before acting.

04 同仓库更多 Skills