hu-wentao/skills

host-governance

>- Govern shared hosts and infrastructure across projects: authoritative inventory, server bootstrap, Docker/storage, CI runners and Jenkins, PostgreSQL, Tailscale, Caddy, Cloudflare, resource inspection, and reusable host operations. Use for shared-host reads, plans, authorized mutations, verification, rollback, or incident recovery.

First seen Aug 2, 2026

Installation

$ npx skills add hu-wentao/skills --skill host-governance

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from hu-wentao/skills · top by installs.

npx skills add hu-wentao/skills

Browse all from hu-wentao/skills

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Not declared
Cursor Not declared
Codex Not declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

Default branch main
Open issues 0
Status Active

Skill metadata

Parsed from SKILL.md frontmatter.

More metadata
context-budget
router

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 7,713 B
  • docs SUMMARY.md 356 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 17 installs

SKILL.md

Host Governance

Own shared compute, CI, network, ingress, DNS, and host transactions. Keep application identity and build intent in the consuming project; keep host facts and controllers in the authoritative host repository.

Locate Authority

Use the first unambiguous source: current user path, authority already established for the task, HOSTINFRAROOT, then the locator in [context.md](references/context.md). Do not scan or copy infrastructure inventory.

Repository context is declared evidence, not live state. Use contracted catalog, search, get, or current-device; never infer a device, alias, address, credential, or freshness claim. Resolve control separately for live reads or writes.

Resolve and Execute

From the authoritative or consuming project, resolve the selected task and operation:

uv run python <skill-root>/scripts/resolve.py \
  --cwd <project-root> --task <task> --operation <operation> --format json

For host-governance.config.v2, read the returned policy references and execute only through the validated runner:

uv run python <skill-root>/scripts/host-governance.py \
  --cwd <project-root> control <operation> [contracted arguments]

Use --authorized only when current user intent covers the non-read-only operation. The runner validates argv, parameters, mutability, authorization, secret environment requirements, outputs, exit states, and transitions. Never bypass a v2 contract with its underlying command. Legacy v1 output is instruction-only. Read [projectconfig.md](references/projectconfig.md) before changing a profile.

Route the Operation

  • Generic transaction planning and ownership: [control.md](references/control.md).
  • Explicit request to turn a procedure into a reusable contracted function: [procedure-productization.md](references/procedure-productization.md).
  • Initial Linux server bootstrap: [server-bootstrap.md](references/server-bootstrap.md).
  • Docker Engine/Compose installation and bounded BuildKit cleanup: [docker-install.md](references/docker-install.md).
  • Docker volume leaks or runtime storage incidents: [docker-storage-maintenance.md](references/docker-storage-maintenance.md).
  • GitHub Actions runner: [github-actions-runner.md](references/github-actions-runner.md).
  • Jenkins controller, agents, plugins, credentials, jobs, backup, upgrade, or mobile packaging: [jenkins.md](references/jenkins.md).
  • PostgreSQL deployment, sizing, parameters, or lifecycle: [postgresql.md](references/postgresql.md); run scripts/postgres_sizing.py before rendering parameters.
  • Tailscale installation, update posture, policy, or node settings: [tailscale.md](references/tailscale.md).
  • Caddy and HTTP/TLS ingress: [caddy.md](references/caddy.md).
  • Cloudflare accounts, DNS, WAF/Rulesets, or selective 403 diagnosis: [cloudflare.md](references/cloudflare.md).
  • Cloudflare Tunnel and Access publication: [cloudflare-tunnel.md](references/cloudflare-tunnel.md).

For deployed service, listener, limit, persistence, or capacity inventory, use only the project-owned read-only inventory operation for one exact device. Collect secret-safe bounded facts; never environment values, full command lines, logs, keys, tokens, or container secret content. Treat snapshots as dated observations, not desired state.

Authorization and Safety

Read [authorization-and-safety.md](references/authorization-and-safety.md) before any host or provider mutation, bootstrap, credential use, emergency manual operation, or conversation-scoped SSH execution. Core invariants:

  • inspect, plan, and verify are read-only. Writes use an authorized contracted transaction; rollback and materially broader effects require their own current scope.
  • One explicit 授权手动进行紧急操作 grants only the current bounded execution round. It expires on completion, blocker, or return to the user and never creates persistent arbitrary-command capability.
  • Conversation-scoped SSH authorization covers only SSH actions for the established task and target; it does not authorize provider consoles, APIs, billing, identity changes, release, or new targets.
  • Never expose or persist tokens, passwords, private keys, session data, environment dumps, secret request bodies, or interactive login URLs.
  • Use only manifest-declared device IDs, SSH aliases, ports, and jump routes. Never guess or bypass a jump host.
  • Snapshot current state, lock one serialized host transaction, re-read under lock, reject drift/collisions, validate the complete candidate, preserve recovery, and verify positive and negative paths.
  • Do not broaden network access to make validation pass. Keep destructive, billable, privileged, exposure-changing, and identity-changing outcomes separate.

Separate One-Time Operations from Productization

A request to execute, repair, recover, clean up, migrate, or order steps for one current target is a one-time operation, even when the method could be reused. Selecting a plan or sequence authorizes only that execution path; governance requirements, repetition history, summaries, and prior assistant proposals do not authorize a durable capability.

For a one-time operation, resolve and compose existing contracted operations. Do not edit the host repository, controllers, contracts, tests, documentation, or versions, and do not install or publish a capability. If no existing operation can safely complete it, report the missing capability or use only an explicitly authorized one-round manual exception.

Productize only when the current user explicitly asks to create, automate, contract, or retain a reusable long-term host capability. Then first resolve an existing contracted controller; otherwise follow [procedure-productization.md](references/procedure-productization.md). Never create a generic shell-fragment or arbitrary-package executor.

Control Workflow

  1. Establish project intent, host authority, exact target, current live state, and owner for every desired fact.
  2. Build one matrix covering current/desired state, executor, authorization, validation, exposure, recovery, and rollback.
  3. Inspect all coupled products before the first write; a repository declaration is not a runtime claim.
  4. Apply only the authorized contracted operations under the host-owned lock. Stop on drift, collision, unexpected output, or failed verification.
  5. Verify each component immediately, then run end-to-end and negative-path checks.
  6. End bootstrap or recovery only as verified complete, explicitly blocked, or rolled back. Partial setup is not completion.

Product-specific completion details remain in their owning references. Examples include hardware-aware PostgreSQL sizing, root-equivalent Docker group access, separate authorization for volume deletion, fresh administrator-key bootstrap verification, immediate return of a Tailscale interactive login URL, and coupled Cloudflare origin/tunnel/DNS/Access verification.

Report

Report repository and live changes separately; exact target, transaction/generation, validation evidence, secret-safe credential requirements, exposure, recovery/rollback state, compatibility, and every unverified gap. Do not promote observations to desired declarations without separate repository-write authority.