跳轉到

Agent Harness —— Agent 殼層

Terminology rule (zh-TW pages)

技術名詞首次出現以「中文 (English original)」格式呈現,例:依賴注入 (dependency injection)。不自創翻譯——若無公認譯名直接保留英文 (如 embedding、tokenizer)。代碼、API 名、CLI flag、套件名、檔名一律不翻。

規格驅動開發 (spec-driven development, SDD) 框架與 agent harness —— 位於 agent skill 之上的層級,負責 requirements → spec → plan → tasks → execute → verify 這個 loop,或管理 context window、執行 sub-agent。

本 repo 聚焦在 skill,刻意 (intentionally) 不 ship SDD framework 或 harness。不過會提供一個設計/審查 LLM application layer 的 local skill,並收錄 Herdr 的官方控制 adapter;兩者都不會把此 collection 變成 harness runtime。

此 repo 內的 skill

Local

Skill 一句話 備註
12-factor-agent-design-review 用 HumanLayer 12-Factor Agents 原則設計或 evidence-review production LLM application 審查 harness 下方的 workflow/application layer:prompt、context、typed decision、state、control flow、pause/resume、human、retry 與 replay;不提供 runtime。

Vendored

Skill Upstream Series
herdr herdrdev/herdr flat

External skills(手動安裝)

Status enum (single source of truth across the catalog)

Every external-skill or MCP entry below carries a status: value:

Status Meaning Where it links
vendored Already in vendor.yaml. Link to the skills/vendor/<name>/ (or series) entry.
deferred Open TODO P? item — under consideration. Link to the TODO.md anchor.
skipped Looked at, chose not to vendor. Inline reason required (1 sentence).
evaluated Read but no decision recorded. Effectively "tracked." Inline 1-line note describing what was learned.
wishlist Surfaced but not yet evaluated. No link required; default for fresh discoveries.

Status changes are explicit edits — see the Adding catalog entries workflow doc (under Workflows in the nav) for the wishlist → deferred → vendored recipe.

Manual install for non-vendored skills

When a skill is not in vendor.yaml, you can still install it directly from upstream via the npx skills CLI:

# Whole-repo install (if SKILL.md sits at repo root or skills/)
npx skills@latest add <owner>/<repo>

# Subpath install (when the skill is nested)
npx skills@latest add <owner>/<repo>/<path-to-skill-dir>

# Cherry-pick by name
npx skills@latest add <owner>/<repo> -s <skill-name>

# Force deep discovery (skill nested >1 level inside skills/)
npx skills@latest add <owner>/<repo> --full-depth -s <skill-name>

For Anthropic-style plugin marketplaces (e.g. anthropics/financial-services, anthropics/knowledge-work-plugins), use the Claude Code plugin commands instead:

claude plugin marketplace add <owner>/<repo>
claude plugin install <plugin-name>@<repo-name>

For MCP servers, see the per-entry "Auth & install" section in the MCP wiki — install commands vary by transport (HTTP / stdio / SSE) and host (Claude Code / Claude Desktop / Cursor / Managed Agents).

Skill / Plugin Upstream Status 為何此狀態 安裝提示
spec-kit(skills mode) github/spec-kit evaluated 事實上 (de facto) 的 SDD framework(95.5k ⭐)。支援 30+ agent。有 --skills 安裝模式可把 slash command ship 成 agent skill。文件見 reference/sdd-and-harnesses.md。未 vendor —— 與本 repo 的 per-skill 哲學會競爭。 uvx --from git+https://github.com/github/spec-kit specify init <project> --integration claude-code --integration-options="--skills"
get-shit-done (gsd) gsd-build/get-shit-done evaluated 較早、較輕量的 SDD(61.4k ⭐)。六指令 loop(gsd-new-project → gsd-discuss-phase → gsd-plan-phase → gsd-execute-phase → gsd-verify-work → gsd-ship)。 (見 upstream README)
gsd-2 gsd-build/gsd-2 evaluated 獨立 harness(不只是 SDD framework)。 (見 upstream README)
OpenClaw openclaw/openclaw evaluated 控制 agent session 的獨立 CLI / runtime。同時也產出 gstack-openclaw-* skill(已歸在 product-planning series vendored)。 (見 upstream README)
Pi SDK badlogic/pi-mono evaluated Agent harness 的另一選項。 (見 upstream README)
agent-architecture-analysis existential-birds/beagle skipped Evidence-first 12-Factor review 有價值,但 rubric 綁定 Python 與特定 implementation;local skill 只引用可攜的 evidence-gate 構想。 npx skills add existential-birds/beagle --skill agent-architecture-analysis
12-factor agent skill pack tika/12-factor-agent-skills skipped 五個 skill 的廣泛組合,但 cross-skill reference 與 scanner self-match 問題使它不適合原樣 vendor。 npx skills add tika/12-factor-agent-skills
pstack(poteto-mode) cursor/plugins/pstack evaluated 強主張的 Cursor workflow pack:router、playbook、principle skill 與 subagent fan-out。耦合度高且更新太快,不適合 vendor。Claude Code 由 upstream 生成的 adapter 支援;Codex 等其他 agent 可用社群移植(Go7hic/ystack、TheOnlyFusionCube/potetos-for-everyone)。見 skill collections。 /plugin marketplace add cursor/plugins then /plugin install pstack@cursor-plugins

MCP servers

名稱 Upstream Status Auth 紀錄
不適用 (not applicable) —— harness 用底層 agent 能用的 MCP

Backlog(TODO P? 條目)

  • Harness runtime 仍不在範圍;只有可重用的 skill adapter 才適合 vendor。

另見

  • docs/reference/sdd-and-harnesses.md —— SDD framework 與 agent harness 的完整 survey,附三層分層 (skill vs SDD framework vs harness)說明。
  • product-planning series 在 docs/skills/index.md —— vendored 的 OpenClaw skill(gstack-openclaw-*),這些是在本 repo 範圍內。
  • herdr —— 官方控制 skill,以及為何安裝 binary 時應優先用 binary-emitted copy 取得精確版本對齊。
  • 12-factor-agent-design-review —— 何時應設計/審查 LLM application layer,而不是導入 harness。