agentic-sop-to-work
Loop Engineering for Claude Code agent loops — deterministic bounded termination (stall + budget), observable health, and bounded state — and Graph Engineering for when one loop isn't enough: typed handoff edges, unique field ownership, bounded back-edges, and static topology validation. Ships agentic-sop-kit (human SOP → controlled loop or bounded graph) plus cross-domain review companions: Second Opinion (honesty reviewer), alcoa-guard (ALCOA+ data-integrity), plugin-forge (plugin linter/scaffolder).
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/s0912758806p--agentic-sop-to-work)HTML
<a href="/listings/s0912758806p--agentic-sop-to-work"><img src="/api/listings/s0912758806p--agentic-sop-to-work/badge.svg" alt="findsafeskills safety rating" /></a>Install
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 65 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.
- {
- "name": "agentic-sop-to-work",
- "owner": {
- "name": "s0912758806p",
- "url": "https://github.com/s0912758806p"
- },
- "description": "Loop Engineering for Claude Code agent loops — deterministic bounded termination (stall + budget), observable health, and bounded state — and Graph Engineering for when one loop isn't enough: typed handoff edges, unique field ownership, bounded back-edges, and static topology validation. Ships agentic-sop-kit (human SOP → controlled loop or bounded graph) plus cross-domain review companions: Second Opinion (honesty reviewer), alcoa-guard (ALCOA+ data-integrity), plugin-forge (plugin linter/scaffolder).",
- "plugins": [
- {
- "name": "agentic-sop-kit",
- "source": "./plugins/agentic-sop-kit",
- "description": "Human SOP → a controlled agentic loop, or a bounded graph of them: typed handoff edges, one declaring writer per state field, bounded back-edges, and a static topology gate that refuses an illegal graph before it runs. Methodology skills + /sop-flow + a project-scoped regression Stop-hook + the portable kit.",
- "category": "development",
- "tags": [
- "agentic-workflow",
- "sop",
- "loop-engineering",
- "graph-engineering",
- "deterministic-gates",
- "stdlib"
- ]
- },
- {
- "name": "second-opinion-honesty",
- "source": "./plugins/second-opinion-honesty",
- "description": "Independent adversarial honesty reviewer: scrutinizes a DRAFT output's claims-vs-evidence (no fabrication / overreach / invented IDs) before a human approves. Snaps onto agentic-sop-kit run dirs (FULL, uses the trace) or reviews any plain document (DEGRADED). Deterministic checks + a capped advisory LLM pass; the report is a DRAFT — the human owns the verdict.",
- "category": "quality",
- "tags": [
- "honesty-review",
- "fact-check",
- "claims-vs-evidence",
- "draft-review",
- "advisory",
- "stdlib"
- ]
- },
- {
- "name": "alcoa-guard",
- "source": "./plugins/alcoa-guard",
- "description": "Deterministic ALCOA+ data-integrity linter: checks Attributable / Contemporaneous / Complete / Accurate / Consistent in code and surfaces the human-judgment slice as a 【待補】 checklist it never auto-concludes. DEGRADED lints any CSV/JSON record against a declared contract; FULL snaps onto an agentic-sop-kit run dir. The report is a DRAFT — the human owns the verdict. Pure stdlib.",
- "category": "compliance",
- "tags": [
- "alcoa-plus",
- "data-integrity",
- "gmp",
- "gxp",
- "audit-trail",
- "stdlib"
- ]
- },
- {
- "name": "plugin-forge",
- "source": "./plugins/plugin-forge",
- "description": "Claude Code plugin linter + scaffolder: `lint` validates a plugin or whole marketplace against a house grammar (a strict superset of the manifest/frontmatter checks plus stdlib-only, hook-protocol, and test-harness invariants); `scaffold` generates a grammar-conformant plugin skeleton. Pure stdlib; self-hosting; standalone on any plugin.",
- "category": "development",
- "tags": [
- "plugin-linter",
- "scaffolder",
- "marketplace",
- "ci",
- "stdlib"
- ]
- }
- ]
- }
README.md · README · 113 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.
- # agentic-sop-to-work
- [](https://claude.com/claude-code) [](LICENSE) [](plugins/agentic-sop-kit/.claude-plugin/plugin.json) [](plugins)
- > Turn a process you do by hand into steps a program runs — and never let the AI make up facts, loop forever, or have the final say.
- > 把你手動做的流程,變成程式一步步執行的流程——而且不讓 AI 編造事實、無限打轉,或替你做最後決定。
- ```
- /plugin marketplace add s0912758806p/agentic-sop-to-work
- /plugin install agentic-sop-kit@agentic-sop-to-work
- /reload-plugins
- ```
- **🌐 [English](#english) ・ [繁體中文](#繁體中文)**
- 
- *Follow the dot: 🟢 passed · 🔴 blocked · 🟠 retried or sent back, always with a limit · 🟣 handed to review.<br>跟著圓點走:🟢 通過 · 🔴 擋下 · 🟠 重試或退回(一定有上限) · 🟣 交給審查。*
- ---
- ## English
- **How it works** — the five things the animation shows:
- 1. **Check the plan first.** `run.py --plan` reads the whole flow and refuses one with a dead end or an endless loop *before anything runs*.
- 2. **Code runs every step.** One step = one tool. If a tool fails or writes nothing, the run stops. To also verify *what* it wrote, attach a gate (`cmd_gate` · `schema_gate` · `trace_gate` · `recompute_gate` · `residual_gate`) — plain code, no AI grading itself.
- 3. **Retries have limits.** `/sop-flow` auto-fixes and re-runs a failed step up to 3 times (configurable), and stops early if it is going in circles. A review step can send work back, but only as many times as the flow declares.
- 4. **The result is a DRAFT.** Missing facts are marked `【待補】`, never invented. Before a human signs off, `second-opinion-honesty` and `alcoa-guard` can check the draft.
- 5. **Every code change is re-tested.** In a project that uses the kit, a Stop hook runs the regression tests when Claude finishes, and blocks if something broke.
- **What's in the box** — four plugins, pure Python standard library, every report is a DRAFT:
- | Plugin | In plain words |
- |---|---|
- | `agentic-sop-kit` — **4 Skills** | `six-rung-ladder`: should we build this at all? · `agentic-sop`: turn an SOP into a flow · `agentic-workflow-audit`: catch a "do-everything" agent · `/agentic-sop-kit:sop-flow`: run the flow, get a DRAFT |
- | `second-opinion-honesty` | Checks a DRAFT's claims against its own evidence: made-up numbers, invented IDs, overreach. Findings are hard on a kit run, advisory on a plain document. |
- | `alcoa-guard` | ALCOA+ data-integrity checks — who, when, complete, accurate, consistent. What needs human judgment becomes a `【待補】` checklist. |
- | `plugin-forge` | Lints and scaffolds Claude Code plugins: manifests, skills, hooks, stdlib-only code, vendored-core drift, and current platform features (`agents/`, `userConfig`, exec-form hooks). |
- **Install** — the three lines at the top. Needs **Python 3.8+** as `python3`. Check: `/help` lists `/agentic-sop-kit:sop-flow`. Settings when you enable the plugin: `max_fix_retries` · `stall_window` · `keep_runs`.
- **Compatibility** — what CI actually runs on every PR, for all four plugins:
- | | Python 3.8 | Python 3.x (latest) |
- |---|---|---|
- | Linux (ubuntu) | ✅ tested (22.04) | ✅ tested |
- | macOS | — | ✅ tested |
- | Windows | — | ⚠️ runs in CI, non-blocking |
- **Go deeper** — [`kit/README.md`](plugins/agentic-sop-kit/kit/README.md): the engine, gates, graph features, settings · [`kit/SOP.md`](plugins/agentic-sop-kit/kit/SOP.md): the method · [`CONTRIBUTING.md`](CONTRIBUTING.md): tests, versioning, releases · [`MIGRATION.md`](MIGRATION.md): what v2.0 removes · [`CHANGELOG.md`](CHANGELOG.md).
- **Update** — `/plugin marketplace update agentic-sop-to-work`, then `/reload-plugins`. Each release is a git tag (`vX.Y.Z` for the kit, `<plugin>-vX.Y.Z` for the others) with a GitHub release.
- ---
- ## 繁體中文
- **怎麼運作** — 動畫裡的五件事:
- 1. **先檢查流程。** `run.py --plan` 讀完整張流程圖;有死路或無限迴圈,*還沒開跑*就拒絕。
- 2. **每一步都由程式執行。** 一步=一個工具。工具失敗或沒有產出,流程就停。要再驗證產出*內容*,就在那一步掛上閘門(`cmd_gate` · `schema_gate` · `trace_gate` · `recompute_gate` · `residual_gate`)——全是程式,不讓 AI 自己打分數。
- 3. **重試有上限。** `/sop-flow` 遇到失敗會自動修、重跑,最多 3 次(可調),原地打轉就提早停。審查步驟可以退件重做,但次數不能超過流程宣告的上限。
- 4. **產出只是草稿。** 缺的事實標 `【待補】`,絕不編造。交給人核准之前,可以先用 `second-opinion-honesty` 和 `alcoa-guard` 檢查。
- 5. **每次改程式都重新測試。** 在導入 kit 的專案裡,Claude 收工時 Stop hook 會自動跑回歸測試,壞了就擋下。
- **你會得到** — 四個 plugin,純 Python 標準庫,所有報告都是草稿:
- | Plugin | 白話說明 |
- |---|---|
- | `agentic-sop-kit` — **4 支 Skills** | `six-rung-ladder`:這東西到底該不該做?· `agentic-sop`:把 SOP 變成流程 · `agentic-workflow-audit`:抓出「什麼都做」的 agent · `/agentic-sop-kit:sop-flow`:跑流程、拿到草稿 |
- | `second-opinion-honesty` | 拿草稿的說法去對它自己的證據:編造的數字、發明的編號、說過頭。審 kit 的 run 是硬性發現,審一般文件只給建議。 |
- | `alcoa-guard` | ALCOA+ 資料完整性檢查——誰做的、何時、完整、正確、一致。需要人判斷的部分變成 `【待補】` 清單。 |
- | `plugin-forge` | 檢查並產生 Claude Code plugin:manifest、skill、hook、只用標準庫、共用核心有沒有漂移,以及新版平台功能(`agents/`、`userConfig`、exec-form hook)。 |
- **安裝** — 最上面那三行。需要 **Python 3.8+**,且能用 `python3` 呼叫。確認:`/help` 看得到 `/agentic-sop-kit:sop-flow`。啟用 plugin 時可設定:`max_fix_retries` · `stall_window` · `keep_runs`。
- **相容性** — CI 每個 PR 對四個 plugin 實際跑的組合:
- | | Python 3.8 | Python 3.x(最新) |
- |---|---|---|
- | Linux(ubuntu) | ✅ 已測(22.04) | ✅ 已測 |
- | macOS | — | ✅ 已測 |
- | Windows | — | ⚠️ CI 有跑,不擋合併 |
- **深入了解** — [`kit/README.md`](plugins/agentic-sop-kit/kit/README.md):引擎、閘門、圖狀流程、設定 · [`kit/SOP.md`](plugins/agentic-sop-kit/kit/SOP.md):方法論 · [`CONTRIBUTING.md`](CONTRIBUTING.md):測試、版本、發版 · [`MIGRATION.md`](MIGRATION.md):v2.0 會移除什麼 · [`CHANGELOG.md`](CHANGELOG.md)。
- **更新** — `/plugin marketplace update agentic-sop-to-work`,再 `/reload-plugins`。每個版本都有 git tag(kit 是 `vX.Y.Z`,其他是 `<plugin>-vX.Y.Z`)和 GitHub release。
- ---
- <details>
- <summary>A real flow, drawn from code · 一個真實流程(由程式畫出)</summary>
- <!-- BEGIN GENERATED: topology -->
- ```mermaid
- flowchart LR
- n0["draft<br/>draft@1"]
- n1["verdict<br/>verdict@1"]
- n2{{gate}}
- n3["accept<br/>accepted@1"]
- n0 --> n1
- n1 --> n2
- n2 -.->|"verdict == 'reject' (≤2)"| n0
- n2 -->|"default"| n3
- ```
- <!-- END GENERATED: topology -->
- *The shipped `graph.json` example: draft → verdict → a reject sends it back (at most twice) → accept. Generated by `run.py --graph` and byte-compared by `kit/tests/test_graph_docs.py`, so the picture cannot drift from the flow.<br>出貨的 `graph.json` 範例:起草 → 判定 → 退件就回頭(最多兩次)→ 接受。由 `run.py --graph` 產生、測試逐位元組比對,圖不可能和流程不一致。*
- </details>
- **License** — MIT ([`LICENSE`](LICENSE) · [`NOTICE`](NOTICE)); keep the notice in copies, including single files. · 授權 MIT,副本(含單一檔案)需保留聲明。
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.