FDE Scope

The complete on-site operating system for a Forward Deployed Engineer.
From first gemba walk to signed-off handoff — across software/SaaS and
embodied-robotics / manufacturing. Built on AgentScope 2.0.

FDE is 2026's hottest AI role (OpenAI, Anthropic, Google, Palantir all build FDE
teams; listings up ~7× YoY). But every FDE shows up to a customer site and
rebuilds the same workflow from scratch — and most teams only model the "build"
half, skipping the pre-engagement, operationalization, and handoff zones that
actually decide whether the engagement produces value.

FDE Scope codifies the full FDE standard operating procedure — 4 zones,
18 phases — into an executable, gate-enforced state machine, with a parallel
industrial overlay (FAT/SAT, functional safety, CE/EU-AI-Act, works council,
air-gap, shift handover) for manufacturing/robotics deployments.

flowchart LR
    subgraph ZA["Zone A · Pre-engagement — 定边界"]
        direction TB
        P1["1 立项框定 qualification"] --> P2["2 现场勘察 site survey 🏭<br/>gate: site_survey"]
        P2 --> P3["3 干系人地图 stakeholder map"]
        P3 --> P4["4 成功标准契约化 success criteria<br/>gate: success_criteria"]
    end
    subgraph ZB["Zone B · Build — 造出来"]
        direction TB
        P5["5 数据接入 connect 🏭<br/>gate: air_gap"] --> P6["6 语料锻造 corpus"]
        P6 --> P7["7 真实数据原型 prototype-real"]
        P7 --> P8["8 干系人验证 validate"]
        P8 --> P9["9 部署上线 deploy 🏭<br/>gates: fat_sat · functional_safety · conformity"]
        P9 --> P10["10 评估交付 eval"]
    end
    subgraph ZC["Zone C · Operationalization — 稳得住"]
        direction TB
        P11["11 SLO/SLA + on-call<br/>gates: slo · shift_handover 🏭"] --> P12["12 Runbook 编写"]
        P12 --> P13["13 监控与漂移 monitoring / drift"]
        P13 --> P14["14 变更管理 + 培训 🏭<br/>gate: works_council"]
        P14 --> P15["15 飞轮 → 产品化 flywheel"]
    end
    subgraph ZD["Zone D · Handoff — 交得出去"]
        direction TB
        P16["16 运维移交 ops handoff"] --> P17["17 知识转移 knowledge transfer"]
        P17 --> P18["18 退场 disengage · 客户签收<br/>gate: handoff_signoff"]
    end
    ZA --> ZB --> ZC --> ZD

    classDef ind fill:#fdecea,stroke:#c0392b;
    class P2,P5,P9,P11,P14 ind

🏭 红色 = industrial-only,仅 manufacturing profile 执行(ticket 跑 15 个阶段);
每个 gate: 挂在"离开该阶段"的动作上,advance 时实时重评,不达标不放行。
逐阶段说明见 docs/features.md。

It is not an agent application — it's an FDE's workbench: a CLI, a Web UI,
and a Python library that turns the SOP from a checklist into enforced engineering.

⏱ 最短安装路径 · The shortest path in

先说结论，因为它不好听：本项目没有"一条命令装完"的路径。实测于 2026-10-08（工具实测档）：
PyPI 上 fde-scope 与 fde_scope 均返回 404（未发布），仓库 releases 列表为空（无预打包产物）。
对照开源同行的 npx 一键分发，这条差距被如实记在 course-level-1/24 §6 第 2/3 条，
并作为短板登记在 course-level-1/13 §6 第 11 项。

按"你想干什么"选一条最短的：

你想干什么	最短路径	前置
只读课程，什么都不装	直接在 GitHub 读 `course-level-1/` 下的 Markdown（Level 1 全本，¥0 永久公开）	会点链接
跑通一条链路（学员 / 评估者）	`git clone` → `uv sync` → `uv run fde-scope course doctor`	Python ≥3.11 + uv
没有 uv	`git clone` → `python -m venv .venv && source .venv/bin/activate && pip install -e ".[dev]"`	Python ≥3.11
要界面	上面任一条之后：`fde-scope web` → `http://127.0.0.1:8080`	—

course doctor 是开工前自检：它检查环境、示例数据与零 key 路径，跑完给你 G0 的通过/提醒/失败计数——
不通的地方它会说缺什么。全程不要 Docker、不要 GPU、不要 API key
（课程侧的承诺见 11 §4.1 与 13 §1.3）。

关于 agent 运行时中立：本项目只服务自己的 CLI / Web / 库，没有为 Cursor、Copilot 之类的
第三方 agent 提供适配层。课程产物本身是 Markdown 与 JSON，可以带进任何工具——
但"带进去之后还帮你判门禁"这件事，目前只有 fde-scope 自己做。差距与计划见
24 §6 第 3 条。

📸 At a glance

FDE Scope engagement console — engagement dashboard, gate inspector, context cards, corpus forge, KPI explorer and report browser in one page

📑 Table of contents

What makes it different
Quick start · Web UI · Mock demo data
Feature checklist
Architecture · The 10 phase gates · Repository layout
Engineering best practices
CLI reference
多 Agent 生产（AgentScope 2.0）
QwenPaw 集成
Skill 沉淀（Skills / Methodology capture）
FDE 工作台（Workbench + 现场记录）
LLM 接入（可选）
Tests
Installation · macOS App
Documentation
加强计划（Enhancement Plan）
Roadmap
Contributing
License
Acknowledgements

✨ What makes it different

Full SOP, not just "5 steps". The naive connect→corpus→deploy→eval→flywheel
model covers only the Build zone. FDE Scope adds the 4 pre-engagement phases,
5 operationalization phases, and 3 handoff phases that most teams skip — and
enforces them with phase gates. The manufacturing profile runs the full
18 phases; the ticket profile runs the 15 SaaS-relevant ones. See
docs/fde_sop_full.md.
A real industrial overlay. Manufacturing/robotics engagements hit gates
that SaaS never does: FAT/SAT acceptance, functional safety
(ISO 13849 / IEC 61508 / ISO 10218), CE/EU-AI-Act conformity, German works-council
co-determination (BetrVG §87), air-gapped deployment, shift handover. Each is
an executable Gate.check() — it blocks advancement when blockers are present.
Scenario profiles. ticket (customer service) and manufacturing
(embodied-robotics factory) share one engine but differ in connectors, KPIs,
and which gates apply. Manufacturing reports real production KPIs: OEE, MTBF/
MTTR, FPY/DPMO, grasp success rate, collision/intervention rate.
Built on real AgentScope 2.0. The runtime layer maps to the actual
2.0.4.post1–2.0.x API — not the fictional HarnessAgent/SequentialPipeline/
EventSystem.on that appear in many design docs. See
docs/agentscope_api_mapping.md.
Zero-config runnable. The core data + engagement layers have no
AgentScope dependency — no LLM key, no Docker. pip install -e ".[dev]" and
pytest is green. The Web UI is one extra.
A real workbench, not a mock. Ship with examples/seed_mock_engagements.py
to demo the console with ten realistic but entirely fictional enterprise
engagements (EV/factory, ride-hailing, retail-tea-chain, outbound-SaaS …),
each with engine-generated gate records, runbooks and corpus reports — see
Mock demo data.

🚀 Quick start

# install (core + dev + web; no agentscope/docker/api-key needed)
pip install -e ".[dev]"

# shell completion + version (Typer built-ins)
fde-scope --install-completion       # bash/zsh/fish
fde-scope --version

# scriptable output: every data command has --json (plain JSON on stdout)
fde-scope deploy --tenant acme --dry-run --json | jq '.permissions.allow'
fde-scope engage list --json

# ── Scenario 1: customer-service tickets (the original) ──
fde-scope connect --type csv --source examples/quickstart_csv/sample_tickets.csv
fde-scope corpus --input examples/quickstart_csv/sample_tickets.csv --out reports/corpus_report.html

# ── Scenario 2: embodied-robotics factory (the new one) ──
fde-scope profiles                                    # see ticket + manufacturing
fde-scope engage init --customer "Aurora Motors" --profile manufacturing
fde-scope kpi <engagement-id> --samples examples/quickstart_manufacturing/station_samples.jsonl
fde-scope gate list --profile manufacturing           # 10 gates, 7 industrial-only

# ── Walk the full 18-phase SOP with gate enforcement ──
fde-scope engage advance <id>          # gate fails → refuses, shows blockers
fde-scope engage status <id>           # phase / zone / gate state / progress
fde-scope handoff <id> --accept        # Zone D: assemble the handoff package

# ── Or do it all in the browser ──
fde-scope web                          # → http://127.0.0.1:8080

Web UI

pip install -e ".[web]"
fde-scope web                          # --host 127.0.0.1 --port 8080 --reload

A single-page engagement console — no build step, backed by JSON APIs:

Engagement dashboard — list / create engagements, advance through the
18-phase SOP (gates block in red when unmet), roll back, inspect blockers.
Gate inspector — evaluate any gate live against an engagement
(/api/engagements/{eid}/gates) with blocker/warning breakdown.
Guided mode（引导） — engagement 详情页「引导」tab，让没有 FDE 驻场的客户
自己走完 SOP：当前阶段白话说明 + 产出清单；gate blocker 翻译成「下一步做
什么」；技能手册库本地深链（/catalog/）。AI 起草 context（现场/干系人/
成功标准/SLO，diff 预览后确认写入）与目标→阶段任务清单（勾选自动保存）；
一切产出仍须过 gate 才算数。无 AI key 时回退手动填写表单。
Context tab — six structured cards per engagement: site survey, stakeholder
map (with SPONSOR badges), success criteria, SLOs, functional-safety posture,
and artifacts (runbook / corpus report / eval metrics / drift monitoring).
Corpus forge — upload a CSV (≤ 10 MiB) and forge an auditable corpus report.
KPI explorer — compute profile-specific KPIs over a sample set.
Report browser — /reports/ serves generated runbooks and corpus HTML.
Agent deploy preflight — 🤖 Agent 部署 view (/console#deploy): render
the per-tenant crew plan (role buckets, bound / unbound tools with reasons,
permission rules) via the same build_deploy_plan — no AgentScope import,
no model call. Linked from the overview page hero and Deploy tile.
Ontology 本体库 — read-only view (/console#ontology): browse the built-in
TBox schemas (classes / properties / SKOS concept trees, cycle-safe rendering),
workspace ABox stores, and preview JSON-LD 1.1 exports via the same
implementation as fde-scope ontology export.
Reading pages, all local — SOP full text (/sop), Level-1 course
(/course), glossary (/glossary), certification centre (/certification),
repo docs (/docs/, README plus 20 docs/ pages rendered locally), the skills
handbook portal (/catalog/) and the GTM site (/GTM/) are all mounted or
rendered in-app — the sidebar has no off-site jumps.

实机截图（macOS App · universal2 DMG，与 fde-scope web 同一控制台）：

Engagement 控制台

功能总览

🧪 Mock demo data

pip install -e ".[dev]"
python examples/seed_mock_engagements.py   # idempotent; re-run to re-seed
fde-scope web                              # open the console, pick any engagement

Seeds .fde_scope/engagements/ with ten realistic engagements. All customer
names are fictional (Aurora Motors, Atlas Auto Alliance, Southbay Motors,
Northern Auto, Swift Ride, Adnova, Teaverse, Horizon Auto, Riverbend Spirits,
Whitelake Dairy) — invented for product demo only, not affiliated with any real
company. Every gate record is produced by the engine
itself (Engagement.evaluate_gate), so the console's gate panels and the
engagement context are always self-consistent. Each engagement also writes a
rendered runbook (reports/<slug>-runbook.md) and corpus report
(reports/<slug>-corpus.html) with real train/eval/test splits.

Customer	Profile	Phase	Story
茶语万象 Teaverse (tea-chain)	ticket	15/15 disengage	Full handoff package, customer accepted
亚特联汽车 Atlas Auto Alliance (automaker)	manufacturing	11/18 runbook	9/9 industrial gates passed (3 shifts + works council + PL d/SIL 2 + HAZOP)
南湾汽车 Southbay Motors (automaker)	manufacturing	9/18 deploy	FAT/SAT + functional safety + conformity genuinely BLOCKED → advance refused
北疆汽车 Northern Auto (automaker)	ticket	13/15 flywheel	Productization flywheel running
迅捷出行 Swift Ride (ride-hailing)	ticket	8/15 prototype	Real-data prototyping, SLO defined
星澜科技 Adnova (outbound SaaS)	ticket	3/15 success_criteria	Single-sponsor blocker — the Sponsor Collapse anti-pattern

✅ Feature checklist

Everything below is executable code with tests — the
architecture guard suite pins the counts,
and docs/architecture-model/architecture-map.md
carries the full evidence index. 完整的中文功能清单(逐项说明、连接器成熟度、
"功能 → 优点"速查表)见 docs/features.md。

SOP engine

[x] 18 phases across 4 zones (pre-engagement → build → operationalization → handoff); ticket runs 15, manufacturing runs all 18
[x] 10 executable gates, live re-evaluated on every advance() — stale passes never count, --force still evaluates and records
[x] On-site journal (research / implementation / optimization) with one-click skill capture

Surfaces

[x] CLI — fde-scope (connect / corpus / deploy / eval / flywheel / engage / gate / skill / ontology / qwenpaw / handoff / kpi / web)
[x] Web console — FastAPI, 86 routes: workbench + engagement dashboard + guided mode（引导 tab：阶段引导 / AI 起草 context / 目标拆解）+ journal + skills + corpus forge + KPI explorer + ontology 本体库
[x] QwenPaw PawApp — desktop plugin, 18 routes under /api/fde-scope, 3 agent tools, skill provider
[x] macOS app — double-click DMG (universal2, signed/notarized release path, same console as fde-scope web)

Data & intelligence

[x] 10 connectors — CSV, Zammad, Salesforce, MySQL, OPC UA, MQTT-Sparkplug, ROS2 bag, MES (ISA-95), Historian, document parsers (PDF/Word/Excel/PPT)
[x] Corpus forge — PII scrub → dedup → quality gate → coverage → gap-targeted synthesis → auditable HTML report
[x] Eval — ticket metrics + industrial KPIs (OEE / MTBF / FPY / DPMO / grasp success / collision rate) + bad-case miner
[x] Flywheel — concept→real event mapping, corpus backflow, retrain scheduler
[x] Skills library — draft → published → archived, AgentScope + QwenPaw dual-format export
[x] Ontology 语义层 — TBox schemas (fde-core + mfg-overlay ISA-95) + SKOS concept scheme + SHACL-lite validation (ONTO-* codes) + JSON-LD 1.1 export; wired into corpus coverage and skill search
[x] Deploy — three-pillar runtime assembly on real AgentScope 2.0 (multi-agent, subagent templates, permission context, honest manifest)
[x] Handoff — runbook + eval report + SLO + training material, customer sign-off package
[x] Optional LLM — MiMo Token Plan or any OpenAI-compatible endpoint, via env var only (get_llm_client() factory); every entry point falls back to deterministic rules

🏗 Architecture

flowchart TB
    subgraph surfaces["Surfaces · 四个入口,同一引擎同一数据根"]
        CLI["fde-scope CLI<br/>13 组命令"]
        WEB["web/ FastAPI 控制台<br/>86 路由 · 单页"]
        PAW["pawapp/ QwenPaw 插件<br/>/api/fde-scope · 18 路由"]
        MAC["macOS App<br/>universal2 DMG"]
    end

    subgraph engine["SOP engine · engagement/ + profiles/ · 零 AgentScope"]
        PH["18-phase 状态机<br/>4 zones · ticket 15 / mfg 18"]
        GT["10 executable gates<br/>每次 advance 实时重评"]
    end

    subgraph layers["Data & intelligence · 五层流水线"]
        L1["1 connectors/<br/>10 连接器"]
        L2["2 corpus/<br/>PII→去重→质量→补盲"]
        L3["3 deploy/<br/>AgentScope 2.0 装配 lazy"]
        L4["4 eval/<br/>工单指标 + 工业 KPI"]
        L5["5 flywheel/<br/>回流 + 重训调度"]
        L1 -->|"samples + schema"| L2
        L2 -->|"语料 + splits"| L3
        L3 -->|"被测 Agent"| L4
        L4 -->|"bad cases"| L5
        L5 -.->|"corpus 回流 · retrain"| L2
    end

    subgraph cross["Cross-cutting · 横切,全部可选"]
        SK["6 skills/<br/>方法论沉淀"]
        INT["7 integrations/<br/>QwenPaw + ACP"]
        ONT["8 ontology/<br/>TBox + ABox + JSON-LD"]
        LLM["llm.py<br/>可选 LLM 出口"]
    end

    CLI --> PH
    WEB --> PH
    PAW --> PH
    MAC --> PH
    PH --> GT
    PH -->|"按阶段调用"| L1
    ONT -.->|"概念对齐:覆盖率 / 技能检索"| L2
    ONT -.-> SK
    SK -.->|"skills_dirs 注册进 Toolkit"| L3
    LLM -.->|"失败回退规则"| L3
    INT -.-> PAW

怎么读这张图

实线 = 主数据流:连接器接原始数据 → corpus 锻成语料 → deploy 装配出真实
Agent → eval 用 KPI 证明价值 → flywheel 把现场学习回流到 corpus(虚线回边,
闭环重训)。五层各自可独立使用,组合起来是完整闭环。
四个表面 → 同一 SOP 引擎:CLI / Web / PawApp / macOS App 操作同一份
engagement 状态与数据根,任何入口的预检与门禁结果零分歧。
虚线 = 可选/横切:ontology、skills、LLM 都不挡主流程——没有它们流水线照样
跑通,这是"零门槛可验证"的架构来源。
PH --> GT 门禁内嵌在状态机里:advance 是唯一推进入口,门禁评估没有旁路。

Layer	Module	Role	agentscope?
SOP	`engagement/`	18-phase state machine + 10 enforceable gates + handoff	❌
Profiles	`profiles/`	Scenario selector (ticket / manufacturing)	❌
1	`connectors/`	CSV, Zammad, Salesforce, MySQL, OPC UA, MQTT-Sparkplug, ROS2, MES, Historian + document parsers	❌ (documents: lazy)
2	`corpus/`	PII scrub, dedup, quality gate, coverage, synthesis, report	❌
3	`deploy/`	Three-pillar runtime assembly (Agent / Permission / Workspace) on real AgentScope 2.0	✅ lazy
4	`eval/`	Ticket metrics + manufacturing KPIs (OEE/MTBF/FPY/…) + bad-case miner	❌
5	`flywheel/`	Concept→real event mapping + collectors + retrain scheduler	❌
6	`skills/`	Methodology capture (draft → published → archived, dual-format export)	❌
7	`integrations/`	QwenPaw bundle export + ACP adapter + manifest validator	❌
8	`ontology/`	Semantic layer: TBox schemas + SKOS concepts + ABox stores, SHACL-lite validation, JSON-LD 1.1 export	❌
Web	`web/`	FastAPI engagement console (86 routes, single-page, JSON APIs)	❌
PawApp	`pawapp/`	QwenPaw desktop plugin (18 routes under `/api/fde-scope`)	❌
Cross	`llm.py` · `config.py` · `templates/` · `paths.py`	Optional LLM egress (fallback), tenant config, Jinja reports, single data-root resolution	❌

Repository layout

fde-scope/
├── fde_scope/            # the Python package — layers map to the table above
│   ├── engagement/       # 18-phase SOP state machine + 10 gates + handoff (no AgentScope)
│   ├── profiles/         # ticket / manufacturing scenario profiles
│   ├── connectors/       # 10 data-source connectors (CSV … Historian + document parsers)
│   ├── corpus/           # PII scrub → dedup → quality gate → synthesis → report
│   ├── deploy/           # three-pillar runtime assembly on AgentScope 2.0 (lazy import)
│   ├── eval/             # ticket metrics + industrial KPIs + bad-case miner
│   ├── flywheel/         # concept→real event mapping + retrain scheduler
│   ├── skills/           # methodology capture (draft → published → archived)
│   ├── integrations/     # QwenPaw bundle export + ACP adapter + manifest validator
│   ├── ontology/         # TBox schemas + SKOS concepts + ABox stores + JSON-LD export (data/*.yaml built-ins)
│   └── web/              # FastAPI engagement console (86 routes)
├── pawapp/               # QwenPaw desktop plugin (18 routes under /api/fde-scope)
├── appbuild/             # macOS app packaging (PyInstaller, universal2 DMG)
├── GTM/                  # product / GTM landing page (single-file HTML)
├── docs/                 # architecture, SOP, API mapping + 101-page skills catalog
│                         #   + github-resources/ (verbatim mirror + static portal)
├── examples/             # quickstart data + seed_mock_engagements.py
├── samples/              # sample datasets
├── reports/              # generated runbooks / corpus reports (produced by corpus & seed scripts)
├── scripts/              # guard scripts (skills-catalog check, …)
├── tests/                # pytest suite incl. tests/test_architecture_guard.py contracts
└── .github/workflows/    # CI + release pipelines

The 10 phase gates

Gate	Applies	Checks
`site_survey`	🏭	Site location required; empty assets / multi-shift without networks → warning
`success_criteria`	🏢	Written criteria + ≥ 2 sponsors; sponsor without a success metric → warning
`fat_sat`	🏭	FAT and SAT both passed, each signed off
`functional_safety`	🏭	Achieved PL/SIL ≥ required; ISO 10218 assessed; hazard analysis mandatory (blocker)
`conformity`	🏭	EU-AI-Act high-risk ⇒ CE marking + technical construction file + hazard analysis
`works_council`	🏭	Works-council approval signed off when representatives exist
`air_gap`	🏭	Offline deployment posture settled at connect time (not phone-home)
`shift_handover`	🏭	Multi-shift sites need digital handover log + per-shift runbook
`slo`	🏢	SLO/SLA + error budget + alert route (missing route → warning)
`handoff_signoff`	🏢	Handoff package: runbook + eval report + SLO + training material, customer accepted; boundary (G7③) / termination (G7④) missing → warning, never a blocker

Gates are executable code, not checklists, and the difference is mechanical
rather than rhetorical:

A checklist is a record; a gate is a predicate. A ticked box states what
was once true. Gate.check(ctx) re-derives the truth from the engagement
context at the moment you try to advance.
They decay in opposite directions. A checklist is freshest the day it is
filled in and rots as reality drifts. Engagement.advance() re-evaluates every
gate of the current phase unconditionally: a stale pass record never grants
passage, because the context may have changed since.
The enforcement point moves from discipline to mechanism. With blockers
present, advance refuses and the state machine cannot move. --force
overrides authority but still evaluates and records, so even exceptions leave
an audit trail instead of becoming indistinguishable from a clean pass.

flowchart TD
    A["fde-scope engage advance(可带 --force)"] --> B["evaluate_phase_gates(ctx)<br/>当前阶段所有门禁 · 无条件实时重评"]
    B --> C{"blockers 为空?"}
    C -->|"是"| D["✅ 记录 pass → 放行进入下一阶段"]
    C -->|"否"| E{"带了 --force?"}
    E -->|"否"| F["⛔ 拒绝推进 · 返回 blockers 清单<br/>自动生成 gate_hint 技能草稿(skill review 完善)"]
    E -->|"是"| G["⚠️ 强制放行 · 仍评估并记录<br/>例外留下审计痕迹,不混同于干净通过"]

门禁执行流:advance 是唯一推进入口,评估内联其中——每一条路径都会评估并记录,
包括被 --force 打开的那条。

Concretely, Engagement(ctx).evaluate_gate(slug) returns
(passed, blockers, warnings) and blocks advance until clean.
See docs/architecture.md and
docs/fde_sop_full.md.

📐 Engineering best practices (enforced, not aspirational)

Seven invariants every change must preserve — AGENTS.md is the
machine-readable source, and the guard tests fail when they drift:

Gate re-evaluation is live. Engagement.advance() re-evaluates the
current phase's gates on every advance; no caching or short-circuits
around evaluate_phase_gates.
IDs are server-generated. SkillRecord.id / EngagementContext.id are
assigned by the owning service — external input never sets them.
Credentials live in env vars only. Keys flow through
fde_scope/config.py; never persisted into manifests, reports,
engagement JSON or logs.
Persisted records are written atomically. Every on-disk user record
goes through fsutil.atomic_write_text (temp file + os.replace).
Rule grants are the only permission channel. build_toolkit never
flags a deployed tool is_read_only (upstream ≥2.0.5 auto-allows read-only
calls); unmatched tools fall back by mode: DEFAULT→ASK, DONT_ASK→DENY.
The AgentScope window is measured, not aspirational. pyproject.toml
and docs/agentscope_api_mapping.md must quote the same specifier
verbatim (>=2.0.4.post1,<3 today); re-measure per version before widening.
Web mutating 路由必须过认证依赖。 新增 web/ 路由时确认未被 auth
白名单（fde_scope/web/auth.py 的 PUBLIC_PATHS）误放；token 只从
FDE_SCOPE_API_TOKEN env 读。认证关闭只对回环生效：未设 token 时，
从非 loopback 地址到达的请求必须 403（中间件逐请求判 scope["server"]，
覆盖绕过 CLI 直接起服务的情形），CLI 另外拒绝无 token 的非回环绑定。

Verification anchors: make test ·
pytest tests/test_architecture_guard.py (contract pins) ·
pytest -m agentscope (real-library runtime) ·
ruff check && ruff format --check · data-root contract in tests/test_paths.py.
Architecture evidence model: docs/architecture-model/architecture-map.md.

🛠 CLI reference

connect   [Layer 1] Connect to a data source and preview schema + samples
corpus    [Layer 2] Forge a raw sample into an auditable, gap-aware corpus
deploy    [Layer 3] Assemble (and optionally start) a multi-tenant agent (multi-agent via --agent)
eval      [Layer 4] Run the FDE benchmark over a test set
flywheel  [Layer 5] Start (or replay into) the data flywheel
handoff   [Zone D]  Assemble the handoff / knowledge-transfer package
kpi       Profile-specific KPIs over a sample set
profiles  List available deployment scenario profiles
web       Launch the Web UI (workbench + skills + engagement console)
engage    [SOP] init / status / advance / rollback / journal / list
gate      [SOP] list (per profile) / check (per engagement)
skill     [Skills] 技能/方法论沉淀库（add / list / show / edit / publish / archive / review / export）
ontology  [Ontology] 本体语义层（list / validate / check / export — JSON-LD 1.1）
qwenpaw   [QwenPaw] Export / validate QwenPaw-compatible bundles (agents + skills + corpus)
license   [商业] 查看当前许可 tier / 到期 / 功能开关；license-issue 签发客户 key
storage-migrate [存储] engagement 持久化 file → sqlite 单向迁移

🤖 多 Agent 生产（AgentScope 2.0）

装配即真实 Agent。 tenant_config.yaml 的 agents 段（或 CLI
--agent 名字:角色[:模型]）声明一个 tenant 的多个 Agent；fde-scope deploy
为每个 spec 组装一个真实的 agentscope.agent.Agent：模型 wiring（spec.model，
缺省留运行时注入）+ 角色工具真 Toolkit + AgentState(permission_context=...)
规则注入 + ReActConfig(max_iters=20)，manifest 携带完整 agents 段与
subagent_templates。

FDE 自己开 Agent，三个通道（名字任意，中文可用）：

# 通道 1 · CLI：--agent 可重复；角色是自由文本，自动归入 数据/日志/文件 工具桶
fde-scope deploy --tenant acme \
  --agent 数据员:数据分析 --agent 日志员:日志分析:qwen3-14b \
  --connector-source csv=examples/quickstart_csv/sample_tickets.csv
# 通道 2 · YAML：TenantConfig.agents 声明 + TenantConfig.sources 绑连接器数据源
# 通道 3 · 单 Agent 覆盖：spec.toolkit["sources"] / ["skills_dirs"]
#          （该 Agent 专属数据源与挂载技能，技能经 Toolkit(skills_or_loaders=...) 注册）

角色未匹配任何桶时退化为 corpus-only；没配数据源的连接器工具在 manifest 里
诚实标注 unbound 及原因，不假装可用。

协调走 2.0 官方机制：每个 Agent 同时导出 agentscope.app.SubAgentTemplate
蓝图进真实 create_app(custom_subagent_templates=...)；leader 通过
AgentCreate 按 subagent_type 派活、成员经 TeamSay 回报（CI 真库测试覆盖）。

装配 ≠ 服务：不带 --serve，deploy 输出真对象装配 + 诚实 manifest
（dry-run / 审计友好）；deploy --serve（build_app() + uvicorn）起真实
多 Agent 服务，需要 .[agentscope] extra + Redis + 可达模型。生命周期：
2.0 无 Agent.stop，TenantDeployer.stop() 关闭 workspace/engine 句柄并
标记 manifest。

子 Agent 配置方法速查

配置项	位置	说明
Agent 列表	`TenantConfig.agents` / CLI `--agent`	每条 `AgentSpec(name, role, system_prompt?, model?, toolkit?)`
数据源	`TenantConfig.sources` / CLI `--connector-source slug=value`	connector slug → 物理源（文件路径 / DSN / URL）
单 Agent 覆盖	`spec.toolkit["sources"]` / `["skills_dirs"]`	该 Agent 专属数据源与挂载技能
审批策略	`TenantConfig.approval_policy.mode`	`conservative` / `balanced` / `autonomous`
预检	Web `POST /api/deploy/plan` · PawApp `/deploy/plan` · `deploy --dry-run`	同一 `build_deploy_plan`，纯数据不 import AgentScope

注意事项

角色是自由文本，但决定工具桶。 canonical_role() 按中文关键词把角色
归入数据/日志/文件桶并派生连接器集；匹配不上任何桶时退化为
corpus-only（只有语料工具）。想要某个连接器，角色词要沾边或显式覆盖
spec.toolkit。
数据源解析有优先级：per-agent spec.toolkit["sources"] → tenant
sources → 隐式字段（ticket profile 下 ticket_api 隐式喂给
zammad/salesforce）。三处都没配的连接器工具诚实标注 unbound 及原因，
不进 Toolkit、不假装可用——先看 plan 再上现场。
审批模式决定未匹配工具的命运：绑定的工具自动获得显式 ALLOW 规则；
未匹配工具在 conservative/balanced 下走 DEFAULT→ASK（HITL 确认），
autonomous 下走 DONT_ASK→DENY。默认 deny 恒含
access_other_tenant / delete_any / exec_shell。绑定工具不标
read-only（上游 ≥2.0.5 会先于 allow 规则自动放行只读调用）。
skills_dirs 必须指向真实存在、含 SKILL.md 的目录；技能经
Toolkit(skills_or_loaders=[...]) 注册，路径错了装配期即可发现。
凭据只走环境变量（FDE_SCOPE_MIMO_API_KEY 等）——永远不进
tenant_config.yaml / manifest / 报告。
连接器 sample 工具每调用最多返回 MAX_TOOL_ROWS = 50 行，大表预览
先小后大，不要拿它当导出通道。
--serve 不是默认：需要 .[agentscope] extra（窗口
>=2.0.4.post1,<3，扩窗前先实测）+ Redis 后端 + 可达模型；纯装配 /
dry-run / plan 零 AgentScope 依赖。2.0 无 Agent.stop，停服用
TenantDeployer.stop()。

入口（三个渠道，同一 build_deploy_plan，输出零分歧）：

Web 控制台 — fde-scope web → /console#deploy「🤖 Agent 部署」：
表单点名 Agent / 选 profile / 填数据源，直接出预检清单；总览页 hero 按钮
与 Deploy 模块卡都链向此处；
macOS App（QwenPaw PawApp） — 同款「🤖 Agent 部署」标签页，走
/api/fde-scope/deploy/plan，与 Web 控制台同一引擎；
CLI — fde-scope deploy --tenant acme … --dry-run，同一预检的命令行形态。

🐾 QwenPaw 集成

fde-scope qwenpaw export --tenant acme --agent "researcher:调研员" --out qwenpaw-out
把 tenant 的 agent 拓扑导出为 QwenPaw 兼容产物（config.json 的 agents.profiles
多 Agent 声明 + workspaces/{agent_id}/agent.json + AGENTS.md persona +
published 技能包 + corpus 说明），qwenpaw validate --out ... 按官方规则校验
（必填字段 / agent id 规则 / SKILL.md frontmatter）。macOS App（PawApp）
内的「🤖 Agent 部署」标签页提供同款多 Agent 预检。ACP 适配：AcpEndpoint
基类把 FDE 能力暴露为 QwenPaw ACP runner（delegate_external_agent）。
详见 docs/qwenpaw_integration.md。

💡 Skill 沉淀（Skills / Methodology capture）

FDE 在现场的每一步（调研、实施、调优、方法论）都可以沉淀为可复用的技能，跨 engagement、跨客户复用。技能存为文件库 .fde_scope/skills/（每个技能一个目录：skill.md + meta.json），随项目携带，零数据库依赖。

三种沉淀入口：

入口	触发方式	说明
手动	`fde-scope skill add`	随时沉淀，四类分类（research / implementation / optimization / methodology）+ 标签
gate 阻塞提示	`engage advance` / `gate check` 失败时自动	生成预填草稿（gate_hint），`skill review` 完善
操作自动捕获	`engage advance` 成功 / `gate check` 通过时自动	轻量记录（auto_capture），`skill review` 完善

生命周期： draft → published → archived。skill review 审阅草稿队列 → skill publish → 可检索、可导出。

导出给 Agent： 支持 AgentScope 2.0 与 QwenPaw 两种格式（Anthropic Agent Skills 规范：<skill-name>/SKILL.md，frontmatter 含 name + description）。AgentScope 侧经 Toolkit(skills_or_loaders=[...])（部署装配的 AgentSpec.toolkit.skills_dirs）注册进真实 Toolkit；QwenPaw 侧放入 customized_skills 目录自动发现。

fde-scope skill add --title "OPC UA 连接踩坑" --category implementation --tags opcua --body "# 步骤..."
fde-scope skill review                  # 审阅自动捕获/gate 提示产生的草稿队列
fde-scope skill publish <skill-id>
fde-scope skill list --category implementation --tag opcua
fde-scope skill export <skill-id> --format qwenpaw --out exports/   # 或 --format agentscope

Web UI 暴露完整的 skills API：GET/POST /api/skills、GET/PATCH /api/skills/{id}、POST /api/skills/{id}/publish|archive|export、GET /api/skills/drafts。

🖥️ FDE 工作台（Workbench + 现场记录）

fde-scope web 打开的单页控制台现在是一个跨项目工作台：

工作台首页（/console）：全局统计条（进行中项目 / 阶段分布 / 待审草稿 / 技能总数）、
跨项目矩阵（客户/阶段/zone/门禁/最近更新，点击进入详情）、最近沉淀技能；
技能库页（/console#skills）：搜索 + 分类/状态筛选 + 技能卡片 +
新建技能表单 + 草稿审阅队列（一键发布/编辑）；
现场记录（详情页 tab）：research / implementation / optimization 三类记录，
任意一条可一键“沉淀为技能”（kind → category 自动映射、关联来源 engagement）。

CLI 等价入口：fde-scope engage journal <id> --kind research --note "..." [--link-skill <sid>]。

🤖 LLM 接入（可选）

核心层零 LLM、零额外依赖：语料合成、质量评分、eval 回复、runbook 生成全部默认走确定性规则路径，只有设置了环境变量才启用 MiMo Token Plan（OpenAI 兼容协议，标准库 urllib 直连，无新依赖）。

# Token Plan 凭据（tp- 前缀，与按量付费 sk- 不通用）
export FDE_SCOPE_MIMO_API_KEY="tp-..."
# 可选：默认已是 https://token-plan-cn.xiaomimimo.com/v1 与 mimo-v2.5-pro
export FDE_SCOPE_MIMO_BASE_URL="https://token-plan-cn.xiaomimimo.com/v1"
export FDE_SCOPE_MIMO_MODEL="mimo-v2.5-pro"

# 或：任意 OpenAI 兼容端点（vLLM / OpenAI / DeepSeek …，标准 Bearer 头）。
# 两个变量都设置时优先于 MiMo；模型缺省 gpt-4o-mini
export FDE_SCOPE_LLM_BASE_URL="http://localhost:11434/v1"
export FDE_SCOPE_LLM_API_KEY="sk-..."
export FDE_SCOPE_LLM_MODEL="qwen3:14b"

# 日志级别（默认 WARNING）：DEBUG 可看到 gate 评估/forge 细节
export FDE_SCOPE_LOG_LEVEL="INFO"

# Zammad 连接器（ticket profile）：两个都设置才启用真实 HTTP API
# （/api/v1/tickets 分页拉取）；只设一个或都不设时回退本地 .jsonl source
export FDE_SCOPE_ZAMMAD_BASE_URL="https://support.example.com"
export FDE_SCOPE_ZAMMAD_API_TOKEN="..."

# Salesforce 连接器（ticket profile）：SOQL 只读查询 Case 对象；
# access token 直配（在进程外获取），两个都设置才启用真实 REST API
export FDE_SCOPE_SF_INSTANCE_URL="https://acme.my.salesforce.com"
export FDE_SCOPE_SF_ACCESS_TOKEN="..."

# MES / Historian 连接器（manufacturing profile）：两个变量都设置才启用真实 REST API
export FDE_SCOPE_MES_BASE_URL="https://mes.example.com"
export FDE_SCOPE_MES_API_TOKEN="..."
export FDE_SCOPE_HISTORIAN_BASE_URL="https://historian.example.com"
export FDE_SCOPE_HISTORIAN_API_TOKEN="..."

# 商业许可（可选）：不设置即 community tier；签发走 fde-scope license-issue
export FDE_SCOPE_LICENSE_SECRET="..."   # 签发方持有，验签用
export FDE_SCOPE_LICENSE_KEY="..."      # 客户侧设置

# 存储后端（可选）：默认 file（JSON 原子写）；sqlite 启用 WAL 单文件库
export FDE_SCOPE_STORAGE="sqlite"

# 飞轮训练后端（可选）：不配置则 noop（如实标注）
export FDE_SCOPE_TRAINING_URL="https://training.example.com"
export FDE_SCOPE_TRAINING_TOKEN="..."

所有 LLM 入口都是失败自动回退规则路径——网络错误、坏 JSON、无 key 都不会中断流水线：

入口	启用方式	效果
语料合成	`fde-scope corpus --input … --llm`	每个缺口类别由 LLM 生成真实感样本（参考种子风格，JSON 输出）
质量评分	同上（合成后自动）	LLM 对样本打 1-5 质量分，规则分兜底
Eval 基准	`fde-scope eval --agent mimo --test-set …`	用 MiMo 充当被评测的客服 Agent（`MiMoReplyFn`）
部署清单	`fde-scope deploy --tenant …`	manifest 携带 `llm` 段（provider/model/base_url）
Runbook	`fde-scope handoff <id> --llm`	按客户/profile/阶段/SLO 生成 runbook，失败回退模板
Web 控制台	自动	`/api/forge` 有 key 即启用 LLM 合成，无需配置

FDE_SCOPE_MIMO_API_KEY 未设置时：corpus --llm / eval --agent mimo / handoff --llm 会明确报错（exit 2）提示配置，其余命令静默走规则模式。凭据只经环境变量传递，请勿写入任何 git 文件。

协议细节、六个接入点的回退语义、成本模型、测试策略与多 provider 扩展指南见
docs/llm_integration.md；LLM 路径的加固记录与评审案例库见
docs/code_review_checklist.md。

🧪 Tests

pip install -e ".[dev]"
pytest                 # CI badge (top) shows current pass/skip status.
                       # Coverage: core + engagement + gates + skills + web + integrations + LLM mock + PawApp.
                       # No agentscope/mysql/opcua/LLM-key/QwenPaw needed locally (skips are integration-only);
                       # real-AgentScope runtime tests live under the `agentscope` marker (CI dedicated job).
pytest -m agentscope   # 仅真库运行时测试（需 pip install -e ".[agentscope]"）
pytest --cov=fde_scope --cov-report=term-missing   # ~89% coverage

CI (GitHub Actions) runs the same suite on Python 3.11 + 3.12 with ruff lint
(check + format --check), mypy, and an sdist/wheel build — see
.github/workflows/ci.yml.

📦 Installation

pip install -e "."                 # core only
pip install -e ".[dev]"            # + pytest + web test client
pip install -e ".[web]"            # + FastAPI/uvicorn console
pip install -e ".[agentscope]"    # + real AgentScope 2.0 runtime layer
pip install -e ".[mysql]"          # + MySQL connector driver
pip install -e ".[opcua]"          # + OPC UA connector driver (asyncua)
pip install -e ".[mqtt]"           # + MQTT-Sparkplug live broker driver (paho-mqtt)
pip install -e ".[ros2]"           # + ROS2 bag replay driver (rosbags)
pip install -e ".[full]"           # everything (dev + agentscope + mysql + opcua + mqtt + ros2 + web)

Requires Python ≥ 3.11. Entry point: fde-scope (or python -m fde_scope.cli).

macOS App（双击即用）

./appbuild/build_release.sh          # universal2 app + DMG → appbuild/dist/

产出 ad-hoc 签名的 FDE-Scope-<ver>-universal2.dmg（含同版本 .zip、SHA256SUMS 与
manifest.json 溯源清单；单实例守护、健康检查、崩溃兜底、SIGTERM 优雅退出）。
ad-hoc 包首次打开被 Gatekeeper 拦时右键 → 打开；配置 DEVELOPER_ID_P12_* 与
NOTARY_API_KEY_* 后脚本自动改为 Developer ID 签名 + 公证。推送 v* tag 时
CI 自动出包（.github/workflows/release.yml）。

与 Web UI / QwenPaw PawApp 的关系： Mac App 没有自己的界面——
launcher.py 直接加载 fde_scope.web.app 的 FastAPI
对象跑在 127.0.0.1:8737，模板就是 fde_scope/templates 同一份，页面与
fde-scope web 完全一致；它只是加了一层双击入口（打包运行时 + 单实例守护 +
DMG 分发）。QwenPaw PawApp（pawapp/）则是同一引擎、同一数据根
目录之上的另一个原生前端：自带 APIRouter（宿主挂载 /api/fde-scope，
18 条路由）与 ui/index.js 面板，不复用 Web 页面。一句话：Mac App = web
控制台换个双击入口；PawApp = 同一引擎的另一个前端。

📄 Documentation

docs/enhancement-plan.md — 全面加强计划与进度台账：逐页逐功能加强项、每条带 file:line 证据、分期路线图（详见加强计划）
docs/features.md — 完整功能清单：18 阶段逐条说明 / 10 门禁执行流 / 连接器成熟度 / "功能 → 优点"速查表（English: docs/features-en.md）
docs/user-guide.md — 客户侧快速上手：5 分钟跑通、核心概念、常用操作、env 速查、故障 FAQ
docs/api-reference.md — Web 控制台 REST API 参考（认证 + 全路由逐条对照代码）
docs/onboarding-customer.md — 客户现场 onboarding：部署形态、安全基线、数据/LLM 接入决策树、POC 验收模板
docs/fde_sop_full.md — the 18-phase SOP, 12 anti-patterns, sources
docs/manufacturing_scenario.md — embodied-robotics factory end-to-end walkthrough
docs/architecture.md — layered design + data flow
docs/architecture-model/architecture-map.md — evidence-backed architecture views (L1/L2/L3) + health report + canonical DSL/DOT sources
docs/skills.md — 技能沉淀系统：数据模型 / 生命周期 / 三种入口 / 双格式导出
docs/skills-catalog/README.md — 101 页 skill 手册库（Zone A–D + 横切 + MCP + 工具），门户 site/index.html，make check-catalog 门禁 / make build-site 重建
GTM/index.html — 产品 / GTM 落地页（仓库内单文件：定位、功能/架构/实践清单、实机截图）
docs/llm_integration.md — MiMo Token Plan integration: 6 touchpoints, fallback semantics, cost model, testing
docs/qwenpaw_integration.md — QwenPaw 集成：配置两层结构 / 导出与校验 / PawApp 桌面形态
pawapp/README.md — QwenPaw PawApp 安装/开发/API 契约
docs/code_review_checklist.md — repeatable review workflow + bug case library
docs/agentscope_api_mapping.md — design-doc fiction vs. real 2.0.x API (measured version matrix)
docs/fde_playbook.md — on-site 72h playbook
examples/README.md — CSV cold-start walkthrough (quickstart data included)

🧭 加强计划（Enhancement Plan）

完整台账见 docs/enhancement-plan.md：逐页逐功能的加强项，每条都带
file:line 证据与验收动作。这里只放骨架与当前状态。

基线（实测，非估计）：867 passed / 4 skipped / 92% 覆盖率 / 20.6s；ruff check 绿；
ruff format --check 与 mypy fde_scope 红（CI 阻塞步骤）；Web 路由真实 57 条而文档写 56。

P0 收口后复测（同组命令）：876 passed / 4 skipped / exit 0（+9 条新守卫）；
ruff format --check → 174 files already formatted；mypy fde_scope → Success: no issues found in 101 source files；
路由数 11 处宣称全部对齐 57，且装饰器扫描与 app.openapi() 逐条对账；
干净 venv 装 wheel 后 /course 正常渲染（zh 770,509 字符），发行版不再空页。

本轮最高价值发现：evidence / handoff / audit / archive / delete 五项能力
后端已实现、测试已覆盖，但控制台前端从不调用；rollback 更是领域层与 CLI 都有、Web 层连端点都没有
——而 docs/features.md 已经宣称看板支持「回滚」。所以 P1 几乎全是「给已测后端接前端」，
不引入新领域逻辑，风险最低、收益最高。

期	状态	工作量	范围	出口标准
P0 止血	✅ 已完成（9/9）	0.5–1 天	9 项：格式/类型红、GTM 图片 401、GTM GitHub 假链、`course-level-1/` 未打包、路由守卫漏 `patch`、favicon、嵌套 SVG、误提交的 `.apk.idsig`	四条回归锚全绿 ✅
P1 补齐已实现能力	✅ 已完成（§1.1–§1.3）	3–5 天	回滚端点、11 tab、四确认框、skills 四筛选+facets、deploy 服务端解析、ontology 三子视图、Context JSON 编辑器、workbench 筛选、运维视图（连接器/eval/flywheel/license，env 只报布尔）、flywheel CLI 假计划修复、鉴权全量枚举守卫、favicon 死链修复（落地记录见计划 §1.1–§1.3）	每项都有 UI 和测试；文档宣称变成事实 ✅
P2 页面质量	✅ 已完成	5–8 天	glossary 搜索+转义、course 进度/勾选/搜索/token 流统计、sop token rail+EN 标签、docs TOC+href 白名单+mdrender 抽取、catalog a11y+暗色+⚰️ chip、首页磁贴注册表化、i18n 键漂移守卫、CLI C1–C12（`--version`、9 命令 `--json`、spinner、友好错误、公共 agent 解析/门禁访问器）（落地记录见计划 P2 节）；C13 help 全文双语留独立批次	每页至少一项导航增强 ✅；CLI 可脚本化 ✅
P3 工程化	🔄 主体完成（Q2/Q4/Q5/§12 部分待立项）	8–12 天	✅ 覆盖率门禁（fail_under=90，实测 92.08%）、未测路由/分支补齐（连带修 context 字段路径缺陷）、macOS CI 冒烟、文档零漂移（5 处）；⬜ CI service containers、前端 JS 抽出+DOM 测试、`app.py` 拆分、PawApp 共享层（落地记录见计划 P3 节）	门禁生效 ✅、文档不再漂移 ✅、巨石拆分 ⬜
P4 独立立项	[-] 本轮不做	—	多租户/RBAC/SSO、PII 字段级加密与保留期、数字签名、真实训练管道、真实连接器 POC	不在本轮

红线：所有加强都受 AGENTS.md 七条不变式约束。计划文档 §15 逐条列出
「哪条不变式约束了哪条加强」，以及本轮明确不做的 6 件事（含「审计证据是 sha256 文件哈希，
不是数字签名，不得在文档里说成签名」）。

🤝 Contributing

Issues and pull requests are welcome. Fastest path to a green PR:

pip install -e ".[dev]"
pytest -q                                    # full suite
ruff check fde_scope tests && ruff format --check fde_scope tests
pytest tests/test_architecture_guard.py      # contract guards

Ground rules — AGENTS.md is the machine-readable source:

Don't weaken the seven invariants above; the guard tests fail on purpose when they drift.
engagement/phases.py (18 phases) and the default gate registry (10 gates) are contract-pinned: changing them requires a deliberate contract update and a description in docs/.
Credentials flow through environment variables only — never persisted into manifests, reports, engagement JSON or logs.
Large or risky changes: open an issue / draft PR first so the invariants can be discussed before code lands.

🗺 Roadmap

[x] OPC UA 真实工业 IO（asyncua 驱动，mock 测试覆盖）
[x] Category-stratified train/eval/test split（CorpusReport.splits）
[x] MQTT-Sparkplug 真实 broker IO（paho-mqtt 驱动，mock + 真 broker 测试）
[x] rosbag2 真实回放（rosbags）
[x] Ontology 模块（横切语义层：TBox + ABox、SKOS 概念体系、SHACL-lite 校验 ONTO-* 错误码、JSON-LD 1.1 标准导出、零新依赖）— 设计文档 · 使用指南
[x] Real LLM corpus synthesis & quality scoring (drop-in behind existing signatures)
[x] 小米 MiMo Token Plan 接入（fde_scope/llm.py，env 配置，失败回退规则）
[x] Skill 沉淀库（四类分类 + 生命周期 + AgentScope/QwenPaw 双格式导出）
[x] 多 Agent 生产（真实 Agent 装配 + 模型 wiring + SubAgentTemplate 蓝图 + FDE 三通道自开 Agent）
[x] QwenPaw 集成（qwenpaw export/validate + PawApp 桌面应用，真机验证通过）
[x] deploy --serve 端到端实测（Redis 后端 + fake OpenAI 兼容模型，2026-09-28 本机 PASS，agentscope 2.0.8；见 deploy/e2e/）
[x] Zammad HTTP API 连接器（/api/v1/tickets + Token 认证，env 配置，JSONL 回放兜底）
[x] Salesforce REST 连接器（SOQL query + Bearer 认证，只读，env 配置，JSONL 回放兜底）
[x] Full MES / Historian HTTP/SQL implementations（MES work-orders REST、Historian 时序 samples REST，均 env 配置 + JSONL 回放兜底）
[x] 离线 License 模块（HMAC-SHA256 签名 key，community/pro/enterprise 三档功能开关，fde-scope license / license-issue，web 层审计导出与证据留存门控 pro+）
[x] 存储后端抽象（FDE_SCOPE_STORAGE=file|sqlite；SQLite 走 stdlib + WAL，fde-scope storage-migrate 迁移工具）
[x] 飞轮训练后端接口（TrainingBackend Protocol：HTTP webhook 后端 env 配置 + 显式 noop 兜底）
[ ] AgentScope Studio (npm @agentscope/studio) integration

📄 License

MIT.

🙏 Acknowledgements

AgentScope 2.0 — the multi-agent runtime under the deploy layer (three-pillar assembly).
QwenPaw — the desktop PawApp host targeted by our plugin (see QwenPaw 集成).
MiMo Token Plan — optional LLM egress, env-var configured with deterministic fallback (see LLM 接入).
Ruff — lint & format enforcement across the codebase.