FDE · Field Delivery Engine

The complete field workbench for FDEs

From the first Gemba walk to signed handoff — covering software/SaaS and embodied-robotics/manufacturing scenarios. An 18-phase SOP, 10 executable compliance gates, 10 data connectors, real production KPIs.

Open the engagement console Agent deploy preflight Quick start
18
SOP phases (4 zones)
10
Executable gates
10
Data connectors
CI
Tests green

Full SOP · 18 phases · 4 zones

01
Zone A
Pre-engagement
  1. qualification Problem framing
  2. site_survey Gemba
  3. stakeholder_map Dual sponsor
  4. success_criteria Contracted
Zone B
Build
  1. connect Data ingestion
  2. corpus Corpus forging
  3. prototype Real-data prototype
  4. validate Validation
  5. deploy FAT/SAT
  6. eval Evaluation
Zone C
Industrialization
  1. slo_sla SLO+on-call
  2. runbook Emergency runbook
  3. monitoring Drift detection
  4. change_mgmt Works council
  5. flywheel Flywheel productization
Zone D
Handoff & exit
  1. ops_handoff Ops handoff
  2. knowledge_transfer Knowledge transfer
  3. disengage Signed exit

10 executable compliance gates (industrial overlay)

02

A gate is a predicate, not a checklist: checklists record "was once checked" and rot one-way as reality drifts; gates re-evaluate on every advance and a stale pass never counts. --force only waives the block while still evaluating and recording; exceptions land in the audit log.

site_survey
Site-survey record check
success_criteria
Dual sponsor + measurable done
fat_sat
FAT/SAT acceptance sign-off
functional_safety
ISO 13849 / IEC 61508 / ISO 10218
conformity
CE / EU AI Act conformity
works_council
German BetrVG §87 works-council co-determination
air_gap
Air-gapped deployment checklist
shift_handover
24/7 shift-handover integration
slo
SLO + on-call definition
handoff_signoff
Handoff-package sign-off

10 data connectors

03
CSV
Universal fallback, cold start
Live
MySQL
Direct relational-DB access
Live
MQTT-Sparkplug
Factory device telemetry
JSONL
MES (ISA-95)
Work orders / quality / downtime
Live
OPC UA
PLC tag reads (asyncua driver)
Live
ROS2 Bag
Robot trajectory replay
stub
Historian
Time-series historian
Live
Zammad
Ticketing system
Live
Salesforce
CRM
Live
Documents
PDF/Word/Excel/PPT parsing (agentscope.rag, lazy import)
needs [agentscope]

Ontology semantic layer · TBox + ABox

04
Dual TBox schemas
fde-core (Skill/Connector/Engagement core concepts) + mfg-overlay (ISA-95 manufacturing overlay, reused via imports)
Built-in
SKOS concept schemes
Two-way fde-corpus-taxonomy integration: the corpus engine samples by concept coverage, skill search expands recall via concepts
Built-in
SHACL-lite validation
ONTO-* error-code contract (domain/range, cycles, undeclared prefixes); a failed check blocks
10 error codes
JSON-LD 1.1 export
schemas export directly; store export merges its TBox context; CLI and Web API share one implementation.→ Ontology
zero new deps

6 functional modules

05
Corpus Engine
PII masking → dedup → quality gate → coverage analysis → gap detection → targeted synthesis → report. Core differentiator: synthesis fills blind spots, it doesn't pad counts.
Eval
ticket metrics + manufacturing KPIs (OEE/MTBF/pick rate/collision rate) + bad-case mining + auto suggestions.
Deploy
Real role-agent assembly: sandboxed workspace + permission context (injected via AgentState) + Toolkit tool binding + SubAgentTemplate blueprints — all real AgentScope 2.0 APIs.→ Deploy preflight
Flywheel
Concept-event → real-event mapping + corpus backflow + weekly incremental retraining.
SOP state machine
18-phase advance/rollback; a failed gate blocks.
Web console
Interactive engagement dashboard + gates + forge + KPIs + agent deploy preflight (/en/console#deploy) + ontology (/en/console#ontology).

Architecture map · 4 layers + cross-cutting

06

Evidence-based modeling: docs/architecture-model/.

UI layer
CLI (Typer, fully lazy imports) · Web console (43 routes) · QwenPaw PawApp (18 routes /api/fde-scope) · macOS app (double-click DMG — this console).
SOP layer
engagement/: 18 phases · 4 zones · 10 executable gates (re-evaluated live on advance) + handoff; profiles/: ticket · manufacturing scenario selectors.
Capability layer
connectors · corpus · deploy · eval · flywheel · integrations · ontology · skills — the core data pipeline, rules-first with optional LLM enhancement.
Cross-cutting
llm.py (the only LLM egress, falls back to rule paths) · config.py · templates/ (Jinja reports and runbooks) · paths.py (single data_root path resolution).
Files as source of truth
Zero database: .fde_scope/ (engagements · skills · uploads) + reports/; every write goes through fsutil.atomic_write_text.
External deps (all optional)
AgentScope 2.0.x (extra, measured window >=2.0.4.post1,<3; deploy pillars + lazy documents) · MiMo LLM (credentials via env vars only) · QwenPaw host.

Engineering best practices · seven invariants

07

The AGENTS.md contract, pinned by architecture-guard tests.

1Live gate re-evaluation
advance() re-evaluates every gate of the current phase each time; stale passes never count; forced advances still evaluate and record. No caching or short-circuits.
2Server-generated IDs
SkillRecord.id / EngagementContext.id are assigned by the owning service; external input can never set them.
3Credentials via env vars only
API keys never land in manifests / reports / engagement JSON / logs.
4Atomic writes
All user state goes through fsutil.atomic_write_text (temp file + os.replace); bare write_text is banned.
5Rule grants are the only channel
build_toolkit never flags is_read_only (upstream ≥2.0.5 auto-allows read-only first); unmatched tools fall back by mode: DEFAULT→ASK / DONT_ASK→DENY.
6Measured version window
pyproject and docs quote the same agentscope specifier verbatim; widen only after per-version runs against the real library.
7Web mutating routes require auth
New routes must not slip into the auth allowlist by mistake; the token is read from the FDE_SCOPE_API_TOKEN env var only — never in files / logs / responses.

Verification anchors: make test · pytest tests/test_architecture_guard.py (6 contracts) · pytest -m agentscope (real-library runtime) · ruff check + format --check · architecture evidence model docs/architecture-model/architecture-map.md

Manufacturing KPIs (fictional demo data)

08
OEE
Overall Equipment Effectiveness0.724
World-class ≥0.85
Pick success
pick success0.793
DexNet benchmark ~0.80
MTBF
Mean time between failures (h)39.8
Higher is better
Collision / intervention rate
per cycle1.08%
Lower is better

Quick start

09
CLI
# Install (core layer has zero deps)
pip install -e ".[dev]"

# Start a manufacturing engagement
fde-scope engage init --customer aurora-motors --profile manufacturing

# Advance the SOP (gate enforcement)
fde-scope engage advance <id>
fde-scope gate check <id> --gate fat_sat

# Corpus forging + KPIs
fde-scope corpus --input alarm.jsonl --out r.html
fde-scope kpi <id> --samples station.jsonl

# Generate the handoff package
fde-scope handoff <id> --accept
Web UI
# Start the interactive console
pip install -e ".[web]"
fde-scope web

→ http://127.0.0.1:8080

# Console features:
· Create an engagement (pick a profile)
· 18-phase visualization + gate blocking
· CSV → corpus forging (HTML report)
· KPI computation (ticket / manufacturing)
· Advance/roll back the SOP state machine
· Agent deploy preflight (/en/console#deploy)

Two scenarios

10
ticket · support
CSV/Zammad/Salesforce/MySQL ingestion. Intent accuracy, reply adoption and escalation-rate eval. No industrial gates.
Fully available
manufacturing · embodied robotics
OPC UA/MQTT/ROS2/MES/Historian ingestion. OEE/MTBF/pick rate/collision rate + 6 industrial compliance gates.
Core available · industrial IO partly stub