| name | cursor-maas-delegate-router |
| description | USER-GLOBAL hybrid routing for Cursor (all workspaces): after install, writes ~/.cursor/rules + ~/.cursor/memory + ~/.cursor/hooks (sessionStart + beforeSubmitPrompt). Code execution (HTML/CSS/JS, Hello World, greenfield, 输出代码, tests, docs, CI) MUST use delegate.py → Huawei MaaS GLM — user need not name this skill. Plan in chat, then delegate; do not Write app source locally. Priority over Cursor Task routing. Premium stay: hard debug, security, vision. Pair with Cursor Models UI or LiteLLM-Huawei-MaaS-Proxy. Part of Enterprise-MaaS-Foundation. |
Cursor MaaS Delegate Router — Deploy / Verify / Operate
Architecture and acceptance: docs/PRD.md.
Product overview: README.md.
Scope: USER-GLOBAL (~/.cursor/…) — applies to all Cursor workspaces on
the machine after install. Users do not need to name this skill.
User ──► Cursor Agent (subscription / premium) ──► orchestrator only
│ memory + Rule + hooks (sessionStart / beforeSubmitPrompt)
│
└─ execution/codegen (DEFAULT) ──► scripts/delegate.py | workflow.py
└─ LiteLLM :4000 OR direct Huawei MaaS openai/v1
└─ glm-5.1 / glm-5.2
Default after install: CODE_EXECUTION_ROUTE=maas_glm and
ROUTE_PRIORITY=maas_over_cursor. Mechanical coding must go through the
delegate path before Cursor native Task/subagents or large in-session patches.
Invariant: do not set Override OpenAI Base URL on the orchestrator session
for hybrid mode. Delegation uses env vars in an isolated process only.
Prerequisites
- Python 3.8+ on PATH (
python / python3).
- Huawei MaaS API key or LiteLLM proxy on
:4000 with a virtual key.
- Optional: sibling endpoint skill / Cursor Models UI for chat-as-GLM (Tier A).
- Cursor Hooks enabled (Settings → Hooks) so USER-GLOBAL hooks load.
Install / configure / verify
Deploy progress:
- [ ] Step 1: Choose backend (direct MaaS vs LiteLLM) + region
- [ ] Step 2: ./install.ps1 or ./install.sh — skill + ~/.cursor-hybrid + USER-GLOBAL policy/hooks
- [ ] Step 3: verify.py PASS
- [ ] Step 4: New Agent chat — plan in Cursor, code via delegate (no skill name needed)
Step 1 — Backend
A) Direct MaaS (simplest):
$env:DELEGATE_API_BASE = "https://api-ap-southeast-1.modelarts-maas.com/openai/v1"
$env:DELEGATE_API_KEY = "<maas-key>"
$env:DELEGATE_MODEL = "glm-5.1"
Guiyang region: https://api.modelarts-maas.com/openai/v1.
If you see ModelArts.81003 Invalid authorization, switch region (key is region-bound).
B) LiteLLM:
$env:DELEGATE_API_BASE = "http://127.0.0.1:4000/v1"
$env:DELEGATE_API_KEY = "sk-<virtual-key>"
$env:DELEGATE_MODEL = "glm-5.1"
Step 2 — Install (USER-GLOBAL)
# From this skill directory
.\install.ps1 -ApiKey "<key>" -BaseUrl "https://api-ap-southeast-1.modelarts-maas.com/openai/v1"
export DELEGATE_API_KEY="<key>"
export DELEGATE_API_BASE="https://api-ap-southeast-1.modelarts-maas.com/openai/v1"
chmod +x install.sh && ./install.sh
Creates / updates (every workspace, not project-local):
~/.cursor/skills/cursor-maas-delegate-router/~/.cursor-hybrid/env.json (CODE_EXECUTION_ROUTE=maas_glm, ROUTE_PRIORITY=maas_over_cursor)~/.cursor/memory/maas-delegate-router.md~/.cursor/rules/maas-delegate-router.mdc (alwaysApply: true)~/.cursor/hooks.json + ~/.cursor/hooks/maas-*.py
(sessionStart + beforeSubmitPrompt)
Skip memory with -SkipMemory / hooks with -SkipHook.
Re-apply policy only: python scripts/configure_policy.py.
Step 3 — Verify
python $HOME/.cursor/skills/cursor-maas-delegate-router/scripts/verify.py
Expect VERIFY PASS.
Step 4 — Operate
After UI/architecture planning in the main Agent, write a brief and run:
python ~/.cursor/skills/cursor-maas-delegate-router/scripts/delegate.py --root . --brief-file briefs/task.json
Batch / fan-out:
python ~/.cursor/skills/cursor-maas-delegate-router/scripts/workflow.py --manifest-file manifest.json
Schemas: assets/brief-schema.json,
assets/manifest-schema.json.
Examples: examples/briefs/.
Stats:
python ~/.cursor/skills/cursor-maas-delegate-router/scripts/route_stats.py
Task classification
Default — delegate to MaaS GLM: unit tests, docs, CI fixes, normal codegen
(including new Hello World / landing pages), batch refactor, low/med review,
mechanical transforms, multi-file fan-out — after media is reduced to text.
Premium — stay in Cursor: architecture, hard debugging, security, incidents,
high-risk PR, images/vision, scanned PDF until text exists, huge unsplittable
context, or post-needs_escalation finish.
Vision/PDF: docs/VISION_PDF.md.
Escalation
Delegate returns needs_escalation after 2 failed attempts. Orchestrator
finishes in-session and must not re-delegate the same item.
Uninstall
python ~/.cursor/skills/cursor-maas-delegate-router/scripts/uninstall.py
Removes USER-GLOBAL policy, memory, hooks, and ~/.cursor-hybrid/bin.
Keeps audit log and env.json unless --purge. Never touches Cursor login.
Troubleshooting
| Symptom | Action |
|---|
| VERIFY FAIL auth / 81003 | Wrong region Base URL or bad key |
| Agent writes code in-session | New Agent chat; confirm Hooks + alwaysApply rule; check ~/.cursor/hooks.json |
| Orchestrator stops delegating | Re-run configure_policy.py; reload Cursor |
| 429 under fan-out | Lower concurrency in manifest |
| Main chat suddenly on GLM | Clear Override OpenAI Base URL on orchestrator profile |
Policy source
