jason-easyazz zoe-capability-extender
Add new capabilities to Zoe: new intents, MCP tools, or OpenClaw skills.
Trigger conditions
This skill activates when the system message begins with [ZOE_SELF_BUILD: capability], or when the user asks to add a new ability, command, feature, or integration to Zoe.
Prerequisites
- Caller must have admin role. Check via
zoe_self_capabilitiestool. If not admin, stop. - Read the relevant architectural files before writing anything:
services/zoe-data/intent_router.py— for intent pattern additionsservices/zoe-data/mcp_server.py— for MCP tool additions/home/zoe/.openclaw/workspace/skills/— for skill-level additions
- No hardcoded NLU in
chat.py: never addif/elsekeyword detection to the production chat router.
Identifying the right layer
Choose the layer that matches the capability scope:
| Layer | When to use | Where to edit |
|---|---|---|
| Intent router pattern | New phrase/intent that maps to existing logic | intent_router.py — add a pattern + handler |
| MCP tool | New structured action Zoe can take (API call, DB write, etc.) | mcp_server.py — add a @mcp_tool function |
| OpenClaw skill | Multi-step reasoning task or domain knowledge | New SKILL.md in /home/zoe/.openclaw/workspace/skills/<name>/ |
| Proactive trigger | Background check that fires unprompted | New Trigger class in services/zoe-data/proactive/triggers/ + register in main.py |
Step-by-step workflow
1. Understand and classify the request
Ask if ambiguous:
- Is this a new phrase Zoe should understand, or a new action she should take?
- Should it run in the background, or only when asked?
- Does it need external data (API, DB, HA)?
2. Spec + confirm
Draft a brief spec (≤5 bullet points) describing:
- Which layer will be used and why
- What the new capability does end-to-end
- Any new data sources or credentials needed
Present the spec to the user and wait for confirmation.
3. Implement minimally
Intent router pattern (edit intent_router.py):
# Add to the relevant pattern group — do NOT create a new if/else block in chat.py
{"pattern": r"(?i)\bnew phrase\b", "intent": "new_intent", "handler": existing_or_new_handler}
MCP tool (edit mcp_server.py):
@mcp_tool(name="tool_name", description="One-line description")
async def tool_name(param: str) -> dict:
...
OpenClaw skill (create SKILL.md):
- Follow the format of
zoe-widget-builder/SKILL.md - Describe trigger conditions, step-by-step workflow, output contract, and error handling
- Place in
/home/zoe/.openclaw/workspace/skills/<skill-name>/SKILL.md
Proactive trigger (create .py + register):
- Subclass
ProactiveTriggerfromproactive.triggers.base - Implement
async def check(self, db) -> list[TriggerResult] - Register in
main.pylifespan withregister_trigger(MyTrigger())
4. Test
- For intent/MCP changes: restart
zoe-data(systemctl --user restart zoe-data.service) and exercise the new path manually - For skills: test via OpenClaw with a representative prompt
- For triggers: check
journalctl --user -u zoe-data.service -n 30for "Registered trigger" and confirm no import errors
5. Confirm with user
Report exactly what was added, which file was changed, and how to exercise the new capability. Do not add speculative follow-on features.
Constraints
- No NLU if/else in
chat.py— this is a hard rule from the design charter - Minimum change for the stated request; no speculative abstractions
- Touch only the files the task requires
- All new credentials go in
.env, never hardcoded - New MCP tools must be documented with a clear
description=string — Zoe uses this for routing