Skip to content

Latest commit

 

History

History
193 lines (141 loc) · 7.53 KB

File metadata and controls

193 lines (141 loc) · 7.53 KB

OpenClaw Install — Growth Coach as a standalone agent

This guide walks you through installing Growth Coach as a first-class OpenClaw agent (not just a skill dropped into your main agent). The result: a dedicated growth-coach agent with its own workspace, memory, sandbox, and routing — your daily venting channel stays clean and isolated from your main assistant.

If you'd rather keep things simple and just install the skill into your existing main agent, jump to § Skill-only install at the bottom.

Prerequisites

  • OpenClaw ≥ 2026.x installed (npm install -g openclaw@latest)
  • Authenticated with at least one model provider (openclaw auth login)
  • Feishu / Lark channel already configured (recommended — see OpenClaw Feishu setup)

Verify with:

openclaw doctor
openclaw agents list   # should show at least `main`

One-command install

# Clone the repo (or download the install script directly)
git clone https://github.com/Tangweiwei227/growth-coach ~/growth-coach
cd ~/growth-coach

# Run the scaffold script — installs agent + skill + bootstrap files + openclaw.json patch
bash install/openclaw-agent.sh

The script will:

  1. Create the agent growth-coach with its own workspace at ~/.openclaw/workspace-growth-coach/
  2. Install the growth-coach skill into that workspace's skills/ dir
  3. Copy the bootstrap files (IDENTITY.md / SOUL.md / USER.md / TOOLS.md / AGENTS.md / BOOTSTRAP.md) into the workspace
  4. Patch ~/.openclaw/openclaw.json with:
    • agents.list[] entry (id, model, skills allowlist, tool policy)
    • bindings[] entry for Feishu DMs (only if you pass --bind)
  5. Restart the gateway

Options

# Custom workspace location
bash install/openclaw-agent.sh --workspace ~/.coaching/workspace

# Different model (override the default anthropic/claude-sonnet-4-6)
OPENCLAW_GROWTH_COACH_MODEL=openai/gpt-5 bash install/openclaw-agent.sh

# Bind to a specific Feishu open_id (find it via openclaw contacts)
bash install/openclaw-agent.sh \
  --bind feishu:ou_d738e0d578a074814982527784c703b1

# Multiple bindings? Run the script twice with different --bind values

What you get

After the script finishes, your ~/.openclaw/workspace-growth-coach/ looks like:

workspace-growth-coach/
├── AGENTS.md          # runtime contract (every-session read)
├── BOOTSTRAP.md       # first-run ritual — agent deletes after onboarding
├── IDENTITY.md        # name, vibe, channels, what-I-am-NOT
├── MEMORY.md          # (empty) curated long-term memory index
├── SOUL.md            # persona contract + hard limits
├── TOOLS.md           # local cheat-sheet for the agent
├── USER.md            # placeholder — agent fills during bootstrap
├── memory/            # daily slugged notes
└── skills/
    └── growth-coach/
        └── SKILL.md   # the actual system prompt

First-run ritual

Send any message to the agent (via the bound channel — Feishu DM by default). The agent loads BOOTSTRAP.md and asks you four onboarding questions:

  1. What should I call you?
  2. One line about who you are (your work, family, anything that helps me understand your context)
  3. One thing you'd want me to know before we ever talk about anything heavy
  4. One thing you NEVER want me to do

Answer in any format. The agent will:

  • Write your answers into USER.md (replacing the placeholder)
  • Initialize MEMORY.md with one bootstrap entry
  • Delete BOOTSTRAP.md (it served its purpose)
  • Say something warm and short, like: "好了。我准备好了。今天想聊什么?"

Then start venting. Try @Growth Coach 今天被老板当众批评方案很烂,我整个人都不好了.

Daily use

👤 You (via Feishu DM):
   今天又被老板挑刺了

🤖 Growth Coach:
   听上去是那种"明明已经很小心了,还是被挑出来"的憋屈,对吗?
   [high-perspective analysis ...]
   [specific blessing in disguise ...]
   微行动:下班路上买一杯平时舍不得的咖啡,跟自己说"今天扛了一件很难的事"。

The agent will:

  • ✅ Read MEMORY.md and any prior growth-{slug}.md notes to recall context
  • ✅ Respond in three parts (emotion naming + analysis + blessing-or-honest-admission)
  • ✅ Close with one 5-minute micro-action
  • ✅ Write a memory/{YYYY-MM-DD}-{slug}.md if you committed to something
  • Never say "加油" / "你可以的" / "一切都会好的"
  • 🚨 Stop coaching and route to crisis hotlines if you express self-harm intent

Inspecting memory

# Curated long-term memory (one bullet per stable preference)
cat ~/.openclaw/workspace-growth-coach/MEMORY.md

# Daily notes
ls ~/.openclaw/workspace-growth-coach/memory/

# Engine-indexed search across all your notes
openclaw memory search "被批评"

Tool policy (what the agent can do)

The Growth Coach agent is explicitly gated to a minimal tool set:

Allowed Purpose
group:fs Read/write memory files
group:memory Search across session history
group:sessions List active sessions (for context)
group:messaging Reply via Feishu/IM
Denied Why
group:runtime (exec/process) A coaching agent should never run shell commands
group:ui (browser/canvas) Not needed for text-based venting
group:media (image/video/music gen) Distracts from the coaching work
group:nodes (paired-device control) Scope creep — your coach doesn't need to control your devices
group:automation (cron/gateway) This is a real-time conversational agent, not a scheduled task

This is intentional — a coaching agent doesn't need to run shell commands or generate images. If you ever need a more powerful variant (e.g., Growth Coach that can search the web for you), edit ~/.openclaw/openclaw.json and add group:web to tools.allow.

Uninstalling

bash install/openclaw-agent.sh --uninstall

This removes the agent, uninstalls the skill, and deletes the workspace. You'll need to manually remove the growth-coach entry from ~/.openclaw/openclaw.json agents.list[].


Skill-only install (alternative)

If you don't want a dedicated agent and just want the skill available in your existing main agent:

# Install into the shared managed pool (visible to all agents on this machine)
openclaw skills install https://github.com/Tangweiwei227/growth-coach --as growth-coach --global

# Or install into your active workspace only
openclaw skills install https://github.com/Tangweiwei227/growth-coach --as growth-coach

Then in ~/.openclaw/openclaw.json, ensure agents.defaults.skills (or agents.list[] for your specific agent) includes "growth-coach".

Restart the gateway: openclaw gateway restart.

The skill will auto-load when the user vents, but it will share main's workspace and memory — your coaching notes will be visible to the rest of your main agent's context.

Publishing to ClawHub (for maintainers)

When you're ready to publish:

clawhub login
clawhub skill publish ./agents/openclaw \
  --slug growth-coach \
  --name "Growth Coach" \
  --version 0.1.0 \
  --changelog "Initial release" \
  --tags "coaching,emotional-support,personal-growth,reframing"

# Verify after publish
openclaw skills verify @<your-handle>/growth-coach
openclaw skills verify @<your-handle>/growth-coach --card

Then other OpenClaw users can install with openclaw skills install @<your-handle>/growth-coach without cloning this repo.