Notify AI Assistant of Completed Work

Send a structured notification to the user's personal AI assistant so it stays informed about project progress. Currently supports OpenClaw; designed to be extensible to other assistants.

Why This Matters

Personal AI assistants like OpenClaw maintain long-term context about your projects, preferences, and goals. When Claude Code finishes a feature, your assistant doesn't know unless you manually tell it. This skill bridges that gap automatically --- your coding agent informs your personal agent, so your assistant can track progress, remind you of next steps, or answer questions about recent changes without you playing messenger.

Detection

Check for OpenClaw in this order:

1. •CLI available: which openclaw succeeds
2. •Config exists: ~/.openclaw/openclaw.json is present
3. •Gateway running: curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:18789/ 2>/dev/null returns a response

If none of these pass, skip notification silently --- the user may not have an assistant installed.

Sending the Notification

Via CLI (preferred --- simplest, works if daemon is running)

openclaw agent --message "PROJECT UPDATE: [message]"

Via message send (if targeting a specific channel)

openclaw message send --message "PROJECT UPDATE: [message]"

Message Format

Keep notifications concise and structured. The assistant needs facts, not prose:

PROJECT UPDATE: [repo-name]

Completed: [what was done — 1-2 sentences]
Files changed: [count] ([key files])
Tests: [passing/failing] ([count] tests)
Commit: [short hash] on [branch]
Next: [what's planned next, if known]

Example:

PROJECT UPDATE: xswarm-tdd

Completed: Added BDD/Playwright skill with ARIA-first locators and axe-core accessibility audits.
Files changed: 4 (agents/TDD.md, agents/reviewer.md, agents/tester.md, skills/bdd-playwright/SKILL.md)
Tests: all passing (47 tests)
Commit: 2e95414 on main
Next: Security scan integration with xswarm-ai-sanitize

When to Use

• •After completing a major feature or task group
• •After a deployment or release
• •After a significant refactor
• •When the user explicitly asks to notify their assistant

Workflow

1. •Detect --- Check if OpenClaw (or another assistant) is available
2. •Compose --- Build a structured message from the current git state and task context
3. •Send --- Deliver via CLI
4. •Confirm --- Report to the user that the notification was sent (or skipped if no assistant found)

Quick recipe for composing from git state

# Get repo name
REPO=$(basename "$(git rev-parse --show-toplevel)")

# Get last commit info
COMMIT=$(git log -1 --format="%h %s")

# Get changed files count
FILES=$(git diff --name-only HEAD~1 | wc -l | tr -d ' ')

# Get key files
KEY_FILES=$(git diff --name-only HEAD~1 | head -5 | tr '\n' ', ')

Extensibility

This skill is designed to support additional assistants. To add a new one:

1. •Add a detection check (CLI, config file, or running process)
2. •Add a send method (CLI command, API call, or WebSocket message)
3. •Follow the same structured message format

Potential future targets: ntfy.sh, Home Assistant, custom local services.

Two-Way Question Routing (Stuck Agent Integration)

The most powerful use of this skill is full round-trip communication: the stuck agent sends a question through OpenClaw to whatever messaging app you're on (WhatsApp, Telegram, Slack, etc.), you reply from your phone, and the answer flows back to Claude Code to unblock the pipeline. You never need to be at the terminal.

How It Works

Claude Code (stuck)                    OpenClaw                         You (on phone)
      |                                   |                                  |
      |-- openclaw agent --message ------>|                                  |
      |   "QUESTION: which DB?"           |-- forwards to WhatsApp --------->|
      |                                   |                                  |
      |                                   |<-- "Use PostgreSQL" -------------|
      |<-- returns answer to stdout ------|                                  |
      |                                   |                                  |
      | (captures response, unblocks)     |                                  |

1. •Stuck agent detects a problem and formats the question with options
2. •Checks for OpenClaw: which openclaw — if not found, falls back to terminal AskUserQuestion
3. •Sends question and waits for response:
   bash

   RESPONSE=$(openclaw agent --message "CLAUDE CODE NEEDS YOUR INPUT
   
   Project: [repo-name]
   Problem: [brief description]
   
   Options:
   1. [option A] — [explanation]
   2. [option B] — [explanation]
   3. [option C] — [explanation]
   
   Reply with a number or your own answer.")
   

4. •OpenClaw relays the question to the user's preferred messaging channel
5. •User replies from their phone (e.g., "2" or "Use PostgreSQL")
6. •OpenClaw returns the user's response to stdout
7. •Stuck agent captures the response and relays it back to the calling agent
8. •Pipeline unblocks — no terminal interaction needed

Message Format for Questions

CLAUDE CODE NEEDS YOUR INPUT

Project: [repo-name]
Problem: [1-2 sentence description]
Context: [error message or key detail]

Options:
1. [option] — [brief explanation]
2. [option] — [brief explanation]
3. [option] — [brief explanation]

Reply with a number or your own answer.

Fallback Chain

If OpenClaw is unavailable or the command fails, the stuck agent falls back gracefully:

1. •Try OpenClaw — openclaw agent --message (two-way, works remotely)
2. •Fall back to terminal — AskUserQuestion (requires user at keyboard)

This means the system works whether or not OpenClaw is installed. OpenClaw just makes it work remotely.

Rules

• •Never send notifications without the user requesting it or the workflow calling for it. Unsolicited messages to external systems violate the principle of least surprise.
• •Skip silently if no assistant is detected. Not everyone has OpenClaw installed; failing loudly would be annoying noise.
• •Keep messages factual and structured. The receiving assistant needs parseable information, not conversational filler.
• •Never include secrets, tokens, or credentials in notifications. Messages may traverse external channels; treat them as potentially public.
• •Fall back to AskUserQuestion if OpenClaw is unavailable. The pipeline must not proceed without a confirmed human decision, regardless of which channel delivers it.