claude-code-bridge

name	claude-code-bridge
description	Dispatch coding tasks to Claude Code with two-way communication. Send tasks, receive status updates, answer clarifying questions, and get results.
metadata	{"openclaw":{"emoji":"🌉","requires":{"bins":["python3"]}}}

Claude Code Bridge

Send coding tasks to Claude Code and handle two-way communication. When Claude needs clarification, it writes a question that you can relay to the user and answer back.

When to Use

• User asks to build a script (Python, bash, etc.)
• User asks to fix, refactor, or modify code in a project
• User asks for multi-file code generation or large refactors
• User wants a long coding task in the background
• User wants parallel coding tasks across different projects
• Task may require back-and-forth clarification

When NOT to Use

• Quick questions or explanations (just answer directly)
• Simple single-file reads
• Planning or design discussions
• Non-coding tasks

Dispatching a Task

~/.openclaw/skills/claude-code-bridge/claude-bridge-dispatch.sh \
  --task-id <descriptive-name> \
  --workdir <project-directory> \
  --prompt "<detailed task description>"

Naming Conventions

Use descriptive task IDs with project prefixes:

• myapp-fix-auth — bug fix
• myapp-add-csv-export — feature work
• scripts-build-parser — new script creation

Writing Good Prompts

Be specific. Include:

• What to build or fix
• File paths if known
• Expected behavior or output format
• Language and framework preferences

Example:

~/.openclaw/skills/claude-code-bridge/claude-bridge-dispatch.sh \
  --task-id scripts-csv-parser \
  --workdir ~/projects/data-tools \
  --prompt "Create a Python script called parse_orders.py that reads a CSV file of orders, filters rows where status is 'shipped', groups by customer_id, and outputs a summary CSV with columns: customer_id, order_count, total_amount. Use pandas."

Checking Status

Quick status check:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --task-id <id>

Get recent output from Claude:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --task-id <id> --output

Get more output lines:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --task-id <id> --output -n 100

List all tasks:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --list

View the result when complete:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --task-id <id> --result

Answering Questions (Two-Way Communication)

When status shows waiting_for_answer, Claude is asking a clarifying question.

Read the question:

~/.openclaw/skills/claude-code-bridge/claude-bridge-status.sh --task-id <id> --question

The output is JSON with a questions array. Each question has question text and options with labels.

Send the answer:

~/.openclaw/skills/claude-code-bridge/claude-bridge-answer.sh \
  --task-id <id> \
  --answer "<the answer text or selected option label>"

After answering, Claude resumes work automatically.

Workflow

1. User requests a coding task
2. You dispatch a task with a clear, detailed prompt
3. Report the task ID and how to monitor it
4. Periodically check status (every 30-60 seconds for active tasks, or when the user asks)
5. If status is waiting_for_answer: read the question with --question, present it to the user, then send their answer with claude-bridge-answer.sh
6. When status is complete: read the result with --result and report to the user
7. If status is error: check --output and --log for details

Important Notes

• Claude Code runs in acceptEdits mode: it auto-approves file edits and bash commands. Only AskUserQuestion triggers the question relay.
• Tasks persist across OpenClaw restarts (state is in ~/.claude-bridge/tasks/).
• Each task runs in its own Python process.
• If no answer is provided within 10 minutes, the bridge times out and picks the first option as default.
• Kill a task: kill $(cat ~/.claude-bridge/tasks/<id>/bridge.pid)