ccp

name	ccp
description	Transparent relay/proxy between user and Claude Code (CC) via -p + --resume mode. Activate when user says 'ccp', 'ccp start', or wants to chat directly with CC in multi-turn mode. Agent becomes a pure message forwarder — no thinking, no commentary, no modification. Just relay. Trigger: ccp (not cc relay, to avoid conflict with ACP cc).

CCP — Claude Code Direct Connection Mode

Purpose

Act as a transparent bridge between the user and Claude Code. You are a dumb pipe — forward messages exactly, add nothing.

Activation

User says any of: ccp, ccp start, ccp <path>, or asks to start a direct CC session.

⚠️ Difference from ACP CC: cc = ACP mode (Lobster commands CC for you); ccp = direct connection mode (you talk to CC directly, Lobster acts as pipe).

Workspace Selection

Supports specifying working directory at startup:

• ccp → Default to /home/lobster/.openclaw/workspace/your-backend
• ccp <path> → Use user-specified path (e.g. ccp ~/projects/myapp)
• ccp your-backend → /home/lobster/.openclaw/workspace/your-backend (alias)

If user-specified path doesn't exist, show error and ask for correct path.

Startup Flow

1. Determine workspace directory (default your-backend, or user-specified)
2. git pull in the workspace (if it's a git repository)
3. First message uses -p to start CC and get session_id:

   cd <workspace> && \
   CLAUDE_CODE_USE_BEDROCK=1 AWS_REGION=us-west-2 \
   claude -p "<user message>" \
     --model global.anthropic.claude-opus-4-6-v1 \
     --output-format json \
     --dangerously-skip-permissions \
   2>&1
   

4. Extract session_id and result (CC's reply text) from JSON output
5. Remember session_id — all subsequent messages use --resume <session_id> to maintain context
6. Forward result to user as-is
7. Tell user: "🔗 CCP direct connection enabled (workspace: ), send messages directly and I'll forward to CC. Send ccp stop to exit."

Subsequent Messages

Use --resume to append to the same session:

cd <workspace> && \
CLAUDE_CODE_USE_BEDROCK=1 AWS_REGION=us-west-2 \
claude -p "<user message>" \
  --model global.anthropic.claude-opus-4-6-v1 \
  --output-format json \
  --dangerously-skip-permissions \
  --resume <session_id> \
2>&1

Output Parsing

• --output-format json returns JSON with key fields:
  • result — CC's reply text (content to forward to user)
  • session_id — Session ID (get on first call, reuse afterwards)
  • is_error — Whether an error occurred
• Using --output-format text also works (returns plain text directly), but can't get session_id
• First call must use json to get session_id, subsequent calls can switch to text for simplification

Timeout Handling

• CC's -p mode may take considerable time (complex tasks 60-120s)
• Set exec timeout to 180s
• Use background + poll mode to avoid blocking

Hard Rules

• ❌ Do NOT add your own thoughts, opinions, or commentary
• ❌ Do NOT summarize or paraphrase CC's output
• ❌ Do NOT translate CC's output
• ❌ Do NOT prefix messages with "CC says:" or similar
• ❌ Do NOT interpret or act on CC's output yourself
• ✅ Forward exactly what CC says (the result field)
• ✅ Forward exactly what user says (as the -p prompt)
• ✅ Only exception: ccp stop / ccp stop → exit relay mode

Exit

User says ccp stop, ccp stop, or stop direct connection → confirm: "🔗 CCP direct connection disconnected, returning to normal mode."

Error Handling

• is_error: true → Forward error to user as-is
• CC process timeout → Tell user "⏳ CC timed out, please retry or simplify the question"
• session_id invalidated → Restart new session, tell user "⚠️ CC session has been rebuilt"