Self Repair

This skill enables crocbot to safely repair and refactor its own codebase using Claude Code (claude) as a high-thinking-budget sub-agent.

Overview

Self-repair uses a specialized command configuration to ensure Claude Code operates with maximum analytical depth and bypasses routine permission prompts to maintain momentum during autonomous sessions.

Workspace Awareness

The sub-agent is aware of two critical paths:

1. •Source Code: {projectRoot} (where repairs/exploration happen)
2. •Identity Workspace: {workspaceDir} (where research, reports, and persistent documentation are stored)

Workflow

1. Launching a Repair or Research Session

Always use the following command structure. Replace <instructions> with the specific task.

# Launch Claude Code with high thinking budget and identity workspace awareness
# Always use pty:true and background:true for autonomous work
bash pty:true workdir:{projectRoot} background:true command:"claude -p 'Use maximum thinking budget for this prompt and make sure you output a detailed summary end of response. You have permission to write files directly to my identity workspace at {workspaceDir}/ for reporting or research purposes. Instructions: <instructions>' --dangerously-skip-permissions"

2. Full Lifecycle (Repair, Build, Redeploy)

When a sub-agent is tasked with modifying the code, it must follow this completion sequence:

1. •Modify Code: Apply the requested fixes/refactors.
2. •Build: Run pnpm build to ensure the changes are valid and compiled.
3. •Redeploy: Once the build succeeds, it must restart the gateway service:
   bash

   systemctl --user restart crocbot-gateway
   

4. •Continuity: The sub-agent should understand that restarting the gateway will terminate its parent session. It should finalize its report/summary BEFORE the restart command.

2. Monitoring Progress

Once the session is launched, monitor the output using the process tool:

# Check logs
process action:log sessionId:XXX

# Check if finished
process action:poll sessionId:XXX

3. Verification

Once the sub-agent completes the task, review the summary provided by Claude Code to verify the changes.

Safety Rules

• •Targeted Scope: Always set the workdir to {projectRoot}.
• •Deduplication: Before starting a new repair session, check process action:list to ensure a similar session isn't already running.
• •Reporting: Always provide the user with the final "detailed summary" generated by Claude Code.