Coding Agent

An agentic coding assistant with human-in-the-loop review.

Workflow

1. •Explore — Gathers codebase context
2. •Plan — Generates implementation plan → human reviews
3. •Execute — Implements changes
4. •Verify — Reviews changes → human approves
5. •Apply — Writes changes to files

Invocation

$HOME/.flatagents/skills/coding_agent/run.sh "<task>" --cwd "<working_directory>" --claude

Claude Code Integration

When running as a Claude Code skill, use the --claude flag. This enables a checkpoint/exit approval flow instead of interactive input() prompts.

Approval Gates

This agent requires human approval at two points:

• •Plan review - Before implementing changes
• •Result review - Before applying changes to files

Exit Code 2: Approval Needed

When the agent needs approval, it:

1. •Writes approval request to stderr starting with APPROVAL_NEEDED: <type> (where type is plan or result)
2. •Saves state to .coding_agent_checkpoint.json
3. •Exits with code 2

Handling Approval

When you see exit code 2:

1. •Parse stderr to find APPROVAL_NEEDED: plan or APPROVAL_NEEDED: result
2. •Display the content from stderr to the user
3. •Use AskUserQuestion to ask the user to approve or provide feedback
4. •Resume with the decision via environment variable (task arg not needed on resume):

# Approve plan (task retrieved from checkpoint)
CODING_AGENT_APPROVAL_PLAN=approved $HOME/.flatagents/skills/coding_agent/run.sh --cwd "<dir>" --claude

# Reject plan with feedback
CODING_AGENT_APPROVAL_PLAN="<user feedback here>" $HOME/.flatagents/skills/coding_agent/run.sh --cwd "<dir>" --claude

# Approve result
CODING_AGENT_APPROVAL_RESULT=approved $HOME/.flatagents/skills/coding_agent/run.sh --cwd "<dir>" --claude

Environment Variables

The agent restores state (including the task) from the checkpoint file and continues from where it paused.