copilot-loop

This directory contains a small CLI tool used to run an iterative PDCA-style agent loop backed by the GitHub Copilot SDK.

Files

• src/index.ts - Entrypoint script (#!/usr/bin/env bun). Parses CLI args, loads YAML prompt config, and starts the SweAgentInteraction loop.
• src/SweAgentInteraction.ts - Implements the interactive PDCA loop with YOLO and CONFIRM modes, console colors, and signal handling.

Usage

• Run copilot-loop [config-file] [options]
• CLI flags:
  • [config-file] Load YAML config (positional) or use --config <file>
  • -p Execute a prompt in non-interactive mode
  • -a Append additional prompt text to config prompt
  • --resume [sessionId] Resume from a previous session (optionally specify session ID; without ID, resumes last session from /tmp/copilot-loop-last-session)
  • -r Shorthand for --resume (sets confirm mode)
  • --model Specify the AI model to use
  • --think Set reasoning effort (choices: low, medium, high, xhigh)
  • --max Set maximum iterations for agent loop
  • --promise Set completion promise phrase for task completion
  • --timeout Set timeout in seconds (default: 604800 / 7 days)
  • --persona Deploy a specific persona to activate and maintain persistence
  • --debug Use confirm mode for permission prompts instead of automatic approval
  • -h, --help Display help for command

Behavior notes

• The tool implements a PDCA (Plan-Do-Check-Act) loop and expects the assistant to emit a completion promise phrase to terminate (configurable via --promise, no hardcoded default).
• Signal handling (SIGINT/SIGTERM) attempts graceful session abort and client cleanup; the script also listens for stdout stream destruction and Copilot disconnect errors.
• The interactive loop supports two modes: yolo (auto-iterate without prompts) and confirm (ask before executing commands). Use --debug flag or --resume without session ID to enable confirm mode.
• Sessions are tracked via Snowflake-like IDs (generated from timestamp + random bits). Last session ID is saved to /tmp/copilot-loop-last-session for easy resumption.
• Execution logs are written to /tmp/copilot-loop-[sessionId]-log.txt, /tmp/copilot-loop-[sessionId]-error.txt for debugging.
• Health checks run every 3 seconds via client.ping() to detect server hangs; timeout per prompt is configurable (default 7 days).

Examples

Start with a config file:

$ copilot-loop config.yaml

Execute a prompt directly (non-interactive):

$ copilot-loop -p "Fix the bug in main.js"

Resume the last session in confirm mode:

$ copilot-loop --resume

Resume a specific session:

$ copilot-loop --resume abc123def456

Use specific model and reasoning with custom timeout:

$ copilot-loop --model gpt-5-mini --think high --timeout 3600 -p "Optimize this code"

Enable debug mode with persona deployment:

$ copilot-loop config.yaml --debug --persona "JAMES"

YAML configuration file example (config.yaml):

mcpServers:
  counter:
    type: http
    url: http://localhost:3000/mcp
    tools: ["*"]

model: "gpt-5-mini"
reasoningEffort: "high"
max-iterations: 5
timeout: 3600
promise: "TASK_COMPLETED_SUCCESSFULLY"
persona: "JAMES"

prompt: |
  Say Hi, and exit the loop.

Configuration field reference:

• mcpServers: MCP server configuration (passed to Copilot SDK)
• model: AI model identifier
• reasoningEffort: Reasoning level (low, medium, high, xhigh)
• max-iterations: Maximum loop iterations before timeout
• timeout: RPC timeout in seconds (default: 604800)
• promise: Completion signal phrase for the agent
• persona: Persona name to deploy at session start
• prompt: Initial prompt for the agent loop

References

• GitHub Copilot SDK