Add iterative agent loop support for ralph-loop style refinement

[Copilot]

Implements iterative agent execution where each iteration appends the previous result to the original prompt, enabling self-refinement and multi-pass analysis patterns.

Changes

• Engine configuration: Added iterations field to EngineConfig struct and JSON schema

  • Accepts integer or string values (minimum 1)
  • Only activates when > 1 (single runs bypass loop overhead)

• Loop implementation: Created run_agent_loop.sh in actions/setup/sh/

  • Saves original prompt, augments with previous outputs for subsequent runs
  • Per-iteration logging with timing and progress tracking
  • Stops on first failure, preserves all intermediate results

• Engine integration:

  • Copilot, Claude, Codex: Wrapped execution commands with WrapCommandWithIterationLoop() helper
  • Custom engine: Exposes GH_AW_ITERATIONS environment variable for user-defined steps

Usage

---
engine:
  id: copilot
  iterations: 3
---

# Iteratively refine analysis

Review the codebase and suggest improvements.
On each iteration, build on previous findings.

The workflow executes 3 times, with iterations 2-3 receiving the original prompt plus all previous outputs in context.

Original prompt

Add support for specifying a number of iteration to run the agentic engine in a loop. On each iteration, the result of the previous exploration is appended to the original prompt and the agent is started again.

See reference https://github.com/anthropics/claude-plugins-official/tree/main/plugins/ralph-loop

• implemented as a bash script in actions/setup/sh
• extensive logging to support debugging
• not activated unless the number of iteration is greater than 1
• works for all agentic engine.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.