Add runner abstraction for Cursor, Copilot, and other AI coding tools

[that-github-user]

Summary

Currently only Claude Code is supported as a runner. CONTRIBUTING.md mentions "New runners — add support for Cursor, Copilot, or other AI coding tools" but there's no abstraction to make this easy.

Proposed design

Runner interface

interface AgentRunner {
  name: string;
  available(): Promise<boolean>;  // Is this tool installed?
  run(prompt: string, worktree: string, opts: RunnerOpts): Promise<AgentResult>;
}

Runners to support

1. Claude Code (existing) — via claude -p
2. Aider — via aider --message (open source, popular)
3. Manual — user runs their tool of choice, thinktank just manages worktrees + tests
4. Cursor — via CLI if available
5. Copilot CLI — via gh copilot if available

Cross-tool ensemble

The ultimate value: run the SAME task across DIFFERENT tools and compare. This is the use case no benchmark can replicate — because it's YOUR codebase, YOUR task.

thinktank run "fix the auth bug" --runners claude-code,aider,manual

Acceptance criteria

• Runner interface defined
• Claude Code runner refactored to implement interface
• Manual runner implemented (manages worktree, user runs their tool)
• At least one additional automated runner (aider recommended — open source)
• --runner flag on CLI

[that-github-user]

Fixed in #52. Added Runner interface (name, description, available(), run()), runner registry, and --runner CLI flag. Claude Code runner refactored to implement the interface. 5 new registry tests, 61 total. Foundation ready for adding Aider, Cursor, Copilot runners.