feat: add autonomous coach-player mode for improved code quality (It's the Goose 3 CLI coding model)

[chaollapark]

What does this PR do?

Adds an autonomous coach-player adversarial feedback loop to OpenCode CLI (opencode autonomous) to improve implementation quality via iterative critique.

Inspired by Block's Goose 3 "adversarial cooperation" approach:
https://block.xyz/documents/adversarial-cooperation-in-code-synthesis.pdf

Roles

• Player agent: implements requirements, writes code, creates files
• Coach agent: reviews, critiques, finds gaps, suggests fixes (fresh instance each turn to reduce context drift)

Key features

• --max-turns configurable (default: 5)
• Separate models for player and coach
• Per-turn metrics (token usage + duration)
• Explicit approval signal (IMPLEMENTATION_APPROVED) for reliable stop condition
• --format json output for scripting/automation

Usage

opencode autonomous requirements.md
opencode autonomous "Build a CLI" --max-turns 10
opencode autonomous req.md -p anthropic/claude-3-5-sonnet -c openai/gpt-4o

Code layout

• Added: packages/opencode/src/autonomous/* (types, player, coach, loop/index)
• Added: packages/opencode/src/cli/cmd/autonomous.ts
• Added: packages/opencode/test/autonomous/autonomous.test.ts
• Modified: packages/opencode/src/index.ts (command registration)

---

How did you verify your code works?

Automated

• Ran unit tests (including packages/opencode/test/autonomous/autonomous.test.ts)
• Ran typecheck
• Ran lint/format

Manual smoke tests

• opencode autonomous requirements.md completes successfully
• --max-turns 1 stops after 1 turn
• Coach approval (IMPLEMENTATION_APPROVED) exits early
• --format json prints valid JSON and includes per-turn metrics
• Invalid args / missing file paths return a clear error + non-zero exit code

Notes

• Coach is instantiated fresh each turn to avoid accumulating bias/context.
• Network/provider calls are mocked in tests (so CI doesn't spend money or flake).

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate Found

PR #9312: feat: add ralph-loop for autonomous task completion

• feat: add ralph-loop for autonomous task completion #9312

Why it might be related: This PR also addresses autonomous task completion with an iterative loop mechanism. While it may use different implementation details ("ralph-loop" vs "coach-player mode"), both PRs appear to tackle similar autonomous execution patterns. Worth reviewing to ensure there's no overlapping functionality or conflicting approaches.