feat(model-optimizer): static model rankings with model-config --apply

[qwertystars]

Summary

• Adds model-config CLI command that shows optimal model configuration based on static rankings
• Adds --apply flag to automatically write config to the user's oh-my-opencode.json
• No API calls required - completely free model optimization!

Usage

# Show optimal config
bunx oh-my-opencode model-config

# Show with full rankings
bunx oh-my-opencode model-config --verbose

# Auto-apply to config
bunx oh-my-opencode model-config --apply

Help Wanted 🙏

We need help ranking AI models!

The static rankings in src/cli/model-optimizer/rankings.ts are based on general observations, but we'd love community input to improve them. If you have experience with these models and can help refine the rankings, please:

1. Open an issue with your suggestions
2. Submit a PR updating the rankings

Categories that need ranking help:

• orchestrator - Best overall models for complex tasks
• reasoning - Best for debugging/architecture
• fast - Fast + capable for quick lookups
• coding - Best coding models
• instruction - Best instruction-following
• multimodal - Vision-capable models
• creative - Creative/artistic models
• free - Free/cheap fallback models

Test Results

✓ TypeScript type check: Passes
✓ Build: Passes
✓ CLI tests: 184 pass, 0 fail
✓ Model-optimizer tests: 45 pass, 0 fail

[cubic-dev-ai]

1 issue found across 9 files

Confidence score: 4/5

• This PR looks safe to merge with minimal risk; the main concern is limited to config path handling in the test command.
• src/cli/model-optimizer/test-command.ts writes oh-my-opencode.json to a hard-coded ~/.config path, which can bypass shared config resolution (XDG/APPDATA/OPENCODE_CONFIG_DIR) and lead to applied config not being read.
• Pay close attention to src/cli/model-optimizer/test-command.ts - hard-coded config path may bypass shared resolution and cause applied config to be ignored.

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/cli/model-optimizer/test-command.ts">

<violation number="1" location="src/cli/model-optimizer/test-command.ts:131">
P2: --apply writes oh-my-opencode.json to a hard-coded ~/.config path, bypassing the shared config resolution (XDG/APPDATA/OPENCODE_CONFIG_DIR) used elsewhere, so applied config may not be read.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[qwertystars]

Fixed the issue identified by cubic-dev-ai in commit 059cf67:

Issue: --apply was writing to a hardcoded ~/.config/opencode path, bypassing the shared config resolution.

Fix: Now uses getOpenCodeConfigPaths() from shared utilities to respect:

• OPENCODE_CONFIG_DIR environment variable
• XDG_CONFIG_HOME on Linux
• APPDATA on Windows

This ensures the applied config is written to the same location that oh-my-opencode reads from.

[GollyJer]

Thinking out loud...

The best community input you can get is from https://lmarena.ai/leaderboard for general categories beyond just programming.
https://www.swebench.com/bash-only.html is great for programming tasks. It's based on solving github issues.
https://artificialanalysis.ai/ is decent.
https://www.vellum.ai/llm-leaderboard?utm_source=direct&utm_medium=none is also good.

We could probably pick different categories across these ranking sites and tie them to different agents.
Like the "text" category would be "document-writer" as an example.
Once all the categories are mapped to the agents, we just run some code to scrape the results and let all the benchmarks update the model recommendations.