OpenCode model discovery fails — gentle-ai doesn't detect models from SQLite database

[naujrevilo]

Pre-flight Checklist

• I have searched existing issues and this is not a duplicate
• I understand that PRs will be rejected if the linked issue does not have status:approved

📝 Bug Description

Gentle-AI fails to discover OpenCode models (including MiniMax, Xiaomi, and other custom providers) when configuring models via gentle-ai install or the TUI. The model picker shows no available models, even though opencode providers list and the /models command in OpenCode correctly display all connected providers and their models.

🔄 Steps to Reproduce

1. Install and configure OpenCode with custom API providers (e.g., MiniMax, Xiaomi)
2. Verify providers are connected: run opencode providers list → shows all providers correctly
3. Run gentle-ai install or launch gentle-ai TUI
4. Navigate to: Configure Models → Configure OpenCode models
5. The model picker shows no available providers/models

✅ Expected Behavior

Gentle-AI should discover and display all OpenCode models (including custom providers like MiniMax, Xiaomi) in the model configuration picker.

❌ Actual Behavior

• The OpenCode model picker shows: "OpenCode has not been run yet — model cache not found."
• No providers or models appear in the selection list
• However, opencode providers list correctly shows all 4 providers: MiniMax Token Plan, Xiaomi, GitHub Copilot, MiniMax

Gentle AI Version

• Gentle AI Version: 1.43.3

Operating System

Windows (WSL)

AI Agent / Client

OpenCode

📋 Affected Area

Other

💡 Logs / Error Output

**Gentle-AI detection output:**

Detected Configs
  claude-code: present
  opencode: present
  kilocode: missing
  gemini-cli: missing
  cursor: missing
  vscode-copilot: present
  codex: missing
  ...


**OpenCode providers (working correctly):**

┌  Credentials ~/.local/share/opencode/auth.json
│
●  MiniMax Token Plan (minimax.io) api
●  Xiaomi api
●  GitHub Copilot oauth
●  MiniMax (minimax.io) api
└  4 credentials


**Gentle-AI model picker error:**

OpenCode has not been run yet — model cache not found.
Run 'opencode' once, then re-run 'gentle-ai sync' to assign models.
Using default model assignments for now.

Additional Context

Root Cause Analysis

Gentle-AI searches for OpenCode models in the old models.cache location:

~/.config/opencode/models.cache  ← File does not exist

However, OpenCode stores provider credentials and model information in:

~/.local/share/opencode/auth.json       ← Contains API keys and provider configs
~/.local/share/opencode/opencode.db    ← SQLite database with model data

File System Evidence

Existing gentle-ai search path (no models found):

~/.config/opencode/
├── AGENTS.md
├── opencode.json
├── tui.json
├── skills/
├── profiles/
├── node_modules/
└── NO models.cache file exists

Actual OpenCode data locations (working):

~/.config/opencode/
├── opencode.json     ← SDD orchestrator config only
└── ...

~/.local/share/opencode/
├── auth.json         ← API credentials (minimax, xiaomi, github-copilot)
├── opencode.db       ← SQLite database with models/providers
├── opencode.db-wal
├── opencode.db-shm
└── ...

Code Reference

The gentle-ai code in internal/tui/screens/model_picker.go loads models from:

providers, err := opencode.LoadModelsOrEmpty(cachePath)

Where cachePath appears to point to models.cache in ~/.config/opencode/, which no longer exists in current OpenCode versions.

Suggested Fix Direction

Gentle-AI should:

1. Query the SQLite database ~/.local/share/opencode/opencode.db for available models, OR
2. Detect that OpenCode uses the new data storage format and fall back to opencode providers list CLI output, OR
3. Support reading from ~/.local/share/opencode/ as an additional search path

---

Related: This also explains why Trae IDE models from subscriptions are not visible — the same model discovery mechanism would need to query Trae's own data storage.

[Alan-TheGentleman]

This is clear and in scope. Adding status:approved. A PR linking this issue with Closes #1015 is welcome.

[samar-syd7]

I'd like to take this one.

I'll investigate the current OpenCode model discovery flow, verify the storage assumptions against the current OpenCode layout, and implement a compatible discovery path while preserving existing behavior for older installations. I'll submit a PR with the necessary tests once the implementation is validated.

[samar-syd7]

@Alan-TheGentleman , @naujrevilo
While tracing the current implementation on main, I noticed that the model picker already uses ~/.cache/opencode/models.json together with ~/.local/share/opencode/auth.json for provider availability, and the OpenSpec documents also describe the missing models.json case as expected behavior.

Before I implement a fix, I'd like to confirm the intended discovery source for current OpenCode versions.

Should model discovery continue to rely on the generated ~/.cache/opencode/models.json catalog (with an updated way of generating/populating it), or is the intended direction to discover models from the current OpenCode runtime/storage (e.g. SQLite or another supported source) when the cache is absent?

I want to make sure the implementation matches the intended architecture rather than introducing an alternative discovery path.