[Bug]: Category delegation fails to resolve Ollama models despite correct configuration

[rooftop-Owl]

Bug Description

Category-based delegation using delegate_task(category='quick') fails with "Model not configured for category" error, even though the category is correctly configured with an Ollama model (ollama/ministral-3:14b-32k-agent).

Direct agent routing using delegate_task(subagent_type='explore') works perfectly with the same Ollama model — only category-based routing fails.

---

Steps to Reproduce

1. Configure Ollama provider in OpenCode (~/.config/opencode/opencode.json)
2. Set category model in oh-my-opencode config:

   "categories": {
     "quick": {
       "agents": ["explore", "librarian"],
       "default_model": "ollama/ministral-3:14b-32k-agent"
     }
   }

3. Populate provider cache with Ollama models (required because OpenCode SDK doesn't auto-discover local models)
4. Attempt category delegation:

   delegate_task(category='quick', load_skills=[], run_in_background=true, prompt='List files')

5. Observe error: "Model not configured for category"
6. Compare with direct agent routing (works):

   delegate_task(subagent_type='explore', load_skills=[], run_in_background=true, prompt='List files')

---

Expected Behavior

Category delegation should resolve ollama/ministral-3:14b-32k-agent from categories.quick.default_model and delegate to the explore agent using Ollama, the same way direct agent routing does.

---

Actual Behavior

Error returned:

Model not configured for category "quick".

Configure in one of:
1. OpenCode: Set "model" in opencode.json
2. Oh-My-OpenCode: Set category model in oh-my-opencode.json
3. Provider: Connect a provider with available models

Current category: quick
Available categories: visual-engineering, ultrabrain, deep, artistry, quick, ...

Note: "quick" IS listed in available categories, confirming config was loaded. The failure is in model resolution, not category discovery.

---

Root Cause

fetchAvailableModels() in src/shared/model-availability.ts line 196 assumes modelIds is string[], but manually-populated Ollama caches use object[] format with metadata ({id, provider, context, output}). This causes:

// Line 196 — string concatenation with object produces garbage
modelSet.add(`${providerId}/${modelId}`)
// Actual result: "ollama/[object Object]"

The availableModels Set then contains invalid entries → resolveModelPipeline() can't find the model → error at line 869.

Why direct routing works: delegate_task(subagent_type='explore') bypasses fetchAvailableModels() entirely, going straight to agent config.

---

Doctor Output

Manual environment info:
- oh-my-opencode: 3.2.3
- @opencode-ai/plugin: 1.1.51
- OpenCode SDK: 1.1.19
- Ollama: running on http://localhost:11434
- Models: ministral-3:14b-32k-agent, qwen3-coder:32k-agent, lfm2.5-thinking:agent
- Provider cache: manually populated (6 Ollama models)

---

Error Logs

Error originates at src/tools/delegate-task/executor.ts line 869:

if (!categoryModel && !actualModel) {
  return {
    error: `Model not configured for category "${args.category}". ...`
  }
}

Both categoryModel and actualModel are undefined after resolveModelPipeline() runs at line 810.

---

Configuration

{
  "categories": {
    "quick": {
      "agents": ["explore", "librarian"],
      "default_model": "ollama/ministral-3:14b-32k-agent",
      "description": "Fast trivial tasks - local Ollama (free)"
    }
  },
  "agents": {
    "explore": {
      "model": "ollama/ministral-3:14b-32k-agent",
      "temperature": 0.2,
      "stream": false,
      "description": "Codebase exploration - Ollama local model"
    }
  }
}

Provider cache (~/.cache/oh-my-opencode/provider-models.json):

{
  "models": {
    "ollama": [
      { "id": "ministral-3:14b-32k-agent", "provider": "ollama", "context": 32768, "output": 8192 }
    ]
  },
  "connected": ["anthropic", "google", "ollama", "openai", "opencode"]
}

---

Fix Submitted

PR #1509: Handles both string[] and object[] cache formats in fetchAvailableModels().

• Backward compatible with existing string[] format
• Extracts .id from object entries when present
• Skips invalid entries gracefully
• 4 new test cases, 48 total passing

---

Additional Context — Upstream Root Cause

Manual cache population is required because OpenCode SDK's client.model.list() does not return local Ollama models. This is a known issue at the OpenCode level:

OpenCode SDK Issues (Root Cause)

• Local Ollama models not included in model listing, breaking plugin model discovery anomalyco/opencode#12243 — Our report: local Ollama models not in model listing
• Auto-discover models from OpenAI-compatible provider endpoints anomalyco/opencode#6231 — Canonical feature request: auto-discover models from OpenAI-compatible endpoints (assigned to @thdxr)
• [FEATURE]: Ollama - Multiple Ollama Cloud Models are missing anomalyco/opencode#7873 — Ollama Cloud models missing
• Model List differ between the CLI and Web. anomalyco/opencode#9581 — Model list differs between CLI and Web
• The cache of the Provider‘s Models cannot be updatable anomalyco/opencode#7714 — Provider's Models cache not updatable

Active Upstream Development

OpenCode is actively building auto-discovery on the models-endpoint feature branch:

• feat(opencode): add auto model detection for OpenAI-compatible providers anomalyco/opencode#8359 [MERGED → models-endpoint, Jan 30] — Auto model detection via /v1/models
• opencode: added logic to probe loaded models from lmstudio, ollama an… anomalyco/opencode#11951 [OPEN → models-endpoint, Feb 3] — Ollama/LM Studio/llama.cpp probing with dedicated provider/local/ollama.ts

Feature branch has not yet been promoted to dev, so PR #1509 serves as a bridge fix covering the gap until OpenCode ships auto-discovery.

Prior Contribution

• PR docs: add Ollama streaming NDJSON issue guide and workaround #1197 (merged 2026-01-28) — NDJSON parser for Ollama streaming responses (different issue)

---

Operating System

Linux

OpenCode Version

1.1.51

[rooftop-Owl]

Related to PR #1197: This issue is separate from the NDJSON streaming parser (which was successfully merged). PR #1197 addressed Ollama's streaming response format, while this issue (#1508) concerns category-based model resolution in the delegation pipeline.

Both issues affect Ollama integration but at different layers:

• PR docs: add Ollama streaming NDJSON issue guide and workaround #1197: Response parsing (NDJSON format handling)
• Issue [Bug]: Category delegation fails to resolve Ollama models despite correct configuration #1508: Model discovery and category routing

Direct agent routing works (showing PR #1197's fix is effective), but category delegation still fails due to model resolution issues in executor.ts.