[BUG] Model selection reverts to hardcoded Anthropic models instead of preserving user's OAuth/configured model

[tctinh]

Description

When using Oh-My-OpenCode with OAuth providers (e.g., Google Antigravity) or custom model configurations, the system frequently reverts to hardcoded Anthropic models instead of preserving the user's selected model across agent handoffs and background task continuations.
This happens particularly when:

• Specialized agents (Planner/Sisyphus, Librarian, Oracle) are invoked
• Background agents complete and "kick" work back to the main session
• Agent continuation/handoff occurs

Steps to Reproduce

1. Configure Oh-My-OpenCode with a non-Anthropic model provider (e.g., Google OAuth with google/claude-opus-4-5-thinking)
2. Start a session using your configured model
3. Trigger a specialized agent (e.g., run a planning task that uses Planner-Sisyphus)
4. Observe the model being used in the agent's response
5. Notice that subsequent messages revert to anthropic/claude-opus-4-5 or anthropic/claude-sonnet-4-5

Expected Behavior

The system should preserve the user's selected model across:

• Agent handoffs
• Background task completions
• Session continuations
• Specialized agent invocations
  If a user is authenticated via Google OAuth and using google/claude-opus-4-5-thinking, all agents and continuations should respect this choice unless explicitly overridden.

Actual Behavior

The system reverts to hardcoded models:

• Sisyphus (Planner): anthropic/claude-opus-4-5
• Librarian: anthropic/claude-sonnet-4-5
• Oracle: openai/gpt-5.2
• Background agent continuations lose model context entirely

Root Cause Analysis

After investigating the codebase, I identified three primary issues:

1. Hardcoded Agent Models

Files affected:

• src/agents/sisyphus.ts (line ~12)
• src/agents/librarian.ts (line ~10)
• src/agents/oracle.ts (line ~8)
  Each specialized agent has a hardcoded DEFAULT_MODEL constant that overrides user configuration:

// sisyphus.ts
const DEFAULT_MODEL = "anthropic/claude-opus-4-5"
// librarian.ts  
const DEFAULT_MODEL = "anthropic/claude-sonnet-4-5"

2. Background Agent Manager Missing Model Context

File: src/features/background-agent/manager.ts (around line 370)
When background tasks notify the parent session after completion:

const prevMessage = messageDir ? findNearestMessageWithFields(messageDir) : null
await this.client.session.prompt({
  path: { id: task.parentSessionID },
  body: {
    agent: prevMessage?.agent,  // Preserves agent
    // MISSING: model field is not preserved!
    parts: [{ type: "text", text: message }],
  },
})

The model field is not carried forward, causing fallback to default models.

3. Incomplete Model Context Propagation

While src/features/hook-message-injector/injector.ts contains logic to find model context via findNearestMessageWithFields(), it can only work if the model information is actually passed through agent handoffs (which it currently isn't).

Proposed Solutions

Solution 1: Fix Background Agent Manager (Critical)

Update src/features/background-agent/manager.ts to preserve model context:

const prevMessage = messageDir ? findNearestMessageWithFields(messageDir) : null
await this.client.session.prompt({
  path: { id: task.parentSessionID },
  body: {
    agent: prevMessage?.agent,
    model: prevMessage?.model  // ADD THIS
      ? {
          providerID: prevMessage.model.providerID,
          modelID: prevMessage.model.modelID,
        }
      : undefined,
    parts: [{ type: "text", text: message }],
  },
})

Solution 2: Make Agent Models Configurable

Instead of hardcoded defaults, allow agents to respect user configuration:

// Example for sisyphus.ts
const getDefaultModel = (userConfig) => {
  // Priority: OAuth provider > User preference > Fallback default
  if (userConfig.oauthModel?.provider) {
    return `${userConfig.oauthModel.provider}/${userConfig.oauthModel.modelId}`
  }
  return userConfig.preferredModel || "anthropic/claude-opus-4-5"
}

Solution 3: Add Configuration Option

Add a user preference in config:

{
  "modelPersistence": {
    "preserveAcrossAgents": true,
    "respectAgentDefaults": false
  }
}

Environment

• Oh-My-OpenCode Version: latest from main branch
• Node Version: v22.x
• OS: Linux
• OAuth Provider: Google Antigravity
• Configured Model: google/claude-opus-4-5-thinking or google/claude-sonnet-4-5

Impact

This issue affects:

• Users with OAuth providers (Google, custom providers)
• Users with non-Anthropic model preferences
• Multi-agent workflows where model consistency is critical
• Background task continuations

Additional Context

This bug makes it difficult to use Oh-My-OpenCode with alternative model providers, as the system constantly reverts to Anthropic models regardless of user configuration. This is particularly problematic for:

• Users who don't have Anthropic API access
• Users leveraging OAuth authentication with other providers
• Teams standardizing on specific models for consistency

Workarounds

Temporary workarounds while waiting for a fix:

1. Avoid using specialized agents (Planner, Librarian, Oracle)
2. Manually re-select model after each agent handoff (if CLI supports it)
3. Fork the repo and apply the proposed fixes locally

Would you like me to submit a PR?

I'm happy to contribute a fix if the maintainers are open to it. The changes are relatively straightforward and focused on:

1. Preserving model context in background agent manager
2. Making agent model defaults configurable
3. Respecting user OAuth provider settings

[code-yeongyu]

Would you like me to submit a PR?

Yes if you are interested. We are open

[code-yeongyu]

@sisyphus-dev-ai hi investigate this issue btw

[sisyphus-dev-ai]

👋 Hey @code-yeongyu! I'm on it...

[sisyphus-dev-ai]

✅ Fix Implemented!

I've created PR #208 to address this issue: #208

What was fixed:

1. Background Agent Manager Model Preservation

   • Background tasks now capture and preserve the parent session's model context
   • Model info is passed when notifying parent session after task completion
   • Prevents reverting to hardcoded defaults

2. Librarian Agent Factory Pattern

   • Converted from static config to factory function with configurable model
   • Maintains backward compatibility while enabling flexible model selection

3. Background Task Tool Model Capture

   • Automatically captures current model context when launching background tasks
   • Ensures model selection persists through the entire task lifecycle

Testing:

• ✅ All tests pass (20/20)
• ✅ Build successful
• ✅ Backward compatible

The PR is ready for review and targets the dev branch as requested. Thanks for the detailed bug report! 🎉