paseo/MODEL_FETCHING_REPORT.md

Model Fetching Investigation Report

Summary

Investigated the model fetching system to determine what models are actually available for both Claude and Codex agents.

Code Location

The model fetching logic is located in:

• File: packages/server/src/server/agent/model-catalog.ts (lines 22-101)
• Main Function: fetchProviderModelCatalog(provider, options)
• Claude Function: fetchClaudeModelCatalog(options) (lines 35-66)
• Codex Function: fetchCodexModelCatalog() (lines 68-101)

How It Works

Claude Model Fetching

• Uses the Claude Agent SDK's query.supportedModels() method
• Creates a query instance with the given cwd (working directory)
• Returns models from the SDK with their display names and descriptions

Codex Model Fetching

• Spawns the Codex binary (codex app-server) as a subprocess
• Communicates via JSON-RPC protocol
• Sends initialize followed by model/list requests
• Returns models with metadata including reasoning efforts

Actual Models Returned

Claude Models (3 total)

1. Default (recommended) (default)

   • Description: "Sonnet 4.5 · Smartest model for daily use"
   • Is Default: false

2. Opus (opus)

   • Description: "Legacy: Opus 4.1 · Reaches usage limits faster"
   • Is Default: false

3. Haiku (haiku)

   • Description: "Haiku 4.5 · Fastest model for simple tasks"
   • Is Default: false

Codex Models (3 total)

1. gpt-5.1-codex (gpt-5.1-codex) ⭐ DEFAULT

   • Description: "Optimized for codex."
   • Is Default: true
   • Model: gpt-5.1-codex
   • Default Reasoning Effort: medium

2. gpt-5.1-codex-mini (gpt-5.1-codex-mini)

   • Description: "Optimized for codex. Cheaper, faster, but less capable."
   • Is Default: false
   • Model: gpt-5.1-codex-mini
   • Default Reasoning Effort: medium

3. gpt-5.1 (gpt-5.1)

   • Description: "Broad world knowledge with strong general reasoning."
   • Is Default: false
   • Model: gpt-5.1
   • Default Reasoning Effort: medium

Key Observations

1. Claude models: The model IDs returned are simplified (default, opus, haiku) rather than full model identifiers like claude-sonnet-4-5-20250929.

2. Codex models: All three models are available and properly marked with isDefault flag on gpt-5.1-codex.

3. No Sonnet variant: Interestingly, Claude returns a "default" model (which is Sonnet 4.5 based on the description) but doesn't expose a model ID called "sonnet" - only "default", "opus", and "haiku".

4. Model resolution: The model-resolver.ts file uses these catalogs to find the default model when none is specified, preferring models marked with isDefault: true.

Testing Method

Created and ran a test script (test-models.ts) that directly imports and calls:

• fetchClaudeModelCatalog()
• fetchCodexModelCatalog()

Executed with: npx tsx test-models.ts

Related Files

• src/server/agent/model-catalog.ts - Model fetching implementation
• src/server/agent/model-resolver.ts - Default model resolution
• src/server/agent/model-catalog.e2e.test.ts - End-to-end tests
• src/server/session.ts:1589 - Usage in session handling (list_provider_models_response)