[BUG] Codebase indexing fails with OpenAI-compatible embedding providers (API key not saved, hardcoded model list, config resets)

[QEU-B-458]

When configuring codebase indexing to use an OpenAI-compatible embedding provider like Featherless AI, the API key is never saved, the model list is hardcoded to only OpenAI's own models (ignoring the provider's actual models), and the configuration silently resets or fails to persist.

Summary

Roo Code's codebase indexing feature fails to properly configure and persist settings when using an OpenAI-compatible embedding provider (Featherless AI). Three distinct bugs have been identified:

1. API key is not saved/persisted for the code indexing embedder
2. OpenAI-compatible model list is hardcoded and ignores models actually available from the provider
3. Configuration silently resets when attempting to set up a custom embedding provider

---

Bug 1: API Key Not Saved for Codebase Indexing

The codebase indexing configuration does not store an API key field. When the user configures an OpenAI-compatible embedding provider, there is no mechanism to save the API key, causing authentication failures when the indexer attempts to call the embedding API.

The main API configs section does store an API key:

"apiConfigs": {
  "default": {
    "openAiApiKey": "<key>",
    ...
  }
}

But the codebase indexing config section has no API key field at all:

"codebaseIndexConfig": {
  "codebaseIndexEnabled": true,
  "codebaseIndexQdrantUrl": "http://localhost:6333",
  "codebaseIndexEmbedderProvider": "openai-compatible",
  "codebaseIndexEmbedderBaseUrl": "https://api.featherless.ai/v1",
  "codebaseIndexEmbedderModelId": "Qwen/Qwen3-Embedding-8B",
  "codebaseIndexEmbedderModelDimension": 4096,
  ...
}

There should be a codebaseIndexEmbedderApiKey (or similar) field that stores the API key for the embedding provider, separate from the main chat API key. Currently no such field exists, so the indexer either calls the embedding API without authentication (and fails) or falls back to some default behavior silently.

---

Bug 2: Hardcoded OpenAI-Compatible Model List

The codebaseIndexModels section contains a hardcoded list of models for the openai-compatible provider that does not reflect what the actual provider (Featherless AI) offers:

"codebaseIndexModels": {
  "openai-compatible": {
    "text-embedding-3-small": { "dimension": 1536 },
    "text-embedding-3-large": { "dimension": 3072 },
    "text-embedding-ada-002": { "dimension": 1536 },
    "nomic-embed-code": { "dimension": 3584 }
  }
}

The user's actual model (Qwen/Qwen3-Embedding-8B with dimension 4096) is not in this list. Compare this to the openrouter provider which does include Qwen models:

"openrouter": {
  ...
  "qwen/qwen3-embedding-0.6b": { "dimension": 1024 },
  "qwen/qwen3-embedding-4b": { "dimension": 2560 },
  "qwen/qwen3-embedding-8b": { "dimension": 4096 }
}

The openai-compatible provider should either dynamically fetch available models from the provider's /models endpoint, or allow the user to manually specify any model ID and dimension without requiring it to be in a hardcoded list. Instead, the hardcoded list only contains OpenAI's own embedding models plus nomic-embed-code, and any model outside this list is not recognized even though the user has manually entered it.

---

Bug 3: Configuration Silently Resets / Duplicate Fields

When the user attempts to configure the codebase indexing with a custom OpenAI-compatible provider, the settings appear to reset or fail to persist properly:

1. The codebaseIndexOpenAiCompatibleBaseUrl has a leading space: " https://api.featherless.ai/v1" — the field was not properly trimmed/sanitized on save
2. There are two separate base URL fields for what should be the same thing:
   • codebaseIndexEmbedderBaseUrl (correct: "https://api.featherless.ai/v1")
   • codebaseIndexOpenAiCompatibleBaseUrl (has leading space: " https://api.featherless.ai/v1")

This duplication suggests the settings UI and the backend code are out of sync about which field to use, and input sanitization (trimming whitespace) is not being applied.

Context (who is affected and when)

Any user trying to use a third-party OpenAI-compatible embedding provider (like Featherless AI, LocalAI, Together AI, etc.) with Roo Code's codebase indexing feature. The only workaround is to use a provider that has a hardcoded entry in the model list (OpenAI, Ollama, Gemini, Mistral, Vercel AI Gateway, OpenRouter, or Bedrock). This effectively locks users out of using their own embedding infrastructure.

Reproduction steps

1. Open Roo Code settings
2. Navigate to the codebase indexing configuration section
3. Set the embedder provider to "OpenAI Compatible"
4. Enter a custom base URL (e.g., https://api.featherless.ai/v1)
5. Enter a custom model ID (e.g., Qwen/Qwen3-Embedding-8B)
6. Enter the model dimension (e.g., 4096)
7. Note that there is no field to enter an API key for the embedding provider
8. Save the settings
9. Export settings to JSON and observe:
   • No API key field exists in codebaseIndexConfig
   • The model Qwen/Qwen3-Embedding-8B is not in codebaseIndexModels["openai-compatible"]
   • Two conflicting base URL fields exist, one with a leading space
10. Attempt to use codebase indexing — it fails because the API cannot authenticate

Expected result

• An API key field is available and saved for the codebase indexing embedder - The openai-compatible model list either dynamically fetches from the provider or allows free-form model ID entry - A single, unambiguous base URL field exists with proper input sanitization - Settings persist correctly across saves and restarts

Actual result

• No API key field exists for the codebase indexing embedder — authentication is impossible - The openai-compatible model list is hardcoded to only OpenAI's models; custom models are ignored - Two conflicting base URL fields exist (codebaseIndexEmbedderBaseUrl and codebaseIndexOpenAiCompatibleBaseUrl), one with a leading space from lack of input sanitization - Configuration silently fails to work without any error message to the user

Variations tried (optional)

• Tried using the same API key as the main chat provider — no field exists to enter it
• Tried entering the model ID manually — it's saved in codebaseIndexEmbedderModelId but not recognized because it's missing from the hardcoded codebaseIndexModels["openai-compatible"] list
• Tried different model IDs — same result, only hardcoded models are recognized

App Version

v3.53.0

API Provider (optional)

Featherless AI

Model Used (optional)

Qwen/Qwen3-Embedding-8B (for embeddings), deepseek-ai/DeepSeek-V4-Pro (for chat)

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've analyzed this issue against the current main branch and here's what I found:

Current State of the Three Reported Bugs

Bug 1: API Key Not Saved -- Already Fixed on main

The current codebase does have an API key field for the openai-compatible embedding provider. In CodeIndexPopover.tsx, there is a password field for codebaseIndexOpenAiCompatibleApiKey (around line 952). The backend in webviewMessageHandler.ts saves it via storeSecret("codebaseIndexOpenAiCompatibleApiKey", ...), and config-manager.ts reads it back. This appears to have been fixed after v3.53.0.

Bug 2: Hardcoded Model List -- Partially Addressed

The openai-compatible provider in the UI already uses a free-form text field (VSCodeTextField) for model ID, not a dropdown. Users can type any model ID like Qwen/Qwen3-Embedding-8B. There is also a manual dimension input field. So custom models do work in the UI. However, the EMBEDDING_MODEL_PROFILES["openai-compatible"] list in embeddingModels.ts only contains 4 models, which limits auto-dimension detection for known models.

Bug 3: Input Sanitization / Leading Whitespace -- Still Present

There is no input trimming on URL fields when saving. A leading space in a URL like " https://api.featherless.ai/v1" would be saved as-is and cause silent authentication failures. This affects all URL fields across the codebase indexing settings.

Proposed Implementation Plan

Since Bugs 1 and 2 are largely addressed, the remaining work focuses on Bug 3 (input sanitization) plus some quality-of-life improvements:

1. Add input trimming for URL fields on save

File: webview-ui/src/components/chat/CodeIndexPopover.tsx -- handleSaveSettings() function (~line 530)

Trim whitespace from string fields before sending to the backend. Specifically:

• codebaseIndexQdrantUrl
• codebaseIndexEmbedderBaseUrl
• codebaseIndexOpenAiCompatibleBaseUrl
• All other string config values

2. Expand openai-compatible model profiles for better auto-dimension detection

File: src/shared/embeddingModels.ts

Add popular third-party embedding models to the openai-compatible profiles so dimension auto-detection works for common models like Qwen embedding models (already present in the openrouter list).

3. Add corresponding tests

• Test that URL trimming works correctly in handleSaveSettings
• Test that the new model profiles have correct dimensions

This is a focused change -- the core functionality for openai-compatible providers is already working on main, so the fix primarily improves robustness (trimming) and convenience (more known models).

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[QEU-B-458]

@roomote i approve

[roomote-v0]

I've seen the approval and will proceed with the implementation. The plan focuses on:

1. Adding input trimming for URL fields to prevent leading/trailing whitespace issues
2. Expanding the openai-compatible model profiles with popular third-party embedding models

I'll open a PR to share progress shortly.

[roomote-v0]

Opened PR #12367 with my attempt at addressing this. It adds input trimming for URL fields and expands the openai-compatible embedding model profiles with popular third-party models. Feedback welcome!

#12367