MCP ask_codebase rejects explicit languageModel: getLanguageModelKey includes displayName which the MCP schema doesn't expose

Bắt đầu trong packages/web/src/features/mcp/askCodebase.ts và kiểm tra cách languageModel được yêu cầu được đối chiếu với các model đã cấu hình; so sánh đường xử lý đó với getLanguageModelKey trong packages/web/src/features/chat/utils.ts. Hoàn thành có nghĩa là ask_codebase chấp nhận dạng {provider, model} của schema cho các model đã cấu hình, đồng thời giữ nguyên hành vi mặc định hiện có và mọi cách sử dụng khác của getLanguageModelKey.

Summary

When calling the MCP tool ask_codebase with an explicit languageModel parameter, the request is rejected with 400 Language model '<provider>/<model>' is not configured. — even when list_language_models clearly shows that exact model configured.

Version

• Sourcebot v4.16.11 (Helm chart 0.1.76)
• MCP client @sourcebot/mcp@1.0.18 (via npx -y @sourcebot/mcp@latest)

Reproduction

Config in sourcebot.json / helm values sourcebot.config:

models:
  - provider: anthropic
    model: claude-sonnet-4-6
    displayName: Claude Sonnet 4.6
    token:
      env: ANTHROPIC_API_KEY
  - provider: anthropic
    model: claude-opus-4-7
    displayName: Claude Opus 4.7 (slow, highest quality)
    token:
      env: ANTHROPIC_API_KEY

1. Call list_language_models via MCP — returns both models as expected:

[
  {"provider":"anthropic","model":"claude-sonnet-4-6","displayName":"Claude Sonnet 4.6"},
  {"provider":"anthropic","model":"claude-opus-4-7","displayName":"Claude Opus 4.7 (slow, highest quality)"}
]

2. Call ask_codebase with an explicit language model of the exact shape the tool's JSON schema requires ({provider, model}):

{
  "query": "...",
  "languageModel": {"provider": "anthropic", "model": "claude-opus-4-7"}
}

Result:

{"statusCode":400,"errorCode":"INVALID_REQUEST_BODY","message":"Language model 'anthropic/claude-opus-4-7' is not configured."}

Same happens for claude-sonnet-4-6 — every explicit model is rejected, regardless of which one.

ask_codebase without languageModel works fine — the implicit default (first model in config) is used correctly. So the failure is specifically in the explicit-selection path.

Root cause

getLanguageModelKey in packages/web/src/features/chat/utils.ts builds the match key from provider + model + displayName:

export const getLanguageModelKey = (model: LanguageModelInfo) => {
    return `${model.provider}-${model.model}-${model.displayName}`;
}

askCodebase uses this function to match the requested model against configured models in packages/web/src/features/mcp/askCodebase.ts:

const matchingModel = configuredModels.find(
    (m) => getLanguageModelKey(m) === getLanguageModelKey(requestedLanguageModel)
);

The MCP JSON schema for ask_codebase.languageModel only requires {provider, model} — displayName is not part of the schema and clients cannot provide it:

languageModel: {
  provider: (required),
  model: (required)
}

So at match time:

• Configured entry key: anthropic-claude-opus-4-7-Claude Opus 4.7 (slow, highest quality)
• Requested entry key: anthropic-claude-opus-4-7-undefined

Keys never match → 400 for any explicit selection.

Suggested fix

Either:

1. Match on provider+model only in askCodebase.ts (simplest — displayName is a cosmetic label, not an identity key):

   const matchingModel = configuredModels.find(
       (m) => m.provider === requestedLanguageModel.provider &&
              m.model    === requestedLanguageModel.model
   );
   

2. Or drop displayName from getLanguageModelKey so the match is consistent with the externally-visible identity.

Option 1 is the minimal change that keeps getLanguageModelKey untouched if it is used elsewhere where displayName matters.

Impact

• MCP clients cannot route a specific question to a non-default model — the dropdown selection in the Web UI works because the UI presumably submits the full LanguageModelInfo (including displayName), but the MCP tool schema intentionally does not expose displayName, so there is no workaround.
• Defaulting to the first model in config still works, so this is not a total outage, but it removes a documented MCP capability.

Happy to open a PR if helpful.