[Cursor] Grok 4.5 import sends invalid modelId "grok-4.5" to AgentService

[pengpengzi]

Summary

Importing a local Cursor subscription exposes cursor/grok-4.5 in OpenCodex, but requests fail with:

stream disconnected before completion: Cursor 流结束错误：Model name is not valid: "grok-4.5"

The Cursor subscription and authentication are valid. The failure is caused by OpenCodex sending the imported base slug grok-4.5 directly as the Cursor AgentService modelId, while Cursor Agent CLI requires an effort-qualified selector such as cursor-grok-4.5-high.

Environment

• macOS, Apple Silicon
• OpenCodex 1.1.1 installed locally
• Cursor Agent 2026.07.23-e383d2b
• Cursor Desktop 3.14.7
• The same relevant code is still present in OpenCodex v1.1.2 / current master (20b73d09)

Steps to reproduce

1. Import a local Cursor subscription in OpenCodex.
2. The provider test succeeds and reports that Cursor models were fetched.
3. Select cursor/grok-4.5 with reasoning effort high.
4. Send any prompt through Codex Desktop / /v1/responses.
5. The stream ends with Model name is not valid: "grok-4.5".

Cursor model IDs reported for this account

cursor-agent --list-models reports these valid Grok selectors:

cursor-grok-4.5-low
cursor-grok-4.5-low-fast
cursor-grok-4.5-medium
cursor-grok-4.5-medium-fast
cursor-grok-4.5-high
cursor-grok-4.5-high-fast

It does not report a bare grok-4.5 selector.

Root cause

The imported catalog entry is effectively:

{
  "slug": "cursor/grok-4.5",
  "backend_model": "grok-4.5",
  "backend_provider": "cursor",
  "default_reasoning_level": "medium"
}

The routing path does not combine the backend model with the selected reasoning effort:

• router.ts assigns cursorModelForRequest = ... || upstreamModel.
• router.ts passes that value directly to streamCursorChat.
• cursor_protocol.ts writes it directly into ModelDetails.modelId.

As a result, modelId is grok-4.5, and the separately selected reasoning_effort=high never becomes part of Cursor's required model selector.

The current provider test only proves that AvailableModels can be fetched, so it reports a healthy Cursor subscription even though an actual AgentService generation request fails.

Verified local workaround

Changing only the imported catalog entry to:

"backend_model": "cursor-grok-4.5-high"

made the same local /v1/responses request complete successfully with response.completed. This workaround may be overwritten when models are re-imported.

Suggested fix

1. Normalize Cursor AgentService model selectors before streamCursorChat, using the selected reasoning effort. For this model:

   grok-4.5 + low    -> cursor-grok-4.5-low
   grok-4.5 + medium -> cursor-grok-4.5-medium
   grok-4.5 + high   -> cursor-grok-4.5-high
   

2. Prefer preserving or deriving AgentService-valid IDs from Cursor's actual CLI/model metadata instead of assuming the imported base slug is directly callable.

3. Add an end-to-end provider test that performs a minimal AgentService generation, rather than only fetching the model catalog.

4. Add a regression test asserting that cursor/grok-4.5 with high effort sends modelId=cursor-grok-4.5-high.

Expected behavior

Selecting cursor/grok-4.5 should generate a valid Cursor AgentService model selector from the chosen effort and complete the response normally.