mirror of
https://github.com/jwadow/kiro-gateway.git
synced 2026-04-25 01:15:57 +03:00
[GH-ISSUE #59] BUG: Trying to use "auto" model in cursor #39
Labels
No labels
bug
bug
enhancement
enhancement
fixed
fixed
invalid
needs-info
needs-testing
pull-request
question
upstream
wontfix
workaround
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
starred/kiro-gateway-jwadow#39
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @DemonMartin on GitHub (Jan 29, 2026).
Original GitHub issue: https://github.com/jwadow/kiro-gateway/issues/59
Kiro Gateway Version
v2.1
What happened?
Hey, when calling the /models API, it responds with 6 models:
{"object":"list","data":[{"id":"auto","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"},{"id":"claude-3.7-sonnet","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"},{"id":"claude-haiku-4.5","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"},{"id":"claude-opus-4.5","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"},{"id":"claude-sonnet-4","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"},{"id":"claude-sonnet-4.5","object":"model","created":1769707365,"owned_by":"anthropic","description":"Claude model via Kiro API"}]}
I was trying to use "auto" model in cursor (from kiro, not from cursor), but cursor sadly detects the "auto" as their model causing API issues.
I believe a possible fix would be implementing manual model routing. So for example, I can call it "auto-kiro" and it routes to the "auto" model from kiro.
This would be also useless for other IDEs like the Codex CLI, so a user can overwrite "GPT5.2" with Opus or similar.
Debug Logs
Request ID: 1bd12cb0-77c2-4566-91e2-71218dc8dcaf
{"error":57,"details":{"title":"Provider Error","detail":"We're having trouble connecting to the model provider. This might be temporary - please try again in a moment.","isRetryable":true,"additionalInfo":{},"buttons":[],"planChoices":[]},"isExpected":true}
[resource_exhausted] Error
LTe: [resource_exhausted] Error
at Gmf (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:9095:38348)
at Hmf (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:9095:37251)
at rpf (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:9096:4395)
at fva.run (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:9096:8170)
at async Hyt.runAgentLoop (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:34196:57047)
at async Zpc.streamFromAgentBackend (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:34245:7695)
at async Zpc.getAgentStreamResponse (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:34245:8436)
at async FTe.submitChatMaybeAbortCurrent (vscode-file://vscode-app/c:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:9170:14575)
@DemonMartin commented on GitHub (Jan 29, 2026):
Also, what do I do against him just saying "I will do xyz" now, but it doesnt actually do it and gets registered as "completed"?
I tried opus, sonnet and haiku. All do this.
Example:
@DemonMartin commented on GitHub (Jan 29, 2026):
Here are some logs of kiro gateway if that helps:
https://mclo.gs/pPtxTK3
@jwadow commented on GitHub (Jan 30, 2026):
Hey @DemonMartin, I've implemented a model alias system to fix the
automodel conflict.What's fixed:
The gateway now exposes
auto-kiroinstead ofautoin/v1/modelsto avoid the namespace collision. Both names work - useauto-kiroin Cursor, or keep usingautofor backward compatibility.You can also add custom aliases in the config if needed.
About the second issue (models not executing actions):
I looked at your logs. The gateway is working correctly - it's converting orphaned tool results to text as expected. The issue you're describing sounds like Cursor-specific behavior.
I need to be upfront with you: Cursor support in this gateway is nominal at best. Cursor is... complicated. We had a contributor who's a Cursor power user submit a PR (#49) to improve compatibility, and even he struggled to get everything working smoothly. The IDE does some non-standard things with the OpenAI API that make it difficult to support reliably.
The gateway works great with standard OpenAI-compatible clients, but Cursor has its own quirks that are outside our control. If you're seeing models say "I will do X" without actually doing it, that's likely Cursor's agent logic, not the gateway.
If you want to dig deeper, I'd need more info:
curldirectly)?But honestly, if you're having persistent issues with Cursor, you might have better luck with a different client. The gateway itself is solid - tested with 1297 unit tests.
In any case, this needs a separate issue.
@DemonMartin commented on GitHub (Jan 30, 2026):
@jwadow Thank you very much. I really like Cursor and I tried other IDEs like Antigravity, but I couldn't find anything that comes even near Cursors work efficiency (as in: The work I get done and the quality of it, not related to specific models itself).
Any IDE you could recommend? Trying to switch of cursor because its very expensive and doesn't properly allow having custom models...
@jwadow commented on GitHub (Jan 30, 2026):
@DemonMartin
Personally, I've been using Roo Code for a year now. It's an extension for VS Code. I haven't found anything better, even though it's not perfect. You can customize prompts, configure it for each custom model, use any models, MCPs, basically, whatever your heart desires.
But I'm making my own open source extension, abstract "all agents killer," which I'll enjoy working with and which will have a lot of killer features, even compared to Cursor. The release isn't coming soon, so Roo Code is the best free software out there.
P.S. Don't fall for Kilo Code, they cloned Roo Code repo in a bad way and did a ton of "native advertising" on social media (tutorials, community, etc).