anomalyco/opencode
1
0
Fork 0
mirror of https://github.com/anomalyco/opencode.git synced 2026-07-22 10:15:31 -04:00
Code Issues 5.9k Packages Projects Releases 19 Wiki Activity

Reasoning variants don't work with Vercel AI Gateway provider #8797

New Issue
Closed
opened 2026-02-16 18:10:52 -05:00 by yindo · 2 comments
yindo commented 2026-02-16 18:10:52 -05:00
Owner
Copy Link
Copy Source

Originally created by @AndrewBarba on GitHub (Feb 7, 2026).

Originally assigned to: @thdxr on GitHub.

Description

When using the Vercel AI Gateway provider, built-in reasoning variants are defined but have no effect. Reasoning options are passed into providerOptions.gateway, which does not accept or forward reasoning options to the upstream provider.

This means that when using models like Claude Opus 4.6 or GPT 5.2 through AI Gateway, no thinking/reasoning output is produced and the AI Gateway dashboard shows zero reasoning tokens.

Expected Behavior

Reasoning variants (e.g. thinking/extended thinking) should work correctly when using models through the Vercel AI Gateway, just as they do when using the providers directly.

Actual Behavior

  • No thinking/reasoning output is produced when using reasoning variants through AI Gateway
  • AI Gateway dashboard shows zero reasoning tokens
  • The reasoning variant options are silently ignored because providerOptions.gateway doesn't support them

Root Cause

The Vercel AI Gateway passes through providerOptions for upstream providers (docs), but the built-in variant logic doesn't account for this. It sets reasoning options on the gateway provider options directly rather than targeting the upstream provider's options.

Related

Originally created by @AndrewBarba on GitHub (Feb 7, 2026). Originally assigned to: @thdxr on GitHub. ## Description When using the Vercel AI Gateway provider, built-in reasoning variants are defined but have no effect. Reasoning options are passed into `providerOptions.gateway`, which does not accept or forward reasoning options to the upstream provider. This means that when using models like Claude Opus 4.6 or GPT 5.2 through AI Gateway, no thinking/reasoning output is produced and the AI Gateway dashboard shows zero reasoning tokens. ## Expected Behavior Reasoning variants (e.g. thinking/extended thinking) should work correctly when using models through the Vercel AI Gateway, just as they do when using the providers directly. ## Actual Behavior - No thinking/reasoning output is produced when using reasoning variants through AI Gateway - AI Gateway dashboard shows zero reasoning tokens - The reasoning variant options are silently ignored because `providerOptions.gateway` doesn't support them ## Root Cause The Vercel AI Gateway passes through `providerOptions` for upstream providers ([docs](https://ai-sdk.dev/providers/ai-sdk-providers/ai-gateway#provider-specific-options)), but the built-in variant logic doesn't account for this. It sets reasoning options on the gateway provider options directly rather than targeting the upstream provider's options. ## Related - Fix PR: #12391
yindo closed this issue 2026-02-16 18:10:52 -05:00
yindo commented 2026-02-16 18:10:53 -05:00
Author
Owner
Copy Link
Copy Source

@github-actions[bot] commented on GitHub (Feb 7, 2026):

This issue might be a duplicate of or related to existing issues. Please check:

  • #12396: Gateway provider rejecting thinking/reasoning config for paid models (GLM-4.7, Kimi K2.5, MiniMax M2.1 on Zen)
  • #12155: Missing reasoning variants for multiple providers (Together.AI, OpenAI-compatible, OpenRouter)

The root cause appears to be a systematic issue with how reasoning options are passed through gateway/proxy providers that don't natively support the parameters. The fix PR #12391 may address this pattern across multiple providers.

Feel free to ignore if these don't match your specific use case with Vercel AI Gateway.

@github-actions[bot] commented on GitHub (Feb 7, 2026): This issue might be a duplicate of or related to existing issues. Please check: - #12396: Gateway provider rejecting thinking/reasoning config for paid models (GLM-4.7, Kimi K2.5, MiniMax M2.1 on Zen) - #12155: Missing reasoning variants for multiple providers (Together.AI, OpenAI-compatible, OpenRouter) The root cause appears to be a systematic issue with how reasoning options are passed through gateway/proxy providers that don't natively support the parameters. The fix PR #12391 may address this pattern across multiple providers. Feel free to ignore if these don't match your specific use case with Vercel AI Gateway.
yindo commented 2026-02-16 18:10:53 -05:00
Author
Owner
Copy Link
Copy Source

@AdJIa commented on GitHub (Feb 8, 2026):

Related: We encountered a similar pattern with alibaba-cn (DashScope) provider.

Same root cause: reasoning options are not correctly passed to the underlying provider.

In our case, DashScope's OpenAI-compatible API requires enable_thinking: true in the request body for reasoning models (kimi-k2.5, qwen-plus, deepseek-r1, etc.) to return reasoning_content. Without it, all these models produce zero reasoning tokens.

The existing transform.ts logic only handled:

  1. providerID === "opencode" (via chat_template_args.enable_thinking)
  2. @ai-sdk/anthropic SDK (via thinking.type: "enabled")

But alibaba-cn uses @ai-sdk/openai-compatible, so neither path was taken.

We've submitted PR #12772 to fix this for the alibaba-cn provider. The fix adds a condition to set enable_thinking: true for all reasoning-capable models on alibaba-cn.

Temporary workaround for alibaba-cn users until the PR is merged: add "enable_thinking": true to model options in opencode.json:

"alibaba-cn": {
  "models": {
    "kimi-k2.5": { "options": { "enable_thinking": true } },
    "qwen-plus": { "options": { "enable_thinking": true } }
  }
}
@AdJIa commented on GitHub (Feb 8, 2026): Related: We encountered a similar pattern with `alibaba-cn` (DashScope) provider. **Same root cause**: reasoning options are not correctly passed to the underlying provider. In our case, DashScope's OpenAI-compatible API requires `enable_thinking: true` in the request body for reasoning models (kimi-k2.5, qwen-plus, deepseek-r1, etc.) to return `reasoning_content`. Without it, all these models produce zero reasoning tokens. The existing `transform.ts` logic only handled: 1. `providerID === "opencode"` (via `chat_template_args.enable_thinking`) 2. `@ai-sdk/anthropic` SDK (via `thinking.type: "enabled"`) But `alibaba-cn` uses `@ai-sdk/openai-compatible`, so neither path was taken. We've submitted PR #12772 to fix this for the alibaba-cn provider. The fix adds a condition to set `enable_thinking: true` for all reasoning-capable models on alibaba-cn. Temporary workaround for alibaba-cn users until the PR is merged: add `"enable_thinking": true` to model options in `opencode.json`: ```json "alibaba-cn": { "models": { "kimi-k2.5": { "options": { "enable_thinking": true } }, "qwen-plus": { "options": { "enable_thinking": true } } } } ```
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
acp
bad
bug
discussion
docs
documentation
enhancement
extension
good first issue
growth
help-wanted
mcp
model-problem
needs:compliance
nix
opentui
perf
pull-request
Mirrored from GitHub Pull Request
 tui
web
windows
zen
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
yindo
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: anomalyco/opencode#8797
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?