Skip to content

fix(openai): preserve reasoning for local OpenAI-compatible models via R1 toggle #1119

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ch405canova-sudo wants to merge 4 commits into
base: main
Choose a base branch
from
+23 −1
Open

fix(openai): preserve reasoning for local OpenAI-compatible models via R1 toggle #1119

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
beef72e
fix(openai): preserve reasoning for local OpenAI-compatible models vi…
Aug 4, 2026
f4a69cc
Update src/api/providers/__tests__/openai.spec.ts
ch405canova-sudo Aug 9, 2026
60b1b45
Merge branch 'main' into fix/preserve-reasoning-openai-compatible
ch405canova-sudo Sep 10, 2026
de3725b
Merge branch 'main' into fix/preserve-reasoning-openai-compatible
ch405canova-sudo Sep 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions src/api/providers/__tests__/openai.spec.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1081,6 +1081,20 @@ describe("OpenAiHandler", () => {
expect(model.id).toBe("")
expect(model.info).toBeDefined()
})

it("should set preserveReasoning when openAiR1FormatEnabled is on", () => {
const r1Handler = new OpenAiHandler({
...mockOptions,
openAiR1FormatEnabled: true,
})
const model = r1Handler.getModel()
expect(model.info).toEqual({ ...openAiModelInfoSaneDefaults, preserveReasoning: true })

@edelauna edelauna Sep 11, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Both new tests leave openAiCustomModelInfo unset, so info is value-identical to the defaults here — the { ...info } merge only gets exercised on the defaults path. A case combining custom model info (e.g. a distinct contextWindow) with the toggle would pin the merge for the llama.cpp / LM Studio / Ollama setups this PR targets. Worth adding?

})

it("should not set preserveReasoning by default", () => {
const model = handler.getModel()
expect(model.info.preserveReasoning).toBeUndefined()

@edelauna edelauna Sep 11, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This pins the flag, but the consumer it feeds — buildCleanConversationHistory in Task.ts (preserveReasoning === true at Task.ts:5020) — is only covered by tests that re-implement the branch (reasoning-preservation.test.ts:223, 290, 347, 398), so deleting the production gate would still pass everything. Would a regression test calling the real method be worth adding here?

})
})

describe("Azure AI Inference Service", () => {
Expand Down
10 changes: 9 additions & 1 deletion src/api/providers/openai.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,7 +303,15 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
settings: { ...this.options, reasoningEffort: info.reasoningEffort },
defaultTemperature: 0,
})
return { id, info, ...params }
// Local OpenAI-compatible reasoning models (llama.cpp, LM Studio, Ollama)
// stream reasoning_content, but Roo strips it from the follow-up context
// unless info.preserveReasoning is set. When the user enables the R1 format
// toggle, treat the model as preserving reasoning so the chain is fed back.
return {
id,
info: this.options.openAiR1FormatEnabled ? { ...info, preserveReasoning: true } : info,

@edelauna edelauna Aug 6, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

createMessage() treats a model as R1-format via modelId.includes("deepseek-reasoner") || enabledR1Format (line 90), but this gate only checks the toggle. If someone points openAiModelId at a custom endpoint whose id contains deepseek-reasoner without flipping the toggle, wouldn't createMessage still use R1 conversion while preserveReasoning never gets set here — leaving the original bug open for that case?

@edelauna edelauna Aug 6, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This forces preserveReasoning: true whenever the toggle is on, even if openAiCustomModelInfo.preserveReasoning was explicitly set to false. Should an explicit value win here, e.g. info.preserveReasoning ?? true?

@edelauna edelauna Sep 11, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Line 95 derives the R1 path from modelId.includes("deepseek-reasoner") || enabledR1Format, so a model id containing deepseek-reasoner goes through convertToR1Format while Task.ts still strips its reasoning (preserveReasoning stays unset here). Should both gates share one predicate?

Suggested change
info: this.options.openAiR1FormatEnabled ? { ...info, preserveReasoning: true } : info,
info: (id.includes("deepseek-reasoner") || this.options.openAiR1FormatEnabled) ? { ...info, preserveReasoning: true } : info,

(id is already in scope in getModel; createMessage could reuse the same condition.)

...params,
}
}

async completePrompt(prompt: string, options?: CompletePromptOptions): Promise<string> {
Expand Down
Loading