Skip to content

feat: add temperature and other common settings to model-level configuration - #388

Merged
robert-j-y merged 1 commit into
mainfrom
devin/1769809800-issue-387-temperature-settings
Apr 15, 2026
Merged

feat: add temperature and other common settings to model-level configuration#388
robert-j-y merged 1 commit into
mainfrom
devin/1769809800-issue-387-temperature-settings

Conversation

@robert-j-y

@robert-j-y robert-j-y commented Jan 30, 2026

Copy link
Copy Markdown
Contributor

Description

Fixes #387

Adds support for setting default values for common model parameters at model creation time. Previously, settings like temperature could only be passed at call time via generateText/streamText. Now users can set defaults when creating the model:

// Before: TypeScript error - temperature not allowed
const model = openrouter('google/gemini-3-flash-preview', {
  temperature: 0 // ❌ Error
});

// After: Works correctly
const model = openrouter('google/gemini-3-flash-preview', {
  temperature: 0,
  topP: 0.9,
  maxTokens: 100,
});

Call-level options still override model-level settings when provided.

New settings added to OpenRouterSharedSettings:

  • temperature - Controls randomness (0-2)
  • topP - Nucleus sampling (0-1)
  • topK - Top-k sampling
  • frequencyPenalty - Penalize repeated tokens (-2 to 2)
  • presencePenalty - Penalize tokens based on presence (-2 to 2)
  • maxTokens - Maximum tokens to generate

Key areas for review:

Checklist

  • I have run pnpm stylecheck and pnpm typecheck
  • I have run pnpm test and all tests pass
  • I have added tests for my changes (if applicable)
  • I have updated documentation (if applicable)

Changeset

  • I have run pnpm changeset to create a changeset file

Link to Devin run: https://app.devin.ai/sessions/ad14476ac9334b8a962f771f56e48975
Requested by: Robert Yeakel (@robert-j-y)

@seannetlife

seannetlife commented Feb 2, 2026

Copy link
Copy Markdown

ah, this is gonna be awesome to have. ty!

louisgv
louisgv approved these changes Feb 10, 2026
View reviewed changes

@louisgv louisgv left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

LGTM!

@devin-ai-integration @robert-j-y
feat: add temperature and other common settings to model-level config…
89d7000 
…uration

Fixes #387

This change adds support for setting default values for common model parameters
at the model creation level:
- temperature
- topP
- topK
- frequencyPenalty
- presencePenalty
- maxTokens

These settings can be overridden at call time via generateText/streamText options.

Example:
```typescript
const model = openrouter('google/gemini-3-flash-preview', {
  temperature: 0,
  topP: 0.9,
  maxTokens: 100,
});
```

Co-Authored-By: Robert Yeakel <robert.yeakel@openrouter.ai>
@robert-j-y
robert-j-y force-pushed the devin/1769809800-issue-387-temperature-settings branch from 0eb69f3 to 89d7000 Compare April 15, 2026 21:18
@robert-j-y
robert-j-y merged commit 916c4c1 into main Apr 15, 2026
2 checks passed
@robert-j-y
robert-j-y deleted the devin/1769809800-issue-387-temperature-settings branch April 15, 2026 21:45
@github-actions github-actions Bot mentioned this pull request Apr 15, 2026
Merged
robert-j-y added a commit that referenced this pull request Apr 26, 2026
@devin-ai-integration @robert-j-y
fix: restore #419 usage fallback, settings fallbacks, eager_input_str…
3d2c93f 
…eaming, and #413 reasoning-delta metadata fix

Restores four pieces of behavior that were inadvertently removed when the
empty reasoning_details fix was applied:

- #419: Defensive usage fallback in stream finish handler that copies
  openrouterUsage promptTokens/completionTokens into usage.inputTokens.total
  /outputTokens.total when the standard fields are still undefined after
  computeTokenUsage().
- #388: Model-level settings fallbacks (?? this.settings.X) for max_tokens,
  temperature, top_p, frequency_penalty, presence_penalty, top_k.
- #443: eager_input_streaming forwarding from tool providerOptions to the
  function tool request body.
- #413: Stop attaching accumulated reasoning_details snapshots to
  reasoning-start/reasoning-delta events to avoid payload bloat. The
  always-include behavior remains on reasoning-end and the stream finish
  event, which is the only place opencode-style consumers read it from.

Keeps the legitimate empty reasoning_details fix in place so DeepSeek V4
and similar providers continue to receive the empty array on follow-up
turns.

Co-Authored-By: Robert Yeakel <robert.yeakel@openrouter.ai>
robert-j-y added a commit that referenced this pull request Apr 26, 2026
@louisgv @devin-ai-integration @robert-j-y
fix: preserve empty reasoning_details arrays in multi-turn conversati…
4588197 
…ons (#487)

* fix: preserve empty reasoning_details arrays in multi-turn conversations

Always include reasoning_details in finish event metadata and reasoning-end
providerMetadata, even when the accumulated array is empty. DeepSeek V4 and
similar providers require the field to be sent back in subsequent turns to
maintain conversation state.

* fix: treat empty reasoning_details array as explicit metadata signal

Replace length>0 checks with Array.isArray checks so that an explicit []
from message-level providerOptions is treated as 'metadata present but empty'
rather than 'metadata absent'. Preserve [] through dedup/signature-filter path.

* test(perry): update assertions for empty reasoning_details preservation

* test(perry): update signature-roundtrip assertions for empty reasoning_details

* test(perry): update stream usage accounting assertion for reasoning_details

* test(perry): update doStream finish event snapshots for reasoning_details

* chore(perry): add patch changeset for empty reasoning_details fix

* fix: restore #419 usage fallback, settings fallbacks, eager_input_streaming, and #413 reasoning-delta metadata fix

Restores four pieces of behavior that were inadvertently removed when the
empty reasoning_details fix was applied:

- #419: Defensive usage fallback in stream finish handler that copies
  openrouterUsage promptTokens/completionTokens into usage.inputTokens.total
  /outputTokens.total when the standard fields are still undefined after
  computeTokenUsage().
- #388: Model-level settings fallbacks (?? this.settings.X) for max_tokens,
  temperature, top_p, frequency_penalty, presence_penalty, top_k.
- #443: eager_input_streaming forwarding from tool providerOptions to the
  function tool request body.
- #413: Stop attaching accumulated reasoning_details snapshots to
  reasoning-start/reasoning-delta events to avoid payload bloat. The
  always-include behavior remains on reasoning-end and the stream finish
  event, which is the only place opencode-style consumers read it from.

Keeps the legitimate empty reasoning_details fix in place so DeepSeek V4
and similar providers continue to receive the empty array on follow-up
turns.

Co-Authored-By: Robert Yeakel <robert.yeakel@openrouter.ai>

* test: restore deleted #419/eager_input_streaming/#413 test coverage

Restores the seven tests that were inadvertently deleted along with the
empty reasoning_details fix and flips the reasoning-start/reasoning-delta
providerMetadata assertions back to toBeUndefined() to lock in the
restored #413 behavior.

- 3 tests for eager_input_streaming forwarding from tool providerOptions.
- 4 tests for the #419 defensive usage fallback in the stream finish handler.
- Adjusts the two streaming reasoning_details tests to assert that
  reasoning-start and reasoning-delta events do NOT carry providerMetadata
  (the snapshot is exposed only on reasoning-end and the finish event).

Co-Authored-By: Robert Yeakel <robert.yeakel@openrouter.ai>

---------

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Robert Yeakel <robert.yeakel@openrouter.ai>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@louisgv louisgv louisgv approved these changes
@ping-Toven ping-Toven Awaiting requested review from ping-Toven

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Support temperature settings in OpenRouterChatSettings. (Missing)

3 participants

@robert-j-y @seannetlife @louisgv