Remove hardcoded temperature=0 default in prepareCallSettings — breaks models like Kimi K2.5

[hikariming]

Bug Description

prepareCallSettings() in packages/core/src/core/prompt/prepare-call-settings.ts defaults temperature to 0 when not explicitly set:

// TODO v5 remove default 0 for temperature
temperature: temperature != null ? temperature : 0,

This causes hard failures with models that restrict temperature values. For example, Moonshot's Kimi K2.5 only accepts temperature=1 and rejects any other value with:

400 Bad Request: "invalid temperature: only 1 is allowed for this model"

Since most users don't explicitly set temperature (expecting the model's own default), every agent.stream() / agent.generate() call silently forces temperature=0, which breaks these models.

Reproduction

import { Agent } from '@mastra/core/agent';
import { createOpenAI } from '@ai-sdk/openai';

const provider = createOpenAI({
  baseURL: 'https://api.moonshot.cn/v1',
  apiKey: '...',
});

const agent = new Agent({
  name: 'test',
  model: provider.chat('kimi-k2-5'),
  instructions: 'You are a helpful assistant.',
});

// Fails: 400 "invalid temperature: only 1 is allowed for this model"
const result = await agent.stream('Hello');

Expected Behavior

When temperature is not set by the user, it should be undefined so that the AI SDK provider omits it from the request body, letting the model use its own default.

Proposed Fix

One-line change — the TODO comment already exists in the code:

- // TODO v5 remove default 0 for temperature
- temperature: temperature != null ? temperature : 0,
+ temperature,

This is safe because:

• AI SDK providers (@ai-sdk/openai, etc.) correctly strip undefined fields from requests
• Users who explicitly set temperature: 0 still get deterministic behavior
• Only the "didn't set anything" case changes: from forced 0 to model default
• Aligns with how every other parameter (topP, topK, presencePenalty, etc.) is already handled in the same function — none of them have a forced default

Current Workaround

We intercept the fetch call and strip temperature: 0 from the request body before sending, which is fragile and shouldn't be necessary.

Environment

• @mastra/core: 1.22.0
• Model: Kimi K2.5 (Moonshot AI) via @ai-sdk/openai compatible provider

I'm happy to submit a PR for this if the team is open to it — it's a one-line change.

[daneatmastra]

Thank you for reporting this issue! We have labeled it for trio-tb and we will look into it as soon as possible.

If you're experiencing an error, please provide a minimal reproducible example whenever possible to help us resolve it quickly.

[hikariming]

Root cause investigation

After digging deeper, the prepareCallSettings function with temperature ?? 0 is not Mastra's own code — it's bundled from ai@4.3.19 (Vercel AI SDK v4) during @mastra/core's build process.

The current AI SDK v6 (ai@6.x) has already removed this default:

// ai@6.x — fixed, no default
return { temperature, ... };

But @mastra/core bundles the older ai@4.3.19 via @internal/ai-sdk-v4, which still has:

// ai@4.3.19 — hardcoded default
temperature: temperature != null ? temperature : 0,

Possible fixes

1. Upgrade the bundled ai version in @internal/ai-sdk-v4 to a version that doesn't default temperature to 0
2. Patch the prepareCallSettings call in Mastra's own code to explicitly pass temperature: undefined in a way that survives the bundled SDK's default
3. Add a post-processing step in the Agent's stream()/generate() methods to strip temperature: 0 when the user didn't explicitly set it

Option 1 seems cleanest. Happy to help with a PR if you can point me to the right approach for updating the internal AI SDK bundle.

[hikariming]

Root Cause Update

After deeper investigation, the temperature ?? 0 default is not in Mastra's own source code — it comes from the vendored AI SDK v4 dependency:

Dependency chain:

packages/core/src/llm/model/model.ts
  → imports from @internal/ai-sdk-v4 (packages/_vendored/ai_v4)
    → re-exports from ai@^4.3.19
      → ai/core/prompt/prepare-call-settings.ts contains:
        temperature: temperature != null ? temperature : 0  // TODO v5 remove default 0 for temperature

Key findings:

• ai@4.3.19 is the last v4 release — this will never be fixed in v4
• Vercel AI SDK v5/v6 already removed this default (just passes temperature through as-is)
• The TODO comment in the AI SDK v4 source confirms this was always known tech debt
• Since @internal/ai-sdk-v4 is vendored via packages/_vendored/ai_v4, the Mastra team controls the bundling

Possible fix approaches:

1. pnpm patch on ai@4.3.19 to remove the ?? 0 fallback
2. Post-build transform in the vendored package's tsup config to strip the default
3. Migrate to AI SDK v5/v6 for @internal/ai-sdk-v4 (if feasible)

I'll leave the implementation approach to the team since it involves internal vendored dependency management. Happy to help with a PR if you can point me to the preferred approach.

[TylerBarnes]

@hikariming we support v4, v5, and v6, does this run for all versions or just v4? (haven't looked myself)