[Claimed #1912] Feat: Add Anthropic CUA adaptive thinking

[github-actions]

Mirrored from external contributor PR #1912 after approval by @miguelg719.

Original author: @chromiebot
Original PR: #1912
Approved source head SHA: 0ea0332c525017c727743d11a68b8eb74f76b646

@chromiebot, please continue any follow-up discussion on this mirrored PR. When the external PR gets new commits, this same internal PR will be marked stale until the latest external commit is approved and refreshed here.

Original description

why

what changed

test plan

---

Summary by cubic

Add adaptive thinking for Anthropic Claude 4.6 models in the CUA client with effort controls and automatic temperature=1. Keeps legacy thinkingBudget for older models and improves model detection and tool versioning.

• New Features

  • Detects 4.6 models (e.g., claude-opus-4-6*, claude-sonnet-4-6*, incl. provider-prefixed) and sends thinking: { type: "adaptive" } with output_config.effort (defaults to "medium").
  • Adds ThinkingEffort and thinkingEffort in ClientOptions ("none" | "low" | "medium" | "high" | "max"); "none" disables adaptive thinking.
  • Forces temperature: 1 with adaptive thinking and logs when overriding a user value; logs when thinkingBudget is provided on 4.6 models.
  • Uses computer_20251124 for 4.6 and claude-opus-4-5-20251101; older models continue with thinking: { type: "enabled", budget_tokens }.

• Migration

  • For 4.6 models, set thinkingEffort; thinkingBudget is ignored.
  • Adaptive thinking forces temperature: 1; set thinkingEffort: "none" to disable.

^{Written for commit ebe95e7. Summary will update on new commits. Review in cubic}

[github-actions]

This mirrored PR has been merged into main. The original external PR #1912 is now completed.

[changeset-bot]

🦋 Changeset detected

Latest commit: ebe95e7

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 4 packages

Name	Type
@browserbasehq/stagehand	Patch
@browserbasehq/stagehand-evals	Patch
@browserbasehq/stagehand-server-v3	Patch
@browserbasehq/stagehand-server-v4	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[cubic-dev-ai]

1 issue found across 3 files

Confidence score: 5/5

• This looks low risk to merge: the only reported issue is a documentation/behavior mismatch with low severity (3/10), not a functional break in core logic.
• In packages/core/lib/v3/types/public/model.ts, docs say ThinkingEffort defaults to high, while runtime behavior uses medium when thinkingEffort is unset, which could mislead integrators about default model behavior.
• Pay close attention to packages/core/lib/v3/types/public/model.ts - align the ThinkingEffort default in docs with the actual adaptive-thinking fallback (medium).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="packages/core/lib/v3/types/public/model.ts">

<violation number="1" location="packages/core/lib/v3/types/public/model.ts:101">
P3: The new `ThinkingEffort` docs claim `high` is the default, but the implementation defaults adaptive thinking to `medium` when `thinkingEffort` is unset.</violation>
</file>

Architecture diagram

sequenceDiagram
    participant App as Application Logic
    participant Client as AnthropicCUAClient
    participant SDK as Anthropic SDK (Beta)
    participant API as Anthropic API

    Note over App,API: Runtime Flow for Adaptive Thinking

    App->>Client: getAction(messages)
    
    Client->>Client: Detect model version (4.6 vs older)

    alt NEW: Model is Claude 4.6 (Opus or Sonnet)
        Client->>Client: NEW: Set thinking.type = "adaptive"
        Client->>Client: NEW: Set output_config.effort = thinkingEffort (default: "medium")
        Client->>Client: NEW: Force temperature = 1
        Note right of Client: Required for adaptive thinking mode
    else CHANGED: Older Claude models (e.g., 4.5)
        opt thinkingBudget provided
            Client->>Client: CHANGED: Set thinking.type = "enabled"
            Client->>Client: CHANGED: Set budget_tokens = thinkingBudget
        end
    end

    Client->>SDK: beta.messages.create({ model, messages, thinking, ... })
    
    Note over SDK,API: Uses computer_20251124 header
    
    SDK->>API: POST /v1/messages
    API-->>SDK: Response (with thinking blocks)
    SDK-->>Client: Message Object
    Client-->>App: Action Result

Loading

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[filip-michalsky]

will this not clash with Opus 4.7 not supporting temp?

[miguelg719]

https://platform.claude.com/docs/en/api/cli/messages/create#create.temperature

[pirate]

needs a test for claude-opus-4-7 to make sure temperature/adapting thinking doesn't break

[miguelg719]

@filip-michalsky @pirate this pr doesn't add support for opus 4.7 on CUA. Regardless, a temperature for default of 1.0 is still supported but deprecated per the documentation here. Any non-default value will throw a 400, but we should scope removing passing temperature as its own PR

[filip-michalsky]

@filip-michalsky @pirate this pr doesn't add support for opus 4.7 on CUA. Regardless, a temperature for default of 1.0 is still supported but deprecated per the documentation here. Any non-default value will throw a 400, but we should scope removing passing temperature as its own PR

sounds good

requested changes are oos for this pr