Anthropic provider: ToolChoiceNone returns nil, tool_choice:{type:“none”} never sent to API

[birdmanmandbir]

Description

Bug

When ToolChoiceNone is set, the Anthropic provider's toTools function returns nil for anthropicToolChoice instead of setting OfNone on the ToolChoiceUnionParam. The caller checks if toolChoice != nil before assigning to params.ToolChoice, so tool_choice: {type: "none"} is never included in the API request. The Anthropic API defaults to auto, meaning the model may still call tools — the opposite of the intended behavior.

Where

providers/anthropic/anthropic.go, in the toTools method:

case fantasy.ToolChoiceNone:
    return anthropicTools, anthropicToolChoice, warnings  // anthropicToolChoice is nil

The caller at the prepareParams level:

if toolChoice != nil {
    params.ToolChoice = *toolChoice
}

Since toolChoice is nil, params.ToolChoice is never set.

Expected

The ToolChoiceNone case should set OfNone using the SDK's constructor:

case fantasy.ToolChoiceNone:
    none := anthropic.NewToolChoiceNoneParam()
    anthropicToolChoice = &anthropic.ToolChoiceUnionParam{
        OfNone: &none,
    }

The Anthropic Go SDK already supports ToolChoiceNoneParam — it just isn't wired up.

How we found it

We're building logos, a Bash-only reasoning engine on top of fantasy. LLMs think in plain text and act with ! rg main.go commands — no tool call at all. Since logos intentionally has no tools, we use ToolChoiceNone to explicitly enforce that the model never attempts tool calls. We discovered the model was still making tool calls on the Anthropic provider despite ToolChoiceNone being set, and traced it to this nil return.

Other providers

• OpenAI provider handles this correctly — sets OfAuto with value "none"
• Google provider handles this correctly
• Anthropic provider is the only one that returns nil for ToolChoiceNone

Version

v0.14.0

Environment

linux, macos

[andreynering]

Thank you for the detailed report! This was just fixed at #178.