Skip to content

fix: drop sampling parameters when Anthropic thinking is enabled (WI-1798) #424

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

Merged
stefan-ernst merged 1 commit into from
Nov 2, 2026
Merged

fix: drop sampling parameters when Anthropic thinking is enabled (WI-1798) #424

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
40 changes: 34 additions & 6 deletions internal/llm/fantasy_adapter.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -228,11 +228,12 @@ func (c *fantasyClient) Complete(ctx context.Context, req CompletionRequest) (*C
c.replay.set(replays, responseRequestOverrides(req, c.protocol))

maxTokens := int64(req.MaxTokens)
providerOptions, thinkingEnabled := c.callProviderOptions(req)
call := fantasy.Call{
Prompt: prompt,
Tools: toFantasyTools(req.Tools, req.EnablePromptCache),
ToolChoice: toFantasyToolChoice(req.ToolChoice),
ProviderOptions: c.callProviderOptions(req),
ProviderOptions: providerOptions,
}
if req.MaxTokens > 0 && (c.protocol != APIContractChatCompletions || c.completionTokens.useLegacy.Load()) {
call.MaxOutputTokens = &maxTokens
Expand All @@ -242,6 +243,13 @@ func (c *fantasyClient) Complete(ctx context.Context, req CompletionRequest) (*C
call.Temperature = &temperature
}
applyGenericCallOptions(&call, c.providerConfig)
if thinkingEnabled {
// Fantasy clears sampling parameters only on its explicit-budget
// thinking path. The effort path leaves them set, and Anthropic rejects
// temperature/top_p alongside thinking.
call.Temperature = nil
call.TopP = nil
}

if req.StructuredOutput != nil && len(req.StructuredOutput.Schema) > 0 && len(req.Tools) == 0 {
return c.generateObject(ctx, call, req.StructuredOutput)
Expand All @@ -251,7 +259,7 @@ func (c *fantasyClient) Complete(ctx context.Context, req CompletionRequest) (*C
mapped := mapFantasyError(err)
if rejectsModernCompletionTokenParameter(mapped) && !c.completionTokens.useLegacy.Load() {
c.completionTokens.useLegacy.Store(true)
call.ProviderOptions = c.callProviderOptions(req)
call.ProviderOptions, _ = c.callProviderOptions(req)
call.MaxOutputTokens = &maxTokens
response, err = c.generate(ctx, call)
}
Expand Down Expand Up @@ -520,7 +528,11 @@ func toFantasyToolChoice(choice any) *fantasy.ToolChoice {
return &mapped
}

func (c *fantasyClient) callProviderOptions(req CompletionRequest) fantasy.ProviderOptions {
// callProviderOptions builds the provider options for a request. The second
// result reports whether this adapter enabled provider thinking (Anthropic
// effort), so the caller can drop sampling parameters the provider rejects
// alongside thinking.
func (c *fantasyClient) callProviderOptions(req CompletionRequest) (fantasy.ProviderOptions, bool) {
if c.protocol == APIContractResponses {
store := false
options := &openai.ResponsesProviderOptions{
Expand All @@ -541,7 +553,7 @@ func (c *fantasyClient) callProviderOptions(req CompletionRequest) fantasy.Provi
}
}
applyResponsesProviderConfig(options, c.providerConfig)
return openai.NewResponsesProviderOptions(options)
return openai.NewResponsesProviderOptions(options), false
}
if c.protocol == APIContractAnthropic {
options := &anthropic.ProviderOptions{}
Expand All @@ -561,14 +573,30 @@ func (c *fantasyClient) callProviderOptions(req CompletionRequest) fantasy.Provi
sendReasoning := true
options.SendReasoning = &sendReasoning
}
return anthropic.NewProviderOptions(options)
return anthropic.NewProviderOptions(options), options.Effort != nil || thinkingOverrideActive(options.ExtraBody)
}
options := &openai.ProviderOptions{}
if req.MaxTokens > 0 && !c.completionTokens.useLegacy.Load() {
maxTokens := int64(req.MaxTokens)
options.MaxCompletionTokens = &maxTokens
}
return openai.NewProviderOptions(options)
return openai.NewProviderOptions(options), false
}

// thinkingOverrideActive reports whether an operator-supplied provider-native
// thinking override enables Anthropic thinking. A missing type, or any type
// other than "disabled", counts as enabled.
func thinkingOverrideActive(extraBody map[string]any) bool {
raw, ok := extraBody["thinking"]
if !ok {
return false
}
config, ok := raw.(map[string]any)
if !ok {
return true
}
kind, _ := config["type"].(string)
return !strings.EqualFold(strings.TrimSpace(kind), "disabled")
}

func applyGenericCallOptions(call *fantasy.Call, raw string) {
Expand Down
Loading