ollie/cmd/olliesrv/internal/backend/reasoning_test.go

package backend

import (
	"context"
	"encoding/json"
	"io"
	"net/http"
	"net/http/httptest"
	"net/url"
	"testing"
)

func TestValidReasoningEffort(t *testing.T) {
	cases := map[string]bool{
		"":       true, // unset means "use default"
		"low":    true,
		"medium": true,
		"high":   true,
		"LOW":    false, // case-sensitive; profiles use lowercase
		"max":    false,
		"hihg":   false, // typo must be rejected, not silently ignored
	}
	for in, want := range cases {
		if got := ValidReasoningEffort(in); got != want {
			t.Errorf("ValidReasoningEffort(%q) = %v, want %v", in, got, want)
		}
	}
}

func TestEffortThinkingBudget(t *testing.T) {
	cases := map[string]int{
		EffortLow:    4096,
		EffortMedium: 8192,
		EffortHigh:   16384,
		"":           0,
		"bogus":      0,
	}
	for in, want := range cases {
		if got := EffortThinkingBudget(in); got != want {
			t.Errorf("EffortThinkingBudget(%q) = %d, want %d", in, got, want)
		}
	}
}

func TestResolvedThinkingBudget(t *testing.T) {
	// Explicit budget wins over effort.
	p := GenerationParams{ThinkingBudget: 5000, ReasoningEffort: EffortHigh}
	if got := p.ResolvedThinkingBudget(); got != 5000 {
		t.Errorf("explicit budget: got %d, want 5000", got)
	}
	// Derived from effort when budget unset.
	p = GenerationParams{ReasoningEffort: EffortHigh}
	if got := p.ResolvedThinkingBudget(); got != 16384 {
		t.Errorf("derived budget: got %d, want 16384", got)
	}
	// Neither set → 0 (no thinking).
	p = GenerationParams{}
	if got := p.ResolvedThinkingBudget(); got != 0 {
		t.Errorf("unset: got %d, want 0", got)
	}
}

// captureAnthropicRequest drives AnthropicBackend.ChatStream against a test
// server and returns the decoded request body.
func captureAnthropicRequest(t *testing.T, params GenerationParams) anthropicRequest {
	t.Helper()
	var raw []byte
	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		raw, _ = io.ReadAll(r.Body)
		w.Header().Set("Content-Type", "text/event-stream")
		// Minimal well-formed stream so the reader terminates cleanly.
		w.Write([]byte("event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n"))
	}))
	defer srv.Close()

	b, err := NewAnthropic("key")
	if err != nil {
		t.Fatal(err)
	}
	u, _ := url.Parse(srv.URL)
	b.baseURL = u

	ch, err := b.ChatStream(context.Background(), []Message{{Role: "user", Content: "audit this"}}, nil, params)
	if err != nil {
		t.Fatal(err)
	}
	for range ch { //nolint:revive
	}

	var areq anthropicRequest
	if err := json.Unmarshal(raw, &areq); err != nil {
		t.Fatalf("decode request: %v (body=%s)", err, raw)
	}
	return areq
}

func TestAnthropic_EffortEnablesThinking(t *testing.T) {
	temp := 0.3
	areq := captureAnthropicRequest(t, GenerationParams{
		ReasoningEffort: EffortHigh,
		MaxTokens:       8192,
		Temperature:     &temp,
	})

	if areq.Thinking == nil {
		t.Fatal("thinking not enabled for reasoningEffort=high")
	}
	if areq.Thinking.BudgetTokens != 16384 {
		t.Errorf("budget_tokens = %d, want 16384", areq.Thinking.BudgetTokens)
	}
	// max_tokens must be grown above the thinking budget (budget < max_tokens).
	if areq.MaxTokens <= areq.Thinking.BudgetTokens {
		t.Errorf("max_tokens (%d) must exceed thinking budget (%d)", areq.MaxTokens, areq.Thinking.BudgetTokens)
	}
	// Extended thinking requires temperature=1.
	if areq.Temperature == nil || *areq.Temperature != 1.0 {
		t.Errorf("temperature = %v, want 1.0 when thinking enabled", areq.Temperature)
	}
}

func TestAnthropic_NoEffortNoThinking(t *testing.T) {
	temp := 0.3
	areq := captureAnthropicRequest(t, GenerationParams{MaxTokens: 8192, Temperature: &temp})
	if areq.Thinking != nil {
		t.Errorf("thinking should be disabled when no effort/budget set")
	}
	// Temperature passes through untouched when thinking is off.
	if areq.Temperature == nil || *areq.Temperature != 0.3 {
		t.Errorf("temperature = %v, want 0.3", areq.Temperature)
	}
}

func TestAnthropic_ExplicitBudgetOverridesEffort(t *testing.T) {
	areq := captureAnthropicRequest(t, GenerationParams{
		ReasoningEffort: EffortLow,
		ThinkingBudget:  20000,
		MaxTokens:       8192,
	})
	if areq.Thinking == nil || areq.Thinking.BudgetTokens != 20000 {
		t.Fatalf("budget_tokens = %v, want 20000 (explicit overrides effort)", areq.Thinking)
	}
	if areq.MaxTokens <= 20000 {
		t.Errorf("max_tokens (%d) must exceed explicit budget (20000)", areq.MaxTokens)
	}
}