feat(llm): client interface and fake test double

internal/llm/client.go

@@ -0,0 +1,36 @@
+// Package llm defines a provider-agnostic LLM client for vision and audio
+// transcription, with concrete HTTP implementations per provider.
+package llm
+
+import "context"
+
+// Image is a decoded image to send to a vision model.
+type Image struct {
+	MediaType string // e.g. "image/png", "image/jpeg"
+	Data      []byte
+}
+
+// Options are per-call overrides. Zero values mean "use the client default".
+type Options struct {
+	Model       string
+	Language    string
+	Temperature float64
+	MaxTokens   int
+}
+
+// Client is the abstraction every provider implements and every handler depends on.
+type Client interface {
+	// Vision sends a prompt plus images and returns the model's text answer.
+	Vision(ctx context.Context, prompt string, images []Image, opts Options) (string, error)
+	// Transcribe converts audio bytes to text. filename carries the extension
+	// for providers that infer the audio format from it. Returns ErrUnsupported
+	// if the provider has no audio capability.
+	Transcribe(ctx context.Context, audio []byte, filename string, opts Options) (string, error)
+}
+
+// ErrUnsupported is returned by Transcribe on providers without audio support.
+type ErrUnsupported struct{ Provider string }
+
+func (e ErrUnsupported) Error() string {
+	return "audio transcription unsupported by provider " + e.Provider
+}

internal/llm/fake.go

@@ -0,0 +1,39 @@
+package llm
+
+import "context"
+
+// Fake is a deterministic Client for tests.
+type Fake struct {
+	VisionResp      string
+	TranscribeResp  string
+	VisionErr       error
+	TranscribeErr   error
+	VisionCalls     []VisionCall
+	TranscribeCalls []TranscribeCall
+}
+
+// VisionCall records a Vision invocation.
+type VisionCall struct {
+	Prompt string
+	Opts   Options
+	Images []Image
+}
+
+// TranscribeCall records a Transcribe invocation.
+type TranscribeCall struct {
+	Filename string
+	Opts     Options
+	Audio    []byte
+}
+
+// Vision implements Client.
+func (f *Fake) Vision(_ context.Context, prompt string, images []Image, opts Options) (string, error) {
+	f.VisionCalls = append(f.VisionCalls, VisionCall{Prompt: prompt, Images: images, Opts: opts})
+	return f.VisionResp, f.VisionErr
+}
+
+// Transcribe implements Client.
+func (f *Fake) Transcribe(_ context.Context, audio []byte, filename string, opts Options) (string, error) {
+	f.TranscribeCalls = append(f.TranscribeCalls, TranscribeCall{Audio: audio, Filename: filename, Opts: opts})
+	return f.TranscribeResp, f.TranscribeErr
+}

internal/llm/fake_test.go

@@ -0,0 +1,21 @@
+package llm
+
+import (
+	"context"
+	"testing"
+)
+
+func TestFakeClient(t *testing.T) {
+	f := &Fake{VisionResp: "hello", TranscribeResp: "spoken"}
+	got, err := f.Vision(context.Background(), "p", []Image{{MediaType: "image/png", Data: []byte{1}}}, Options{})
+	if err != nil || got != "hello" {
+		t.Fatalf("vision got %q err %v", got, err)
+	}
+	if len(f.VisionCalls) != 1 || f.VisionCalls[0].Prompt != "p" {
+		t.Fatalf("vision call not recorded: %+v", f.VisionCalls)
+	}
+	tr, err := f.Transcribe(context.Background(), []byte{1}, "audio.mp3", Options{Language: "en"})
+	if err != nil || tr != "spoken" {
+		t.Fatalf("transcribe got %q err %v", tr, err)
+	}
+}