README.md

@@ -175,7 +175,9 @@ LLM_LS=llama-swaps://token@swap.example.com # https → TLS-fronted instance
 [llama-swap](https://github.com/mostlygeek/llama-swap) is a model-swapping proxy
 over llama.cpp. Its chat API is OpenAI-compatible (majordomo reuses the openai
 client), and the `*llamaswap.Provider` adds management methods
-(`ListModels`/`Running`/`Unload`) plus image generation (see below). A cold
+(`ListModels`/`Running`/`Unload`/`ListVoices`), a cheap liveness probe
+(`Health`, GET /health — the proxy answers without touching a model), image
+generation and editing, and speech synthesis/transcription (see below). A cold
 model swap can take many seconds — bound calls with a context deadline, not a
 client timeout.
 
@@ -211,8 +213,9 @@ resp, err := m.Generate(ctx, majordomo.Request{
 Text-to-image is a separate contract (`imagegen`) from chat, because it shares
 none of the message/tool/stream machinery. Generated images come back as
 `llm.ImagePart`, so they drop straight back into a chat turn. The first backend
-is llama-swap (OpenAI `/v1/images/generations` → a stable-diffusion.cpp
-upstream).
+is llama-swap (the A1111-style `/sdapi/v1/txt2img` on a stable-diffusion.cpp
+upstream — chosen over OpenAI `/v1/images/generations` because that route
+ignores `seed` there).
 
 ```go
 ls := llamaswap.New(llamaswap.WithBaseURL("http://box.local:8080"))
@@ -224,9 +227,45 @@ res, err := im.Generate(ctx, imagegen.Request{Prompt: "a red bicycle"},
 // majordomo.UserParts(majordomo.Text("describe this"), res.Images[0])
 ```
 
+Image-to-image editing is the optional `imagegen.Editor` interface (ADR-0018)
+— llama-swap's image models implement it via `/sdapi/v1/img2img`:
+
+```go
+ed := im.(imagegen.Editor)
+res, err := ed.Edit(ctx, imagegen.EditRequest{
+    Prompt: "make it night",
+    Init:   res.Images[0],           // any llm.ImagePart
+}, imagegen.WithEditStrength(0.6))   // 0..1: how far to depart from Init
+```
+
 `*llamaswap.Provider` also exposes management methods: `ListModels` (what
 llama-swap can serve), `Running` (what's loaded), and `Unload` (free a model).
 
+## Speech: synthesis + transcription
+
+Text-to-speech and speech-to-text live in the `audio` package (ADR-0017),
+mirroring imagegen: small `SpeechModel`/`TranscriptionModel` contracts,
+zero values mean backend defaults, bytes in/out (never URLs). First backend:
+llama-swap (OpenAI `/v1/audio/speech` + `/v1/audio/transcriptions` routed to
+kokoro/whisper.cpp-style upstreams).
+
+```go
+ls := llamaswap.New(llamaswap.WithBaseURL("http://box.local:8080"))
+
+sm, _ := ls.SpeechModel("kokoro")
+speech, err := sm.Speak(ctx, audio.SpeechRequest{Input: "hello world"},
+    audio.WithVoice("af_heart"), audio.WithFormat("mp3"))
+// speech.Audio ([]byte) + speech.MIME ("audio/mpeg")
+
+tm, _ := ls.TranscriptionModel("whisper-large-v3-turbo")
+tr, err := tm.Transcribe(ctx, audio.TranscriptionRequest{
+    Audio: speech.Audio, MIME: speech.MIME,
+})
+// tr.Text
+
+voices, err := ls.ListVoices(ctx, "kokoro") // []string of voice ids
+```
+
 ## Tool calls
 
 ```go
@@ -358,8 +397,9 @@ response as a single delta plus final event.
 ² llama-swap's chat is OpenAI-compatible and reuses the openai client, so these
 capabilities are present at the client level; whether a given call succeeds
 depends on the llama.cpp model llama-swap loads. llama-swap also provides
-**image generation** (a separate `imagegen` axis, not shown above) and
-management methods on `*llamaswap.Provider`.
+**image generation + editing** (`imagegen`), **speech synthesis +
+transcription** (`audio`) — separate axes, not shown above — plus a `Health`
+probe and management methods on `*llamaswap.Provider`.
 
 Notes: Ollama has no native tool_choice — `"none"` drops the tools;
 `"required"`/named choices are best-effort ignored there. Ollama Cloud

audio/audio.go

@@ -0,0 +1,186 @@
+// Package audio is majordomo's canonical speech surface: text-to-speech
+// (SpeechModel) and audio transcription (TranscriptionModel). Like imagegen,
+// it is a deliberately separate contract from the llm package — synthesis and
+// transcription share none of the chat message/tool/stream machinery, so they
+// get their own small Provider/Model interfaces rather than overloading
+// llm.Model (ADR-0017).
+//
+// Zero values mean "backend default" throughout, mirroring imagegen: an empty
+// Voice uses the model's default voice, an empty Format the backend's default
+// container, a zero Speed the natural rate.
+//
+// The first implementation is provider/llamaswap, which targets the OpenAI
+// /v1/audio/speech and /v1/audio/transcriptions endpoints routed to
+// kokoro/whisper.cpp-style upstreams.
+package audio
+
+import "context"
+
+// SpeechRequest is a text-to-speech request.
+type SpeechRequest struct {
+	// Input is the text to speak.
+	Input string
+
+	// Voice selects the voice; "" = the model's default voice.
+	Voice string
+
+	// Format is the audio container ("mp3", "wav", "opus", ...);
+	// "" = backend default.
+	Format string
+
+	// Speed is the playback-rate multiplier; 0 = backend default (1.0).
+	Speed float64
+}
+
+// SpeechResult is the canonical synthesis result: raw audio bytes plus the
+// MIME type reported (or implied) by the backend.
+type SpeechResult struct {
+	// Audio is the encoded audio.
+	Audio []byte
+
+	// MIME is the audio MIME type, e.g. "audio/mpeg".
+	MIME string
+
+	// Raw is the provider-native response object, an escape hatch for
+	// provider-specific fields. May be nil; never required for normal use.
+	Raw any
+}
+
+// SpeechOption mutates a SpeechRequest before it is sent. Options passed to
+// Speak are applied to a copy, so a request value can be reused.
+type SpeechOption func(*SpeechRequest)
+
+// WithVoice selects the voice.
+func WithVoice(v string) SpeechOption { return func(r *SpeechRequest) { r.Voice = v } }
+
+// WithFormat sets the audio container format (e.g. "mp3", "wav").
+func WithFormat(f string) SpeechOption { return func(r *SpeechRequest) { r.Format = f } }
+
+// WithSpeed sets the playback-rate multiplier.
+func WithSpeed(s float64) SpeechOption { return func(r *SpeechRequest) { r.Speed = s } }
+
+// Apply returns a copy of the request with all options applied. Providers
+// call this once at the top of Speak.
+func (r SpeechRequest) Apply(opts ...SpeechOption) SpeechRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// SpeechModel synthesizes speech from text.
+type SpeechModel interface {
+	// Speak renders the request's input text as audio.
+	Speak(ctx context.Context, req SpeechRequest, opts ...SpeechOption) (*SpeechResult, error)
+}
+
+// SpeechModelOption configures a SpeechModel at construction time. Reserved
+// for future per-model settings; present so the interface is
+// forward-compatible (mirrors imagegen.ModelOption).
+type SpeechModelOption func(*SpeechModelConfig)
+
+// SpeechModelConfig carries per-model construction settings.
+type SpeechModelConfig struct{}
+
+// ApplySpeechModelOptions folds options into a config.
+func ApplySpeechModelOptions(opts []SpeechModelOption) SpeechModelConfig {
+	var cfg SpeechModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// SpeechProvider mints speech models bound to one backend.
+type SpeechProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// SpeechModel returns a SpeechModel bound to the given id (passed through
+	// to the backend verbatim; no catalog validation).
+	SpeechModel(id string, opts ...SpeechModelOption) (SpeechModel, error)
+}
+
+// TranscriptionRequest is a speech-to-text request. Audio is carried as bytes
+// (never a URL), mirroring llm.ImagePart's bytes-only contract.
+type TranscriptionRequest struct {
+	// Audio is the encoded audio to transcribe.
+	Audio []byte
+
+	// MIME is the audio MIME type (e.g. "audio/mpeg"); "" = let the backend
+	// sniff it.
+	MIME string
+
+	// Filename is the multipart filename hint some backends key their format
+	// detection on; "" derives one from MIME ("audio.mp3") or falls back to
+	// "audio".
+	Filename string
+
+	// Language is a BCP-47/ISO-639 hint (e.g. "en"); "" = auto-detect.
+	Language string
+
+	// Prompt is optional context or vocabulary to bias decoding; "" = none.
+	Prompt string
+}
+
+// TranscriptionResult is the canonical transcription result.
+type TranscriptionResult struct {
+	// Text is the transcript.
+	Text string
+
+	// Raw is the provider-native response object. May be nil.
+	Raw any
+}
+
+// TranscriptionOption mutates a TranscriptionRequest before it is sent.
+type TranscriptionOption func(*TranscriptionRequest)
+
+// WithLanguage sets the language hint (e.g. "en").
+func WithLanguage(l string) TranscriptionOption {
+	return func(r *TranscriptionRequest) { r.Language = l }
+}
+
+// WithPrompt sets the decoding context/vocabulary hint.
+func WithPrompt(p string) TranscriptionOption {
+	return func(r *TranscriptionRequest) { r.Prompt = p }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r TranscriptionRequest) Apply(opts ...TranscriptionOption) TranscriptionRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// TranscriptionModel transcribes audio to text.
+type TranscriptionModel interface {
+	// Transcribe converts the request's audio into text.
+	Transcribe(ctx context.Context, req TranscriptionRequest, opts ...TranscriptionOption) (*TranscriptionResult, error)
+}
+
+// TranscriptionModelOption configures a TranscriptionModel at construction
+// time. Reserved for future per-model settings.
+type TranscriptionModelOption func(*TranscriptionModelConfig)
+
+// TranscriptionModelConfig carries per-model construction settings.
+type TranscriptionModelConfig struct{}
+
+// ApplyTranscriptionModelOptions folds options into a config.
+func ApplyTranscriptionModelOptions(opts []TranscriptionModelOption) TranscriptionModelConfig {
+	var cfg TranscriptionModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// TranscriptionProvider mints transcription models bound to one backend.
+type TranscriptionProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// TranscriptionModel returns a TranscriptionModel bound to the given id
+	// (passed through to the backend verbatim; no catalog validation).
+	TranscriptionModel(id string, opts ...TranscriptionModelOption) (TranscriptionModel, error)
+}

audio/audio_test.go

@@ -0,0 +1,38 @@
+package audio
+
+import "testing"
+
+func TestSpeechRequestApply(t *testing.T) {
+	base := SpeechRequest{Input: "hello"}
+	got := base.Apply(WithVoice("af_heart"), WithFormat("wav"), WithSpeed(1.5))
+
+	if got.Input != "hello" {
+		t.Errorf("Input = %q, want %q", got.Input, "hello")
+	}
+	if got.Voice != "af_heart" || got.Format != "wav" || got.Speed != 1.5 {
+		t.Errorf("got = %+v", got)
+	}
+
+	// Apply must not mutate the receiver (options apply to a copy).
+	if base.Voice != "" || base.Format != "" || base.Speed != 0 {
+		t.Errorf("base mutated: %+v", base)
+	}
+}
+
+func TestTranscriptionRequestApply(t *testing.T) {
+	base := TranscriptionRequest{Audio: []byte{1, 2, 3}}
+	got := base.Apply(WithLanguage("en"), WithPrompt("names"))
+
+	if got.Language != "en" || got.Prompt != "names" {
+		t.Errorf("got = %+v", got)
+	}
+	if base.Language != "" || base.Prompt != "" {
+		t.Errorf("base mutated: %+v", base)
+	}
+}
+
+func TestApplyModelOptions(t *testing.T) {
+	// No options yet; just verify they return usable zero configs.
+	_ = ApplySpeechModelOptions(nil)
+	_ = ApplyTranscriptionModelOptions(nil)
+}

docs/adr/0017-audio-interfaces.md

@@ -0,0 +1,53 @@
+# ADR-0017: audio — canonical speech synthesis + transcription interfaces
+
+**Status:** Accepted — 2026-07-11
+
+## Context
+
+mort is growing agent tools that speak (TTS) and transcribe audio through a
+llama-swap host whose upstreams expose the OpenAI `/v1/audio/speech` and
+`/v1/audio/transcriptions` endpoints (kokoro, chatterbox, whisper.cpp). Like
+image generation before it (ADR-0016), speech shares none of the chat
+contract's message/tool/stream machinery, and majordomo had no speech surface
+— an earlier migration doc explicitly scoped transcription out of the llm
+package. The same reasoning that produced `imagegen` applies.
+
+## Decision
+
+- One new canonical **leaf package `audio`** holding both directions —
+  synthesis and transcription — rather than two packages; they are one
+  modality and will share future types (voice metadata, audio formats).
+  Root re-exports mirror imagegen (`SpeechModel`, `SpeechProvider`,
+  `SpeechRequest`, `SpeechResult`, `TranscriptionModel`, ...).
+- Minimal v1 surface, imagegen conventions throughout (zero value = backend
+  default, functional options + `Apply`, `Raw any` escape hatch):
+  - `SpeechRequest{ Input; Voice; Format; Speed }` →
+    `SpeechResult{ Audio []byte; MIME string; Raw }`;
+    `SpeechModel.Speak(ctx, req, ...opts)`;
+    `SpeechProvider.SpeechModel(id, ...)`.
+  - `TranscriptionRequest{ Audio []byte; MIME; Filename; Language; Prompt }` →
+    `TranscriptionResult{ Text string; Raw }`;
+    `TranscriptionModel.Transcribe(ctx, req, ...opts)`;
+    `TranscriptionProvider.TranscriptionModel(id, ...)`.
+- **Bytes in/out, never URLs** — mirrors `llm.ImagePart`'s bytes-only
+  contract; fetching is the caller's concern.
+- Providers are split (`SpeechProvider` vs `TranscriptionProvider`) so a
+  backend can implement either half; llamaswap implements both.
+- First implementation: `provider/llamaswap` — `/v1/audio/speech` (JSON body,
+  raw audio response; MIME from Content-Type with a format-based fallback),
+  `/v1/audio/transcriptions` (multipart, `response_format=json`), plus
+  `ListVoices(ctx, model)` (GET `/v1/audio/voices?model=`, tolerant of the
+  string-list and object-list shapes upstreams use) as a llamaswap management
+  method, not part of the canonical contract.
+- Out of scope for v1 (designed-for, deferred): streaming synthesis,
+  word-level timestamps/segments, translation, voice cloning inputs, and
+  registry-level DSN resolution for audio models.
+
+## Consequences
+
+- Speech is provider-agnostic from day one; an OpenAI or Google speech
+  backend implements the same interfaces.
+- `SpeechResult.Audio` is a plain byte slice, so results flow into any file
+  store or attachment pipeline without a majordomo dependency.
+- The `audio` package name is the modality, not the direction; if music/sfx
+  generation ever lands it has a home.

docs/adr/0018-imagegen-editor.md

@@ -0,0 +1,36 @@
+# ADR-0018: imagegen.Editor — image-to-image as a separate optional interface
+
+**Status:** Accepted — 2026-07-11
+
+## Context
+
+ADR-0016 shipped text-to-image and explicitly deferred img2img. mort's new
+llama-swap media tools need "edit this image under this prompt"
+(image-to-image with a denoising strength). Two shape questions: does Edit
+belong on `imagegen.Model`, and which llama-swap endpoint carries it —
+OpenAI-style `/v1/images/edits` (multipart) or A1111-style `/sdapi/v1/img2img`
+(JSON)?
+
+## Decision
+
+- **`Editor` is a separate, optional interface** (`Edit(ctx, EditRequest,
+  ...EditOption) (*Result, error)`), not a new method on `Model`. Existing
+  `Model` implementations keep compiling; callers type-assert
+  (`m.(imagegen.Editor)`) or require the capability explicitly.
+- `EditRequest` = the generation knobs (prompt, N, size, steps, cfg, negative
+  prompt, sampler, seed) plus `Init Image` (required) and `Strength *float64`
+  (denoising strength in [0,1]; nil = backend default). Same option/Apply
+  conventions; result type is the shared `Result`.
+- llama-swap implements it via **`/sdapi/v1/img2img`**, not
+  `/v1/images/edits`: the same sd-server build that ignores `seed` on the
+  OpenAI images route (the reason txt2img went SDAPI in ADR-0016's
+  implementation) applies; the JSON shape is txt2img's plus
+  `init_images: ["<b64>"]` + `denoising_strength`, so it reuses `doJSON`
+  verbatim, where the OpenAI route is multipart.
+
+## Consequences
+
+- Backends that can't edit simply don't implement `Editor`; no stub methods.
+- The init image travels base64-inline in JSON (~33% overhead) — acceptable at
+  chat-image sizes; a future backend needing multipart can still satisfy the
+  same interface.

docs/adr/README.md

@@ -20,3 +20,5 @@ One decision per file, append-only; supersede rather than rewrite.
 | [0014](0014-conversion-driven-extensions.md) | Conversion-driven extensions (resolvers, typed tools, hooks, ops controls) | Accepted |
 | [0015](0015-llama-swap-provider.md) | llama-swap provider — reuse openai for chat, tailored management + image | Accepted |
 | [0016](0016-imagegen-interface.md) | imagegen — a canonical text-to-image interface | Accepted |
+| [0017](0017-audio-interfaces.md) | audio — canonical speech synthesis + transcription interfaces | Accepted |
+| [0018](0018-imagegen-editor.md) | imagegen.Editor — image-to-image as a separate optional interface | Accepted |

imagegen/edit.go

@@ -0,0 +1,91 @@
+package imagegen
+
+import "context"
+
+// EditRequest is an image-to-image (edit) request: a prompt applied to an
+// initial image. As with Request, zero values mean "backend default"
+// (ADR-0018).
+type EditRequest struct {
+	// Prompt is the text description of the desired edit.
+	Prompt string
+
+	// Init is the initial image the edit starts from. Required.
+	Init Image
+
+	// Strength is the denoising strength in [0,1] — how far the result may
+	// depart from Init (0 = return the input, 1 = ignore it); nil = backend
+	// default.
+	Strength *float64
+
+	// N is the number of images to generate; 0 = provider default.
+	N int
+
+	// Size is the requested resolution, e.g. "1024x1024"; "" = provider
+	// default (usually the init image's own resolution).
+	Size string
+
+	// Steps is the number of diffusion steps; nil = backend default.
+	Steps *int
+
+	// CFGScale is the classifier-free-guidance scale; nil = backend default.
+	CFGScale *float64
+
+	// NegativePrompt steers generation away from concepts; "" = none.
+	NegativePrompt string
+
+	// Sampler selects the sampling method (e.g. "euler", "euler_a");
+	// "" = backend default.
+	Sampler string
+
+	// Seed fixes the RNG seed for reproducible output; nil = random.
+	Seed *int64
+}
+
+// EditOption mutates an EditRequest before it is sent. Options passed to Edit
+// are applied to a copy of the request, so an EditRequest value can be reused.
+type EditOption func(*EditRequest)
+
+// WithEditStrength sets the denoising strength in [0,1].
+func WithEditStrength(s float64) EditOption { return func(r *EditRequest) { r.Strength = &s } }
+
+// WithEditN sets the number of images to generate.
+func WithEditN(n int) EditOption { return func(r *EditRequest) { r.N = n } }
+
+// WithEditSize sets the requested resolution (e.g. "1024x1024").
+func WithEditSize(size string) EditOption { return func(r *EditRequest) { r.Size = size } }
+
+// WithEditSteps overrides the number of diffusion steps.
+func WithEditSteps(n int) EditOption { return func(r *EditRequest) { r.Steps = &n } }
+
+// WithEditCFGScale overrides the classifier-free-guidance scale.
+func WithEditCFGScale(s float64) EditOption { return func(r *EditRequest) { r.CFGScale = &s } }
+
+// WithEditNegativePrompt sets a negative prompt.
+func WithEditNegativePrompt(s string) EditOption {
+	return func(r *EditRequest) { r.NegativePrompt = s }
+}
+
+// WithEditSampler overrides the sampling method.
+func WithEditSampler(s string) EditOption { return func(r *EditRequest) { r.Sampler = s } }
+
+// WithEditSeed fixes the RNG seed for reproducible output.
+func WithEditSeed(seed int64) EditOption { return func(r *EditRequest) { r.Seed = &seed } }
+
+// Apply returns a copy of the request with all options applied. Providers
+// call this once at the top of Edit.
+func (r EditRequest) Apply(opts ...EditOption) EditRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Editor is the image-to-image surface. It is a separate, optional interface
+// rather than a method on Model so existing Model implementations keep
+// compiling; callers type-assert (`m.(imagegen.Editor)`) or require it
+// explicitly.
+type Editor interface {
+	// Edit produces one or more images derived from the request's init image
+	// under the request's prompt.
+	Edit(ctx context.Context, req EditRequest, opts ...EditOption) (*Result, error)
+}

imagegen/edit_test.go

@@ -0,0 +1,26 @@
+package imagegen
+
+import "testing"
+
+func TestEditRequestApply(t *testing.T) {
+	base := EditRequest{Prompt: "make it night", Init: Image{MIME: "image/png", Data: []byte{1}}}
+	got := base.Apply(WithEditStrength(0.7), WithEditN(2), WithEditSeed(42))
+
+	if got.Prompt != "make it night" || len(got.Init.Data) != 1 {
+		t.Errorf("got = %+v", got)
+	}
+	if got.Strength == nil || *got.Strength != 0.7 {
+		t.Errorf("Strength = %v, want 0.7", got.Strength)
+	}
+	if got.N != 2 {
+		t.Errorf("N = %d, want 2", got.N)
+	}
+	if got.Seed == nil || *got.Seed != 42 {
+		t.Errorf("Seed = %v, want 42", got.Seed)
+	}
+
+	// Apply must not mutate the receiver (options apply to a copy).
+	if base.Strength != nil || base.N != 0 || base.Seed != nil {
+		t.Errorf("base mutated: %+v", base)
+	}
+}

majordomo.go

@@ -26,6 +26,7 @@ import (
 	"encoding/json"
 	"sync"
 
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
 )
@@ -67,6 +68,26 @@ type (
 	ImageResult      = imagegen.Result
 	ImageOption      = imagegen.Option
 	ImageModelOption = imagegen.ModelOption
+	ImageEditor      = imagegen.Editor
+	ImageEditRequest = imagegen.EditRequest
+	ImageEditOption  = imagegen.EditOption
+)
+
+// Re-exported canonical speech types. See the audio package for
+// documentation. Speech synthesis and transcription are separate contracts
+// from llm, mirroring imagegen (ADR-0017); the first backend is
+// provider/llamaswap.
+type (
+	SpeechModel           = audio.SpeechModel
+	SpeechProvider        = audio.SpeechProvider
+	SpeechRequest         = audio.SpeechRequest
+	SpeechResult          = audio.SpeechResult
+	SpeechOption          = audio.SpeechOption
+	TranscriptionModel    = audio.TranscriptionModel
+	TranscriptionProvider = audio.TranscriptionProvider
+	TranscriptionRequest  = audio.TranscriptionRequest
+	TranscriptionResult   = audio.TranscriptionResult
+	TranscriptionOption   = audio.TranscriptionOption
 )
 
 // Re-exported role and finish-reason constants.

progress.md

@@ -248,3 +248,19 @@ alias-in-chain failover, permanent-policy override) and wires anything the
 tests flush out.
 
 **Next:** Phase 2 — exhaustive health/chain test matrix.
+
+## 2026-07-11 — audio surfaces + image editing (ADR-0017, ADR-0018)
+
+- New leaf package `audio`: `SpeechModel`/`SpeechProvider` (TTS) and
+  `TranscriptionModel`/`TranscriptionProvider` (STT), imagegen conventions
+  (zero value = backend default, options + Apply, bytes in/out). Root
+  re-exports added.
+- `imagegen.Editor` — optional img2img interface (`EditRequest` with
+  `Init Image` + `Strength`); shared `Result`.
+- provider/llamaswap implements all of it: `/v1/audio/speech`,
+  `/v1/audio/transcriptions` (multipart), `ListVoices`
+  (`/v1/audio/voices?model=`, shape-tolerant), `/sdapi/v1/img2img`
+  (txt2img wire + init_images/denoising_strength, shared decode), and a
+  cheap `Health(ctx)` probe (GET /health) for often-offline hosts.
+- Hermetic httptest coverage for every new wire shape + validation errors.
+- Consumer: mort's llamaswap media tool cluster (status/image/TTS/STT tools).

provider/llamaswap/audio.go

@@ -0,0 +1,278 @@
+package llamaswap
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"mime"
+	"mime/multipart"
+	"net/http"
+	"net/url"
+	"strings"
+
+	majaudio "gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// SpeechModel implements audio.SpeechProvider, binding a text-to-speech model
+// served by llama-swap (routed to a kokoro/chatterbox-style OpenAI-compatible
+// upstream). The id is passed through verbatim and selects which upstream
+// llama-swap loads.
+func (p *Provider) SpeechModel(id string, opts ...majaudio.SpeechModelOption) (majaudio.SpeechModel, error) {
+	if p.baseURL == "" {
+		return nil, fmt.Errorf("llama-swap provider %q: no base URL configured (set one via WithBaseURL or an LLM_* env DSN)", p.name)
+	}
+	_ = majaudio.ApplySpeechModelOptions(opts)
+	return &speechModel{p: p, id: id}, nil
+}
+
+type speechModel struct {
+	p  *Provider
+	id string
+}
+
+// speechRequest is the OpenAI /v1/audio/speech request shape. llama-swap
+// routes by the `model` field in the body.
+type speechRequest struct {
+	Model          string  `json:"model"`
+	Input          string  `json:"input"`
+	Voice          string  `json:"voice,omitempty"`
+	ResponseFormat string  `json:"response_format,omitempty"`
+	Speed          float64 `json:"speed,omitempty"`
+}
+
+// Speak implements audio.SpeechModel via POST {base}/v1/audio/speech.
+func (m *speechModel) Speak(ctx context.Context, req majaudio.SpeechRequest, opts ...majaudio.SpeechOption) (*majaudio.SpeechResult, error) {
+	req = req.Apply(opts...)
+	if strings.TrimSpace(req.Input) == "" {
+		return nil, fmt.Errorf("%w: speech synthesis requires input text", llm.ErrUnsupported)
+	}
+	wire := speechRequest{
+		Model:          m.id,
+		Input:          req.Input,
+		Voice:          req.Voice,
+		ResponseFormat: req.Format,
+		Speed:          req.Speed,
+	}
+	body, err := json.Marshal(wire)
+	if err != nil {
+		return nil, fmt.Errorf("llama-swap: encode speech request: %w", err)
+	}
+	audioBytes, contentType, err := m.p.doRaw(ctx, http.MethodPost, "/v1/audio/speech", m.id, "application/json", bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+	if len(audioBytes) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "speech response contained no audio"}
+	}
+	return &majaudio.SpeechResult{Audio: audioBytes, MIME: speechMIME(contentType, req.Format)}, nil
+}
+
+// speechMIME resolves the result MIME type: the response Content-Type when it
+// is a concrete audio type, else a mapping from the requested format, else
+// audio/mpeg (the OpenAI endpoint's default container is mp3).
+func speechMIME(contentType, format string) string {
+	if mt, _, err := mime.ParseMediaType(contentType); err == nil {
+		if strings.HasPrefix(mt, "audio/") || strings.HasPrefix(mt, "video/") {
+			return mt
+		}
+	}
+	switch strings.ToLower(strings.TrimSpace(format)) {
+	case "", "mp3":
+		return "audio/mpeg"
+	case "wav":
+		return "audio/wav"
+	case "opus":
+		return "audio/ogg"
+	case "aac":
+		return "audio/aac"
+	case "flac":
+		return "audio/flac"
+	case "pcm":
+		return "audio/pcm"
+	default:
+		return "audio/mpeg"
+	}
+}
+
+// TranscriptionModel implements audio.TranscriptionProvider, binding a
+// speech-to-text model served by llama-swap (routed to a whisper.cpp-style
+// upstream).
+func (p *Provider) TranscriptionModel(id string, opts ...majaudio.TranscriptionModelOption) (majaudio.TranscriptionModel, error) {
+	if p.baseURL == "" {
+		return nil, fmt.Errorf("llama-swap provider %q: no base URL configured (set one via WithBaseURL or an LLM_* env DSN)", p.name)
+	}
+	_ = majaudio.ApplyTranscriptionModelOptions(opts)
+	return &transcriptionModel{p: p, id: id}, nil
+}
+
+type transcriptionModel struct {
+	p  *Provider
+	id string
+}
+
+// Transcribe implements audio.TranscriptionModel via POST
+// {base}/v1/audio/transcriptions (multipart/form-data — llama-swap routes by
+// the `model` form field).
+func (m *transcriptionModel) Transcribe(ctx context.Context, req majaudio.TranscriptionRequest, opts ...majaudio.TranscriptionOption) (*majaudio.TranscriptionResult, error) {
+	req = req.Apply(opts...)
+	if len(req.Audio) == 0 {
+		return nil, fmt.Errorf("%w: transcription requires audio bytes", llm.ErrUnsupported)
+	}
+
+	var buf bytes.Buffer
+	w := multipart.NewWriter(&buf)
+	fw, err := w.CreateFormFile("file", transcriptionFilename(req))
+	if err != nil {
+		return nil, fmt.Errorf("llama-swap: build transcription form: %w", err)
+	}
+	if _, err := fw.Write(req.Audio); err != nil {
+		return nil, fmt.Errorf("llama-swap: build transcription form: %w", err)
+	}
+	fields := map[string]string{
+		"model":           m.id,
+		"language":        req.Language,
+		"prompt":          req.Prompt,
+		"response_format": "json",
+	}
+	for k, v := range fields {
+		if v == "" {
+			continue
+		}
+		if err := w.WriteField(k, v); err != nil {
+			return nil, fmt.Errorf("llama-swap: build transcription form: %w", err)
+		}
+	}
+	if err := w.Close(); err != nil {
+		return nil, fmt.Errorf("llama-swap: build transcription form: %w", err)
+	}
+
+	raw, _, err := m.p.doRaw(ctx, http.MethodPost, "/v1/audio/transcriptions", m.id, w.FormDataContentType(), &buf)
+	if err != nil {
+		return nil, err
+	}
+	var out struct {
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &out); err != nil {
+		return nil, fmt.Errorf("llama-swap: decode transcription response: %w", err)
+	}
+	return &majaudio.TranscriptionResult{Text: out.Text, Raw: json.RawMessage(raw)}, nil
+}
+
+// transcriptionFilename picks the multipart filename hint: the caller's, else
+// one derived from the MIME subtype ("audio.mp3"), else "audio".
+func transcriptionFilename(req majaudio.TranscriptionRequest) string {
+	if req.Filename != "" {
+		return req.Filename
+	}
+	switch strings.ToLower(req.MIME) {
+	case "audio/mpeg", "audio/mp3":
+		return "audio.mp3"
+	case "audio/wav", "audio/x-wav", "audio/wave":
+		return "audio.wav"
+	case "audio/ogg", "audio/opus":
+		return "audio.ogg"
+	case "audio/flac", "audio/x-flac":
+		return "audio.flac"
+	case "audio/mp4", "audio/m4a", "audio/x-m4a":
+		return "audio.m4a"
+	case "audio/webm", "video/webm":
+		return "audio.webm"
+	default:
+		return "audio"
+	}
+}
+
+// ListVoices returns the voices a TTS model offers (GET
+// /v1/audio/voices?model=...). The decode is tolerant: upstreams answer with
+// either a bare string list or a list of {id|name} objects.
+func (p *Provider) ListVoices(ctx context.Context, model string) ([]string, error) {
+	if model == "" {
+		return nil, fmt.Errorf("llama-swap: ListVoices requires a model id")
+	}
+	raw, _, err := p.doRaw(ctx, http.MethodGet, "/v1/audio/voices?model="+url.QueryEscape(model), model, "", nil)
+	if err != nil {
+		return nil, err
+	}
+	return parseVoices(raw)
+}
+
+// parseVoices extracts voice names from the various shapes upstreams use:
+// {"voices":[...]} or {"data":[...]} envelopes (or a bare array), holding
+// either strings or objects keyed by id/name/voice_id.
+func parseVoices(raw []byte) ([]string, error) {
+	var env struct {
+		Voices json.RawMessage `json:"voices"`
+		Data   json.RawMessage `json:"data"`
+	}
+	list := json.RawMessage(raw)
+	if err := json.Unmarshal(raw, &env); err == nil {
+		if len(env.Voices) > 0 {
+			list = env.Voices
+		} else if len(env.Data) > 0 {
+			list = env.Data
+		}
+	}
+
+	var names []string
+	if err := json.Unmarshal(list, &names); err == nil {
+		return names, nil
+	}
+	// A failed decode above may have partially populated names — start fresh.
+	names = nil
+	var objs []struct {
+		ID      string `json:"id"`
+		Name    string `json:"name"`
+		VoiceID string `json:"voice_id"`
+	}
+	if err := json.Unmarshal(list, &objs); err == nil {
+		for _, o := range objs {
+			switch {
+			case o.ID != "":
+				names = append(names, o.ID)
+			case o.Name != "":
+				names = append(names, o.Name)
+			case o.VoiceID != "":
+				names = append(names, o.VoiceID)
+			}
+		}
+		return names, nil
+	}
+	return nil, fmt.Errorf("llama-swap: unrecognized voices payload shape")
+}
+
+// doRaw performs a request to a llama-swap endpoint and returns the raw
+// response body and its Content-Type — the sibling of doJSON for endpoints
+// whose success payload is not JSON (audio bytes) or whose shape varies.
+// contentType sets the request Content-Type when body is non-nil.
+func (p *Provider) doRaw(ctx context.Context, method, path, model, contentType string, body io.Reader) ([]byte, string, error) {
+	if p.baseURL == "" {
+		return nil, "", fmt.Errorf("llama-swap provider %q: no base URL configured (set one via WithBaseURL or an LLM_* env DSN)", p.name)
+	}
+	req, err := http.NewRequestWithContext(ctx, method, p.baseURL+path, body)
+	if err != nil {
+		return nil, "", fmt.Errorf("llama-swap: build request: %w", err)
+	}
+	if body != nil && contentType != "" {
+		req.Header.Set("Content-Type", contentType)
+	}
+	if p.token != "" {
+		req.Header.Set("Authorization", "Bearer "+p.token)
+	}
+	resp, err := p.client.Do(req)
+	if err != nil {
+		return nil, "", fmt.Errorf("llama-swap: do request: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode/100 != 2 {
+		return nil, "", p.apiError(resp, model)
+	}
+	data, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
+	if err != nil {
+		return nil, "", fmt.Errorf("llama-swap: read response: %w", err)
+	}
+	return data, resp.Header.Get("Content-Type"), nil
+}

provider/llamaswap/audio_test.go

@@ -0,0 +1,247 @@
+package llamaswap
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+func TestSpeak(t *testing.T) {
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/audio/speech" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		if r.Header.Get("Authorization") != "Bearer tok" {
+			t.Errorf("auth = %q", r.Header.Get("Authorization"))
+		}
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		w.Header().Set("Content-Type", "audio/mpeg")
+		_, _ = w.Write([]byte("MP3BYTES"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithToken("tok"), WithHTTPClient(srv.Client()))
+	sm, err := p.SpeechModel("kokoro")
+	if err != nil {
+		t.Fatalf("SpeechModel: %v", err)
+	}
+	res, err := sm.Speak(context.Background(),
+		audio.SpeechRequest{Input: "hello world"},
+		audio.WithVoice("af_heart"), audio.WithFormat("mp3"), audio.WithSpeed(1.2),
+	)
+	if err != nil {
+		t.Fatalf("Speak: %v", err)
+	}
+	if string(res.Audio) != "MP3BYTES" || res.MIME != "audio/mpeg" {
+		t.Errorf("result = %q %q", res.Audio, res.MIME)
+	}
+	want := map[string]any{"model": "kokoro", "input": "hello world", "voice": "af_heart", "response_format": "mp3", "speed": 1.2}
+	for k, w := range want {
+		if gotBody[k] != w {
+			t.Errorf("%s = %v, want %v", k, gotBody[k], w)
+		}
+	}
+}
+
+func TestSpeakDefaultsOmitted(t *testing.T) {
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		_, _ = w.Write([]byte("x"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("kokoro")
+	res, err := sm.Speak(context.Background(), audio.SpeechRequest{Input: "hi"})
+	if err != nil {
+		t.Fatalf("Speak: %v", err)
+	}
+	for _, k := range []string{"voice", "response_format", "speed"} {
+		if v, ok := gotBody[k]; ok {
+			t.Errorf("unset request sent %q = %v, want omitted", k, v)
+		}
+	}
+	// No usable Content-Type and no requested format → mp3 default.
+	if res.MIME != "audio/mpeg" {
+		t.Errorf("MIME = %q, want audio/mpeg fallback", res.MIME)
+	}
+}
+
+func TestSpeakEmptyInput(t *testing.T) {
+	p := New(WithBaseURL("http://example.invalid"))
+	sm, _ := p.SpeechModel("kokoro")
+	if _, err := sm.Speak(context.Background(), audio.SpeechRequest{Input: "  "}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestSpeechMIME(t *testing.T) {
+	cases := []struct {
+		contentType, format, want string
+	}{
+		{"audio/ogg; codecs=opus", "mp3", "audio/ogg"}, // concrete header wins
+		{"application/octet-stream", "wav", "audio/wav"},
+		{"", "", "audio/mpeg"},
+		{"", "opus", "audio/ogg"},
+		{"text/plain", "flac", "audio/flac"},
+	}
+	for _, tc := range cases {
+		if got := speechMIME(tc.contentType, tc.format); got != tc.want {
+			t.Errorf("speechMIME(%q, %q) = %q, want %q", tc.contentType, tc.format, got, tc.want)
+		}
+	}
+}
+
+func TestTranscribe(t *testing.T) {
+	var gotFields map[string]string
+	var gotFile []byte
+	var gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/audio/transcriptions" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotFields = map[string]string{}
+		for k, v := range r.MultipartForm.Value {
+			gotFields[k] = v[0]
+		}
+		f, hdr, err := r.FormFile("file")
+		if err != nil {
+			t.Fatalf("form file: %v", err)
+		}
+		defer f.Close()
+		gotFile, _ = io.ReadAll(f)
+		gotFilename = hdr.Filename
+		_, _ = w.Write([]byte(`{"text":"hello there"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	tm, err := p.TranscriptionModel("whisper")
+	if err != nil {
+		t.Fatalf("TranscriptionModel: %v", err)
+	}
+	res, err := tm.Transcribe(context.Background(),
+		audio.TranscriptionRequest{Audio: []byte("AUDIO"), MIME: "audio/mpeg"},
+		audio.WithLanguage("en"), audio.WithPrompt("robot names"),
+	)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if res.Text != "hello there" {
+		t.Errorf("text = %q", res.Text)
+	}
+	if string(gotFile) != "AUDIO" || gotFilename != "audio.mp3" {
+		t.Errorf("file = %q name = %q", gotFile, gotFilename)
+	}
+	want := map[string]string{"model": "whisper", "language": "en", "prompt": "robot names", "response_format": "json"}
+	for k, w := range want {
+		if gotFields[k] != w {
+			t.Errorf("%s = %q, want %q", k, gotFields[k], w)
+		}
+	}
+}
+
+func TestTranscribeEmptyAudio(t *testing.T) {
+	p := New(WithBaseURL("http://example.invalid"))
+	tm, _ := p.TranscriptionModel("whisper")
+	if _, err := tm.Transcribe(context.Background(), audio.TranscriptionRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestListVoices(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/audio/voices" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		if got := r.URL.Query().Get("model"); got != "kokoro" {
+			t.Errorf("model = %q", got)
+		}
+		_, _ = w.Write([]byte(`{"voices":["af_heart","af_bella"]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	voices, err := p.ListVoices(context.Background(), "kokoro")
+	if err != nil {
+		t.Fatalf("ListVoices: %v", err)
+	}
+	if !reflect.DeepEqual(voices, []string{"af_heart", "af_bella"}) {
+		t.Errorf("voices = %v", voices)
+	}
+
+	if _, err := p.ListVoices(context.Background(), ""); err == nil {
+		t.Error("empty model: want error")
+	}
+}
+
+func TestParseVoicesShapes(t *testing.T) {
+	cases := []struct {
+		name string
+		raw  string
+		want []string
+	}{
+		{"envelope strings", `{"voices":["a","b"]}`, []string{"a", "b"}},
+		{"bare array", `["a","b"]`, []string{"a", "b"}},
+		{"objects by id", `{"voices":[{"id":"a"},{"id":"b"}]}`, []string{"a", "b"}},
+		{"objects by name", `{"data":[{"name":"a"},{"voice_id":"b"}]}`, []string{"a", "b"}},
+	}
+	for _, tc := range cases {
+		got, err := parseVoices([]byte(tc.raw))
+		if err != nil {
+			t.Errorf("%s: %v", tc.name, err)
+			continue
+		}
+		if !reflect.DeepEqual(got, tc.want) {
+			t.Errorf("%s: got %v, want %v", tc.name, got, tc.want)
+		}
+	}
+	if _, err := parseVoices([]byte(`"just a string"`)); err == nil {
+		t.Error("unparseable shape: want error")
+	}
+}
+
+func TestAudioAPIError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusServiceUnavailable)
+		_, _ = w.Write([]byte(`{"error":{"message":"model is loading"}}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("kokoro")
+	_, err := sm.Speak(context.Background(), audio.SpeechRequest{Input: "x"})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %T %v, want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusServiceUnavailable || apiErr.Message != "model is loading" || apiErr.Model != "kokoro" {
+		t.Errorf("apiErr = %+v", apiErr)
+	}
+}
+
+func TestAudioNoBaseURL(t *testing.T) {
+	p := New()
+	if _, err := p.SpeechModel("kokoro"); err == nil {
+		t.Error("SpeechModel: want error without base URL")
+	}
+	if _, err := p.TranscriptionModel("whisper"); err == nil {
+		t.Error("TranscriptionModel: want error without base URL")
+	}
+	if _, err := p.ListVoices(context.Background(), "kokoro"); err == nil {
+		t.Error("ListVoices: want error without base URL")
+	}
+}

provider/llamaswap/edit_test.go

@@ -0,0 +1,110 @@
+package llamaswap
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+func editInit(t *testing.T) imagegen.Image {
+	t.Helper()
+	raw, err := base64.StdEncoding.DecodeString(onePixelPNG)
+	if err != nil {
+		t.Fatalf("decode fixture: %v", err)
+	}
+	return imagegen.Image{MIME: "image/png", Data: raw}
+}
+
+func TestImageEdit(t *testing.T) {
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/sdapi/v1/img2img" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		_, _ = w.Write([]byte(`{"images":["` + onePixelPNG + `"]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	im, _ := p.ImageModel("sd")
+	ed, ok := im.(imagegen.Editor)
+	if !ok {
+		t.Fatal("imageModel does not implement imagegen.Editor")
+	}
+
+	res, err := ed.Edit(context.Background(),
+		imagegen.EditRequest{Prompt: "make it night", Init: editInit(t)},
+		imagegen.WithEditStrength(0.6),
+	)
+	if err != nil {
+		t.Fatalf("Edit: %v", err)
+	}
+	if len(res.Images) != 1 || res.Images[0].MIME != "image/png" {
+		t.Fatalf("images = %+v", res.Images)
+	}
+
+	if gotBody["model"] != "sd" || gotBody["prompt"] != "make it night" {
+		t.Errorf("model/prompt = %v/%v", gotBody["model"], gotBody["prompt"])
+	}
+	inits, ok := gotBody["init_images"].([]any)
+	if !ok || len(inits) != 1 || inits[0] != onePixelPNG {
+		t.Errorf("init_images = %v, want the b64 fixture", gotBody["init_images"])
+	}
+	if gotBody["denoising_strength"] != 0.6 {
+		t.Errorf("denoising_strength = %v, want 0.6", gotBody["denoising_strength"])
+	}
+}
+
+func TestImageEditOmitsUnsetOverrides(t *testing.T) {
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		_, _ = w.Write([]byte(`{"images":["` + onePixelPNG + `"]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	im, _ := p.ImageModel("sd")
+	ed := im.(imagegen.Editor)
+	if _, err := ed.Edit(context.Background(), imagegen.EditRequest{Prompt: "x", Init: editInit(t)}); err != nil {
+		t.Fatalf("Edit: %v", err)
+	}
+	for _, k := range []string{"denoising_strength", "steps", "cfg_scale", "negative_prompt", "sample_method", "seed", "width", "height"} {
+		if v, ok := gotBody[k]; ok {
+			t.Errorf("unset request sent %q = %v, want omitted", k, v)
+		}
+	}
+}
+
+func TestImageEditValidation(t *testing.T) {
+	p := New(WithBaseURL("http://example.invalid"))
+	im, _ := p.ImageModel("sd")
+	ed := im.(imagegen.Editor)
+
+	cases := []struct {
+		name string
+		req  imagegen.EditRequest
+	}{
+		{"empty prompt", imagegen.EditRequest{Prompt: " ", Init: imagegen.Image{Data: []byte{1}}}},
+		{"missing init", imagegen.EditRequest{Prompt: "x"}},
+		{"negative N", imagegen.EditRequest{Prompt: "x", Init: imagegen.Image{Data: []byte{1}}, N: -1}},
+	}
+	for _, tc := range cases {
+		if _, err := ed.Edit(context.Background(), tc.req); !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("%s: err = %v, want ErrUnsupported", tc.name, err)
+		}
+	}
+
+	bad := 1.5
+	if _, err := ed.Edit(context.Background(), imagegen.EditRequest{Prompt: "x", Init: imagegen.Image{Data: []byte{1}}, Strength: &bad}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("out-of-range strength: err = %v, want ErrUnsupported", err)
+	}
+}

provider/llamaswap/health.go

@@ -0,0 +1,45 @@
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+)
+
+// Health reports whether the llama-swap instance is reachable (GET
+// {base}/health, which answers "OK" without touching any model). A nil error
+// means the proxy itself is up — it says nothing about how long a subsequent
+// request will take (a cold model swap can still block for minutes). Bound
+// the probe with a short context deadline; the client has no timeout by
+// design.
+func (p *Provider) Health(ctx context.Context) error {
+	if p.baseURL == "" {
+		return fmt.Errorf("llama-swap provider %q: no base URL configured (set one via WithBaseURL or an LLM_* env DSN)", p.name)
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, p.baseURL+"/health", nil)
+	if err != nil {
+		return fmt.Errorf("llama-swap: build request: %w", err)
+	}
+	if p.token != "" {
+		req.Header.Set("Authorization", "Bearer "+p.token)
+	}
+	resp, err := p.client.Do(req)
+	if err != nil {
+		return fmt.Errorf("llama-swap: health probe: %w", err)
+	}
+	defer resp.Body.Close()
+	_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, 1<<10))
+	if resp.StatusCode/100 != 2 {
+		return &apiHealthError{status: resp.StatusCode}
+	}
+	return nil
+}
+
+// apiHealthError distinguishes "reachable but unhealthy" from transport
+// failure without pulling in the llm error taxonomy for a probe.
+type apiHealthError struct{ status int }
+
+func (e *apiHealthError) Error() string {
+	return fmt.Sprintf("llama-swap: health endpoint returned status %d", e.status)
+}

provider/llamaswap/health_test.go

@@ -0,0 +1,59 @@
+package llamaswap
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestHealth(t *testing.T) {
+	var gotPath, gotAuth string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotAuth = r.Header.Get("Authorization")
+		_, _ = w.Write([]byte("OK"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithToken("tok"), WithHTTPClient(srv.Client()))
+	if err := p.Health(context.Background()); err != nil {
+		t.Fatalf("Health: %v", err)
+	}
+	if gotPath != "/health" || gotAuth != "Bearer tok" {
+		t.Errorf("path/auth = %q/%q", gotPath, gotAuth)
+	}
+}
+
+func TestHealthUnhealthyStatus(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusBadGateway)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	err := p.Health(context.Background())
+	if err == nil || !strings.Contains(err.Error(), "502") {
+		t.Errorf("err = %v, want status-502 error", err)
+	}
+}
+
+func TestHealthUnreachable(t *testing.T) {
+	// A closed server: transport error, not an HTTP status.
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {}))
+	url := srv.URL
+	srv.Close()
+
+	p := New(WithBaseURL(url))
+	if err := p.Health(context.Background()); err == nil {
+		t.Error("want error for unreachable host")
+	}
+}
+
+func TestHealthNoBaseURL(t *testing.T) {
+	p := New()
+	if err := p.Health(context.Background()); err == nil {
+		t.Error("want error without base URL")
+	}
+}

provider/llamaswap/image.go

@@ -85,8 +85,13 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 	if err := m.p.doJSON(ctx, http.MethodPost, "/sdapi/v1/txt2img", m.id, &wire, &resp); err != nil {
 		return nil, err
 	}
+	return decodeImages(m.p.name, m.id, &resp)
+}
 
-	out := &imagegen.Result{Raw: &resp}
+// decodeImages converts an SDAPI response's base64 images into an
+// imagegen.Result, erroring when nothing decodable came back.
+func decodeImages(provider, model string, resp *txt2imgResponse) (*imagegen.Result, error) {
+	out := &imagegen.Result{Raw: resp}
 	for i, b64 := range resp.Images {
 		if b64 == "" {
 			continue
@@ -99,14 +104,70 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 	}
 	if len(out.Images) == 0 {
 		return nil, &llm.APIError{
-			Provider: m.p.name,
-			Model:    m.id,
+			Provider: provider,
+			Model:    model,
 			Message:  "image response contained no images",
 		}
 	}
 	return out, nil
 }
 
+// img2imgRequest is the stable-diffusion.cpp sd-server A1111 request shape
+// (POST /sdapi/v1/img2img): txt2img's fields plus the init image(s) and
+// denoising strength. Same endpoint-family choice as txt2img — the OpenAI
+// /v1/images/edits route is multipart and drops `seed` on this sd-server
+// build, while the SDAPI shape reuses doJSON and keeps seed parity.
+type img2imgRequest struct {
+	txt2imgRequest
+	InitImages        []string `json:"init_images"`
+	DenoisingStrength *float64 `json:"denoising_strength,omitempty"`
+}
+
+// Edit implements imagegen.Editor via POST {base}/sdapi/v1/img2img.
+func (m *imageModel) Edit(ctx context.Context, req imagegen.EditRequest, opts ...imagegen.EditOption) (*imagegen.Result, error) {
+	req = req.Apply(opts...)
+	if strings.TrimSpace(req.Prompt) == "" {
+		return nil, fmt.Errorf("%w: image edit requires a prompt", llm.ErrUnsupported)
+	}
+	if len(req.Init.Data) == 0 {
+		return nil, fmt.Errorf("%w: image edit requires an init image", llm.ErrUnsupported)
+	}
+	if req.N < 0 {
+		return nil, fmt.Errorf("%w: image count N must be >= 0, got %d", llm.ErrUnsupported, req.N)
+	}
+	if req.Strength != nil && (*req.Strength < 0 || *req.Strength > 1) {
+		return nil, fmt.Errorf("%w: edit strength must be in [0,1], got %g", llm.ErrUnsupported, *req.Strength)
+	}
+
+	width, height, err := parseSize(req.Size)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", llm.ErrUnsupported, err)
+	}
+
+	wire := img2imgRequest{
+		txt2imgRequest: txt2imgRequest{
+			Model:          m.id,
+			Prompt:         req.Prompt,
+			NegativePrompt: req.NegativePrompt,
+			Seed:           req.Seed,
+			Steps:          req.Steps,
+			CFGScale:       req.CFGScale,
+			Width:          width,
+			Height:         height,
+			SampleMethod:   req.Sampler,
+			BatchCount:     req.N,
+		},
+		InitImages:        []string{base64.StdEncoding.EncodeToString(req.Init.Data)},
+		DenoisingStrength: req.Strength,
+	}
+
+	var resp txt2imgResponse
+	if err := m.p.doJSON(ctx, http.MethodPost, "/sdapi/v1/img2img", m.id, &wire, &resp); err != nil {
+		return nil, err
+	}
+	return decodeImages(m.p.name, m.id, &resp)
+}
+
 // parseSize splits a "WxH" string into width/height pointers. "" yields
 // (nil, nil) so the model's own default resolution applies.
 func parseSize(size string) (*int, *int, error) {