docs/adr/0021-musicgen-interface.md

@@ -0,0 +1,35 @@
+# ADR-0021: musicgen interface (blocking Generate over an async job queue)
+
+Status: Accepted (2026-07-12)
+
+## Context
+
+The llama-swap host gained ACE-Step 1.5 (full songs with vocals, seconds per
+clip warm on the reference GPU). Its API is an async job queue —
+`POST /release_task` → poll `POST /query_result` → `GET` the result file —
+unlike every other majordomo media backend, which is one blocking call.
+
+## Decision
+
+- New `musicgen` leaf package, ADR-0016→0020 conventions: `Request{Prompt,
+  Lyrics, DurationSeconds, Format, Steps, Seed}`, `Result{Audio{Data, MIME},
+  Raw}`, functional options, zero value = backend default, bytes-only.
+- **`Model.Generate` blocks, polling internally** (2s interval, ctx-bounded).
+  The one-call contract is the package's value: callers budget the whole job
+  with a context deadline exactly like imagegen/videogen, and the async
+  mechanics stay a provider detail. An async job surface (mirroring the
+  deliberately-deferred videogen one, ADR-0019) can come later if a caller
+  ever needs progress.
+- provider/llamaswap reaches ACE-Step through the `/upstream/<model>/`
+  passthrough (ADR-0020). Envelope parsing is tolerant (`data` wrapper or
+  bare array; `result` arrives as a double-encoded JSON string) and the
+  result-file URL is routed back through the same upstream. `audio_duration`
+  is the v1 param name — an unknown field degrades to default clip length
+  upstream, never an error; verify at smoke.
+
+## Consequences
+
+- Music jobs occupy the exclusive GPU group like video; callers own the
+  timeout budget (mort keeps the tool timeout under its agent ceiling).
+- The double-encoded `result` string and envelope shapes are pinned by the
+  netherstorm image build; host smoke tests are the drift defence.

docs/adr/0022-embeddings-rerank-interface.md

@@ -0,0 +1,41 @@
+# ADR-0022: embeddings + rerank interface
+
+Status: Accepted (2026-07-12)
+
+## Context
+
+majordomo had no embedding or reranking surface at all. The llama-swap host
+now runs two persistent CPU-only llama-server members (Qwen3-Embedding-0.6B
+via `/v1/embeddings`, bge-reranker-v2-m3 via `/v1/rerank`), and mort wants a
+reranking stage in memory retrieval with embedding-backed retrieval as a
+later step.
+
+## Decision
+
+- New `embeddings` leaf package with TWO half-surfaces, split like audio's
+  Speech/Transcription: `EmbedModel`/`EmbedProvider` and
+  `RerankModel`/`RerankProvider`. They are separate mints because on the
+  reference host they are two DIFFERENT server instances — llama-server with
+  `--embeddings` and `--rerank` together returns all-zero embeddings
+  (llama.cpp #20085) — and because rerankers are cross-encoders, not
+  embedders.
+- `EmbedResult.Vectors [][]float32` in input order (provider must order by
+  the response's `index`, never trust wire order). No `dimensions` param:
+  llama-server doesn't implement it; Matryoshka truncation is caller-side.
+- `InstructedQuery(task, query)` helper encodes the instruction-aware
+  asymmetry (queries wrapped, documents bare) so call sites can't silently
+  degrade retrieval by forgetting the prefix.
+- `RerankResult` sorted by descending score; parser reads ONLY
+  `results[].index` and `results[].relevance_score` because llama-server
+  documents the shape as subject to change. Scores are model-specific —
+  comparable within one response only.
+
+## Consequences
+
+- Callers get vectors/scores with strict validation (count mismatch, index
+  out of range, empty vector are hard errors — a silently missing vector is
+  a retrieval bug factory).
+- llama-server's rerank scoring has open correctness issues for some models
+  (llama.cpp #16407); consumers must validate against a fixture before
+  trusting scores in production (mort gates its memory-rerank convar on
+  exactly that).

embeddings/embeddings.go

@@ -0,0 +1,166 @@
+// Package embeddings is majordomo's canonical text-embedding and reranking
+// surface (ADR-0022, following the ADR-0016→0021 leaf-contract lineage).
+// Two small halves, split like audio's Speech/Transcription so backends can
+// implement either:
+//
+//   - EmbedModel turns texts into dense vectors (/v1/embeddings-style).
+//   - RerankModel scores documents against a query with a cross-encoder
+//     (/v1/rerank-style) — usually a DIFFERENT backend model than the
+//     embedder, hence a separate mint.
+//
+// Instruction-aware embedders (Qwen3-Embedding et al.) want queries wrapped
+// as "Instruct: {task}\nQuery: {query}" while documents go in bare;
+// InstructedQuery encodes that so callers don't hand-roll (and silently
+// degrade retrieval) at each site.
+package embeddings
+
+import "context"
+
+// EmbedRequest is a batch embedding request. Inputs are embedded
+// independently; the result vector order matches the input order.
+type EmbedRequest struct {
+	// Inputs are the texts to embed. Required (at least one).
+	Inputs []string
+}
+
+// EmbedResult is the canonical embedding result.
+type EmbedResult struct {
+	// Vectors holds one embedding per input, in input order. Backends
+	// normalize per their own convention (llama-server: Euclidean-normalized).
+	Vectors [][]float32
+
+	// Raw is the provider-native response object. May be nil.
+	Raw any
+}
+
+// EmbedOption mutates an EmbedRequest before it is sent. Reserved: the
+// request shape is deliberately minimal today.
+type EmbedOption func(*EmbedRequest)
+
+// Apply returns a copy of the request with all options applied.
+func (r EmbedRequest) Apply(opts ...EmbedOption) EmbedRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// InstructedQuery wraps a retrieval QUERY for instruction-aware embedding
+// models. Documents must NOT be wrapped — the asymmetry is the point, and
+// getting it wrong silently costs retrieval quality. An empty task uses the
+// generic web-search instruction the reference model was trained with.
+func InstructedQuery(task, query string) string {
+	if task == "" {
+		task = "Given a web search query, retrieve relevant passages that answer the query"
+	}
+	return "Instruct: " + task + "\nQuery: " + query
+}
+
+// EmbedModel embeds texts as dense vectors.
+type EmbedModel interface {
+	// Embed returns one vector per input, in input order.
+	Embed(ctx context.Context, req EmbedRequest, opts ...EmbedOption) (*EmbedResult, error)
+}
+
+// EmbedModelOption configures an EmbedModel at construction time. Reserved
+// for future per-model settings.
+type EmbedModelOption func(*EmbedModelConfig)
+
+// EmbedModelConfig carries per-model construction settings.
+type EmbedModelConfig struct{}
+
+// ApplyEmbedModelOptions folds options into a config.
+func ApplyEmbedModelOptions(opts []EmbedModelOption) EmbedModelConfig {
+	var cfg EmbedModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// EmbedProvider mints embedding models bound to one backend.
+type EmbedProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// EmbedModel returns an EmbedModel bound to the given id (passed through
+	// to the backend verbatim; no catalog validation).
+	EmbedModel(id string, opts ...EmbedModelOption) (EmbedModel, error)
+}
+
+// RerankRequest scores documents against a query.
+type RerankRequest struct {
+	// Query is the search query. Required.
+	Query string
+
+	// Documents are the candidate texts to score. Required (at least one).
+	Documents []string
+
+	// TopN limits how many results the backend returns; 0 = all.
+	TopN int
+}
+
+// RerankItem is one scored document.
+type RerankItem struct {
+	// Index is the document's position in the request's Documents slice.
+	Index int
+
+	// Score is the backend's relevance score (higher = more relevant).
+	// Scales are model-specific — compare within one response only.
+	Score float64
+}
+
+// RerankResult is the canonical rerank result, sorted by descending Score.
+type RerankResult struct {
+	// Results are the scored documents (top-N when the request bounded it).
+	Results []RerankItem
+
+	// Raw is the provider-native response object. May be nil.
+	Raw any
+}
+
+// RerankOption mutates a RerankRequest before it is sent.
+type RerankOption func(*RerankRequest)
+
+// WithTopN limits how many results the backend returns.
+func WithTopN(n int) RerankOption { return func(r *RerankRequest) { r.TopN = n } }
+
+// Apply returns a copy of the request with all options applied.
+func (r RerankRequest) Apply(opts ...RerankOption) RerankRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// RerankModel scores documents against a query with a cross-encoder.
+type RerankModel interface {
+	// Rerank returns scored documents sorted by descending relevance.
+	Rerank(ctx context.Context, req RerankRequest, opts ...RerankOption) (*RerankResult, error)
+}
+
+// RerankModelOption configures a RerankModel at construction time. Reserved
+// for future per-model settings.
+type RerankModelOption func(*RerankModelConfig)
+
+// RerankModelConfig carries per-model construction settings.
+type RerankModelConfig struct{}
+
+// ApplyRerankModelOptions folds options into a config.
+func ApplyRerankModelOptions(opts []RerankModelOption) RerankModelConfig {
+	var cfg RerankModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// RerankProvider mints rerank models bound to one backend.
+type RerankProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// RerankModel returns a RerankModel bound to the given id (passed
+	// through to the backend verbatim; no catalog validation).
+	RerankModel(id string, opts ...RerankModelOption) (RerankModel, error)
+}

musicgen/musicgen.go

@@ -0,0 +1,120 @@
+// Package musicgen is majordomo's canonical music-generation surface (full
+// songs from a text prompt, optionally with lyrics). Like imagegen/audio/
+// videogen/meshgen, it is a deliberately separate leaf contract from the llm
+// package (ADR-0021, following the ADR-0016→0020 lineage: functional
+// options, zero values = backend default, bytes-only I/O, Raw escape hatch).
+//
+// The first implementation is provider/llamaswap, which targets an
+// ACE-Step-1.5-style async job API reached through the /upstream
+// passthrough; Generate blocks (polling internally) so callers get the
+// imagegen-style one-call contract — bound it with a context deadline.
+package musicgen
+
+import "context"
+
+// Audio is one generated piece: raw encoded bytes plus a MIME type.
+type Audio struct {
+	// Data is the encoded audio container.
+	Data []byte
+
+	// MIME is the audio MIME type, e.g. "audio/mpeg".
+	MIME string
+}
+
+// Request is a music-generation request. Zero values mean "backend default".
+type Request struct {
+	// Prompt describes the music (genre, mood, instrumentation, tempo...).
+	Prompt string
+
+	// Lyrics are optional song lyrics for models that sing; "" =
+	// instrumental or model-written lyrics, per backend behavior.
+	Lyrics string
+
+	// DurationSeconds is the requested clip length; 0 = backend default.
+	DurationSeconds int
+
+	// Format is the audio container ("mp3", "wav", "flac", "opus", "aac");
+	// "" = backend default (mp3 on ACE-Step).
+	Format string
+
+	// Steps is the number of inference steps; nil = backend default.
+	Steps *int
+
+	// Seed fixes the RNG seed for reproducible output; nil = backend
+	// default (random).
+	Seed *int64
+}
+
+// Result is the canonical music-generation result.
+type Result struct {
+	// Audio is the generated piece.
+	Audio Audio
+
+	// Raw is the provider-native response object (e.g. the job result with
+	// bpm/keyscale metadata). May be nil.
+	Raw any
+}
+
+// Option mutates a Request before it is sent. Options passed to Generate are
+// applied to a copy of the request, so a Request value can be reused.
+type Option func(*Request)
+
+// WithLyrics sets song lyrics.
+func WithLyrics(l string) Option { return func(r *Request) { r.Lyrics = l } }
+
+// WithDuration sets the requested clip length in seconds.
+func WithDuration(seconds int) Option {
+	return func(r *Request) { r.DurationSeconds = seconds }
+}
+
+// WithFormat sets the audio container format.
+func WithFormat(f string) Option { return func(r *Request) { r.Format = f } }
+
+// WithSteps overrides the number of inference steps.
+func WithSteps(n int) Option { return func(r *Request) { r.Steps = &n } }
+
+// WithSeed fixes the RNG seed.
+func WithSeed(seed int64) Option { return func(r *Request) { r.Seed = &seed } }
+
+// Apply returns a copy of the request with all options applied. Providers
+// call this once at the top of Generate.
+func (r Request) Apply(opts ...Option) Request {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Model generates music from text.
+type Model interface {
+	// Generate renders the request as one audio clip. It blocks until the
+	// backend finishes (or ctx expires) even when the backend is an async
+	// job queue — polling is the provider's business, not the caller's.
+	Generate(ctx context.Context, req Request, opts ...Option) (*Result, error)
+}
+
+// ModelOption configures a Model at construction time. Reserved for future
+// per-model settings.
+type ModelOption func(*ModelConfig)
+
+// ModelConfig carries per-model construction settings.
+type ModelConfig struct{}
+
+// ApplyModelOptions folds options into a config.
+func ApplyModelOptions(opts []ModelOption) ModelConfig {
+	var cfg ModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// Provider mints music models bound to one backend.
+type Provider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// MusicModel returns a Model bound to the given id (passed through to
+	// the backend verbatim; no catalog validation).
+	MusicModel(id string, opts ...ModelOption) (Model, error)
+}

provider/llamaswap/embed.go

@@ -0,0 +1,150 @@
+// embed.go implements embeddings.EmbedProvider and embeddings.RerankProvider
+// against llama-server instances behind llama-swap (ADR-0022):
+//
+//	POST /v1/embeddings   {model, input: [...]}                (OpenAI shape)
+//	POST /v1/rerank       {model, query, documents, top_n}     (Jina-ish shape)
+//
+// Both paths are in llama-swap's normal model-routed tables — no /upstream
+// needed. The two surfaces are minted separately because they are two
+// DIFFERENT server instances on the host: llama-server with --embeddings
+// and --rerank enabled together returns all-zero embeddings (llama.cpp
+// #20085), so the host runs one of each and the ids differ.
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"sort"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/embeddings"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// EmbedModel implements embeddings.EmbedProvider. The id selects which
+// upstream llama-swap loads (a persistent CPU member on the reference host,
+// so calls are cheap and never evict GPU models).
+func (p *Provider) EmbedModel(id string, opts ...embeddings.EmbedModelOption) (embeddings.EmbedModel, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = embeddings.ApplyEmbedModelOptions(opts)
+	return &embedModel{p: p, id: id}, nil
+}
+
+type embedModel struct {
+	p  *Provider
+	id string
+}
+
+// Embed implements embeddings.EmbedModel via POST {base}/v1/embeddings.
+func (m *embedModel) Embed(ctx context.Context, req embeddings.EmbedRequest, opts ...embeddings.EmbedOption) (*embeddings.EmbedResult, error) {
+	req = req.Apply(opts...)
+	if len(req.Inputs) == 0 {
+		return nil, fmt.Errorf("%w: embedding requires at least one input", llm.ErrUnsupported)
+	}
+	for i, in := range req.Inputs {
+		if strings.TrimSpace(in) == "" {
+			return nil, fmt.Errorf("%w: embedding input %d is empty", llm.ErrUnsupported, i)
+		}
+	}
+	wire := struct {
+		Model string   `json:"model"`
+		Input []string `json:"input"`
+	}{Model: m.id, Input: req.Inputs}
+
+	var resp struct {
+		Data []struct {
+			Index     int       `json:"index"`
+			Embedding []float32 `json:"embedding"`
+		} `json:"data"`
+	}
+	if err := m.p.doJSON(ctx, http.MethodPost, "/v1/embeddings", m.id, &wire, &resp); err != nil {
+		return nil, err
+	}
+	if len(resp.Data) != len(req.Inputs) {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("embeddings response has %d vectors for %d inputs", len(resp.Data), len(req.Inputs))}
+	}
+	// The OpenAI shape carries an index per entry; order by it rather than
+	// trusting response order.
+	vectors := make([][]float32, len(req.Inputs))
+	for _, d := range resp.Data {
+		if d.Index < 0 || d.Index >= len(vectors) || len(d.Embedding) == 0 {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("embeddings response entry index %d invalid or empty", d.Index)}
+		}
+		if vectors[d.Index] != nil {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("embeddings response repeats index %d", d.Index)}
+		}
+		vectors[d.Index] = d.Embedding
+	}
+	for i, v := range vectors {
+		if v == nil {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("embeddings response missing vector for input %d", i)}
+		}
+	}
+	return &embeddings.EmbedResult{Vectors: vectors, Raw: &resp}, nil
+}
+
+// RerankModel implements embeddings.RerankProvider.
+func (p *Provider) RerankModel(id string, opts ...embeddings.RerankModelOption) (embeddings.RerankModel, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = embeddings.ApplyRerankModelOptions(opts)
+	return &rerankModel{p: p, id: id}, nil
+}
+
+type rerankModel struct {
+	p  *Provider
+	id string
+}
+
+// Rerank implements embeddings.RerankModel via POST {base}/v1/rerank. The
+// response parser reads only results[].index and results[].relevance_score —
+// llama-server documents the shape as "might change", so stay minimal.
+func (m *rerankModel) Rerank(ctx context.Context, req embeddings.RerankRequest, opts ...embeddings.RerankOption) (*embeddings.RerankResult, error) {
+	req = req.Apply(opts...)
+	if strings.TrimSpace(req.Query) == "" {
+		return nil, fmt.Errorf("%w: rerank requires a query", llm.ErrUnsupported)
+	}
+	if len(req.Documents) == 0 {
+		return nil, fmt.Errorf("%w: rerank requires at least one document", llm.ErrUnsupported)
+	}
+	if req.TopN < 0 {
+		return nil, fmt.Errorf("%w: rerank top_n must be >= 0, got %d", llm.ErrUnsupported, req.TopN)
+	}
+	wire := struct {
+		Model     string   `json:"model"`
+		Query     string   `json:"query"`
+		Documents []string `json:"documents"`
+		TopN      int      `json:"top_n,omitempty"`
+	}{Model: m.id, Query: req.Query, Documents: req.Documents, TopN: req.TopN}
+
+	var resp struct {
+		Results []struct {
+			Index          int     `json:"index"`
+			RelevanceScore float64 `json:"relevance_score"`
+		} `json:"results"`
+	}
+	if err := m.p.doJSON(ctx, http.MethodPost, "/v1/rerank", m.id, &wire, &resp); err != nil {
+		return nil, err
+	}
+	if len(resp.Results) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "rerank response contained no results"}
+	}
+	out := &embeddings.RerankResult{Raw: &resp}
+	for _, r := range resp.Results {
+		if r.Index < 0 || r.Index >= len(req.Documents) {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("rerank result index %d out of range", r.Index)}
+		}
+		out.Results = append(out.Results, embeddings.RerankItem{Index: r.Index, Score: r.RelevanceScore})
+	}
+	sort.SliceStable(out.Results, func(i, j int) bool { return out.Results[i].Score > out.Results[j].Score })
+	return out, nil
+}

provider/llamaswap/music.go

@@ -0,0 +1,265 @@
+// music.go implements musicgen.Provider against an ACE-Step-1.5-style API
+// server reached through llama-swap's /upstream passthrough (ADR-0021):
+//
+//	POST /upstream/<id>/release_task   {prompt, lyrics, ...} -> {task_id}
+//	POST /upstream/<id>/query_result   {task_id_list: [...]} -> status+result
+//	GET  /upstream/<id>/<result file URL>                    -> audio bytes
+//
+// The backend is an async job queue; Generate wraps it into the blocking
+// one-call contract by polling, so a context deadline is the caller's
+// budget for the whole job (mort's tool timeout sits well under its
+// agent-runtime ceiling for exactly this reason).
+package llamaswap
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/musicgen"
+)
+
+// musicPollInterval is the delay between query_result polls. Long enough to
+// be polite to the queue, short enough that a ~10s xl-turbo song isn't
+// dominated by poll latency. A var so tests can shrink it.
+var musicPollInterval = 2 * time.Second
+
+// musicPollMaxConsecutiveFailures bounds how many consecutive BAD polls
+// (transport error, unparseable payload, task momentarily absent) are
+// tolerated before aborting. A multi-minute GPU job must not die to one
+// blip; a genuinely broken upstream still fails within ~5 intervals.
+const musicPollMaxConsecutiveFailures = 5
+
+// MusicModel implements musicgen.Provider. The id selects which upstream
+// llama-swap loads.
+func (p *Provider) MusicModel(id string, opts ...musicgen.ModelOption) (musicgen.Model, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = musicgen.ApplyModelOptions(opts)
+	return &musicModel{p: p, id: id}, nil
+}
+
+type musicModel struct {
+	p  *Provider
+	id string
+}
+
+// releaseTaskRequest is the ACE-Step POST /release_task shape. audio_duration
+// is the v1 param name; verify against the upstream ACE-Step-1.5 repo's
+// docs/en/API.md at smoke time — an
+// unknown field is ignored upstream, degrading to the default clip length,
+// never an error.
+type releaseTaskRequest struct {
+	Prompt         string `json:"prompt"`
+	Lyrics         string `json:"lyrics,omitempty"`
+	AudioFormat    string `json:"audio_format,omitempty"`
+	TaskType       string `json:"task_type"`
+	AudioDuration  int    `json:"audio_duration,omitempty"`
+	InferenceSteps *int   `json:"inference_steps,omitempty"`
+	Seed           *int64 `json:"seed,omitempty"`
+}
+
+// queryItem is one task's poll state. `result` arrives as a JSON-encoded
+// STRING (the ACE-Step API double-encodes it).
+type queryItem struct {
+	TaskID string `json:"task_id"`
+	Status int    `json:"status"` // 0 queued/running, 1 succeeded, 2 failed
+	Result string `json:"result"`
+}
+
+// musicFormatMIME resolves the clip MIME from the response Content-Type
+// or the requested format, reusing speechMIME's table (one format->MIME
+// map for the whole provider). wav32 is ACE-Step-specific: normalize it
+// to wav before the shared lookup.
+func musicFormatMIME(contentType, format string) string {
+	format = strings.ToLower(strings.TrimSpace(format))
+	if format == "wav32" {
+		format = "wav"
+	}
+	return speechMIME(contentType, format)
+}
+
+// Generate implements musicgen.Model.
+func (m *musicModel) Generate(ctx context.Context, req musicgen.Request, opts ...musicgen.Option) (*musicgen.Result, error) {
+	req = req.Apply(opts...)
+	if strings.TrimSpace(req.Prompt) == "" {
+		return nil, fmt.Errorf("%w: music generation requires a prompt", llm.ErrUnsupported)
+	}
+	if req.DurationSeconds < 0 {
+		return nil, fmt.Errorf("%w: duration must be >= 0, got %d", llm.ErrUnsupported, req.DurationSeconds)
+	}
+	if req.Steps != nil && *req.Steps <= 0 {
+		return nil, fmt.Errorf("%w: inference steps must be > 0, got %d", llm.ErrUnsupported, *req.Steps)
+	}
+
+	taskID, err := m.releaseTask(ctx, req)
+	if err != nil {
+		return nil, err
+	}
+	item, err := m.pollResult(ctx, taskID)
+	if err != nil {
+		return nil, err
+	}
+	return m.fetchResult(ctx, req.Format, item)
+}
+
+// releaseTask submits the job and returns its task id.
+func (m *musicModel) releaseTask(ctx context.Context, req musicgen.Request) (string, error) {
+	path, err := upstreamPath(m.id, "/release_task")
+	if err != nil {
+		return "", err
+	}
+	wire := releaseTaskRequest{
+		Prompt:         req.Prompt,
+		Lyrics:         req.Lyrics,
+		AudioFormat:    req.Format,
+		TaskType:       "text2music",
+		AudioDuration:  req.DurationSeconds,
+		InferenceSteps: req.Steps,
+		Seed:           req.Seed,
+	}
+	// Tolerant envelope: {"data": {"task_id": ...}} per the docs, with a
+	// top-level fallback in case the wrapper changes.
+	var resp struct {
+		Data struct {
+			TaskID string `json:"task_id"`
+		} `json:"data"`
+		TaskID string `json:"task_id"`
+	}
+	if err := m.p.doJSON(ctx, http.MethodPost, path, m.id, &wire, &resp); err != nil {
+		return "", err
+	}
+	taskID := resp.Data.TaskID
+	if taskID == "" {
+		taskID = resp.TaskID
+	}
+	if taskID == "" {
+		return "", &llm.APIError{Provider: m.p.name, Model: m.id, Message: "release_task returned no task_id"}
+	}
+	return taskID, nil
+}
+
+// pollResult polls query_result until the task succeeds, fails, or ctx
+// expires. Transient trouble — a transport blip, a momentarily
+// unparseable payload, the task briefly absent from the response — is
+// tolerated up to musicPollMaxConsecutiveFailures in a row: a
+// multi-minute exclusive-GPU job must not die to one flaky poll. Only an
+// explicit status=2, a run of consecutive failures, or the ctx deadline
+// aborts.
+func (m *musicModel) pollResult(ctx context.Context, taskID string) (*queryItem, error) {
+	path, err := upstreamPath(m.id, "/query_result")
+	if err != nil {
+		return nil, err
+	}
+	body := map[string]any{"task_id_list": []string{taskID}}
+	ticker := time.NewTicker(musicPollInterval)
+	defer ticker.Stop()
+	failures := 0
+	var lastErr error
+	for {
+		item, pollErr := m.pollOnce(ctx, path, body, taskID)
+		switch {
+		case pollErr != nil:
+			// Ctx expiry is never transient — bail with the deadline error.
+			if ctx.Err() != nil {
+				return nil, fmt.Errorf("llama-swap: music generation: %w", ctx.Err())
+			}
+			failures++
+			lastErr = pollErr
+			if failures >= musicPollMaxConsecutiveFailures {
+				return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+					Message: fmt.Sprintf("music poll failed %d times in a row: %v", failures, lastErr)}
+			}
+		case item.Status == 1:
+			return item, nil
+		case item.Status == 2:
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: "music generation failed upstream: " + truncateForError([]byte(item.Result))}
+		default:
+			failures = 0 // healthy queued/running poll
+		}
+		select {
+		case <-ctx.Done():
+			return nil, fmt.Errorf("llama-swap: music generation: %w", ctx.Err())
+		case <-ticker.C:
+		}
+	}
+}
+
+// pollOnce performs one query_result round trip and locates the task.
+func (m *musicModel) pollOnce(ctx context.Context, path string, body any, taskID string) (*queryItem, error) {
+	// Tolerant envelope: items under "data" or a bare array.
+	var raw json.RawMessage
+	if err := m.p.doJSON(ctx, http.MethodPost, path, m.id, body, &raw); err != nil {
+		return nil, err
+	}
+	return findQueryItem(raw, taskID)
+}
+
+// findQueryItem digs the task's entry out of the query_result payload,
+// tolerating {"data": [...]}, {"data": {...}}, and bare-array envelopes.
+func findQueryItem(raw json.RawMessage, taskID string) (*queryItem, error) {
+	var env struct {
+		Data json.RawMessage `json:"data"`
+	}
+	candidates := raw
+	if json.Unmarshal(raw, &env) == nil && len(env.Data) > 0 {
+		candidates = env.Data
+	}
+	var items []queryItem
+	if err := json.Unmarshal(candidates, &items); err != nil {
+		var one queryItem
+		if err := json.Unmarshal(candidates, &one); err != nil {
+			return nil, fmt.Errorf("unrecognized query_result payload shape")
+		}
+		items = []queryItem{one}
+	}
+	for i := range items {
+		// Single-item responses without a task_id echo are assumed to be
+		// ours — we only ever poll one task; a mismatch surfaces as a
+		// transient miss and is retried by the caller.
+		if items[i].TaskID == taskID || (items[i].TaskID == "" && len(items) == 1) {
+			return &items[i], nil
+		}
+	}
+	return nil, fmt.Errorf("query_result did not include task %s", taskID)
+}
+
+// fetchResult downloads the finished clip named by the job's result blob.
+func (m *musicModel) fetchResult(ctx context.Context, format string, item *queryItem) (*musicgen.Result, error) {
+	var result struct {
+		File string `json:"file"`
+	}
+	if err := json.Unmarshal([]byte(item.Result), &result); err != nil || result.File == "" {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: "music result blob missing file URL: " + truncateForError([]byte(item.Result))}
+	}
+	// The file URL is server-relative (e.g. "/v1/audio?path=..."); route it
+	// back through the same upstream. upstreamPath additionally refuses
+	// dot-dot/scheme smuggling in this SERVER-SUPPLIED value.
+	if !strings.HasPrefix(result.File, "/") {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: "music result file URL is not server-relative: " + truncateForError([]byte(result.File))}
+	}
+	path, err := upstreamPath(m.id, result.File)
+	if err != nil {
+		return nil, err
+	}
+	raw, contentType, err := m.p.doRaw(ctx, http.MethodGet, path, m.id, "", nil, maxResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "music response contained no audio"}
+	}
+	mimeType := musicFormatMIME(contentType, format)
+	return &musicgen.Result{
+		Audio: musicgen.Audio{Data: raw, MIME: mimeType},
+		Raw:   json.RawMessage(item.Result),
+	}, nil
+}

provider/llamaswap/music_embed_test.go

@@ -0,0 +1,295 @@
+package llamaswap
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/embeddings"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/musicgen"
+)
+
+// aceStepStub emulates the ACE-Step job API: one queued poll, then success.
+func aceStepStub(t *testing.T, mp3 []byte) (*httptest.Server, *atomic.Int32) {
+	t.Helper()
+	var polls atomic.Int32
+	var gotRelease map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/upstream/musicgen-acestep/release_task":
+			_ = json.NewDecoder(r.Body).Decode(&gotRelease)
+			if gotRelease["task_type"] != "text2music" {
+				t.Errorf("task_type = %v", gotRelease["task_type"])
+			}
+			_, _ = w.Write([]byte(`{"data": {"task_id": "t-1", "status": "queued"}}`))
+		case "/upstream/musicgen-acestep/query_result":
+			n := polls.Add(1)
+			if n == 1 {
+				_, _ = w.Write([]byte(`{"data": [{"task_id": "t-1", "status": 0, "result": ""}]}`))
+				return
+			}
+			_, _ = w.Write([]byte(`{"data": [{"task_id": "t-1", "status": 1,
+				"result": "{\"file\": \"/v1/audio?path=out.mp3\", \"metas\": {\"bpm\": 120}}"}]}`))
+		case "/upstream/musicgen-acestep/v1/audio":
+			if got := r.URL.Query().Get("path"); got != "out.mp3" {
+				t.Errorf("audio path = %q", got)
+			}
+			w.Header().Set("Content-Type", "audio/mpeg")
+			_, _ = w.Write(mp3)
+		default:
+			t.Errorf("unexpected path %q", r.URL.Path)
+			w.WriteHeader(404)
+		}
+	}))
+	return srv, &polls
+}
+
+func TestMusicGenerate(t *testing.T) {
+	old := musicPollInterval
+	musicPollInterval = 5 * time.Millisecond
+	defer func() { musicPollInterval = old }()
+
+	mp3 := []byte("ID3fakeaudio")
+	srv, polls := aceStepStub(t, mp3)
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mm, err := p.MusicModel("musicgen-acestep")
+	if err != nil {
+		t.Fatalf("MusicModel: %v", err)
+	}
+	// Shrink the poll interval indirectly by bounding the whole call: the
+	// stub succeeds on poll #2, so a generous deadline still finishes fast.
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	res, err := mm.Generate(ctx,
+		musicgen.Request{Prompt: "chiptune anthem about mortbux"},
+		musicgen.WithLyrics("mort mort mort"), musicgen.WithDuration(30))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if polls.Load() < 2 {
+		t.Errorf("polls = %d, want >= 2 (queued then done)", polls.Load())
+	}
+	if res.Audio.MIME != "audio/mpeg" || string(res.Audio.Data) != string(mp3) {
+		t.Fatalf("audio = %q/%d bytes", res.Audio.MIME, len(res.Audio.Data))
+	}
+}
+
+func TestMusicGenerateUpstreamFailure(t *testing.T) {
+	old := musicPollInterval
+	musicPollInterval = 5 * time.Millisecond
+	defer func() { musicPollInterval = old }()
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/upstream/musicgen-acestep/release_task":
+			_, _ = w.Write([]byte(`{"data": {"task_id": "t-2"}}`))
+		case "/upstream/musicgen-acestep/query_result":
+			_, _ = w.Write([]byte(`{"data": [{"task_id": "t-2", "status": 2, "result": "{\"error\": \"OOM\"}"}]}`))
+		}
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mm, _ := p.MusicModel("musicgen-acestep")
+	_, err := mm.Generate(context.Background(), musicgen.Request{Prompt: "p"})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for failed job", err)
+	}
+}
+
+func TestMusicGenerateRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	mm, _ := p.MusicModel("musicgen-acestep")
+	if _, err := mm.Generate(context.Background(), musicgen.Request{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no prompt: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestEmbed(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/embeddings" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		// Deliberately out of order: the client must sort by index.
+		_, _ = w.Write([]byte(`{"object":"list","data":[
+			{"object":"embedding","index":1,"embedding":[0.3,0.4]},
+			{"object":"embedding","index":0,"embedding":[0.1,0.2]}
+		]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	em, err := p.EmbedModel("embed-qwen3-0.6b")
+	if err != nil {
+		t.Fatalf("EmbedModel: %v", err)
+	}
+	res, err := em.Embed(context.Background(), embeddings.EmbedRequest{Inputs: []string{"a", "b"}})
+	if err != nil {
+		t.Fatalf("Embed: %v", err)
+	}
+	if gotBody["model"] != "embed-qwen3-0.6b" {
+		t.Errorf("model = %v", gotBody["model"])
+	}
+	if len(res.Vectors) != 2 || res.Vectors[0][0] != 0.1 || res.Vectors[1][0] != 0.3 {
+		t.Fatalf("vectors = %+v (index ordering broken?)", res.Vectors)
+	}
+}
+
+func TestEmbedCountMismatchIsError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"data":[{"index":0,"embedding":[0.1]}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	em, _ := p.EmbedModel("embed-qwen3-0.6b")
+	_, err := em.Embed(context.Background(), embeddings.EmbedRequest{Inputs: []string{"a", "b"}})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for count mismatch", err)
+	}
+}
+
+func TestEmbedRejectsEmptyInputs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	em, _ := p.EmbedModel("embed-qwen3-0.6b")
+	if _, err := em.Embed(context.Background(), embeddings.EmbedRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no inputs: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := em.Embed(context.Background(), embeddings.EmbedRequest{Inputs: []string{"a", " "}}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("blank input: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestRerank(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/rerank" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		// Out of score order: the client must sort descending.
+		_, _ = w.Write([]byte(`{"results":[
+			{"index":0,"relevance_score":0.11},
+			{"index":2,"relevance_score":0.93},
+			{"index":1,"relevance_score":0.42}
+		]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	rm, err := p.RerankModel("rerank-bge-v2-m3")
+	if err != nil {
+		t.Fatalf("RerankModel: %v", err)
+	}
+	res, err := rm.Rerank(context.Background(),
+		embeddings.RerankRequest{Query: "what is a panda?", Documents: []string{"a", "b", "c"}},
+		embeddings.WithTopN(3))
+	if err != nil {
+		t.Fatalf("Rerank: %v", err)
+	}
+	if gotBody["top_n"] != float64(3) || gotBody["query"] != "what is a panda?" {
+		t.Errorf("top_n/query = %v/%v", gotBody["top_n"], gotBody["query"])
+	}
+	if len(res.Results) != 3 || res.Results[0].Index != 2 || res.Results[2].Index != 0 {
+		t.Fatalf("results = %+v (descending sort broken?)", res.Results)
+	}
+}
+
+func TestRerankRejectsOutOfRangeIndex(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"results":[{"index":7,"relevance_score":0.9}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	rm, _ := p.RerankModel("rerank-bge-v2-m3")
+	_, err := rm.Rerank(context.Background(),
+		embeddings.RerankRequest{Query: "q", Documents: []string{"a"}})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for out-of-range index", err)
+	}
+}
+
+func TestInstructedQuery(t *testing.T) {
+	got := embeddings.InstructedQuery("", "how tall is everest")
+	want := "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: how tall is everest"
+	if got != want {
+		t.Errorf("InstructedQuery = %q", got)
+	}
+}
+
+func TestMusicGenerateSurvivesTransientPollFailures(t *testing.T) {
+	old := musicPollInterval
+	musicPollInterval = 5 * time.Millisecond
+	defer func() { musicPollInterval = old }()
+
+	mp3 := []byte("ID3fakeaudio")
+	var polls atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/upstream/musicgen-acestep/release_task":
+			_, _ = w.Write([]byte(`{"data": {"task_id": "t-3"}}`))
+		case "/upstream/musicgen-acestep/query_result":
+			switch polls.Add(1) {
+			case 1:
+				w.WriteHeader(http.StatusBadGateway) // transient transport blip
+			case 2:
+				_, _ = w.Write([]byte(`{"data": []}`)) // task momentarily absent
+			default:
+				_, _ = w.Write([]byte(`{"data": [{"task_id": "t-3", "status": 1,
+					"result": "{\"file\": \"/v1/audio?path=out.mp3\"}"}]}`))
+			}
+		case "/upstream/musicgen-acestep/v1/audio":
+			w.Header().Set("Content-Type", "audio/mpeg")
+			_, _ = w.Write(mp3)
+		}
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mm, _ := p.MusicModel("musicgen-acestep")
+	res, err := mm.Generate(context.Background(), musicgen.Request{Prompt: "p"})
+	if err != nil {
+		t.Fatalf("Generate should survive 2 transient failures: %v", err)
+	}
+	if len(res.Audio.Data) == 0 {
+		t.Fatal("no audio")
+	}
+}
+
+func TestMusicGenerateRejectsHostileFileURL(t *testing.T) {
+	old := musicPollInterval
+	musicPollInterval = 5 * time.Millisecond
+	defer func() { musicPollInterval = old }()
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/upstream/musicgen-acestep/release_task":
+			_, _ = w.Write([]byte(`{"data": {"task_id": "t-4"}}`))
+		case "/upstream/musicgen-acestep/query_result":
+			_, _ = w.Write([]byte(`{"data": [{"task_id": "t-4", "status": 1,
+				"result": "{\"file\": \"/v1/audio?path=../../api/models/unload\"}"}]}`))
+		}
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mm, _ := p.MusicModel("musicgen-acestep")
+	_, err := mm.Generate(context.Background(), musicgen.Request{Prompt: "p"})
+	if err == nil {
+		t.Fatal("dot-dot result file URL accepted")
+	}
+}