Merge pull request 'fix(agent): recover the front-loaded answer over a summary closer' (#25) from fix/finalize-summary-closer into main

All tidiness, no behavior change: the leading-marker class is one shared
constant for citationLabelRe and summaryCloserRe (hand-copying it is how
'+' went missing the first time); the deliberate 'all' duplication across
summaryCopulas/summaryArticle is now stated at both sites; the weak-final
switch case assigns modeBackRef explicitly; test comments state the
constraint they guard instead of which reviewer asked for them.

Co-Authored-By: Claude Fable 5 <[email protected]>

.env.example

@@ -6,6 +6,7 @@ OLLAMA_API_KEY=your-ollama-cloud-key-here
 
 # Built-in provider keys (each optional; only needed for the providers you use).
 #OPENAI_API_KEY=sk-...
+#KIMI_API_KEY=sk-...           # Moonshot AI (Kimi); provider name "kimi"
 #ANTHROPIC_API_KEY=sk-ant-...
 #GOOGLE_API_KEY=...
 

.gitea/workflows/adversarial-review.yml

@@ -38,10 +38,11 @@ jobs:
           && (github.actor == 'steve'
               || github.actor == 'fizi'
               || github.actor == 'dazed'))
-    # Tracks gadfly's v1 release tag — a curated pointer re-moved on each release
-    # (unlike @main, which moves on every push). Central swarm tuning propagates
-    # here automatically; the tradeoff vs a full sha pin is that v1 is mutable.
-    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@5007597cf921dc3f0a83c708878facfe65fd8e8b
+    # Pinned to an immutable gadfly commit (not @v1): our act_runners are long-lived
+    # and cache the reusable-workflow ref, so a moved v1 tag keeps resolving to the
+    # stale cached copy. A unique sha forces a cache miss → fresh fetch. Bump this
+    # sha to adopt central swarm changes.
+    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@c9dab69d143cb614c1840a5b06d6ffc358f4752d
     # Least privilege: forward only the review secrets (not `secrets: inherit`,
     # which would expose every repo secret). GITEA_TOKEN is the automatic token.
     secrets:
@@ -52,3 +53,6 @@ jobs:
     with:
       # Consumer-specific allow-list; everything else is inherited.
       allowed_users: "steve,fizi,dazed"
+      # Gitea >= 1.27 does not propagate dispatch inputs into a called workflow's
+      # github.event — thread the PR number explicitly (empty on non-dispatch events).
+      pr_number: ${{ github.event.inputs.pr_number }}

README.md

@@ -121,6 +121,7 @@ Chains are health-tracked per target:
 | Provider | Spec name | Key env var | Default endpoint |
 |----------|-----------|-------------|------------------|
 | OpenAI (+compatible) | `openai` | `OPENAI_API_KEY` | https://api.openai.com/v1 |
+| Kimi (Moonshot AI) | `kimi` | `KIMI_API_KEY` | https://api.moonshot.ai/v1 |
 | Anthropic (+compatible) | `anthropic` | `ANTHROPIC_API_KEY` | https://api.anthropic.com |
 | Google (Gemini) | `google` | `GOOGLE_API_KEY` / `GEMINI_API_KEY` | Gemini API (official SDK) |
 | Ollama Cloud | `ollama-cloud` | `OLLAMA_API_KEY` | https://ollama.com |
@@ -128,6 +129,11 @@ Chains are health-tracked per target:
 | foreman | `foreman` | — (token via DSN) | requires an LLM_* DSN or `ollama.Foreman(url, token)` |
 | llama-swap | `llama-swap` | — (token via DSN) | requires an LLM_* DSN or `llamaswap.New(...)` |
 
+Kimi is Moonshot AI's OpenAI-compatible Chat Completions endpoint, so it reuses
+the openai client (like llama-swap). The `kimi` built-in defaults to the
+international endpoint; reach the China endpoint (or any other host) with a
+`kimi://` DSN, e.g. `LLM_KCN=kimi://[email protected]/v1`.
+
 OpenAI-compatible / Anthropic-compatible endpoints: construct the provider
 with a name and base URL and register it —
 
@@ -159,7 +165,7 @@ m, _ := reg.Parse("m5/qwen3:30b,m1/qwen3:30b,thinking")
 ```
 
 DSN format: `scheme://[token@]host[/path]`, scheme ∈ `foreman`, `ollama`,
-`ollama-cloud`, `openai`, `anthropic`, `google`/`gemini`, `llama-swap`,
+`ollama-cloud`, `openai`, `kimi`, `anthropic`, `google`/`gemini`, `llama-swap`,
 `llama-swaps`, or any scheme you add with `RegisterScheme`. The token is the
 credential (bearer token / API key); the base URL is always `https://host[/path]`
 — except `llama-swap`, which builds `http://host[:port]` since it's local-first
@@ -400,6 +406,7 @@ to build one.
 | Provider | Resolve/Parse | Chat | Streaming | Tools | Structured | Images | Env DSN |
 |----------------------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
 | OpenAI (+compatible) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Kimi (Moonshot AI) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅³ | ✅ |
 | Anthropic (+compat) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Google (Gemini) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Ollama Cloud | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
@@ -420,6 +427,10 @@ transcription** (`audio`), **video generation** (`videogen`) — separate
 axes, not shown above — plus a `Health`
 probe and management methods on `*llamaswap.Provider`.
 
+³ Kimi reuses the openai client, so image *inputs* are supported at the client
+level; whether a call succeeds depends on the Moonshot model — only the vision
+variants (e.g. `moonshot-v1-8k-vision-preview`) accept images.
+
 Notes: Ollama has no native tool_choice — `"none"` drops the tools;
 `"required"`/named choices are best-effort ignored there. Ollama Cloud
 ignores the `format` field (verified live), so the provider also states

agent/agent.go

@@ -125,10 +125,12 @@ func WithCompactor(fn func(ctx context.Context, msgs []llm.Message) ([]llm.Messa
 }
 
 // WithToolErrorLimits installs loop guards: maxConsecutiveErrors bounds
-// successive steps whose tool results were ALL errors, and
-// maxSameCallRepeats bounds identical (name + arguments) tool calls within
-// one run. Either guard tripping ends the run with ErrToolLoop and the
-// partial result. Zero disables a guard.
+// successive steps whose tool results were ALL errors, and maxSameCallRepeats
+// bounds identical (name + arguments) tool calls that ALSO return an unchanged
+// result within one run — a call whose result keeps advancing (e.g. polling a
+// long-running background job) is progress and never trips this guard. Either
+// guard tripping ends the run with ErrToolLoop and the partial result. Zero
+// disables a guard.
 func WithToolErrorLimits(maxConsecutiveErrors, maxSameCallRepeats int) Option {
 	return func(a *Agent) {
 		a.maxConsecutiveToolErrors = maxConsecutiveErrors
@@ -251,6 +253,15 @@ func (a *Agent) mergedTools() (map[string]llm.Tool, []llm.Tool, error) {
 // Run executes the loop: send the conversation; while the model requests
 // tools, execute them and feed results back; stop on a final answer,
 // MaxSteps, or an unrecoverable model error.
+// repeatState is the per-signature bookkeeping for the progress-aware same-call
+// guard: count is the run-length of consecutive identical calls that returned
+// lastResult (the previous call's encoded result). A changed result resets count
+// to 1. See the guard block in Run.
+type repeatState struct {
+	count      int
+	lastResult string
+}
+
 func (a *Agent) Run(ctx context.Context, input string, opts ...RunOption) (*Result, error) {
 	var rc runConfig
 	for _, opt := range opts {
@@ -274,9 +285,12 @@ func (a *Agent) Run(ctx context.Context, input string, opts ...RunOption) (*Resu
 	reqOpts := append(append([]llm.Option(nil), a.reqOpts...), rc.reqOpts...)
 	system := a.systemPrompt()
 
-	// Loop-guard state (WithToolErrorLimits).
+	// Loop-guard state (WithToolErrorLimits). repeatStates tracks, per identical
+	// (name+arguments) signature, the run-length of consecutive calls that
+	// returned the same result and that last result — see the same-call guard
+	// below.
 	consecutiveErrorSteps := 0
-	callCounts := make(map[string]int)
+	repeatStates := make(map[string]*repeatState)
 
 	maxSteps := func() int {
 		if a.maxStepsFunc != nil {
@@ -330,13 +344,6 @@ func (a *Agent) Run(ctx context.Context, input string, opts ...RunOption) (*Resu
 				result.Messages = msgs
 				return result, err
 			}
-			if a.maxSameCallRepeats > 0 {
-				sig := call.Name + "\x00" + string(call.Arguments)
-				callCounts[sig]++
-				if callCounts[sig] > a.maxSameCallRepeats {
-					repeatTripped = call.Name
-				}
-			}
 			tool, ok := byName[call.Name]
 			if !ok {
 				results = append(results, llm.ToolResult{
@@ -356,6 +363,42 @@ func (a *Agent) Run(ctx context.Context, input string, opts ...RunOption) (*Resu
 		a.notify(rc, step)
 		msgs = append(msgs, llm.ToolResultsMessage(results...))
 
+		// Same-call repeat guard (progress-aware). An identical (name+arguments)
+		// call only counts toward the loop trip when its result is unchanged from
+		// the previous identical call: a call whose result keeps advancing —
+		// canonically polling a long-running background job — is progress and
+		// resets its count, while a genuinely stuck call returning the same output
+		// trips once it exceeds the ceiling. Result equality is exact-string on
+		// the full encoded content, chosen to err toward NOT tripping: a hung job
+		// whose poll still reports a ticking field is left to MaxRuntime / the
+		// job's own ceiling rather than risking a false kill of real progress.
+		// results[i] pairs with resp.ToolCalls[i] — every call appends exactly one
+		// result (unknown tools append an error result before continue) and the
+		// only early exit above is a full return on ctx cancellation.
+		if a.maxSameCallRepeats > 0 {
+			for i, call := range resp.ToolCalls {
+				sig := call.Name + "\x00" + string(call.Arguments)
+				resKey := results[i].Content
+				if results[i].IsError {
+					resKey = "e\x00" + resKey
+				}
+				st := repeatStates[sig]
+				if st == nil {
+					st = &repeatState{}
+					repeatStates[sig] = st
+				}
+				if st.count > 0 && st.lastResult == resKey {
+					st.count++
+				} else {
+					st.count = 1
+				}
+				st.lastResult = resKey
+				if st.count > a.maxSameCallRepeats {
+					repeatTripped = call.Name
+				}
+			}
+		}
+
 		if repeatTripped != "" {
 			result.Messages = msgs
 			return result, fmt.Errorf("%w: %q called identically more than %d times",

agent/finalize.go

@@ -14,8 +14,8 @@ import (
 // their answer to the final, tool-free turn. But some models — notably several
 // open-weight ones — "front-load" their full answer into an earlier turn that
 // ALSO calls a tool (e.g. answer text alongside a citation call), then close
-// with a degenerate terminal turn that is not itself the answer. Two shapes are
-// recovered from the transcript (zero extra model calls):
+// with a degenerate terminal turn that is not itself the answer. Three shapes
+// are recovered from the transcript (zero extra model calls):
 //
 //   - a trivial back-reference ("(Already answered above.)", "see above", …):
 //     the real answer sits earlier, so recover it and DISCARD the worthless
@@ -25,25 +25,44 @@ import (
 //     glm-5.2 "cite" pattern behind mort issue #1418). The citations are real,
 //     useful content — unlike a back-reference — so recover the prior answer and
 //     KEEP the citations, appended below it.
+//   - a bookkeeping closer ("Citations are logged. Short version: …"): the
+//     model acknowledged the citation round and compressed the answer it had
+//     already written into a one-liner (mort run b3cb9ee9 — a 2,089-char answer
+//     shrank to a 153-byte closer at delivery). The compression is strictly
+//     poorer than the front-loaded answer, so recover the prior turn and
+//     DISCARD the closer — but only when the prior turn clearly dwarfs it,
+//     because unlike a back-reference this closer DOES carry answer content
+//     (see modeSummary).
 //
-// A citations addendum is tested first and wins over the back-reference test (a
-// short terminal can be both), so its links are never discarded. When the
-// terminal text stands on its own it is returned unchanged; when it is
-// degenerate but nothing better can be recovered, it is returned as-is (a bare
-// sources list still beats nothing).
+// A citations addendum is tested first and wins over the other two (a short
+// terminal can match more than one shape), so its links are never discarded.
+// The back-reference test wins over the summary-closer test: a terminal
+// matching both ("Citations are logged. As I said above…") carries no answer
+// content of its own, so the looser back-ref recovery bar — not the summary
+// closer's dwarf ratio — is the right one. When the terminal text stands
+// on its own it is returned unchanged; when it is degenerate but nothing
+// better can be recovered, it is returned as-is (a compressed answer still
+// beats nothing).
 //
 // msgs must already include the terminal assistant message as its last element
 // (the loop appends it before calling this); terminal is that message's text.
 func finalOutput(msgs []llm.Message, terminal string) string {
-	citations := isCitationsOnly(terminal)
-	if !citations && !isWeakFinal(terminal) {
+	mode := modeBackRef
+	switch {
+	case isCitationsOnly(terminal):
+		mode = modeCitations
+	case isWeakFinal(terminal):
+		mode = modeBackRef
+	case isSummaryCloser(terminal):
+		mode = modeSummary
+	default:
 		return terminal
 	}
-	rec, ok := lastSubstantiveAssistantText(msgs, terminal, citations)
+	rec, ok := lastSubstantiveAssistantText(msgs, terminal, mode)
 	if !ok {
 		return terminal
 	}
-	if citations {
+	if mode == modeCitations {
 		// Preserve the citations addendum below the recovered answer, unless the
 		// recovered turn already carries it (guards against a duplicate sources
 		// block when the front-loaded turn included its own citations). The
@@ -65,11 +84,64 @@ func stripURLAngles(s string) string {
 	return strings.NewReplacer("<", "", ">", "").Replace(s)
 }
 
+// recoveryMode selects the bar a prior assistant turn must clear to replace
+// the terminal turn (see isSubstantiveAnswer) and what finalOutput does with
+// the terminal once recovery succeeds.
+type recoveryMode int
+
+const (
+	// modeBackRef: the terminal is empty or a pure back-reference — worthless
+	// on its own, so any real prior answer replaces it and it is discarded.
+	modeBackRef recoveryMode = iota
+	// modeCitations: the terminal is a sources-only addendum — not a rival
+	// answer, so the dwarf ratio is skipped and the addendum is kept, appended
+	// below the recovered answer.
+	modeCitations
+	// modeSummary: the terminal acknowledges the citation round and may carry
+	// a short compression of the front-loaded answer. Unlike a back-reference
+	// it DOES contain answer content, so it is only replaced when a prior turn
+	// clearly dwarfs it — the ratio is mandatory at every length, the recovery
+	// scan stops at the most recent user message (a compression can only be of
+	// THIS turn's answer; never resurrect one from an earlier question), and
+	// the closer is discarded (its content is a strict subset of what it
+	// replaced).
+	modeSummary
+)
+
 // backRefRe matches a terminal turn that merely points back to an earlier
 // message instead of stating the answer ("(Already answered above.)",
 // "see above", "as I said", ...).
 var backRefRe = regexp.MustCompile(`(?i)(already answered|see above|as (i )?(said|mentioned|stated|noted)|answered (that )?above|per my (previous|earlier))`)
 
+// summaryCloserRe matches a terminal turn that OPENS with a bookkeeping
+// acknowledgment of the citation round — "Citations are logged.", "Sources
+// cited.", "Logged the citations." — the shape a model produces when it
+// front-loaded its answer into an earlier cite-call turn and closes by
+// acknowledging the tool results, often followed by a "Short version: …"
+// compression of the answer it already wrote. The ack clause must end at a
+// sentence terminator ([.!]) DIRECTLY after the verb: "The citations are
+// recorded in the court transcript…" is a real answer about citations, not
+// bookkeeping, and must never match. A compression marker without the ack
+// ("Short version: no.") is deliberately out of scope — a user who asked for
+// brevity would be answered with exactly that shape, and misclassifying it
+// would hijack a legitimate answer; an unmatched closer merely keeps today's
+// behavior (fail closed). Assembled from named fragments so the alternations
+// stay legible and extendable.
+const (
+	summaryPreface = `((done|all set|ok(ay)?)[\s,.!:—-]+)?` // optional "Done —" style opener
+	summaryNouns   = `(citations?|sources?|references?|claims?)`
+	// "all" appears here AND in summaryArticle on purpose: as a quantifier
+	// between noun and verb ("Citations all logged.") and as a determiner
+	// before the noun ("All claims cited.", "Logged all the citations.").
+	summaryCopulas = `((are|were|have\s+been|all)\s+)*`
+	summaryVerbs   = `(logged|recorded|cited|saved|noted|captured|filed)`
+	summaryArticle = `((all|the)\s+)*` // star, not ?: "Logged all the citations."
+)
+
+var summaryCloserRe = regexp.MustCompile(`(?i)^` + leadingMarkers + summaryPreface +
+	`(` + summaryArticle + summaryNouns + `\s+` + summaryCopulas + summaryVerbs +
+	`|logged\s+` + summaryArticle + summaryNouns + `)[.!]`)
+
 // preambleRe matches intent-announcing prefixes ("Let me search...", "I'll
 // check...") so a preamble is never mistaken for the answer during recovery.
 var preambleRe = regexp.MustCompile(`(?i)^(let me|let'?s|i'?ll|i will|first[, ]|sure[,. ]|okay[,. ]|on it|checking)`)
@@ -83,7 +155,15 @@ var preambleRe = regexp.MustCompile(`(?i)^(let me|let'?s|i'?ll|i will|first[, ]|
 // the colon/dash separator. Anchored at ^ so a normal answer that merely
 // mentions "sources" mid-sentence, or ends with a "Sources:" section AFTER its
 // prose, is never matched.
-var citationLabelRe = regexp.MustCompile(`(?i)^[\s>#*_+-]*(sources?|references?|citations?|works cited|further reading)\b[\s*_]*[:：\-—]`)
+var citationLabelRe = regexp.MustCompile(`(?i)^` + leadingMarkers +
+	`(sources?|references?|citations?|works cited|further reading)\b[\s*_]*[:：\-—]`)
+
+// leadingMarkers tolerates markdown noise before a label: emphasis (*, _),
+// list (-, +, *), block-quote (>), and ATX-heading (#) markers, with their
+// whitespace. Shared by citationLabelRe and summaryCloserRe so the two
+// classifiers cannot drift apart (the first draft of the summary class
+// dropped '+' by hand-copying this set).
+const leadingMarkers = `[\s>#*_+-]*`
 
 // linkRe matches a whole markdown link "[label](url)" or a bare URL. Used both
 // to require that a citations terminal carries at least one link and to strip
@@ -114,6 +194,12 @@ const (
 	// be at most len/N of the whole, so a prose answer that merely opens with
 	// "Source:" and cites a URL mid-sentence is not mistaken for a bare list.
 	citationDominatedDivisor = 3
+	// summaryCloserMaxChars bounds a summary closer: room for the ack sentence
+	// plus a couple of compression sentences (the b3cb9ee9 closer was 153
+	// bytes — Go len(), which is what every threshold here compares). Beyond
+	// this the "short version" is substantial enough that replacing it risks
+	// losing content the front-loaded turn never had.
+	summaryCloserMaxChars = 300
 )
 
 // isWeakFinal reports whether a terminal turn's text fails to stand on its own
@@ -152,14 +238,39 @@ func isCitationsOnly(s string) bool {
 	return len(residue) <= len(t)/citationDominatedDivisor
 }
 
+// isSummaryCloser reports whether a terminal turn is a bookkeeping closer: it
+// opens with a complete "citations are logged"-style ack sentence (see
+// summaryCloserRe) and is short enough that whatever follows the ack can only
+// be a compression of an earlier, fuller answer. Whether that fuller answer
+// actually exists is modeSummary's job — the dwarf ratio in
+// isSubstantiveAnswer keeps a matching closer in place when nothing earlier
+// clearly outweighs it.
+func isSummaryCloser(s string) bool {
+	t := strings.TrimSpace(s)
+	if t == "" || len(t) > summaryCloserMaxChars {
+		return false
+	}
+	return summaryCloserRe.MatchString(t)
+}
+
 // lastSubstantiveAssistantText scans msgs newest→oldest (skipping the terminal
 // turn and empty tool-only turns) for the most recent assistant turn whose text
-// reads like a real answer. citations selects the recovery bar (see
+// reads like a real answer. mode selects the recovery bar (see
 // isSubstantiveAnswer). Returns ("", false) when nothing qualifies.
-func lastSubstantiveAssistantText(msgs []llm.Message, terminal string, citations bool) (string, bool) {
+func lastSubstantiveAssistantText(msgs []llm.Message, terminal string, mode recoveryMode) (string, bool) {
 	tt := strings.TrimSpace(terminal)
 	for i := len(msgs) - 1; i >= 0; i-- {
 		m := msgs[i]
+		if mode == modeSummary && m.Role == llm.RoleUser {
+			// A summary closer compresses THIS turn's front-loaded answer, so
+			// the scan must not cross into an earlier question: once the dwarf
+			// ratio has rejected the current turn's text, walking further back
+			// would resurrect a stale answer to a DIFFERENT question — strictly
+			// worse than keeping the closer. (A mid-run steer message is also a
+			// user-role boundary; recovery then fails closed, which is fine.)
+			// The other modes keep their historical unbounded scan.
+			break
+		}
 		if m.Role != llm.RoleAssistant {
 			continue
 		}
@@ -167,7 +278,7 @@ func lastSubstantiveAssistantText(msgs []llm.Message, terminal string, citations
 		if txt == "" || txt == tt {
 			continue // the terminal turn itself, or an empty tool-only turn
 		}
-		if isSubstantiveAnswer(txt, tt, citations) {
+		if isSubstantiveAnswer(txt, tt, mode) {
 			return txt, true
 		}
 	}
@@ -177,20 +288,29 @@ func lastSubstantiveAssistantText(msgs []llm.Message, terminal string, citations
 // isSubstantiveAnswer reports whether txt (a prior assistant turn) reads like a
 // real answer rather than a preamble, relative to the terminal text.
 //
-// A sufficiently long turn (>= recoverMinChars) is accepted unconditionally: a
-// multi-hundred-char turn is an answer even when it opens conversationally
-// ("Sure, here's…", "Let me explain: …"), so the preamble filter is NOT applied
-// to it — applying it there would drop a legitimate long front-loaded answer.
-// Only in the borderline band does a turn have to clear a floor, not read like a
-// short planning preamble ("Let me look that up…"), and — unless the terminal is
-// a citations addendum (not a rival answer, so its length is irrelevant) — also
-// clearly dwarf the terminal.
-func isSubstantiveAnswer(txt, terminal string, citations bool) bool {
+// modeSummary demands the dwarf ratio FIRST, at every length: a summary closer
+// carries a real (compressed) answer, so replacing it is only justified when
+// the prior turn is clearly the fuller original it was compressed from.
+//
+// A sufficiently long turn (>= recoverMinChars) is otherwise accepted
+// unconditionally: a multi-hundred-char turn is an answer even when it opens
+// conversationally ("Sure, here's…", "Let me explain: …"), so the preamble
+// filter is NOT applied to it — applying it there would drop a legitimate long
+// front-loaded answer. Only in the borderline band does a turn have to clear a
+// floor, not read like a short planning preamble ("Let me look that up…"),
+// and — for modeBackRef only — also clearly dwarf the terminal (a citations
+// addendum is not a rival answer, so its length is irrelevant; a summary
+// closer already proved the ratio above).
+func isSubstantiveAnswer(txt, terminal string, mode recoveryMode) bool {
+	dwarfs := len(txt) >= recoverRatio*len(terminal)
+	if mode == modeSummary && !dwarfs {
+		return false
+	}
 	if len(txt) >= recoverMinChars {
 		return true
 	}
 	if len(txt) < recoverFloorChars || preambleRe.MatchString(txt) {
 		return false
 	}
-	return citations || len(txt) >= recoverRatio*len(terminal)
+	return mode != modeBackRef || dwarfs
 }

agent/finalize_test.go

@@ -72,6 +72,46 @@ func TestIsCitationsOnly(t *testing.T) {
 	}
 }
 
+// b3cb9ee9Closer is the verbatim terminal turn from mort run b3cb9ee9: a
+// 2,089-char answer was front-loaded into the cite-call turn and this 153-byte
+// compression (151 runes — the em dash is 3 bytes, and byte length is what the
+// thresholds compare) was all that got delivered.
+const b3cb9ee9Closer = "Citations are logged. Short version: the bulk of that ~$64M was AIPAC and dark-money super PACs, not the party committees — and it still wasn't enough."
+
+func TestIsSummaryCloser(t *testing.T) {
+	cases := []struct {
+		name string
+		in   string
+		want bool
+	}{
+		{"b3cb9ee9-verbatim", b3cb9ee9Closer, true},
+		{"ack-only", "Citations are logged.", true},
+		{"ack-no-copula", "Citations logged.", true},
+		{"claims-cited", "All claims cited.", true},
+		{"verb-first", "Logged the citations.", true},
+		{"done-prefix", "Done — citations logged.", true},
+		{"ack-then-tldr", "Sources have been recorded! TL;DR: the GPU was the bottleneck.", true},
+		{"references-noted", "References noted. In short: yes, it ships Tuesday.", true},
+		{"plus-list-marker", "+ Citations are logged.", true},
+		{"logged-all-the", "Logged all the citations.", true},
+
+		{"empty", "", false},
+		{"ack-continues-midsentence", "The citations are recorded in the court transcript, which shows the filing dates.", false},
+		{"ack-verb-then-clause", "Citations are logged in Zotero whenever you click the save button.", false},
+		{"compression-without-ack", "Short version: yes.", false}, // deliberately out of scope
+		{"mentions-citations-midsentence", "The paper's citations are what got it retracted.", false},
+		{"crisp-number", "42", false},
+		{"over-cap", "Citations are logged. " + strings.Repeat("The long version has many more details worth keeping. ", 6), false}, // >300: too substantial to replace
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := isSummaryCloser(c.in); got != c.want {
+				t.Errorf("isSummaryCloser(%q) = %v, want %v", c.in, got, c.want)
+			}
+		})
+	}
+}
+
 func asst(text string, tools ...llm.ToolCall) llm.Message {
 	m := llm.Message{Role: llm.RoleAssistant}
 	if text != "" {
@@ -84,6 +124,7 @@ func asst(text string, tools ...llm.ToolCall) llm.Message {
 func TestFinalOutput(t *testing.T) {
 	cite := []llm.ToolCall{{ID: "c1", Name: "cite", Arguments: json.RawMessage(`{}`)}}
 	longAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 6))  // >200
+	hugeAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 12)) // >3x the b3cb9ee9 closer
 	// A sources/citations-only terminal — the glm-5.2 "cite" shape behind mort
 	// issue #1418: the prose answer was front-loaded into the tool-call turn and
 	// the terminal turn carried only the citations.
@@ -103,6 +144,9 @@ func TestFinalOutput(t *testing.T) {
 	// A >=200-byte real answer that merely OPENS with a conversational word
 	// ("Sure,"). The preamble filter must NOT veto it (gadfly regression guard).
 	longConversationalAnswer := "Sure, here's the rundown: it currently sells for about $2,700 used on eBay, typically $2,400 to $2,900 depending on condition and bundle, with the sealed Founders Edition commanding the top of that range while used AIB cards go a bit lower."
+	// Matches BOTH the summary ack and backRefRe, within the 120-byte weak cap,
+	// and long enough (>~92 bytes) that longAnswer would fail the summary bar.
+	bothMatchCloser := "Citations are logged. As I mentioned above, the full detail on the money sources is in my earlier message."
 
 	tests := []struct {
 		name     string
@@ -256,6 +300,109 @@ func TestFinalOutput(t *testing.T) {
 			terminal: sources,
 			want:     longConversationalAnswer + "\n\n" + sources,
 		},
+		{
+			// The b3cb9ee9 shape: full answer front-loaded into the cite turn,
+			// then a summary closer. The closer is discarded — its content is a
+			// strict compression of the recovered answer.
+			name: "summary closer discarded when the front-loaded answer dwarfs it",
+			msgs: []llm.Message{
+				llm.UserText("where did the $64M come from?"),
+				asst(hugeAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(b3cb9ee9Closer),
+			},
+			terminal: b3cb9ee9Closer,
+			want:     hugeAnswer,
+		},
+		{
+			// The dwarf ratio is mandatory for a summary closer at EVERY length:
+			// a prior turn that is longer but not clearly the fuller original
+			// (here ~275 chars vs a 151-char closer, under the 3x bar) must not
+			// displace a closer that carries real answer content.
+			name: "summary closer kept when the prior turn does not dwarf it",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(longAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(b3cb9ee9Closer),
+			},
+			terminal: b3cb9ee9Closer,
+			want:     b3cb9ee9Closer,
+		},
+		{
+			// An ack-only closer ("Citations are logged.") is tiny, so even a
+			// modest front-loaded answer clears the ratio and replaces it.
+			name: "ack-only summary closer recovered over a modest answer",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(longAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst("Citations are logged."),
+			},
+			terminal: "Citations are logged.",
+			want:     longAnswer,
+		},
+		{
+			name: "summary closer with only a preamble prior keeps the closer",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst("Let me gather the numbers.", cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(b3cb9ee9Closer),
+			},
+			terminal: b3cb9ee9Closer,
+			want:     b3cb9ee9Closer,
+		},
+		{
+			// The modeSummary scan must stop at the most recent user message.
+			// Here the current turn's answer sits in the 1x-3x band (rejected
+			// by the ratio) while a dwarfing answer to a DIFFERENT question
+			// sits in history — resurrecting it would be strictly worse than
+			// keeping the closer.
+			name: "summary closer never resurrects a stale answer across the user boundary",
+			msgs: []llm.Message{
+				llm.UserText("earlier, unrelated question?"),
+				asst(hugeAnswer),
+				llm.UserText("q?"),
+				asst(conciseAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(b3cb9ee9Closer),
+			},
+			terminal: b3cb9ee9Closer,
+			want:     b3cb9ee9Closer,
+		},
+		{
+			// The boundary must not break the legitimate multi-turn case: the
+			// dwarfing front-loaded answer in THIS turn's window is recovered
+			// even with history behind it.
+			name: "summary closer recovery still works with history present",
+			msgs: []llm.Message{
+				llm.UserText("earlier, unrelated question?"),
+				asst(longAnswer),
+				llm.UserText("q?"),
+				asst(hugeAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(b3cb9ee9Closer),
+			},
+			terminal: b3cb9ee9Closer,
+			want:     hugeAnswer,
+		},
+		{
+			// A closer matching BOTH the ack shape and a back-reference
+			// carries no answer content, so the back-ref test must win and the
+			// ordinary recovery bar apply — under the summary bar this
+			// ~106-byte terminal would demand a ~318-byte prior and wrongly
+			// keep the closer over longAnswer.
+			name: "back-reference wins over the summary ack when both match",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(longAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(bothMatchCloser),
+			},
+			terminal: bothMatchCloser,
+			want:     longAnswer,
+		},
 	}
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
@@ -326,6 +473,37 @@ func TestRun_HealthyTerminalUnchanged(t *testing.T) {
 	}
 }
 
+// TestRun_RecoversFrontLoadedAnswerOverSummaryCloser reproduces mort run
+// b3cb9ee9 end-to-end: the model front-loads its full answer into the
+// cite-call turn, the cite results come back, and the terminal turn is only a
+// bookkeeping ack plus a one-line compression. The delivered output must be
+// the front-loaded answer, with no extra model call.
+func TestRun_RecoversFrontLoadedAnswerOverSummaryCloser(t *testing.T) {
+	hugeAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 12))
+	fp := fake.New("fp")
+	fp.Enqueue("test-model",
+		fake.ReplyWith(llm.Response{
+			Parts:        []llm.Part{llm.Text(hugeAnswer)},
+			ToolCalls:    []llm.ToolCall{{ID: "c1", Name: "cite", Arguments: json.RawMessage(`{}`)}},
+			FinishReason: llm.FinishToolCalls,
+			Usage:        llm.Usage{InputTokens: 10, OutputTokens: 5},
+		}),
+		fake.Reply(b3cb9ee9Closer),
+	)
+
+	a := New(newModel(t, fp), "sys", WithToolbox(citeToolbox(t)))
+	res, err := a.Run(context.Background(), "where did the $64M come from?")
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if res.Output != hugeAnswer {
+		t.Errorf("Output = %q, want recovered front-loaded answer", res.Output)
+	}
+	if n := len(fp.Calls()); n != 2 {
+		t.Errorf("model calls = %d, want 2 (no extra nudge turn)", n)
+	}
+}
+
 // TestRun_RecoversFrontLoadedAnswerWithCitations reproduces mort issue #1418
 // end-to-end: the model front-loads the prose answer into the tool-call turn
 // and closes with a sources-only terminal turn. The delivered output must be

agent/hooks_test.go

@@ -173,3 +173,51 @@ func TestSameCallRepeatGuard(t *testing.T) {
 		t.Errorf("varied calls must not trip the guard: %v", err)
 	}
 }
+
+// TestSameCallRepeatGuardProgressAware: identical (name+args) calls whose
+// RESULT keeps changing — canonically polling a long-running background job
+// whose progress advances — do not trip the repeat guard even well past the
+// limit; but identical calls returning an unchanged result still trip it.
+func TestSameCallRepeatGuardProgressAware(t *testing.T) {
+	// A poll tool that advances every call, so identical args yield a
+	// different result each time.
+	calls := 0
+	polling := llm.NewToolbox("jobs", llm.Tool{
+		Name: "poll",
+		Handler: func(context.Context, json.RawMessage) (any, error) {
+			calls++
+			return map[string]any{"status": "running", "elapsed": calls}, nil
+		},
+	})
+	n := 0
+	fp := fake.New("fp", fake.WithDefault(func(string, llm.Request) fake.Step {
+		n++
+		if n > 6 { // six identical polls, well past the limit of 3
+			return fake.Reply("done")
+		}
+		return toolCallReply("c", "poll", `{"job":"x"}`)
+	}))
+	a := New(newModel(t, fp), "", WithToolbox(polling), WithToolErrorLimits(0, 3), WithMaxSteps(20))
+	res, err := a.Run(context.Background(), "go")
+	if err != nil {
+		t.Fatalf("advancing-result polls must not trip the guard: %v", err)
+	}
+	if res.Output != "done" {
+		t.Errorf("output = %q, want run to complete after polling", res.Output)
+	}
+
+	// A call that returns an UNCHANGED result each time still trips the guard.
+	frozen := llm.NewToolbox("jobs", llm.Tool{
+		Name: "poll",
+		Handler: func(context.Context, json.RawMessage) (any, error) {
+			return map[string]any{"status": "running"}, nil // never advances
+		},
+	})
+	fp2 := fake.New("fp", fake.WithDefault(func(string, llm.Request) fake.Step {
+		return toolCallReply("c", "poll", `{"job":"x"}`)
+	}))
+	a2 := New(newModel(t, fp2), "", WithToolbox(frozen), WithToolErrorLimits(0, 3), WithMaxSteps(20))
+	if _, err := a2.Run(context.Background(), "go"); !errors.Is(err, ErrToolLoop) {
+		t.Fatalf("frozen identical result must still trip the guard: %v", err)
+	}
+}

audio/audio.go

@@ -30,6 +30,17 @@ type SpeechRequest struct {
 
 	// Speed is the playback-rate multiplier; 0 = backend default (1.0).
 	Speed float64
+
+	// ReferenceAudio is a short voice sample for zero-shot voice cloning
+	// (chatterbox style): when set, the model speaks Input in the sampled
+	// voice instead of a named Voice. nil = normal synthesis. Backends
+	// without cloning support must reject a reference-carrying request
+	// rather than silently ignoring it.
+	ReferenceAudio []byte
+
+	// ReferenceMIME is the reference audio's MIME type (e.g. "audio/wav");
+	// "" = let the backend sniff it.
+	ReferenceMIME string
 }
 
 // SpeechResult is the canonical synthesis result: raw audio bytes plus the
@@ -59,6 +70,11 @@ func WithFormat(f string) SpeechOption { return func(r *SpeechRequest) { r.Forma
 // WithSpeed sets the playback-rate multiplier.
 func WithSpeed(s float64) SpeechOption { return func(r *SpeechRequest) { r.Speed = s } }
 
+// WithReferenceAudio provides a voice sample for zero-shot voice cloning.
+func WithReferenceAudio(data []byte, mime string) SpeechOption {
+	return func(r *SpeechRequest) { r.ReferenceAudio, r.ReferenceMIME = data, mime }
+}
+
 // 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 {
@@ -121,6 +137,13 @@ type TranscriptionRequest struct {
 
 	// Prompt is optional context or vocabulary to bias decoding; "" = none.
 	Prompt string
+
+	// Translate requests an English translation of the speech instead of a
+	// same-language transcript (whisper's translate task). When set and no
+	// Language is given, providers must force source-language auto-detection
+	// — a backend whose default language is "en" would otherwise skip
+	// translation entirely.
+	Translate bool
 }
 
 // TranscriptionResult is the canonical transcription result.
@@ -145,6 +168,11 @@ func WithPrompt(p string) TranscriptionOption {
 	return func(r *TranscriptionRequest) { r.Prompt = p }
 }
 
+// WithTranslate requests an English translation instead of a transcript.
+func WithTranslate() TranscriptionOption {
+	return func(r *TranscriptionRequest) { r.Translate = true }
+}
+
 // Apply returns a copy of the request with all options applied.
 func (r TranscriptionRequest) Apply(opts ...TranscriptionOption) TranscriptionRequest {
 	for _, opt := range opts {

audio/enhance.go

@@ -0,0 +1,66 @@
+package audio
+
+import "context"
+
+// EnhancementRequest asks a speech-enhancement backend (DeepFilterNet style)
+// to denoise a recording. Audio is carried as bytes (never a URL), mirroring
+// TranscriptionRequest (ADR-0024).
+type EnhancementRequest struct {
+	// Audio is the encoded audio to enhance.
+	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 or falls back to "audio".
+	Filename string
+}
+
+// EnhancementOption mutates an EnhancementRequest before it is sent.
+// Reserved for future request settings (the reference backend takes no
+// parameters).
+type EnhancementOption func(*EnhancementRequest)
+
+// Apply returns a copy of the request with all options applied.
+func (r EnhancementRequest) Apply(opts ...EnhancementOption) EnhancementRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// SpeechEnhancer denoises speech recordings. The result reuses SpeechResult
+// (audio bytes + MIME + Raw) — enhancement is audio-in/audio-out exactly
+// like synthesis.
+type SpeechEnhancer interface {
+	// Enhance returns the denoised audio.
+	Enhance(ctx context.Context, req EnhancementRequest, opts ...EnhancementOption) (*SpeechResult, error)
+}
+
+// SpeechEnhancerModelOption configures a SpeechEnhancer at construction time.
+// Reserved for future per-model settings.
+type SpeechEnhancerModelOption func(*SpeechEnhancerModelConfig)
+
+// SpeechEnhancerModelConfig carries per-model construction settings.
+type SpeechEnhancerModelConfig struct{}
+
+// ApplySpeechEnhancerModelOptions folds options into a config.
+func ApplySpeechEnhancerModelOptions(opts []SpeechEnhancerModelOption) SpeechEnhancerModelConfig {
+	var cfg SpeechEnhancerModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// SpeechEnhancementProvider mints SpeechEnhancers bound to one backend.
+type SpeechEnhancementProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// SpeechEnhancerModel returns a SpeechEnhancer bound to the given id
+	// (passed through to the backend verbatim; no catalog validation).
+	SpeechEnhancerModel(id string, opts ...SpeechEnhancerModelOption) (SpeechEnhancer, error)
+}

audio/stems.go

@@ -0,0 +1,113 @@
+package audio
+
+import "context"
+
+// StemSeparationRequest asks a source-separation backend (Demucs style) to
+// split a mix into stems. Audio is carried as bytes (never a URL), mirroring
+// TranscriptionRequest. Zero values mean "backend default" (ADR-0024).
+type StemSeparationRequest struct {
+	// Audio is the encoded mix to separate.
+	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 or falls back to "audio".
+	Filename string
+
+	// Mode selects the split: "two" (vocals + accompaniment) or "four"
+	// (vocals/drums/bass/other); "" = backend default (four).
+	Mode string
+
+	// Model selects the separator's internal network where the backend
+	// offers several (Demucs: "htdemucs", "htdemucs_ft"); "" = backend
+	// default. This is NOT the provider model id — that is fixed when the
+	// StemSeparator is minted (mirrors BackgroundRemovalRequest.Net).
+	Model string
+
+	// Format is the per-stem audio container ("mp3" or "wav");
+	// "" = backend default.
+	Format string
+}
+
+// Stem is one separated source.
+type Stem struct {
+	// Name is the stem's name ("vocals", "drums", "bass", "other",
+	// "no_vocals", ...), taken from the backend's own labelling.
+	Name string
+
+	// Audio is the encoded stem.
+	Audio []byte
+
+	// MIME is the stem's audio MIME type, e.g. "audio/mpeg".
+	MIME string
+}
+
+// StemSeparationResult is the canonical separation result.
+type StemSeparationResult struct {
+	// Stems are the separated sources, in the order the backend returned
+	// them.
+	Stems []Stem
+}
+
+// StemSeparationOption mutates a StemSeparationRequest before it is sent.
+type StemSeparationOption func(*StemSeparationRequest)
+
+// WithStemMode selects the split ("two" or "four").
+func WithStemMode(m string) StemSeparationOption {
+	return func(r *StemSeparationRequest) { r.Mode = m }
+}
+
+// WithStemModel selects the separator's internal network.
+func WithStemModel(m string) StemSeparationOption {
+	return func(r *StemSeparationRequest) { r.Model = m }
+}
+
+// WithStemFormat sets the per-stem audio container ("mp3", "wav").
+func WithStemFormat(f string) StemSeparationOption {
+	return func(r *StemSeparationRequest) { r.Format = f }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r StemSeparationRequest) Apply(opts ...StemSeparationOption) StemSeparationRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// StemSeparator splits a mix into stems.
+type StemSeparator interface {
+	// SeparateStems returns the separated sources. Separation is CPU-bound
+	// and slow (minutes for a full song); bound the call with a context
+	// deadline.
+	SeparateStems(ctx context.Context, req StemSeparationRequest, opts ...StemSeparationOption) (*StemSeparationResult, error)
+}
+
+// StemSeparatorModelOption configures a StemSeparator at construction time.
+// Reserved for future per-model settings.
+type StemSeparatorModelOption func(*StemSeparatorModelConfig)
+
+// StemSeparatorModelConfig carries per-model construction settings.
+type StemSeparatorModelConfig struct{}
+
+// ApplyStemSeparatorModelOptions folds options into a config.
+func ApplyStemSeparatorModelOptions(opts []StemSeparatorModelOption) StemSeparatorModelConfig {
+	var cfg StemSeparatorModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// StemSeparationProvider mints StemSeparators bound to one backend.
+type StemSeparationProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// StemSeparatorModel returns a StemSeparator bound to the given id
+	// (passed through to the backend verbatim; no catalog validation).
+	StemSeparatorModel(id string, opts ...StemSeparatorModelOption) (StemSeparator, error)
+}

builtin.go

@@ -2,6 +2,7 @@ package majordomo
 
 import (
 	"net/http"
+	"strings"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/anthropic"
@@ -14,6 +15,10 @@ import (
 // Built-in provider names.
 const (
 	ProviderOpenAI = "openai"
+	// ProviderKimi is Moonshot AI's Kimi models over their OpenAI-compatible
+	// Chat Completions endpoint. Reuses the openai client (like llama-swap);
+	// keyed by KIMI_API_KEY, default base URL kimiBaseURL.
+	ProviderKimi        = "kimi"
 	ProviderAnthropic   = "anthropic"
 	ProviderGoogle      = "google"
 	ProviderOllama      = "ollama"
@@ -28,6 +33,10 @@ const (
 	ProviderLlamaSwapTLS = "llama-swaps"
 )
 
+// kimiBaseURL is Moonshot AI's international OpenAI-compatible endpoint. The
+// China endpoint (api.moonshot.cn/v1) is reachable via a kimi:// LLM_* DSN.
+const kimiBaseURL = "https://api.moonshot.ai/v1"
+
 // registerBuiltins installs the built-in providers and env-DSN scheme
 // factories into a fresh registry. httpClient, when non-nil, is used by
 // every provider and factory the registry itself constructs.
@@ -74,6 +83,33 @@ func registerBuiltins(r *Registry, httpClient *http.Client) {
 		)...), nil
 	}
 
+	// Kimi (Moonshot AI): OpenAI-compatible Chat Completions, so it reuses the
+	// openai client (like llama-swap). Defaults to Moonshot's international
+	// endpoint and the KIMI_API_KEY credential. WithAPIKey is passed
+	// unconditionally — even empty — so an unset KIMI_API_KEY can never fall
+	// through to the openai client's OPENAI_API_KEY default; WithAPIKeyName
+	// makes the missing-key error name KIMI_API_KEY.
+	r.providers[ProviderKimi] = openai.New(openaiOpts(
+		openai.WithName(ProviderKimi),
+		openai.WithBaseURL(kimiBaseURL),
+		openai.WithAPIKey(r.envLookup("KIMI_API_KEY")),
+		openai.WithAPIKeyName("KIMI_API_KEY"),
+	)...)
+	// kimi:// DSN scheme: an OpenAI-compatible target labeled kimi, base URL
+	// from the DSN host (e.g. kimi://[email protected]/v1 for China). Its
+	// credential is the DSN token, not KIMI_API_KEY, so the missing-key hint
+	// names the LLM_<NAME> env var that defines this provider (matching the
+	// lazy-resolution key form in providerFor) — the fix for a keyless target
+	// here is adding a token to that DSN.
+	r.schemes[ProviderKimi] = func(name string, dsn DSN) (llm.Provider, error) {
+		return openai.New(openaiOpts(
+			openai.WithName(name),
+			openai.WithBaseURL(dsn.BaseURL()),
+			openai.WithAPIKey(dsn.Token),
+			openai.WithAPIKeyName("LLM_"+strings.ToUpper(strings.ReplaceAll(name, "-", "_"))),
+		)...), nil
+	}
+
 	// llama-swap: OpenAI-compatible chat + image generation + management
 	// endpoints over a model-swapping proxy. Chat reuses the openai client
 	// (provider/llamaswap delegates). Two schemes: "llama-swap" builds an

builtin_kimi_test.go

@@ -0,0 +1,171 @@
+package majordomo
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// kimiResponse is a minimal valid Chat Completions body so Generate returns a
+// non-empty response (an empty one would trigger failover, not a clean pass).
+const kimiResponse = `{"id":"c1","object":"chat.completion","choices":[` +
+	`{"index":0,"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}]}`
+
+// captureRT records the last request and returns a canned response without
+// touching the network, so these tests stay hermetic while still exercising
+// the real openai client the kimi built-in reuses (base URL + auth header).
+type captureRT struct {
+	req  *http.Request
+	body string
+}
+
+func (c *captureRT) RoundTrip(r *http.Request) (*http.Response, error) {
+	c.req = r
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(c.body)),
+		Header:     make(http.Header),
+		Request:    r,
+	}, nil
+}
+
+// TestKimiBuiltin: the built-in "kimi" provider resolves in Parse, targets
+// Moonshot's default endpoint, and authenticates with KIMI_API_KEY.
+func TestKimiBuiltin(t *testing.T) {
+	rt := &captureRT{body: kimiResponse}
+	r := newTestRegistry(t,
+		WithEnvLookup(func(k string) string {
+			if k == "KIMI_API_KEY" {
+				return "kimi-secret"
+			}
+			return ""
+		}),
+		WithHTTPClient(&http.Client{Transport: rt}),
+	)
+
+	if p, ok := r.Provider(ProviderKimi); !ok {
+		t.Fatal("built-in kimi provider not registered")
+	} else if p.Name() != ProviderKimi {
+		t.Errorf("name = %q, want %q", p.Name(), ProviderKimi)
+	}
+
+	m, err := r.Parse("kimi/kimi-k2-0711-preview")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	if got := targetsOf(t, m); len(got) != 1 || got[0] != "kimi/kimi-k2-0711-preview" {
+		t.Fatalf("targets = %v", got)
+	}
+
+	if _, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}}); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if rt.req == nil {
+		t.Fatal("no request captured")
+	}
+	if want := "https://api.moonshot.ai/v1/chat/completions"; rt.req.URL.String() != want {
+		t.Errorf("URL = %q, want %q", rt.req.URL.String(), want)
+	}
+	if want := "Bearer kimi-secret"; rt.req.Header.Get("Authorization") != want {
+		t.Errorf("Authorization = %q, want %q", rt.req.Header.Get("Authorization"), want)
+	}
+}
+
+// TestKimiBuiltinMissingKey: with no KIMI_API_KEY the built-in fails fast with a
+// synthetic 401 whose hint names KIMI_API_KEY — never OPENAI_API_KEY (proving
+// the credential does not fall through to the openai client's default), and
+// without hitting the network.
+func TestKimiBuiltinMissingKey(t *testing.T) {
+	rt := &captureRT{body: kimiResponse}
+	r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+
+	m, err := r.Parse("kimi/kimi-k2-0711-preview")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	_, err = m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusUnauthorized || apiErr.Code != "missing_api_key" {
+		t.Errorf("Status/Code = %d/%q, want 401/missing_api_key", apiErr.Status, apiErr.Code)
+	}
+	if !strings.Contains(apiErr.Message, "KIMI_API_KEY") {
+		t.Errorf("message = %q, want it to name KIMI_API_KEY", apiErr.Message)
+	}
+	if strings.Contains(apiErr.Message, "OPENAI_API_KEY") {
+		t.Errorf("message = %q, must not name OPENAI_API_KEY", apiErr.Message)
+	}
+	if rt.req != nil {
+		t.Error("network was hit despite missing key")
+	}
+}
+
+// TestKimiScheme: a kimi:// LLM_* DSN defines a named provider on any Moonshot
+// host (here the China endpoint) that is first-class in Parse and carries the
+// DSN token as its bearer credential.
+func TestKimiScheme(t *testing.T) {
+	rt := &captureRT{body: kimiResponse}
+	r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+	if err := r.LoadEnv(map[string]string{
+		"LLM_KCN": "kimi://[email protected]/v1",
+	}); err != nil {
+		t.Fatalf("LoadEnv: %v", err)
+	}
+
+	m, err := r.Parse("kcn/moonshot-v1-8k")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	if _, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}}); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if rt.req == nil {
+		t.Fatal("no request captured")
+	}
+	if want := "https://api.moonshot.cn/v1/chat/completions"; rt.req.URL.String() != want {
+		t.Errorf("URL = %q, want %q", rt.req.URL.String(), want)
+	}
+	if want := "Bearer tok"; rt.req.Header.Get("Authorization") != want {
+		t.Errorf("Authorization = %q, want %q", rt.req.Header.Get("Authorization"), want)
+	}
+}
+
+// TestKimiSchemeMissingToken: a kimi:// DSN with no token is fixed by adding one
+// to the DSN, not by setting KIMI_API_KEY — so the missing-key hint names the
+// defining LLM_<NAME> env var, never KIMI_API_KEY (which does nothing for a
+// DSN-defined provider).
+func TestKimiSchemeMissingToken(t *testing.T) {
+	rt := &captureRT{body: kimiResponse}
+	r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+	if err := r.LoadEnv(map[string]string{
+		"LLM_KCN": "kimi://api.moonshot.cn/v1", // no token
+	}); err != nil {
+		t.Fatalf("LoadEnv: %v", err)
+	}
+
+	m, err := r.Parse("kcn/moonshot-v1-8k")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	_, err = m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+	}
+	if !strings.Contains(apiErr.Message, "LLM_KCN") {
+		t.Errorf("message = %q, want it to name LLM_KCN", apiErr.Message)
+	}
+	if strings.Contains(apiErr.Message, "KIMI_API_KEY") {
+		t.Errorf("message = %q, must not name KIMI_API_KEY for a DSN provider", apiErr.Message)
+	}
+	if rt.req != nil {
+		t.Error("network was hit despite missing token")
+	}
+}

docs/adr/0023-image-doc-surfaces.md

@@ -0,0 +1,59 @@
+# ADR-0023: Wave-3 image + document surfaces (segmentation, colorize, face restore, OCR)
+
+Status: Accepted (2026-07-16)
+
+## Context
+
+The llama-swap host is gaining four wave-3 capabilities (spec:
+mort docs/specs/2026-07-16-llamaswap-wave3.md): promptable segmentation
+(GroundingDINO + SAM 2.1), photo colorization (DDColor), face restoration
+(GFPGAN), and document OCR (Surya). All ride the `/upstream/<model>/<path>`
+passthrough (ADR-0020); none of their native APIs are OpenAI-shaped.
+
+## Decision
+
+1. **Grow `imagegen` with three optional interfaces**, ADR-0016→0022
+   conventions (functional options, zero value = backend default, bytes-only
+   I/O, provider mints model):
+   - `imagegen.Segmenter` / `SegmentationProvider` — `SegmentationRequest
+     {Image, Prompt, Threshold}` → one grayscale mask image, WHITE = the
+     prompted region. Same polarity as `EditRequest.Mask` (white = repaint),
+     so the mask feeds inpainting directly; cutouts are derived client-side
+     (mask-as-alpha), one host call serving both. The client always sends
+     `output=mask` — the shim's `cutout`/`boxes` modes are not exposed.
+   - `imagegen.Colorizer` / `ColorizeProvider` — `ColorizeRequest{Image}`,
+     no knobs (the reference backend takes none; options reserved).
+   - `imagegen.FaceRestorer` / `FaceRestoreProvider` —
+     `FaceRestoreRequest{Image, Upscale}` (1 or 2, 0 = backend default).
+2. **New `ocr` leaf package** rather than a method on an existing surface:
+   OCR is document-shaped (multi-page, PDFs), not image-generation-shaped.
+   `Request{Document, MIME, Filename, Languages, MaxPages}` →
+   `Result{Text, Pages []Page{Number, Text}, Raw}`. `Result.Text` is the
+   pages joined with a blank line; the per-line bbox/confidence/layout
+   detail stays in `Raw` (json.RawMessage) — exact text is the contract,
+   geometry is the escape hatch.
+3. **provider/llamaswap wire shapes** (pinned by the netherstorm image
+   builds; host smoke tests are the drift defence):
+   - `POST /upstream/<id>/v1/segment` multipart `file`,`prompt`
+     [,`threshold`],`output=mask` → mask PNG.
+   - `POST /upstream/<id>/v1/colorize` multipart `file` → PNG.
+   - `POST /upstream/<id>/v1/restore_faces` multipart `file`[,`upscale`] → PNG.
+   - `POST /upstream/<id>/v1/ocr` multipart `file`[,`langs` (comma-joined),
+     `max_pages`] → JSON `{pages:[{number,text,lines,layout}]}`. The decode
+     is tolerant: a page without aggregate `text` joins its line texts; a
+     missing page `number` defaults to position. Zero pages is an error
+     (a blank page still arrives as a page), mirroring the "no transcript"
+     honesty rule.
+4. **Binary success bodies are validated before wrapping** (ADR-0020 rule):
+   the three image surfaces reuse `singleImageResult` (positive evidence of
+   image-ness required), OCR requires decodable JSON.
+
+## Consequences
+
+- imagegen grows from five to eight optional surfaces; consumers
+  type-assert or use the provider methods directly, as before.
+- `ocr` is the seventh leaf media package (imagegen, audio, videogen,
+  meshgen, musicgen, embeddings, ocr); the conventions have held across all
+  of them.
+- PDF handling lives host-side (the shim rasterizes via pypdfium2);
+  majordomo ships bytes and never needs a PDF dependency.

docs/adr/0024-audio-wave3-surfaces.md

@@ -0,0 +1,65 @@
+# ADR-0024: Wave-3 audio surfaces (stems, SFX, speech enhance, voice clone, translate)
+
+Status: Accepted (2026-07-16)
+
+## Context
+
+The llama-swap host is gaining wave-3 audio capabilities (spec: mort
+docs/specs/2026-07-16-llamaswap-wave3.md): Demucs stem separation and
+DeepFilterNet speech enhancement (both on the CPU `audioutils` shim), Stable
+Audio Open sound effects (`sfxgen`), plus two upgrades to existing models —
+chatterbox's stateless voice-clone route and whisper.cpp's per-request
+translate flag (both verified against the live images 2026-07-16).
+
+## Decision
+
+1. **`audio.StemSeparator` / `StemSeparationProvider`** —
+   `StemSeparationRequest{Audio, MIME, Filename, Mode, Model, Format}` →
+   `StemSeparationResult{Stems []Stem{Name, Audio, MIME}}`.
+   - `Mode` is the caller-facing split: `"two"` (vocals + accompaniment,
+     sent as Demucs' `two_stems=vocals`) or `"four"`; `""` = backend
+     default. `Model` selects the Demucs variant (`htdemucs`/`htdemucs_ft`),
+     mirroring `BackgroundRemovalRequest.Net`.
+   - **The wire format is a ZIP** (`POST /upstream/<id>/v1/stems`): four WAV
+     stems would blow any JSON-of-base64 budget. Entry name → stem name,
+     extension → MIME; entries may sit under a per-model directory. Unpacking
+     is bounded per entry (zip-bomb guard) and a non-zip 2xx body fails loud.
+2. **`SFXModel` reuses `musicgen`** — a sound effect is a short audio clip
+   from a text prompt; only the provider method differs. The sfxgen route
+   (`POST /upstream/<id>/v1/sfx`, JSON `{prompt, seconds, steps?, cfg_scale?,
+   seed?}`) is SYNCHRONOUS (WAV body), unlike ACE-Step's job queue.
+   `musicgen.Request` gains `CFGScale *float64` for it; lyrics and non-wav
+   formats are rejected (the model can't honor them). The ~11s model ceiling
+   is the backend's to enforce.
+3. **`audio.SpeechEnhancer` / `SpeechEnhancementProvider`** —
+   `EnhancementRequest{Audio, MIME, Filename}` →
+   `POST /upstream/<id>/v1/enhance` → WAV. The result reuses `SpeechResult`;
+   audio-in/audio-out needs no new result type.
+4. **Voice cloning is a `SpeechRequest` field, not a new surface** —
+   `ReferenceAudio []byte` + `ReferenceMIME`. When set, the llamaswap
+   speech model switches from JSON `/v1/audio/speech` to multipart
+   `POST /upstream/<id>/v1/audio/speech/upload` (fields `input` +
+   `voice_file`; verified live: stateless per-request cloning, no voice
+   library). Voice/format/speed still ride when set; the clone route's MIME
+   fallback is wav (not the JSON route's mp3). Backends without cloning must
+   reject a reference-carrying request rather than silently ignore it.
+5. **Translation is a `TranscriptionRequest` bool** — `Translate` maps to
+   whisper.cpp's `translate=true` form field (server.cpp:534). Because that
+   server's language DEFAULT is `en` (which silently skips translation), the
+   provider forces `language=auto` when translating without an explicit
+   language hint; an explicit hint wins.
+6. **Binary success bodies are validated before wrapping** (ADR-0020 rule):
+   sfx/enhance require positive evidence of audio-ness (declared audio/*
+   Content-Type or sniffed RIFF/WAVE), stems require a parseable zip with
+   at least one stem.
+
+## Consequences
+
+- `audio` grows from three surfaces to five; the SFX surface adds zero new
+  types (musicgen reuse) — one format→MIME table and one multipart builder
+  keep serving every audio endpoint.
+- The clone-route switch means one `SpeechModel` can answer over two wire
+  shapes; tests pin both routes so a regression can't silently drop cloning.
+- `two_stems` is hardwired to vocals: "isolate X vs the rest" for other
+  sources is a backend capability not exposed in v1 (add a field when a
+  caller needs it, not before).

docs/adr/0025-videogen-wave3-surfaces.md

@@ -0,0 +1,67 @@
+# ADR-0025: Wave-3 video surfaces (lipsync, video matte, video upscale, chain jobs)
+
+Status: Accepted (2026-07-16)
+
+## Context
+
+The llama-swap host is gaining wave-3 video capabilities (spec: mort
+docs/specs/2026-07-16-llamaswap-wave3.md): SadTalker talking heads, Robust
+Video Matting and per-frame Real-ESRGAN on the mediautils shim, and a
+videoutils orchestrator that generates LONG videos as a chain of i2v
+segments (generate → extract last frame → continue → concat → optional RIFE
+smoothing), exposed as an async job API following the ACE-Step precedent.
+
+## Decision
+
+1. **Three new optional `videogen` interfaces**, ADR-0016→0024 conventions:
+   - `videogen.LipSyncer` / `LipsyncProvider` — `LipsyncRequest{Image,
+     Audio, AudioMIME, AudioFilename, Still, Enhance, Preprocess}` →
+     `POST /upstream/<id>/v1/talking_head` (multipart `image` + `audio`
+     file parts + optional `still`/`enhance`/`preprocess` fields) → mp4.
+     Sync and minutes-slow (Hunyuan precedent); context deadline is the
+     budget.
+   - `videogen.VideoBackgroundRemover` / `VideoBackgroundRemovalProvider` —
+     `VideoBackgroundRemovalRequest{Video, MIME, Filename, Output}` →
+     `POST /upstream/<id>/v1/video/matte`. `Output` ∈
+     `greenscreen_mp4` (universally playable) | `alpha_webm` (true
+     transparency); "" = backend default.
+   - `videogen.VideoUpscaler` / `VideoUpscaleProvider` —
+     `VideoUpscaleRequest{Video, MIME, Filename, Scale}` (2 or 4) →
+     `POST /upstream/<id>/v1/video/upscale` → mp4.
+2. **The chain client is deliberately ASYNC** (`videogen.Chainer` /
+   `ChainerProvider`), unlike musicgen's blocking Generate (ADR-0021): a
+   chain holds the GPU through multiple model swaps for many minutes, and
+   the caller must be able to poll progress AND fetch completed segments
+   after a mid-chain failure — partial delivery is mandatory (never discard
+   multi-minute GPU output), which a blocking one-call contract cannot
+   express.
+   - `SubmitChain(ctx, ChainRequest{Segments[{Prompt,Seconds}], InitImage,
+     SmoothJoins, Size}) (jobID, error)` — JSON
+     `POST /upstream/<id>/v1/video/chain`, init image as base64
+     `init_image_b64` (JSON submit, not multipart, per the pinned host
+     contract).
+   - `ChainStatus(ctx, jobID) (*ChainJob{Status, Segment, Total,
+     SegmentIDs, Raw})` — `GET /v1/jobs/{id}`; polling doubles as a
+     liveness signal for the shim's idle TTL. Segment entries are tolerated
+     as strings or `{id|segment_id}` objects.
+   - `ChainResult(ctx, jobID)` / `ChainSegmentResult(ctx, jobID, n)` —
+     `GET /v1/jobs/{id}/result` and `/v1/jobs/{id}/segments/{n}`.
+   - Job ids are echoed server input: job paths reject ids carrying path
+     structure (`/?#`, `..`), the upstreamPath smuggling rule.
+3. **Binary success bodies are validated before wrapping** (ADR-0020 rule):
+   all four clip-returning calls go through a shared `singleVideoResult`
+   (positive evidence of video-ness — declared video/* Content-Type or
+   sniffed mp4/webm magic — so a JSON status page or proxy error can never
+   become "the clip").
+
+## Consequences
+
+- videogen grows from two surfaces (Model, Interpolator) to six; the
+  chain client is the package's first async surface — the job-API shape
+  deferred in ADR-0019/0021 now exists where the workload actually
+  demands it.
+- The wire shapes are pinned by the netherstorm videoutils/mediautils/
+  SadTalker image builds; host smoke tests are the drift defence.
+- mort's long-video tool owns the poll loop, timeout budget, and
+  partial-result envelope; majordomo only guarantees the artifacts stay
+  fetchable.

docs/adr/0026-kimi-builtin.md

@@ -0,0 +1,60 @@
+# ADR-0026: Kimi (Moonshot AI) built-in provider
+
+**Status:** Accepted — 2026-07-18
+
+## Context
+
+Moonshot AI's Kimi models (Kimi K2, `moonshot-v1-*`, and the vision variants)
+are served over an OpenAI-compatible Chat Completions API at
+`https://api.moonshot.ai/v1` (`https://api.moonshot.cn/v1` for China),
+authenticated with a bearer key. mort wants Kimi as a first-class failover
+tier, so `kimi/kimi-k2-...` should parse, chain, and alias out of the box with
+a dedicated `KIMI_API_KEY` env var — the same ergonomics as `openai`,
+`anthropic`, and `google`.
+
+Two tensions:
+
+- The wire protocol is byte-for-byte OpenAI Chat Completions, so a hand-rolled
+  client would duplicate `provider/openai` for zero gain (ADR-0007 forbids it),
+  exactly as ADR-0015 found for llama-swap.
+- The README's current stance is that arbitrary OpenAI-compatible endpoints
+  (Groq, Together, …) are *consumer-registered*, not baked in. Blessing Kimi as
+  a built-in is a deliberate, narrow exception justified by the north star:
+  mort names Kimi directly in its tiers, and a built-in with `KIMI_API_KEY`
+  keeps mort's config free of boilerplate `openai.New(WithName/WithBaseURL)`
+  wiring.
+
+## Decision
+
+- **No new package.** The `kimi` built-in and `kimi://` DSN scheme both
+  construct `provider/openai` pointed at the Moonshot base URL — the chat path
+  inherits every openai feature/fix automatically (like llama-swap's chat).
+- The built-in reads its key through the registry's injected `envLookup`
+  (`KIMI_API_KEY` only — no `MOONSHOT_API_KEY` alias, per the project owner) so
+  it stays hermetically testable via `WithEnvLookup`.
+- **`WithAPIKey` is passed unconditionally, even when empty.** `openai.New`
+  defaults its key to `OPENAI_API_KEY`; without an explicit override an unset
+  `KIMI_API_KEY` would silently authenticate Kimi with the OpenAI key. Passing
+  the (possibly empty) lookup result severs that fallthrough.
+- New `openai.WithAPIKeyName("KIMI_API_KEY")` option customizes only the
+  synthetic-401 missing-key hint (default `OPENAI_API_KEY`), so a keyless kimi
+  call tells the operator the *right* variable to set.
+- The default endpoint is the international host (`kimiBaseURL`). The China
+  endpoint (or any other host) is reachable with a `kimi://` DSN, e.g.
+  `LLM_KCN=kimi://[email protected]/v1`. The `kimi://` scheme is an
+  OpenAI-compatible target labeled `kimi` with the same key-name hint; it is
+  intentionally near-identical to `openai://` — its value is a clear name in
+  specs and error reporting.
+
+## Consequences
+
+- `kimi/<model>` is first-class in Parse, chains, aliases, and health/failover
+  with no consumer wiring; model ids pass through verbatim (no catalog).
+- Chat, streaming, tools, and structured output ride the openai client. Image
+  *inputs* work at the client level but only the Moonshot vision models accept
+  them (matrix footnote ³).
+- `WithAPIKeyName` is a small, generally useful addition to `provider/openai`;
+  the default preserves existing behavior for every other openai-compat target.
+- Blessing one third-party endpoint as a built-in sets a precedent; future ones
+  should clear the same bar (a named consumer needs it in-config), not be added
+  reflexively — `RegisterProvider`/`LLM_*` remain the path for the rest.

docs/adr/README.md

@@ -23,3 +23,10 @@ One decision per file, append-only; supersede rather than rewrite.
 | [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 |
 | [0019](0019-videogen-interface.md) | videogen — canonical video-generation surface | Accepted |
+| [0020](0020-upstream-passthrough-media-surfaces.md) | Upstream-passthrough media surfaces (mask, upscale, background removal, interpolation, diarization, meshgen) | Accepted |
+| [0021](0021-musicgen-interface.md) | musicgen — blocking Generate over an async job queue | Accepted |
+| [0022](0022-embeddings-rerank-interface.md) | embeddings + rerank interface | Accepted |
+| [0023](0023-image-doc-surfaces.md) | Wave-3 image + document surfaces (segmentation, colorize, face restore, OCR) | Accepted |
+| [0024](0024-audio-wave3-surfaces.md) | Wave-3 audio surfaces (stems, SFX, speech enhance, voice clone, translate) | Accepted |
+| [0025](0025-videogen-wave3-surfaces.md) | Wave-3 video surfaces (lipsync, video matte, video upscale, chain jobs) | Accepted |
+| [0026](0026-kimi-builtin.md) | Kimi (Moonshot AI) built-in provider — reuse openai client, KIMI_API_KEY | Accepted |

env.go

@@ -26,8 +26,8 @@ var ErrUnknownProvider = errors.New("unknown provider")
 // authenticated with the bearer token "test-token".
 type DSN struct {
 	// Scheme selects the provider implementation: "foreman", "ollama",
-	// "ollama-cloud", "openai", "anthropic", "google"/"gemini", or any
-	// custom scheme registered with RegisterScheme.
+	// "ollama-cloud", "openai", "kimi", "anthropic", "google"/"gemini", or
+	// any custom scheme registered with RegisterScheme.
 	Scheme string
 	// Token is the provider secret (bearer token or API key); empty = none.
 	Token string

imagegen/colorize.go

@@ -0,0 +1,54 @@
+package imagegen
+
+import "context"
+
+// ColorizeRequest asks a colorization backend (DDColor style) to add color to
+// a grayscale/faded photo (ADR-0023).
+type ColorizeRequest struct {
+	// Image is the image to colorize. Required.
+	Image Image
+}
+
+// ColorizeOption mutates a ColorizeRequest before it is sent. Reserved for
+// future request settings (the reference backend takes no parameters).
+type ColorizeOption func(*ColorizeRequest)
+
+// Apply returns a copy of the request with all options applied.
+func (r ColorizeRequest) Apply(opts ...ColorizeOption) ColorizeRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Colorizer adds color to grayscale images.
+type Colorizer interface {
+	// Colorize returns the colorized image as a one-image Result.
+	Colorize(ctx context.Context, req ColorizeRequest, opts ...ColorizeOption) (*Result, error)
+}
+
+// ColorizeModelOption configures a Colorizer at construction time. Reserved
+// for future per-model settings.
+type ColorizeModelOption func(*ColorizeModelConfig)
+
+// ColorizeModelConfig carries per-model construction settings.
+type ColorizeModelConfig struct{}
+
+// ApplyColorizeModelOptions folds options into a config.
+func ApplyColorizeModelOptions(opts []ColorizeModelOption) ColorizeModelConfig {
+	var cfg ColorizeModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// ColorizeProvider mints Colorizers bound to one backend.
+type ColorizeProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// ColorizeModel returns a Colorizer bound to the given id (passed through
+	// to the backend verbatim; no catalog validation).
+	ColorizeModel(id string, opts ...ColorizeModelOption) (Colorizer, error)
+}

imagegen/edit.go

@@ -9,9 +9,32 @@ 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 is the initial image the edit starts from. Required, EXCEPT when
+	// RefImages is set — see there.
 	Init Image
 
+	// RefImages carries reference images for INSTRUCTION-EDIT models
+	// (FLUX.1 Kontext, Qwen-Image-Edit), which are a different kind of edit
+	// from img2img and reach the model by a different path.
+	//
+	// img2img noises Init and denoises it back under the prompt: the prompt
+	// describes the DESIRED IMAGE, and how much of the original survives is a
+	// function of Strength. An instruction-edit model instead takes the
+	// picture as conditioning and the prompt as an INSTRUCTION about it
+	// ("change the sign to read OPEN"), leaving everything it was not asked
+	// to touch bit-for-bit intact — no mask, no strength, no compositing.
+	//
+	// Sending one of these models an Init instead of a RefImage does not
+	// degrade gracefully, it silently does the wrong thing: measured against
+	// FLUX.1-Kontext on 2026-07-30, "change the blue rectangle to green" via
+	// init_images left the rectangle blue and drifted every other region,
+	// while the same prompt via a reference image turned it green and left
+	// the rest of the frame numerically unchanged.
+	//
+	// When RefImages is non-empty, Init/Mask/Strength are IGNORED: they
+	// describe a pipeline this model does not run.
+	RefImages []Image
+
 	// Mask restricts the edit to a region (inpainting): a single-channel or
 	// RGB image the same size as Init where WHITE pixels are repainted and
 	// BLACK pixels are kept. Empty = whole-image edit. Backends without mask
@@ -54,6 +77,13 @@ type EditOption func(*EditRequest)
 // WithEditMask restricts the edit to a region (white = repaint, black = keep).
 func WithEditMask(m Image) EditOption { return func(r *EditRequest) { r.Mask = m } }
 
+// WithEditRefImages supplies reference images for an instruction-edit model
+// (Kontext / Qwen-Image-Edit). See EditRequest.RefImages — this selects a
+// different edit path, not a variation on img2img.
+func WithEditRefImages(imgs ...Image) EditOption {
+	return func(r *EditRequest) { r.RefImages = imgs }
+}
+
 // WithEditStrength sets the denoising strength in [0,1].
 func WithEditStrength(s float64) EditOption { return func(r *EditRequest) { r.Strength = &s } }
 

imagegen/facerestore.go

@@ -0,0 +1,62 @@
+package imagegen
+
+import "context"
+
+// FaceRestoreRequest asks a face-restoration backend (GFPGAN style) to repair
+// degraded faces in a photo. Zero values mean "backend default" (ADR-0023).
+type FaceRestoreRequest struct {
+	// Image is the image to restore. Required.
+	Image Image
+
+	// Upscale is the output enlargement factor (1 or 2 on the reference
+	// backend); 0 = backend default.
+	Upscale int
+}
+
+// FaceRestoreOption mutates a FaceRestoreRequest before it is sent.
+type FaceRestoreOption func(*FaceRestoreRequest)
+
+// WithFaceRestoreUpscale sets the output enlargement factor.
+func WithFaceRestoreUpscale(n int) FaceRestoreOption {
+	return func(r *FaceRestoreRequest) { r.Upscale = n }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r FaceRestoreRequest) Apply(opts ...FaceRestoreOption) FaceRestoreRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// FaceRestorer repairs degraded/blurry faces in photos.
+type FaceRestorer interface {
+	// RestoreFaces returns the restored image as a one-image Result.
+	RestoreFaces(ctx context.Context, req FaceRestoreRequest, opts ...FaceRestoreOption) (*Result, error)
+}
+
+// FaceRestoreModelOption configures a FaceRestorer at construction time.
+// Reserved for future per-model settings.
+type FaceRestoreModelOption func(*FaceRestoreModelConfig)
+
+// FaceRestoreModelConfig carries per-model construction settings.
+type FaceRestoreModelConfig struct{}
+
+// ApplyFaceRestoreModelOptions folds options into a config.
+func ApplyFaceRestoreModelOptions(opts []FaceRestoreModelOption) FaceRestoreModelConfig {
+	var cfg FaceRestoreModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// FaceRestoreProvider mints FaceRestorers bound to one backend.
+type FaceRestoreProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// FaceRestoreModel returns a FaceRestorer bound to the given id (passed
+	// through to the backend verbatim; no catalog validation).
+	FaceRestoreModel(id string, opts ...FaceRestoreModelOption) (FaceRestorer, error)
+}

imagegen/faceswap.go

@@ -0,0 +1,137 @@
+package imagegen
+
+import "context"
+
+// FaceSwapRequest transfers an identity from Source into Target.
+//
+// This is a DIFFERENT OPERATION from Edit, not a better-tuned one. Measured
+// against the instruction-edit models on 2026-07-31, asking a diffusion model
+// to put a specific person's face into a photo does not work by any route —
+// by name, by description, or by supplying the portrait as a reference image.
+// Face swapping is a dedicated detect/align/blend pipeline; a provider that
+// cannot do it should not pretend Edit is a substitute.
+type FaceSwapRequest struct {
+	// Target is the photo to edit — the pose, expression, lighting and
+	// everything outside the face are preserved from it.
+	Target Image
+
+	// Source is a photo of the face to put in. Only the identity travels;
+	// the source's own pose and expression do not.
+	Source Image
+
+	// Index selects WHICH face in Target, in the provider's documented
+	// ordering (llamaswap: left to right by box centre, as reported by
+	// ListFaces). nil = the largest face, which is right for a portrait and
+	// wrong for a group — enumerate first when it matters.
+	Index *int
+
+	// All swaps every detected face and ignores Index.
+	All bool
+}
+
+// FaceSwapOption mutates a FaceSwapRequest before it is sent.
+type FaceSwapOption func(*FaceSwapRequest)
+
+// WithFaceIndex selects which face in the target to replace.
+func WithFaceIndex(i int) FaceSwapOption { return func(r *FaceSwapRequest) { r.Index = &i } }
+
+// WithAllFaces swaps every detected face.
+func WithAllFaces() FaceSwapOption { return func(r *FaceSwapRequest) { r.All = true } }
+
+// Apply returns a copy of the request with all options applied.
+func (r FaceSwapRequest) Apply(opts ...FaceSwapOption) FaceSwapRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// DetectedFace is one face located in an image, in PIXEL coordinates.
+type DetectedFace struct {
+	// Index is the face's position in the provider's stable ordering, and
+	// the value FaceSwapRequest.Index expects.
+	Index int
+	// Box is [x0, y0, x1, y1].
+	Box [4]int
+	// Score is the detector's confidence, 0-1.
+	Score float64
+
+	// Yaw is how far the head is turned from camera, in degrees, or nil when
+	// the provider does not report pose. Exposed on ENUMERATION, not just
+	// after the fact, because it is how a caller picks a face a swap will
+	// actually work on: past roughly ±45° the features carrying identity are
+	// edge-on, and the result reads as a generic person however good the
+	// transfer is. A bounding box cannot show this.
+	Yaw *float64
+}
+
+// Size returns the box dimensions. Derived rather than stored: carrying
+// width/height alongside Box is two sources of truth for one fact, and the
+// pair can disagree after any transform.
+func (f DetectedFace) Size() (w, h int) {
+	return f.Box[2] - f.Box[0], f.Box[3] - f.Box[1]
+}
+
+// SwappedFace is the MEASURED outcome for one face the provider replaced.
+//
+// It exists because "the call returned an image" and "the likeness
+// transferred" are different claims that look identical from outside, and a
+// caller that cannot tell them apart will go looking for another way to
+// check. The one it reaches for — asking a vision model who the result looks
+// like — is wrong in exactly the cases that matter: a VLM shown a jogger in a
+// Georgetown cap holding McDonald's cups answers "Bill Clinton" whoever's
+// face is on him, so it reports failure on a correct swap.
+type SwappedFace struct {
+	// Index is the face's position in the provider's left-to-right ordering.
+	Index int
+
+	// Width, Height are the replaced face's pixel size in the TARGET.
+	// Meaningful only against ImageWidth/ImageHeight: a 138px face is large
+	// in a 400px picture and nearly invisible in a 2000px one, and it is the
+	// ratio, not the absolute size, that decides whether a person notices.
+	Width, Height int
+
+	// ImageWidth, ImageHeight are the target image's dimensions, repeated on
+	// every entry so a single face is self-describing without the caller
+	// holding onto the rest of the response.
+	ImageWidth, ImageHeight int
+
+	// Yaw is how far the head is turned from camera, in degrees, or nil when
+	// the provider does not report pose. The best single predictor of whether
+	// a swap will READ as the source person: past roughly ±45° the features
+	// carrying identity are edge-on and the result looks like a generic
+	// person rather than a specific one.
+	Yaw *float64
+
+	// IdentitySimilarity is cosine similarity between the source face and the
+	// face actually present in the result, 0-1, or nil when the provider
+	// could not measure it. Above ~0.5 the identity transferred; a LOW value
+	// is the only evidence that a swap genuinely failed.
+	IdentitySimilarity *float64
+}
+
+// FractionOfImage is the swapped face's width as a share of the image's, 0-1.
+// The number that predicts whether a person will SEE the change: the swap
+// that prompted all this replaced a 138px face in a 1010px-wide photo — 14%,
+// correct by every measure and invisible at a glance — while the same code on
+// a 168px face in a 385px picture (44%) is unmistakable. Returns 0 when the
+// dimensions are unknown.
+func (f SwappedFace) FractionOfImage() float64 {
+	if f.ImageWidth <= 0 || f.Width <= 0 {
+		return 0
+	}
+	return float64(f.Width) / float64(f.ImageWidth)
+}
+
+// FaceSwapper is the optional face-transfer surface. Separate interface so
+// existing providers keep compiling; callers type-assert.
+type FaceSwapper interface {
+	// ListFaces enumerates the faces in an image, in the SAME ordering
+	// FaceSwapRequest.Index uses. Exposed because a caller asked to change
+	// "the man on the right" needs a way to name one face and to check its
+	// own choice against pixel boxes.
+	ListFaces(ctx context.Context, img Image) ([]DetectedFace, error)
+
+	// FaceSwap transfers Source's identity into Target.
+	FaceSwap(ctx context.Context, req FaceSwapRequest, opts ...FaceSwapOption) (*Result, error)
+}

imagegen/imagegen.go

@@ -68,6 +68,15 @@ type Result struct {
 	// Images are the generated images, in the order the backend returned them.
 	Images []Image
 
+	// SwappedFaces is the measured outcome of a FaceSwap, one entry per face
+	// replaced. Empty for every other operation, and empty from a provider
+	// that does not measure. TYPED rather than tucked into Raw: a caller has
+	// to act on this — it is the only way to distinguish a swap that
+	// transferred the likeness from one that returned an image and nothing
+	// more — and a value reachable solely by type-asserting an `any` is one
+	// nobody discovers in time to use.
+	SwappedFaces []SwappedFace
+
 	// Raw is the provider-native response object, an escape hatch for
 	// provider-specific fields. May be nil; never required for normal use.
 	Raw any

imagegen/segment.go

@@ -0,0 +1,70 @@
+package imagegen
+
+import "context"
+
+// SegmentationRequest asks a promptable-segmentation backend (GroundingDINO +
+// SAM style) for the mask of a text-described region. Zero values mean
+// "backend default" (ADR-0023).
+type SegmentationRequest struct {
+	// Image is the image to segment. Required.
+	Image Image
+
+	// Prompt is the text description of the region to segment (e.g. "the red
+	// car"). Required — promptless segmentation is a different capability.
+	Prompt string
+
+	// Threshold is the detection confidence threshold in (0,1]; 0 = backend
+	// default.
+	Threshold float64
+}
+
+// SegmentationOption mutates a SegmentationRequest before it is sent.
+type SegmentationOption func(*SegmentationRequest)
+
+// WithSegmentationThreshold sets the detection confidence threshold.
+func WithSegmentationThreshold(t float64) SegmentationOption {
+	return func(r *SegmentationRequest) { r.Threshold = t }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r SegmentationRequest) Apply(opts ...SegmentationOption) SegmentationRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Segmenter extracts a prompted region's mask from an image. The result is a
+// single grayscale mask image where WHITE marks the prompted region — the
+// same polarity as EditRequest.Mask (white = repaint), so it feeds inpainting
+// directly; derive a cutout client-side by applying the mask as alpha.
+type Segmenter interface {
+	// Segment returns the region mask as a one-image Result.
+	Segment(ctx context.Context, req SegmentationRequest, opts ...SegmentationOption) (*Result, error)
+}
+
+// SegmentationModelOption configures a Segmenter at construction time.
+// Reserved for future per-model settings (mirrors ModelOption).
+type SegmentationModelOption func(*SegmentationModelConfig)
+
+// SegmentationModelConfig carries per-model construction settings.
+type SegmentationModelConfig struct{}
+
+// ApplySegmentationModelOptions folds options into a config.
+func ApplySegmentationModelOptions(opts []SegmentationModelOption) SegmentationModelConfig {
+	var cfg SegmentationModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// SegmentationProvider mints Segmenters bound to one backend.
+type SegmentationProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// SegmentationModel returns a Segmenter bound to the given id (passed
+	// through to the backend verbatim; no catalog validation).
+	SegmentationModel(id string, opts ...SegmentationModelOption) (Segmenter, error)
+}

meshgen/meshgen.go

@@ -150,3 +150,23 @@ type Provider interface {
 	// backend verbatim; no catalog validation).
 	MeshModel(id string, opts ...ModelOption) (Model, error)
 }
+
+// Converter re-encodes a mesh between containers (GLB/STL/OBJ). A separate
+// optional surface because conversion typically runs on a DIFFERENT backend
+// than generation (the reference host converts on its mediautils shim —
+// Hunyuan3D's live server always emits GLB regardless of the requested
+// format).
+type Converter interface {
+	// Convert returns the mesh re-encoded in the given format.
+	Convert(ctx context.Context, mesh Mesh, format string) (*Result, error)
+}
+
+// ConverterProvider mints Converters bound to one backend.
+type ConverterProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// MeshConverter returns a Converter bound to the given id (passed
+	// through to the backend verbatim; no catalog validation).
+	MeshConverter(id string) (Converter, error)
+}

musicgen/musicgen.go

@@ -40,6 +40,11 @@ type Request struct {
 	// Steps is the number of inference steps; nil = backend default.
 	Steps *int
 
+	// CFGScale is the classifier-free-guidance scale; nil = backend default.
+	// Architecture-sensitive, so prefer leaving it nil unless the caller
+	// knows the target model. Backends without the knob ignore it.
+	CFGScale *float64
+
 	// Seed fixes the RNG seed for reproducible output; nil = backend
 	// default (random).
 	Seed *int64
@@ -73,6 +78,9 @@ 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 } }
 
+// WithCFGScale overrides the classifier-free-guidance scale.
+func WithCFGScale(s float64) Option { return func(r *Request) { r.CFGScale = &s } }
+
 // WithSeed fixes the RNG seed.
 func WithSeed(seed int64) Option { return func(r *Request) { r.Seed = &seed } }
 

ocr/ocr.go

@@ -0,0 +1,114 @@
+// Package ocr is majordomo's canonical document text-recognition surface.
+// Like imagegen/audio/videogen/musicgen, it is a deliberately separate leaf
+// contract from the llm package (ADR-0023, following the ADR-0016→0022
+// lineage: functional options, zero values = backend default, bytes-only
+// I/O, Raw escape hatch). OCR is not chat-vision: it targets dedicated
+// detection+recognition models (Surya style) that return exact per-page,
+// per-line text rather than a model's paraphrase.
+//
+// The first implementation is provider/llamaswap, which posts the document
+// to a Surya shim through the /upstream passthrough; the shim rasterizes
+// PDFs itself, so Document may be an image (png/jpg/webp) or a PDF.
+package ocr
+
+import "context"
+
+// Request is a text-recognition request. Zero values mean "backend default".
+type Request struct {
+	// Document is the encoded document to recognize: an image (png/jpg/webp)
+	// or a PDF. Required. Carried as bytes, never a URL.
+	Document []byte
+
+	// MIME is the document MIME type (e.g. "image/png", "application/pdf");
+	// "" = let the backend sniff it.
+	MIME string
+
+	// Filename is the multipart filename hint some backends key their format
+	// detection on; "" derives one from MIME ("document.pdf") or falls back
+	// to "document".
+	Filename string
+
+	// Languages are ISO-639 hints for the recognizer (e.g. "en", "de");
+	// nil = backend default (auto/multilingual).
+	Languages []string
+
+	// MaxPages caps how many pages of a multi-page document are recognized;
+	// 0 = backend default (all pages, up to the backend's own ceiling).
+	MaxPages int
+}
+
+// Page is the recognized text of one page.
+type Page struct {
+	// Number is the 1-based page number.
+	Number int
+
+	// Text is the page's recognized text.
+	Text string
+}
+
+// Result is the canonical text-recognition result.
+type Result struct {
+	// Text is the full recognized text, pages joined in order.
+	Text string
+
+	// Pages are the per-page results in page order.
+	Pages []Page
+
+	// Raw is the provider-native response object (e.g. the per-line
+	// bbox/confidence/layout detail), an escape hatch for provider-specific
+	// fields. May be nil; never required for normal use.
+	Raw any
+}
+
+// Option mutates a Request before it is sent. Options passed to Recognize are
+// applied to a copy of the request, so a Request value can be reused.
+type Option func(*Request)
+
+// WithLanguages sets the recognizer language hints.
+func WithLanguages(langs ...string) Option {
+	return func(r *Request) { r.Languages = langs }
+}
+
+// WithMaxPages caps how many pages are recognized.
+func WithMaxPages(n int) Option { return func(r *Request) { r.MaxPages = n } }
+
+// Apply returns a copy of the request with all options applied. Providers
+// call this once at the top of Recognize.
+func (r Request) Apply(opts ...Option) Request {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Model recognizes text in documents.
+type Model interface {
+	// Recognize extracts the document's text, page by page.
+	Recognize(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 OCR models bound to one backend.
+type Provider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// OCRModel returns a Model bound to the given id (passed through to the
+	// backend verbatim; no catalog validation).
+	OCRModel(id string, opts ...ModelOption) (Model, error)
+}

parse_test.go

@@ -213,7 +213,9 @@ func TestBuiltinsResolve(t *testing.T) {
 	r := newTestRegistry(t)
 	// All built-in provider names resolve even before their client
 	// implementations land (stub providers error only on use).
-	for _, name := range []string{"openai", "anthropic", "google", "ollama", "ollama-cloud", "foreman"} {
+	// Note: llama-swap is intentionally excluded — its no-URL built-in errors
+	// at Model() construction (not just on use), so it can't resolve here.
+	for _, name := range []string{"openai", "kimi", "anthropic", "google", "ollama", "ollama-cloud", "foreman"} {
 		if _, err := r.Parse(name + "/anything"); err != nil {
 			t.Errorf("Parse(%s/anything): %v", name, err)
 		}

progress.md

@@ -264,3 +264,24 @@ tests flush out.
   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).
+
+## 2026-07-18 — Kimi (Moonshot AI) built-in provider (ADR-0026)
+
+- New built-in `kimi` provider + `kimi://` DSN scheme: Moonshot's
+  OpenAI-compatible Chat Completions endpoint, so both reuse `provider/openai`
+  (no new client, mirrors llama-swap's chat path). Default base URL
+  `https://api.moonshot.ai/v1`; China endpoint via
+  `LLM_KCN=kimi://[email protected]/v1`.
+- Credential is `KIMI_API_KEY` (read through the registry's injected envLookup,
+  so it's hermetically testable). `WithAPIKey` is passed unconditionally so an
+  unset `KIMI_API_KEY` can never fall through to the openai client's
+  `OPENAI_API_KEY` default.
+- New `openai.WithAPIKeyName` option customizes the missing-key error hint
+  (default `OPENAI_API_KEY`); the kimi built-in/scheme name `KIMI_API_KEY`.
+- Hermetic tests (capturing RoundTripper): built-in base URL + bearer, missing
+  key names KIMI_API_KEY with no OPENAI fallthrough and no network hit, and the
+  kimi:// scheme round-trips against the China host.
+- Docs kept in sync: README built-in table + DSN scheme list + support matrix
+  (footnote ³), `.env.example`, ADR-0026 (+ index; also backfilled the missing
+  0024/0025 index rows).
+- Consumer: mort names Kimi as a failover tier.

provider/llamaswap/audio.go

@@ -9,6 +9,7 @@ import (
 	"mime"
 	"net/http"
 	"net/url"
+	"strconv"
 	"strings"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
@@ -42,7 +43,11 @@ type speechRequest struct {
 	Speed          float64 `json:"speed,omitempty"`
 }
 
-// Speak implements audio.SpeechModel via POST {base}/v1/audio/speech.
+// Speak implements audio.SpeechModel via POST {base}/v1/audio/speech, or —
+// when the request carries reference audio for voice cloning — via multipart
+// POST {base}/upstream/<id>/v1/audio/speech/upload (the chatterbox clone
+// route; verified live 2026-07-16: stateless per-request cloning, fields
+// `input` + `voice_file`).
 func (m *speechModel) Speak(ctx context.Context, req audio.SpeechRequest, opts ...audio.SpeechOption) (*audio.SpeechResult, error) {
 	req = req.Apply(opts...)
 	if strings.TrimSpace(req.Input) == "" {
@@ -51,6 +56,9 @@ func (m *speechModel) Speak(ctx context.Context, req audio.SpeechRequest, opts .
 	if req.Speed < 0 {
 		return nil, fmt.Errorf("%w: speech speed must be >= 0, got %g", llm.ErrUnsupported, req.Speed)
 	}
+	if len(req.ReferenceAudio) > 0 {
+		return m.speakWithReference(ctx, req)
+	}
 	wire := speechRequest{
 		Model:          m.id,
 		Input:          req.Input,
@@ -72,6 +80,48 @@ func (m *speechModel) Speak(ctx context.Context, req audio.SpeechRequest, opts .
 	return &audio.SpeechResult{Audio: audioBytes, MIME: speechMIME(contentType, req.Format)}, nil
 }
 
+// speakWithReference performs zero-shot voice cloning via the upstream
+// passthrough clone route. The reference sample rides as the `voice_file`
+// part and the text as an `input` field; voice/format/speed still go on the
+// wire when set (upstreams ignore fields they don't understand). The result
+// MIME comes from the response header or content sniffing (audioResultMIME,
+// same validation as the sfx/enhance surfaces) — a JSON soft error or an
+// HTML proxy page must never be wrapped up as audio bytes.
+func (m *speechModel) speakWithReference(ctx context.Context, req audio.SpeechRequest) (*audio.SpeechResult, error) {
+	upPath, err := upstreamPath(m.id, "/v1/audio/speech/upload")
+	if err != nil {
+		return nil, err
+	}
+	speed := ""
+	if req.Speed != 0 {
+		speed = strconv.FormatFloat(req.Speed, 'g', -1, 64)
+	}
+	body, formType, err := buildMultipart("build speech clone form",
+		filePart{field: "voice_file", filename: transcriptionFilename("", req.ReferenceMIME), data: req.ReferenceAudio},
+		[]formField{
+			{"input", req.Input, true},
+			{"voice", req.Voice, false},
+			{"response_format", req.Format, false},
+			{"speed", speed, false},
+		})
+	if err != nil {
+		return nil, err
+	}
+	audioBytes, contentType, err := m.p.doRaw(ctx, http.MethodPost, upPath, m.id, formType, body, maxAudioResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(audioBytes) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "speech clone response contained no audio"}
+	}
+	mimeType := audioResultMIME(contentType, audioBytes)
+	if mimeType == "" {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("speech clone response is not audio (Content-Type %q): %s", contentType, truncateForError(audioBytes))}
+	}
+	return &audio.SpeechResult{Audio: audioBytes, MIME: mimeType}, 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).
@@ -93,6 +143,28 @@ func speechMIME(contentType, format string) string {
 	}
 }
 
+// audioResultMIME resolves an audio body's MIME type: the response
+// Content-Type when it is a concrete audio type, else content sniffing
+// (RIFF/WAVE, Ogg and friends), else "" — the caller treats undetectable as
+// an upstream error, mirroring videoMIME. Sniffed "audio/wave" is normalized
+// to the conventional "audio/wav", and Ogg's container type
+// "application/ogg" to "audio/ogg". Shared by the clone, sfx, and enhance
+// surfaces, which all answer raw audio bodies.
+func audioResultMIME(contentType string, data []byte) string {
+	if mt := mimeFromContentType(contentType, "audio/"); mt != "" {
+		return mt
+	}
+	switch mt := http.DetectContentType(data); {
+	case mt == "audio/wave":
+		return "audio/wav"
+	case mt == "application/ogg":
+		return "audio/ogg"
+	case strings.HasPrefix(mt, "audio/"):
+		return mt
+	}
+	return ""
+}
+
 // TranscriptionModel implements audio.TranscriptionProvider, binding a
 // speech-to-text model served by llama-swap (routed to a whisper.cpp-style
 // upstream).
@@ -118,13 +190,26 @@ func (m *transcriptionModel) Transcribe(ctx context.Context, req audio.Transcrip
 		return nil, fmt.Errorf("%w: transcription requires audio bytes", llm.ErrUnsupported)
 	}
 
+	// Translation is a per-request bool form field on whisper.cpp's server
+	// (server.cpp:534, verified 2026-07-16). Its language DEFAULT is "en",
+	// which would make translate a no-op — so an explicit language hint
+	// wins, but an unset one is forced to auto-detection.
+	translate := ""
+	language := req.Language
+	if req.Translate {
+		translate = "true"
+		if language == "" {
+			language = "auto"
+		}
+	}
 	buf, formType, err := buildMultipart("build transcription form",
 		filePart{field: "file", filename: transcriptionFilename(req.Filename, req.MIME), data: req.Audio},
 		[]formField{
 			{"model", m.id, true},
 			{"response_format", "json", true},
-			{"language", req.Language, false},
+			{"language", language, false},
 			{"prompt", req.Prompt, false},
+			{"translate", translate, false},
 		})
 	if err != nil {
 		return nil, err
@@ -177,10 +262,13 @@ func transcriptionFilename(filename, mimeType string) string {
 }
 
 // sanitizeFilename strips characters that would corrupt or inject into the
-// multipart Content-Disposition header. Quotes and backslashes are escaped
-// by mime/multipart itself; CR/LF are not — they must go.
+// multipart Content-Disposition header, or smuggle directory structure to
+// the receiving side. Quotes and backslashes are escaped by mime/multipart
+// itself, but a file-writing shim decodes them right back — so CR/LF/NUL
+// and both path separators are dropped: an upload-metadata filename must
+// never traverse ("../x", "a/b", "C:\x").
 func sanitizeFilename(name string) string {
-	name = strings.NewReplacer("\r", "", "\n", "").Replace(name)
+	name = strings.NewReplacer("\r", "", "\n", "", "\x00", "", "/", "", "\\", "").Replace(name)
 	return strings.TrimSpace(name)
 }
 
@@ -250,27 +338,41 @@ func parseVoices(raw []byte) ([]string, error) {
 // varies. contentType sets the request Content-Type when body is non-nil.
 // A response larger than maxBytes is an error, never a silent truncation.
 func (p *Provider) doRaw(ctx context.Context, method, path, model, contentType string, body io.Reader, maxBytes int64) ([]byte, string, error) {
-	if err := p.requireBaseURL(); err != nil {
+	data, hdr, err := p.doRawHeaders(ctx, method, path, model, contentType, body, maxBytes)
+	if err != nil {
 		return nil, "", err
 	}
+	return data, hdr.Get("Content-Type"), nil
+}
+
+// doRawHeaders is doRaw with the WHOLE response header rather than just
+// Content-Type. Only the face swap needs it — the shim reports whether the
+// likeness actually transferred in X-Swap-Report, and that answer would be
+// thrown away by a function that keeps one header — so doRaw stays the
+// signature 25 other call sites use and delegates here. Two bodies would be
+// two places for the size cap and the status check to drift apart.
+func (p *Provider) doRawHeaders(ctx context.Context, method, path, model, contentType string, body io.Reader, maxBytes int64) ([]byte, http.Header, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, nil, err
+	}
 	req, err := p.newRequest(ctx, method, path, contentType, body)
 	if err != nil {
-		return nil, "", err
+		return nil, nil, err
 	}
 	resp, err := p.client.Do(req)
 	if err != nil {
-		return nil, "", fmt.Errorf("llama-swap: do request: %w", err)
+		return nil, nil, fmt.Errorf("llama-swap: do request: %w", err)
 	}
 	defer resp.Body.Close()
 	if resp.StatusCode/100 != 2 {
-		return nil, "", p.apiError(resp, model)
+		return nil, nil, p.apiError(resp, model)
 	}
 	data, err := io.ReadAll(io.LimitReader(resp.Body, maxBytes+1))
 	if err != nil {
-		return nil, "", fmt.Errorf("llama-swap: read response: %w", err)
+		return nil, nil, fmt.Errorf("llama-swap: read response: %w", err)
 	}
 	if int64(len(data)) > maxBytes {
-		return nil, "", fmt.Errorf("llama-swap: response exceeds %d bytes", maxBytes)
+		return nil, nil, fmt.Errorf("llama-swap: response exceeds %d bytes", maxBytes)
 	}
-	return data, resp.Header.Get("Content-Type"), nil
+	return data, resp.Header, nil
 }

provider/llamaswap/clone_translate_test.go

@@ -0,0 +1,215 @@
+package llamaswap
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+func TestSpeakWithReferenceUsesCloneRoute(t *testing.T) {
+	var gotPath, gotInput, gotVoice, gotFilename string
+	var gotRef []byte
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotInput = r.FormValue("input")
+		gotVoice = r.FormValue("voice")
+		f, hdr, err := r.FormFile("voice_file")
+		if err != nil {
+			t.Fatalf("voice_file part: %v", err)
+		}
+		defer f.Close()
+		gotRef, _ = io.ReadAll(f)
+		gotFilename = hdr.Filename
+		w.Header().Set("Content-Type", "audio/wav")
+		_, _ = w.Write(wavFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("chatterbox")
+	res, err := sm.Speak(context.Background(),
+		audio.SpeechRequest{Input: "hello in my voice", Voice: "narrator"},
+		audio.WithReferenceAudio([]byte("REFWAV"), "audio/wav"))
+	if err != nil {
+		t.Fatalf("Speak: %v", err)
+	}
+	if gotPath != "/upstream/chatterbox/v1/audio/speech/upload" {
+		t.Errorf("path = %q, want clone route", gotPath)
+	}
+	if gotInput != "hello in my voice" || gotVoice != "narrator" {
+		t.Errorf("input/voice = %q/%q", gotInput, gotVoice)
+	}
+	if string(gotRef) != "REFWAV" || gotFilename != "audio.wav" {
+		t.Errorf("ref = %q name = %q", gotRef, gotFilename)
+	}
+	if res.MIME != "audio/wav" || len(res.Audio) == 0 {
+		t.Errorf("result = %q/%d bytes", res.MIME, len(res.Audio))
+	}
+}
+
+func TestSpeakWithReferenceSniffsHeaderlessWav(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		for _, k := range []string{"voice", "response_format", "speed"} {
+			if v, ok := r.MultipartForm.Value[k]; ok {
+				t.Errorf("unset request sent %q = %v, want omitted", k, v)
+			}
+		}
+		w.Header()["Content-Type"] = nil // no declared type at all
+		_, _ = w.Write(wavFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("chatterbox")
+	res, err := sm.Speak(context.Background(),
+		audio.SpeechRequest{Input: "hi"},
+		audio.WithReferenceAudio([]byte("REF"), ""))
+	if err != nil {
+		t.Fatalf("Speak: %v", err)
+	}
+	// Headerless RIFF sniffs audio/wave, normalized to the conventional wav.
+	if res.MIME != "audio/wav" {
+		t.Errorf("MIME = %q, want sniffed audio/wav", res.MIME)
+	}
+}
+
+func TestSpeakWithReferenceRejectsNonAudioResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// A FastAPI-style 2xx soft error must not be wrapped up as audio.
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"detail":"reference audio too short"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("chatterbox")
+	_, err := sm.Speak(context.Background(),
+		audio.SpeechRequest{Input: "hi"},
+		audio.WithReferenceAudio([]byte("REF"), "audio/wav"))
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-audio clone body", err)
+	}
+}
+
+func TestSanitizeFilenameStripsPathAndControlBytes(t *testing.T) {
+	for in, want := range map[string]string{
+		"song.mp3":                "song.mp3",
+		"../../etc/passwd":        "....etcpasswd",
+		"a\r\nContent-Type: evil": "aContent-Type: evil",
+		"..\\..\\boot.ini":        "....boot.ini",
+		"nul\x00byte.wav":         "nulbyte.wav",
+		"  /  ":                   "",
+	} {
+		if got := sanitizeFilename(in); got != want {
+			t.Errorf("sanitizeFilename(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestSpeakWithoutReferenceKeepsJSONRoute(t *testing.T) {
+	var gotPath, gotContentType string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotContentType = r.Header.Get("Content-Type")
+		_, _ = w.Write([]byte("MP3"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SpeechModel("chatterbox")
+	if _, err := sm.Speak(context.Background(), audio.SpeechRequest{Input: "hi"}); err != nil {
+		t.Fatalf("Speak: %v", err)
+	}
+	if gotPath != "/v1/audio/speech" || gotContentType != "application/json" {
+		t.Errorf("path/content-type = %q/%q, want JSON route", gotPath, gotContentType)
+	}
+}
+
+func TestTranscribeTranslateForcesAutoLanguage(t *testing.T) {
+	var gotTranslate, gotLanguage string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotTranslate = r.FormValue("translate")
+		gotLanguage = r.FormValue("language")
+		_, _ = w.Write([]byte(`{"text":"hello"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	tm, _ := p.TranscriptionModel("whisper-large-v3-turbo")
+	res, err := tm.Transcribe(context.Background(),
+		audio.TranscriptionRequest{Audio: []byte("AUDIO")},
+		audio.WithTranslate())
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	// whisper.cpp's server default language is "en", which would skip
+	// translation — translate must force auto-detection when no explicit
+	// language hint was given.
+	if gotTranslate != "true" || gotLanguage != "auto" {
+		t.Errorf("translate/language = %q/%q, want true/auto", gotTranslate, gotLanguage)
+	}
+	if res.Text != "hello" {
+		t.Errorf("text = %q", res.Text)
+	}
+}
+
+func TestTranscribeTranslateKeepsExplicitLanguage(t *testing.T) {
+	var gotTranslate, gotLanguage string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotTranslate = r.FormValue("translate")
+		gotLanguage = r.FormValue("language")
+		_, _ = w.Write([]byte(`{"text":"hello"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	tm, _ := p.TranscriptionModel("whisper-large-v3-turbo")
+	if _, err := tm.Transcribe(context.Background(),
+		audio.TranscriptionRequest{Audio: []byte("AUDIO"), Language: "de", Translate: true}); err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if gotTranslate != "true" || gotLanguage != "de" {
+		t.Errorf("translate/language = %q/%q, want true/de", gotTranslate, gotLanguage)
+	}
+}
+
+func TestTranscribeWithoutTranslateOmitsFields(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		for _, k := range []string{"translate", "language"} {
+			if v, ok := r.MultipartForm.Value[k]; ok {
+				t.Errorf("unset request sent %q = %v, want omitted", k, v)
+			}
+		}
+		_, _ = w.Write([]byte(`{"text":"hi"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	tm, _ := p.TranscriptionModel("whisper-large-v3-turbo")
+	if _, err := tm.Transcribe(context.Background(),
+		audio.TranscriptionRequest{Audio: []byte("AUDIO")}); err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+}

provider/llamaswap/edit_test.go

@@ -152,3 +152,76 @@ func TestImageEditWithoutMaskOmitsField(t *testing.T) {
 		t.Error("mask field sent for unmasked edit; want omitted")
 	}
 }
+
+// TestImageEditByReferenceUsesTxt2ImgExtraImages pins the instruction-edit
+// wire shape. It is a DIFFERENT endpoint and a DIFFERENT field from img2img,
+// and the difference is not cosmetic: measured against FLUX.1-Kontext on
+// 2026-07-30, the same prompt sent as init_images left the thing it was told
+// to change untouched and drifted everything else, while extra_images changed
+// exactly what was asked and left the rest of the frame numerically
+// unchanged. Routing a reference edit down the img2img path would look like
+// a working call and silently produce the wrong picture.
+func TestImageEditByReferenceUsesTxt2ImgExtraImages(t *testing.T) {
+	var gotPath string
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = 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("imagegen-flux-kontext")
+	ed := im.(imagegen.Editor)
+
+	ref := editInit(t)
+	if _, err := ed.Edit(context.Background(), imagegen.EditRequest{
+		Prompt: "make the sign read OPEN",
+		// Init/Mask/Strength are set and must be IGNORED — they describe a
+		// pipeline this model does not run.
+		Init:     ref,
+		Mask:     ref,
+		Strength: func() *float64 { s := 0.75; return &s }(),
+	}, imagegen.WithEditRefImages(ref)); err != nil {
+		t.Fatalf("reference edit: %v", err)
+	}
+
+	if gotPath != "/sdapi/v1/txt2img" {
+		t.Errorf("path = %q, want /sdapi/v1/txt2img (there is no init latent to denoise)", gotPath)
+	}
+	extra, ok := gotBody["extra_images"].([]any)
+	if !ok || len(extra) != 1 {
+		t.Fatalf("extra_images = %v, want the one reference image", gotBody["extra_images"])
+	}
+	if _, present := gotBody["init_images"]; present {
+		t.Error("init_images must NOT be sent on the reference path — it re-noises the picture")
+	}
+	if _, present := gotBody["denoising_strength"]; present {
+		t.Error("denoising_strength must NOT be sent on the reference path")
+	}
+	if _, present := gotBody["mask"]; present {
+		t.Error("mask must NOT be sent on the reference path")
+	}
+}
+
+// TestImageEditByReferenceRejectsEmptyRefs guards the case that would
+// otherwise silently become a plain txt2img: a reference edit whose only
+// reference carries no bytes has nothing to condition on, and rendering the
+// prompt from scratch is not what the caller asked for.
+func TestImageEditByReferenceRejectsEmptyRefs(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_, _ = w.Write([]byte(`{"images":["` + onePixelPNG + `"]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	im, _ := p.ImageModel("imagegen-flux-kontext")
+	ed := im.(imagegen.Editor)
+
+	_, err := ed.Edit(context.Background(), imagegen.EditRequest{Prompt: "anything"},
+		imagegen.WithEditRefImages(imagegen.Image{MIME: "image/png"}))
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported for an all-empty reference set", err)
+	}
+}

provider/llamaswap/enhance.go

@@ -0,0 +1,61 @@
+// enhance.go implements audio.SpeechEnhancementProvider against a
+// DeepFilterNet shim (audioutils) reached through llama-swap's /upstream
+// passthrough (ADR-0024):
+//
+//	POST /upstream/<id>/v1/enhance   multipart file -> WAV
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// SpeechEnhancerModel implements audio.SpeechEnhancementProvider. The id
+// selects which upstream llama-swap loads.
+func (p *Provider) SpeechEnhancerModel(id string, opts ...audio.SpeechEnhancerModelOption) (audio.SpeechEnhancer, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = audio.ApplySpeechEnhancerModelOptions(opts)
+	return &speechEnhancerModel{p: p, id: id}, nil
+}
+
+type speechEnhancerModel struct {
+	p  *Provider
+	id string
+}
+
+// Enhance implements audio.SpeechEnhancer. The endpoint always answers WAV.
+func (m *speechEnhancerModel) Enhance(ctx context.Context, req audio.EnhancementRequest, opts ...audio.EnhancementOption) (*audio.SpeechResult, error) {
+	req = req.Apply(opts...)
+	if len(req.Audio) == 0 {
+		return nil, fmt.Errorf("%w: speech enhancement requires audio bytes", llm.ErrUnsupported)
+	}
+	path, err := upstreamPath(m.id, "/v1/enhance")
+	if err != nil {
+		return nil, err
+	}
+	body, contentType, err := buildMultipart("build enhance form",
+		filePart{field: "file", filename: transcriptionFilename(req.Filename, req.MIME), data: req.Audio},
+		nil)
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxAudioResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "enhance response contained no audio"}
+	}
+	mimeType := audioResultMIME(respType, raw)
+	if mimeType == "" {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("enhance response is not audio (Content-Type %q): %s", respType, truncateForError(raw))}
+	}
+	return &audio.SpeechResult{Audio: raw, MIME: mimeType}, nil
+}

provider/llamaswap/faceswap.go

@@ -0,0 +1,236 @@
+// faceswap.go implements imagegen.FaceSwapper against the InsightFace shim
+// (buffalo_l + inswapper_128) reached through llama-swap's /upstream
+// passthrough (ADR-0024):
+//
+//	POST /upstream/<id>/v1/faces     multipart file            -> JSON
+//	POST /upstream/<id>/v1/faceswap  multipart target, source  -> PNG
+package llamaswap
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"mime"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// maxFaceSwapResponseBytes bounds the returned PNG. Generous: the shim echoes
+// the target's dimensions, and a 4K photo round-trips as a large lossless PNG.
+const maxFaceSwapResponseBytes = 64 << 20
+
+// FaceSwapModel implements the face-transfer surface. The id selects which
+// upstream llama-swap loads.
+func (p *Provider) FaceSwapModel(id string) (imagegen.FaceSwapper, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	return &faceSwapModel{p: p, id: id}, nil
+}
+
+type faceSwapModel struct {
+	p  *Provider
+	id string
+}
+
+// facesResponse mirrors the shim's /v1/faces body.
+type facesResponse struct {
+	Count int `json:"count"`
+	Faces []struct {
+		Index  int      `json:"index"`
+		Box    []int    `json:"box"`
+		Score  float64  `json:"score"`
+		Width  int      `json:"width"`
+		Height int      `json:"height"`
+		Yaw    *float64 `json:"yaw"`
+	} `json:"faces"`
+}
+
+// ListFaces implements imagegen.FaceSwapper.
+func (m *faceSwapModel) ListFaces(ctx context.Context, img imagegen.Image) ([]imagegen.DetectedFace, error) {
+	if len(img.Data) == 0 {
+		return nil, fmt.Errorf("%w: face detection requires image bytes", llm.ErrUnsupported)
+	}
+	path, err := upstreamPath(m.id, "/v1/faces")
+	if err != nil {
+		return nil, err
+	}
+	body, contentType, err := buildMultipart("build faces form",
+		filePart{field: "file", filename: imageFilename(img.MIME, "image"), data: img.Data}, nil)
+	if err != nil {
+		return nil, err
+	}
+	raw, _, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxFaceSwapResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	var parsed facesResponse
+	if err := json.Unmarshal(raw, &parsed); err != nil {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("faces response is not JSON: %s", truncateForError(raw))}
+	}
+	out := make([]imagegen.DetectedFace, 0, len(parsed.Faces))
+	for _, f := range parsed.Faces {
+		df := imagegen.DetectedFace{Index: f.Index, Score: f.Score, Yaw: f.Yaw}
+		// A short box would silently index out of range below; treat a
+		// malformed entry as a protocol error rather than zero-filling it,
+		// because a wrong box sends the caller at the wrong face.
+		if len(f.Box) != 4 {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("face %d has a %d-element box, want 4", f.Index, len(f.Box))}
+		}
+		copy(df.Box[:], f.Box)
+		out = append(out, df)
+	}
+	return out, nil
+}
+
+// FaceSwap implements imagegen.FaceSwapper. The endpoint always answers PNG.
+func (m *faceSwapModel) FaceSwap(ctx context.Context, req imagegen.FaceSwapRequest, opts ...imagegen.FaceSwapOption) (*imagegen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Target.Data) == 0 {
+		return nil, fmt.Errorf("%w: face swap requires a target image", llm.ErrUnsupported)
+	}
+	if len(req.Source.Data) == 0 {
+		return nil, fmt.Errorf("%w: face swap requires a source image", llm.ErrUnsupported)
+	}
+	// Only when it will actually be sent: under All the index is documented
+	// as ignored, so rejecting a negative one there would fail a request that
+	// is perfectly well formed.
+	if !req.All && req.Index != nil && *req.Index < 0 {
+		return nil, fmt.Errorf("%w: face index must be >= 0, got %d", llm.ErrUnsupported, *req.Index)
+	}
+	path, err := upstreamPath(m.id, "/v1/faceswap")
+	if err != nil {
+		return nil, err
+	}
+
+	var fields []formField
+	if req.All {
+		fields = append(fields, formField{key: "all", value: "true", required: true})
+	} else if req.Index != nil {
+		// Only sent when NOT swapping all: the shim ignores index under
+		// all=true, and sending both would imply a precedence the caller
+		// cannot see.
+		fields = append(fields, formField{key: "index", value: strconv.Itoa(*req.Index), required: true})
+	}
+
+	body, contentType, err := buildMultipartFiles("build faceswap form",
+		[]filePart{
+			{field: "target", filename: imageFilename(req.Target.MIME, "target"), data: req.Target.Data},
+			{field: "source", filename: imageFilename(req.Source.MIME, "source"), data: req.Source.Data},
+		}, fields)
+	if err != nil {
+		return nil, err
+	}
+	raw, respHdr, err := m.p.doRawHeaders(ctx, http.MethodPost, path, m.id, contentType, body, maxFaceSwapResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	respType := respHdr.Get("Content-Type")
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "face swap response contained no image"}
+	}
+	// Validate the BYTES, not the header. sniffImageMIME falls back to
+	// image/png when detection is inconclusive, so trusting it here would
+	// label a JSON error body as a PNG and return it as a successful image —
+	// and a header check alone misses the case where the response carries no
+	// Content-Type at all. The shim answers JSON on a semantic miss (no face
+	// found), which is exactly the body that would sail through.
+	detected := http.DetectContentType(raw)
+	if !strings.HasPrefix(detected, "image/") {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("face swap response is not an image (sniffed %q, Content-Type %q): %s",
+				detected, respType, truncateForError(raw))}
+	}
+	// Prefer the server's own label when it is an image type (it knows
+	// subtypes the sniffer does not), else what the bytes actually are.
+	mimeType := detected
+	if hdr := mimeFromContentType(respType, "image/"); hdr != "" {
+		mimeType = hdr
+	}
+	return &imagegen.Result{
+		Images:       []llm.ImagePart{{MIME: mimeType, Data: raw}},
+		SwappedFaces: parseSwapReport(respHdr.Get("X-Swap-Report")),
+	}, nil
+}
+
+// swapReport mirrors the shim's X-Swap-Report header.
+type swapReport struct {
+	Image []int `json:"image"`
+	Faces []struct {
+		Index              int      `json:"index"`
+		Size               []int    `json:"size"`
+		Yaw                *float64 `json:"yaw"`
+		IdentitySimilarity *float64 `json:"identity_similarity"`
+	} `json:"faces"`
+}
+
+// parseSwapReport decodes the measured outcome. A missing or malformed header
+// yields nil rather than an error: an older shim does not send it, and a swap
+// that produced a good image must not fail because the diagnostics alongside
+// it were unreadable.
+func parseSwapReport(header string) []imagegen.SwappedFace {
+	header = strings.TrimSpace(header)
+	if header == "" {
+		return nil
+	}
+	var rep swapReport
+	if err := json.Unmarshal([]byte(header), &rep); err != nil {
+		return nil
+	}
+	out := make([]imagegen.SwappedFace, 0, len(rep.Faces))
+	for _, f := range rep.Faces {
+		sf := imagegen.SwappedFace{
+			Index:              f.Index,
+			Yaw:                f.Yaw,
+			IdentitySimilarity: f.IdentitySimilarity,
+		}
+		if len(f.Size) == 2 {
+			sf.Width, sf.Height = f.Size[0], f.Size[1]
+		}
+		if len(rep.Image) == 2 {
+			sf.ImageWidth, sf.ImageHeight = rep.Image[0], rep.Image[1]
+		}
+		out = append(out, sf)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+// imageFilename picks a multipart filename for an image part. The shim reads
+// bytes, not names, but a plausible extension keeps server-side sniffing and
+// request logs honest. base distinguishes the parts of a multi-file form
+// ("target"/"source") so a log line says which one was malformed.
+//
+// initImageFilename (video.go) is this function with base fixed to "frame"
+// and delegates here — two copies of one extension table is how they drift.
+func imageFilename(mimeType, base string) string {
+	if base == "" {
+		base = "image"
+	}
+	mt := strings.ToLower(strings.TrimSpace(mimeType))
+	if parsed, _, err := mime.ParseMediaType(mt); err == nil {
+		mt = parsed
+	}
+	switch mt {
+	case "image/jpeg", "image/jpg":
+		return base + ".jpg"
+	case "image/webp":
+		return base + ".webp"
+	case "image/gif":
+		return base + ".gif"
+	case "image/bmp":
+		return base + ".bmp"
+	default:
+		// PNG is the safe default: every caller in this repo either sends PNG
+		// or sends something the decoder identifies by magic bytes anyway.
+		return base + ".png"
+	}
+}

provider/llamaswap/faceswap_test.go

@@ -0,0 +1,344 @@
+package llamaswap
+
+import (
+	"context"
+	"encoding/base64"
+	"errors"
+	"io"
+	"mime"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// parseParts pulls the multipart form a handler received.
+func parseParts(t *testing.T, r *http.Request) (files map[string][]byte, fields map[string]string) {
+	t.Helper()
+	_, params, err := mime.ParseMediaType(r.Header.Get("Content-Type"))
+	if err != nil {
+		t.Fatalf("content type: %v", err)
+	}
+	mr := multipart.NewReader(r.Body, params["boundary"])
+	files, fields = map[string][]byte{}, map[string]string{}
+	for {
+		p, err := mr.NextPart()
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			t.Fatalf("next part: %v", err)
+		}
+		body, _ := io.ReadAll(p)
+		if p.FileName() != "" {
+			files[p.FormName()] = body
+		} else {
+			fields[p.FormName()] = string(body)
+		}
+	}
+	return files, fields
+}
+
+// TestFaceSwapSendsBothFiles pins the two-file wire shape. A face swap is the
+// first endpoint in this provider taking more than one file, so buildMultipart
+// grew a sibling; getting the field NAMES wrong would reach the shim as a
+// missing-argument 422 rather than anything self-explanatory.
+func TestFaceSwapSendsBothFiles(t *testing.T) {
+	var gotPath string
+	var files map[string][]byte
+	var fields map[string]string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		files, fields = parseParts(t, r)
+		w.Header().Set("Content-Type", "image/png")
+		raw, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+		_, _ = w.Write(raw)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, err := p.FaceSwapModel("faceswap")
+	if err != nil {
+		t.Fatalf("model: %v", err)
+	}
+	img := editInit(t)
+	res, err := m.FaceSwap(context.Background(), imagegen.FaceSwapRequest{Target: img, Source: img},
+		imagegen.WithFaceIndex(2))
+	if err != nil {
+		t.Fatalf("faceswap: %v", err)
+	}
+	if len(res.Images) != 1 {
+		t.Fatalf("images = %d, want 1", len(res.Images))
+	}
+	if !strings.HasSuffix(gotPath, "/upstream/faceswap/v1/faceswap") {
+		t.Errorf("path = %q", gotPath)
+	}
+	for _, want := range []string{"target", "source"} {
+		if len(files[want]) == 0 {
+			t.Errorf("no %q file part — the shim requires both", want)
+		}
+	}
+	if fields["index"] != "2" {
+		t.Errorf("index = %q, want 2", fields["index"])
+	}
+	if _, ok := fields["all"]; ok {
+		t.Error("all sent alongside index — the shim ignores index under all=true, so sending both implies a precedence the caller cannot see")
+	}
+}
+
+// TestFaceSwapAllSuppressesIndex: same reasoning from the other side.
+func TestFaceSwapAllSuppressesIndex(t *testing.T) {
+	var fields map[string]string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, fields = parseParts(t, r)
+		w.Header().Set("Content-Type", "image/png")
+		raw, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+		_, _ = w.Write(raw)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	img := editInit(t)
+	if _, err := m.FaceSwap(context.Background(),
+		imagegen.FaceSwapRequest{Target: img, Source: img, Index: new(int), All: true}); err != nil {
+		t.Fatalf("faceswap: %v", err)
+	}
+	if fields["all"] != "true" {
+		t.Errorf("all = %q, want true", fields["all"])
+	}
+	if _, ok := fields["index"]; ok {
+		t.Error("index sent under all=true")
+	}
+}
+
+// TestFaceSwapRejectsMissingImages: both are required, and the error should
+// name which one rather than surfacing a shim 422.
+func TestFaceSwapRejectsMissingImages(t *testing.T) {
+	p := New(WithBaseURL("http://example.invalid"))
+	m, _ := p.FaceSwapModel("faceswap")
+	img := editInit(t)
+	for _, tc := range []struct {
+		name string
+		req  imagegen.FaceSwapRequest
+		want string
+	}{
+		{"no target", imagegen.FaceSwapRequest{Source: img}, "target"},
+		{"no source", imagegen.FaceSwapRequest{Target: img}, "source"},
+	} {
+		_, err := m.FaceSwap(context.Background(), tc.req)
+		if err == nil || !strings.Contains(err.Error(), tc.want) {
+			t.Errorf("%s: err = %v, want one naming %q", tc.name, err, tc.want)
+		}
+	}
+}
+
+// TestFaceSwapRejectsNonImageResponse: the shim answers JSON on a semantic
+// miss (no face found). Returning those bytes as an "image" would hand the
+// caller a file that is not a picture and call it success.
+func TestFaceSwapRejectsNonImageResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"detail":{"error":"no_face_in_source"}}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	img := editInit(t)
+	_, err := m.FaceSwap(context.Background(), imagegen.FaceSwapRequest{Target: img, Source: img})
+	if err == nil {
+		t.Fatal("a JSON body was accepted as an image")
+	}
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Errorf("err = %T, want *llm.APIError so callers can classify it", err)
+	}
+	if !strings.Contains(err.Error(), "no_face_in_source") {
+		t.Errorf("err = %v, want it to relay the shim's reason", err)
+	}
+}
+
+// TestListFacesParsesOrdering: the shim's left-to-right index is the contract
+// callers select against, so it must survive decoding intact.
+func TestListFacesParsesOrdering(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if !strings.HasSuffix(r.URL.Path, "/upstream/faceswap/v1/faces") {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"count":2,"faces":[
+			{"index":0,"box":[10,20,30,40],"score":0.9,"width":20,"height":20},
+			{"index":1,"box":[50,20,90,60],"score":0.8,"width":40,"height":40}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	faces, err := m.ListFaces(context.Background(), editInit(t))
+	if err != nil {
+		t.Fatalf("list: %v", err)
+	}
+	if len(faces) != 2 || faces[0].Index != 0 || faces[1].Index != 1 {
+		t.Fatalf("faces = %+v", faces)
+	}
+	if faces[1].Box != [4]int{50, 20, 90, 60} {
+		t.Errorf("box = %v", faces[1].Box)
+	}
+}
+
+// TestListFacesRejectsShortBox: a malformed box would send a caller at the
+// wrong face, which is worse than an error.
+func TestListFacesRejectsShortBox(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"count":1,"faces":[{"index":0,"box":[1,2],"score":0.9}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	if _, err := m.ListFaces(context.Background(), editInit(t)); err == nil {
+		t.Fatal("a 2-element box was accepted")
+	}
+}
+
+// TestFaceSwapRejectsHeaderlessNonImage is the regression for gadfly's
+// blocking finding on #23, agreed by both models. sniffImageMIME falls back
+// to image/png when detection is inconclusive, and the original guard only
+// looked at Content-Type — so a JSON error body sent WITHOUT a Content-Type
+// header was labelled a PNG and returned as a successful image. The shim
+// answers JSON on a semantic miss, which is precisely the body that would
+// have sailed through.
+func TestFaceSwapRejectsHeaderlessNonImage(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		// Explicitly no Content-Type — Go only sets one if we write before
+		// deleting it, so clear it to model a bare upstream response.
+		w.Header()["Content-Type"] = nil
+		_, _ = w.Write([]byte(`{"detail":{"error":"no_face_in_target"}}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	img := editInit(t)
+	_, err := m.FaceSwap(context.Background(), imagegen.FaceSwapRequest{Target: img, Source: img})
+	if err == nil {
+		t.Fatal("a headerless JSON body was accepted and would have been returned as image/png")
+	}
+	if !strings.Contains(err.Error(), "no_face_in_target") {
+		t.Errorf("err = %v, want it to relay what actually came back", err)
+	}
+}
+
+// TestFaceSwapAllowsNegativeIndexUnderAll: index is documented as ignored
+// when all=true, so validating it there would reject a well-formed request.
+func TestFaceSwapAllowsNegativeIndexUnderAll(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "image/png")
+		raw, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+		_, _ = w.Write(raw)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, _ := p.FaceSwapModel("faceswap")
+	img := editInit(t)
+	neg := -1
+	if _, err := m.FaceSwap(context.Background(),
+		imagegen.FaceSwapRequest{Target: img, Source: img, Index: &neg, All: true}); err != nil {
+		t.Fatalf("negative index rejected under all=true, where it is ignored: %v", err)
+	}
+	// ...but still rejected when it WOULD be sent.
+	if _, err := m.FaceSwap(context.Background(),
+		imagegen.FaceSwapRequest{Target: img, Source: img, Index: &neg}); err == nil {
+		t.Error("negative index accepted when it would actually be sent")
+	}
+}
+
+// TestFaceSwapParsesSwapReport: the measured outcome is the whole reason the
+// header exists — a caller that cannot tell "the likeness transferred" from
+// "an image came back" goes and asks a vision model, which is wrong in
+// exactly the cases that matter.
+func TestFaceSwapParsesSwapReport(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "image/png")
+		w.Header().Set("X-Swap-Report",
+			`{"image":[1010,1200],"faces":[{"index":2,"size":[138,172],"yaw":-82.2,"identity_similarity":0.791}]}`)
+		raw, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+		_, _ = w.Write(raw)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	m, err := p.FaceSwapModel("faceswap")
+	if err != nil {
+		t.Fatalf("model: %v", err)
+	}
+	img := editInit(t)
+	res, err := m.FaceSwap(context.Background(), imagegen.FaceSwapRequest{Target: img, Source: img})
+	if err != nil {
+		t.Fatalf("FaceSwap: %v", err)
+	}
+	if len(res.SwappedFaces) != 1 {
+		t.Fatalf("SwappedFaces = %d, want 1 — the measurement was dropped", len(res.SwappedFaces))
+	}
+	f := res.SwappedFaces[0]
+	if f.Index != 2 || f.Width != 138 || f.Height != 172 {
+		t.Errorf("face = %+v, want index 2 at 138x172", f)
+	}
+	if f.Yaw == nil || *f.Yaw != -82.2 {
+		t.Errorf("yaw = %v, want -82.2 — the pose signal is how a caller knows a profile swap will not read", f.Yaw)
+	}
+	if f.IdentitySimilarity == nil || *f.IdentitySimilarity != 0.791 {
+		t.Errorf("identity_similarity = %v, want 0.791", f.IdentitySimilarity)
+	}
+	// 138/1010 — the number that says "correct, and invisible at a glance".
+	if got := f.FractionOfImage(); got < 0.13 || got > 0.14 {
+		t.Errorf("FractionOfImage = %.3f, want ~0.137", got)
+	}
+}
+
+// TestFaceSwapSurvivesMissingReport: an older shim sends no header at all. A
+// swap that produced a good image must not fail because the diagnostics
+// beside it were absent or malformed.
+func TestFaceSwapSurvivesMissingReport(t *testing.T) {
+	for name, hdr := range map[string]string{
+		"absent":    "",
+		"garbage":   "not json at all",
+		"wrongtype": `{"faces":"nope"}`,
+	} {
+		t.Run(name, func(t *testing.T) {
+			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				w.Header().Set("Content-Type", "image/png")
+				if hdr != "" {
+					w.Header().Set("X-Swap-Report", hdr)
+				}
+				raw, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+				_, _ = w.Write(raw)
+			}))
+			defer srv.Close()
+
+			p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+			m, err := p.FaceSwapModel("faceswap")
+			if err != nil {
+				t.Fatalf("model: %v", err)
+			}
+			img := editInit(t)
+			res, err := m.FaceSwap(context.Background(), imagegen.FaceSwapRequest{Target: img, Source: img})
+			if err != nil {
+				t.Fatalf("a %s report failed the whole swap: %v", name, err)
+			}
+			if len(res.Images) != 1 {
+				t.Fatal("image lost")
+			}
+			if res.SwappedFaces != nil {
+				t.Errorf("SwappedFaces = %+v, want nil for a %s report", res.SwappedFaces, name)
+			}
+		})
+	}
+}

provider/llamaswap/image.go

@@ -135,9 +135,30 @@ type img2imgRequest struct {
 	Mask string `json:"mask,omitempty"`
 }
 
-// Edit implements imagegen.Editor via POST {base}/sdapi/v1/img2img.
+// refEditRequest is the wire shape for an INSTRUCTION-EDIT model. sd-server
+// exposes reference images as `extra_images` on the shared img-gen request
+// builder (routes_sdapi.cpp lands them in gen_params.ref_images — the same
+// place the CLI's -r/--ref-image goes), and that field is read on BOTH
+// /txt2img and /img2img.
+//
+// It posts to /txt2img because there is no init latent to denoise: the
+// reference IS the conditioning, so an init image plus a denoising strength
+// would only add noise to a pipeline that does not want any. Output
+// resolution follows the reference image.
+type refEditRequest struct {
+	txt2imgRequest
+	ExtraImages []string `json:"extra_images"`
+}
+
+// Edit implements imagegen.Editor. Two different pipelines live behind it,
+// selected by the request: RefImages routes to an instruction-edit model via
+// /sdapi/v1/txt2img + extra_images, everything else is img2img. See
+// imagegen.EditRequest.RefImages for why they are not interchangeable.
 func (m *imageModel) Edit(ctx context.Context, req imagegen.EditRequest, opts ...imagegen.EditOption) (*imagegen.Result, error) {
 	req = req.Apply(opts...)
+	if len(req.RefImages) > 0 {
+		return m.editByReference(ctx, req)
+	}
 	if len(req.Init.Data) == 0 {
 		return nil, fmt.Errorf("%w: image edit requires an init image", llm.ErrUnsupported)
 	}
@@ -164,6 +185,34 @@ func (m *imageModel) Edit(ctx context.Context, req imagegen.EditRequest, opts ..
 	return decodeImages(m.p.name, m.id, &resp)
 }
 
+// editByReference runs the instruction-edit path. Mask and Strength are
+// deliberately NOT rejected when set: a caller that hands the same
+// EditRequest to whichever model is configured should get the better result
+// on a Kontext-class model, not an error, and both fields describe a
+// pipeline that simply does not exist here.
+func (m *imageModel) editByReference(ctx context.Context, req imagegen.EditRequest) (*imagegen.Result, error) {
+	base, err := m.sdWire("reference edit", req.Prompt, req.NegativePrompt, req.Sampler, req.Size, req.Seed, req.Steps, req.CFGScale, req.N)
+	if err != nil {
+		return nil, err
+	}
+	wire := refEditRequest{txt2imgRequest: base}
+	for _, ref := range req.RefImages {
+		if len(ref.Data) == 0 {
+			continue
+		}
+		wire.ExtraImages = append(wire.ExtraImages, base64.StdEncoding.EncodeToString(ref.Data))
+	}
+	if len(wire.ExtraImages) == 0 {
+		return nil, fmt.Errorf("%w: reference edit requires at least one non-empty reference image", llm.ErrUnsupported)
+	}
+
+	var resp txt2imgResponse
+	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)
+}
+
 // 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) {

provider/llamaswap/lipsync.go

@@ -0,0 +1,97 @@
+// lipsync.go implements videogen.LipsyncProvider against a SadTalker shim
+// reached through llama-swap's /upstream passthrough (ADR-0025):
+//
+//	POST /upstream/<id>/v1/talking_head   multipart image,audio[,still,enhance,preprocess]
+//
+// The response body IS the encoded clip (same contract as /v1/videos/sync),
+// hence the video-sized response cap and the same MIME resolution rules.
+// Generation is sync and slow (minutes) — bound calls with a context
+// deadline (Hunyuan precedent).
+package llamaswap
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"mime/multipart"
+	"net/http"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/videogen"
+)
+
+// LipsyncModel implements videogen.LipsyncProvider. The id selects which
+// upstream llama-swap loads.
+func (p *Provider) LipsyncModel(id string, opts ...videogen.LipsyncModelOption) (videogen.Lipsyncer, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = videogen.ApplyLipsyncModelOptions(opts)
+	return &lipsyncModel{p: p, id: id}, nil
+}
+
+type lipsyncModel struct {
+	p  *Provider
+	id string
+}
+
+// Lipsync implements videogen.Lipsyncer.
+func (m *lipsyncModel) Lipsync(ctx context.Context, req videogen.LipsyncRequest, opts ...videogen.LipsyncOption) (*videogen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Image.Data) == 0 {
+		return nil, fmt.Errorf("%w: lipsync requires a portrait image", llm.ErrUnsupported)
+	}
+	if len(req.Audio) == 0 {
+		return nil, fmt.Errorf("%w: lipsync requires audio bytes", llm.ErrUnsupported)
+	}
+	if req.Preprocess != "" && req.Preprocess != "crop" && req.Preprocess != "full" {
+		return nil, fmt.Errorf("%w: lipsync preprocess must be \"crop\" or \"full\", got %q", llm.ErrUnsupported, req.Preprocess)
+	}
+	path, err := upstreamPath(m.id, "/v1/talking_head")
+	if err != nil {
+		return nil, err
+	}
+
+	// Two file parts — buildMultipart handles exactly one, so assemble by
+	// hand (mirrors videoModel.Generate).
+	var buf bytes.Buffer
+	w := multipart.NewWriter(&buf)
+	fw, err := w.CreateFormFile("image", initImageFilename(req.Image.MIME))
+	if err != nil {
+		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
+	}
+	if _, err := fw.Write(req.Image.Data); err != nil {
+		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
+	}
+	fw, err = w.CreateFormFile("audio", transcriptionFilename(req.AudioFilename, req.AudioMIME))
+	if err != nil {
+		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
+	}
+	if _, err := fw.Write(req.Audio); err != nil {
+		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
+	}
+	still := ""
+	if req.Still {
+		still = "true"
+	}
+	enhance := ""
+	if req.Enhance {
+		enhance = "true"
+	}
+	if err := writeFormFields(w, "build lipsync form", []formField{
+		{"still", still, false},
+		{"enhance", enhance, false},
+		{"preprocess", req.Preprocess, false},
+	}); err != nil {
+		return nil, err
+	}
+	if err := w.Close(); err != nil {
+		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
+	}
+
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, w.FormDataContentType(), &buf, maxVideoResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleVideoResult(m.p.name, m.id, "lipsync", raw, respType)
+}

provider/llamaswap/lipsync_videoutil_test.go

@@ -0,0 +1,295 @@
+package llamaswap
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/videogen"
+)
+
+// webmFixture is a minimal EBML header so http.DetectContentType sniffs
+// video/webm.
+func webmFixture() []byte {
+	return append([]byte{0x1A, 0x45, 0xDF, 0xA3}, make([]byte, 20)...)
+}
+
+func TestLipsync(t *testing.T) {
+	png := pngFixture(t)
+	var gotPath, gotStill, gotEnhance, gotPreprocess string
+	var gotImage, gotAudio []byte
+	var gotImageName, gotAudioName string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotStill = r.FormValue("still")
+		gotEnhance = r.FormValue("enhance")
+		gotPreprocess = r.FormValue("preprocess")
+		if f, hdr, err := r.FormFile("image"); err == nil {
+			gotImage, _ = io.ReadAll(f)
+			gotImageName = hdr.Filename
+			f.Close()
+		} else {
+			t.Errorf("image part: %v", err)
+		}
+		if f, hdr, err := r.FormFile("audio"); err == nil {
+			gotAudio, _ = io.ReadAll(f)
+			gotAudioName = hdr.Filename
+			f.Close()
+		} else {
+			t.Errorf("audio part: %v", err)
+		}
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ls, err := p.LipsyncModel("lipsync-sadtalker")
+	if err != nil {
+		t.Fatalf("LipsyncModel: %v", err)
+	}
+	res, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{
+			Image:     videogen.Image{MIME: "image/png", Data: png},
+			Audio:     []byte("SPEECH"),
+			AudioMIME: "audio/wav",
+		},
+		videogen.WithLipsyncStill(), videogen.WithLipsyncEnhance(), videogen.WithLipsyncPreprocess("full"))
+	if err != nil {
+		t.Fatalf("Lipsync: %v", err)
+	}
+	if gotPath != "/upstream/lipsync-sadtalker/v1/talking_head" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotStill != "true" || gotEnhance != "true" || gotPreprocess != "full" {
+		t.Errorf("still/enhance/preprocess = %q/%q/%q", gotStill, gotEnhance, gotPreprocess)
+	}
+	if string(gotImage) != string(png) || gotImageName != "frame.png" {
+		t.Errorf("image bytes/name = %d bytes/%q", len(gotImage), gotImageName)
+	}
+	if string(gotAudio) != "SPEECH" || gotAudioName != "audio.wav" {
+		t.Errorf("audio = %q name = %q", gotAudio, gotAudioName)
+	}
+	if res.Video.MIME != "video/mp4" || len(res.Video.Data) == 0 {
+		t.Fatalf("video = %q/%d bytes", res.Video.MIME, len(res.Video.Data))
+	}
+}
+
+func TestLipsyncOmitsUnsetFlags(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		for _, k := range []string{"still", "enhance", "preprocess"} {
+			if v, ok := r.MultipartForm.Value[k]; ok {
+				t.Errorf("unset request sent %q = %v, want omitted", k, v)
+			}
+		}
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ls, _ := p.LipsyncModel("lipsync-sadtalker")
+	if _, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{Image: videogen.Image{Data: pngFixture(t)}, Audio: []byte("A")}); err != nil {
+		t.Fatalf("Lipsync: %v", err)
+	}
+}
+
+func TestLipsyncRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	ls, _ := p.LipsyncModel("lipsync-sadtalker")
+	if _, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{Audio: []byte("A")}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no image: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{Image: videogen.Image{Data: []byte{1}}}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no audio: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{Image: videogen.Image{Data: []byte{1}}, Audio: []byte{1}, Preprocess: "zoom"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("preprocess zoom: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestLipsyncRejectsNonVideoResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html>proxy error page</html>"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ls, _ := p.LipsyncModel("lipsync-sadtalker")
+	_, err := ls.Lipsync(context.Background(),
+		videogen.LipsyncRequest{Image: videogen.Image{Data: pngFixture(t)}, Audio: []byte("A")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-video body", err)
+	}
+}
+
+func TestRemoveVideoBackground(t *testing.T) {
+	var gotPath, gotOutput, gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotOutput = r.FormValue("output")
+		if _, hdr, err := r.FormFile("file"); err == nil {
+			gotFilename = hdr.Filename
+		} else {
+			t.Errorf("file part: %v", err)
+		}
+		w.Header().Set("Content-Type", "video/webm")
+		_, _ = w.Write(webmFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vb, err := p.VideoBackgroundRemoverModel("mediautils")
+	if err != nil {
+		t.Fatalf("VideoBackgroundRemoverModel: %v", err)
+	}
+	res, err := vb.RemoveVideoBackground(context.Background(),
+		videogen.VideoBackgroundRemovalRequest{Video: mp4Fixture(), MIME: "video/mp4"},
+		videogen.WithVideoBackgroundOutput("alpha_webm"))
+	if err != nil {
+		t.Fatalf("RemoveVideoBackground: %v", err)
+	}
+	if gotPath != "/upstream/mediautils/v1/video/matte" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotOutput != "alpha_webm" || gotFilename != "video.mp4" {
+		t.Errorf("output/filename = %q/%q", gotOutput, gotFilename)
+	}
+	if res.Video.MIME != "video/webm" || len(res.Video.Data) == 0 {
+		t.Fatalf("video = %q/%d bytes", res.Video.MIME, len(res.Video.Data))
+	}
+}
+
+func TestRemoveVideoBackgroundOmitsDefaultOutput(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		if v, ok := r.MultipartForm.Value["output"]; ok {
+			t.Errorf("output field sent for default: %v; want omitted", v)
+		}
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vb, _ := p.VideoBackgroundRemoverModel("mediautils")
+	if _, err := vb.RemoveVideoBackground(context.Background(),
+		videogen.VideoBackgroundRemovalRequest{Video: mp4Fixture()}); err != nil {
+		t.Fatalf("RemoveVideoBackground: %v", err)
+	}
+}
+
+func TestRemoveVideoBackgroundRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	vb, _ := p.VideoBackgroundRemoverModel("mediautils")
+	if _, err := vb.RemoveVideoBackground(context.Background(),
+		videogen.VideoBackgroundRemovalRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no video: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := vb.RemoveVideoBackground(context.Background(),
+		videogen.VideoBackgroundRemovalRequest{Video: []byte{1}, Output: "gif"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("output gif: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestUpscaleVideo(t *testing.T) {
+	var gotPath, gotScale string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotScale = r.FormValue("scale")
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vu, err := p.VideoUpscalerModel("mediautils")
+	if err != nil {
+		t.Fatalf("VideoUpscalerModel: %v", err)
+	}
+	res, err := vu.UpscaleVideo(context.Background(),
+		videogen.VideoUpscaleRequest{Video: mp4Fixture(), MIME: "video/mp4"},
+		videogen.WithVideoUpscaleScale(2))
+	if err != nil {
+		t.Fatalf("UpscaleVideo: %v", err)
+	}
+	if gotPath != "/upstream/mediautils/v1/video/upscale" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotScale != "2" {
+		t.Errorf("scale = %q", gotScale)
+	}
+	if res.Video.MIME != "video/mp4" || len(res.Video.Data) == 0 {
+		t.Fatalf("video = %q/%d bytes", res.Video.MIME, len(res.Video.Data))
+	}
+}
+
+func TestUpscaleVideoRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	vu, _ := p.VideoUpscalerModel("mediautils")
+	if _, err := vu.UpscaleVideo(context.Background(), videogen.VideoUpscaleRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no video: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := vu.UpscaleVideo(context.Background(),
+		videogen.VideoUpscaleRequest{Video: []byte{1}, Scale: 3}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("scale 3: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestUpscaleVideoRejectsNonVideoResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header()["Content-Type"] = nil // NO Content-Type at all
+		_, _ = w.Write([]byte("502 bad gateway"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vu, _ := p.VideoUpscalerModel("mediautils")
+	_, err := vu.UpscaleVideo(context.Background(), videogen.VideoUpscaleRequest{Video: mp4Fixture()})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for headerless non-video body", err)
+	}
+}
+
+func TestVideoInputFilename(t *testing.T) {
+	cases := []struct {
+		filename, mime, want string
+	}{
+		{"clip.mp4", "", "clip.mp4"},
+		{"evil\r\nclip.mp4", "", "evilclip.mp4"},
+		{"", "video/mp4", "video.mp4"},
+		{"", "video/webm", "video.webm"},
+		{"", "video/quicktime", "video.mov"},
+		{"", "", "video"},
+	}
+	for _, tc := range cases {
+		if got := videoInputFilename(tc.filename, tc.mime); got != tc.want {
+			t.Errorf("videoInputFilename(%q, %q) = %q, want %q", tc.filename, tc.mime, got, tc.want)
+		}
+	}
+}

provider/llamaswap/llamaswap.go

@@ -53,6 +53,12 @@ const maxResponseBytes = 64 << 20
 // can't allocate without limit.
 const maxVideoResponseBytes = 512 << 20
 
+// maxAudioResponseBytes caps bodies that ARE a single encoded audio clip
+// (voice clone, speech enhancement, sfx): a long uncompressed WAV
+// legitimately passes the 64MB JSON cap. Still bounded so a buggy upstream
+// can't allocate without limit.
+const maxAudioResponseBytes = 256 << 20
+
 // Provider is a llama-swap client. It satisfies llm.Provider (chat, delegated
 // to provider/openai) and imagegen.Provider (image generation), and exposes
 // llama-swap's management endpoints as concrete methods.

provider/llamaswap/mediautil.go

@@ -207,12 +207,5 @@ func (m *interpolatorModel) Interpolate(ctx context.Context, req videogen.Interp
 	if err != nil {
 		return nil, err
 	}
-	if len(raw) == 0 {
-		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "interpolate response contained no video"}
-	}
-	mimeType := videoMIME(respType, raw)
-	if mimeType == "" {
-		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "interpolate response is not a video"}
-	}
-	return &videogen.Result{Video: videogen.Video{Data: raw, MIME: mimeType}}, nil
+	return singleVideoResult(m.p.name, m.id, "interpolate", raw, respType)
 }

provider/llamaswap/mediautil_test.go

@@ -238,7 +238,7 @@ func TestInterpolateRejectsBadArgs(t *testing.T) {
 }
 
 func TestUpstreamPathRejectsSeparators(t *testing.T) {
-	for _, bad := range []string{"", "a/b", "a?b", "a#b"} {
+	for _, bad := range []string{"", "a/b", "a?b", "a#b", "a%2Fb", "%2e%2e", "a%b"} {
 		if _, err := upstreamPath(bad, "/x"); err == nil {
 			t.Errorf("upstreamPath(%q) succeeded; want error", bad)
 		}

provider/llamaswap/mesh.go

@@ -42,7 +42,11 @@ type meshModel struct {
 
 // hunyuanGenerateRequest is the Hunyuan3D api_server /generate shape.
 // Optional fields are pointers/omitempty so unset values fall back to the
-// server's defaults (mirrors the sd-server wire structs).
+// server's defaults (mirrors the sd-server wire structs). Type is sent for
+// forward-compat but the LIVE server's GenerationRequest has no such field
+// and always returns GLB (verified 2026-07-14) — the response format is
+// therefore SNIFFED, and non-GLB output is the caller's conversion problem
+// (meshgen.Converter / the mediautils shim).
 type hunyuanGenerateRequest struct {
 	Image             string   `json:"image"`
 	Type              string   `json:"type,omitempty"`
@@ -123,7 +127,24 @@ func (m *meshModel) Generate(ctx context.Context, req meshgen.Request, opts ...m
 		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
 			Message: "mesh response is JSON, not mesh bytes: " + truncateForError(raw)}
 	}
-	return &meshgen.Result{Mesh: meshgen.Mesh{Data: raw, Format: format, MIME: mimeType}}, nil
+	// Label the result by what the bytes ARE, not what was requested: the
+	// live server ignores the format field entirely.
+	actualFormat, actualMIME := sniffMeshFormat(raw, format, mimeType)
+	return &meshgen.Result{Mesh: meshgen.Mesh{Data: raw, Format: actualFormat, MIME: actualMIME}}, nil
+}
+
+// sniffMeshFormat identifies the mesh container from magic bytes, falling
+// back to the requested format only when the bytes are ambiguous (binary
+// STL has no magic).
+func sniffMeshFormat(raw []byte, requested, requestedMIME string) (string, string) {
+	switch {
+	case len(raw) >= 4 && string(raw[:4]) == "glTF":
+		return "glb", meshFormats["glb"]
+	case len(raw) >= 6 && strings.EqualFold(string(raw[:6]), "solid "):
+		return "stl", meshFormats["stl"]
+	default:
+		return requested, requestedMIME
+	}
 }
 
 // truncateForError bounds a payload quoted into an error message.

provider/llamaswap/mesh_convert.go

@@ -0,0 +1,64 @@
+// mesh_convert.go implements meshgen.ConverterProvider against the
+// mediautils shim's POST /v1/convert_mesh (multipart file + target),
+// reached through the /upstream passthrough (ADR-0020). Exists because
+// Hunyuan3D's api_server always returns GLB — STL for the printer
+// pipeline is produced by this conversion hop.
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/meshgen"
+)
+
+// MeshConverter implements meshgen.ConverterProvider.
+func (p *Provider) MeshConverter(id string) (meshgen.Converter, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	return &meshConverter{p: p, id: id}, nil
+}
+
+type meshConverter struct {
+	p  *Provider
+	id string
+}
+
+// Convert implements meshgen.Converter.
+func (m *meshConverter) Convert(ctx context.Context, mesh meshgen.Mesh, format string) (*meshgen.Result, error) {
+	if len(mesh.Data) == 0 {
+		return nil, fmt.Errorf("%w: mesh conversion requires mesh bytes", llm.ErrUnsupported)
+	}
+	format = strings.ToLower(strings.TrimSpace(format))
+	mimeType, ok := meshFormats[format]
+	if !ok {
+		return nil, fmt.Errorf("%w: unsupported mesh format %q (want glb, stl, or obj)", llm.ErrUnsupported, format)
+	}
+	path, err := upstreamPath(m.id, "/v1/convert_mesh")
+	if err != nil {
+		return nil, err
+	}
+	filename := "mesh." + mesh.Format
+	if mesh.Format == "" {
+		filename = "mesh"
+	}
+	body, contentType, err := buildMultipart("build mesh-convert form",
+		filePart{field: "file", filename: filename, data: mesh.Data},
+		[]formField{{"target", format, true}})
+	if err != nil {
+		return nil, err
+	}
+	raw, _, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxMeshResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "mesh conversion returned no data"}
+	}
+	actualFormat, actualMIME := sniffMeshFormat(raw, format, mimeType)
+	return &meshgen.Result{Mesh: meshgen.Mesh{Data: raw, Format: actualFormat, MIME: actualMIME}}, nil
+}

provider/llamaswap/music.go

@@ -72,6 +72,36 @@ type queryItem struct {
 	Result string `json:"result"`
 }
 
+// musicResult is the useful subset of ACE-Step's double-encoded result
+// blob.
+type musicResult struct {
+	File string `json:"file"`
+}
+
+// parseMusicResult decodes the `result` string, tolerating two live-API
+// quirks (observed 2026-07-14): the payload is an ARRAY of result objects
+// (not a bare object), and string values can contain RAW control
+// characters (a literal newline in timing fields) that strict JSON
+// rejects. Control chars can only legally sit inside string values in the
+// double-encoded blob, so replacing them with spaces preserves structure.
+func parseMusicResult(blob string) (musicResult, bool) {
+	sanitized := strings.Map(func(r rune) rune {
+		if r < 0x20 {
+			return ' '
+		}
+		return r
+	}, blob)
+	var arr []musicResult
+	if err := json.Unmarshal([]byte(sanitized), &arr); err == nil && len(arr) > 0 {
+		return arr[0], true
+	}
+	var one musicResult
+	if err := json.Unmarshal([]byte(sanitized), &one); err == nil {
+		return one, true
+	}
+	return musicResult{}, false
+}
+
 // 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
@@ -232,10 +262,8 @@ func findQueryItem(raw json.RawMessage, taskID string) (*queryItem, error) {
 
 // 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 == "" {
+	result, ok := parseMusicResult(item.Result)
+	if !ok || result.File == "" {
 		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
 			Message: "music result blob missing file URL: " + truncateForError([]byte(item.Result))}
 	}

provider/llamaswap/music_embed_test.go

@@ -12,6 +12,7 @@ import (
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/embeddings"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/meshgen"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/musicgen"
 )
 
@@ -293,3 +294,75 @@ func TestMusicGenerateRejectsHostileFileURL(t *testing.T) {
 		t.Fatal("dot-dot result file URL accepted")
 	}
 }
+
+func TestParseMusicResultLiveShapes(t *testing.T) {
+	// Live ACE-Step (2026-07-14): result is an ARRAY and carries raw
+	// control characters inside string values.
+	arrayWithCtrl := "[{\"file\": \"/v1/audio?path=x.mp3\", \"prompt\": \"line1\nline2\"}]"
+	res, ok := parseMusicResult(arrayWithCtrl)
+	if !ok || res.File != "/v1/audio?path=x.mp3" {
+		t.Fatalf("array+ctrl: ok=%v res=%+v", ok, res)
+	}
+	// Docs shape (bare object) still parses.
+	res, ok = parseMusicResult(`{"file": "/v1/audio?path=y.mp3"}`)
+	if !ok || res.File != "/v1/audio?path=y.mp3" {
+		t.Fatalf("object: ok=%v res=%+v", ok, res)
+	}
+	if _, ok := parseMusicResult("not json"); ok {
+		t.Fatal("garbage parsed")
+	}
+}
+
+func TestMeshResultSniffsActualFormat(t *testing.T) {
+	// Live Hunyuan3D always returns GLB regardless of the requested
+	// format — the result must be labelled by its magic bytes.
+	glb := append([]byte("glTF"), []byte("\x02\x00\x00\x00rest")...)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/octet-stream")
+		_, _ = w.Write(glb)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mm, _ := p.MeshModel("image3d-hunyuan21")
+	res, err := mm.Generate(context.Background(),
+		meshgen.Request{Image: meshgen.Image{Data: []byte{1}}, Format: "stl"})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if res.Mesh.Format != "glb" || res.Mesh.MIME != "model/gltf-binary" {
+		t.Fatalf("mesh labelled %s/%s, want glb (sniffed)", res.Mesh.Format, res.Mesh.MIME)
+	}
+}
+
+func TestMeshConverter(t *testing.T) {
+	stl := []byte("solid m\nendsolid m\n")
+	var gotTarget, gotPath string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("multipart: %v", err)
+		}
+		gotTarget = r.FormValue("target")
+		w.Header().Set("Content-Type", "model/stl")
+		_, _ = w.Write(stl)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	mc, err := p.MeshConverter("mediautils")
+	if err != nil {
+		t.Fatalf("MeshConverter: %v", err)
+	}
+	res, err := mc.Convert(context.Background(),
+		meshgen.Mesh{Data: []byte("glTFxxxx"), Format: "glb"}, "stl")
+	if err != nil {
+		t.Fatalf("Convert: %v", err)
+	}
+	if gotPath != "/upstream/mediautils/v1/convert_mesh" || gotTarget != "stl" {
+		t.Errorf("path/target = %q/%q", gotPath, gotTarget)
+	}
+	if res.Mesh.Format != "stl" {
+		t.Errorf("format = %q", res.Mesh.Format)
+	}
+}

provider/llamaswap/ocr.go

@@ -0,0 +1,140 @@
+// ocr.go implements ocr.Provider against a Surya-style shim reached through
+// llama-swap's /upstream passthrough (ADR-0023):
+//
+//	POST /upstream/<id>/v1/ocr   multipart file[,langs,max_pages]
+//
+// The document may be an image or a PDF (the shim rasterizes PDFs itself).
+// The response is per-page JSON: {pages:[{number,text,lines,layout}]}. The
+// decode is tolerant — when a page carries no aggregate `text`, its line
+// texts are joined instead — and the full payload survives in Result.Raw for
+// callers that want bboxes/confidence/layout.
+package llamaswap
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"mime"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/ocr"
+)
+
+// OCRModel implements ocr.Provider. The id selects which upstream llama-swap
+// loads.
+func (p *Provider) OCRModel(id string, opts ...ocr.ModelOption) (ocr.Model, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = ocr.ApplyModelOptions(opts)
+	return &ocrModel{p: p, id: id}, nil
+}
+
+type ocrModel struct {
+	p  *Provider
+	id string
+}
+
+// ocrResponse is the Surya shim's /v1/ocr shape (the subset this package
+// relies on; per-line bbox/confidence and layout stay in Raw).
+type ocrResponse struct {
+	Pages []struct {
+		Number int    `json:"number"`
+		Text   string `json:"text"`
+		Lines  []struct {
+			Text string `json:"text"`
+		} `json:"lines"`
+	} `json:"pages"`
+}
+
+// Recognize implements ocr.Model.
+func (m *ocrModel) Recognize(ctx context.Context, req ocr.Request, opts ...ocr.Option) (*ocr.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Document) == 0 {
+		return nil, fmt.Errorf("%w: ocr requires document bytes", llm.ErrUnsupported)
+	}
+	if req.MaxPages < 0 {
+		return nil, fmt.Errorf("%w: ocr max pages must be >= 0, got %d", llm.ErrUnsupported, req.MaxPages)
+	}
+	path, err := upstreamPath(m.id, "/v1/ocr")
+	if err != nil {
+		return nil, err
+	}
+	maxPages := ""
+	if req.MaxPages != 0 {
+		maxPages = strconv.Itoa(req.MaxPages)
+	}
+	body, contentType, err := buildMultipart("build ocr form",
+		filePart{field: "file", filename: documentFilename(req.Filename, req.MIME), data: req.Document},
+		[]formField{
+			{"langs", strings.Join(req.Languages, ","), false},
+			{"max_pages", maxPages, false},
+		})
+	if err != nil {
+		return nil, err
+	}
+	raw, _, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+
+	var out ocrResponse
+	if err := json.Unmarshal(raw, &out); err != nil {
+		return nil, fmt.Errorf("llama-swap: decode ocr response: %w", err)
+	}
+	if len(out.Pages) == 0 {
+		// A blank page still comes back as a page with empty text; zero pages
+		// is API drift or a soft error, never "the document was empty".
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: "ocr response contained no pages: " + truncateForError(raw)}
+	}
+	res := &ocr.Result{Raw: json.RawMessage(raw)}
+	texts := make([]string, 0, len(out.Pages))
+	for i, pg := range out.Pages {
+		text := pg.Text
+		if text == "" && len(pg.Lines) > 0 {
+			lines := make([]string, 0, len(pg.Lines))
+			for _, ln := range pg.Lines {
+				lines = append(lines, ln.Text)
+			}
+			text = strings.Join(lines, "\n")
+		}
+		number := pg.Number
+		if number == 0 {
+			number = i + 1
+		}
+		res.Pages = append(res.Pages, ocr.Page{Number: number, Text: text})
+		texts = append(texts, text)
+	}
+	res.Text = strings.Join(texts, "\n\n")
+	return res, nil
+}
+
+// documentFilename picks the multipart filename hint for an OCR document: the
+// caller's (sanitized), else one derived from the MIME subtype
+// ("document.pdf"), else "document". MIME parameters are stripped before
+// matching, mirroring transcriptionFilename.
+func documentFilename(filename, mimeType string) string {
+	if name := sanitizeFilename(filename); name != "" {
+		return name
+	}
+	mt := strings.ToLower(strings.TrimSpace(mimeType))
+	if parsed, _, err := mime.ParseMediaType(mt); err == nil {
+		mt = parsed
+	}
+	switch mt {
+	case "application/pdf":
+		return "document.pdf"
+	case "image/png":
+		return "document.png"
+	case "image/jpeg", "image/jpg":
+		return "document.jpg"
+	case "image/webp":
+		return "document.webp"
+	default:
+		return "document"
+	}
+}

provider/llamaswap/ocr_test.go

@@ -0,0 +1,200 @@
+package llamaswap
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/ocr"
+)
+
+func TestRecognize(t *testing.T) {
+	var gotPath, gotLangs, gotMaxPages, gotFilename string
+	var gotFile []byte
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotLangs = r.FormValue("langs")
+		gotMaxPages = r.FormValue("max_pages")
+		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(`{"pages":[
+			{"number":1,"text":"page one","lines":[{"text":"page","bbox":[0,0,1,1],"confidence":0.9},{"text":"one"}],"layout":{}},
+			{"number":2,"text":"page two"}
+		]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, err := p.OCRModel("ocr-surya")
+	if err != nil {
+		t.Fatalf("OCRModel: %v", err)
+	}
+	res, err := om.Recognize(context.Background(),
+		ocr.Request{Document: []byte("%PDF"), MIME: "application/pdf"},
+		ocr.WithLanguages("en", "de"), ocr.WithMaxPages(5))
+	if err != nil {
+		t.Fatalf("Recognize: %v", err)
+	}
+	if gotPath != "/upstream/ocr-surya/v1/ocr" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotLangs != "en,de" || gotMaxPages != "5" {
+		t.Errorf("langs/max_pages = %q/%q", gotLangs, gotMaxPages)
+	}
+	if string(gotFile) != "%PDF" || gotFilename != "document.pdf" {
+		t.Errorf("file = %q name = %q", gotFile, gotFilename)
+	}
+	if len(res.Pages) != 2 || res.Pages[0].Number != 1 || res.Pages[0].Text != "page one" ||
+		res.Pages[1].Number != 2 || res.Pages[1].Text != "page two" {
+		t.Errorf("pages = %+v", res.Pages)
+	}
+	if res.Text != "page one\n\npage two" {
+		t.Errorf("text = %q", res.Text)
+	}
+	if res.Raw == nil {
+		t.Error("Raw = nil, want raw payload")
+	}
+}
+
+func TestRecognizeJoinsLinesWhenPageTextAbsent(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"pages":[{"lines":[{"text":"first line"},{"text":"second line"}]}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, _ := p.OCRModel("ocr-surya")
+	res, err := om.Recognize(context.Background(), ocr.Request{Document: []byte("img")})
+	if err != nil {
+		t.Fatalf("Recognize: %v", err)
+	}
+	if len(res.Pages) != 1 || res.Pages[0].Text != "first line\nsecond line" {
+		t.Errorf("pages = %+v", res.Pages)
+	}
+	// Missing page number defaults to position.
+	if res.Pages[0].Number != 1 {
+		t.Errorf("number = %d, want 1", res.Pages[0].Number)
+	}
+	if res.Text != "first line\nsecond line" {
+		t.Errorf("text = %q", res.Text)
+	}
+}
+
+func TestRecognizeOmitsUnsetFields(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		for _, k := range []string{"langs", "max_pages"} {
+			if v, ok := r.MultipartForm.Value[k]; ok {
+				t.Errorf("unset request sent %q = %v, want omitted", k, v)
+			}
+		}
+		if _, hdr, err := r.FormFile("file"); err == nil {
+			if hdr.Filename != "document" {
+				t.Errorf("filename = %q, want document fallback", hdr.Filename)
+			}
+		}
+		_, _ = w.Write([]byte(`{"pages":[{"number":1,"text":"x"}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, _ := p.OCRModel("ocr-surya")
+	if _, err := om.Recognize(context.Background(), ocr.Request{Document: []byte("img")}); err != nil {
+		t.Fatalf("Recognize: %v", err)
+	}
+}
+
+func TestRecognizeRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	om, _ := p.OCRModel("ocr-surya")
+	if _, err := om.Recognize(context.Background(), ocr.Request{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no document: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := om.Recognize(context.Background(),
+		ocr.Request{Document: []byte{1}, MaxPages: -1}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("negative max pages: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestRecognizeRejectsZeroPages(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"pages":[]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, _ := p.OCRModel("ocr-surya")
+	_, err := om.Recognize(context.Background(), ocr.Request{Document: []byte("img")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for zero pages", err)
+	}
+}
+
+func TestRecognizeRejectsNonJSONResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html>proxy error page</html>"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, _ := p.OCRModel("ocr-surya")
+	if _, err := om.Recognize(context.Background(), ocr.Request{Document: []byte("img")}); err == nil ||
+		!strings.Contains(err.Error(), "decode ocr response") {
+		t.Fatalf("err = %v, want decode error for non-JSON body", err)
+	}
+}
+
+func TestRecognizeSurfacesAPIError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write([]byte(`{"error":{"message":"unsupported file type"}}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	om, _ := p.OCRModel("ocr-surya")
+	_, err := om.Recognize(context.Background(), ocr.Request{Document: []byte("bad")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %T %v, want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusBadRequest || apiErr.Message != "unsupported file type" || apiErr.Model != "ocr-surya" {
+		t.Errorf("apiErr = %+v", apiErr)
+	}
+}
+
+func TestDocumentFilename(t *testing.T) {
+	cases := []struct {
+		filename, mime, want string
+	}{
+		{"scan.pdf", "", "scan.pdf"},
+		{"evil\r\nname.pdf", "", "evilname.pdf"},
+		{"", "application/pdf", "document.pdf"},
+		{"", "image/png", "document.png"},
+		{"", "image/jpeg", "document.jpg"},
+		{"", "image/webp; charset=binary", "document.webp"},
+		{"", "", "document"},
+	}
+	for _, tc := range cases {
+		if got := documentFilename(tc.filename, tc.mime); got != tc.want {
+			t.Errorf("documentFilename(%q, %q) = %q, want %q", tc.filename, tc.mime, got, tc.want)
+		}
+	}
+}

provider/llamaswap/restore.go

@@ -0,0 +1,106 @@
+// restore.go implements the imagegen.ColorizeProvider and
+// imagegen.FaceRestoreProvider surfaces against the mediautils shim reached
+// through llama-swap's /upstream passthrough (ADR-0023):
+//
+//	colorize       POST /upstream/<id>/v1/colorize        (DDColor)
+//	restore_faces  POST /upstream/<id>/v1/restore_faces   (GFPGAN)
+//
+// Both are one-file multipart in, one PNG out, mirroring mediautil.go.
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strconv"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// --- colorize ---
+
+// ColorizeModel implements imagegen.ColorizeProvider against the mediautils
+// shim's POST /v1/colorize. The id selects which upstream llama-swap loads.
+func (p *Provider) ColorizeModel(id string, opts ...imagegen.ColorizeModelOption) (imagegen.Colorizer, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = imagegen.ApplyColorizeModelOptions(opts)
+	return &colorizeModel{p: p, id: id}, nil
+}
+
+type colorizeModel struct {
+	p  *Provider
+	id string
+}
+
+// Colorize implements imagegen.Colorizer.
+func (m *colorizeModel) Colorize(ctx context.Context, req imagegen.ColorizeRequest, opts ...imagegen.ColorizeOption) (*imagegen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Image.Data) == 0 {
+		return nil, fmt.Errorf("%w: colorization requires an image", llm.ErrUnsupported)
+	}
+	path, err := upstreamPath(m.id, "/v1/colorize")
+	if err != nil {
+		return nil, err
+	}
+	body, contentType, err := buildMultipart("build colorize form",
+		filePart{field: "file", filename: "image.png", data: req.Image.Data},
+		nil)
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxImageResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleImageResult(m.p.name, m.id, "colorize", raw, respType)
+}
+
+// --- face restoration ---
+
+// FaceRestoreModel implements imagegen.FaceRestoreProvider against the
+// mediautils shim's POST /v1/restore_faces.
+func (p *Provider) FaceRestoreModel(id string, opts ...imagegen.FaceRestoreModelOption) (imagegen.FaceRestorer, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = imagegen.ApplyFaceRestoreModelOptions(opts)
+	return &faceRestoreModel{p: p, id: id}, nil
+}
+
+type faceRestoreModel struct {
+	p  *Provider
+	id string
+}
+
+// RestoreFaces implements imagegen.FaceRestorer.
+func (m *faceRestoreModel) RestoreFaces(ctx context.Context, req imagegen.FaceRestoreRequest, opts ...imagegen.FaceRestoreOption) (*imagegen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Image.Data) == 0 {
+		return nil, fmt.Errorf("%w: face restoration requires an image", llm.ErrUnsupported)
+	}
+	if req.Upscale != 0 && req.Upscale != 1 && req.Upscale != 2 {
+		return nil, fmt.Errorf("%w: face-restore upscale must be 1 or 2, got %d", llm.ErrUnsupported, req.Upscale)
+	}
+	path, err := upstreamPath(m.id, "/v1/restore_faces")
+	if err != nil {
+		return nil, err
+	}
+	upscale := ""
+	if req.Upscale != 0 {
+		upscale = strconv.Itoa(req.Upscale)
+	}
+	body, contentType, err := buildMultipart("build restore-faces form",
+		filePart{field: "file", filename: "image.png", data: req.Image.Data},
+		[]formField{{"upscale", upscale, false}})
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxImageResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleImageResult(m.p.name, m.id, "face restoration", raw, respType)
+}

provider/llamaswap/segment.go

@@ -0,0 +1,79 @@
+// segment.go implements imagegen.SegmentationProvider against a
+// GroundingDINO+SAM shim (segment-langsam) reached through llama-swap's
+// /upstream passthrough (ADR-0023):
+//
+//	POST /upstream/<id>/v1/segment   multipart file,prompt[,threshold],output=mask
+//
+// The response is a single grayscale mask PNG where WHITE marks the prompted
+// region — directly usable as imagegen.EditRequest.Mask (white = repaint).
+// The shim also offers output=cutout|boxes; this client always requests the
+// mask, because a cutout is derivable client-side from mask+original with no
+// second GPU call.
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"math"
+	"net/http"
+	"strconv"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// SegmentationModel implements imagegen.SegmentationProvider. The id selects
+// which upstream llama-swap loads.
+func (p *Provider) SegmentationModel(id string, opts ...imagegen.SegmentationModelOption) (imagegen.Segmenter, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = imagegen.ApplySegmentationModelOptions(opts)
+	return &segmentationModel{p: p, id: id}, nil
+}
+
+type segmentationModel struct {
+	p  *Provider
+	id string
+}
+
+// Segment implements imagegen.Segmenter.
+func (m *segmentationModel) Segment(ctx context.Context, req imagegen.SegmentationRequest, opts ...imagegen.SegmentationOption) (*imagegen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Image.Data) == 0 {
+		return nil, fmt.Errorf("%w: segmentation requires an image", llm.ErrUnsupported)
+	}
+	if req.Prompt == "" {
+		return nil, fmt.Errorf("%w: segmentation requires a prompt", llm.ErrUnsupported)
+	}
+	// NaN fails every comparison, so it would sail through a bare range
+	// check and reach the upstream as the literal string "NaN".
+	if math.IsNaN(req.Threshold) || req.Threshold < 0 || req.Threshold > 1 {
+		return nil, fmt.Errorf("%w: segmentation threshold must be in [0,1], got %g", llm.ErrUnsupported, req.Threshold)
+	}
+	path, err := upstreamPath(m.id, "/v1/segment")
+	if err != nil {
+		return nil, err
+	}
+	threshold := ""
+	if req.Threshold != 0 {
+		threshold = strconv.FormatFloat(req.Threshold, 'g', -1, 64)
+	}
+	body, contentType, err := buildMultipart("build segment form",
+		filePart{field: "file", filename: "image.png", data: req.Image.Data},
+		[]formField{
+			{"prompt", req.Prompt, true},
+			{"threshold", threshold, false},
+			// Always the mask: white = prompted region, EditRequest.Mask
+			// polarity. Cutouts are derived client-side.
+			{"output", "mask", true},
+		})
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxImageResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleImageResult(m.p.name, m.id, "segmentation", raw, respType)
+}

provider/llamaswap/segment_restore_test.go

@@ -0,0 +1,281 @@
+package llamaswap
+
+import (
+	"context"
+	"errors"
+	"math"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+func TestSegment(t *testing.T) {
+	png := pngFixture(t)
+	var gotPath, gotPrompt, gotThreshold, gotOutput, gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("parse multipart: %v", err)
+		}
+		gotPrompt = r.FormValue("prompt")
+		gotThreshold = r.FormValue("threshold")
+		gotOutput = r.FormValue("output")
+		if f, hdr, err := r.FormFile("file"); err == nil {
+			gotFilename = hdr.Filename
+			f.Close()
+		} else {
+			t.Errorf("file part: %v", err)
+		}
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(png)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sg, err := p.SegmentationModel("segment-langsam")
+	if err != nil {
+		t.Fatalf("SegmentationModel: %v", err)
+	}
+	res, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Image: imagegen.Image{MIME: "image/png", Data: png}, Prompt: "the red car"},
+		imagegen.WithSegmentationThreshold(0.35))
+	if err != nil {
+		t.Fatalf("Segment: %v", err)
+	}
+	if gotPath != "/upstream/segment-langsam/v1/segment" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotPrompt != "the red car" || gotThreshold != "0.35" || gotOutput != "mask" || gotFilename != "image.png" {
+		t.Errorf("prompt/threshold/output/filename = %q/%q/%q/%q", gotPrompt, gotThreshold, gotOutput, gotFilename)
+	}
+	if len(res.Images) != 1 || res.Images[0].MIME != "image/png" {
+		t.Fatalf("images = %+v", res.Images)
+	}
+}
+
+func TestSegmentOmitsDefaultThreshold(t *testing.T) {
+	png := pngFixture(t)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("parse multipart: %v", err)
+		}
+		if _, ok := r.MultipartForm.Value["threshold"]; ok {
+			t.Error("threshold field sent for default; want omitted")
+		}
+		if got := r.FormValue("output"); got != "mask" {
+			t.Errorf("output = %q, want mask (always sent)", got)
+		}
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(png)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sg, _ := p.SegmentationModel("segment-langsam")
+	if _, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Image: imagegen.Image{Data: png}, Prompt: "dog"}); err != nil {
+		t.Fatalf("Segment: %v", err)
+	}
+}
+
+func TestSegmentRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	sg, _ := p.SegmentationModel("segment-langsam")
+	if _, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Prompt: "dog"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no image: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Image: imagegen.Image{Data: []byte{1}}}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no prompt: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Image: imagegen.Image{Data: []byte{1}}, Prompt: "dog", Threshold: 1.5}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("threshold 1.5: err = %v, want ErrUnsupported", err)
+	}
+	for name, bad := range map[string]float64{
+		"NaN":  math.NaN(),
+		"+Inf": math.Inf(1),
+		"-Inf": math.Inf(-1),
+	} {
+		if _, err := sg.Segment(context.Background(),
+			imagegen.SegmentationRequest{Image: imagegen.Image{Data: []byte{1}}, Prompt: "dog", Threshold: bad}); !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("threshold %s: err = %v, want ErrUnsupported", name, err)
+		}
+	}
+}
+
+func TestSegmentRejectsNonImageResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html>proxy error page</html>"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sg, _ := p.SegmentationModel("segment-langsam")
+	_, err := sg.Segment(context.Background(),
+		imagegen.SegmentationRequest{Image: imagegen.Image{Data: pngFixture(t)}, Prompt: "dog"})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-image body", err)
+	}
+}
+
+func TestColorize(t *testing.T) {
+	png := pngFixture(t)
+	var gotPath, gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("parse multipart: %v", err)
+		}
+		if f, hdr, err := r.FormFile("file"); err == nil {
+			gotFilename = hdr.Filename
+			f.Close()
+		} else {
+			t.Errorf("file part: %v", err)
+		}
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(png)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	cl, err := p.ColorizeModel("mediautils")
+	if err != nil {
+		t.Fatalf("ColorizeModel: %v", err)
+	}
+	res, err := cl.Colorize(context.Background(),
+		imagegen.ColorizeRequest{Image: imagegen.Image{Data: png}})
+	if err != nil {
+		t.Fatalf("Colorize: %v", err)
+	}
+	if gotPath != "/upstream/mediautils/v1/colorize" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotFilename != "image.png" {
+		t.Errorf("filename = %q", gotFilename)
+	}
+	if len(res.Images) != 1 || res.Images[0].MIME != "image/png" {
+		t.Fatalf("images = %+v", res.Images)
+	}
+}
+
+func TestColorizeRejectsEmptyImage(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	cl, _ := p.ColorizeModel("mediautils")
+	if _, err := cl.Colorize(context.Background(), imagegen.ColorizeRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestColorizeSurfacesAPIError(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()))
+	cl, _ := p.ColorizeModel("mediautils")
+	_, err := cl.Colorize(context.Background(),
+		imagegen.ColorizeRequest{Image: imagegen.Image{Data: pngFixture(t)}})
+	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 != "mediautils" {
+		t.Errorf("apiErr = %+v", apiErr)
+	}
+}
+
+func TestRestoreFaces(t *testing.T) {
+	png := pngFixture(t)
+	var gotPath, gotUpscale string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("parse multipart: %v", err)
+		}
+		gotUpscale = r.FormValue("upscale")
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(png)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	fr, err := p.FaceRestoreModel("mediautils")
+	if err != nil {
+		t.Fatalf("FaceRestoreModel: %v", err)
+	}
+	res, err := fr.RestoreFaces(context.Background(),
+		imagegen.FaceRestoreRequest{Image: imagegen.Image{Data: png}},
+		imagegen.WithFaceRestoreUpscale(2))
+	if err != nil {
+		t.Fatalf("RestoreFaces: %v", err)
+	}
+	if gotPath != "/upstream/mediautils/v1/restore_faces" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotUpscale != "2" {
+		t.Errorf("upscale = %q", gotUpscale)
+	}
+	if len(res.Images) != 1 {
+		t.Fatalf("images = %+v", res.Images)
+	}
+}
+
+func TestRestoreFacesOmitsDefaultUpscale(t *testing.T) {
+	png := pngFixture(t)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Errorf("parse multipart: %v", err)
+		}
+		if _, ok := r.MultipartForm.Value["upscale"]; ok {
+			t.Error("upscale field sent for default; want omitted")
+		}
+		w.Header().Set("Content-Type", "image/png")
+		_, _ = w.Write(png)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	fr, _ := p.FaceRestoreModel("mediautils")
+	if _, err := fr.RestoreFaces(context.Background(),
+		imagegen.FaceRestoreRequest{Image: imagegen.Image{Data: png}}); err != nil {
+		t.Fatalf("RestoreFaces: %v", err)
+	}
+}
+
+func TestRestoreFacesRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	fr, _ := p.FaceRestoreModel("mediautils")
+	if _, err := fr.RestoreFaces(context.Background(), imagegen.FaceRestoreRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no image: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := fr.RestoreFaces(context.Background(),
+		imagegen.FaceRestoreRequest{Image: imagegen.Image{Data: []byte{1}}, Upscale: 3}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("upscale 3: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestRestoreFacesRejectsNonImageResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header()["Content-Type"] = nil // NO Content-Type at all
+		_, _ = w.Write([]byte("502 bad gateway"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	fr, _ := p.FaceRestoreModel("mediautils")
+	_, err := fr.RestoreFaces(context.Background(),
+		imagegen.FaceRestoreRequest{Image: imagegen.Image{Data: pngFixture(t)}})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for headerless non-image body", err)
+	}
+}

provider/llamaswap/sfx.go

@@ -0,0 +1,100 @@
+// sfx.go implements a second musicgen.Model factory against a Stable Audio
+// Open shim (sfxgen) reached through llama-swap's /upstream passthrough
+// (ADR-0024):
+//
+//	POST /upstream/<id>/v1/sfx   JSON {prompt, seconds?, steps?, cfg_scale?, seed?}
+//
+// Unlike the ACE-Step music path (music.go), /v1/sfx is SYNCHRONOUS: the
+// response body is the finished WAV clip — no job queue, no polling. The
+// surface reuses the musicgen types (a sound effect is a short audio clip
+// from a text prompt); only the provider method differs.
+package llamaswap
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/musicgen"
+)
+
+// SFXModel returns a musicgen.Model bound to a sync sound-effect backend.
+// The id selects which upstream llama-swap loads.
+func (p *Provider) SFXModel(id string, opts ...musicgen.ModelOption) (musicgen.Model, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = musicgen.ApplyModelOptions(opts)
+	return &sfxModel{p: p, id: id}, nil
+}
+
+type sfxModel struct {
+	p  *Provider
+	id string
+}
+
+// sfxRequest is the sfxgen shim's /v1/sfx shape. Optional fields stay off
+// the wire so the model's own defaults apply.
+type sfxRequest struct {
+	Prompt   string   `json:"prompt"`
+	Seconds  int      `json:"seconds,omitempty"`
+	Steps    *int     `json:"steps,omitempty"`
+	CFGScale *float64 `json:"cfg_scale,omitempty"`
+	Seed     *int64   `json:"seed,omitempty"`
+}
+
+// Generate implements musicgen.Model. The clip-length ceiling (~11s for
+// Stable Audio Open Small) is the backend's to enforce, not this client's.
+func (m *sfxModel) 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: sfx generation requires a prompt", llm.ErrUnsupported)
+	}
+	if req.Lyrics != "" {
+		return nil, fmt.Errorf("%w: sfx generation does not support lyrics", 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)
+	}
+	// The endpoint emits WAV only; a caller asking for another container
+	// would silently get mislabelled bytes — reject instead.
+	if f := strings.ToLower(strings.TrimSpace(req.Format)); f != "" && f != "wav" {
+		return nil, fmt.Errorf("%w: sfx output is wav only, got format %q", llm.ErrUnsupported, req.Format)
+	}
+
+	path, err := upstreamPath(m.id, "/v1/sfx")
+	if err != nil {
+		return nil, err
+	}
+	wire := sfxRequest{
+		Prompt:   req.Prompt,
+		Seconds:  req.DurationSeconds,
+		Steps:    req.Steps,
+		CFGScale: req.CFGScale,
+		Seed:     req.Seed,
+	}
+	encoded, err := json.Marshal(wire)
+	if err != nil {
+		return nil, fmt.Errorf("llama-swap: encode sfx request: %w", err)
+	}
+	raw, contentType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, "application/json", bytes.NewReader(encoded), maxAudioResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "sfx response contained no audio"}
+	}
+	mimeType := audioResultMIME(contentType, raw)
+	if mimeType == "" {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("sfx response is not audio (Content-Type %q): %s", contentType, truncateForError(raw))}
+	}
+	return &musicgen.Result{Audio: musicgen.Audio{Data: raw, MIME: mimeType}}, nil
+}

provider/llamaswap/stems.go

@@ -0,0 +1,165 @@
+// stems.go implements audio.StemSeparationProvider against a Demucs shim
+// (audioutils) reached through llama-swap's /upstream passthrough (ADR-0024):
+//
+//	POST /upstream/<id>/v1/stems   multipart file[,model,two_stems,format]
+//
+// The response is a ZIP of the separated stems — one entry per stem, entry
+// name = stem name, extension = container. Zip transport keeps a 4-stem WAV
+// result (hundreds of MB decoded) off the JSON-of-base64 path entirely.
+package llamaswap
+
+import (
+	"archive/zip"
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"path"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// maxStemsResponseBytes caps the /v1/stems zip body: four WAV stems of a
+// long song legitimately pass the 64MB JSON cap. Bounded so a buggy
+// upstream can't allocate without limit.
+const maxStemsResponseBytes = 512 << 20
+
+// maxStemEntryBytes caps ONE decompressed zip entry — the zip-bomb guard.
+// A single WAV stem of even a very long song stays far under this.
+const maxStemEntryBytes = 256 << 20
+
+// maxStemEntries caps how many stem entries are unpacked: Demucs emits at
+// most six, so anything past a small multiple of that is a hostile or
+// broken archive, not a result.
+const maxStemEntries = 16
+
+// maxStemsTotalBytes caps the AGGREGATE decompressed size across entries —
+// the per-entry bound alone would still let a many-entry bomb multiply up.
+const maxStemsTotalBytes = 1 << 30
+
+// StemSeparatorModel implements audio.StemSeparationProvider. The id selects
+// which upstream llama-swap loads.
+func (p *Provider) StemSeparatorModel(id string, opts ...audio.StemSeparatorModelOption) (audio.StemSeparator, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = audio.ApplyStemSeparatorModelOptions(opts)
+	return &stemSeparatorModel{p: p, id: id}, nil
+}
+
+type stemSeparatorModel struct {
+	p  *Provider
+	id string
+}
+
+// SeparateStems implements audio.StemSeparator.
+func (m *stemSeparatorModel) SeparateStems(ctx context.Context, req audio.StemSeparationRequest, opts ...audio.StemSeparationOption) (*audio.StemSeparationResult, error) {
+	req = req.Apply(opts...)
+	if len(req.Audio) == 0 {
+		return nil, fmt.Errorf("%w: stem separation requires audio bytes", llm.ErrUnsupported)
+	}
+	mode := strings.ToLower(strings.TrimSpace(req.Mode))
+	if mode != "" && mode != "two" && mode != "four" {
+		return nil, fmt.Errorf("%w: stem mode must be \"two\" or \"four\", got %q", llm.ErrUnsupported, req.Mode)
+	}
+	format := strings.ToLower(strings.TrimSpace(req.Format))
+	if format != "" && format != "mp3" && format != "wav" {
+		return nil, fmt.Errorf("%w: stem format must be \"mp3\" or \"wav\", got %q", llm.ErrUnsupported, req.Format)
+	}
+	upPath, err := upstreamPath(m.id, "/v1/stems")
+	if err != nil {
+		return nil, err
+	}
+	// Demucs' two-stem mode is "isolate one source vs the rest"; vocals is
+	// the split this surface promises ("two" = vocals + accompaniment).
+	twoStems := ""
+	if mode == "two" {
+		twoStems = "vocals"
+	}
+	body, contentType, err := buildMultipart("build stems form",
+		filePart{field: "file", filename: transcriptionFilename(req.Filename, req.MIME), data: req.Audio},
+		[]formField{
+			{"model", req.Model, false},
+			{"two_stems", twoStems, false},
+			{"format", format, false},
+		})
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, upPath, m.id, contentType, body, maxStemsResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "stems response contained no data"}
+	}
+	zr, err := zip.NewReader(bytes.NewReader(raw), int64(len(raw)))
+	if err != nil {
+		// A non-zip 2xx body is a misconfigured upstream (an HTML error page
+		// behind a proxy, a JSON soft error) — fail loud, quoting a slice.
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: fmt.Sprintf("stems response is not a zip (Content-Type %q): %s", respType, truncateForError(raw))}
+	}
+	res := &audio.StemSeparationResult{}
+	var totalBytes int64
+	for _, f := range zr.File {
+		if f.FileInfo().IsDir() {
+			continue
+		}
+		if len(res.Stems) >= maxStemEntries {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("stems zip holds more than %d entries", maxStemEntries)}
+		}
+		data, err := readZipEntry(f)
+		if err != nil {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("stems zip entry %q: %v", f.Name, err)}
+		}
+		totalBytes += int64(len(data))
+		if totalBytes > maxStemsTotalBytes {
+			return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+				Message: fmt.Sprintf("stems zip decompresses past %d bytes", int64(maxStemsTotalBytes))}
+		}
+		// Entry name → stem name; extension → MIME. Entries may sit in a
+		// per-model directory ("htdemucs/vocals.mp3"), so use the base name.
+		base := path.Base(f.Name)
+		ext := path.Ext(base)
+		res.Stems = append(res.Stems, audio.Stem{
+			Name:  strings.TrimSuffix(base, ext),
+			Audio: data,
+			MIME:  stemMIME(ext),
+		})
+	}
+	if len(res.Stems) == 0 {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "stems zip contained no stems"}
+	}
+	return res, nil
+}
+
+// readZipEntry decompresses one entry, bounded by maxStemEntryBytes so a
+// zip bomb can't allocate without limit.
+func readZipEntry(f *zip.File) ([]byte, error) {
+	rc, err := f.Open()
+	if err != nil {
+		return nil, err
+	}
+	defer rc.Close()
+	data, err := io.ReadAll(io.LimitReader(rc, maxStemEntryBytes+1))
+	if err != nil {
+		return nil, err
+	}
+	if int64(len(data)) > maxStemEntryBytes {
+		return nil, fmt.Errorf("decompressed entry exceeds %d bytes", int64(maxStemEntryBytes))
+	}
+	return data, nil
+}
+
+// stemMIME maps a stem file extension to its MIME type, reusing speechMIME's
+// format table ("" and unknown extensions land on the mp3 default — Demucs'
+// own default container).
+func stemMIME(ext string) string {
+	return speechMIME("", strings.TrimPrefix(strings.ToLower(ext), "."))
+}

provider/llamaswap/stems_sfx_enhance_test.go

@@ -0,0 +1,355 @@
+package llamaswap
+
+import (
+	"archive/zip"
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/audio"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/musicgen"
+)
+
+// stemsZipFixture builds a Demucs-style stems zip: entries under a per-model
+// directory, entry name = stem name, extension = container.
+func stemsZipFixture(t *testing.T, entries map[string]string) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	zw := zip.NewWriter(&buf)
+	for name, data := range entries {
+		w, err := zw.Create(name)
+		if err != nil {
+			t.Fatalf("zip create: %v", err)
+		}
+		if _, err := w.Write([]byte(data)); err != nil {
+			t.Fatalf("zip write: %v", err)
+		}
+	}
+	if err := zw.Close(); err != nil {
+		t.Fatalf("zip close: %v", err)
+	}
+	return buf.Bytes()
+}
+
+func TestSeparateStems(t *testing.T) {
+	zipBody := stemsZipFixture(t, map[string]string{
+		"htdemucs/vocals.mp3":    "VOX",
+		"htdemucs/no_vocals.mp3": "ACC",
+	})
+	var gotPath, gotTwoStems, gotModel, gotFormat, gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		gotTwoStems = r.FormValue("two_stems")
+		gotModel = r.FormValue("model")
+		gotFormat = r.FormValue("format")
+		if _, hdr, err := r.FormFile("file"); err == nil {
+			gotFilename = hdr.Filename
+		} else {
+			t.Errorf("file part: %v", err)
+		}
+		w.Header().Set("Content-Type", "application/zip")
+		_, _ = w.Write(zipBody)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ss, err := p.StemSeparatorModel("audioutils")
+	if err != nil {
+		t.Fatalf("StemSeparatorModel: %v", err)
+	}
+	res, err := ss.SeparateStems(context.Background(),
+		audio.StemSeparationRequest{Audio: []byte("SONG"), MIME: "audio/mpeg"},
+		audio.WithStemMode("two"), audio.WithStemModel("htdemucs_ft"), audio.WithStemFormat("mp3"))
+	if err != nil {
+		t.Fatalf("SeparateStems: %v", err)
+	}
+	if gotPath != "/upstream/audioutils/v1/stems" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotTwoStems != "vocals" || gotModel != "htdemucs_ft" || gotFormat != "mp3" || gotFilename != "audio.mp3" {
+		t.Errorf("two_stems/model/format/filename = %q/%q/%q/%q", gotTwoStems, gotModel, gotFormat, gotFilename)
+	}
+	if len(res.Stems) != 2 {
+		t.Fatalf("stems = %+v", res.Stems)
+	}
+	byName := map[string]audio.Stem{}
+	for _, s := range res.Stems {
+		byName[s.Name] = s
+	}
+	if v := byName["vocals"]; string(v.Audio) != "VOX" || v.MIME != "audio/mpeg" {
+		t.Errorf("vocals = %+v", v)
+	}
+	if a := byName["no_vocals"]; string(a.Audio) != "ACC" || a.MIME != "audio/mpeg" {
+		t.Errorf("no_vocals = %+v", a)
+	}
+}
+
+func TestSeparateStemsOmitsUnsetFields(t *testing.T) {
+	zipBody := stemsZipFixture(t, map[string]string{"vocals.wav": "V"})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		for _, k := range []string{"two_stems", "model", "format"} {
+			if v, ok := r.MultipartForm.Value[k]; ok {
+				t.Errorf("unset request sent %q = %v, want omitted", k, v)
+			}
+		}
+		_, _ = w.Write(zipBody)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ss, _ := p.StemSeparatorModel("audioutils")
+	res, err := ss.SeparateStems(context.Background(),
+		audio.StemSeparationRequest{Audio: []byte("SONG")})
+	if err != nil {
+		t.Fatalf("SeparateStems: %v", err)
+	}
+	// Top-level entry, wav extension.
+	if len(res.Stems) != 1 || res.Stems[0].Name != "vocals" || res.Stems[0].MIME != "audio/wav" {
+		t.Errorf("stems = %+v", res.Stems)
+	}
+}
+
+func TestSeparateStemsRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	ss, _ := p.StemSeparatorModel("audioutils")
+	if _, err := ss.SeparateStems(context.Background(), audio.StemSeparationRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no audio: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ss.SeparateStems(context.Background(),
+		audio.StemSeparationRequest{Audio: []byte{1}, Mode: "three"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("mode three: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ss.SeparateStems(context.Background(),
+		audio.StemSeparationRequest{Audio: []byte{1}, Format: "flac"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("format flac: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestSeparateStemsRejectsNonZip(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html>proxy error page</html>"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ss, _ := p.StemSeparatorModel("audioutils")
+	_, err := ss.SeparateStems(context.Background(), audio.StemSeparationRequest{Audio: []byte("SONG")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-zip body", err)
+	}
+}
+
+func TestSeparateStemsRejectsEmptyZip(t *testing.T) {
+	zipBody := stemsZipFixture(t, map[string]string{})
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/zip")
+		_, _ = w.Write(zipBody)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ss, _ := p.StemSeparatorModel("audioutils")
+	_, err := ss.SeparateStems(context.Background(), audio.StemSeparationRequest{Audio: []byte("SONG")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for stem-less zip", err)
+	}
+}
+
+func TestSeparateStemsRejectsTooManyEntries(t *testing.T) {
+	entries := map[string]string{}
+	for i := 0; i <= maxStemEntries; i++ {
+		entries[fmt.Sprintf("stem%02d.wav", i)] = "X"
+	}
+	zipBody := stemsZipFixture(t, entries)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/zip")
+		_, _ = w.Write(zipBody)
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ss, _ := p.StemSeparatorModel("audioutils")
+	_, err := ss.SeparateStems(context.Background(), audio.StemSeparationRequest{Audio: []byte("SONG")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) || !strings.Contains(apiErr.Message, "entries") {
+		t.Fatalf("err = %v, want APIError for over-long stems zip", err)
+	}
+}
+
+// wavFixture is a minimal RIFF/WAVE header so http.DetectContentType sniffs
+// audio/wave.
+func wavFixture() []byte {
+	return []byte("RIFF\x24\x00\x00\x00WAVEfmt ")
+}
+
+func TestSFXGenerate(t *testing.T) {
+	var gotPath string
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		w.Header().Set("Content-Type", "audio/wav")
+		_, _ = w.Write(wavFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, err := p.SFXModel("sfxgen-stableaudio")
+	if err != nil {
+		t.Fatalf("SFXModel: %v", err)
+	}
+	res, err := sm.Generate(context.Background(),
+		musicgen.Request{Prompt: "glass shattering", DurationSeconds: 8},
+		musicgen.WithSteps(50), musicgen.WithCFGScale(7), musicgen.WithSeed(42))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if gotPath != "/upstream/sfxgen-stableaudio/v1/sfx" {
+		t.Errorf("path = %q", gotPath)
+	}
+	want := map[string]any{"prompt": "glass shattering", "seconds": 8.0, "steps": 50.0, "cfg_scale": 7.0, "seed": 42.0}
+	for k, w := range want {
+		if gotBody[k] != w {
+			t.Errorf("%s = %v, want %v", k, gotBody[k], w)
+		}
+	}
+	if res.Audio.MIME != "audio/wav" || len(res.Audio.Data) == 0 {
+		t.Errorf("audio = %q/%d bytes", res.Audio.MIME, len(res.Audio.Data))
+	}
+}
+
+func TestSFXOmitsDefaults(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)
+		// No Content-Type: the RIFF sniff must still label the clip.
+		w.Header()["Content-Type"] = nil
+		_, _ = w.Write(wavFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SFXModel("sfxgen-stableaudio")
+	res, err := sm.Generate(context.Background(), musicgen.Request{Prompt: "boom"})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	for _, k := range []string{"seconds", "steps", "cfg_scale", "seed"} {
+		if v, ok := gotBody[k]; ok {
+			t.Errorf("unset request sent %q = %v, want omitted", k, v)
+		}
+	}
+	if res.Audio.MIME != "audio/wav" {
+		t.Errorf("MIME = %q, want sniffed audio/wav", res.Audio.MIME)
+	}
+}
+
+func TestSFXRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	sm, _ := p.SFXModel("sfxgen-stableaudio")
+	if _, err := sm.Generate(context.Background(), musicgen.Request{Prompt: "  "}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("empty prompt: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := sm.Generate(context.Background(),
+		musicgen.Request{Prompt: "boom", Lyrics: "la la"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("lyrics: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := sm.Generate(context.Background(),
+		musicgen.Request{Prompt: "boom", Format: "mp3"}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("format mp3: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestSFXRejectsNonAudioResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"detail":"queue full"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	sm, _ := p.SFXModel("sfxgen-stableaudio")
+	_, err := sm.Generate(context.Background(), musicgen.Request{Prompt: "boom"})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-audio body", err)
+	}
+}
+
+func TestEnhance(t *testing.T) {
+	var gotPath, gotFilename string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		if err := r.ParseMultipartForm(1 << 20); err != nil {
+			t.Fatalf("parse multipart: %v", err)
+		}
+		if _, hdr, err := r.FormFile("file"); err == nil {
+			gotFilename = hdr.Filename
+		} else {
+			t.Errorf("file part: %v", err)
+		}
+		w.Header().Set("Content-Type", "audio/wav")
+		_, _ = w.Write(wavFixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	en, err := p.SpeechEnhancerModel("audioutils")
+	if err != nil {
+		t.Fatalf("SpeechEnhancerModel: %v", err)
+	}
+	res, err := en.Enhance(context.Background(),
+		audio.EnhancementRequest{Audio: []byte("NOISY"), MIME: "audio/ogg"})
+	if err != nil {
+		t.Fatalf("Enhance: %v", err)
+	}
+	if gotPath != "/upstream/audioutils/v1/enhance" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotFilename != "audio.ogg" {
+		t.Errorf("filename = %q", gotFilename)
+	}
+	if res.MIME != "audio/wav" || len(res.Audio) == 0 {
+		t.Errorf("result = %q/%d bytes", res.MIME, len(res.Audio))
+	}
+}
+
+func TestEnhanceRejectsEmptyAudio(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	en, _ := p.SpeechEnhancerModel("audioutils")
+	if _, err := en.Enhance(context.Background(), audio.EnhancementRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestEnhanceRejectsNonAudioResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/html")
+		_, _ = w.Write([]byte("<html>oops</html>"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	en, _ := p.SpeechEnhancerModel("audioutils")
+	_, err := en.Enhance(context.Background(), audio.EnhancementRequest{Audio: []byte("NOISY")})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-audio body", err)
+	}
+}

provider/llamaswap/upstream.go

@@ -16,12 +16,14 @@ import (
 //
 // Why reject rather than escape: same rationale as Unload — model ids
 // legitimately contain ":" but never path-structure characters, and escaping
-// would mask a config error instead of surfacing it.
+// would mask a config error instead of surfacing it. '%' is rejected too:
+// ids never legitimately carry percent-escapes, and %2F/%2E%2E would decode
+// back into path structure on the server side.
 func upstreamPath(model, rest string) (string, error) {
 	if strings.TrimSpace(model) == "" {
 		return "", fmt.Errorf("llama-swap: upstream call requires a model id")
 	}
-	if strings.ContainsAny(model, "/?#") || strings.Contains(model, "..") {
+	if strings.ContainsAny(model, "/?#%") || strings.Contains(model, "..") {
 		return "", fmt.Errorf("llama-swap: invalid model id %q for upstream call (contains a path separator)", model)
 	}
 	if !strings.HasPrefix(rest, "/") {
@@ -49,8 +51,20 @@ type filePart struct {
 // writeFormFields). wrap labels errors. Returns the body and its content
 // type.
 func buildMultipart(wrap string, file filePart, fields []formField) (*bytes.Buffer, string, error) {
+	return buildMultipartFiles(wrap, []filePart{file}, fields)
+}
+
+// buildMultipartFiles is buildMultipart for endpoints taking SEVERAL files
+// (face swap sends a target and a source). Files are written in the given
+// order, then the fields. One writer loop serves both so the two cannot drift
+// in how they escape names or terminate the body.
+func buildMultipartFiles(wrap string, files []filePart, fields []formField) (*bytes.Buffer, string, error) {
+	if len(files) == 0 {
+		return nil, "", fmt.Errorf("llama-swap: %s: no file parts", wrap)
+	}
 	var buf bytes.Buffer
 	w := multipart.NewWriter(&buf)
+	for _, file := range files {
 		fw, err := w.CreateFormFile(file.field, file.filename)
 		if err != nil {
 			return nil, "", fmt.Errorf("llama-swap: %s: %w", wrap, err)
@@ -58,6 +72,7 @@ func buildMultipart(wrap string, file filePart, fields []formField) (*bytes.Buff
 		if _, err := fw.Write(file.data); err != nil {
 			return nil, "", fmt.Errorf("llama-swap: %s: %w", wrap, err)
 		}
+	}
 	if err := writeFormFields(w, wrap, fields); err != nil {
 		return nil, "", err
 	}

provider/llamaswap/video.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"fmt"
-	"mime"
 	"mime/multipart"
 	"net/http"
 	"strconv"
@@ -100,22 +99,7 @@ func (m *videoModel) Generate(ctx context.Context, req videogen.Request, opts ..
 	if err != nil {
 		return nil, err
 	}
-	if len(videoBytes) == 0 {
-		return nil, &llm.APIError{Provider: m.p.name, Model: m.id, Message: "video response contained no video"}
-	}
-	mimeType := videoMIME(contentType, videoBytes)
-	if mimeType == "" {
-		// A 2xx body that is neither declared nor sniffable as video is a
-		// misconfigured upstream (a JSON job envelope, an HTML error page
-		// behind a proxy) — fail loud rather than hand back garbage as a
-		// playable clip.
-		return nil, &llm.APIError{
-			Provider: m.p.name,
-			Model:    m.id,
-			Message:  fmt.Sprintf("video response is not a video (content-type %q)", contentType),
-		}
-	}
-	return &videogen.Result{Video: videogen.Video{Data: videoBytes, MIME: mimeType}}, nil
+	return singleVideoResult(m.p.name, m.id, "video", videoBytes, contentType)
 }
 
 // videoMIME resolves the result MIME type: the response Content-Type when it
@@ -132,22 +116,29 @@ func videoMIME(contentType string, data []byte) string {
 	return ""
 }
 
+// singleVideoResult wraps one raw video body into a videogen.Result,
+// requiring positive evidence of video-ness (declared video/* Content-Type
+// or sniffed mp4/webm magic) so a 2xx body that is anything else — a JSON
+// job envelope, an HTML error page behind a proxy — fails loud instead of
+// coming back as "the clip". The video sibling of singleImageResult, shared
+// by every surface whose response body IS the encoded clip.
+func singleVideoResult(provider, model, verb string, raw []byte, contentType string) (*videogen.Result, error) {
+	if len(raw) == 0 {
+		return nil, &llm.APIError{Provider: provider, Model: model, Message: verb + " response contained no video"}
+	}
+	mimeType := videoMIME(contentType, raw)
+	if mimeType == "" {
+		return nil, &llm.APIError{Provider: provider, Model: model,
+			Message: fmt.Sprintf("%s response is not a video (Content-Type %q)", verb, contentType)}
+	}
+	return &videogen.Result{Video: videogen.Video{Data: raw, MIME: mimeType}}, nil
+}
+
 // initImageFilename picks the multipart filename hint for the conditioning
 // frame from its MIME subtype. The name is provider-chosen (never
 // caller-supplied), so no sanitization is needed.
 func initImageFilename(mimeType string) string {
-	mt := strings.ToLower(strings.TrimSpace(mimeType))
-	if parsed, _, err := mime.ParseMediaType(mt); err == nil {
-		mt = parsed
-	}
-	switch mt {
-	case "image/jpeg", "image/jpg":
-		return "frame.jpg"
-	case "image/webp":
-		return "frame.webp"
-	default: // unknown MIME — PNG is the safe hint
-		return "frame.png"
-	}
+	return imageFilename(mimeType, "frame")
 }
 
 // formatInt renders an optional int pointer for a form field; nil = "" (omit).

provider/llamaswap/videochain.go

@@ -0,0 +1,217 @@
+// videochain.go implements videogen.ChainerProvider against the videoutils
+// chain orchestrator reached through llama-swap's /upstream passthrough
+// (ADR-0025):
+//
+//	POST /upstream/<id>/v1/video/chain          JSON submit -> {job_id}
+//	GET  /upstream/<id>/v1/jobs/{id}            -> {status,segment,total,segments}
+//	GET  /upstream/<id>/v1/jobs/{id}/result     -> encoded clip
+//	GET  /upstream/<id>/v1/jobs/{id}/segments/{n} -> encoded clip
+//
+// Unlike the ACE-Step music path this client does NOT hide the job queue
+// behind a blocking call: a chain runs through multiple GPU swaps for many
+// minutes, and the caller (mort's long-video tool) owns the poll loop so it
+// can deliver PARTIAL results — completed segments survive a mid-chain
+// failure and stay fetchable via ChainSegmentResult.
+package llamaswap
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/videogen"
+)
+
+// ChainerModel implements videogen.ChainerProvider. The id selects which
+// upstream llama-swap loads (the videoutils orchestrator, which in turn
+// drives the generation model through llama-swap itself).
+func (p *Provider) ChainerModel(id string, opts ...videogen.ChainerModelOption) (videogen.Chainer, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = videogen.ApplyChainerModelOptions(opts)
+	return &chainerModel{p: p, id: id}, nil
+}
+
+type chainerModel struct {
+	p  *Provider
+	id string
+}
+
+// chainSubmitRequest is the videoutils POST /v1/video/chain shape. The init
+// image rides as base64 in the JSON body (`init_image_b64`) — the submit is
+// JSON, not multipart, per the pinned host contract.
+type chainSubmitRequest struct {
+	Segments     []chainSegmentWire `json:"segments"`
+	InitImageB64 string             `json:"init_image_b64,omitempty"`
+	SmoothJoins  bool               `json:"smooth_joins,omitempty"`
+	Size         string             `json:"size,omitempty"`
+}
+
+type chainSegmentWire struct {
+	Prompt  string  `json:"prompt"`
+	Seconds float64 `json:"seconds,omitempty"`
+}
+
+// SubmitChain implements videogen.Chainer.
+func (m *chainerModel) SubmitChain(ctx context.Context, req videogen.ChainRequest) (string, error) {
+	if len(req.Segments) == 0 {
+		return "", fmt.Errorf("%w: video chain requires at least one segment", llm.ErrUnsupported)
+	}
+	wire := chainSubmitRequest{
+		SmoothJoins: req.SmoothJoins,
+		Size:        strings.TrimSpace(req.Size),
+	}
+	for i, seg := range req.Segments {
+		if strings.TrimSpace(seg.Prompt) == "" {
+			return "", fmt.Errorf("%w: video chain segment %d requires a prompt", llm.ErrUnsupported, i)
+		}
+		// NaN/±Inf would otherwise surface as an obscure json.Marshal error
+		// (NaN fails every comparison; +Inf passes the >= 0 check).
+		if seg.Seconds < 0 || math.IsNaN(seg.Seconds) || math.IsInf(seg.Seconds, 0) {
+			return "", fmt.Errorf("%w: video chain segment %d seconds must be a finite value >= 0, got %g", llm.ErrUnsupported, i, seg.Seconds)
+		}
+		wire.Segments = append(wire.Segments, chainSegmentWire{Prompt: seg.Prompt, Seconds: seg.Seconds})
+	}
+	if len(req.InitImage) > 0 {
+		wire.InitImageB64 = base64.StdEncoding.EncodeToString(req.InitImage)
+	}
+	path, err := upstreamPath(m.id, "/v1/video/chain")
+	if err != nil {
+		return "", err
+	}
+	// Tolerant envelope: {"job_id": ...} per the contract, with data-wrapped
+	// and bare-id fallbacks (musicgen release_task precedent).
+	var resp struct {
+		JobID string `json:"job_id"`
+		ID    string `json:"id"`
+		Data  struct {
+			JobID string `json:"job_id"`
+			ID    string `json:"id"`
+		} `json:"data"`
+	}
+	if err := m.p.doJSON(ctx, http.MethodPost, path, m.id, &wire, &resp); err != nil {
+		return "", err
+	}
+	for _, id := range []string{resp.JobID, resp.Data.JobID, resp.ID, resp.Data.ID} {
+		if id != "" {
+			return id, nil
+		}
+	}
+	return "", &llm.APIError{Provider: m.p.name, Model: m.id, Message: "video chain submit returned no job_id"}
+}
+
+// chainJobResponse is the GET /v1/jobs/{id} shape. `segments` entries are
+// tolerated as bare strings or objects keyed by id/segment_id.
+type chainJobResponse struct {
+	Status   string            `json:"status"`
+	Segment  int               `json:"segment"`
+	Total    int               `json:"total"`
+	Segments []json.RawMessage `json:"segments"`
+}
+
+// ChainStatus implements videogen.Chainer.
+func (m *chainerModel) ChainStatus(ctx context.Context, jobID string) (*videogen.ChainJob, error) {
+	path, err := m.jobPath(jobID, "")
+	if err != nil {
+		return nil, err
+	}
+	var raw json.RawMessage
+	if err := m.p.doJSON(ctx, http.MethodGet, path, m.id, nil, &raw); err != nil {
+		return nil, err
+	}
+	var out chainJobResponse
+	if err := json.Unmarshal(raw, &out); err != nil {
+		return nil, fmt.Errorf("llama-swap: decode chain job status: %w", err)
+	}
+	job := &videogen.ChainJob{
+		Status:  out.Status,
+		Segment: out.Segment,
+		Total:   out.Total,
+		Raw:     raw,
+	}
+	// Entries that carry no usable id — JSON null (which unmarshals into a
+	// string as a no-op, leaving ""), an empty string, or an object with
+	// neither key — are SKIPPED, never appended as "": SegmentIDs promises
+	// fetchable artifacts, and the full payload stays in Raw for callers
+	// that want the unfiltered list.
+	for _, entry := range out.Segments {
+		var s string
+		if json.Unmarshal(entry, &s) == nil {
+			if s != "" {
+				job.SegmentIDs = append(job.SegmentIDs, s)
+			}
+			continue
+		}
+		var obj struct {
+			ID        string `json:"id"`
+			SegmentID string `json:"segment_id"`
+		}
+		if json.Unmarshal(entry, &obj) == nil {
+			switch {
+			case obj.ID != "":
+				job.SegmentIDs = append(job.SegmentIDs, obj.ID)
+			case obj.SegmentID != "":
+				job.SegmentIDs = append(job.SegmentIDs, obj.SegmentID)
+			}
+		}
+	}
+	if job.Status == "" {
+		return nil, &llm.APIError{Provider: m.p.name, Model: m.id,
+			Message: "chain job status payload carried no status: " + truncateForError(raw)}
+	}
+	return job, nil
+}
+
+// ChainResult implements videogen.Chainer.
+func (m *chainerModel) ChainResult(ctx context.Context, jobID string) (*videogen.Result, error) {
+	path, err := m.jobPath(jobID, "/result")
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodGet, path, m.id, "", nil, maxVideoResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleVideoResult(m.p.name, m.id, "video chain result", raw, respType)
+}
+
+// ChainSegmentResult implements videogen.Chainer.
+func (m *chainerModel) ChainSegmentResult(ctx context.Context, jobID string, n int) (*videogen.Result, error) {
+	if n < 0 {
+		return nil, fmt.Errorf("%w: chain segment index must be >= 0, got %d", llm.ErrUnsupported, n)
+	}
+	path, err := m.jobPath(jobID, "/segments/"+strconv.Itoa(n))
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodGet, path, m.id, "", nil, maxVideoResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleVideoResult(m.p.name, m.id, "video chain segment", raw, respType)
+}
+
+// jobPath builds /upstream/<model>/v1/jobs/<jobID><suffix>, refusing job ids
+// that carry path structure. The id is SERVER-SUPPLIED (echoed back from
+// SubmitChain), so like upstreamPath's rest-component checks this rejects
+// rather than escapes — a hostile/buggy upstream must not be able to steer
+// the follow-up request at another proxy endpoint.
+func (m *chainerModel) jobPath(jobID, suffix string) (string, error) {
+	if strings.TrimSpace(jobID) == "" {
+		return "", fmt.Errorf("llama-swap: chain job call requires a job id")
+	}
+	// '%' is rejected alongside the literal path characters: job ids never
+	// legitimately carry percent-escapes, and %2F/%2E%2E would decode back
+	// into path structure server-side.
+	if strings.ContainsAny(jobID, "/?#%") || strings.Contains(jobID, "..") {
+		return "", fmt.Errorf("llama-swap: invalid chain job id %q (contains path structure)", jobID)
+	}
+	return upstreamPath(m.id, "/v1/jobs/"+jobID+suffix)
+}

provider/llamaswap/videochain_test.go

@@ -0,0 +1,322 @@
+package llamaswap
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"math"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/videogen"
+)
+
+func TestSubmitChain(t *testing.T) {
+	var gotPath string
+	var gotBody map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		_ = json.NewDecoder(r.Body).Decode(&gotBody)
+		_, _ = w.Write([]byte(`{"job_id":"job-123"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, err := p.ChainerModel("videoutils")
+	if err != nil {
+		t.Fatalf("ChainerModel: %v", err)
+	}
+	jobID, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+		Segments: []videogen.ChainSegment{
+			{Prompt: "a cat walks in", Seconds: 5},
+			{Prompt: "the cat sits down", Seconds: 5},
+		},
+		InitImage:   []byte("IMG"),
+		SmoothJoins: true,
+		Size:        "1280x704",
+	})
+	if err != nil {
+		t.Fatalf("SubmitChain: %v", err)
+	}
+	if jobID != "job-123" {
+		t.Errorf("jobID = %q", jobID)
+	}
+	if gotPath != "/upstream/videoutils/v1/video/chain" {
+		t.Errorf("path = %q", gotPath)
+	}
+	segments, _ := gotBody["segments"].([]any)
+	if len(segments) != 2 {
+		t.Fatalf("segments = %v", gotBody["segments"])
+	}
+	first, _ := segments[0].(map[string]any)
+	if first["prompt"] != "a cat walks in" || first["seconds"] != 5.0 {
+		t.Errorf("segment[0] = %v", first)
+	}
+	if gotBody["init_image_b64"] != base64.StdEncoding.EncodeToString([]byte("IMG")) {
+		t.Errorf("init_image_b64 = %v", gotBody["init_image_b64"])
+	}
+	if gotBody["smooth_joins"] != true || gotBody["size"] != "1280x704" {
+		t.Errorf("smooth_joins/size = %v/%v", gotBody["smooth_joins"], gotBody["size"])
+	}
+}
+
+func TestSubmitChainOmitsUnsetFields(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(`{"data":{"job_id":"job-9"}}`)) // data-wrapped envelope
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	jobID, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+		Segments: []videogen.ChainSegment{{Prompt: "a dog"}},
+	})
+	if err != nil {
+		t.Fatalf("SubmitChain: %v", err)
+	}
+	if jobID != "job-9" {
+		t.Errorf("jobID = %q, want data-wrapped id", jobID)
+	}
+	for _, k := range []string{"init_image_b64", "smooth_joins", "size"} {
+		if v, ok := gotBody[k]; ok {
+			t.Errorf("unset request sent %q = %v, want omitted", k, v)
+		}
+	}
+	seg, _ := gotBody["segments"].([]any)
+	if first, _ := seg[0].(map[string]any); first == nil {
+		t.Fatalf("segments = %v", gotBody["segments"])
+	} else if _, ok := first["seconds"]; ok {
+		t.Error("zero seconds sent; want omitted")
+	}
+}
+
+func TestSubmitChainRejectsBadArgs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	ch, _ := p.ChainerModel("videoutils")
+	if _, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("no segments: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+		Segments: []videogen.ChainSegment{{Prompt: "  "}},
+	}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("empty prompt: err = %v, want ErrUnsupported", err)
+	}
+	if _, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+		Segments: []videogen.ChainSegment{{Prompt: "x", Seconds: -1}},
+	}); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("negative seconds: err = %v, want ErrUnsupported", err)
+	}
+	for name, bad := range map[string]float64{
+		"NaN":  math.NaN(),
+		"+Inf": math.Inf(1),
+		"-Inf": math.Inf(-1),
+	} {
+		if _, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+			Segments: []videogen.ChainSegment{{Prompt: "x", Seconds: bad}},
+		}); !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("%s seconds: err = %v, want ErrUnsupported", name, err)
+		}
+	}
+}
+
+func TestSubmitChainRejectsMissingJobID(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"ok":true}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	_, err := ch.SubmitChain(context.Background(), videogen.ChainRequest{
+		Segments: []videogen.ChainSegment{{Prompt: "x"}},
+	})
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for missing job_id", err)
+	}
+}
+
+func TestChainStatus(t *testing.T) {
+	var gotPath string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		_, _ = w.Write([]byte(`{"status":"running","segment":2,"total":3,"segments":["seg-0"]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	job, err := ch.ChainStatus(context.Background(), "job-123")
+	if err != nil {
+		t.Fatalf("ChainStatus: %v", err)
+	}
+	if gotPath != "/upstream/videoutils/v1/jobs/job-123" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if job.Status != "running" || job.Segment != 2 || job.Total != 3 {
+		t.Errorf("job = %+v", job)
+	}
+	if !reflect.DeepEqual(job.SegmentIDs, []string{"seg-0"}) {
+		t.Errorf("SegmentIDs = %v", job.SegmentIDs)
+	}
+	if job.Raw == nil {
+		t.Error("Raw = nil, want raw payload")
+	}
+}
+
+func TestChainStatusToleratesObjectSegments(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"status":"done","segment":2,"total":2,"segments":[{"id":"seg-0"},{"segment_id":"seg-1"}]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	job, err := ch.ChainStatus(context.Background(), "job-123")
+	if err != nil {
+		t.Fatalf("ChainStatus: %v", err)
+	}
+	if !reflect.DeepEqual(job.SegmentIDs, []string{"seg-0", "seg-1"}) {
+		t.Errorf("SegmentIDs = %v", job.SegmentIDs)
+	}
+}
+
+func TestChainStatusSkipsIDLessSegments(t *testing.T) {
+	// null, "", an id-less object, and a mistyped entry must all be
+	// skipped — never appended as "" (SegmentIDs promises fetchable
+	// artifacts; the unfiltered list stays in Raw).
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"status":"running","segment":3,"total":5,` +
+			`"segments":[null,"seg-0","",{},{"id":"seg-1"},{"segment_id":""},42]}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	job, err := ch.ChainStatus(context.Background(), "job-123")
+	if err != nil {
+		t.Fatalf("ChainStatus: %v", err)
+	}
+	if !reflect.DeepEqual(job.SegmentIDs, []string{"seg-0", "seg-1"}) {
+		t.Errorf("SegmentIDs = %v, want [seg-0 seg-1]", job.SegmentIDs)
+	}
+}
+
+func TestChainStatusRejectsStatuslessPayload(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"detail":"no such job"}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	_, err := ch.ChainStatus(context.Background(), "job-123")
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for statusless payload", err)
+	}
+}
+
+func TestChainJobPathRejectsHostileIDs(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	ch, _ := p.ChainerModel("videoutils")
+	for _, bad := range []string{"", "a/b", "a?b", "a#b", "..", "a..b", "a%2Fb", "%2e%2e", "a%b"} {
+		if _, err := ch.ChainStatus(context.Background(), bad); err == nil {
+			t.Errorf("ChainStatus(%q) succeeded; want error", bad)
+		}
+		if _, err := ch.ChainResult(context.Background(), bad); err == nil {
+			t.Errorf("ChainResult(%q) succeeded; want error", bad)
+		}
+	}
+}
+
+func TestChainResult(t *testing.T) {
+	var gotPath string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	res, err := ch.ChainResult(context.Background(), "job-123")
+	if err != nil {
+		t.Fatalf("ChainResult: %v", err)
+	}
+	if gotPath != "/upstream/videoutils/v1/jobs/job-123/result" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if res.Video.MIME != "video/mp4" || len(res.Video.Data) == 0 {
+		t.Fatalf("video = %q/%d bytes", res.Video.MIME, len(res.Video.Data))
+	}
+}
+
+func TestChainSegmentResult(t *testing.T) {
+	var gotPath string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write(mp4Fixture())
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	res, err := ch.ChainSegmentResult(context.Background(), "job-123", 1)
+	if err != nil {
+		t.Fatalf("ChainSegmentResult: %v", err)
+	}
+	if gotPath != "/upstream/videoutils/v1/jobs/job-123/segments/1" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if res.Video.MIME != "video/mp4" {
+		t.Fatalf("video = %q", res.Video.MIME)
+	}
+
+	if _, err := ch.ChainSegmentResult(context.Background(), "job-123", -1); !errors.Is(err, llm.ErrUnsupported) {
+		t.Errorf("negative segment: err = %v, want ErrUnsupported", err)
+	}
+}
+
+func TestChainResultRejectsNonVideoResponse(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"status":"running"}`)) // a status page, not the clip
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	_, err := ch.ChainResult(context.Background(), "job-123")
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %v, want APIError for non-video body", err)
+	}
+}
+
+func TestChainSurfacesAPIError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+		_, _ = w.Write([]byte(`{"error":{"message":"job not found"}}`))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	ch, _ := p.ChainerModel("videoutils")
+	_, err := ch.ChainStatus(context.Background(), "job-void")
+	var apiErr *llm.APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("err = %T %v, want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusNotFound || apiErr.Message != "job not found" || apiErr.Model != "videoutils" {
+		t.Errorf("apiErr = %+v", apiErr)
+	}
+}

provider/llamaswap/videoutil.go

@@ -0,0 +1,139 @@
+// videoutil.go implements the videogen.VideoBackgroundRemovalProvider and
+// videogen.VideoUpscaleProvider surfaces against the mediautils shim reached
+// through llama-swap's /upstream passthrough (ADR-0025):
+//
+//	matte    POST /upstream/<id>/v1/video/matte    (Robust Video Matting)
+//	upscale  POST /upstream/<id>/v1/video/upscale  (per-frame Real-ESRGAN)
+//
+// Both are one-file multipart in, encoded clip out — the video siblings of
+// mediautil.go's still-image surfaces.
+package llamaswap
+
+import (
+	"context"
+	"fmt"
+	"mime"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/videogen"
+)
+
+// --- video background removal (matting) ---
+
+// VideoBackgroundRemoverModel implements
+// videogen.VideoBackgroundRemovalProvider against the mediautils shim's
+// POST /v1/video/matte. The id selects which upstream llama-swap loads.
+func (p *Provider) VideoBackgroundRemoverModel(id string, opts ...videogen.VideoBackgroundRemoverModelOption) (videogen.VideoBackgroundRemover, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = videogen.ApplyVideoBackgroundRemoverModelOptions(opts)
+	return &videoMatteModel{p: p, id: id}, nil
+}
+
+type videoMatteModel struct {
+	p  *Provider
+	id string
+}
+
+// RemoveVideoBackground implements videogen.VideoBackgroundRemover.
+func (m *videoMatteModel) RemoveVideoBackground(ctx context.Context, req videogen.VideoBackgroundRemovalRequest, opts ...videogen.VideoBackgroundRemovalOption) (*videogen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Video) == 0 {
+		return nil, fmt.Errorf("%w: video background removal requires a video", llm.ErrUnsupported)
+	}
+	if req.Output != "" && req.Output != "greenscreen_mp4" && req.Output != "alpha_webm" {
+		return nil, fmt.Errorf("%w: video matte output must be \"greenscreen_mp4\" or \"alpha_webm\", got %q", llm.ErrUnsupported, req.Output)
+	}
+	path, err := upstreamPath(m.id, "/v1/video/matte")
+	if err != nil {
+		return nil, err
+	}
+	body, contentType, err := buildMultipart("build video-matte form",
+		filePart{field: "file", filename: videoInputFilename(req.Filename, req.MIME), data: req.Video},
+		[]formField{{"output", req.Output, false}})
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxVideoResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleVideoResult(m.p.name, m.id, "video matte", raw, respType)
+}
+
+// --- video upscale ---
+
+// VideoUpscalerModel implements videogen.VideoUpscaleProvider against the
+// mediautils shim's POST /v1/video/upscale.
+func (p *Provider) VideoUpscalerModel(id string, opts ...videogen.VideoUpscalerModelOption) (videogen.VideoUpscaler, error) {
+	if err := p.requireBaseURL(); err != nil {
+		return nil, err
+	}
+	_ = videogen.ApplyVideoUpscalerModelOptions(opts)
+	return &videoUpscaleModel{p: p, id: id}, nil
+}
+
+type videoUpscaleModel struct {
+	p  *Provider
+	id string
+}
+
+// UpscaleVideo implements videogen.VideoUpscaler.
+func (m *videoUpscaleModel) UpscaleVideo(ctx context.Context, req videogen.VideoUpscaleRequest, opts ...videogen.VideoUpscaleOption) (*videogen.Result, error) {
+	req = req.Apply(opts...)
+	if len(req.Video) == 0 {
+		return nil, fmt.Errorf("%w: video upscale requires a video", llm.ErrUnsupported)
+	}
+	if req.Scale != 0 && req.Scale != 2 && req.Scale != 4 {
+		return nil, fmt.Errorf("%w: video upscale scale must be 2 or 4, got %d", llm.ErrUnsupported, req.Scale)
+	}
+	path, err := upstreamPath(m.id, "/v1/video/upscale")
+	if err != nil {
+		return nil, err
+	}
+	scale := ""
+	if req.Scale != 0 {
+		scale = strconv.Itoa(req.Scale)
+	}
+	body, contentType, err := buildMultipart("build video-upscale form",
+		filePart{field: "file", filename: videoInputFilename(req.Filename, req.MIME), data: req.Video},
+		[]formField{{"scale", scale, false}})
+	if err != nil {
+		return nil, err
+	}
+	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxVideoResponseBytes)
+	if err != nil {
+		return nil, err
+	}
+	return singleVideoResult(m.p.name, m.id, "video upscale", raw, respType)
+}
+
+// videoInputFilename picks the multipart filename hint for a caller-supplied
+// video: the caller's (sanitized — upload metadata is untrusted), else one
+// derived from the MIME subtype ("video.mp4"), else "video". Mirrors
+// transcriptionFilename.
+func videoInputFilename(filename, mimeType string) string {
+	if name := sanitizeFilename(filename); name != "" {
+		return name
+	}
+	mt := strings.ToLower(strings.TrimSpace(mimeType))
+	if parsed, _, err := mime.ParseMediaType(mt); err == nil {
+		mt = parsed
+	}
+	switch mt {
+	case "video/mp4":
+		return "video.mp4"
+	case "video/webm":
+		return "video.webm"
+	case "video/quicktime":
+		return "video.mov"
+	case "video/x-matroska":
+		return "video.mkv"
+	default:
+		return "video"
+	}
+}

provider/openai/model.go

@@ -82,7 +82,7 @@ func (m *model) do(ctx context.Context, req llm.Request, stream bool) (*http.Res
 			Model:    m.id,
 			Status:   http.StatusUnauthorized,
 			Code:     "missing_api_key",
-			Message:  "no API key configured: set OPENAI_API_KEY or use WithAPIKey",
+			Message:  "no API key configured: set " + m.p.apiKeyName + " or use WithAPIKey",
 		}
 	}
 	body, err := json.Marshal(m.buildRequest(req, stream))

provider/openai/openai.go

@@ -34,6 +34,7 @@ const defaultBaseURL = "https://api.openai.com/v1"
 type Provider struct {
 	name            string
 	apiKey          string
+	apiKeyName      string
 	baseURL         string
 	client          *http.Client
 	caps            llm.Capabilities
@@ -64,6 +65,14 @@ func WithHTTPClient(c *http.Client) Option {
 	}
 }
 
+// WithAPIKeyName sets the environment-variable name shown in the missing-key
+// error (default "OPENAI_API_KEY"). Why: the same client serves compat
+// endpoints keyed by other env vars (e.g. KIMI_API_KEY), and the error should
+// name the one the operator actually needs to set.
+func WithAPIKeyName(name string) Option {
+	return func(p *Provider) { p.apiKeyName = name }
+}
+
 // WithName overrides the registry name ("openai" by default). Why: the same
 // client serves many OpenAI-compatible endpoints, and each needs a distinct
 // name in "provider/model" specs and error reporting.
@@ -106,6 +115,7 @@ func New(opts ...Option) *Provider {
 	p := &Provider{
 		name:       "openai",
 		apiKey:     os.Getenv("OPENAI_API_KEY"),
+		apiKeyName: "OPENAI_API_KEY",
 		baseURL:    defaultBaseURL,
 		client:     http.DefaultClient,
 		caps:       defaultCapabilities(),

videogen/chain.go

@@ -0,0 +1,103 @@
+package videogen
+
+import "context"
+
+// ChainSegment is one prompt in a multi-segment ("long video") chain.
+type ChainSegment struct {
+	// Prompt describes this segment. Required.
+	Prompt string
+
+	// Seconds is the segment's requested length; 0 = backend default.
+	Seconds float64
+}
+
+// ChainRequest asks a chain orchestrator to generate a long video as a
+// sequence of segments, each continuing from the previous segment's last
+// frame. Zero values mean "backend default" (ADR-0025).
+type ChainRequest struct {
+	// Segments are the per-segment prompts in order. At least one required.
+	Segments []ChainSegment
+
+	// InitImage optionally conditions the FIRST segment on a starting frame
+	// (image-to-video); nil = pure text-to-video.
+	InitImage []byte
+
+	// SmoothJoins asks the orchestrator to interpolate across segment
+	// boundaries (RIFE-style) so cuts don't pop.
+	SmoothJoins bool
+
+	// Size is the requested resolution, e.g. "1280x704"; "" = backend
+	// default.
+	Size string
+}
+
+// ChainJob is a chain job's progress snapshot.
+type ChainJob struct {
+	// Status is the backend's job state, passed through verbatim
+	// (e.g. "queued", "running", "done", "failed").
+	Status string
+
+	// Segment is the segment currently being generated (1-based); Total is
+	// the segment count.
+	Segment int
+	Total   int
+
+	// SegmentIDs name the COMPLETED per-segment artifacts, in order. A
+	// mid-chain failure still leaves those segments retrievable via
+	// ChainSegmentResult — note it takes the segment's index in the chain,
+	// not an id string; this list tells you WHICH segments completed. So
+	// multi-minute GPU output is never discarded. Entries the backend
+	// reports without a usable id are skipped (the unfiltered list survives
+	// in Raw).
+	SegmentIDs []string
+
+	// Raw is the provider-native job payload. May be nil.
+	Raw any
+}
+
+// Chainer drives a multi-segment video-chain job. Unlike Model.Generate it
+// is deliberately ASYNC — a chain runs through multiple GPU loads for many
+// minutes, so callers submit, poll, and fetch instead of holding one
+// blocking call open.
+type Chainer interface {
+	// SubmitChain starts a chain job and returns its job id.
+	SubmitChain(ctx context.Context, req ChainRequest) (string, error)
+
+	// ChainStatus reports the job's progress. Polling also signals liveness
+	// to backends that unload idle orchestrators.
+	ChainStatus(ctx context.Context, jobID string) (*ChainJob, error)
+
+	// ChainResult fetches the finished, concatenated clip.
+	ChainResult(ctx context.Context, jobID string) (*Result, error)
+
+	// ChainSegmentResult fetches one completed segment's clip (n indexes the
+	// job's segment list) — the partial-delivery path when a chain dies
+	// mid-run.
+	ChainSegmentResult(ctx context.Context, jobID string, n int) (*Result, error)
+}
+
+// ChainerModelOption configures a Chainer at construction time. Reserved for
+// future per-model settings.
+type ChainerModelOption func(*ChainerModelConfig)
+
+// ChainerModelConfig carries per-model construction settings.
+type ChainerModelConfig struct{}
+
+// ApplyChainerModelOptions folds options into a config.
+func ApplyChainerModelOptions(opts []ChainerModelOption) ChainerModelConfig {
+	var cfg ChainerModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// ChainerProvider mints Chainers bound to one backend.
+type ChainerProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// ChainerModel returns a Chainer bound to the given id (passed through
+	// to the backend verbatim; no catalog validation).
+	ChainerModel(id string, opts ...ChainerModelOption) (Chainer, error)
+}

videogen/lipsync.go

@@ -0,0 +1,93 @@
+package videogen
+
+import "context"
+
+// LipsyncRequest asks a talking-head backend (SadTalker style) to animate a
+// still portrait so it speaks the given audio. Zero values mean "backend
+// default" (ADR-0025).
+type LipsyncRequest struct {
+	// Image is the portrait to animate. Required.
+	Image Image
+
+	// Audio is the encoded speech the head lip-syncs to. Required. Carried
+	// as bytes (never a URL), mirroring audio.TranscriptionRequest.
+	Audio []byte
+
+	// AudioMIME is the audio MIME type (e.g. "audio/wav"); "" = let the
+	// backend sniff it.
+	AudioMIME string
+
+	// AudioFilename is the multipart filename hint some backends key their
+	// format detection on; "" derives one from AudioMIME or falls back to
+	// "audio".
+	AudioFilename string
+
+	// Still reduces head motion to blinks and lip movement (less uncanny on
+	// formal portraits); false = backend default motion.
+	Still bool
+
+	// Enhance runs the backend's face enhancer over the output frames.
+	Enhance bool
+
+	// Preprocess selects how the backend frames the face: "crop" (animate
+	// the face crop) or "full" (paste the animated face back into the whole
+	// image); "" = backend default.
+	Preprocess string
+}
+
+// LipsyncOption mutates a LipsyncRequest before it is sent.
+type LipsyncOption func(*LipsyncRequest)
+
+// WithLipsyncStill reduces head motion to blinks and lip movement.
+func WithLipsyncStill() LipsyncOption { return func(r *LipsyncRequest) { r.Still = true } }
+
+// WithLipsyncEnhance runs the backend's face enhancer over the output.
+func WithLipsyncEnhance() LipsyncOption { return func(r *LipsyncRequest) { r.Enhance = true } }
+
+// WithLipsyncPreprocess selects the face framing ("crop" or "full").
+func WithLipsyncPreprocess(p string) LipsyncOption {
+	return func(r *LipsyncRequest) { r.Preprocess = p }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r LipsyncRequest) Apply(opts ...LipsyncOption) LipsyncRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// Lipsyncer animates still portraits into talking-head clips. Its own small
+// interface rather than a method on Model: lip-syncers are not text-to-video
+// generators — they bind to a different backend id entirely.
+type Lipsyncer interface {
+	// Lipsync returns the talking-head clip. Generation is slow (minutes);
+	// bound the call with a context deadline.
+	Lipsync(ctx context.Context, req LipsyncRequest, opts ...LipsyncOption) (*Result, error)
+}
+
+// LipsyncModelOption configures a Lipsyncer at construction time. Reserved
+// for future per-model settings.
+type LipsyncModelOption func(*LipsyncModelConfig)
+
+// LipsyncModelConfig carries per-model construction settings.
+type LipsyncModelConfig struct{}
+
+// ApplyLipsyncModelOptions folds options into a config.
+func ApplyLipsyncModelOptions(opts []LipsyncModelOption) LipsyncModelConfig {
+	var cfg LipsyncModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// LipsyncProvider mints Lipsyncers bound to one backend.
+type LipsyncProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// LipsyncModel returns a Lipsyncer bound to the given id (passed through
+	// to the backend verbatim; no catalog validation).
+	LipsyncModel(id string, opts ...LipsyncModelOption) (Lipsyncer, error)
+}

videogen/matte.go

@@ -0,0 +1,80 @@
+package videogen
+
+import "context"
+
+// VideoBackgroundRemovalRequest asks a video-matting backend (Robust Video
+// Matting style) to separate the foreground subject from the background of a
+// clip. Zero values mean "backend default" (ADR-0025).
+type VideoBackgroundRemovalRequest struct {
+	// Video is the encoded clip to matte. Required. Carried as bytes (never
+	// a URL).
+	Video []byte
+
+	// MIME is the video MIME type (e.g. "video/mp4"); "" = let the backend
+	// sniff it.
+	MIME string
+
+	// Filename is the multipart filename hint some backends key their format
+	// detection on; "" derives one from MIME ("video.mp4") or falls back to
+	// "video".
+	Filename string
+
+	// Output selects the delivery container: "greenscreen_mp4" (subject over
+	// solid green, universally playable) or "alpha_webm" (true transparency,
+	// VP9 alpha channel); "" = backend default.
+	Output string
+}
+
+// VideoBackgroundRemovalOption mutates a VideoBackgroundRemovalRequest
+// before it is sent.
+type VideoBackgroundRemovalOption func(*VideoBackgroundRemovalRequest)
+
+// WithVideoBackgroundOutput selects the delivery container
+// ("greenscreen_mp4" or "alpha_webm").
+func WithVideoBackgroundOutput(o string) VideoBackgroundRemovalOption {
+	return func(r *VideoBackgroundRemovalRequest) { r.Output = o }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r VideoBackgroundRemovalRequest) Apply(opts ...VideoBackgroundRemovalOption) VideoBackgroundRemovalRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// VideoBackgroundRemover mattes the subject out of video clips — the moving-
+// picture sibling of imagegen.BackgroundRemover.
+type VideoBackgroundRemover interface {
+	// RemoveVideoBackground returns the matted clip. Matting is slow
+	// (per-frame inference); bound the call with a context deadline.
+	RemoveVideoBackground(ctx context.Context, req VideoBackgroundRemovalRequest, opts ...VideoBackgroundRemovalOption) (*Result, error)
+}
+
+// VideoBackgroundRemoverModelOption configures a VideoBackgroundRemover at
+// construction time. Reserved for future per-model settings.
+type VideoBackgroundRemoverModelOption func(*VideoBackgroundRemoverModelConfig)
+
+// VideoBackgroundRemoverModelConfig carries per-model construction settings.
+type VideoBackgroundRemoverModelConfig struct{}
+
+// ApplyVideoBackgroundRemoverModelOptions folds options into a config.
+func ApplyVideoBackgroundRemoverModelOptions(opts []VideoBackgroundRemoverModelOption) VideoBackgroundRemoverModelConfig {
+	var cfg VideoBackgroundRemoverModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// VideoBackgroundRemovalProvider mints VideoBackgroundRemovers bound to one
+// backend.
+type VideoBackgroundRemovalProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// VideoBackgroundRemoverModel returns a VideoBackgroundRemover bound to
+	// the given id (passed through to the backend verbatim; no catalog
+	// validation).
+	VideoBackgroundRemoverModel(id string, opts ...VideoBackgroundRemoverModelOption) (VideoBackgroundRemover, error)
+}

videogen/vupscale.go

@@ -0,0 +1,74 @@
+package videogen
+
+import "context"
+
+// VideoUpscaleRequest asks a super-resolution backend (per-frame Real-ESRGAN
+// style) to enlarge a clip. Zero values mean "backend default" (ADR-0025).
+type VideoUpscaleRequest struct {
+	// Video is the encoded clip to upscale. Required. Carried as bytes
+	// (never a URL).
+	Video []byte
+
+	// MIME is the video MIME type (e.g. "video/mp4"); "" = let the backend
+	// sniff it.
+	MIME string
+
+	// Filename is the multipart filename hint some backends key their format
+	// detection on; "" derives one from MIME ("video.mp4") or falls back to
+	// "video".
+	Filename string
+
+	// Scale is the enlargement factor (2 or 4 on the reference backend);
+	// 0 = backend default.
+	Scale int
+}
+
+// VideoUpscaleOption mutates a VideoUpscaleRequest before it is sent.
+type VideoUpscaleOption func(*VideoUpscaleRequest)
+
+// WithVideoUpscaleScale sets the enlargement factor.
+func WithVideoUpscaleScale(s int) VideoUpscaleOption {
+	return func(r *VideoUpscaleRequest) { r.Scale = s }
+}
+
+// Apply returns a copy of the request with all options applied.
+func (r VideoUpscaleRequest) Apply(opts ...VideoUpscaleOption) VideoUpscaleRequest {
+	for _, opt := range opts {
+		opt(&r)
+	}
+	return r
+}
+
+// VideoUpscaler enlarges video clips frame by frame — the moving-picture
+// sibling of imagegen.Upscaler.
+type VideoUpscaler interface {
+	// UpscaleVideo returns the enlarged clip. Upscaling is slow (per-frame
+	// inference); bound the call with a context deadline.
+	UpscaleVideo(ctx context.Context, req VideoUpscaleRequest, opts ...VideoUpscaleOption) (*Result, error)
+}
+
+// VideoUpscalerModelOption configures a VideoUpscaler at construction time.
+// Reserved for future per-model settings.
+type VideoUpscalerModelOption func(*VideoUpscalerModelConfig)
+
+// VideoUpscalerModelConfig carries per-model construction settings.
+type VideoUpscalerModelConfig struct{}
+
+// ApplyVideoUpscalerModelOptions folds options into a config.
+func ApplyVideoUpscalerModelOptions(opts []VideoUpscalerModelOption) VideoUpscalerModelConfig {
+	var cfg VideoUpscalerModelConfig
+	for _, opt := range opts {
+		opt(&cfg)
+	}
+	return cfg
+}
+
+// VideoUpscaleProvider mints VideoUpscalers bound to one backend.
+type VideoUpscaleProvider interface {
+	// Name is the registry identifier for the provider.
+	Name() string
+
+	// VideoUpscalerModel returns a VideoUpscaler bound to the given id
+	// (passed through to the backend verbatim; no catalog validation).
+	VideoUpscalerModel(id string, opts ...VideoUpscalerModelOption) (VideoUpscaler, error)
+}