chore: ignore .claude/ (agent worktrees)

A `git add -A` in a checkout with a worktree under .claude/worktrees/ records
it as a gitlink — a submodule pointer no clone can resolve. That happened in
gadfly today; this is the same one-line prevention, applied before it happens
here.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

.env.example

@@ -7,6 +7,7 @@ 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"
+#QWEN_API_KEY=sk-...           # Alibaba Model Studio (Qwen); provider name "qwen"
 #ANTHROPIC_API_KEY=sk-ant-...
 #GOOGLE_API_KEY=...
 

.gitignore

@@ -28,3 +28,6 @@ go.work.sum
 
 # macOS
 .DS_Store
+
+# Local worktrees created for agent work — never part of the repo.
+.claude/

README.md

@@ -122,6 +122,7 @@ Chains are health-tracked per target:
 |----------|-----------|-------------|------------------|
 | OpenAI (+compatible) | `openai` | `OPENAI_API_KEY` | https://api.openai.com/v1 |
 | Kimi (Moonshot AI) | `kimi` | `KIMI_API_KEY` | https://api.moonshot.ai/v1 |
+| Qwen (Alibaba) | `qwen` | `QWEN_API_KEY` | https://dashscope-intl.aliyuncs.com/compatible-mode/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 |
@@ -134,6 +135,19 @@ 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`.
 
+Qwen is the same shape: Alibaba Model Studio's OpenAI-compatible mode, reusing
+the openai client. The `qwen` built-in defaults to the international
+(Singapore) host; reach the China host or a workspace-scoped regional one with
+a `qwen://` DSN, e.g.
+`LLM_QCN=qwen://[email protected]/compatible-mode/v1`. Model Studio
+also fronts the same models with an Anthropic-compatible `/v1/messages` shim —
+majordomo does **not** use it, because on that surface `reasoning_effort` is
+dropped, `Request.Schema` stops being enforced, and cached-token accounting
+disappears; see [ADR-0027](docs/adr/0027-qwen-builtin.md). Two Alibaba-side
+quirks are worth knowing: thinking is on by default for some models (e.g.
+`qwen3.7-plus`), and the Qwen3 open-source models require streaming while
+thinking, so buffered `Generate` calls want a Max/Plus model.
+
 OpenAI-compatible / Anthropic-compatible endpoints: construct the provider
 with a name and base URL and register it —
 
@@ -165,7 +179,7 @@ m, _ := reg.Parse("m5/qwen3:30b,m1/qwen3:30b,thinking")
 ```
 
 DSN format: `scheme://[token@]host[/path]`, scheme ∈ `foreman`, `ollama`,
-`ollama-cloud`, `openai`, `kimi`, `anthropic`, `google`/`gemini`, `llama-swap`,
+`ollama-cloud`, `openai`, `kimi`, `qwen`, `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
@@ -272,17 +286,30 @@ tr, err := tm.Transcribe(ctx, audio.TranscriptionRequest{
 voices, err := ls.ListVoices(ctx, "kokoro") // []string of voice ids
 ```
 
-## Video: text-to-video + image-to-video
+## Video: text-to-video, image-to-video, first-last-frame
 
 Video generation lives in the `videogen` package (ADR-0019), mirroring
 imagegen/audio: one small `Model` contract, zero values mean backend
-defaults, bytes in/out. Text-to-video and image-to-video are one surface —
-a nil `InitImage` is a pure text prompt; setting it conditions generation
-on that frame (hybrid checkpoints like Wan 2.2 TI2V serve both). First
-backend: llama-swap (blocking `/v1/videos/sync`, vLLM-Omni style — the
+defaults, bytes in/out. All modes are one surface, selected by which
+keyframes are set rather than by a mode flag:
+
+| `InitImage` | `LastImage` | mode |
+|---|---|---|
+| nil | nil | text-to-video |
+| set | nil | image-to-video (hybrid checkpoints like Wan 2.2 TI2V serve both) |
+| set | set | first-last-frame — both ends pinned |
+| nil | set | pin the destination, model invents the approach |
+
+First backend: llama-swap (blocking `/v1/videos/sync`, vLLM-Omni style — the
 response body is the encoded clip, so `Result` carries a single `Video`).
 Generation runs for minutes; bound the call with a context deadline.
 
+**`LastImage` support is per-model and cannot be detected.** A backend that
+does not understand a trailing keyframe ignores the part and returns an
+ordinary clip — indistinguishable from success. There is no capability bit,
+because the contract has no way to learn one, so a caller depending on the
+pin must establish support out of band.
+
 ```go
 vm, _ := ls.VideoModel("videogen-wan22-5b")
 res, err := vm.Generate(ctx, videogen.Request{Prompt: "a cat surfing"},
@@ -407,6 +434,7 @@ to build one.
 |----------------------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
 | OpenAI (+compatible) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Kimi (Moonshot AI) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅³ | ✅ |
+| Qwen (Alibaba) | ✅ | ✅ | ✅ | ✅ | ✅⁴ | ✅⁴ | ✅ |
 | Anthropic (+compat) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Google (Gemini) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Ollama Cloud | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
@@ -431,6 +459,13 @@ probe and management methods on `*llamaswap.Provider`.
 level; whether a call succeeds depends on the Moonshot model — only the vision
 variants (e.g. `moonshot-v1-8k-vision-preview`) accept images.
 
+⁴ Qwen also reuses the openai client (ADR-0027), so both columns are present at
+the client level and gated by the Model Studio model you name: `json_schema`
+structured output is on the Max/Plus families, image inputs on the `qwen-vl-*`
+/ `qwen3-vl-*` models. `reasoning_effort` rides through as a top-level field —
+one reason the built-in speaks OpenAI-compat rather than Model Studio's
+Anthropic-compat shim.
+
 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/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 != "" {
@@ -83,7 +123,8 @@ 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
+	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

builtin.go

@@ -2,7 +2,6 @@ package majordomo
 
 import (
 	"net/http"
-	"strings"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/anthropic"
@@ -18,7 +17,13 @@ const (
 	// 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"
+	ProviderKimi = "kimi"
+	// ProviderQwen is Alibaba's Qwen models over Model Studio's
+	// OpenAI-compatible Chat Completions endpoint. Reuses the openai client
+	// (like kimi and llama-swap); keyed by QWEN_API_KEY, default base URL
+	// qwenBaseURL. ADR-0027 records why the OpenAI surface and not the
+	// Anthropic-compatible one Model Studio also exposes.
+	ProviderQwen        = "qwen"
 	ProviderAnthropic   = "anthropic"
 	ProviderGoogle      = "google"
 	ProviderOllama      = "ollama"
@@ -37,6 +42,55 @@ const (
 // China endpoint (api.moonshot.cn/v1) is reachable via a kimi:// LLM_* DSN.
 const kimiBaseURL = "https://api.moonshot.ai/v1"
 
+// qwenBaseURL is Alibaba Model Studio's international (Singapore) endpoint in
+// OpenAI-compatible mode. The China endpoint
+// (dashscope.aliyuncs.com/compatible-mode/v1) and any regional host are
+// reachable via a qwen:// LLM_* DSN.
+const qwenBaseURL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+
+// openaiCompatScheme builds the DSN factory shared by every built-in that is
+// "the openai client pointed somewhere else" (kimi, qwen, ...). The provider
+// is named after the LLM_<NAME> var that defined it, takes its credential from
+// the DSN token — not the built-in's own env var, which does nothing for a
+// DSN-defined provider — and so names that same LLM_<NAME> var in the
+// missing-key hint, matching the lazy-resolution key form in providerFor.
+//
+// wrap is the caller's option-decorator (it injects the registry's HTTP
+// client), so a DSN provider is built exactly like the eager built-ins.
+func openaiCompatScheme(wrap func(...openai.Option) []openai.Option) SchemeFactory {
+	return func(name string, dsn DSN) (llm.Provider, error) {
+		return openai.New(wrap(
+			openai.WithName(name),
+			openai.WithBaseURL(dsn.BaseURL()),
+			openai.WithAPIKey(dsn.Token),
+			openai.WithAPIKeyName(envKeyForProvider(name)),
+		)...), nil
+	}
+}
+
+// registerOpenAICompatBuiltin installs BOTH halves of an OpenAI-compat
+// built-in: the eager provider under name (credential from keyEnv) and the
+// matching name:// DSN scheme. Why both in one call: the two halves are a pair
+// — a built-in whose scheme is missing resolves as a spec but not from an
+// LLM_* DSN, and the credential rules below have to hold identically in each.
+// Adding the next one is a single line rather than six lines to copy.
+//
+// The two credential rules, holding by construction for every caller:
+//   - WithAPIKey is passed UNCONDITIONALLY, even when the lookup comes back
+//     empty. openai.New defaults its key to OPENAI_API_KEY, so anything less
+//     lets an unset keyEnv silently authenticate as OpenAI.
+//   - WithAPIKeyName makes the synthetic-401 hint name keyEnv, so a keyless
+//     call tells the operator the variable that actually fixes it.
+func registerOpenAICompatBuiltin(r *Registry, wrap func(...openai.Option) []openai.Option, name, baseURL, keyEnv string) {
+	r.providers[name] = openai.New(wrap(
+		openai.WithName(name),
+		openai.WithBaseURL(baseURL),
+		openai.WithAPIKey(r.envLookup(keyEnv)),
+		openai.WithAPIKeyName(keyEnv),
+	)...)
+	r.schemes[name] = openaiCompatScheme(wrap)
+}
+
 // 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.
@@ -83,32 +137,19 @@ 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
-	}
+	// Third-party endpoints that ARE the openai client at another base URL —
+	// no new package, mirroring llama-swap's chat path. Each gets the eager
+	// built-in plus its name:// DSN scheme, and the credential rules hold by
+	// construction (see registerOpenAICompatBuiltin).
+	//
+	// kimi (ADR-0026): Moonshot's international endpoint; China host via
+	// kimi://[email protected]/v1.
+	registerOpenAICompatBuiltin(r, openaiOpts, ProviderKimi, kimiBaseURL, "KIMI_API_KEY")
+	// qwen (ADR-0027): Alibaba Model Studio's international host. Model Studio
+	// also exposes an Anthropic-compatible endpoint; the ADR records why the
+	// OpenAI one is the built-in. China / workspace-scoped regional hosts via
+	// qwen://[email protected]/compatible-mode/v1.
+	registerOpenAICompatBuiltin(r, openaiOpts, ProviderQwen, qwenBaseURL, "QWEN_API_KEY")
 
 	// llama-swap: OpenAI-compatible chat + image generation + management
 	// endpoints over a model-swapping proxy. Chat reuses the openai client

builtin_kimi_test.go

@@ -1,171 +0,0 @@
-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")
-	}
-}

builtin_openaicompat_test.go

@@ -0,0 +1,240 @@
+package majordomo
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// Shared fixtures and the shared contract for the built-ins that are "the
+// openai client pointed somewhere else" (kimi, qwen, ...). They live here
+// rather than in any one provider's test file so a new OpenAI-compat built-in
+// has nothing to copy — the same reason registerOpenAICompatBuiltin exists on
+// the production side.
+
+// chatCompletionOK 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 chatCompletionOK = `{"id":"c1","object":"chat.completion","choices":[` +
+	`{"index":0,"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}]}`
+
+// captureRT records the last request (and the bytes of its body) and returns a
+// canned response without touching the network, so these tests stay hermetic
+// while still exercising the real openai client the built-ins reuse: base URL,
+// auth header, and the JSON actually put on the wire.
+type captureRT struct {
+	req     *http.Request
+	reqBody []byte
+	body    string
+}
+
+func (c *captureRT) RoundTrip(r *http.Request) (*http.Response, error) {
+	c.req = r
+	// Drain and close the request body: a RoundTripper owns it, and those
+	// bytes are what wire-shape assertions read.
+	c.reqBody = nil
+	if r.Body != nil {
+		c.reqBody, _ = io.ReadAll(r.Body)
+		_ = r.Body.Close()
+	}
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(c.body)),
+		Header:     make(http.Header),
+		Request:    r,
+	}, nil
+}
+
+// singleKeyEnv builds a WithEnvLookup function that knows exactly one variable
+// and returns "" for everything else. The empty default has teeth: a built-in
+// that reached for any other variable name gets nothing, so the request 401s
+// and the test fails rather than quietly authenticating off the wrong key.
+func singleKeyEnv(key, value string) func(string) string {
+	return func(k string) string {
+		if k == key {
+			return value
+		}
+		return ""
+	}
+}
+
+// openAICompatBuiltin describes one built-in for the shared contract below.
+// Adding an OpenAI-compat built-in means adding a row here — not copying a
+// test file, which is how kimi's and qwen's suites became near-identical.
+type openAICompatBuiltin struct {
+	name    string // registry name and spec prefix
+	keyEnv  string // the credential variable this built-in reads
+	model   string // a current model id for that endpoint
+	wantURL string // chat-completions URL the default endpoint must produce
+
+	// The name:// DSN case: an alternate host (regional/China endpoint)
+	// reached through an LLM_<dsnVar> definition.
+	dsnVar     string
+	dsnHost    string
+	wantDSNURL string
+}
+
+var openAICompatBuiltins = []openAICompatBuiltin{
+	{
+		name:       ProviderKimi,
+		keyEnv:     "KIMI_API_KEY",
+		model:      "kimi-k2-0711-preview",
+		wantURL:    "https://api.moonshot.ai/v1/chat/completions",
+		dsnVar:     "LLM_KCN",
+		dsnHost:    "api.moonshot.cn/v1",
+		wantDSNURL: "https://api.moonshot.cn/v1/chat/completions",
+	},
+	{
+		name:       ProviderQwen,
+		keyEnv:     "QWEN_API_KEY",
+		model:      "qwen3.8-max",
+		wantURL:    "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions",
+		dsnVar:     "LLM_QCN",
+		dsnHost:    "dashscope.aliyuncs.com/compatible-mode/v1",
+		wantDSNURL: "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions",
+	},
+}
+
+// TestOpenAICompatBuiltins is the whole contract an OpenAI-compat built-in
+// owes, asserted identically for every one of them: it resolves in Parse and
+// targets its own endpoint with its own key; a missing key fails closed naming
+// the right variable and never reaching the network; its name:// DSN reaches
+// any other host on the DSN token; and a keyless DSN names the LLM_<NAME> that
+// actually fixes it rather than the built-in's variable, which does nothing
+// for a DSN-defined provider.
+func TestOpenAICompatBuiltins(t *testing.T) {
+	for _, tc := range openAICompatBuiltins {
+		t.Run(tc.name+"/builtin", func(t *testing.T) {
+			rt := &captureRT{body: chatCompletionOK}
+			secret := tc.name + "-secret"
+			r := newTestRegistry(t,
+				WithEnvLookup(singleKeyEnv(tc.keyEnv, secret)),
+				WithHTTPClient(&http.Client{Transport: rt}),
+			)
+
+			if p, ok := r.Provider(tc.name); !ok {
+				t.Fatalf("built-in %q not registered", tc.name)
+			} else if p.Name() != tc.name {
+				t.Errorf("name = %q, want %q", p.Name(), tc.name)
+			}
+
+			spec := tc.name + "/" + tc.model
+			m, err := r.Parse(spec)
+			if err != nil {
+				t.Fatalf("Parse(%q): %v", spec, err)
+			}
+			if got := targetsOf(t, m); len(got) != 1 || got[0] != spec {
+				t.Fatalf("targets = %v, want [%q]", got, spec)
+			}
+
+			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 rt.req.URL.String() != tc.wantURL {
+				t.Errorf("URL = %q, want %q", rt.req.URL.String(), tc.wantURL)
+			}
+			if want := "Bearer " + secret; rt.req.Header.Get("Authorization") != want {
+				t.Errorf("Authorization = %q, want %q", rt.req.Header.Get("Authorization"), want)
+			}
+		})
+
+		t.Run(tc.name+"/builtin missing key", func(t *testing.T) {
+			rt := &captureRT{body: chatCompletionOK}
+			r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+
+			m, err := r.Parse(tc.name + "/" + tc.model)
+			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, tc.keyEnv) {
+				t.Errorf("message = %q, want it to name %s", apiErr.Message, tc.keyEnv)
+			}
+			// The load-bearing half: openai.New defaults its key to
+			// OPENAI_API_KEY, so a built-in that stopped passing WithAPIKey
+			// unconditionally would authenticate as OpenAI instead of failing.
+			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")
+			}
+		})
+
+		t.Run(tc.name+"/dsn scheme", func(t *testing.T) {
+			rt := &captureRT{body: chatCompletionOK}
+			r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+			if err := r.LoadEnv(map[string]string{
+				tc.dsnVar: tc.name + "://tok@" + tc.dsnHost,
+			}); err != nil {
+				t.Fatalf("LoadEnv: %v", err)
+			}
+
+			dsnName := strings.ToLower(strings.TrimPrefix(tc.dsnVar, "LLM_"))
+			m, err := r.Parse(dsnName + "/" + tc.model)
+			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 rt.req.URL.String() != tc.wantDSNURL {
+				t.Errorf("URL = %q, want %q", rt.req.URL.String(), tc.wantDSNURL)
+			}
+			if want := "Bearer tok"; rt.req.Header.Get("Authorization") != want {
+				t.Errorf("Authorization = %q, want %q", rt.req.Header.Get("Authorization"), want)
+			}
+		})
+
+		t.Run(tc.name+"/dsn scheme missing token", func(t *testing.T) {
+			rt := &captureRT{body: chatCompletionOK}
+			r := newTestRegistry(t, WithHTTPClient(&http.Client{Transport: rt}))
+			if err := r.LoadEnv(map[string]string{
+				tc.dsnVar: tc.name + "://" + tc.dsnHost, // no token
+			}); err != nil {
+				t.Fatalf("LoadEnv: %v", err)
+			}
+
+			dsnName := strings.ToLower(strings.TrimPrefix(tc.dsnVar, "LLM_"))
+			m, err := r.Parse(dsnName + "/" + tc.model)
+			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)
+			}
+			// A keyless DSN is fixed by adding a token to that DSN, so the
+			// hint must name the defining variable — never the built-in's own
+			// key, which does nothing for a DSN-defined provider.
+			if !strings.Contains(apiErr.Message, tc.dsnVar) {
+				t.Errorf("message = %q, want it to name %s", apiErr.Message, tc.dsnVar)
+			}
+			if strings.Contains(apiErr.Message, tc.keyEnv) {
+				t.Errorf("message = %q, must not name %s for a DSN provider", apiErr.Message, tc.keyEnv)
+			}
+			if rt.req != nil {
+				t.Error("network was hit despite missing token")
+			}
+		})
+	}
+}

builtin_qwen_test.go

@@ -0,0 +1,86 @@
+package majordomo
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// The contract qwen shares with every other OpenAI-compat built-in (endpoint,
+// credential isolation, its qwen:// DSN) is asserted by the table in
+// builtin_openaicompat_test.go. What remains here is qwen-specific: the
+// reverse-leak direction, and the wire claim ADR-0027 turns on.
+
+// TestQwenBuiltinKeyDoesNotLeakToOpenAI: QWEN_API_KEY is the qwen built-in's
+// credential and nothing else's. Why this direction too: the shared table's
+// missing-key case only proves qwen never borrows OPENAI_API_KEY; this proves
+// the reverse — a registry that can see QWEN_API_KEY must not hand it to the
+// openai built-in, which would send an Alibaba key to api.openai.com.
+func TestQwenBuiltinKeyDoesNotLeakToOpenAI(t *testing.T) {
+	// Set before newTestRegistry: the openai built-in reads OPENAI_API_KEY at
+	// construction. Giving it a real key is what keeps this test honest — a
+	// keyless openai target would 401 before any request, and the assertion
+	// below would pass without a single byte reaching the wire.
+	t.Setenv("OPENAI_API_KEY", "openai-secret")
+
+	rt := &captureRT{body: chatCompletionOK}
+	r := newTestRegistry(t,
+		WithEnvLookup(singleKeyEnv("QWEN_API_KEY", "qwen-secret")),
+		WithHTTPClient(&http.Client{Transport: rt}),
+	)
+
+	m, err := r.Parse("openai/gpt-4o-mini")
+	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 := "Bearer openai-secret"; rt.req.Header.Get("Authorization") != want {
+		t.Errorf("Authorization = %q, want %q — the qwen credential must not reach the openai built-in",
+			rt.req.Header.Get("Authorization"), want)
+	}
+}
+
+// TestQwenReasoningEffortReachesWire is the load-bearing test for ADR-0027's
+// central claim: Model Studio's OpenAI-compatible surface takes reasoning as a
+// top-level "reasoning_effort" body field, which the openai client already
+// sends — so llm.WithReasoningEffort survives the trip on qwen with no
+// qwen-specific code. Routing qwen through the anthropic client instead would
+// drop it silently (provider/anthropic ignores ReasoningEffort by design), and
+// that difference would be invisible without asserting on the wire body.
+func TestQwenReasoningEffortReachesWire(t *testing.T) {
+	rt := &captureRT{body: chatCompletionOK}
+	r := newTestRegistry(t,
+		WithEnvLookup(singleKeyEnv("QWEN_API_KEY", "qwen-secret")),
+		WithHTTPClient(&http.Client{Transport: rt}),
+	)
+
+	m, err := r.Parse("qwen/qwen3.8-max")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	_, err = m.Generate(context.Background(), llm.Request{
+		Messages:        []llm.Message{llm.UserText("hi")},
+		ReasoningEffort: "high",
+	})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if rt.reqBody == nil {
+		t.Fatal("no request body captured")
+	}
+	var sent map[string]any
+	if err := json.Unmarshal(rt.reqBody, &sent); err != nil {
+		t.Fatalf("decode request body: %v", err)
+	}
+	if got := sent["reasoning_effort"]; got != "high" {
+		t.Errorf("reasoning_effort = %v, want %q (body: %s)", got, "high", rt.reqBody)
+	}
+}

docs/adr/0027-qwen-builtin.md

@@ -0,0 +1,102 @@
+# ADR-0027: Qwen (Alibaba) built-in provider — OpenAI-compat, not Anthropic-compat
+
+**Status:** Accepted — 2026-08-12
+
+## Context
+
+Alibaba's Qwen models (`qwen3.8-max`, `qwen3.7-plus`, the `qwen3-vl-*` vision
+variants, …) are served from Model Studio / DashScope, and mort wants them as a
+first-class failover tier with a dedicated `QWEN_API_KEY` — the same ergonomics
+ADR-0026 gave Kimi.
+
+Unlike Kimi, Model Studio exposes the same models over **two** protocols:
+
+| | OpenAI-compatible | Anthropic-compatible |
+|---|---|---|
+| Base URL | `https://dashscope-intl.aliyuncs.com/compatible-mode/v1` | `https://dashscope-intl.aliyuncs.com/apps/anthropic` |
+| Endpoints | full Chat Completions surface | `/v1/messages` only (no `/v1/models`) |
+| Purpose | the documented developer API | a shim, documented around hosting Claude Code |
+
+So the question this ADR answers is not "which client do we reuse" but
+"which of Alibaba's two wire protocols does the built-in speak".
+
+## Decision
+
+**The `qwen` built-in and the `qwen://` DSN scheme speak OpenAI-compat**, over
+`provider/openai` — no new package, mirroring ADR-0026 (kimi) and ADR-0015
+(llama-swap chat). Default base URL is the international host; the China host
+(`dashscope.aliyuncs.com/compatible-mode/v1`) and workspace-scoped regional
+hosts are reachable with a `qwen://` DSN.
+
+Credential handling is copied from kimi verbatim, because both of its rules
+are load-bearing: `WithAPIKey` is passed unconditionally (even empty) so an
+unset `QWEN_API_KEY` can never fall through to `openai.New`'s `OPENAI_API_KEY`
+default, and `WithAPIKeyName("QWEN_API_KEY")` makes the synthetic-401 hint name
+the variable the operator actually has to set.
+
+The kimi and qwen DSN factories were identical, so they now share one
+`openaiCompatScheme` helper — the next OpenAI-compat built-in gets the
+credential and key-hint rules by construction rather than by copy.
+
+### Why not the Anthropic-compatible endpoint
+
+Every concrete difference favors OpenAI-compat *for this codebase*:
+
+- **Reasoning survives the trip.** Model Studio takes `reasoning_effort` as a
+  top-level field on the OpenAI surface, which `provider/openai` already sends
+  — `llm.WithReasoningEffort` works on qwen with zero qwen-specific code
+  (`TestQwenReasoningEffortReachesWire` asserts it on the wire). Down the
+  anthropic client it would be dropped in silence: `provider/anthropic`
+  deliberately ignores `Request.ReasoningEffort`, because first-party Claude
+  has no such knob.
+- **Structured output would regress.** `provider/anthropic` implements
+  `Request.Schema` with the first-party GA `output_config.format` mechanism.
+  Alibaba's shim does not implement it; a compat endpoint that ignores an
+  unknown field returns unconstrained prose while still reporting success.
+  The OpenAI path sends `response_format: json_schema`, which Model Studio
+  supports natively on the Max/Plus families.
+- **Cache accounting already lands.** Model Studio's implicit prefix cache
+  reports hits in `usage.prompt_tokens_details.cached_tokens`, which the openai
+  client already maps to `llm.Usage.CacheReadTokens`. The anthropic client
+  reads `cache_read_input_tokens`, a field the shim has no reason to emit.
+- **Thinking content is discarded on the anthropic path anyway.**
+  `provider/anthropic` skips `thinking` blocks in both the buffered and
+  streaming decoders, so the shim's headline feature — first-class
+  `thinking: {type: "enabled", budget_tokens: N}` — buys majordomo nothing
+  today.
+- **Smaller blast radius.** The anthropic client has no `WithAPIKeyName`
+  option, so a keyless qwen would tell the operator to set `ANTHROPIC_API_KEY`;
+  fixing that means changing the first-party Anthropic client to serve a
+  third-party shim.
+- **It is the less-exercised surface.** The Anthropic endpoint is documented as
+  Messages-only, with a temperature range that differs from Anthropic's own
+  ([0, 2) vs [0.0, 1.0]) — i.e. it is Qwen semantics wearing an Anthropic
+  envelope, not an Anthropic-equivalent target.
+
+The one thing the Anthropic surface offers that OpenAI-compat does not is
+explicit `cache_control` breakpoints reached through `Request.PromptCache`.
+That is not a reason to route Qwen through it: Model Studio's implicit cache is
+automatic and already metered, and if explicit breakpoints ever matter they
+belong in `provider/openai` (Model Studio accepts `cache_control` on content
+blocks there too), where every OpenAI-compat target would get them.
+
+## Consequences
+
+- `qwen/<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, structured output, reasoning effort, and cached-token
+  accounting all ride the openai client and inherit its fixes.
+- Image *inputs* work at the client level, but only the `qwen-vl-*` /
+  `qwen3-vl-*` models accept them (matrix footnote ⁴; ³ is kimi's).
+- Two model-side quirks are Alibaba's, not majordomo's, and are left to the
+  caller rather than papered over: thinking is **on by default** on some models
+  (e.g. `qwen3.7-plus`), and Qwen3 *open-source* models require streaming when
+  thinking is enabled — a buffered `Generate` against one of those needs a
+  model that supports non-streaming thinking (the Max/Plus families do).
+- If a future consumer genuinely needs the Anthropic surface, it is reachable
+  today without library changes:
+  `LLM_QWEN_ANTHROPIC=anthropic://[email protected]/apps/anthropic`
+  — with the reasoning/structured-output caveats above.
+- Second third-party built-in after kimi. The ADR-0026 bar still holds: a named
+  consumer needs it in-config. `RegisterProvider`/`LLM_*` remain the path for
+  everything else.

docs/adr/README.md

@@ -30,3 +30,4 @@ One decision per file, append-only; supersede rather than rewrite.
 | [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 |
+| [0027](0027-qwen-builtin.md) | Qwen (Alibaba) built-in provider — OpenAI-compat, not Model Studio's Anthropic-compat endpoint | Accepted |

env.go

@@ -26,8 +26,9 @@ 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", "kimi", "anthropic", "google"/"gemini", or
-	// any custom scheme registered with RegisterScheme.
+	// "ollama-cloud", "openai", "kimi", "qwen", "anthropic",
+	// "google"/"gemini", "llama-swap"/"llama-swaps", or any custom scheme
+	// registered with RegisterScheme.
 	Scheme string
 	// Token is the provider secret (bearer token or API key); empty = none.
 	Token string
@@ -40,6 +41,19 @@ type DSN struct {
 // env-defined providers always speak TLS).
 func (d DSN) BaseURL() string { return "https://" + d.Host }
 
+// envKeyForProvider returns the LLM_* variable that defines the provider named
+// name: "m1" → LLM_M1, "my-prov" → LLM_MY_PROV.
+//
+// This is the single definition on purpose. Two call sites need byte-identical
+// output and would drift apart in silence: lazy resolution reads this variable
+// to find an unregistered provider, and openaiCompatScheme names it in the
+// missing-key hint so a keyless DSN target tells the operator which variable to
+// set. Those two were separate copies with a comment asserting they matched —
+// a comment is not enforcement, this function is.
+func envKeyForProvider(name string) string {
+	return "LLM_" + strings.ToUpper(strings.ReplaceAll(name, "-", "_"))
+}
+
 // ParseDSN parses a raw DSN string. The algorithm matches go-llm exactly:
 // split on "://", then an optional "@" separates the token from the host;
 // trailing slashes on the host are trimmed.

imagegen/faceswap.go

@@ -55,6 +55,14 @@ type DetectedFace struct {
 	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
@@ -64,6 +72,57 @@ 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 {

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

progress.md

@@ -285,3 +285,39 @@ tests flush out.
   (footnote ³), `.env.example`, ADR-0026 (+ index; also backfilled the missing
   0024/0025 index rows).
 - Consumer: mort names Kimi as a failover tier.
+
+## 2026-08-12 — Qwen (Alibaba) built-in provider (ADR-0027)
+
+- New built-in `qwen` provider + `qwen://` DSN scheme over Alibaba Model
+  Studio's OpenAI-compatible mode, reusing `provider/openai` (no new client,
+  mirrors kimi/llama-swap). Default base URL
+  `https://dashscope-intl.aliyuncs.com/compatible-mode/v1`; China/regional
+  hosts via `LLM_QCN=qwen://[email protected]/compatible-mode/v1`.
+- Credential is `QWEN_API_KEY` (via the registry's injected envLookup).
+  `WithAPIKey` passed unconditionally so an unset key cannot fall through to
+  `OPENAI_API_KEY`; `WithAPIKeyName` names `QWEN_API_KEY` in the 401 hint.
+- **Chose OpenAI-compat over Model Studio's Anthropic-compatible
+  `/apps/anthropic` shim** (ADR-0027): on the anthropic client
+  `ReasoningEffort` is ignored by design, `Request.Schema` rides
+  `output_config.format` (which the shim does not implement), and cached-token
+  accounting reads Anthropic-only usage fields. The shim is still reachable
+  ad hoc via an `anthropic://` DSN.
+- `registerOpenAICompatBuiltin` installs BOTH halves of an OpenAI-compat
+  built-in (eager provider + `name://` DSN scheme via the shared
+  `openaiCompatScheme`), so the two credential rules — unconditional
+  `WithAPIKey`, and `WithAPIKeyName` naming that same variable — hold by
+  construction. kimi and qwen are one line each.
+- `envKeyForProvider` is the single definition of the `LLM_<NAME>` form,
+  shared by lazy resolution (`registry.go`) and the DSN missing-key hint. They
+  were separate copies with a comment asserting they matched.
+- The shared contract is ONE table (`builtin_openaicompat_test.go`), run
+  identically for every OpenAI-compat built-in: endpoint + bearer, missing key
+  fails closed naming its own variable with no network hit, the `name://` DSN
+  reaching another host, and a keyless DSN naming `LLM_<NAME>` rather than the
+  built-in's key. Adding a built-in is a table row that immediately owes all
+  four; `builtin_kimi_test.go` was retired into it. Qwen-only tests: the
+  reverse credential leak, and `reasoning_effort` asserted on the wire body
+  (the ADR's load-bearing claim).
+- Docs in sync: README built-in table + Qwen paragraph + DSN scheme list +
+  support matrix (footnote ⁴), `.env.example`, ADR-0027 (+ index).
+- Consumer: mort wants Qwen as a failover tier.

provider/llamaswap/audio.go

@@ -338,27 +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/faceswap.go

@@ -41,11 +41,12 @@ type faceSwapModel struct {
 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"`
+		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"`
 }
 
@@ -74,7 +75,7 @@ func (m *faceSwapModel) ListFaces(ctx context.Context, img imagegen.Image) ([]im
 	}
 	out := make([]imagegen.DetectedFace, 0, len(parsed.Faces))
 	for _, f := range parsed.Faces {
-		df := imagegen.DetectedFace{Index: f.Index, Score: f.Score}
+		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.
@@ -126,10 +127,11 @@ func (m *faceSwapModel) FaceSwap(ctx context.Context, req imagegen.FaceSwapReque
 	if err != nil {
 		return nil, err
 	}
-	raw, respType, err := m.p.doRaw(ctx, http.MethodPost, path, m.id, contentType, body, maxFaceSwapResponseBytes)
+	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"}
 	}
@@ -151,16 +153,66 @@ func (m *faceSwapModel) FaceSwap(ctx context.Context, req imagegen.FaceSwapReque
 	if hdr := mimeFromContentType(respType, "image/"); hdr != "" {
 		mimeType = hdr
 	}
-	return &imagegen.Result{Images: []llm.ImagePart{{MIME: mimeType, Data: raw}}}, nil
+	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.
+// ("target"/"source", "frame"/"frame_last") so a log line says which one was
+// malformed — and, for the video keyframes, so a backend that stages uploads
+// by filename cannot have the second overwrite the first.
 //
-// initImageFilename (video.go) is this function with base fixed to "frame"
-// and delegates here — two copies of one extension table is how they drift.
+// Every caller routes through here: two copies of one extension table is how
+// they drift.
 func imageFilename(mimeType, base string) string {
 	if base == "" {
 		base = "image"

provider/llamaswap/faceswap_test.go

@@ -259,3 +259,86 @@ func TestFaceSwapAllowsNegativeIndexUnderAll(t *testing.T) {
 		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/lipsync.go

@@ -56,7 +56,7 @@ func (m *lipsyncModel) Lipsync(ctx context.Context, req videogen.LipsyncRequest,
 	// hand (mirrors videoModel.Generate).
 	var buf bytes.Buffer
 	w := multipart.NewWriter(&buf)
-	fw, err := w.CreateFormFile("image", initImageFilename(req.Image.MIME))
+	fw, err := w.CreateFormFile("image", imageFilename(req.Image.MIME, "frame"))
 	if err != nil {
 		return nil, fmt.Errorf("llama-swap: build lipsync form: %w", err)
 	}

provider/llamaswap/video.go

@@ -38,10 +38,13 @@ type videoModel struct {
 // bound the call with a context deadline.
 //
 // Parameter names follow vLLM-Omni's videos API (num_frames, fps,
-// num_inference_steps, guidance_scale); the conditioning frame is sent as an
-// `input_reference` file part, following OpenAI's videos API. Upstreams
+// num_inference_steps, guidance_scale); the leading conditioning frame is sent
+// as an `input_reference` file part, following OpenAI's videos API, and a
+// trailing keyframe (Request.LastImage) as `input_reference_last`. Upstreams
 // ignore fields they don't understand, and optional fields stay off the wire
-// entirely so the model's own defaults apply.
+// entirely so the model's own defaults apply — which is also why a backend
+// without first-last-frame support returns an ordinary clip here rather than
+// an error.
 func (m *videoModel) Generate(ctx context.Context, req videogen.Request, opts ...videogen.Option) (*videogen.Result, error) {
 	req = req.Apply(opts...)
 	if strings.TrimSpace(req.Prompt) == "" {
@@ -56,6 +59,9 @@ func (m *videoModel) Generate(ctx context.Context, req videogen.Request, opts ..
 	if req.InitImage != nil && len(req.InitImage.Data) == 0 {
 		return nil, fmt.Errorf("%w: video init image has no bytes", llm.ErrUnsupported)
 	}
+	if req.LastImage != nil && len(req.LastImage.Data) == 0 {
+		return nil, fmt.Errorf("%w: video last image has no bytes", llm.ErrUnsupported)
+	}
 	width, height, err := parseSize(req.Size)
 	if err != nil {
 		return nil, fmt.Errorf("%w: %v", llm.ErrUnsupported, err)
@@ -83,12 +89,19 @@ func (m *videoModel) Generate(ctx context.Context, req videogen.Request, opts ..
 		return nil, err
 	}
 	if req.InitImage != nil {
-		fw, err := w.CreateFormFile("input_reference", initImageFilename(req.InitImage.MIME))
-		if err != nil {
-			return nil, fmt.Errorf("llama-swap: build video form: %w", err)
+		if err := writeImagePart(w, "input_reference", "frame", req.InitImage); err != nil {
+			return nil, err
 		}
-		if _, err := fw.Write(req.InitImage.Data); err != nil {
-			return nil, fmt.Errorf("llama-swap: build video form: %w", err)
+	}
+	// The trailing keyframe rides a SEPARATE part rather than a second
+	// `input_reference`: multipart permits repeated names, but the receiving
+	// end would then have to rely on part ORDER to tell first from last, and
+	// an ordering contract that is invisible in the field name is one nobody
+	// can see they have broken. A backend that does not know the name ignores
+	// the part, which is the same degradation as any other unknown field.
+	if req.LastImage != nil {
+		if err := writeImagePart(w, "input_reference_last", "frame_last", req.LastImage); err != nil {
+			return nil, err
 		}
 	}
 	if err := w.Close(); err != nil {
@@ -134,11 +147,27 @@ func singleVideoResult(provider, model, verb string, raw []byte, contentType str
 	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 {
-	return imageFilename(mimeType, "frame")
+// writeImagePart attaches one conditioning frame under the given field name,
+// with a filename derived from nameStem. Shared by the first- and last-frame
+// parts so the two cannot drift in how they encode, which is the usual way a
+// second copy of a block goes wrong.
+//
+// The two frames MUST carry DISTINCT filenames, not merely distinct field
+// names. Backends commonly stage an uploaded frame under a name derived from
+// the filename — our own ComfyUI shim posts to /upload/image with
+// overwrite=true — so two parts sharing "frame.png" would have the second
+// clobber the first, and BOTH keyframe inputs would then resolve to the same
+// stored image. The clip would render clean, pinned at both ends to the same
+// frame, with nothing anywhere reporting a problem.
+func writeImagePart(w *multipart.Writer, field, nameStem string, img *videogen.Image) error {
+	fw, err := w.CreateFormFile(field, imageFilename(img.MIME, nameStem))
+	if err != nil {
+		return fmt.Errorf("llama-swap: build video form: %w", err)
+	}
+	if _, err := fw.Write(img.Data); err != nil {
+		return fmt.Errorf("llama-swap: build video form: %w", err)
+	}
+	return nil
 }
 
 // formatInt renders an optional int pointer for a form field; nil = "" (omit).

provider/llamaswap/video_test.go

@@ -223,3 +223,134 @@ func TestVideoGenerateNonVideoBodyErrors(t *testing.T) {
 		t.Errorf("message = %q, want mention of non-video body", apiErr.Message)
 	}
 }
+
+// Both keyframes reach the wire, under DISTINCT field names.
+//
+// The distinct-name property is the actual contract with the backend shim: the
+// two frames could have shared one repeated `input_reference` name, and then
+// which is first and which is last would depend on multipart part ORDER — an
+// ordering contract invisible in the payload, that nothing would notice
+// breaking. Asserting the names is what pins it.
+func TestVideoGenerateSendsBothKeyframes(t *testing.T) {
+	var gotFirst, gotLast []byte
+	var firstName, lastName string
+	var sawLastPart bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(32 << 20); err != nil {
+			t.Errorf("parse form: %v", err)
+			return
+		}
+		if f, hdr, err := r.FormFile("input_reference"); err == nil {
+			gotFirst, _ = io.ReadAll(f)
+			firstName = hdr.Filename
+			f.Close()
+		}
+		if f, hdr, err := r.FormFile("input_reference_last"); err == nil {
+			sawLastPart = true
+			gotLast, _ = io.ReadAll(f)
+			lastName = hdr.Filename
+			f.Close()
+		}
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write([]byte("fake-mp4-bytes"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vm, err := p.VideoModel("videogen-minimax-h3")
+	if err != nil {
+		t.Fatalf("VideoModel: %v", err)
+	}
+
+	first, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+	last := append(append([]byte{}, first...), 0x00) // distinguishable from first
+
+	if _, err := vm.Generate(context.Background(), videogen.Request{
+		Prompt:    "a cat surfing",
+		InitImage: &videogen.Image{MIME: "image/png", Data: first},
+		LastImage: &videogen.Image{MIME: "image/png", Data: last},
+	}); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+
+	if !sawLastPart {
+		t.Fatal("input_reference_last was not sent — a pinned end frame would be silently dropped")
+	}
+	if string(gotFirst) != string(first) {
+		t.Errorf("input_reference = %d bytes, want %d", len(gotFirst), len(first))
+	}
+	if string(gotLast) != string(last) {
+		t.Errorf("input_reference_last = %d bytes, want %d", len(gotLast), len(last))
+	}
+	// The two must not be the same bytes, or a swap/aliasing bug reads as a pass.
+	if string(gotFirst) == string(gotLast) {
+		t.Error("both parts carry identical bytes — the frames are being aliased")
+	}
+	// DISTINCT FILENAMES, not just distinct field names. Backends stage an
+	// uploaded frame under a name derived from the filename (our ComfyUI shim
+	// posts to /upload/image with overwrite=true), so two parts sharing
+	// "frame.png" would have the second clobber the first and BOTH keyframes
+	// would resolve to the same stored image — a clip pinned at both ends to
+	// the same frame, rendering cleanly with nothing reporting a fault.
+	if firstName == "" || lastName == "" {
+		t.Fatalf("filenames = %q / %q, want both set", firstName, lastName)
+	}
+	if firstName == lastName {
+		t.Errorf("both parts use filename %q — the second upload would clobber the first", firstName)
+	}
+}
+
+// LastImage alone (no InitImage) is a legitimate request: pin the destination
+// and let the model invent the approach. It must not require a first frame.
+func TestVideoGenerateLastImageAloneIsAllowed(t *testing.T) {
+	var sawFirst, sawLast bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(32 << 20); err != nil {
+			t.Errorf("parse form: %v", err)
+			return
+		}
+		if f, _, err := r.FormFile("input_reference"); err == nil {
+			sawFirst = true
+			f.Close()
+		}
+		if f, _, err := r.FormFile("input_reference_last"); err == nil {
+			sawLast = true
+			f.Close()
+		}
+		w.Header().Set("Content-Type", "video/mp4")
+		_, _ = w.Write([]byte("fake-mp4-bytes"))
+	}))
+	defer srv.Close()
+
+	p := New(WithBaseURL(srv.URL), WithHTTPClient(srv.Client()))
+	vm, _ := p.VideoModel("videogen-minimax-h3")
+	frame, _ := base64.StdEncoding.DecodeString(onePixelPNG)
+
+	if _, err := vm.Generate(context.Background(),
+		videogen.Request{Prompt: "arrive here"},
+		videogen.WithLastImage(videogen.Image{MIME: "image/png", Data: frame}),
+	); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if sawFirst {
+		t.Error("input_reference sent, want omitted")
+	}
+	if !sawLast {
+		t.Error("input_reference_last omitted, want sent")
+	}
+}
+
+// An empty LastImage is rejected before the request is built, matching
+// InitImage's existing contract — a zero-byte frame reaching the backend is a
+// confusing upstream error instead of a clear local one.
+func TestVideoGenerateRejectsEmptyLastImage(t *testing.T) {
+	p := New(WithBaseURL("http://unused"))
+	vm, _ := p.VideoModel("videogen-minimax-h3")
+	_, err := vm.Generate(context.Background(), videogen.Request{
+		Prompt:    "x",
+		LastImage: &videogen.Image{MIME: "image/png"},
+	})
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want llm.ErrUnsupported", err)
+	}
+}

registry.go

@@ -263,7 +263,7 @@ func (r *Registry) providerFor(name string) (llm.Provider, error) {
 		return nil, envErr
 	}
 
-	envKey := "LLM_" + strings.ToUpper(strings.ReplaceAll(name, "-", "_"))
+	envKey := envKeyForProvider(name)
 	envVal := r.envLookup(envKey)
 	if envVal == "" {
 		return nil, fmt.Errorf("%w: %q (checked registry and %s env var)", ErrUnknownProvider, name, envKey)

videogen/videogen.go

@@ -12,6 +12,8 @@
 // InitImage is a pure text prompt, a non-nil InitImage conditions generation
 // on that frame. Hybrid models (e.g. Wan 2.2 TI2V) serve both from the same
 // checkpoint, so unlike imagegen there is no separate Editor-style interface.
+// LastImage extends the same surface to the other end of the clip, so one
+// Request covers t2v, i2v, and first-last-frame-to-video without a mode flag.
 //
 // The first implementation is provider/llamaswap, which targets the blocking
 // OpenAI/vLLM-Omni-style POST /v1/videos/sync endpoint: the response body is
@@ -51,6 +53,19 @@ type Request struct {
 	// nil = pure text-to-video.
 	InitImage *Image
 
+	// LastImage conditions generation on an ENDING frame. With InitImage it
+	// pins both ends (first-last-frame-to-video); alone it pins only the
+	// destination and lets the backend invent the approach.
+	//
+	// Support is per-model and NOT advertised anywhere in this contract: a
+	// backend that does not understand a trailing keyframe ignores it and
+	// returns an ordinary clip, which is indistinguishable from success.
+	// There is no capability bit to consult, because the contract has no way
+	// to learn one. A caller that needs to know whether the pin actually took
+	// effect must establish that out of band — by configuration it controls,
+	// not by inspecting the result.
+	LastImage *Image
+
 	// Size is the requested resolution, e.g. "1280x704"; "" = backend default.
 	Size string
 
@@ -92,6 +107,10 @@ type Option func(*Request)
 // WithInitImage conditions generation on a starting frame (image-to-video).
 func WithInitImage(img Image) Option { return func(r *Request) { r.InitImage = &img } }
 
+// WithLastImage conditions generation on an ending frame. Combined with
+// WithInitImage this pins both ends of the clip.
+func WithLastImage(img Image) Option { return func(r *Request) { r.LastImage = &img } }
+
 // WithSize sets the requested resolution (e.g. "1280x704").
 func WithSize(size string) Option { return func(r *Request) { r.Size = size } }