fix(agent): recover front-loaded answer past a citations-only terminal turn (#11)

finalOutput now recovers the front-loaded answer when the terminal turn is a
sources/citations-only addendum ("Sources: [x](url), ..."), not just when it is
empty or a back-reference. It recovers the prior substantive answer and appends
the (real) citations below it. Guards: citation-DOMINANCE (a prose answer that
merely opens with "Source: ... http://..." is left as the answer), ^-anchored
heading, citations recovery decoupled from the terminal-length ratio (concise
answers recover too), preamble filter applied only in the borderline band
(long answers opening with "Sure,"/"Let me" are not vetoed), and a dedup that
ignores <url> angle-bracket wrappers. Healthy terminal answers unchanged; zero
extra model calls.

Fixes mort #1418. Gadfly-reviewed (6 reviewers) + adversarially pre-verified;
all findings graded, real ones addressed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitea/workflows/adversarial-review.yml

@@ -41,7 +41,7 @@ jobs:
     # Tracks gadfly's v1 release tag — a curated pointer re-moved on each release
     # (unlike @main, which moves on every push). Central swarm tuning propagates
     # here automatically; the tradeoff vs a full sha pin is that v1 is mutable.
-    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@7bc3c982fa7b72367034c673f7812bf05e9c503e
+    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@5007597cf921dc3f0a83c708878facfe65fd8e8b
     # Least privilege: forward only the review secrets (not `secrets: inherit`,
     # which would expose every repo secret). GITEA_TOKEN is the automatic token.
     secrets:

agent/finalize.go

@@ -14,21 +14,55 @@ 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 trivial pointer such as "(Already answered above.)". Returning only
+// with a degenerate terminal turn that is not itself the answer. Two shapes are
-// the terminal text would discard the real answer, which is still present
+// recovered from the transcript (zero extra model calls):
-// earlier in the transcript. When the terminal text is weak (empty, or a short
+//
-// back-reference) fall back to the last substantive assistant content in msgs.
+//   - a trivial back-reference ("(Already answered above.)", "see above", …):
+//     the real answer sits earlier, so recover it and DISCARD the worthless
+//     closer.
+//   - a sources/citations-only addendum ("Sources: [x](…), [y](…)"): the model
+//     front-loaded the prose answer and closed with just its citations (the
+//     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 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).
 //
 // 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 {
-	if !isWeakFinal(terminal) {
+	citations := isCitationsOnly(terminal)
+	if !citations && !isWeakFinal(terminal) {
 		return terminal
 	}
-	if rec, ok := lastSubstantiveAssistantText(msgs, terminal); ok {
+	rec, ok := lastSubstantiveAssistantText(msgs, terminal, citations)
+	if !ok {
+		return terminal
+	}
+	if citations {
+		// 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
+		// containment test ignores <url> angle-bracket wrappers so a turn that
+		// listed the same sources unwrapped still suppresses the duplicate.
+		if tail := strings.TrimSpace(terminal); !strings.Contains(stripURLAngles(rec), stripURLAngles(tail)) {
+			return rec + "\n\n" + tail
+		}
+	}
 	return rec
+}
+
+// stripURLAngles removes the <…> wrappers Discord uses to suppress link embeds,
+// so the citations dedup compares URLs regardless of that formatting delta.
+func stripURLAngles(s string) string {
+	if !strings.ContainsAny(s, "<>") {
+		return s
 	}
-	return terminal
+	return strings.NewReplacer("<", "", ">", "").Replace(s)
 }
 
 // backRefRe matches a terminal turn that merely points back to an earlier
@@ -40,19 +74,46 @@ var backRefRe = regexp.MustCompile(`(?i)(already answered|see above|as (i )?(sai
 // 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)`)
 
+// citationLabelRe matches a terminal turn that OPENS with a sources/citations
+// heading — the shape a model produces when it front-loads its answer into an
+// earlier tool-call turn and closes with only its sources. Leading markdown
+// emphasis (*, _), list (-, +, *), block-quote (>), and ATX-heading (#) markers
+// — with their whitespace, since \s is in the class — are tolerated before the
+// label, as is closing emphasis (** / __) plus whitespace between the label and
+// 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*_]*[:：\-—]`)
+
+// 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
+// links out when measuring how citation-dominated the terminal is.
+var linkRe = regexp.MustCompile(`\[[^\]]*\]\([^)]*\)|https?://\S+`)
+
+// citationResidueCutset is trimmed from the ends of a citations terminal's
+// non-link remainder before measuring it — list bullets, separators, and the
+// short per-source annotations models add in parentheses.
+const citationResidueCutset = " \t\r\n,;.:|·•*_()[]—-"
+
 const (
 	// weakFinalMaxChars bounds how long a back-reference closer can be. A
 	// genuine final answer that merely contains "as I said" mid-sentence is
 	// longer than this, so it is never treated as weak.
 	weakFinalMaxChars = 120
 	// recoverMinChars: a prior assistant turn this long is treated as a real
-	// answer regardless of how it opens.
+	// answer regardless of how it opens (the preamble filter is not applied at
+	// this length — see isSubstantiveAnswer).
 	recoverMinChars = 200
 	// recoverFloorChars / recoverRatio gate the borderline band: a shorter
-	// prior turn must still clearly dwarf the (very short) terminal and not
+	// prior turn must clear the floor and — unless the terminal is a citations
-	// look like a preamble.
+	// addendum, which is not a rival answer — also clearly dwarf the (very
+	// short) terminal. See isSubstantiveAnswer.
 	recoverFloorChars = 80
 	recoverRatio      = 3
+	// citationDominatedDivisor: a citations terminal's non-link remainder must
+	// 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
 )
 
 // isWeakFinal reports whether a terminal turn's text fails to stand on its own
@@ -65,10 +126,37 @@ func isWeakFinal(s string) bool {
 	return len(t) <= weakFinalMaxChars && backRefRe.MatchString(t)
 }
 
+// isCitationsOnly reports whether a terminal turn is essentially just a
+// sources/citations addendum: it OPENS with a citations heading, carries at
+// least one link, and — once the heading and links are removed — is dominated
+// by that citation structure (only list punctuation and short per-source
+// annotations remain). The dominance check is what separates a bare sources
+// list (recover the front-loaded answer, keep the links) from a real prose
+// answer that merely opens with "Source:" and references a URL mid-sentence
+// (leave it as the answer). Unlike a back-reference closer the links are worth
+// keeping, so finalOutput appends them to the recovered answer.
+//
+// A citations terminal whose sources are bare domains (no scheme, no markdown
+// link) is intentionally out of scope — there is no reliable link signal, so it
+// is left as-is rather than risk misclassifying prose.
+func isCitationsOnly(s string) bool {
+	t := strings.TrimSpace(s)
+	if !citationLabelRe.MatchString(t) {
+		return false
+	}
+	body := citationLabelRe.ReplaceAllString(t, "")
+	if !linkRe.MatchString(body) {
+		return false
+	}
+	residue := strings.Trim(linkRe.ReplaceAllString(body, ""), citationResidueCutset)
+	return len(residue) <= len(t)/citationDominatedDivisor
+}
+
 // 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. Returns ("", false) when nothing qualifies.
+// reads like a real answer. citations selects the recovery bar (see
-func lastSubstantiveAssistantText(msgs []llm.Message, terminal string) (string, bool) {
+// isSubstantiveAnswer). Returns ("", false) when nothing qualifies.
+func lastSubstantiveAssistantText(msgs []llm.Message, terminal string, citations bool) (string, bool) {
 	tt := strings.TrimSpace(terminal)
 	for i := len(msgs) - 1; i >= 0; i-- {
 		m := msgs[i]
@@ -79,7 +167,7 @@ func lastSubstantiveAssistantText(msgs []llm.Message, terminal string) (string,
 		if txt == "" || txt == tt {
 			continue // the terminal turn itself, or an empty tool-only turn
 		}
-		if isSubstantiveAnswer(txt, tt) {
+		if isSubstantiveAnswer(txt, tt, citations) {
 			return txt, true
 		}
 	}
@@ -88,11 +176,21 @@ func lastSubstantiveAssistantText(msgs []llm.Message, terminal string) (string,
 
 // isSubstantiveAnswer reports whether txt (a prior assistant turn) reads like a
 // real answer rather than a preamble, relative to the terminal text.
-func isSubstantiveAnswer(txt, terminal string) bool {
+//
+// 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 {
 	if len(txt) >= recoverMinChars {
 		return true
 	}
-	return len(txt) >= recoverFloorChars &&
+	if len(txt) < recoverFloorChars || preambleRe.MatchString(txt) {
-		len(txt) >= recoverRatio*len(terminal) &&
+		return false
-		!preambleRe.MatchString(txt)
+	}
+	return citations || len(txt) >= recoverRatio*len(terminal)
 }

agent/finalize_test.go

@@ -36,6 +36,42 @@ func TestIsWeakFinal(t *testing.T) {
 	}
 }
 
+func TestIsCitationsOnly(t *testing.T) {
+	cases := []struct {
+		name string
+		in   string
+		want bool
+	}{
+		{"sources-md-links", "Sources: [pcprice.watch](https://pcprice.watch/x), [ebay](https://ebay.com/1).", true},
+		{"lowercase-bare-url", "sources: see https://example.com/a", true},
+		{"bold-label-colon-inside", "**Sources:** [a](https://a), [b](https://b)", true},
+		{"bold-label-colon-outside", "**Sources**: [a](https://a), [b](https://b)", true}, // colon after the closing **
+		{"references-dash", "References — [a](https://a)", true},
+		{"citations-label", "Citations: https://x/y", true},
+		{"leading-list-marker", "- Sources: [a](https://a)", true},
+		{"atx-heading", "## Sources: [a](https://a), [b](https://b)", true}, // ATX heading marker + its trailing space
+		{"further-reading", "Further reading: https://example.com/deep-dive", true},
+		{"annotated-multi-source", "Sources: [pcprice.watch](https://a) (tracker), [eBay](https://b) (sold), [bestvaluegpu](https://c) (retail), [resaleprices](https://d) (asking).", true}, // the reported issue-1418 shape
+		{"backref-plus-links-is-citations", "References: as noted above, [pcprice.watch](https://pcprice.watch/x).", true},                                                                   // a back-ref phrase inside a real sources list is still citations
+
+		{"empty", "", false},
+		{"label-but-no-link", "Source: internal analysis, no URL here", false},
+		{"prose-then-sources", "It sells for ~$2,700. Sources: [a](https://a)", false}, // answer first → not a pure addendum
+		{"source-led-prose-answer", "Source: According to https://cdc.gov the flu vaccine is 40-60% effective, and the CDC recommends annual vaccination for everyone over six months old.", false}, // a prose answer that merely opens with a "Source:" label
+		{"mentions-sources-midsentence", "The sources of the leak were never confirmed.", false},
+		{"link-without-label", "Here is the link you asked for: [a](https://a)", false},
+		{"bare-domains-out-of-scope", "Sources: pcprice.watch (used ~$200), ebay.com (sold listings)", false}, // bare domains: no scheme or markdown link to key on
+		{"crisp-number", "42", false},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := isCitationsOnly(c.in); got != c.want {
+				t.Errorf("isCitationsOnly(%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 != "" {
@@ -48,6 +84,25 @@ 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
+	// 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.
+	sources := "Sources: [pcprice.watch](https://pcprice.watch/x), [ebay](https://www.ebay.com/itm/1)."
+	answerWithSources := longAnswer + "\n\n" + sources
+	// A concise (>80, <200 byte) front-loaded answer + a long citations terminal:
+	// the ratio arm can't be met against the long terminal, so citations mode
+	// must fall back to the floor.
+	conciseAnswer := "It sells for about $2,700 used on eBay, typically $2,400 to $2,900 depending on condition and bundle."
+	longSources := "Sources: [pcprice.watch](https://pcprice.watch/gpu/rtx5090) (tracker), [ebay](https://www.ebay.com/sch/rtx5090) (sold), [newegg](https://newegg.com/rtx5090) (retail), [pcpartpicker](https://pcpartpicker.com/rtx5090) (history)."
+	// A substantive answer that merely OPENS with "Source:" (not a bare list).
+	sourceLedAnswer := "Source: https://nvd.nist.gov/vuln/detail/CVE-2024-1234 — this is the authoritative NVD entry for the vulnerability, rated CVSS 9.8 critical."
+	// A borderline-band (80–200 byte) turn that opens like a planning preamble:
+	// it clears the floor, but the preamble filter still vetoes it (the filter
+	// applies only in the borderline band; a >=200-byte turn is accepted as-is).
+	preambleTurn := "Let me look that up across a few different sites and then compile the full comparison for you here."
+	// 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."
 
 	tests := []struct {
 		name     string
@@ -110,6 +165,97 @@ func TestFinalOutput(t *testing.T) {
 			terminal: "(see above)",
 			want:     "(see above)", // preamble excluded; falls back to terminal
 		},
+		{
+			name: "citations-only terminal recovers front-loaded answer and keeps sources",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(longAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(sources),
+			},
+			terminal: sources,
+			want:     answerWithSources, // answer recovered, citations appended
+		},
+		{
+			name: "citations-only terminal but only a preamble prior: keeps the sources",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst("Let me gather the sources.", cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(sources),
+			},
+			terminal: sources,
+			want:     sources, // nothing substantive to recover → keep the addendum
+		},
+		{
+			name: "citations already in the recovered answer are not duplicated",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(answerWithSources, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(sources),
+			},
+			terminal: sources,
+			want:     answerWithSources, // recovered turn already carries the sources
+		},
+		{
+			// #1418 persisted for CONCISE answers: a <200-char front-loaded
+			// answer must still be recovered against a long citations terminal
+			// (the ratio arm is skipped in citations mode).
+			name: "concise front-loaded answer recovered against a long citations terminal",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(conciseAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(longSources),
+			},
+			terminal: longSources,
+			want:     conciseAnswer + "\n\n" + longSources,
+		},
+		{
+			// A substantive answer that merely OPENS with "Source:" and cites a
+			// URL mid-sentence is NOT a citations addendum — return it verbatim,
+			// never prepend the prior planning turn.
+			name: "source-led substantive answer is not hijacked by a prior turn",
+			msgs: []llm.Message{
+				llm.UserText("what's the authoritative URL?"),
+				asst("I'll look up the CVE in the NVD database, cross-reference the vendor advisory, and confirm the canonical URL before I answer.", cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(sourceLedAnswer),
+			},
+			terminal: sourceLedAnswer,
+			want:     sourceLedAnswer,
+		},
+		{
+			// A borderline-length turn that opens like a preamble is vetoed
+			// during recovery; the older real answer is recovered instead. (A
+			// >=200-byte turn would be accepted verbatim — see the next case.)
+			name: "borderline preamble is skipped; older real answer recovered",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(conciseAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(preambleTurn, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c2", Name: "cite", Content: "ok"}),
+				asst(sources),
+			},
+			terminal: sources,
+			want:     conciseAnswer + "\n\n" + sources,
+		},
+		{
+			// Guards the gadfly regression: a LONG (>=200-byte) front-loaded
+			// answer that merely opens with a conversational word ("Sure, …")
+			// must still be recovered — the preamble filter must not veto it.
+			name: "long answer opening with a conversational word is still recovered",
+			msgs: []llm.Message{
+				llm.UserText("q?"),
+				asst(longConversationalAnswer, cite...),
+				llm.ToolResultsMessage(llm.ToolResult{ID: "c1", Name: "cite", Content: "ok"}),
+				asst(sources),
+			},
+			terminal: sources,
+			want:     longConversationalAnswer + "\n\n" + sources,
+		},
 	}
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
@@ -179,3 +325,36 @@ func TestRun_HealthyTerminalUnchanged(t *testing.T) {
 		t.Errorf("Output = %q, want terminal answer unchanged", res.Output)
 	}
 }
+
+// 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
+// the recovered answer with the citations appended (not the bare sources list),
+// with no extra model call.
+func TestRun_RecoversFrontLoadedAnswerWithCitations(t *testing.T) {
+	longAnswer := strings.TrimSpace(strings.Repeat("Free group calls are capped at sixty minutes. ", 6))
+	sources := "Sources: [docs](https://example.com/docs), [pricing](https://example.com/pricing)."
+	fp := fake.New("fp")
+	fp.Enqueue("test-model",
+		fake.ReplyWith(llm.Response{
+			Parts:        []llm.Part{llm.Text(longAnswer)},
+			ToolCalls:    []llm.ToolCall{{ID: "c1", Name: "cite", Arguments: json.RawMessage(`{}`)}},
+			FinishReason: llm.FinishToolCalls,
+			Usage:        llm.Usage{InputTokens: 10, OutputTokens: 5},
+		}),
+		fake.Reply(sources),
+	)
+
+	a := New(newModel(t, fp), "sys", WithToolbox(citeToolbox(t)))
+	res, err := a.Run(context.Background(), "is there a meet time limit?")
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	want := longAnswer + "\n\n" + sources
+	if res.Output != want {
+		t.Errorf("Output = %q, want recovered answer + citations %q", res.Output, want)
+	}
+	if n := len(fp.Calls()); n != 2 {
+		t.Errorf("model calls = %d, want 2 (no extra nudge turn)", n)
+	}
+}

docs/adr/0016-imagegen-interface.md

@@ -52,3 +52,12 @@ contract is unchanged for callers that don't set them. `provider/llamaswap`
 forwards them to sd-server as `steps`/`cfg_scale`/`negative_prompt`/`sample_method`/
 `seed` (omitempty). This realizes the "seeds/steps … additive fields" note above;
 img2img/masks/streaming remain deferred.
+
+## Update — A1111 txt2img endpoint (seed support)
+
+`provider/llamaswap` now POSTs to sd-server's **`/sdapi/v1/txt2img`** (A1111)
+instead of the OpenAI `/v1/images/generations`. That OpenAI endpoint **ignores
+`seed`** on the stable-diffusion.cpp build we run — every render of a prompt is
+byte-identical, so a batch of N collapses to one image. `/sdapi/v1/txt2img`
+honours `seed`, restoring real per-render variety. llama-swap still routes by
+the `model` field in the body; `Size` is split into `width`/`height`.

media/media.go

@@ -5,10 +5,16 @@
 // already satisfies the target's llm.Capabilities. Images that do not fit
 // are decoded, downscaled (never upscaled), and re-encoded into an allowed
 // format and byte budget. Anything that cannot honestly be made to fit —
-// undecodable formats, impossible byte budgets, too many images, images for
+// undecodable formats, impossible byte budgets, images for a text-only
-// a text-only target — fails with an error wrapping llm.ErrUnsupported so a
+// target — fails with an error wrapping llm.ErrUnsupported so a failover
-// failover chain can advance to a more capable target without a health
+// chain can advance to a more capable target without a health penalty.
-// penalty.
+//
+// Over-count is the exception: a request carrying more images than
+// MaxImagesPerReq does NOT fail — the oldest images are replaced with a short
+// text placeholder and the most-recent MaxImagesPerReq are kept, because a hard
+// refuse exhausts a chain whose targets share the same cap (e.g. an agent loop
+// accumulating a preview image per iteration). MaxImagesPerReq remains the
+// per-model knob (0 = no image support).
 //
 // Why a separate package: every provider would otherwise duplicate the same
 // decode/scale/encode pipeline. Providers keep only a cheap capability
@@ -52,15 +58,21 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 	if !caps.SupportsImages() {
 		return llm.Request{}, fmt.Errorf("media: %w: target does not accept image input (request carries %d image(s))", llm.ErrUnsupported, total)
 	}
-	// Why error instead of dropping the overflow: silently removing an image
+	// Over-cap images are elided in the same copy-on-write pass below: the
-	// changes the question the caller asked; the honest move is to refuse and
+	// OLDEST excess are replaced with a placeholder and the most-recent
-	// let a chain try a roomier target.
+	// MaxImagesPerReq kept (see the package doc for why we elide rather than
+	// refuse). toElide is how many of the first images, front-to-back, to drop.
+	toElide := 0
 	if total > caps.MaxImagesPerReq {
-		return llm.Request{}, fmt.Errorf("media: %w: request carries %d images, target allows at most %d per request", llm.ErrUnsupported, total, caps.MaxImagesPerReq)
+		toElide = total - caps.MaxImagesPerReq
 	}
 
+	// Single copy-on-write pass: for each image, the first toElide become a text
+	// placeholder; the rest are size-normalized against caps. The Messages slice
+	// and an affected message's Parts slice are copied at most once.
 	out := req
 	copiedMessages := false
+	seen := 0
 	for mi := range req.Messages {
 		copiedParts := false
 		for pi, part := range req.Messages[mi].Parts {
@@ -68,6 +80,12 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 			if !ok {
 				continue
 			}
+			seen++
+
+			var replacement llm.Part
+			if seen <= toElide {
+				replacement = llm.Text(imageOverflowPlaceholder)
+			} else {
 				norm, changed, err := normalizeImage(ip, caps)
 				if err != nil {
 					return llm.Request{}, fmt.Errorf("media: message %d, part %d: %w", mi, pi, err)
@@ -75,6 +93,9 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 				if !changed {
 					continue
 				}
+				replacement = norm
+			}
+
 			if !copiedMessages {
 				out.Messages = make([]llm.Message, len(req.Messages))
 				copy(out.Messages, req.Messages)
@@ -86,12 +107,17 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 				out.Messages[mi].Parts = parts
 				copiedParts = true
 			}
-			out.Messages[mi].Parts[pi] = norm
+			out.Messages[mi].Parts[pi] = replacement
 		}
 	}
 	return out, nil
 }
 
+// imageOverflowPlaceholder replaces an image elided to fit a target's
+// per-request image cap. It keeps the message turn intact and tells the model
+// an earlier image was omitted rather than silently changing the conversation.
+const imageOverflowPlaceholder = "[earlier image omitted to fit this model's per-request image limit]"
+
 // Info reports an image part's sniffed format ("jpeg", "png", "gif", or
 // "webp") and pixel dimensions. It is a cheap metadata read — the pixels are
 // never decoded. webp is recognized by signature but not decodable with the

media/media_test.go

@@ -149,18 +149,48 @@ func TestNormalizeImagesUnsupported(t *testing.T) {
 	}
 }
 
-func TestNormalizeTooManyImages(t *testing.T) {
+func TestNormalizeOverCount(t *testing.T) {
-	img := llm.Image("image/png", encPNG(t, gradient(4, 4)))
+	// 3 distinguishable images across 2 messages; cap = 2. Over-count no longer
+	// errors — the OLDEST image is replaced with a placeholder and the most-recent
+	// two (the relevant ones in an iterative run) are kept, in order.
+	a := llm.Image("image/png", encPNG(t, gradient(2, 2))).(llm.ImagePart)
+	b := llm.Image("image/png", encPNG(t, gradient(4, 4))).(llm.ImagePart)
+	c := llm.Image("image/png", encPNG(t, gradient(8, 8))).(llm.ImagePart)
 	req := llm.Request{Messages: []llm.Message{
-		llm.UserParts(img, img),
+		llm.UserParts(a, b),
-		llm.UserParts(img),
+		llm.UserParts(c),
 	}}
-	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 2})
+	caps := llm.Capabilities{MaxImagesPerReq: 2, MaxImageDimension: 64, MaxImageBytes: 1 << 20, AllowedImageMIME: []string{"image/png"}}
-	if !errors.Is(err, llm.ErrUnsupported) {
+	out, err := Normalize(req, caps)
-		t.Fatalf("err = %v, want ErrUnsupported", err)
+	if err != nil {
+		t.Fatalf("over-count should not error: %v", err)
 	}
-	if !strings.Contains(err.Error(), "3 images") || !strings.Contains(err.Error(), "at most 2") {
+	var imgs []llm.ImagePart
-		t.Errorf("err message %q lacks the counts", err)
+	placeholders := 0
+	for _, m := range out.Messages {
+		for _, p := range m.Parts {
+			switch v := p.(type) {
+			case llm.ImagePart:
+				imgs = append(imgs, v)
+			case llm.TextPart:
+				if v.Text == imageOverflowPlaceholder {
+					placeholders++
+				}
+			}
+		}
+	}
+	// The exact survivors are the most-recent two, in order: b then c (a elided).
+	if len(imgs) != 2 || !bytes.Equal(imgs[0].Data, b.Data) || !bytes.Equal(imgs[1].Data, c.Data) {
+		t.Fatalf("kept %d images; want exactly [b, c] (the most-recent two)", len(imgs))
+	}
+	if placeholders != 1 {
+		t.Errorf("placeholders = %d, want 1 for the elided oldest image", placeholders)
+	}
+	// Input request untouched (copy-on-write): the first part is still image a,
+	// not a placeholder — a len check alone wouldn't catch in-place substitution.
+	first, ok := req.Messages[0].Parts[0].(llm.ImagePart)
+	if !ok || !bytes.Equal(first.Data, a.Data) {
+		t.Errorf("input request was mutated; first part = %+v", req.Messages[0].Parts[0])
 	}
 }
 

provider/llamaswap/image.go

@@ -5,6 +5,7 @@ import (
 	"encoding/base64"
 	"fmt"
 	"net/http"
+	"strconv"
 	"strings"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/imagegen"
@@ -27,34 +28,32 @@ type imageModel struct {
 	id string
 }
 
-// imageRequest is the OpenAI /v1/images/generations request shape, plus the
+// txt2imgRequest is the stable-diffusion.cpp sd-server A1111 request shape
-// stable-diffusion.cpp extras llama-swap forwards to sd-server. We always
+// (POST /sdapi/v1/txt2img). We use this endpoint rather than the OpenAI
-// request b64_json so the bytes come back inline (no second fetch). The
+// /v1/images/generations one because that endpoint IGNORES `seed` on this
-// optional fields are pointers/omitempty so an unset value is omitted entirely
+// sd-server build — every render of a given prompt comes back byte-identical,
-// and sd-server falls back to the model's own default (a field name a given
+// so a batch of N collapses to one image. /sdapi/v1/txt2img honours `seed`,
-// sd-server build doesn't recognize is simply ignored — harmless).
+// giving real variety. llama-swap still routes by the `model` field in the
-type imageRequest struct {
+// body. Optional fields are pointers/omitempty so an unset value falls back to
+// the model's baked default (the per-model --steps/--cfg-scale/etc. flags).
+type txt2imgRequest struct {
 	Model          string   `json:"model"`
 	Prompt         string   `json:"prompt"`
-	N              int      `json:"n,omitempty"`
+	NegativePrompt string   `json:"negative_prompt,omitempty"`
-	Size           string   `json:"size,omitempty"`
+	Seed           *int64   `json:"seed,omitempty"`
-	ResponseFormat string   `json:"response_format"`
 	Steps          *int     `json:"steps,omitempty"`
 	CFGScale       *float64 `json:"cfg_scale,omitempty"`
-	NegativePrompt string   `json:"negative_prompt,omitempty"`
+	Width          *int     `json:"width,omitempty"`
+	Height         *int     `json:"height,omitempty"`
 	SampleMethod   string   `json:"sample_method,omitempty"`
-	Seed           *int64   `json:"seed,omitempty"`
+	BatchCount     int      `json:"batch_count,omitempty"`
 }
 
-type imageResponse struct {
+type txt2imgResponse struct {
-	Created int64 `json:"created"`
+	Images []string `json:"images"`
-	Data    []struct {
-		B64JSON string `json:"b64_json"`
-		URL     string `json:"url"`
-	} `json:"data"`
 }
 
-// Generate implements imagegen.Model via POST {base}/v1/images/generations.
+// Generate implements imagegen.Model via POST {base}/sdapi/v1/txt2img.
 func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ...imagegen.Option) (*imagegen.Result, error) {
 	req = req.Apply(opts...)
 	if strings.TrimSpace(req.Prompt) == "" {
@@ -64,37 +63,35 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 		return nil, fmt.Errorf("%w: image count N must be >= 0, got %d", llm.ErrUnsupported, req.N)
 	}
 
-	wire := imageRequest{
+	width, height, err := parseSize(req.Size)
-		Model:          m.id,
+	if err != nil {
-		Prompt:         req.Prompt,
+		return nil, fmt.Errorf("%w: %v", llm.ErrUnsupported, err)
-		N:              req.N,
-		Size:           req.Size,
-		ResponseFormat: "b64_json",
-		Steps:          req.Steps,
-		CFGScale:       req.CFGScale,
-		NegativePrompt: req.NegativePrompt,
-		SampleMethod:   req.Sampler,
-		Seed:           req.Seed,
 	}
 
-	var resp imageResponse
+	wire := txt2imgRequest{
-	if err := m.p.doJSON(ctx, http.MethodPost, "/v1/images/generations", m.id, &wire, &resp); err != nil {
+		Model:          m.id,
+		Prompt:         req.Prompt,
+		NegativePrompt: req.NegativePrompt,
+		Seed:           req.Seed,
+		Steps:          req.Steps,
+		CFGScale:       req.CFGScale,
+		Width:          width,
+		Height:         height,
+		SampleMethod:   req.Sampler,
+		BatchCount:     req.N,
+	}
+
+	var resp txt2imgResponse
+	if err := m.p.doJSON(ctx, http.MethodPost, "/sdapi/v1/txt2img", m.id, &wire, &resp); err != nil {
 		return nil, err
 	}
 
 	out := &imagegen.Result{Raw: &resp}
-	for i, d := range resp.Data {
+	for i, b64 := range resp.Images {
-		if d.B64JSON == "" {
+		if b64 == "" {
-			// Why error rather than skip: a url-only entry means the backend
+			continue
-			// ignored response_format; we don't fetch remote content (mirrors
-			// llm.ImagePart's bytes-only contract), so surface it.
-			return nil, &llm.APIError{
-				Provider: m.p.name,
-				Model:    m.id,
-				Message:  fmt.Sprintf("image %d returned no inline b64_json data", i),
 		}
-		}
+		raw, err := base64.StdEncoding.DecodeString(b64)
-		raw, err := base64.StdEncoding.DecodeString(d.B64JSON)
 		if err != nil {
 			return nil, fmt.Errorf("llama-swap: decode image %d: %w", i, err)
 		}
@@ -110,6 +107,25 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 	return out, nil
 }
 
+// parseSize splits a "WxH" string into width/height pointers. "" yields
+// (nil, nil) so the model's own default resolution applies.
+func parseSize(size string) (*int, *int, error) {
+	size = strings.TrimSpace(size)
+	if size == "" {
+		return nil, nil, nil
+	}
+	parts := strings.SplitN(strings.ToLower(size), "x", 2)
+	if len(parts) != 2 {
+		return nil, nil, fmt.Errorf("invalid size %q (want WxH)", size)
+	}
+	w, err1 := strconv.Atoi(strings.TrimSpace(parts[0]))
+	h, err2 := strconv.Atoi(strings.TrimSpace(parts[1]))
+	if err1 != nil || err2 != nil || w <= 0 || h <= 0 {
+		return nil, nil, fmt.Errorf("invalid size %q (want WxH)", size)
+	}
+	return &w, &h, nil
+}
+
 // sniffImageMIME identifies the image format from its leading bytes, defaulting
 // to image/png (stable-diffusion.cpp emits PNG) when detection is inconclusive.
 func sniffImageMIME(data []byte) string {

provider/llamaswap/llamaswap_test.go

@@ -166,11 +166,11 @@ func TestRunningRaw(t *testing.T) {
 func TestImageGenerate(t *testing.T) {
 	var gotBody map[string]any
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/images/generations" {
+		if r.URL.Path != "/sdapi/v1/txt2img" {
 			t.Errorf("path = %q", r.URL.Path)
 		}
 		_ = json.NewDecoder(r.Body).Decode(&gotBody)
-		_, _ = w.Write([]byte(`{"created":1,"data":[{"b64_json":"` + onePixelPNG + `"}]}`))
+		_, _ = w.Write([]byte(`{"images":["` + onePixelPNG + `"]}`))
 	}))
 	defer srv.Close()
 
@@ -192,12 +192,9 @@ func TestImageGenerate(t *testing.T) {
 	if len(res.Images[0].Data) == 0 {
 		t.Error("decoded image has no bytes")
 	}
-	// response_format must be forced to b64_json, and options applied.
+	// Size is split into width/height ints for the A1111 endpoint.
-	if gotBody["response_format"] != "b64_json" {
+	if gotBody["width"] != float64(512) || gotBody["height"] != float64(512) {
-		t.Errorf("response_format = %v, want b64_json", gotBody["response_format"])
+		t.Errorf("width/height = %v/%v, want 512/512", gotBody["width"], gotBody["height"])
-	}
-	if gotBody["size"] != "512x512" {
-		t.Errorf("size = %v, want 512x512", gotBody["size"])
 	}
 }
 
@@ -205,7 +202,7 @@ func TestImageGenerateSettings(t *testing.T) {
 	var gotBody map[string]any
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		_ = json.NewDecoder(r.Body).Decode(&gotBody)
-		_, _ = w.Write([]byte(`{"created":1,"data":[{"b64_json":"` + onePixelPNG + `"}]}`))
+		_, _ = w.Write([]byte(`{"images":["` + onePixelPNG + `"]}`))
 	}))
 	defer srv.Close()