test(media): match the overflow placeholder by const, not substring (gadfly #8)

ragnaros/qwen3.6-27b noted TestNormalizeOverCount matched 'omitted' by substring;
the test is in-package, so assert == imageOverflowPlaceholder instead — robust to
wording changes. No behavior change.

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

.gitea/workflows/adversarial-review.yml

@@ -4,9 +4,9 @@
 # caches :latest, and this build is what carries foreman provider-type support)
 # as a specialist swarm and posts
 # ONE consolidated review comment as gitea-actions. Advisory only — never blocks a
-# merge. This reviews majordomo PRs with 9 ollama-cloud models + the M5 Mac
-# (3-lens suite). Gadfly is a simple system — findings are advisory; always
-# double-check before acting.
+# merge. This reviews majordomo PRs with 6 ollama-cloud models (3-lens suite).
+# Gadfly is a simple system — findings are advisory; always double-check before
+# acting.
 
 name: Adversarial Review (Gadfly)
 
@@ -42,36 +42,26 @@ jobs:
               || github.actor == 'fizi'
               || github.actor == 'dazed'))
     runs-on: ubuntu-latest
-    # Fleet: 9 cloud (lens fan-out) + the M5 Mac via foreman. The slow local
-    # lane dominates wall time, so allow plenty of headroom. (M1 was dropped —
-    # consistently slow for zero real findings.)
-    timeout-minutes: 90
+    # Fleet: 6 ollama-cloud models (lens fan-out), no local Macs. (Trimmed the
+    # weakest reviewers by grade — m5/qwen3.6, gemma4, gpt-oss, kimi-k2.7 — plus
+    # the earlier M1 drop.) Plenty of headroom for the cloud lanes.
+    timeout-minutes: 45
     steps:
       - uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-d7f364d
         env:
           GITEA_API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
           GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
           OLLAMA_CLOUD_API_KEY: ${{ secrets.OLLAMA_CLOUD_API_KEY }}
-          # Local Mac, reached through its foreman queue (native Ollama on the
-          # wire). GADFLY_ENDPOINT_M5 registers provider "m5", a foreman-preset
-          # Ollama client at the secret's URL, of the form:
-          #   foreman|https://<foreman-host>|<token>
-          # Needs an image with foreman provider-type support (this one). If the Mac
-          # is offline that model's comment shows an error and the others still post.
-          # (Gitea secrets aren't auto-exposed — map each explicitly.)
-          GADFLY_ENDPOINT_M5: ${{ secrets.GADFLY_ENDPOINT_M5 }}
-          # Fleet: 9 cloud + M5 Max. Cloud concurrency lives in the LENSES: cloud
-          # models run a few at a time (ollama-cloud=3) with their 3 lenses
-          # concurrent (LENS ollama-cloud=3) so comments land sooner; the Mac runs
-          # one model, lenses serial (its foreman queue serializes anyway). Both
-          # provider lanes run parallel.
-          GADFLY_MODELS: "minimax-m3:cloud,glm-5.2:cloud,glm-5.1:cloud,kimi-k2.7-code:cloud,deepseek-v4-pro:cloud,nemotron-3-super:cloud,gpt-oss:120b-cloud,qwen3-coder:480b-cloud,gemma4:cloud,m5/qwen3.6:35b-mlx"
-          GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=3,m5=1"
+          # Cloud-only fleet (no local Macs). Cloud concurrency lives in the
+          # LENSES: models run a few at a time (ollama-cloud=3) with their 3 lenses
+          # concurrent (LENS ollama-cloud=3) so comments land sooner.
+          GADFLY_MODELS: "minimax-m3:cloud,glm-5.2:cloud,glm-5.1:cloud,deepseek-v4-pro:cloud,nemotron-3-super:cloud,qwen3-coder:480b-cloud"
+          GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=3"
           GADFLY_PROVIDER_LENS_CONCURRENCY: "ollama-cloud=3"
           # Default => the 3-lens suite (security, correctness, error-handling).
           # Set the repo var GADFLY_SPECIALISTS to override (csv / "all" / "auto").
           GADFLY_SPECIALISTS: ${{ vars.GADFLY_SPECIALISTS || 'security,correctness,error-handling' }}
-          # Per-lens deadline + bounded steps so the slow local models stay sane.
+          # Per-lens deadline + bounded steps to keep each reviewer's run sane.
           GADFLY_TIMEOUT_SECS: "600"
           GADFLY_MAX_STEPS: "14"
           # Allow-list for the comment trigger (mirrors the job-level if: guard).

CLAUDE.md

@@ -142,9 +142,9 @@ Ship work through PRs and let Gadfly review it before merge:
 
 - **Push to a PR, never straight to `main`.** Branch, push, open a PR.
   `.gitea/workflows/adversarial-review.yml` runs Gadfly (the standalone
-  agentic adversarial reviewer) — a fleet of 9 ollama-cloud models +
-  the M5 Mac via foreman, each running the 3-lens suite (security,
-  correctness, error-handling). Advisory only; it never blocks the merge.
+  agentic adversarial reviewer) — a fleet of 6 ollama-cloud models, each
+  running the 3-lens suite (security, correctness, error-handling). Advisory
+  only; it never blocks the merge.
 - **Wait for Gadfly to finish, then read its output.** Don't merge while the
   review is still running. Each model posts one consolidated comment; weigh
   every finding on its merits and fix the real ones (Gadfly is a simple

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
-// a text-only target — fails with an error wrapping llm.ErrUnsupported so a
-// failover chain can advance to a more capable target without a health
-// penalty.
+// undecodable formats, impossible byte budgets, images for a text-only
+// target — fails with an error wrapping llm.ErrUnsupported so a failover
+// chain can advance to a more capable target without a health 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
-	// changes the question the caller asked; the honest move is to refuse and
-	// let a chain try a roomier target.
+	// Over-cap images are elided in the same copy-on-write pass below: the
+	// OLDEST excess are replaced with a placeholder and the most-recent
+	// 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 {
@@ -68,13 +80,22 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 			if !ok {
 				continue
 			}
-			norm, changed, err := normalizeImage(ip, caps)
-			if err != nil {
-				return llm.Request{}, fmt.Errorf("media: message %d, part %d: %w", mi, pi, err)
-			}
-			if !changed {
-				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)
+				}
+				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) {
-	img := llm.Image("image/png", encPNG(t, gradient(4, 4)))
+func TestNormalizeOverCount(t *testing.T) {
+	// 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(img),
+		llm.UserParts(a, b),
+		llm.UserParts(c),
 	}}
-	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 2})
-	if !errors.Is(err, llm.ErrUnsupported) {
-		t.Fatalf("err = %v, want ErrUnsupported", err)
+	caps := llm.Capabilities{MaxImagesPerReq: 2, MaxImageDimension: 64, MaxImageBytes: 1 << 20, AllowedImageMIME: []string{"image/png"}}
+	out, err := Normalize(req, caps)
+	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") {
-		t.Errorf("err message %q lacks the counts", err)
+	var imgs []llm.ImagePart
+	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])
 	}
 }