diff --git a/docs/models.md b/docs/models.md
index c837cf293..dd709ed7b 100644
--- a/docs/models.md
+++ b/docs/models.md
@@ -61,6 +61,7 @@ providers:
         api: openai-completions
         reasoning: false
         input: [text]
+        imageInputDecoder: stb # local STB decoder; OMP converts WebP before dispatch
         cost:
           input: 0
           output: 0
@@ -101,6 +102,7 @@ providers:
 - `auth`: `apiKey` (default), `none`, or `oauth`; for `models.yml` custom models, `oauth` is accepted by schema but does not waive the `apiKey` requirement
 - `discovery.type`: `ollama`, `llama.cpp`, `lm-studio`, `openai-models-list`, `proxy`, or `litellm`
 - `transport`: `pi-native` only. When set, every model under that provider is sent to an `omp auth-gateway` compatible `baseUrl` via `POST /v1/pi/stream`; `apiKey` is the gateway bearer.
+- `imageInputDecoder`: `stb` only. Set this on a custom model or `modelOverrides` entry when the serving backend uses an STB-compatible image decoder that cannot accept WebP; OMP converts attached and historical WebP images before provider dispatch.
 
 ## Validation rules (current)
 
@@ -190,7 +192,7 @@ Provider defaults vs per-model overrides:
 
 - Provider `headers`, `compat`, and `remoteCompaction` are baselines.
 - Model `headers` override provider header keys.
-- `modelOverrides` can override model metadata (`name`, `reasoning`, `thinking`, `input`,
+- `modelOverrides` can override model metadata (`name`, `reasoning`, `thinking`, `input`, `imageInputDecoder`,
   `supportsTools`, `cost`, `premiumMultiplier`, `contextWindow`, `maxTokens`,
   `omitMaxOutputTokens`, `headers`, `compat`, `contextPromotionTarget`, `compactionModel`, and
   `remoteCompaction`).
diff --git a/packages/coding-agent/src/utils/image-loading.ts b/packages/coding-agent/src/utils/image-loading.ts
index 0bad86f6a..6b321d6c0 100644
--- a/packages/coding-agent/src/utils/image-loading.ts
+++ b/packages/coding-agent/src/utils/image-loading.ts
@@ -236,7 +236,14 @@ export async function normalizeModelContextMessages(messages: Message[], model:
 		}
 		if (!content) continue;
 		output ??= messages.slice();
-		output[messageIndex] = { ...message, content } as Message;
+		const normalizedMessage = { ...message, content } as Message;
+		if (normalizedMessage.role === "user" || normalizedMessage.role === "developer") {
+			// Native Responses history takes precedence over message content. Once an
+			// image changes, that opaque replay payload is stale and could resend the
+			// original WebP bytes instead of this normalized transport copy.
+			delete normalizedMessage.providerPayload;
+		}
+		output[messageIndex] = normalizedMessage;
 	}
 	return output ?? messages;
 }
diff --git a/packages/coding-agent/test/image-webp-exclusion.test.ts b/packages/coding-agent/test/image-webp-exclusion.test.ts
index 143d3d157..952df6731 100644
--- a/packages/coding-agent/test/image-webp-exclusion.test.ts
+++ b/packages/coding-agent/test/image-webp-exclusion.test.ts
@@ -1,5 +1,6 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import type { Api, Message, Model } from "@oh-my-pi/pi-ai";
+import { buildResponsesInput } from "@oh-my-pi/pi-ai/providers/openai-shared";
 import { buildModel } from "@oh-my-pi/pi-catalog/build";
 import { getBundledModels } from "@oh-my-pi/pi-catalog/models";
 import type { CustomMessage } from "@oh-my-pi/pi-coding-agent/session/messages";
@@ -247,6 +248,52 @@ describe("normalizeModelContextImages model-aware WebP exclusion", () => {
 		expect(messages[0]!.content[1]).toBe(original);
 	});
 
+	test("drops stale native Responses history after rewriting an image", async () => {
+		const model = buildStbVisionModel("managed-primary", "openai-responses");
+		const original = {
+			type: "image" as const,
+			data: await makeRedWebP(200, 200),
+			mimeType: "image/webp",
+		};
+		const providerPayload = {
+			type: "openaiResponsesHistory" as const,
+			provider: model.provider,
+			dt: true,
+			items: [
+				{
+					type: "message",
+					role: "user",
+					content: [{ type: "input_image", image_url: `data:image/webp;base64,${original.data}` }],
+				},
+			],
+		};
+		const message: Message = {
+			role: "user",
+			content: [{ type: "text", text: "inspect" }, original],
+			providerPayload,
+			timestamp: 1,
+		};
+
+		const messages = await normalizeModelContextMessages([message], model);
+		const normalizedMessage = messages[0]!;
+		expect(normalizedMessage.role).toBe("user");
+		if (normalizedMessage.role !== "user") throw new Error("Expected user message");
+		expect(normalizedMessage.providerPayload).toBeUndefined();
+		expect(message.providerPayload).toBe(providerPayload);
+
+		const wire = buildResponsesInput({
+			model,
+			context: { messages },
+			strictResponsesPairing: false,
+			supportsImageDetailOriginal: true,
+			nativeHistory: { replay: true, filterReasoning: false },
+		});
+		const serializedWire = JSON.stringify(wire);
+		expect(serializedWire).toContain("input_image");
+		expect(serializedWire).not.toContain("image/webp");
+		expect(serializedWire).not.toContain(original.data);
+	});
+
 	test("replaces an undecodable historical WebP with an omission note", async () => {
 		const corrupt = {
 			type: "image" as const,