fix(agents): keep a stored context window when the same model is re-picked

src/components/modal/AgentEditorModal.svelte

@@ -187,6 +187,29 @@ function openModelSelectionModal() {
 	}).open();
 }
 
+// The window is filled from provider metadata when a model is picked; this lets the user
+// override it (e.g. to match an Ollama num_ctx their hardware can hold). Re-picking the same
+// model keeps the override — see buildPersistedChatModel (#542).
+// Bumped on a rejected edit to remount the field, which otherwise keeps showing the invalid
+// value it bound locally.
+let contextWindowFieldKey = $state(0);
+
+function updateChatContextWindow(value: number) {
+	const chatModel = selectedAgent?.chatModel;
+	if (!chatModel) return;
+	const contextWindow = Math.round(value);
+	if (!Number.isFinite(contextWindow) || contextWindow < 1) {
+		new Notice("Context window must be a positive number of tokens.");
+		contextWindowFieldKey++;
+		return;
+	}
+	if (contextWindow === chatModel.modelConfig?.contextWindow) return;
+	pluginData.updateAgent(agentId, {
+		chatModel: { ...chatModel, modelConfig: { ...chatModel.modelConfig, contextWindow } },
+	});
+	void applyChanges();
+}
+
 function openSummarizationModelSelectionModal() {
 	const currentSelection = selectedAgent?.summarizationModel
 		? {
@@ -784,6 +807,21 @@ function getServerToolsState(serverId: string): MCPServerToolsState | undefined
           />
         </SettingItem>
 
+        {#if selectedAgent.chatModel}
+          <SettingItem
+            name="Context window"
+            desc="Tokens the chat model can take in. Decides when older history is summarized, and for Ollama it is also the num_ctx sent with each request. Filled in from the provider when you pick a model."
+          >
+            {#key contextWindowFieldKey}
+              <Text
+                inputType="number"
+                value={selectedAgent.chatModel.modelConfig?.contextWindow ?? 128000}
+                onchange={updateChatContextWindow}
+              />
+            {/key}
+          </SettingItem>
+        {/if}
+
         <!-- The agent's whole definition is one note: base instructions, the current-date
              section, and the memory section, with live values written as placeholders that
              assembly substitutes. There is no memory toggle — an agent participates in memory

src/utils/persistedChatModel.ts

@@ -4,9 +4,11 @@ import { useAvailableModels } from "../hooks/useAvailableModels.svelte";
 /**
  * Build the `ChatModel` record persisted on an agent when a model is picked.
  *
- * Hydrated catalogue metadata wins where available; otherwise the previously stored
- * config is preserved so a re-pick of the same model doesn't silently drop a tuned
- * context window or temperature.
+ * Re-picking the model already stored keeps its stored context window: that value may be
+ * user-tuned (e.g. an Ollama `num_ctx` sized to the machine), and catalogue metadata —
+ * which falls back to a 128k default when the provider reports nothing — must not
+ * silently replace it (#542). Picking a different model takes the hydrated catalogue
+ * value. Vision support follows the catalogue where known; temperature is always kept.
  *
  * Shared by every model picker (composer pill, agent editor, onboarding, and the
  * "select a model" notice action) so a model selected from any of them persists
@@ -15,11 +17,14 @@ import { useAvailableModels } from "../hooks/useAvailableModels.svelte";
 export function buildPersistedChatModel(provider: string, model: string, existing?: ChatModel | null): ChatModel {
 	const models = useAvailableModels();
 	const hydrated = models.hydratedChatModelsByKey.get(`${provider}:${model}`);
+	const sameModel = existing?.provider === provider && existing?.model === model;
+	const storedContextWindow = sameModel ? existing?.modelConfig?.contextWindow : undefined;
 	return {
 		provider,
 		model,
 		modelConfig: {
-			contextWindow: hydrated?.contextWindow ?? existing?.modelConfig?.contextWindow ?? 128000,
+			contextWindow:
+				storedContextWindow ?? hydrated?.contextWindow ?? existing?.modelConfig?.contextWindow ?? 128000,
 			supportsVision: hydrated?.capabilities.vision ?? existing?.modelConfig?.supportsVision,
 			temperature: existing?.modelConfig?.temperature,
 		},

test/utils/persistedChatModel.test.ts

@@ -0,0 +1,83 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import type { HydratedChatModelMetadata } from "../../src/types/modelMetadata";
+
+const hydratedChatModelsByKey = new Map<string, HydratedChatModelMetadata>();
+
+vi.mock("../../src/hooks/useAvailableModels.svelte", () => ({
+	useAvailableModels: () => ({ hydratedChatModelsByKey }),
+}));
+
+import { buildPersistedChatModel } from "../../src/utils/persistedChatModel";
+
+function hydrated(provider: string, variantKey: string, contextWindow: number, vision?: boolean) {
+	hydratedChatModelsByKey.set(`${provider}:${variantKey}`, {
+		kind: "chat",
+		provider,
+		variantKey,
+		displayName: variantKey,
+		contextWindow,
+		capabilities: { vision },
+	});
+}
+
+describe("buildPersistedChatModel", () => {
+	beforeEach(() => {
+		hydratedChatModelsByKey.clear();
+	});
+
+	it("keeps a stored context window when the same model is re-picked (#542)", () => {
+		hydrated("ollama", "qwen3:8b", 128000);
+		const existing = { provider: "ollama", model: "qwen3:8b", modelConfig: { contextWindow: 16384 } };
+
+		const result = buildPersistedChatModel("ollama", "qwen3:8b", existing);
+
+		expect(result.modelConfig.contextWindow).toBe(16384);
+	});
+
+	it("takes the catalogue context window when a different model is picked", () => {
+		hydrated("ollama", "llama3.2:3b", 131072);
+		const existing = { provider: "ollama", model: "qwen3:8b", modelConfig: { contextWindow: 16384 } };
+
+		const result = buildPersistedChatModel("ollama", "llama3.2:3b", existing);
+
+		expect(result.modelConfig.contextWindow).toBe(131072);
+	});
+
+	it("treats the same model name on another provider as a different model", () => {
+		hydrated("openrouter", "qwen3:8b", 40960);
+		const existing = { provider: "ollama", model: "qwen3:8b", modelConfig: { contextWindow: 16384 } };
+
+		const result = buildPersistedChatModel("openrouter", "qwen3:8b", existing);
+
+		expect(result.modelConfig.contextWindow).toBe(40960);
+	});
+
+	it("uses the catalogue value when the same model has no stored context window", () => {
+		hydrated("ollama", "qwen3:8b", 40960);
+		const existing = { provider: "ollama", model: "qwen3:8b", modelConfig: {} };
+
+		const result = buildPersistedChatModel("ollama", "qwen3:8b", existing);
+
+		expect(result.modelConfig.contextWindow).toBe(40960);
+	});
+
+	it("falls back to the previous value, then 128000, when the catalogue has no entry", () => {
+		const existing = { provider: "ollama", model: "qwen3:8b", modelConfig: { contextWindow: 16384 } };
+
+		expect(buildPersistedChatModel("ollama", "llama3.2:3b", existing).modelConfig.contextWindow).toBe(16384);
+		expect(buildPersistedChatModel("ollama", "llama3.2:3b", null).modelConfig.contextWindow).toBe(128000);
+	});
+
+	it("follows the catalogue for vision and keeps the stored temperature", () => {
+		hydrated("ollama", "qwen3:8b", 40960, true);
+		const existing = {
+			provider: "ollama",
+			model: "qwen3:8b",
+			modelConfig: { contextWindow: 16384, supportsVision: false, temperature: 0.2 },
+		};
+
+		const result = buildPersistedChatModel("ollama", "qwen3:8b", existing);
+
+		expect(result.modelConfig).toEqual({ contextWindow: 16384, supportsVision: true, temperature: 0.2 });
+	});
+});