fix(compaction): scale summary window floor to the model's context

The absolute 16,384-token floor plus the carried summary and output
reserves exceeds windows below ~58k outright, and overflow recovery then
bailed at the very floor that caused the rejection, leaving compaction
unusable on small-context models. Scale the floor to window/8 (min 1k)
and use the same floor in overflow recovery.

packages/agent/src/compaction/compaction.ts

@@ -895,9 +895,19 @@ function createSnapcompactArchiveMigrationMessage(archiveText: string): Message
  */
 const DEFAULT_SUMMARY_INPUT_WINDOW = 200_000;
 
-/** Floor for one summarization window, so a tiny model still makes progress. */
+/**
+ * Floor for one summarization window, so a tiny model still makes progress.
+ * Scaled down (never below 1k) for models whose window cannot host the full
+ * floor next to the carried summary and output reserves.
+ */
 const MIN_SUMMARY_INPUT_TOKENS = 16_384;
 
+/** Smallest window worth planning for `model`; below this, overflow recovery gives up. */
+function minSummaryInputTokens(model: Model): number {
+	const window = model.contextWindow && model.contextWindow > 0 ? model.contextWindow : DEFAULT_SUMMARY_INPUT_WINDOW;
+	return Math.min(MIN_SUMMARY_INPUT_TOKENS, Math.max(1_024, Math.floor(window / 8)));
+}
+
 /**
  * Usable conversation input for ONE summarization call: the summarizer's window
  * minus the summary it must emit, the previous summary it carries forward, and
@@ -909,7 +919,7 @@ function summaryInputBudgetTokens(model: Model, maxTokens: number): number {
 	// 0.8, not "window minus reserves": provider tokenizers disagree with the
 	// local cl100k estimate by a few percent, and being wrong here is a hard
 	// 400 on the one call that is supposed to rescue an oversized session.
-	return Math.max(MIN_SUMMARY_INPUT_TOKENS, Math.floor(window * 0.8) - maxTokens - MAX_SUMMARY_TOKENS);
+	return Math.max(minSummaryInputTokens(model), Math.floor(window * 0.8) - maxTokens - MAX_SUMMARY_TOKENS);
 }
 
 /**
@@ -1012,7 +1022,7 @@ export async function generateSummary(
 			// the rejection proves the plan was fiction, so converging on the real
 			// cap must not spend a call per level of an imaginary ladder.
 			const halved = Math.floor(Math.min(window.budgetTokens, windowTokens) / 2);
-			if (!AIError.is(AIError.classify(error), AIError.Flag.ContextOverflow) || halved < MIN_SUMMARY_INPUT_TOKENS) {
+			if (!AIError.is(AIError.classify(error), AIError.Flag.ContextOverflow) || halved < minSummaryInputTokens(model)) {
 				throw error;
 			}
 			pending.splice(

packages/agent/test/compaction-oversized-input.test.ts

@@ -124,6 +124,31 @@ describe("summarization input budget", () => {
 		}
 	});
 
+	test("keeps the window floor inside a small model's context", async () => {
+		// The absolute 16,384-token floor plus the carried summary and output
+		// reserves exceeds a 40k window outright, and overflow recovery would then
+		// bail at the very floor that caused the rejection. The floor scales with
+		// the window instead, so a small-context model still folds successfully.
+		const providerCapChars = 26_000; // what a 40k window can host next to the reserves
+		const spy = vi.spyOn(ai, "completeSimple").mockImplementation(async (_model, context) => {
+			const prompt = promptTextOf([_model, context]);
+			if (prompt.length > providerCapChars) {
+				throw new Error(`400 prompt is too long: ${prompt.length} tokens > ${providerCapChars} maximum`);
+			}
+			return createAssistantMessage("summary");
+		});
+		try {
+			const messages = Array.from({ length: 12 }, (_, i) => turn(i, 4_000)).flat();
+			const summary = await generateSummary(messages, getModel(40_000), 16_384, "test-key");
+			expect(summary).toBe("summary");
+			for (const prompt of spy.mock.calls.map(promptTextOf)) {
+				expect(prompt.length).toBeLessThanOrEqual(providerCapChars);
+			}
+		} finally {
+			spy.mockRestore();
+		}
+	});
+
 	test("propagates a non-overflow failure instead of shrinking", async () => {
 		let calls = 0;
 		const spy = vi.spyOn(ai, "completeSimple").mockImplementation(async () => {