fix(ai): require OpenAI Responses terminal events

packages/agent/CHANGELOG.md

@@ -8,6 +8,10 @@
 - `compact()`, `generateSummary()`, and `generateBranchSummary()` take a `Models` parameter and no longer accept explicit `apiKey`/`headers`.
 - `StreamFn` is defined structurally (`(model, context, options?) => AssistantMessageEventStream | Promise<...>`); `Models.streamSimple` satisfies it.
 
+### Fixed
+
+- Fixed harness compaction estimates to ignore malformed all-zero assistant usage after truncated responses ([#5526](https://github.com/earendil-works/pi/pull/5526) by [@dmmulroy](https://github.com/dmmulroy)).
+
 ## [0.79.10] - 2026-06-22
 
 ## [0.79.9] - 2026-06-20

packages/agent/src/harness/compaction/compaction.ts

@@ -121,14 +121,19 @@ export function calculateContextTokens(usage: Usage): number {
 function getAssistantUsage(msg: AgentMessage): Usage | undefined {
 	if (msg.role === "assistant" && "usage" in msg) {
 		const assistantMsg = msg as AssistantMessage;
-		if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
+		if (
+			assistantMsg.stopReason !== "aborted" &&
+			assistantMsg.stopReason !== "error" &&
+			assistantMsg.usage &&
+			calculateContextTokens(assistantMsg.usage) > 0
+		) {
 			return assistantMsg.usage;
 		}
 	}
 	return undefined;
 }
 
-/** Return usage from the last successful assistant message in session entries. */
+/** Return usage from the last valid assistant message in session entries. */
 export function getLastAssistantUsage(entries: SessionTreeEntry[]): Usage | undefined {
 	for (let i = entries.length - 1; i >= 0; i--) {
 		const entry = entries[i];

packages/agent/test/harness/compaction.test.ts

@@ -301,11 +301,28 @@ describe("harness compaction", () => {
 				createMessageEntry({ ...assistant, stopReason: "error" }),
 			]),
 		).toBeUndefined();
+		expect(
+			getLastAssistantUsage([
+				createMessageEntry(createUserMessage("user")),
+				createMessageEntry(assistant),
+				createMessageEntry(createAssistantMessage("partial", createMockUsage(0, 0))),
+			]),
+		).toBe(usage);
 		expect(estimateContextTokens([createUserMessage("no usage")]).lastUsageIndex).toBeNull();
 		expect(estimateContextTokens([assistant, createUserMessage("tail")])).toMatchObject({
 			usageTokens: 20,
 			lastUsageIndex: 0,
 		});
+		const estimate = estimateContextTokens([
+			createUserMessage("Hello"),
+			assistant,
+			createUserMessage("continue"),
+			createAssistantMessage("Partial thinking", createMockUsage(0, 0)),
+		]);
+		expect(estimate.usageTokens).toBe(20);
+		expect(estimate.lastUsageIndex).toBe(1);
+		expect(estimate.trailingTokens).toBeGreaterThan(0);
+		expect(estimate.tokens).toBe(20 + estimate.trailingTokens);
 	});
 
 	it("builds session context with a compaction entry", () => {