fix(ai): require OpenAI Responses terminal events

packages/coding-agent/CHANGELOG.md

@@ -14,6 +14,7 @@
 
 ### Fixed
 
+- Fixed context usage and 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)).
 - Fixed `pi --resume` to load user package themes and resolve automatic light/dark theme settings.
 - Fixed `models.json` custom providers so stored credentials can satisfy auth without a redundant provider-level `apiKey` ([#5953](https://github.com/earendil-works/pi/issues/5953)).
 

packages/coding-agent/src/core/agent-session.ts

@@ -1874,10 +1874,12 @@ export class AgentSession {
 		}
 
 		// Case 2: Threshold - context is getting large
-		// For error messages (no usage data), estimate from last successful response.
-		// This ensures sessions that hit persistent API errors (e.g. 529) can still compact.
+		// For error messages or all-zero usage messages, estimate from the last valid response.
+		// This ensures sessions that hit persistent API errors (e.g. 529) or malformed zero-usage
+		// responses can still compact and do not reset context accounting.
 		let contextTokens: number;
-		if (assistantMessage.stopReason === "error") {
+		const directContextTokens = assistantMessage.usage ? calculateContextTokens(assistantMessage.usage) : 0;
+		if (assistantMessage.stopReason === "error" || directContextTokens === 0) {
 			const messages = this.agent.state.messages;
 			const estimate = estimateContextTokens(messages);
 			if (estimate.lastUsageIndex === null) return false; // No usage data at all
@@ -1894,7 +1896,7 @@ export class AgentSession {
 			}
 			contextTokens = estimate.tokens;
 		} else {
-			contextTokens = calculateContextTokens(assistantMessage.usage);
+			contextTokens = directContextTokens;
 		}
 		if (shouldCompact(contextTokens, contextWindow, settings)) {
 			return await this._runAutoCompaction("threshold", false);
@@ -3011,8 +3013,8 @@ export class AgentSession {
 						const contextTokens = calculateContextTokens(assistant.usage);
 						if (contextTokens > 0) {
 							hasPostCompactionUsage = true;
+							break;
 						}
-						break;
 					}
 				}
 			}

packages/coding-agent/src/core/compaction/compaction.ts

@@ -139,12 +139,17 @@ export function calculateContextTokens(usage: Usage): number {
 
 /**
  * Get usage from an assistant message if available.
- * Skips aborted and error messages as they don't have valid usage data.
+ * Skips aborted, error, and all-zero usage messages as they don't have valid usage data.
  */
 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;
 		}
 	}
@@ -152,7 +157,7 @@ function getAssistantUsage(msg: AgentMessage): Usage | undefined {
 }
 
 /**
- * Find the last non-aborted assistant message usage from session entries.
+ * Find the last valid assistant message usage from session entries.
  */
 export function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {
 	for (let i = entries.length - 1; i >= 0; i--) {

packages/coding-agent/test/agent-session-stats.test.ts

@@ -140,4 +140,28 @@ describe("AgentSession.getSessionStats", () => {
 			session.dispose();
 		}
 	});
+
+	it("ignores zero-usage messages when checking for post-compaction context usage", () => {
+		const { session, sessionManager } = createSession();
+
+		try {
+			sessionManager.appendMessage(createUserMessage("first", 1));
+			sessionManager.appendMessage(createAssistantMessage("response1", 180_000, 2));
+			const keptUserId = sessionManager.appendMessage(createUserMessage("second", 3));
+			sessionManager.appendMessage(createAssistantMessage("response2", 195_000, 4));
+			sessionManager.appendCompaction("summary", keptUserId, 195_000);
+			sessionManager.appendMessage(createUserMessage("third", 5));
+			sessionManager.appendMessage(createAssistantMessage("response3", 25_000, 6));
+			sessionManager.appendMessage(createUserMessage("continue", 7));
+			sessionManager.appendMessage(createAssistantMessage("partial", 0, 8));
+			syncAgentMessages(session, sessionManager);
+
+			const stats = session.getSessionStats();
+			expect(stats.contextUsage).toBeDefined();
+			expect(stats.contextUsage?.tokens).not.toBeNull();
+			expect(stats.contextUsage?.tokens ?? 0).toBeGreaterThan(25_000);
+		} finally {
+			session.dispose();
+		}
+	});
 });

packages/coding-agent/test/compaction.test.ts

@@ -9,6 +9,7 @@ import {
 	calculateContextTokens,
 	compact,
 	DEFAULT_COMPACTION_SETTINGS,
+	estimateContextTokens,
 	findCutPoint,
 	getLastAssistantUsage,
 	prepareCompaction,
@@ -217,12 +218,43 @@ describe("getLastAssistantUsage", () => {
 		expect(usage!.input).toBe(100);
 	});
 
+	it("should skip all-zero assistant usage", () => {
+		const entries: SessionEntry[] = [
+			createMessageEntry(createUserMessage("Hello")),
+			createMessageEntry(createAssistantMessage("Hi", createMockUsage(100, 50))),
+			createMessageEntry(createUserMessage("continue")),
+			createMessageEntry(createAssistantMessage("Partial", createMockUsage(0, 0))),
+		];
+
+		const usage = getLastAssistantUsage(entries);
+		expect(usage).not.toBeNull();
+		expect(usage!.input).toBe(100);
+	});
+
 	it("should return undefined if no assistant messages", () => {
 		const entries: SessionEntry[] = [createMessageEntry(createUserMessage("Hello"))];
 		expect(getLastAssistantUsage(entries)).toBeUndefined();
 	});
 });
 
+describe("estimateContextTokens", () => {
+	it("uses the last non-zero assistant usage as the context anchor", () => {
+		const messages: AgentMessage[] = [
+			createUserMessage("Hello"),
+			createAssistantMessage("Hi", createMockUsage(100, 50)),
+			createUserMessage("continue"),
+			createAssistantMessage("Partial thinking", createMockUsage(0, 0)),
+		];
+
+		const estimate = estimateContextTokens(messages);
+
+		expect(estimate.usageTokens).toBe(150);
+		expect(estimate.lastUsageIndex).toBe(1);
+		expect(estimate.trailingTokens).toBeGreaterThan(0);
+		expect(estimate.tokens).toBe(150 + estimate.trailingTokens);
+	});
+});
+
 describe("shouldCompact", () => {
 	it("should return true when context exceeds threshold", () => {
 		const settings: CompactionSettings = {