fix(agent): excluded reasoning from empty-stop diagnosis

packages/coding-agent/CHANGELOG.md

@@ -102,7 +102,7 @@
 - Fixed Ctrl+G external editors failing to launch on Windows because Bun re-quoted the embedded `cmd.exe /c` command line ([#8544](https://github.com/can1357/oh-my-pi/issues/8544)).
 ### Fixed
 
-- Fixed the capped empty-stop failure always naming the context/`/shake images` hint even when the provider billed output tokens. A zero-block `stop` (no content blocks) with `usage.output > 0` means content was generated and dropped downstream (a filter/refusal flattened to `finish_reason: "stop"` by a proxy, or a lossy API translation), so the message now reports the billed output-token count and points at a provider-side filter/translation instead of a context problem, and logs `outputTokens` alongside the existing warning fields. Thinking-only stops keep a thinking block (and bill output for it), so they retain the context hint ([#8511](https://github.com/can1357/oh-my-pi/issues/8511)).
+- Fixed the capped empty-stop failure always naming the context/`/shake images` hint even when the provider billed output tokens. A zero-block `stop` (no content blocks) with billed output beyond any provider-reported reasoning usage means content was generated and dropped downstream (a filter/refusal flattened to `finish_reason: "stop"` by a proxy, or a lossy API translation), so the message now reports the billed output-token count and points at a provider-side filter/translation instead of a context problem, and logs `outputTokens` alongside the existing warning fields. Thinking-only and known reasoning-only stops retain the context hint ([#8511](https://github.com/can1357/oh-my-pi/issues/8511)).
 
 ## [17.3.3] - 2026-08-14
 

packages/coding-agent/src/session/turn-recovery.ts

@@ -678,16 +678,20 @@ export class TurnRecovery {
 		if (this.#emptyStopRetryCount > EMPTY_STOP_MAX_RETRIES) {
 			const attempts = this.#emptyStopRetryCount - 1;
 			const outputTokens = assistantMessage.usage.output;
+			const outputTokensExcludingKnownReasoning = Math.max(
+				0,
+				outputTokens - (assistantMessage.usage.reasoningTokens ?? 0),
+			);
 			let finalError: string;
 			if (providerEmptyOutput) {
 				finalError = "Assistant returned no final output after retry cap; try switching models";
-			} else if (outputTokens > 0 && assistantMessage.content.length === 0) {
-				// Billed output on a truly zero-block stop means content was generated and
-				// then dropped downstream (a filter/refusal flattened to
+			} else if (outputTokensExcludingKnownReasoning > 0 && assistantMessage.content.length === 0) {
+				// Billed non-reasoning output on a truly zero-block stop means content was
+				// generated and then dropped downstream (a filter/refusal flattened to
 				// `finish_reason: "stop"` by a proxy, or a lossy API translation) — the
 				// context/`/shake images` hint is wrong here, so name the billed output
-				// instead. Thinking-only stops keep a thinking block (and bill output for
-				// it), so they fall through to the context hint rather than this path.
+				// instead. Known reasoning-only usage is not evidence that deliverable
+				// content was dropped, and thinking-only stops retain a thinking block.
 				finalError = `Assistant returned an empty stop after retry cap, but the provider billed ${outputTokens} output token${outputTokens === 1 ? "" : "s"} for it; content was generated and then dropped before delivery, which usually points to a provider-side content filter or a lossy API translation rather than a context problem`;
 			} else {
 				finalError =

packages/coding-agent/test/agent-session-empty-stop-guard.test.ts

@@ -72,6 +72,14 @@ function filteredEmptyStop(): MockResponse {
 	};
 }
 
+function reasoningOnlyEmptyStop(): MockResponse {
+	return {
+		content: [],
+		stopReason: "stop",
+		usage: { output: 126, reasoningTokens: 126, cacheRead: 100 },
+	};
+}
+
 function orphanedToolUseStop(): MockResponse {
 	return {
 		content: [{ type: "thinking", thinking: "I should call a tool next." }],
@@ -499,6 +507,31 @@ describe("AgentSession empty stop guard", () => {
 		expect(finalError).not.toContain("/shake images");
 	});
 
+	it("keeps the context hint when a capped zero-block stop billed only reasoning tokens", async () => {
+		const { session, mock } = await createHarness([
+			reasoningOnlyEmptyStop(),
+			reasoningOnlyEmptyStop(),
+			reasoningOnlyEmptyStop(),
+			reasoningOnlyEmptyStop(),
+		]);
+		const retryEndEvents: Array<Extract<AgentSessionEvent, { type: "auto_retry_end" }>> = [];
+		session.subscribe(event => {
+			if (event.type === "auto_retry_end") {
+				retryEndEvents.push(event);
+			}
+		});
+
+		await expectPromptCompletes(session.prompt("think without delivering an answer"));
+		await session.waitForIdle();
+
+		expect(mock.calls).toHaveLength(4);
+		expect(retryEndEvents).toHaveLength(1);
+		expect(retryEndEvents[0]?.success).toBe(false);
+		const finalError = retryEndEvents[0]?.finalError ?? "";
+		expect(finalError).toContain("/shake images");
+		expect(finalError).not.toContain("billed");
+	});
+
 	it("keeps the context hint for a capped thinking-only stop even though it billed output", async () => {
 		const { session, mock } = await createHarness([
 			thinkingOnlyStop(),