retry on compaction/branch summarization retryable failures

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

@@ -8,7 +8,6 @@
 import type { AgentMessage, StreamFn } from "@earendil-works/pi-agent-core";
 import { contentText } from "@earendil-works/pi-ai";
 import type { Model, SimpleStreamOptions } from "@earendil-works/pi-ai/compat";
-import { completeSimple } from "@earendil-works/pi-ai/compat";
 import {
 	convertToLlm,
 	createBranchSummaryMessage,
@@ -17,6 +16,7 @@ import {
 } from "../messages.ts";
 import type { ReadonlySessionManager, SessionEntry } from "../session-manager.ts";
 import { estimateTokens } from "./compaction.ts";
+import { completeSummaryRequest } from "./summary-request.ts";
 import {
 	computeFileLists,
 	createFileOps,
@@ -340,9 +340,7 @@ export async function generateBranchSummary(
 	// without running through agent state/events.
 	const context = { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages };
 	const requestOptions: SimpleStreamOptions = { apiKey, headers, env, signal, maxTokens: 2048 };
-	const response = streamFn
-		? await (await streamFn(model, context, requestOptions)).result()
-		: await completeSimple(model, context, requestOptions);
+	const response = await completeSummaryRequest(model, context, requestOptions, streamFn);
 
 	// Check if aborted or errored
 	if (response.stopReason === "aborted") {

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

@@ -7,15 +7,15 @@
 
 import type { AgentMessage, StreamFn, ThinkingLevel } from "@earendil-works/pi-agent-core";
 import { contentText } from "@earendil-works/pi-ai";
-import type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from "@earendil-works/pi-ai/compat";
-import { completeSimple } from "@earendil-works/pi-ai/compat";
+import type { AssistantMessage, Model, SimpleStreamOptions, Usage } from "@earendil-works/pi-ai/compat";
 import { convertToLlm } from "../messages.ts";
 import {
 	buildSessionContext,
 	type CompactionEntry,
 	type SessionEntry,
 	sessionEntryToContextMessages,
 } from "../session-manager.ts";
+import { completeSummaryRequest } from "./summary-request.ts";
 import {
 	computeFileLists,
 	createFileOps,
@@ -527,19 +527,6 @@ function createSummarizationOptions(
 	return options;
 }
 
-async function completeSummarization(
-	model: Model<any>,
-	context: Context,
-	options: SimpleStreamOptions,
-	streamFn?: StreamFn,
-): Promise<AssistantMessage> {
-	if (!streamFn) {
-		return completeSimple(model, context, options);
-	}
-	const stream = await streamFn(model, context, options);
-	return stream.result();
-}
-
 /**
  * Generate a summary of the conversation using the LLM.
  * If previousSummary is provided, uses the update prompt to merge.
@@ -590,7 +577,7 @@ export async function generateSummary(
 
 	const completionOptions = createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel);
 
-	const response = await completeSummarization(
+	const response = await completeSummaryRequest(
 		model,
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		completionOptions,
@@ -852,7 +839,7 @@ async function generateTurnPrefixSummary(
 		},
 	];
 
-	const response = await completeSummarization(
+	const response = await completeSummaryRequest(
 		model,
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel),

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

@@ -0,0 +1,32 @@
+import type { StreamFn } from "@earendil-works/pi-agent-core";
+import type { Api, AssistantMessage, Context, Model, SimpleStreamOptions } from "@earendil-works/pi-ai/compat";
+import { completeSimple, isRetryableAssistantError } from "@earendil-works/pi-ai/compat";
+
+const MAX_SUMMARY_RETRIES = 1;
+
+/**
+ * Complete a summary request, retrying one retryable stream failure.
+ * Summary requests have no side effects until their final response is used, so
+ * a fresh request is safe after a transient transport or provider failure.
+ */
+export async function completeSummaryRequest(
+	model: Model<Api>,
+	context: Context,
+	options: SimpleStreamOptions,
+	streamFn?: StreamFn,
+): Promise<AssistantMessage> {
+	let response = streamFn
+		? await (await streamFn(model, context, options)).result()
+		: await completeSimple(model, context, options);
+
+	for (let attempt = 0; attempt < MAX_SUMMARY_RETRIES; attempt++) {
+		if (options.signal?.aborted || !isRetryableAssistantError(response)) {
+			break;
+		}
+		response = streamFn
+			? await (await streamFn(model, context, options)).result()
+			: await completeSimple(model, context, options);
+	}
+
+	return response;
+}

packages/coding-agent/test/suite/regressions/6647-compaction-summary-retry.test.ts

@@ -0,0 +1,75 @@
+import { fauxAssistantMessage } from "@earendil-works/pi-ai";
+import { afterEach, describe, expect, it } from "vitest";
+import { createHarness, type Harness } from "../harness.ts";
+
+describe("issue #6647 compaction summary retry", () => {
+	const harnesses: Harness[] = [];
+
+	afterEach(() => {
+		while (harnesses.length > 0) {
+			harnesses.pop()?.cleanup();
+		}
+	});
+
+	it("retries a transient split-turn summary failure", async () => {
+		const harness = await createHarness({ settings: { compaction: { keepRecentTokens: 1 } } });
+		harnesses.push(harness);
+
+		const now = Date.now();
+		harness.sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "A request large enough to split during compaction." }],
+			timestamp: now - 1000,
+		});
+		harness.sessionManager.appendMessage({
+			...fauxAssistantMessage("An assistant response that should be retained."),
+			timestamp: now,
+		});
+		harness.session.agent.state.messages = harness.sessionManager.buildSessionContext().messages;
+
+		harness.setResponses([
+			fauxAssistantMessage("", {
+				stopReason: "error",
+				errorMessage: "OpenAI Responses stream ended before a terminal response event",
+			}),
+			fauxAssistantMessage("## Original Request\nRecovered split-turn summary"),
+		]);
+
+		await expect(harness.session.compact()).resolves.toMatchObject({
+			summary: expect.stringContaining("Recovered split-turn summary"),
+		});
+		expect(harness.faux.state.callCount).toBe(2);
+	});
+
+	it("retries a transient branch summary failure", async () => {
+		const harness = await createHarness();
+		harnesses.push(harness);
+
+		const targetId = harness.sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "Start a branch." }],
+			timestamp: Date.now() - 1000,
+		});
+		harness.sessionManager.appendMessage(fauxAssistantMessage("Initial branch reply."));
+		harness.sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "Abandoned branch work." }],
+			timestamp: Date.now(),
+		});
+		harness.sessionManager.appendMessage(fauxAssistantMessage("Abandoned branch reply."));
+
+		harness.setResponses([
+			fauxAssistantMessage("", {
+				stopReason: "error",
+				errorMessage: "OpenAI Responses stream ended before a terminal response event",
+			}),
+			fauxAssistantMessage("## Goal\nRecovered branch summary"),
+		]);
+
+		const result = await harness.session.navigateTree(targetId, { summarize: true });
+
+		expect(result.cancelled).toBe(false);
+		expect(result.summaryEntry?.summary).toContain("Recovered branch summary");
+		expect(harness.faux.state.callCount).toBe(2);
+	});
+});