compaction & branch summarization follow retry policy

fixes #6647

compaction (auto & manual) and branch summarization retry on transient failures.
use the same retry policy from settings.
emit events for the tui to show indication of retries

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

@@ -41,6 +41,7 @@ import {
 	isContextOverflow,
 	isRetryableAssistantError,
 	modelsAreEqual,
+	type RetryCallbacks,
 	resetApiProviders,
 	streamSimple,
 } from "@earendil-works/pi-ai/compat";
@@ -161,7 +162,21 @@ export type AgentSessionEvent =
 			errorMessage?: string;
 	  }
 	| { type: "auto_retry_start"; attempt: number; maxAttempts: number; delayMs: number; errorMessage: string }
-	| { type: "auto_retry_end"; success: boolean; attempt: number; finalError?: string };
+	| { type: "auto_retry_end"; success: boolean; attempt: number; finalError?: string }
+	| {
+			type: "summarization_retry_start";
+			attempt: number;
+			maxAttempts: number;
+			delayMs: number;
+			errorMessage: string;
+	  }
+	| { type: "summarization_retry_attempt_start"; source: "branchSummary" }
+	| {
+			type: "summarization_retry_attempt_start";
+			source: "compaction";
+			reason: "manual" | "threshold" | "overflow";
+	  }
+	| { type: "summarization_retry_end" };
 
 /** Listener function for agent session events */
 export type AgentSessionEventListener = (event: AgentSessionEvent) => void;
@@ -1838,6 +1853,8 @@ export class AgentSession {
 					this.thinkingLevel,
 					this.agent.streamFunction,
 					env,
+					this.settingsManager.getRetrySettings(),
+					this._summarizationRetryCallbacks({ source: "compaction", reason: "manual" }),
 				);
 				summary = result.summary;
 				firstKeptEntryId = result.firstKeptEntryId;
@@ -2114,6 +2131,8 @@ export class AgentSession {
 					this.thinkingLevel,
 					this.agent.streamFunction,
 					env,
+					this.settingsManager.getRetrySettings(),
+					this._summarizationRetryCallbacks({ source: "compaction", reason }),
 				);
 				summary = compactResult.summary;
 				firstKeptEntryId = compactResult.firstKeptEntryId;
@@ -2620,6 +2639,37 @@ export class AgentSession {
 		return isRetryableAssistantError(message);
 	}
 
+	/**
+	 * Retry policy + callbacks shared by compaction and branch-summary summarization calls.
+	 * Uses the same `settings.retry` budget/backoff as agent-turn retries so a single transient
+	 * stream drop no longer fails the whole operation. `source` carries the context
+	 * the TUI needs to render the retry and recreate the underlying indicator.
+	 */
+	private _summarizationRetryCallbacks(
+		source: { source: "branchSummary" } | { source: "compaction"; reason: "manual" | "threshold" | "overflow" },
+	): RetryCallbacks {
+		return {
+			onRetry: (attempt, maxAttempts, delayMs, errorMessage) => {
+				this._emit({
+					type: "summarization_retry_start",
+					attempt,
+					maxAttempts,
+					delayMs,
+					errorMessage,
+				});
+			},
+			onRetryAttemptStart: () => {
+				this._emit({
+					type: "summarization_retry_attempt_start",
+					...source,
+				});
+			},
+			onRetryEnd: () => {
+				this._emit({ type: "summarization_retry_end" });
+			},
+		};
+	}
+
 	/**
 	 * Prepare a retryable error for continuation with exponential backoff.
 	 * @returns true if the caller should continue the agent, false otherwise
@@ -2934,6 +2984,8 @@ export class AgentSession {
 					replaceInstructions,
 					reserveTokens: branchSummarySettings.reserveTokens,
 					streamFn: this.agent.streamFunction,
+					retry: this.settingsManager.getRetrySettings(),
+					callbacks: this._summarizationRetryCallbacks({ source: "branchSummary" }),
 				});
 				if (result.aborted) {
 					return { cancelled: true, aborted: true };

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

@@ -6,9 +6,9 @@
  */
 
 import type { AgentMessage, StreamFn } from "@earendil-works/pi-agent-core";
+import type { RetryCallbacks, RetryPolicy } from "@earendil-works/pi-ai";
 import { contentText } from "@earendil-works/pi-ai";
 import type { Model, SimpleStreamOptions, Usage } from "@earendil-works/pi-ai/compat";
-import { completeSimple } from "@earendil-works/pi-ai/compat";
 import {
 	convertToLlm,
 	createBranchSummaryMessage,
@@ -16,7 +16,7 @@ import {
 	createCustomMessage,
 } from "../messages.ts";
 import type { ReadonlySessionManager, SessionEntry } from "../session-manager.ts";
-import { estimateTokens } from "./compaction.ts";
+import { completeSummarization, estimateTokens } from "./compaction.ts";
 import {
 	computeFileLists,
 	createFileOps,
@@ -83,6 +83,10 @@ export interface GenerateBranchSummaryOptions {
 	reserveTokens?: number;
 	/** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */
 	streamFn?: StreamFn;
+	/** Retry policy for transient summarization errors. Reuses coding-agent's `settings.retry`. */
+	retry?: RetryPolicy;
+	/** Optional callbacks for retry reporting (e.g. TUI retry indicators). */
+	callbacks?: RetryCallbacks;
 }
 
 // ============================================================================
@@ -300,6 +304,8 @@ export async function generateBranchSummary(
 		replaceInstructions,
 		reserveTokens = 16384,
 		streamFn,
+		retry,
+		callbacks,
 	} = options;
 
 	// Token budget = context window minus reserved space for prompt + response
@@ -338,12 +344,11 @@ export async function generateBranchSummary(
 
 	// Call LLM for summarization. Prefer the session stream function so SDK
 	// request behavior (timeouts, retries, attribution headers) stays consistent
-	// without running through agent state/events.
+	// without running through agent state/events. Retried via completeSummarization
+	// so transient stream drops reuse the configured retry policy.
 	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 completeSummarization(model, context, requestOptions, streamFn, retry, callbacks);
 
 	// Check if aborted or errored
 	if (response.stopReason === "aborted") {

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

@@ -6,7 +6,7 @@
  */
 
 import type { AgentMessage, StreamFn, ThinkingLevel } from "@earendil-works/pi-agent-core";
-import { contentText } from "@earendil-works/pi-ai";
+import { contentText, type RetryCallbacks, type RetryPolicy, retryAssistantCall } 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 { convertToLlm } from "../messages.ts";
@@ -552,17 +552,24 @@ function createSummarizationOptions(
 	return options;
 }
 
-async function completeSummarization(
+/**
+ * Shared choke point for every compaction/branch-summary summarization call. Wraps the
+ * single LLM call in {@link retryAssistantCall} so transient stream drops (e.g.
+ * `terminated`, socket close) honor the configured retry policy instead of failing
+ * the whole compaction on the first attempt. Deterministic errors and aborts return
+ * immediately (see {@link retryAssistantCall}).
+ */
+export async function completeSummarization(
 	model: Model<any>,
 	context: Context,
 	options: SimpleStreamOptions,
 	streamFn?: StreamFn,
+	retry?: RetryPolicy,
+	callbacks?: RetryCallbacks,
 ): Promise<AssistantMessage> {
-	if (!streamFn) {
-		return completeSimple(model, context, options);
-	}
-	const stream = await streamFn(model, context, options);
-	return stream.result();
+	const produce = async (): Promise<AssistantMessage> =>
+		streamFn ? (await streamFn(model, context, options)).result() : completeSimple(model, context, options);
+	return retryAssistantCall(produce, retry, options.signal, callbacks);
 }
 
 /**
@@ -581,6 +588,8 @@ export async function generateSummary(
 	thinkingLevel?: ThinkingLevel,
 	streamFn?: StreamFn,
 	env?: Record<string, string>,
+	retry?: RetryPolicy,
+	callbacks?: RetryCallbacks,
 ): Promise<string> {
 	return (
 		await generateSummaryWithUsage(
@@ -595,6 +604,8 @@ export async function generateSummary(
 			thinkingLevel,
 			streamFn,
 			env,
+			retry,
+			callbacks,
 		)
 	).text;
 }
@@ -612,6 +623,8 @@ export async function generateSummaryWithUsage(
 	thinkingLevel?: ThinkingLevel,
 	streamFn?: StreamFn,
 	env?: Record<string, string>,
+	retry?: RetryPolicy,
+	callbacks?: RetryCallbacks,
 ): Promise<{ text: string; usage: Usage }> {
 	const maxTokens = Math.min(
 		Math.floor(0.8 * reserveTokens),
@@ -651,6 +664,8 @@ export async function generateSummaryWithUsage(
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		completionOptions,
 		streamFn,
+		retry,
+		callbacks,
 	);
 
 	if (response.stopReason === "error") {
@@ -801,6 +816,8 @@ export async function compact(
 	thinkingLevel?: ThinkingLevel,
 	streamFn?: StreamFn,
 	env?: Record<string, string>,
+	retry?: RetryPolicy,
+	callbacks?: RetryCallbacks,
 ): Promise<CompactionResult> {
 	const {
 		firstKeptEntryId,
@@ -833,6 +850,8 @@ export async function compact(
 				thinkingLevel,
 				streamFn,
 				env,
+				retry,
+				callbacks,
 			);
 			historyText = historyResult.text;
 			historyUsage = historyResult.usage;
@@ -847,6 +866,8 @@ export async function compact(
 			signal,
 			thinkingLevel,
 			streamFn,
+			retry,
+			callbacks,
 		);
 		// Merge into single summary
 		summary = `${historyText}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefixResult.text}`;
@@ -865,6 +886,8 @@ export async function compact(
 			thinkingLevel,
 			streamFn,
 			env,
+			retry,
+			callbacks,
 		);
 		summary = result.text;
 		summaryUsage = result.usage;
@@ -900,6 +923,8 @@ async function generateTurnPrefixSummary(
 	signal?: AbortSignal,
 	thinkingLevel?: ThinkingLevel,
 	streamFn?: StreamFn,
+	retry?: RetryPolicy,
+	callbacks?: RetryCallbacks,
 ): Promise<{ text: string; usage: Usage }> {
 	const maxTokens = Math.min(
 		Math.floor(0.5 * reserveTokens),
@@ -921,6 +946,8 @@ async function generateTurnPrefixSummary(
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel),
 		streamFn,
+		retry,
+		callbacks,
 	);
 
 	if (response.stopReason === "error") {

packages/coding-agent/src/modes/interactive/interactive-mode.ts

@@ -3110,6 +3110,32 @@ export class InteractiveMode {
 				this.ui.requestRender();
 				break;
 			}
+
+			case "summarization_retry_start": {
+				this.showError(event.errorMessage);
+				this.showStatusIndicator(
+					new RetryStatusIndicator(this.ui, event.attempt, event.maxAttempts, event.delayMs),
+				);
+				this.ui.requestRender();
+				break;
+			}
+
+			case "summarization_retry_attempt_start": {
+				this.clearStatusIndicator("retry");
+				if (event.source === "branchSummary") {
+					this.showStatusIndicator(new BranchSummaryStatusIndicator(this.ui));
+				} else {
+					this.showStatusIndicator(new CompactionStatusIndicator(this.ui, event.reason));
+				}
+				this.ui.requestRender();
+				break;
+			}
+
+			case "summarization_retry_end": {
+				this.clearStatusIndicator("retry");
+				this.ui.requestRender();
+				break;
+			}
 		}
 	}
 

packages/coding-agent/test/suite/regressions/6647-compaction-retries-transient-stream-drop.test.ts

@@ -0,0 +1,192 @@
+import type { StreamFn } from "@earendil-works/pi-agent-core";
+import { type AssistantMessage, createAssistantMessageEventStream, fauxAssistantMessage } from "@earendil-works/pi-ai";
+import { afterEach, describe, expect, it } from "vitest";
+import { createHarness, type Harness } from "../harness.ts";
+
+/**
+ * Regression for #6647: compaction runs a single non-retried summarization call, so a
+ * transient mid-stream socket death (`terminated`) failed the whole compaction.
+ * Verifies that summarization now reuses `settings.retry` (bounded retries with
+ * exponential backoff gated on isRetryableAssistantError), emits
+ * `summarization_retry_*` events, and that aborts / non-retryable errors are not retried.
+ */
+describe("#6647 compaction retries transient summarization failures", () => {
+	const harnesses: Harness[] = [];
+
+	afterEach(() => {
+		while (harnesses.length > 0) {
+			harnesses.pop()?.cleanup();
+		}
+	});
+
+	function createUsage(totalTokens: number) {
+		return {
+			input: totalTokens,
+			output: 0,
+			cacheRead: 0,
+			cacheWrite: 0,
+			totalTokens,
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+		};
+	}
+
+	function seedCompactableSession(harness: Harness): void {
+		harness.settingsManager.applyOverrides({ compaction: { keepRecentTokens: 1 } });
+		const now = Date.now();
+		harness.sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "message to compact" }],
+			timestamp: now - 1000,
+		});
+		const model = harness.getModel();
+		const assistant: AssistantMessage = {
+			...fauxAssistantMessage("", { stopReason: "stop", timestamp: now - 500 }),
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: createUsage(100),
+		};
+		assistant.content = [{ type: "text", text: "assistant response to compact" }];
+		harness.sessionManager.appendMessage(assistant);
+		harness.session.agent.state.messages = harness.sessionManager.buildSessionContext().messages;
+	}
+
+	/** streamFn that responds with the given sequence of assistant messages across calls. */
+	function useScriptedStreamFn(harness: Harness, script: AssistantMessage[]): () => number {
+		let callCount = 0;
+		const streamFunction: StreamFn = (model) => {
+			const message = script[callCount] ?? script[script.length - 1]!;
+			callCount++;
+			const stream = createAssistantMessageEventStream();
+			queueMicrotask(() => {
+				if (message.stopReason === "error" || message.stopReason === "aborted") {
+					stream.push({
+						type: "error",
+						reason: message.stopReason,
+						error: { ...message, api: model.api, provider: model.provider, model: model.id },
+					});
+				} else {
+					stream.push({
+						type: "done",
+						reason: message.stopReason,
+						message: { ...message, api: model.api, provider: model.provider, model: model.id },
+					});
+				}
+			});
+			return stream;
+		};
+		harness.session.agent.streamFunction = streamFunction;
+		return () => callCount;
+	}
+
+	it("retries a transient `terminated` summarization error and compacts successfully", async () => {
+		const harness = await createHarness({ withConfiguredAuth: false });
+		harnesses.push(harness);
+		seedCompactableSession(harness);
+		harness.settingsManager.applyOverrides({ retry: { enabled: true, maxRetries: 3, baseDelayMs: 0 } });
+
+		const model = harness.getModel();
+		const error = (errorMessage: string): AssistantMessage => ({
+			...fauxAssistantMessage("", { stopReason: "error", errorMessage }),
+			usage: createUsage(10),
+		});
+		const success: AssistantMessage = {
+			...fauxAssistantMessage("recovered summary"),
+			usage: createUsage(10),
+		};
+		const getCallCount = useScriptedStreamFn(harness, [error("terminated"), error("terminated"), success]);
+
+		const result = await harness.session.compact();
+
+		expect(result.summary).toContain("recovered summary");
+		expect(getCallCount()).toBe(3); // 1 initial + 2 retries
+		const starts = harness.eventsOfType("summarization_retry_start");
+		const ends = harness.eventsOfType("summarization_retry_end");
+		expect(starts).toHaveLength(2);
+		expect(ends).toHaveLength(1);
+		expect(starts[0]).toMatchObject({ attempt: 1, maxAttempts: 3, errorMessage: "terminated" });
+		expect(starts[1]).toMatchObject({ attempt: 2, maxAttempts: 3 });
+		expect(ends[0]).toMatchObject({ type: "summarization_retry_end" });
+		// model.* referenced to keep imports honest
+		expect(model.id).toBeTruthy();
+	});
+
+	it("does not retry a non-retryable error (insufficient_quota)", async () => {
+		const harness = await createHarness({ withConfiguredAuth: false });
+		harnesses.push(harness);
+		seedCompactableSession(harness);
+		harness.settingsManager.applyOverrides({ retry: { enabled: true, maxRetries: 3, baseDelayMs: 0 } });
+
+		const error: AssistantMessage = {
+			...fauxAssistantMessage("", { stopReason: "error", errorMessage: "insufficient_quota" }),
+			usage: createUsage(10),
+		};
+		const getCallCount = useScriptedStreamFn(harness, [error]);
+
+		await expect(harness.session.compact()).rejects.toThrow("insufficient_quota");
+		expect(getCallCount()).toBe(1);
+		expect(harness.eventsOfType("summarization_retry_start")).toHaveLength(0);
+	});
+
+	it("does not retry when retry is disabled", async () => {
+		const harness = await createHarness({ withConfiguredAuth: false });
+		harnesses.push(harness);
+		seedCompactableSession(harness);
+		harness.settingsManager.applyOverrides({ retry: { enabled: false, maxRetries: 3, baseDelayMs: 0 } });
+
+		const error: AssistantMessage = {
+			...fauxAssistantMessage("", { stopReason: "error", errorMessage: "terminated" }),
+			usage: createUsage(10),
+		};
+		const getCallCount = useScriptedStreamFn(harness, [error]);
+
+		await expect(harness.session.compact()).rejects.toThrow("terminated");
+		expect(getCallCount()).toBe(1);
+		expect(harness.eventsOfType("summarization_retry_start")).toHaveLength(0);
+	});
+
+	it("stops retrying after maxRetries and reports failure", async () => {
+		const harness = await createHarness({ withConfiguredAuth: false });
+		harnesses.push(harness);
+		seedCompactableSession(harness);
+		harness.settingsManager.applyOverrides({ retry: { enabled: true, maxRetries: 2, baseDelayMs: 0 } });
+
+		const error: AssistantMessage = {
+			...fauxAssistantMessage("", { stopReason: "error", errorMessage: "terminated" }),
+			usage: createUsage(10),
+		};
+		const getCallCount = useScriptedStreamFn(harness, [error, error, error]);
+
+		await expect(harness.session.compact()).rejects.toThrow("terminated");
+		expect(getCallCount()).toBe(3); // 1 initial + 2 retries
+		const starts = harness.eventsOfType("summarization_retry_start");
+		const ends = harness.eventsOfType("summarization_retry_end");
+		expect(starts).toHaveLength(2);
+		expect(ends).toHaveLength(1);
+		expect(ends[0]).toMatchObject({ type: "summarization_retry_end" });
+	});
+
+	it("aborts an in-flight retry backoff via abortCompaction", async () => {
+		const harness = await createHarness({ withConfiguredAuth: false });
+		harnesses.push(harness);
+		seedCompactableSession(harness);
+		harness.settingsManager.applyOverrides({ retry: { enabled: true, maxRetries: 5, baseDelayMs: 30_000 } });
+
+		const error: AssistantMessage = {
+			...fauxAssistantMessage("", { stopReason: "error", errorMessage: "terminated" }),
+			usage: createUsage(10),
+		};
+		useScriptedStreamFn(harness, [error, error, error]);
+
+		const compactPromise = harness.session.compact();
+		// Let the first error resolve and the retry backoff sleep start.
+		await new Promise((resolve) => setTimeout(resolve, 0));
+		harness.session.abortCompaction();
+
+		// The aborted retry backoff rejects with an AbortError (matching the real SDK
+		// abort path), which compaction classifies as aborted.
+		await expect(compactPromise).rejects.toThrow();
+		const compactionEnd = harness.eventsOfType("compaction_end").at(-1);
+		expect(compactionEnd).toMatchObject({ aborted: true });
+	});
+});