@oh-my-pi/pi-agent-core 15.10.12 → 15.11.0

package/CHANGELOG.md

@@ -2,6 +2,30 @@
 
 ## [Unreleased]
 
+## [15.11.0] - 2026-06-10
+### Breaking Changes
+
+- Removed `compaction/index.ts` re-export of snapcompact helpers, so snapcompact utilities are no longer available from the agent compaction barrel and should be imported from `@oh-my-pi/snapcompact`
+- Removed the `convertToLlm` alias export from `compaction/messages` — it duplicated `defaultConvertToLlm` under a second name. Import `defaultConvertToLlm` (array form) or the new `convertMessageToLlm` (single-message form) instead
+
+### Added
+
+- Added `convertMessageToLlm()`: the single-message core transformer behind `defaultConvertToLlm()`. Embedders with app-specific message roles should handle their own roles and delegate every core role (`user`/`developer`/`assistant`/`toolResult`/`custom`/`hookMessage`/`branchSummary`/`compactionSummary`) to it instead of duplicating the conversion — a duplicated `compactionSummary` case is how snapcompact frames once silently dropped off provider requests
+- Added `pruneSupersededToolResults()` and the opt-in `PruneConfig.supersedeKey` hook so harnesses can prune stale tool results superseded by a newer read of the same file; superseded results are pruned ahead of age-based victims during overflow pruning and replaced with a `[Superseded by a newer read of this file]` placeholder. Without the new config, `pruneToolOutputs()` behavior is unchanged.
+- Added `readToolSupersedeKey()` implementing the read-tool path/selector grammar (selector-free reads supersede range reads of the same file; URL-scheme paths exempt). Pruning honors prompt-cache economics: per-turn prunes only fire when the post-candidate suffix is small or the cache is cold (idle gap).
+- Added the `snapcompact` compaction strategy via `@oh-my-pi/snapcompact`: instead of an LLM summary, discarded history is printed onto dense bitmap frames and re-attached to the compaction summary message as image blocks. `CompactionSummaryMessage` gains an optional `images` field, `estimateTokens()` charges per attached frame, and frames persist under `preserveData.snapcompact` with an 8-frame middle-out eviction budget.
+- Snapcompact frames are now rendered in a provider-aware shape (`SNAPCOMPACT_SHAPES` + `resolveSnapcompactShape(api)`), following the snapcompact 200k-token monolithic evals: Anthropic-family and unknown APIs get `8x8r-bw` (unscii-8 square cells, black ink, every line printed twice with the copy on a pale highlight band — read at F1 parity with raw text at ~2x lower cost and the most refusal-robust), Google gets `8x8r-sent` (sentence-hue ink, ~2.9x cheaper), and OpenAI gets `6x6u-sent` (unscii Lanczos-stretched to 6x6 cells — OpenAI bills a flat ~2.9k tokens per image, so frame count is the only cost lever) with `detail: "original"` on the frame images. `snapcompactCompact()` accepts `model`/`shape` options, frames persist their shape metadata, mixed-shape archives (provider switches, legacy 5x8 frames) are flagged in the reading instructions, and `snapcompactGeometry()`/`renderSnapcompactFrame()` now take a shape
+
+### Changed
+
+- Compaction and branch-summary file lists are now a single `<files>` tag instead of `<read-files>`/`<modified-files>`: paths render as the grouped, prefix-folded directory tree the find/search tools emit (`# dir/` headers, bare basenames), each annotated `(Read)`, `(Write)`, or `(RW)` — modified files that were also read get `(RW)`. Legacy tags in summaries written by earlier versions are still stripped and self-heal on the next compaction
+
+### Fixed
+
+- Fixed queued steering messages being drained into an externally aborted run: interrupting mid-tool execution (e.g. Enter with a pending steer) dequeued the steer into the dying run — it landed in history without a response and the post-abort resume saw an empty queue, so the agent stopped instead of continuing. Steering/follow-up/aside queue polls are now skipped once the run's abort signal fires, leaving the queue intact for `Agent.continue()`.
+- Fixed `<read-files>` compaction lists recording the same file once per line-range/raw selector (`src/foo.ts:50-200`, `:raw`, `:1-50:raw`, …): read-tool selectors are now stripped before tracking, so reads dedupe to the base path and match their write/edit path when splitting read-only vs modified lists. Selector-polluted lists stored by earlier compactions self-heal on the next compaction. `readToolSupersedeKey()` now shares the same splitter (`splitReadSelector()`), gaining the `..` range alias and `L`-prefix forms it previously missed.
+- Fixed `estimateTokens()` undercounting thinking-heavy assistant messages on replay: `thinkingSignature` payloads (OpenAI Responses encrypted reasoning items, Anthropic signed thinking blocks, etc.) and `redactedThinking.data` are now charged alongside the visible thinking text, so the local estimate tracks provider-reported usage instead of straddling the threshold on every turn ([#2275](https://github.com/can1357/oh-my-pi/issues/2275)).
+
 ## [15.10.12] - 2026-06-10
 
 ### Added
@@ -664,4 +688,4 @@ Initial release under @oh-my-pi scope. See previous releases at [badlogic/pi-mon
 
 - `Agent` constructor now has all options optional (empty options use defaults).
 
-- `queueMessage()` is now synchronous (no longer returns a Promise).
+- `queueMessage()` is now synchronous (no longer returns a Promise).

package/dist/types/compaction/compaction.d.ts

@@ -30,7 +30,7 @@ export interface CompactionResult<T = unknown> {
 }
 export interface CompactionSettings {
     enabled: boolean;
-    strategy?: "context-full" | "handoff" | "shake" | "off";
+    strategy?: "context-full" | "handoff" | "shake" | "snapcompact" | "off";
     thresholdPercent?: number;
     thresholdTokens?: number;
     reserveTokens: number;

package/dist/types/compaction/messages.d.ts

@@ -33,6 +33,8 @@ export interface CompactionSummaryMessage {
     shortSummary?: string;
     tokensBefore: number;
     providerPayload?: ProviderPayload;
+    /** Snapcompact frames archived by this compaction; appended as image blocks after the summary text. */
+    images?: ImageContent[];
     timestamp: number;
 }
 export type CoreCompactionMessage = CustomMessage | HookMessage | BranchSummaryMessage | CompactionSummaryMessage;
@@ -48,8 +50,19 @@ export type ConvertToLlm = (messages: AgentMessage[]) => Message[];
 export declare function renderBranchSummaryContext(summary: string): string;
 export declare function renderCompactionSummaryContext(summary: string): string;
 export declare function createBranchSummaryMessage(summary: string, fromId: string, timestamp: string): BranchSummaryMessage;
-export declare function createCompactionSummaryMessage(summary: string, tokensBefore: number, timestamp: string, shortSummary?: string, providerPayload?: ProviderPayload): CompactionSummaryMessage;
+export declare function createCompactionSummaryMessage(summary: string, tokensBefore: number, timestamp: string, shortSummary?: string, providerPayload?: ProviderPayload, images?: ImageContent[]): CompactionSummaryMessage;
 export declare function createCustomMessage(customType: string, content: string | (TextContent | ImageContent)[], display: boolean, details: unknown | undefined, timestamp: string, attribution?: MessageAttribution): CustomMessage;
+/**
+ * Transform a single core-domain agent message to its LLM form; `undefined`
+ * drops it from the provider request.
+ *
+ * Single source of truth for the core roles (user/developer/assistant/
+ * toolResult) and the compaction messages owned by this package. Embedders
+ * with their own app messages (e.g. the coding agent) handle their custom
+ * roles and delegate every core role here — duplicating these cases is how
+ * snapcompact frames once silently fell off the provider request.
+ */
+export declare function convertMessageToLlm(message: AgentMessage): Message | undefined;
 /**
  * Default compaction-domain transformer.
  *
@@ -58,4 +71,3 @@ export declare function createCustomMessage(customType: string, content: string
  * core LLM roles and the compaction messages owned by this package.
  */
 export declare function defaultConvertToLlm(messages: AgentMessage[]): Message[];
-export declare const convertToLlm: typeof defaultConvertToLlm;

package/dist/types/compaction/pruning.d.ts

@@ -10,10 +10,58 @@ export interface PruneConfig {
     minimumSavings: number;
     /** Tool-result protection matchers. String entries protect every result from that tool; predicates may inspect the paired tool call. */
     protectedTools: ProtectedToolMatcher[];
+    /**
+     * Optional supersede key function (see {@link SupersedePruneConfig.supersedeKey}).
+     * When provided, superseded tool results are pruned first — even inside the
+     * `protectTokens` window — before age-based victims. Absent, behavior is
+     * unchanged.
+     */
+    supersedeKey?: SupersedeKeyFn;
 }
 export declare const DEFAULT_PRUNE_CONFIG: PruneConfig;
 export interface PruneResult {
     prunedCount: number;
     tokensSaved: number;
 }
+/** Exact placeholder written over a superseded tool result. */
+export declare const SUPERSEDED_NOTICE = "[Superseded by a newer read of this file]";
+/**
+ * Maps a tool call to a supersede key. Results sharing a key form a group in
+ * which every result except the newest is a supersede candidate. A key `K`
+ * additionally supersedes keys with prefix `K + "\u0000"` (selector-free read
+ * supersedes selector-carrying reads of the same base path). Return
+ * `undefined` to exempt a call from supersede grouping.
+ */
+export type SupersedeKeyFn = (toolName: string, args: Record<string, unknown>) => string | undefined;
+export interface SupersedePruneConfig {
+    /** Supersede key function; results sharing a key supersede older ones. */
+    supersedeKey: SupersedeKeyFn;
+    /** Prune a candidate now when all messages after it total at most this many estimated tokens. Default 8 000. */
+    suffixTokenLimit?: number;
+    /** Prune all candidates when the last message is at least this old (prompt cache is cold anyway). Default 30 min. */
+    idleFlushMs?: number;
+    /** Clock override for tests. */
+    now?: number;
+    /** Tool-result protection matchers (same contract as {@link PruneConfig.protectedTools}). */
+    protectedTools: ProtectedToolMatcher[];
+}
+/**
+ * Prune superseded tool results (e.g. stale `read` outputs replaced by a newer
+ * read of the same file). Cheap, incremental, and prompt-cache-aware: a
+ * candidate is pruned now only when the suffix after it is small (tail case —
+ * the read→edit→read loop) or when the context has been idle long enough that
+ * the provider cache is cold anyway (then ALL candidates flush).
+ */
+export declare function pruneSupersededToolResults(entries: SessionEntry[], config: SupersedePruneConfig): PruneResult;
 export declare function pruneToolOutputs(entries: SessionEntry[], config?: PruneConfig): PruneResult;
+/**
+ * Supersede key for the `read` tool: the file path with the trailing line/raw
+ * selector stripped (the read tool's own splitter grammar via
+ * {@link splitReadSelector}, e.g. `src/foo.ts:50-200`, `:2-4:raw`).
+ * Internal/URL-scheme paths (`skill://…`, `https://…`) are exempt.
+ * Selector-free reads key on the bare path; selector-carrying reads key on
+ * `path + "\u0000" + selector`, so two reads collide only when the newer is
+ * selector-free or the selectors are identical (the pass's prefix rule lets a
+ * bare-path read supersede selector-carrying reads of the same file).
+ */
+export declare function readToolSupersedeKey(toolName: string, args: Record<string, unknown>): string | undefined;

package/dist/types/compaction/utils.d.ts

@@ -9,6 +9,22 @@ export interface FileOperations {
     edited: Set<string>;
 }
 export declare function createFileOps(): FileOperations;
+/**
+ * Split a read-tool path into its base path and trailing selector, mirroring the
+ * read tool's own splitter. Single source of the grammar in this package: the
+ * file-operations list strips selectors via {@link stripReadSelector}, and the
+ * supersede-prune pass keys on both parts via `readToolSupersedeKey`.
+ */
+export declare function splitReadSelector(path: string): {
+    path: string;
+    sel?: string;
+};
+/**
+ * Strip a trailing read-tool selector (`:50-200`, `:raw`, `:1-50:raw`, `:conflicts`, …)
+ * so the same file read with different line ranges dedupes to one `<files>` entry
+ * and matches its write/edit path when computing Read/Write/RW markers.
+ */
+export declare function stripReadSelector(path: string): string;
 /**
  * Extract file operations from tool calls in an assistant message.
  */
@@ -21,8 +37,8 @@ export declare function computeFileLists(fileOps: FileOperations): {
     readFiles: string[];
     modifiedFiles: string[];
 };
-export declare function formatFileOperations(readFiles: string[], modifiedFiles: string[]): string;
-export declare function upsertFileOperations(summary: string, readFiles: string[], modifiedFiles: string[]): string;
+export declare function formatFileOperations(readFiles: string[], modifiedFiles: string[], readSet?: ReadonlySet<string>): string;
+export declare function upsertFileOperations(summary: string, readFiles: string[], modifiedFiles: string[], readSet?: ReadonlySet<string>): string;
 /**
  * Serialize LLM messages to text for summarization.
  * This prevents the model from treating it as a conversation to continue.

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "15.10.12",
+	"version": "15.11.0",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,10 +35,11 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "15.10.12",
-		"@oh-my-pi/pi-catalog": "15.10.12",
-		"@oh-my-pi/pi-natives": "15.10.12",
-		"@oh-my-pi/pi-utils": "15.10.12",
+		"@oh-my-pi/pi-ai": "15.11.0",
+		"@oh-my-pi/pi-catalog": "15.11.0",
+		"@oh-my-pi/pi-natives": "15.11.0",
+		"@oh-my-pi/pi-utils": "15.11.0",
+		"@oh-my-pi/snapcompact": "15.11.0",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/agent-loop.ts

@@ -564,8 +564,10 @@ async function runLoopBody(
 	streamFn?: StreamFn,
 ): Promise<void> {
 	let firstTurn = true;
-	// Check for steering messages at start (user may have typed while waiting)
-	let pendingMessages: AgentMessage[] = (await config.getSteeringMessages?.()) || [];
+	// Check for steering messages at start (user may have typed while waiting).
+	// Skip when the run is already externally aborted — dequeuing would strand
+	// the messages in a run that is about to die.
+	let pendingMessages: AgentMessage[] = signal?.aborted ? [] : (await config.getSteeringMessages?.()) || [];
 	let harmonyRetryAttempt = 0;
 	let harmonyTruncateResumeCount = 0;
 
@@ -743,7 +745,12 @@ async function runLoopBody(
 
 			stream.push({ type: "turn_end", message, toolResults });
 
-			const steering = steeringMessagesFromExecution ?? ((await config.getSteeringMessages?.()) || []);
+			// On external abort (user interrupt), leave the steering queue intact: the
+			// session aborts then continues, delivering the queue into a fresh run.
+			// Draining it here would inject the messages right before a model call that
+			// instantly aborts — message lands in history, agent never responds.
+			const steering =
+				steeringMessagesFromExecution ?? (signal?.aborted ? [] : (await config.getSteeringMessages?.()) || []);
 			if (hasMoreToolCalls) {
 				// Mid-work: fold any non-interrupting asides into the next turn alongside steering.
 				const asides = resolveAsides(await config.getAsideMessages?.());
@@ -758,8 +765,9 @@ async function runLoopBody(
 
 		// Agent would stop here. Drain non-interrupting asides + follow-up messages.
 		await config.onBeforeYield?.();
-		const asideMessages = resolveAsides(await config.getAsideMessages?.());
-		const followUpMessages = (await config.getFollowUpMessages?.()) || [];
+		// Skip queue drains when externally aborted (same stranding hazard as above).
+		const asideMessages = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.());
+		const followUpMessages = signal?.aborted ? [] : (await config.getFollowUpMessages?.()) || [];
 		if (asideMessages.length > 0 || followUpMessages.length > 0) {
 			// Set as pending so the inner loop processes them before stopping.
 			pendingMessages = [...asideMessages, ...followUpMessages];
@@ -1253,11 +1261,16 @@ async function executeToolCalls(
 	}));
 
 	const checkSteering = async (): Promise<void> => {
-		if (!shouldInterruptImmediately || !getSteeringMessages || interruptState.triggered) {
+		// `signal` (external/user abort) is checked separately from the internal
+		// steeringAbortController: once the run is externally aborted it is
+		// unwinding, and draining the steering queue here would strand the
+		// messages in the dying run instead of leaving them for the post-abort
+		// continue (interruptAndFlushQueuedMessages → Agent.continue()).
+		if (!shouldInterruptImmediately || !getSteeringMessages || interruptState.triggered || signal?.aborted) {
 			return;
 		}
 		const check = steeringCheckTail.then(async () => {
-			if (interruptState.triggered) return;
+			if (interruptState.triggered || signal?.aborted) return;
 			const steering = await getSteeringMessages();
 			if (steering.length > 0) {
 				steeringMessages = steering;

package/src/compaction/branch-summarization.ts

@@ -13,10 +13,10 @@ import { estimateTokens } from "./compaction";
 import type { ReadonlySessionManager, SessionEntry } from "./entries";
 import {
 	type ConvertToLlm,
-	convertToLlm,
 	createBranchSummaryMessage,
 	createCompactionSummaryMessage,
 	createCustomMessage,
+	defaultConvertToLlm,
 } from "./messages";
 import branchSummaryPrompt from "./prompts/branch-summary.md" with { type: "text" };
 import branchSummaryPreamble from "./prompts/branch-summary-preamble.md" with { type: "text" };
@@ -27,6 +27,7 @@ import {
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
 	serializeConversation,
+	stripReadSelector,
 	upsertFileOperations,
 } from "./utils";
 
@@ -214,7 +215,7 @@ export function prepareBranchEntries(entries: SessionEntry[], tokenBudget: numbe
 		if (entry.type === "branch_summary" && !entry.fromExtension && entry.details) {
 			const details = entry.details as BranchSummaryDetails;
 			if (Array.isArray(details.readFiles)) {
-				for (const f of details.readFiles) fileOps.read.add(f);
+				for (const f of details.readFiles) fileOps.read.add(stripReadSelector(f));
 			}
 			if (Array.isArray(details.modifiedFiles)) {
 				// Modified files go into both edited and written for proper deduplication
@@ -288,7 +289,7 @@ export async function generateBranchSummary(
 
 	// Transform to LLM-compatible messages, then serialize to text
 	// Serialization prevents the model from treating it as a conversation to continue
-	const llmMessages = (options.convertToLlm ?? convertToLlm)(messages);
+	const llmMessages = (options.convertToLlm ?? defaultConvertToLlm)(messages);
 	const conversationText = serializeConversation(llmMessages);
 
 	// Build prompt
@@ -329,7 +330,7 @@ export async function generateBranchSummary(
 
 	// Compute file lists and append to summary
 	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
-	summary = upsertFileOperations(summary, readFiles, modifiedFiles);
+	summary = upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
 
 	return {
 		summary: summary || "No summary generated",

package/src/compaction/compaction.ts

@@ -18,11 +18,12 @@ import {
 import { clampThinkingLevelForModel } from "@oh-my-pi/pi-catalog/model-thinking";
 import { countTokens } from "@oh-my-pi/pi-natives";
 import { logger, prompt } from "@oh-my-pi/pi-utils";
+import { SNAPCOMPACT_FRAME_TOKEN_ESTIMATE } from "@oh-my-pi/snapcompact";
 import { type AgentTelemetry, instrumentedCompleteSimple } from "../telemetry";
 import { ThinkingLevel } from "../thinking";
 import type { AgentMessage } from "../types";
 import type { CompactionEntry, SessionEntry } from "./entries";
-import { type ConvertToLlm, convertToLlm, createBranchSummaryMessage, createCustomMessage } from "./messages";
+import { type ConvertToLlm, createBranchSummaryMessage, createCustomMessage, defaultConvertToLlm } from "./messages";
 import {
 	buildOpenAiNativeHistory,
 	getPreservedOpenAiRemoteCompactionData,
@@ -45,6 +46,7 @@ import {
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
 	serializeConversation,
+	stripReadSelector,
 	upsertFileOperations,
 } from "./utils";
 
@@ -74,7 +76,7 @@ function extractFileOperations(
 		if (!prevCompaction.fromExtension && prevCompaction.details) {
 			const details = prevCompaction.details as CompactionDetails;
 			if (Array.isArray(details.readFiles)) {
-				for (const f of details.readFiles) fileOps.read.add(f);
+				for (const f of details.readFiles) fileOps.read.add(stripReadSelector(f));
 			}
 			if (Array.isArray(details.modifiedFiles)) {
 				for (const f of details.modifiedFiles) fileOps.edited.add(f);
@@ -137,7 +139,7 @@ export interface CompactionResult<T = unknown> {
 
 export interface CompactionSettings {
 	enabled: boolean;
-	strategy?: "context-full" | "handoff" | "shake" | "off";
+	strategy?: "context-full" | "handoff" | "shake" | "snapcompact" | "off";
 	thresholdPercent?: number;
 	thresholdTokens?: number;
 	reserveTokens: number;
@@ -285,9 +287,19 @@ export function estimateTokens(message: AgentMessage): number {
 					fragments.push(block.text);
 				} else if (block.type === "thinking") {
 					fragments.push(block.thinking);
+					// Providers charge for the opaque signature/reasoning payload that
+					// rides alongside the thinking text (OpenAI Responses encrypted
+					// reasoning items, Anthropic signed thinking blocks, etc.). Without
+					// counting it, this estimator can read ~half of the provider-reported
+					// usage on thinking-heavy turns — see #2275 for the resulting
+					// compaction-trigger / post-check metric divergence.
+					if (block.thinkingSignature) fragments.push(block.thinkingSignature);
 				} else if (block.type === "toolCall") {
 					fragments.push(block.name);
 					fragments.push(JSON.stringify(block.arguments));
+				} else if (block.type === "redactedThinking") {
+					// Encrypted reasoning blob the provider still bills for on replay.
+					fragments.push(block.data);
 				}
 			}
 			break;
@@ -310,6 +322,10 @@ export function estimateTokens(message: AgentMessage): number {
 		case "branchSummary":
 		case "compactionSummary": {
 			fragments.push(message.summary);
+			if (message.role === "compactionSummary" && message.images) {
+				// Snapcompact frames render at ≥1568px; providers bill the downscaled cap.
+				extra += message.images.length * SNAPCOMPACT_FRAME_TOKEN_ESTIMATE;
+			}
 			break;
 		}
 		default:
@@ -625,7 +641,7 @@ export async function generateSummary(
 
 	// Serialize conversation to text so model doesn't try to continue it
 	// Convert to LLM messages first (handles custom app messages when caller provides a transformer).
-	const llmMessages = (options?.convertToLlm ?? convertToLlm)(currentMessages);
+	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(currentMessages);
 	const conversationText = serializeConversation(llmMessages);
 
 	// Build the prompt with conversation wrapped in tags
@@ -724,7 +740,7 @@ export async function generateHandoff(
 	options: HandoffOptions,
 	signal?: AbortSignal,
 ): Promise<string> {
-	const llmMessages = (options.convertToLlm ?? convertToLlm)(messages);
+	const llmMessages = (options.convertToLlm ?? defaultConvertToLlm)(messages);
 	const requestMessages: Message[] = [
 		...llmMessages,
 		{
@@ -773,7 +789,7 @@ async function generateShortSummary(
 	options?: SummaryOptions,
 ): Promise<string> {
 	const maxTokens = Math.min(512, Math.floor(0.2 * reserveTokens));
-	const llmMessages = (options?.convertToLlm ?? convertToLlm)(recentMessages);
+	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(recentMessages);
 	const conversationText = serializeConversation(llmMessages);
 
 	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
@@ -1010,7 +1026,7 @@ export async function compact(
 				? previousRemoteCompaction.replacementHistory
 				: undefined;
 		const remoteHistory = buildOpenAiNativeHistory(
-			(summaryOptions.convertToLlm ?? convertToLlm)(remoteMessages),
+			(summaryOptions.convertToLlm ?? defaultConvertToLlm)(remoteMessages),
 			model,
 			previousReplacementHistory,
 		);
@@ -1098,7 +1114,7 @@ export async function compact(
 
 	// Compute file lists and append to summary
 	const { readFiles, modifiedFiles } = computeFileLists(fileOps);
-	summary = upsertFileOperations(summary, readFiles, modifiedFiles);
+	summary = upsertFileOperations(summary, readFiles, modifiedFiles, fileOps.read);
 
 	if (!firstKeptEntryId) {
 		throw new Error("First kept entry has no ID - session may need migration");
@@ -1127,7 +1143,7 @@ async function generateTurnPrefixSummary(
 ): Promise<string> {
 	const maxTokens = Math.floor(0.5 * reserveTokens); // Smaller budget for turn prefix
 
-	const llmMessages = (options?.convertToLlm ?? convertToLlm)(messages);
+	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(messages);
 	const conversationText = serializeConversation(llmMessages);
 	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
 	const summarizationMessages = [

package/src/compaction/messages.ts

@@ -51,6 +51,8 @@ export interface CompactionSummaryMessage {
 	shortSummary?: string;
 	tokensBefore: number;
 	providerPayload?: ProviderPayload;
+	/** Snapcompact frames archived by this compaction; appended as image blocks after the summary text. */
+	images?: ImageContent[];
 	timestamp: number;
 }
 
@@ -98,6 +100,7 @@ export function createCompactionSummaryMessage(
 	timestamp: string,
 	shortSummary?: string,
 	providerPayload?: ProviderPayload,
+	images?: ImageContent[],
 ): CompactionSummaryMessage {
 	return {
 		role: "compactionSummary",
@@ -105,6 +108,7 @@ export function createCompactionSummaryMessage(
 		shortSummary,
 		tokensBefore,
 		providerPayload,
+		images: images && images.length > 0 ? images : undefined,
 		timestamp: new Date(timestamp).getTime(),
 	};
 }
@@ -137,6 +141,79 @@ function isCoreCompactionMessage(message: AgentMessage): message is AgentMessage
 	);
 }
 
+/**
+ * Transform a single core-domain agent message to its LLM form; `undefined`
+ * drops it from the provider request.
+ *
+ * Single source of truth for the core roles (user/developer/assistant/
+ * toolResult) and the compaction messages owned by this package. Embedders
+ * with their own app messages (e.g. the coding agent) handle their custom
+ * roles and delegate every core role here — duplicating these cases is how
+ * snapcompact frames once silently fell off the provider request.
+ */
+export function convertMessageToLlm(message: AgentMessage): Message | undefined {
+	if (isCoreCompactionMessage(message)) {
+		switch (message.role) {
+			case "custom":
+			case "hookMessage": {
+				const content =
+					typeof message.content === "string"
+						? [{ type: "text" as const, text: message.content }]
+						: message.content;
+				return {
+					role: "developer",
+					content,
+					attribution: message.attribution,
+					timestamp: message.timestamp,
+				};
+			}
+			case "branchSummary":
+				return {
+					role: "user",
+					content: [
+						{
+							type: "text" as const,
+							text: renderBranchSummaryContext(message.summary),
+						},
+					],
+					attribution: "agent",
+					timestamp: message.timestamp,
+				};
+			case "compactionSummary":
+				return {
+					role: "user",
+					content: [
+						{
+							type: "text" as const,
+							text: renderCompactionSummaryContext(message.summary),
+						},
+						...(message.images ?? []),
+					],
+					attribution: "agent",
+					providerPayload: message.providerPayload,
+					timestamp: message.timestamp,
+				};
+		}
+	}
+
+	switch (message.role) {
+		case "user":
+			return { ...message, attribution: message.attribution ?? "user" };
+		case "developer":
+			return { ...message, attribution: message.attribution ?? "agent" };
+		case "assistant":
+			return message as AssistantMessage;
+		case "toolResult":
+			return {
+				...message,
+				content: getPrunedToolResultContent(message as ToolResultMessage),
+				attribution: message.attribution ?? "agent",
+			};
+		default:
+			return undefined;
+	}
+}
+
 /**
  * Default compaction-domain transformer.
  *
@@ -145,68 +222,5 @@ function isCoreCompactionMessage(message: AgentMessage): message is AgentMessage
  * core LLM roles and the compaction messages owned by this package.
  */
 export function defaultConvertToLlm(messages: AgentMessage[]): Message[] {
-	return messages
-		.map((message): Message | undefined => {
-			if (isCoreCompactionMessage(message)) {
-				switch (message.role) {
-					case "custom":
-					case "hookMessage": {
-						const content =
-							typeof message.content === "string"
-								? [{ type: "text" as const, text: message.content }]
-								: message.content;
-						return {
-							role: "developer",
-							content,
-							attribution: message.attribution,
-							timestamp: message.timestamp,
-						};
-					}
-					case "branchSummary":
-						return {
-							role: "user",
-							content: [
-								{
-									type: "text" as const,
-									text: renderBranchSummaryContext(message.summary),
-								},
-							],
-							attribution: "agent",
-							timestamp: message.timestamp,
-						};
-					case "compactionSummary":
-						return {
-							role: "user",
-							content: [
-								{
-									type: "text" as const,
-									text: renderCompactionSummaryContext(message.summary),
-								},
-							],
-							attribution: "agent",
-							providerPayload: message.providerPayload,
-							timestamp: message.timestamp,
-						};
-				}
-			}
-
-			switch (message.role) {
-				case "user":
-					return { ...message, attribution: message.attribution ?? "user" };
-				case "developer":
-					return { ...message, attribution: message.attribution ?? "agent" };
-				case "assistant":
-					return message as AssistantMessage;
-				case "toolResult":
-					return {
-						...message,
-						content: getPrunedToolResultContent(message as ToolResultMessage),
-						attribution: message.attribution ?? "agent",
-					};
-				default:
-					return undefined;
-			}
-		})
-		.filter(message => message !== undefined);
+	return messages.map(convertMessageToLlm).filter(message => message !== undefined);
 }
-export const convertToLlm = defaultConvertToLlm;

package/src/compaction/prompts/file-operations.md

@@ -1,10 +1,5 @@
-{{#if readFiles.length}}
-{{#xml "read-files"}}
-{{join readFiles "\n"}}
-{{/xml}}
-{{/if}}
-{{#if modifiedFiles.length}}
-{{#xml "modified-files"}}
-{{join modifiedFiles "\n"}}
+{{#if files}}
+{{#xml "files"}}
+{{files}}
 {{/xml}}
 {{/if}}

package/src/compaction/pruning.ts

@@ -3,7 +3,7 @@
  */
 
 import type { ToolResultMessage } from "@oh-my-pi/pi-ai";
-import type { AgentMessage } from "../types";
+import type { AgentMessage, AgentToolCall } from "../types";
 import { estimateTokens } from "./compaction";
 import type { SessionEntry, SessionMessageEntry } from "./entries";
 import {
@@ -12,6 +12,7 @@ import {
 	isSkillReadToolResult,
 	type ProtectedToolMatcher,
 } from "./tool-protection";
+import { splitReadSelector } from "./utils";
 
 export interface PruneConfig {
 	/** Keep the most recent tool output tokens intact. */
@@ -20,6 +21,13 @@ export interface PruneConfig {
 	minimumSavings: number;
 	/** Tool-result protection matchers. String entries protect every result from that tool; predicates may inspect the paired tool call. */
 	protectedTools: ProtectedToolMatcher[];
+	/**
+	 * Optional supersede key function (see {@link SupersedePruneConfig.supersedeKey}).
+	 * When provided, superseded tool results are pruned first — even inside the
+	 * `protectTokens` window — before age-based victims. Absent, behavior is
+	 * unchanged.
+	 */
+	supersedeKey?: SupersedeKeyFn;
 }
 
 export const DEFAULT_PRUNE_CONFIG: PruneConfig = {
@@ -33,6 +41,34 @@ export interface PruneResult {
 	tokensSaved: number;
 }
 
+/** Exact placeholder written over a superseded tool result. */
+export const SUPERSEDED_NOTICE = "[Superseded by a newer read of this file]";
+
+/**
+ * Maps a tool call to a supersede key. Results sharing a key form a group in
+ * which every result except the newest is a supersede candidate. A key `K`
+ * additionally supersedes keys with prefix `K + "\u0000"` (selector-free read
+ * supersedes selector-carrying reads of the same base path). Return
+ * `undefined` to exempt a call from supersede grouping.
+ */
+export type SupersedeKeyFn = (toolName: string, args: Record<string, unknown>) => string | undefined;
+
+export interface SupersedePruneConfig {
+	/** Supersede key function; results sharing a key supersede older ones. */
+	supersedeKey: SupersedeKeyFn;
+	/** Prune a candidate now when all messages after it total at most this many estimated tokens. Default 8 000. */
+	suffixTokenLimit?: number;
+	/** Prune all candidates when the last message is at least this old (prompt cache is cold anyway). Default 30 min. */
+	idleFlushMs?: number;
+	/** Clock override for tests. */
+	now?: number;
+	/** Tool-result protection matchers (same contract as {@link PruneConfig.protectedTools}). */
+	protectedTools: ProtectedToolMatcher[];
+}
+
+const DEFAULT_SUFFIX_TOKEN_LIMIT = 8_000;
+const DEFAULT_IDLE_FLUSH_MS = 30 * 60_000;
+
 function createPrunedNotice(tokens: number): string {
 	return `[Output truncated - ${tokens} tokens]`;
 }
@@ -44,18 +80,121 @@ function getToolResultMessage(entry: SessionEntry): ToolResultMessage | undefine
 	return message as ToolResultMessage;
 }
 
-function estimatePrunedSavings(tokens: number): number {
-	const noticeTokens = Math.ceil(createPrunedNotice(tokens).length / 4);
+function estimatePrunedSavings(tokens: number, notice: string): number {
+	const noticeTokens = Math.ceil(notice.length / 4);
 	return Math.max(0, tokens - noticeTokens);
 }
 
+interface SupersedeCandidate {
+	entry: SessionMessageEntry;
+	message: ToolResultMessage;
+	/** Index of the entry within the `entries` array. */
+	index: number;
+	tokens: number;
+}
+
+/**
+ * Collect superseded tool results: for every unpruned, unprotected tool result
+ * whose paired call resolves a supersede key, a LATER result with the same key
+ * — or with a key that is the `"\u0000"`-prefix parent of this one — marks it
+ * superseded. Returned in message order.
+ */
+function collectSupersededResults(
+	entries: readonly SessionEntry[],
+	toolCallsById: ReadonlyMap<string, AgentToolCall>,
+	supersedeKey: SupersedeKeyFn,
+	protectedTools: readonly ProtectedToolMatcher[],
+): SupersedeCandidate[] {
+	const candidates: SupersedeCandidate[] = [];
+	const seenKeys = new Set<string>();
+	for (let i = entries.length - 1; i >= 0; i--) {
+		const entry = entries[i];
+		const message = getToolResultMessage(entry);
+		if (!message || message.prunedAt !== undefined) continue;
+		const toolCall = toolCallsById.get(message.toolCallId);
+		if (!toolCall) continue;
+		if (isProtectedToolResult(message, toolCall, protectedTools)) continue;
+		const key = supersedeKey(toolCall.name, toolCall.arguments as Record<string, unknown>);
+		if (key === undefined) continue;
+		const separator = key.indexOf("\u0000");
+		const superseded = seenKeys.has(key) || (separator >= 0 && seenKeys.has(key.slice(0, separator)));
+		seenKeys.add(key);
+		if (!superseded) continue;
+		candidates.push({
+			entry: entry as SessionMessageEntry,
+			message,
+			index: i,
+			tokens: estimateTokens(message as AgentMessage),
+		});
+	}
+	return candidates.reverse();
+}
+
+/**
+ * Prune superseded tool results (e.g. stale `read` outputs replaced by a newer
+ * read of the same file). Cheap, incremental, and prompt-cache-aware: a
+ * candidate is pruned now only when the suffix after it is small (tail case —
+ * the read→edit→read loop) or when the context has been idle long enough that
+ * the provider cache is cold anyway (then ALL candidates flush).
+ */
+export function pruneSupersededToolResults(entries: SessionEntry[], config: SupersedePruneConfig): PruneResult {
+	const toolCallsById = collectToolCallsById(entries);
+	const candidates = collectSupersededResults(entries, toolCallsById, config.supersedeKey, config.protectedTools);
+	if (candidates.length === 0) return { prunedCount: 0, tokensSaved: 0 };
+
+	const now = config.now ?? Date.now();
+	let lastMessageTimestamp: number | undefined;
+	for (let i = entries.length - 1; i >= 0; i--) {
+		const entry = entries[i];
+		if (entry.type !== "message") continue;
+		const timestamp = (entry.message as AgentMessage).timestamp;
+		if (typeof timestamp === "number") lastMessageTimestamp = timestamp;
+		break;
+	}
+	const idle =
+		lastMessageTimestamp !== undefined && now - lastMessageTimestamp >= (config.idleFlushMs ?? DEFAULT_IDLE_FLUSH_MS);
+
+	let toPrune: SupersedeCandidate[];
+	if (idle) {
+		toPrune = candidates;
+	} else {
+		const suffixTokenLimit = config.suffixTokenLimit ?? DEFAULT_SUFFIX_TOKEN_LIMIT;
+		// suffixTokens[i] = estimated tokens of all messages strictly after entry i.
+		const suffixTokens = new Array<number>(entries.length);
+		let accumulated = 0;
+		for (let i = entries.length - 1; i >= 0; i--) {
+			suffixTokens[i] = accumulated;
+			const entry = entries[i];
+			if (entry.type === "message") accumulated += estimateTokens(entry.message as AgentMessage);
+		}
+		toPrune = candidates.filter(candidate => suffixTokens[candidate.index] <= suffixTokenLimit);
+	}
+	if (toPrune.length === 0) return { prunedCount: 0, tokensSaved: 0 };
+
+	const prunedAt = Date.now();
+	let tokensSaved = 0;
+	for (const candidate of toPrune) {
+		candidate.message.content = [{ type: "text", text: SUPERSEDED_NOTICE }];
+		candidate.message.prunedAt = prunedAt;
+		tokensSaved += estimatePrunedSavings(candidate.tokens, SUPERSEDED_NOTICE);
+	}
+	return { prunedCount: toPrune.length, tokensSaved };
+}
+
 export function pruneToolOutputs(entries: SessionEntry[], config: PruneConfig = DEFAULT_PRUNE_CONFIG): PruneResult {
 	let accumulatedTokens = 0;
 	let tokensSaved = 0;
 	let prunedCount = 0;
 
-	const candidates: Array<{ entry: SessionMessageEntry; tokens: number }> = [];
+	const candidates: Array<{ entry: SessionMessageEntry; tokens: number; superseded: boolean }> = [];
 	const toolCallsById = collectToolCallsById(entries);
+	const supersededMessages = config.supersedeKey
+		? new Set(
+				collectSupersededResults(entries, toolCallsById, config.supersedeKey, config.protectedTools).map(
+					candidate => candidate.message,
+				),
+			)
+		: undefined;
 
 	for (let i = entries.length - 1; i >= 0; i--) {
 		const entry = entries[i];
@@ -70,17 +209,23 @@ export function pruneToolOutputs(entries: SessionEntry[], config: PruneConfig =
 			continue;
 		}
 
-		if (accumulatedTokens < config.protectTokens || isProtected) {
+		// Superseded results are pruned first: they bypass the protect window
+		// (a stale copy of re-read content is dead weight at any age).
+		const superseded = supersededMessages?.has(message) ?? false;
+		if (!superseded && (accumulatedTokens < config.protectTokens || isProtected)) {
 			accumulatedTokens += tokens;
 			continue;
 		}
 
-		candidates.push({ entry: entry as SessionMessageEntry, tokens });
+		candidates.push({ entry: entry as SessionMessageEntry, tokens, superseded });
 		accumulatedTokens += tokens;
 	}
 
 	for (const candidate of candidates) {
-		tokensSaved += estimatePrunedSavings(candidate.tokens);
+		tokensSaved += estimatePrunedSavings(
+			candidate.tokens,
+			candidate.superseded ? SUPERSEDED_NOTICE : createPrunedNotice(candidate.tokens),
+		);
 	}
 
 	if (tokensSaved < config.minimumSavings || candidates.length === 0) {
@@ -90,10 +235,31 @@ export function pruneToolOutputs(entries: SessionEntry[], config: PruneConfig =
 	const prunedAt = Date.now();
 	for (const candidate of candidates) {
 		const message = candidate.entry.message as ToolResultMessage;
-		message.content = [{ type: "text", text: createPrunedNotice(candidate.tokens) }];
+		message.content = [
+			{ type: "text", text: candidate.superseded ? SUPERSEDED_NOTICE : createPrunedNotice(candidate.tokens) },
+		];
 		message.prunedAt = prunedAt;
 		prunedCount++;
 	}
 
 	return { prunedCount, tokensSaved };
 }
+
+/**
+ * Supersede key for the `read` tool: the file path with the trailing line/raw
+ * selector stripped (the read tool's own splitter grammar via
+ * {@link splitReadSelector}, e.g. `src/foo.ts:50-200`, `:2-4:raw`).
+ * Internal/URL-scheme paths (`skill://…`, `https://…`) are exempt.
+ * Selector-free reads key on the bare path; selector-carrying reads key on
+ * `path + "\u0000" + selector`, so two reads collide only when the newer is
+ * selector-free or the selectors are identical (the pass's prefix rule lets a
+ * bare-path read supersede selector-carrying reads of the same file).
+ */
+export function readToolSupersedeKey(toolName: string, args: Record<string, unknown>): string | undefined {
+	if (toolName !== "read") return undefined;
+	const path = args.path;
+	if (typeof path !== "string" || path.length === 0) return undefined;
+	if (path.includes("://")) return undefined;
+	const { path: base, sel } = splitReadSelector(path);
+	return sel === undefined ? base : `${base}\u0000${sel}`;
+}

package/src/compaction/utils.ts

@@ -3,7 +3,7 @@
  */
 
 import type { Message } from "@oh-my-pi/pi-ai";
-import { prompt } from "@oh-my-pi/pi-utils";
+import { formatGroupedPaths, prompt } from "@oh-my-pi/pi-utils";
 import type { AgentMessage } from "../types";
 import fileOperationsTemplate from "./prompts/file-operations.md" with { type: "text" };
 import summarizationSystemPrompt from "./prompts/summarization-system.md" with { type: "text" };
@@ -26,6 +26,55 @@ export function createFileOps(): FileOperations {
 	};
 }
 
+// Read-tool selector grammar, mirrored from the conservative filesystem splitter in
+// packages/coding-agent/src/tools/path-utils.ts (splitPathAndSel). Keep in sync.
+// A trailing `:chunk` is a selector only when it is a line-range list
+// (`50`, `50-200`, `50+10`, `5-16,960-973`, `..` alias), `raw`, or `conflicts` —
+// alone or as a `range:raw` / `raw:range` compound.
+const RANGE_CHUNK_SRC = String.raw`L?\d+(?:(?:[-+]|\.\.)L?\d+|-|\.\.)?`;
+const RANGE_LIST_SRC = `${RANGE_CHUNK_SRC}(?:,${RANGE_CHUNK_SRC})*`;
+const READ_SELECTOR_RE = new RegExp(`^(?:${RANGE_LIST_SRC}|raw|conflicts)$`, "i");
+const READ_RANGE_ONLY_RE = new RegExp(`^${RANGE_LIST_SRC}$`, "i");
+const READ_RAW_ONLY_RE = /^raw$/i;
+
+/**
+ * Split a read-tool path into its base path and trailing selector, mirroring the
+ * read tool's own splitter. Single source of the grammar in this package: the
+ * file-operations list strips selectors via {@link stripReadSelector}, and the
+ * supersede-prune pass keys on both parts via `readToolSupersedeKey`.
+ */
+export function splitReadSelector(path: string): { path: string; sel?: string } {
+	const colon = path.lastIndexOf(":");
+	if (colon <= 0) return { path };
+	const candidate = path.slice(colon + 1);
+	if (!READ_SELECTOR_RE.test(candidate)) return { path };
+	let base = path.slice(0, colon);
+	let sel = candidate;
+	// Compound trailing selector: `path:1-50:raw` or `path:raw:1-50`.
+	const inner = base.lastIndexOf(":");
+	if (inner > 0) {
+		const innerCandidate = base.slice(inner + 1);
+		const innerIsRaw = READ_RAW_ONLY_RE.test(innerCandidate);
+		const outerIsRaw = READ_RAW_ONLY_RE.test(candidate);
+		const innerIsRange = READ_RANGE_ONLY_RE.test(innerCandidate);
+		const outerIsRange = READ_RANGE_ONLY_RE.test(candidate);
+		if ((innerIsRaw && outerIsRange) || (innerIsRange && outerIsRaw)) {
+			sel = `${innerCandidate}:${candidate}`;
+			base = base.slice(0, inner);
+		}
+	}
+	return { path: base, sel };
+}
+
+/**
+ * Strip a trailing read-tool selector (`:50-200`, `:raw`, `:1-50:raw`, `:conflicts`, …)
+ * so the same file read with different line ranges dedupes to one `<files>` entry
+ * and matches its write/edit path when computing Read/Write/RW markers.
+ */
+export function stripReadSelector(path: string): string {
+	return splitReadSelector(path).path;
+}
+
 /**
  * Extract file operations from tool calls in an assistant message.
  */
@@ -46,7 +95,7 @@ export function extractFileOpsFromMessage(message: AgentMessage, fileOps: FileOp
 
 		switch (block.name) {
 			case "read":
-				fileOps.read.add(path);
+				fileOps.read.add(stripReadSelector(path));
 				break;
 			case "write":
 				fileOps.written.add(path);
@@ -70,32 +119,48 @@ export function computeFileLists(fileOps: FileOperations): { readFiles: string[]
 }
 
 /**
- * Format file operations as XML tags for summary.
+ * Format file operations as one `<files>` tag: a grouped, prefix-folded
+ * directory tree (find-tool shape — `# dir/` headers, bare basenames) with a
+ * ` (Read)` / ` (Write)` / ` (RW)` marker per file instead of separate
+ * read/modified lists. `readSet` is the cumulative read set (`fileOps.read`),
+ * used to tell modified files that were also read (RW) from blind writes.
  */
 const FILE_OPERATION_SUMMARY_LIMIT = 20;
 
-function truncateFileList(files: string[]): string[] {
-	if (files.length <= FILE_OPERATION_SUMMARY_LIMIT) return files;
-	const omitted = files.length - FILE_OPERATION_SUMMARY_LIMIT;
-	return [...files.slice(0, FILE_OPERATION_SUMMARY_LIMIT), `… (${omitted} more files omitted)`];
-}
-
 function stripFileOperationTags(summary: string): string {
-	const withoutReadFiles = summary.replace(/<read-files>[\s\S]*?<\/read-files>\s*/g, "");
-	const withoutModifiedFiles = withoutReadFiles.replace(/<modified-files>[\s\S]*?<\/modified-files>\s*/g, "");
-	return withoutModifiedFiles.trimEnd();
+	// Legacy <read-files>/<modified-files> tags are still stripped so summaries
+	// written before the combined <files> tag self-heal on the next compaction.
+	return summary
+		.replace(/<files>[\s\S]*?<\/files>\s*/g, "")
+		.replace(/<read-files>[\s\S]*?<\/read-files>\s*/g, "")
+		.replace(/<modified-files>[\s\S]*?<\/modified-files>\s*/g, "")
+		.trimEnd();
 }
-export function formatFileOperations(readFiles: string[], modifiedFiles: string[]): string {
+export function formatFileOperations(
+	readFiles: string[],
+	modifiedFiles: string[],
+	readSet?: ReadonlySet<string>,
+): string {
 	if (readFiles.length === 0 && modifiedFiles.length === 0) return "";
-	return prompt.render(fileOperationsTemplate, {
-		readFiles: truncateFileList(readFiles),
-		modifiedFiles: truncateFileList(modifiedFiles),
-	});
+	const mode = new Map<string, "Read" | "Write" | "RW">();
+	for (const file of readFiles) mode.set(file, "Read");
+	for (const file of modifiedFiles) mode.set(file, readSet?.has(file) ? "RW" : "Write");
+	const all = [...mode.keys()].sort();
+	let files = formatGroupedPaths(all.slice(0, FILE_OPERATION_SUMMARY_LIMIT), path => ` (${mode.get(path)})`);
+	if (all.length > FILE_OPERATION_SUMMARY_LIMIT) {
+		files += `\n… (${all.length - FILE_OPERATION_SUMMARY_LIMIT} more files omitted)`;
+	}
+	return prompt.render(fileOperationsTemplate, { files });
 }
 
-export function upsertFileOperations(summary: string, readFiles: string[], modifiedFiles: string[]): string {
+export function upsertFileOperations(
+	summary: string,
+	readFiles: string[],
+	modifiedFiles: string[],
+	readSet?: ReadonlySet<string>,
+): string {
 	const baseSummary = stripFileOperationTags(summary);
-	const fileOperations = formatFileOperations(readFiles, modifiedFiles);
+	const fileOperations = formatFileOperations(readFiles, modifiedFiles, readSet);
 	if (!fileOperations) return baseSummary;
 	if (!baseSummary) return fileOperations;
 	return `${baseSummary}\n\n${fileOperations}`;