@oh-my-pi/pi-coding-agent 15.7.2 → 15.7.3

package/dist/types/session/agent-session.d.ts

@@ -14,7 +14,7 @@
  */
 import type { InMemorySnapshotStore } from "@oh-my-pi/hashline";
 import { type Agent, type AgentEvent, type AgentMessage, type AgentState, type AgentTool, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
-import { type CompactionResult } from "@oh-my-pi/pi-agent-core/compaction";
+import { type CompactionResult, type ShakeConfig } from "@oh-my-pi/pi-agent-core/compaction";
 import type { AssistantMessage, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
 import { Effort } from "@oh-my-pi/pi-ai";
 import { type AsyncJob, type AsyncJobDeliveryState, AsyncJobManager } from "../async";
@@ -36,7 +36,7 @@ import { type FileSlashCommand } from "../extensibility/slash-commands";
 import { GoalRuntime } from "../goals/runtime";
 import type { Goal, GoalModeState } from "../goals/state";
 import type { HindsightSessionState } from "../hindsight/state";
-import { type MnemosyneSessionState } from "../mnemosyne/state";
+import { type MnemopiSessionState } from "../mnemopi/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import { type SecretObfuscator } from "../secrets/obfuscator";
@@ -47,16 +47,17 @@ import { type TodoItem, type TodoPhase } from "../tools/todo-write";
 import type { ClientBridge } from "./client-bridge";
 import { type CustomMessage } from "./messages";
 import type { BranchSummaryEntry, NewSessionOptions, SessionContext, SessionManager } from "./session-manager";
+import type { ShakeMode, ShakeResult } from "./shake-types";
 import { ToolChoiceQueue } from "./tool-choice-queue";
 import { YieldQueue } from "./yield-queue";
 /** Session-specific events that extend the core AgentEvent */
 export type AgentSessionEvent = AgentEvent | {
     type: "auto_compaction_start";
     reason: "threshold" | "overflow" | "idle" | "incomplete";
-    action: "context-full" | "handoff";
+    action: "context-full" | "handoff" | "shake" | "shake-summary";
 } | {
     type: "auto_compaction_end";
-    action: "context-full" | "handoff";
+    action: "context-full" | "handoff" | "shake" | "shake-summary";
     result: CompactionResult | undefined;
     aborted: boolean;
     willRetry: boolean;
@@ -121,6 +122,7 @@ export interface AsyncJobSnapshot {
     recent: AsyncJobSnapshotItem[];
     delivery: AsyncJobDeliveryState;
 }
+export type { ShakeMode, ShakeResult };
 export interface AgentSessionConfig {
     agent: Agent;
     sessionManager: SessionManager;
@@ -315,7 +317,7 @@ export declare class AgentSession {
     get providerSessionState(): Map<string, ProviderSessionState>;
     getHindsightSessionState(): HindsightSessionState | undefined;
     setHindsightSessionState(state: HindsightSessionState | undefined): HindsightSessionState | undefined;
-    getMnemosyneSessionState(): MnemosyneSessionState | undefined;
+    getMnemopiSessionState(): MnemopiSessionState | undefined;
     /** TTSR manager for time-traveling stream rules */
     get ttsrManager(): TtsrManager | undefined;
     /** Whether a TTSR abort is pending (stream was aborted to inject rules) */
@@ -615,12 +617,13 @@ export declare class AgentSession {
     fork(): Promise<boolean>;
     /**
      * Set model directly.
-     * Validates API key, saves to session and settings.
+     * Validates API key and saves to the active session. Persists settings only when requested.
      * @throws Error if no API key available for the model
      */
     setModel(model: Model, role?: string, options?: {
         selector?: string;
         thinkingLevel?: ThinkingLevel;
+        persist?: boolean;
     }): Promise<void>;
     /**
      * Set model temporarily (for this session only).
@@ -647,28 +650,26 @@ export declare class AgentSession {
      */
     getRoleModelCycle(roleOrder: readonly string[]): RoleModelCycle | undefined;
     /**
-     * Apply a resolved role model as the active model, persisting the choice to
-     * settings under its role. Mirrors the non-temporary branch of
-     * {@link cycleRoleModels} and is shared with the plan-approval model slider.
+     * Apply a resolved role model as the active model without changing global
+     * settings. Shared with role cycling and the plan-approval model slider.
      */
     applyRoleModel(entry: ResolvedRoleModel): Promise<void>;
     /**
      * Cycle through configured role models in a fixed order.
-     * Skips missing roles.
+     * Skips missing roles and changes only the active session model.
      * @param roleOrder - Order of roles to cycle through (e.g., ["slow", "default", "smol"])
-     * @param options - Optional settings: `temporary` to not persist to settings
+     * @param direction - "forward" (default) or "backward"
      */
-    cycleRoleModels(roleOrder: readonly string[], options?: {
-        temporary?: boolean;
-    }): Promise<RoleModelCycleResult | undefined>;
+    cycleRoleModels(roleOrder: readonly string[], direction?: "forward" | "backward"): Promise<RoleModelCycleResult | undefined>;
     /**
      * Get all available models with valid API keys.
      */
     getAvailableModels(): Model[];
     /**
-     * Set the thinking level. `auto` enables per-turn classification (session-level,
-     * never written to the session log); a concrete level clears auto. The effective
-     * metadata-clamped level is saved to the session/settings only when it changes.
+     * Set the thinking level. `auto` enables per-turn classification; the selector
+     * itself is never written to the session log, but resolved concrete levels are
+     * persisted when real user turns are classified so resumed sessions keep the
+     * last resolved effort instead of reverting to pending auto.
      */
     setThinkingLevel(level: ConfiguredThinkingLevel | undefined, persist?: boolean): void;
     /**
@@ -730,6 +731,27 @@ export declare class AgentSession {
     dropImages(): Promise<{
         removed: number;
     }>;
+    /**
+     * Surgically reduce context by dropping heavy content ("shake").
+     *
+     * - `images` delegates to {@link dropImages}.
+     * - `elide` replaces whole tool-call results and large fenced/XML blocks
+     *   with short placeholders that embed an `artifact://` recovery link.
+     * - `summary` extractively compresses the same regions with the configured
+     *   local on-device model (`providers.shakeSummaryModel`), falling back to
+     *   the elide placeholder per region (or wholesale when the local model is
+     *   unavailable). Never calls a remote/cloud LLM.
+     *
+     * Mutates the branch in place, persists via `rewriteEntries`, replays the
+     * rebuilt context through the agent, and tears down provider sessions that
+     * cache message identity — same rewrite contract as {@link dropImages}.
+     *
+     * No-op (zero counts) when nothing is eligible.
+     */
+    shake(mode: ShakeMode, opts?: {
+        config?: ShakeConfig;
+        signal?: AbortSignal;
+    }): Promise<ShakeResult>;
     /**
      * Manually compact the session context.
      * Aborts current agent operation first.
@@ -880,6 +902,7 @@ export declare class AgentSession {
         promptText: string;
         onTextDelta?: (delta: string) => void;
         signal?: AbortSignal;
+        dedupeReply?: boolean;
     }): Promise<{
         replyText: string;
         assistantMessage: AssistantMessage;

package/dist/types/session/session-manager.d.ts

@@ -306,6 +306,24 @@ export declare class SessionManager {
     getCwd(): string;
     /** Get usage statistics across all assistant messages in the session. */
     getUsageStatistics(): UsageStatistics;
+    /**
+     * Open a new per-turn budget window: snapshot the cumulative output baseline,
+     * reset the eval-subagent counter, and set the (optional) ceiling. Called once
+     * per real user message; `total` is null when no `+Nk` directive was present.
+     */
+    beginTurnBudget(total: number | null, hard: boolean): void;
+    /** Record output tokens consumed by an eval-spawned subagent in the current turn. */
+    recordEvalSubagentOutput(output: number): void;
+    /**
+     * Current turn budget for the eval `budget` helper: the ceiling (null = none),
+     * output tokens spent this turn (main loop + eval-spawned subagents, no
+     * double-count), and whether the ceiling is hard.
+     */
+    getTurnBudget(): {
+        total: number | null;
+        spent: number;
+        hard: boolean;
+    };
     getSessionDir(): string;
     getSessionId(): string;
     getSessionFile(): string | undefined;

package/dist/types/session/session-storage.d.ts

@@ -65,6 +65,12 @@ export declare class MemorySessionStorage implements SessionStorage {
     ensureDirSync(_dir: string): void;
     existsSync(path: string): boolean;
     writeTextSync(path: string, content: string): void;
+    /**
+     * Internal O(1) append used by {@link MemorySessionStorageWriter}. Lazily
+     * creates the entry. External callers should go through `openWriter()`
+     * rather than touching the mirror directly.
+     */
+    appendChunkSync(path: string, chunk: string): void;
     readTextSync(path: string): string;
     statSync(path: string): SessionStorageStat;
     listFilesSync(dir: string, pattern: string): string[];

package/dist/types/session/shake-types.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Public shape of the `shake` operation, kept in a dependency-free leaf module
+ * so slash-command registries and controllers can import `formatShakeSummary`
+ * without pulling in the heavy `agent-session` module graph (which would form
+ * an import cycle through the slash-command registry).
+ */
+/** Mode selector for `AgentSession.shake`. */
+export type ShakeMode = "elide" | "summary" | "images";
+/** Outcome of an `AgentSession.shake` run. */
+export interface ShakeResult {
+    mode: ShakeMode;
+    /** Whole tool-call results dropped/compressed. */
+    toolResultsDropped: number;
+    /** Large fenced/XML blocks dropped/compressed. */
+    blocksDropped: number;
+    /** Image blocks removed (images mode only). */
+    imagesDropped?: number;
+    /** Estimated context tokens reclaimed. */
+    tokensFreed: number;
+    /** Session artifact holding the dropped originals, when persisted. */
+    artifactId?: string;
+}
+/** One-line operator summary of a {@link ShakeResult} (shared by TUI + ACP). */
+export declare function formatShakeSummary(result: ShakeResult): string;

package/dist/types/task/executor.d.ts

@@ -11,7 +11,7 @@ import { type Skill } from "../extensibility/skills";
 import type { HindsightSessionState } from "../hindsight/state";
 import type { LocalProtocolOptions } from "../internal-urls";
 import type { MCPManager } from "../mcp/manager";
-import type { MnemosyneSessionState } from "../mnemosyne/state";
+import type { MnemopiSessionState } from "../mnemopi/state";
 import type { ArtifactManager } from "../session/artifacts";
 import type { AuthStorage } from "../session/auth-storage";
 import type { ContextFileEntry } from "../tools";
@@ -65,7 +65,7 @@ export interface ExecutorOptions {
      */
     parentArtifactManager?: ArtifactManager;
     parentHindsightSessionState?: HindsightSessionState;
-    parentMnemosyneSessionState?: MnemosyneSessionState;
+    parentMnemopiSessionState?: MnemopiSessionState;
     /** Parent agent's eval executor session id. Subagents reuse it so eval state is shared. */
     parentEvalSessionId?: string;
     /**

package/dist/types/tiny/models.d.ts

@@ -65,7 +65,7 @@ export declare const ONLINE_MEMORY_MODEL_KEY = "online";
 /** Recommended local model for memory tasks when none is named. */
 export declare const DEFAULT_MEMORY_LOCAL_MODEL_KEY = "qwen3-1.7b";
 /**
- * Local models for Mnemosyne memory tasks (fact extraction + consolidation).
+ * Local models for Mnemopi memory tasks (fact extraction + consolidation).
  * These are larger (1B-1.7B) than the title models: structured extraction and
  * faithful summarization need more capacity than 3-6 word titles. All q4.
  * Ranking/recipe rationale lives in docs/local-models.md.
@@ -113,6 +113,20 @@ export declare const TINY_MEMORY_MODEL_OPTIONS: ({
 })[];
 export declare function isTinyMemoryLocalModelKey(value: string): value is TinyMemoryLocalModelKey;
 export declare function getTinyMemoryModelSpec(key: TinyMemoryLocalModelKey): (typeof TINY_MEMORY_LOCAL_MODELS)[number];
+/**
+ * Shake-summary models. Shake's `summary` mode (and the `shake-summary`
+ * compaction strategy) compress heavy regions strictly on-device — there is no
+ * online/remote option, so this registry reuses the local memory models only.
+ */
+export declare const SHAKE_SUMMARY_MODEL_VALUES: readonly ["qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+export type ShakeSummaryModelKey = (typeof SHAKE_SUMMARY_MODEL_VALUES)[number];
+export declare const SHAKE_SUMMARY_MODEL_OPTIONS: {
+    value: "gemma-3-1b" | "lfm2-1.2b" | "qwen2.5-1.5b" | "qwen3-1.7b";
+    label: "Gemma 3 1B" | "LFM2 1.2B" | "Qwen2.5 1.5B" | "Qwen3 1.7B";
+    description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction." | "Best extraction granularity (atomic facts); weaker consolidation." | "Fastest load; solid all-rounder, slightly noisier extraction labels." | "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+}[];
+/** Default shake-summary local model when none is named. */
+export declare const DEFAULT_SHAKE_SUMMARY_MODEL_KEY: ShakeSummaryModelKey;
 /** Any local model key (title or memory), used by the shared inference worker. */
 export type TinyLocalModelKey = TinyTitleLocalModelKey | TinyMemoryLocalModelKey;
 /** Resolve a local model spec by key across both the title and memory registries. */

package/dist/types/tiny/title-protocol.d.ts

@@ -30,6 +30,10 @@ export type TinyTitleWorkerInbound = {
     modelKey: TinyLocalModelKey;
     prompt: string;
     maxTokens?: number;
+    /** Optional assistant-turn prefix appended after the generation prompt to pin output format. */
+    prefill?: string;
+    /** Optional literal stop string; generation halts once it appears in the decoded tail. */
+    stop?: string;
 } | {
     type: "download";
     id: string;

package/dist/types/tools/index.d.ts

@@ -6,7 +6,9 @@ import type { Settings } from "../config/settings";
 import type { Skill } from "../extensibility/skills";
 import type { GoalModeState, GoalRuntime } from "../goals";
 import type { HindsightSessionState } from "../hindsight/state";
-import type { MnemosyneSessionState } from "../mnemosyne/state";
+import type { LocalProtocolOptions } from "../internal-urls";
+import type { MCPManager } from "../mcp";
+import type { MnemopiSessionState } from "../mnemopi/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import type { ArtifactManager } from "../session/artifacts";
@@ -107,8 +109,8 @@ export interface ToolSession {
     getSessionId?: () => string | null;
     /** Get Hindsight runtime state for this agent session. */
     getHindsightSessionState?: () => HindsightSessionState | undefined;
-    /** Get Mnemosyne runtime state for this agent session. */
-    getMnemosyneSessionState?: () => MnemosyneSessionState | undefined;
+    /** Get Mnemopi runtime state for this agent session. */
+    getMnemopiSessionState?: () => MnemopiSessionState | undefined;
     /** Agent identity used for IRC routing. Returns the registry id (e.g. "0-Main", "0-AuthLoader"). */
     getAgentId?: () => string | null;
     /** Look up a registered tool by name (used by the eval js backend's tool bridge). */
@@ -136,6 +138,10 @@ export interface ToolSession {
     modelRegistry?: import("../config/model-registry").ModelRegistry;
     /** Agent output manager for unique agent:// IDs across task invocations */
     agentOutputManager?: AgentOutputManager;
+    /** MCP manager visible to subagents without relying on the process-global singleton. */
+    mcpManager?: MCPManager;
+    /** Local protocol root to propagate to nested subagents and eval-created agents. */
+    localProtocolOptions?: LocalProtocolOptions;
     /** Settings instance for passing to subagents */
     settings: Settings;
     /** Plan mode state (if active) */
@@ -144,6 +150,16 @@ export interface ToolSession {
     getGoalModeState?: () => GoalModeState | undefined;
     /** Goal runtime for the active agent session. */
     getGoalRuntime?: () => GoalRuntime | undefined;
+    /** Get cumulative session usage statistics (input/output tokens, cost). */
+    getUsageStatistics?: () => import("../session/session-manager").UsageStatistics;
+    /** Current per-turn token budget {total, spent, hard} for the eval `budget` helper. */
+    getTurnBudget?: () => {
+        total: number | null;
+        spent: number;
+        hard: boolean;
+    };
+    /** Record output tokens consumed by an eval-spawned subagent toward the current turn budget. */
+    recordEvalSubagentUsage?: (output: number) => void;
     /** Bridge to the connected client (e.g. ACP editor host). Tools should route fs/terminal/permission requests through this when available. */
     getClientBridge?: () => ClientBridge | undefined;
     /** Get compact conversation context for subagents (excludes tool results, system prompts) */

package/dist/types/tools/memory-edit.d.ts

@@ -32,7 +32,7 @@ export declare class MemoryEditTool implements AgentTool<typeof memoryEditSchema
     }, z.core.$strip>;
     readonly strict = true;
     readonly loadMode = "discoverable";
-    readonly summary = "Update, forget, or invalidate Mnemosyne memories";
+    readonly summary = "Update, forget, or invalidate Mnemopi memories";
     constructor(session: ToolSession);
     static createIf(session: ToolSession): MemoryEditTool | null;
     execute(_id: string, params: MemoryEditParams): Promise<AgentToolResult>;

package/examples/sdk/09-api-keys-and-oauth.ts

@@ -26,7 +26,7 @@ console.log("Session with default auth storage and model registry");
 
 // Custom auth storage location
 const customAuthStorage = await AuthStorage.create("/tmp/my-app/agent.db");
-const customModelRegistry = new ModelRegistry(customAuthStorage, "/tmp/my-app/models.json");
+const customModelRegistry = await ModelRegistry.create(customAuthStorage, "/tmp/my-app/models.json");
 
 await createAgentSession({
 	sessionManager: SessionManager.inMemory(),
@@ -45,7 +45,7 @@ await createAgentSession({
 console.log("Session with runtime API key override");
 
 // No models.json - only built-in models
-const simpleRegistry = new ModelRegistry(authStorage); // null = no models.json
+const simpleRegistry = await ModelRegistry.create(authStorage); // null = no models.json
 await createAgentSession({
 	sessionManager: SessionManager.inMemory(),
 	authStorage,

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "15.7.2",
+	"version": "15.7.3",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,7 +35,7 @@
 		"check": "biome check . && bun run check:types",
 		"check:types": "tsgo -p tsconfig.json --noEmit",
 		"lint": "biome lint .",
-		"test": "bun test",
+		"test": "bun test --parallel",
 		"fix": "biome check --write --unsafe . && bun run format-prompts && bun run generate-docs-index",
 		"fmt": "biome format --write . && bun run format-prompts",
 		"format-prompts": "bun scripts/format-prompts.ts",
@@ -47,14 +47,14 @@
 		"@agentclientprotocol/sdk": "0.22.1",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "15.7.2",
-		"@oh-my-pi/omp-stats": "15.7.2",
-		"@oh-my-pi/pi-agent-core": "15.7.2",
-		"@oh-my-pi/pi-ai": "15.7.2",
-		"@oh-my-pi/pi-mnemosyne": "15.7.2",
-		"@oh-my-pi/pi-natives": "15.7.2",
-		"@oh-my-pi/pi-tui": "15.7.2",
-		"@oh-my-pi/pi-utils": "15.7.2",
+		"@oh-my-pi/hashline": "15.7.3",
+		"@oh-my-pi/omp-stats": "15.7.3",
+		"@oh-my-pi/pi-agent-core": "15.7.3",
+		"@oh-my-pi/pi-ai": "15.7.3",
+		"@oh-my-pi/pi-mnemopi": "15.7.3",
+		"@oh-my-pi/pi-natives": "15.7.3",
+		"@oh-my-pi/pi-tui": "15.7.3",
+		"@oh-my-pi/pi-utils": "15.7.3",
 		"@puppeteer/browsers": "^3.0.4",
 		"@types/turndown": "5.0.6",
 		"@xterm/headless": "^6.0.0",

package/src/autoresearch/tools/run-experiment.ts

@@ -1,12 +1,12 @@
-import * as childProcess from "node:child_process";
 import * as fs from "node:fs";
 import * as path from "node:path";
 import { Text } from "@oh-my-pi/pi-tui";
 import { formatBytes } from "@oh-my-pi/pi-utils";
 import * as z from "zod/v4";
+import { executeBash } from "../../exec/bash-executor";
 import type { ToolDefinition } from "../../extensibility/extensions";
 import type { Theme } from "../../modes/theme/theme";
-import { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, truncateTail } from "../../session/streaming-output";
+import { DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, TailBuffer, truncateTail } from "../../session/streaming-output";
 import { replaceTabs, shortenPath } from "../../tools/render-utils";
 import * as git from "../../utils/git";
 import { parseWorkDirDirtyPaths } from "../git";
@@ -15,7 +15,6 @@ import {
 	EXPERIMENT_MAX_LINES,
 	formatElapsed,
 	formatNum,
-	killTree,
 	parseAsiLines,
 	parseMetricLines,
 	tryGitPrefix,
@@ -117,7 +116,7 @@ export function createRunExperimentTool(
 			let execution: ProcessExecutionResult;
 			try {
 				execution = await executeProcess({
-					command: ["bash", "-lc", resolvedCommand],
+					command: resolvedCommand,
 					cwd: ctx.cwd,
 					logPath: benchmarkLogPath,
 					timeoutMs,
@@ -268,68 +267,18 @@ export function createRunExperimentTool(
 	};
 }
 async function executeProcess(opts: {
-	command: string[];
+	command: string;
 	cwd: string;
 	logPath: string;
 	timeoutMs: number;
 	signal?: AbortSignal;
 	onProgress?(details: ProgressSnapshot): void;
 }): Promise<ProcessExecutionResult> {
-	const { promise, resolve, reject } = Promise.withResolvers<ProcessExecutionResult>();
-	const child = childProcess.spawn(opts.command[0] ?? "bash", opts.command.slice(1), {
-		cwd: opts.cwd,
-		detached: true,
-		stdio: ["ignore", "pipe", "pipe"],
-	});
-
-	const tailChunks: Buffer[] = [];
-	let chunksBytes = 0;
-	let killedByTimeout = false;
-	let resolved = false;
-	let writeStream: fs.WriteStream | undefined = fs.createWriteStream(opts.logPath);
-	let forceKillTimeout: NodeJS.Timeout | undefined;
-
-	const closeWriteStream = (): Promise<void> => {
-		if (!writeStream) return Promise.resolve();
-		const stream = writeStream;
-		writeStream = undefined;
-		return new Promise<void>((resolveClose, rejectClose) => {
-			stream.end((error?: Error | null) => {
-				if (error) {
-					rejectClose(error);
-					return;
-				}
-				resolveClose();
-			});
-		});
-	};
-
-	const cleanup = (): void => {
-		if (progressTimer) clearInterval(progressTimer);
-		if (timeoutHandle) clearTimeout(timeoutHandle);
-		if (forceKillTimeout) clearTimeout(forceKillTimeout);
-		opts.signal?.removeEventListener("abort", abortHandler);
-	};
-
-	const finish = (callback: () => void): void => {
-		if (resolved) return;
-		resolved = true;
-		cleanup();
-		callback();
-	};
-
-	const appendChunk = (data: Buffer): void => {
-		writeStream?.write(data);
-		tailChunks.push(data);
-		chunksBytes += data.length;
-		while (chunksBytes > DEFAULT_MAX_BYTES * 2 && tailChunks.length > 1) {
-			const removed = tailChunks.shift();
-			if (removed) chunksBytes -= removed.length;
-		}
-	};
+	const tailBuffer = new TailBuffer(DEFAULT_MAX_BYTES * 2);
 
+	const startedAt = Date.now();
 	const snapshot = (): ProgressSnapshot => {
-		const tail = truncateTail(Buffer.concat(tailChunks).toString("utf8"), {
+		const tail = truncateTail(tailBuffer.text(), {
 			maxBytes: DEFAULT_MAX_BYTES,
 			maxLines: DEFAULT_MAX_LINES,
 		});
@@ -342,71 +291,54 @@ async function executeProcess(opts: {
 		};
 	};
 
-	const killTreeWithEscalation = (): void => {
-		if (!child.pid) return;
-		killTree(child.pid);
-		forceKillTimeout = setTimeout(() => {
-			if (child.pid) killTree(child.pid, "SIGKILL");
-		}, 1_000);
-		forceKillTimeout.unref?.();
-	};
-
-	const startedAt = Date.now();
 	const progressTimer = opts.onProgress
 		? setInterval(() => {
 				opts.onProgress?.(snapshot());
 			}, 1000)
 		: undefined;
-	const timeoutHandle =
-		opts.timeoutMs > 0
-			? setTimeout(() => {
-					killedByTimeout = true;
-					killTreeWithEscalation();
-				}, opts.timeoutMs)
-			: undefined;
 
-	const abortHandler = (): void => {
-		killTreeWithEscalation();
+	const logSink = Bun.file(opts.logPath).writer();
+	let logSinkClosed = false;
+	const closeLogSink = async (): Promise<void> => {
+		if (logSinkClosed) return;
+		logSinkClosed = true;
+		await logSink.end();
 	};
-	if (opts.signal?.aborted) {
-		abortHandler();
-	} else {
-		opts.signal?.addEventListener("abort", abortHandler, { once: true });
-	}
-
-	child.stdout?.on("data", data => {
-		appendChunk(data);
-	});
-	child.stderr?.on("data", data => {
-		appendChunk(data);
-	});
-	child.on("error", error => {
-		void closeWriteStream().finally(() => {
-			finish(() => reject(error));
+	try {
+		const result = await executeBash(opts.command, {
+			cwd: opts.cwd,
+			sessionKey: `autoresearch:${opts.cwd}`,
+			timeout: opts.timeoutMs > 0 ? opts.timeoutMs : 2_147_000_000,
+			signal: opts.signal,
+			chunkThrottleMs: 0,
+			onChunk: chunk => {
+				tailBuffer.append(chunk);
+				logSink.write(chunk);
+			},
 		});
-	});
-	child.on("close", async code => {
-		try {
-			await closeWriteStream();
-			if (opts.signal?.aborted) {
-				finish(() => reject(new Error("aborted")));
-				return;
-			}
-			const output = await fs.promises.readFile(opts.logPath, "utf8");
-			finish(() =>
-				resolve({
-					exitCode: code,
-					killed: killedByTimeout,
-					logPath: opts.logPath,
-					output,
-				}),
-			);
-		} catch (error) {
-			finish(() => reject(error));
+		await closeLogSink();
+		if (opts.signal?.aborted) {
+			throw new Error("aborted");
 		}
-	});
 
-	return promise;
+		const output = await fs.promises.readFile(opts.logPath, "utf8");
+
+		return {
+			exitCode: result.exitCode ?? null,
+			killed: result.cancelled,
+			logPath: opts.logPath,
+			output,
+		};
+	} finally {
+		if (progressTimer) clearInterval(progressTimer);
+		if (!logSinkClosed) {
+			try {
+				await closeLogSink();
+			} catch {
+				// Preserve the command failure when cleanup is best-effort.
+			}
+		}
+	}
 }
 
 function buildRunText(details: RunDetails, outputPreview: string, bestMetric: number | null): string {