@ferris1225/pi-subagents 4.2.7 → 4.2.8

package/src/format.ts

@@ -1,165 +1,181 @@
-/**
- * Pure formatting/result helpers shared by the subagent tool and completion
- * delivery: usage rendering, completion blocks, synthetic result constructors,
- * and run-id matching.
- */
-
-import type { AgentConfig } from "./agents.ts";
-import { formatTaskSummary } from "./monitor.ts";
-import { emptyUsage } from "./rpc-run.ts";
-import {
-	getResultOutput,
-	isFailedResult,
-	truncateResultOutput,
-	writeResultArtifact,
-	type SingleResult,
-	type UsageStats,
-} from "./spawn.ts";
-
-export function queuedResult(agent: AgentConfig, task: string, thinking?: string): SingleResult {
-	return {
-		agent: agent.name,
-		task,
-		exitCode: -1,
-		messages: [],
-		stderr: "",
-		usage: emptyUsage(),
-		model: agent.model,
-		...(thinking ? { thinking } : {}),
-	};
-}
-
-export function failedStartResult(agentName: string, task: string, errorMessage: string): SingleResult {
-	return {
-		agent: agentName,
-		task,
-		exitCode: 1,
-		messages: [],
-		stderr: errorMessage,
-		usage: emptyUsage(),
-		errorMessage,
-		dispatchFailed: true,
-	};
-}
-
-/** Failed result for a background task that crashed with an exception (spawn
- * infra, delivery API, ...) instead of returning a normal result. */
-export function dispatchFailedResult(agent: AgentConfig, task: string, error: unknown, thinking?: string): SingleResult {
-	const errorMessage = error instanceof Error ? error.message : String(error);
-	return {
-		...queuedResult(agent, task, thinking),
-		exitCode: 1,
-		stderr: errorMessage,
-		stopReason: "error",
-		errorMessage,
-		dispatchFailed: true,
-	};
-}
-
-function formatTokens(count: number): string {
-	if (count >= 1_000_000) return `${(count / 1_000_000).toFixed(1)}M`;
-	if (count >= 1_000) return `${(count / 1_000).toFixed(1)}k`;
-	return String(count);
-}
-
-export function formatUsage(usage: UsageStats): string {
-	const parts: string[] = [];
-	if (usage.turns) parts.push(`${usage.turns} turn${usage.turns > 1 ? "s" : ""}`);
-	if (usage.input) parts.push(`↑${formatTokens(usage.input)}`);
-	if (usage.output) parts.push(`↓${formatTokens(usage.output)}`);
-	if (usage.cacheRead) parts.push(`R${formatTokens(usage.cacheRead)}`);
-	if (usage.cacheWrite) parts.push(`W${formatTokens(usage.cacheWrite)}`);
-	if (usage.cost) parts.push(`$${usage.cost.toFixed(4)}`);
-	return parts.join(" ");
-}
-
-export interface CompletionFormatOptions {
-	/** Project-scoped directory the full result is written to when the output
-	 * is truncated: <projectRoot>/results. */
-	resultRoot?: string;
-}
-
-export function formatCompletionBlock(
-	result: SingleResult,
-	maxResultLines: number,
-	options: CompletionFormatOptions = {},
-): string {
-	const failed = isFailedResult(result);
-	const failedTools = result.failedTools ?? [];
-	const status = failed ? "failed" : "completed";
-	const usage = formatUsage(result.usage);
-	const output = getResultOutput(result);
-	const { text, truncated } = truncateResultOutput(output, maxResultLines);const fallbackNote = result.modelFallbackFrom
-		? ` (selected model ${result.modelFallbackFrom} failed → main ${result.model ?? "dynamic default"})`
-		: "";
-	const startupRetryNote = result.startupRetries
-		? ` (recovered after ${result.startupRetries} startup retr${result.startupRetries === 1 ? "y" : "ies"} — concurrent pi startup race)`
-		: "";
-	const runNote = result.runId !== undefined ? ` · run #${result.runId}` : "";
-	const lines = [`### [${result.agent}] ${status}${usage ? ` (${usage})` : ""}${fallbackNote}${startupRetryNote}${runNote}`, "", `Task: ${formatTaskSummary(result.task, 80, false)}`, ""];
-	if (result.isolation === "worktree") {
-		const isolation =
-			result.integrationStatus === "integrated"
-				? "worktree · changes integrated into the original working tree"
-				: result.integrationStatus === "no_changes"
-					? "worktree · no changes; temporary worktree removed"
-					: result.integrationStatus === "retained"
-						? result.integrationApplied
-							? "worktree · changes applied, but cleanup failed; recovery artifacts retained"
-							: "worktree · integration failed; recovery artifacts retained"
-						: "worktree · isolated";
-		lines.push(`Isolation: ${isolation}`);
-		if (result.integrationWorktreePath) lines.push(`Retained worktree: ${result.integrationWorktreePath}`);
-		if (result.integrationPatchPath) lines.push(`Retained patch: ${result.integrationPatchPath}`);
-		if (result.integrationError) lines.push(`Integration error: ${result.integrationError}`);
-		lines.push("");
-	}
-	lines.push(text);
-	// Failed-tool diagnostics ride along only when the run itself failed: they
-	// explain the failure, while on a successful run a transient failed call
-	// (no-match grep, rejected edit) is noise the agent already worked around.
-	if (failed && failedTools.length > 0) {
-		lines.push(
-			"",
-			`⚠ ${failedTools.length} failed tool call${failedTools.length === 1 ? "" : "s"}:`,
-			...failedTools.map((tool) => `- ${tool.toolName}: ${tool.error.trim() || "(no output)"}`),
-		);
-	}
-	if (truncated) {
-		// The full text lives on disk so the main agent can read it on demand.
-		const artifact = options.resultRoot
-			? writeResultArtifact(output, result.agent, options.resultRoot)
-			: "(result root unavailable)";
-		lines.push("", `(output truncated to ${maxResultLines} lines; full result: ${artifact})`);
-	}
-	return lines.join("\n");
-}
-
-/** Instruction appended to a model-level failure: the sub-agent's provider never
- * produced usable output (or the run stalled), so the task is handed back to the
- * main window instead of being left as a dead failure. When the run preserved a
- * session with earlier work (and the run id is known), steer the main agent to
- * RESUME it in-context once a model is available, instead of re-dispatching
- * fresh (which would re-scan everything). */
-export function modelLevelTakeoverNote(result: SingleResult, opts?: { runId?: number }): string {
-	const retry = result.modelFallbackFrom ? ", and the current main model also failed" : "";
-	const detail = result.errorMessage?.trim();
-	const cause = detail
-		? `its model/provider call failed (${detail})`
-		: "its model was unavailable or failed (or the run stalled)";
-	const sessionPreserved = Boolean(result.sessionDir && result.sessionId) && opts?.runId !== undefined;
-	const recovery = sessionPreserved
-		? ` The sub-agent's earlier work in this run is preserved. Once a model is available again, call subagent_control with { action: "resume", id: ${opts!.runId} } to CONTINUE it in-context (it keeps the same run id and does not re-scan), or execute the task in the main window with your own tools.`
-		: ` Please execute this task in the main window with your own tools; do not re-dispatch it as a sub-agent.`;
-	return `The sub-agent could not complete this task: ${cause}${retry}.${recovery}`;
-}
-
-/** Resolve a run-id request to actual ids: an exact numeric match always wins
- * (so "1" never fans out to 10, 11, …); only when no exact match exists does a
- * prefix match run, as a convenience for partial ids. Keeps single-digit lookups
- * from returning — or, for subagent_stop, acting on — a whole prefix family. */
-export function matchRunIds(ids: number[], requested: string): number[] {
-	const exact = ids.filter((id) => String(id) === requested);
-	if (exact.length > 0) return exact;
-	return ids.filter((id) => String(id).startsWith(requested));
-}
+/**
+ * Pure formatting/result helpers shared by the subagent tool and completion
+ * delivery: usage rendering, completion blocks, synthetic result constructors,
+ * and run-id matching.
+ */
+
+import type { AgentConfig } from "./agents.ts";
+import { runLabel, shrinkRunLabel } from "./monitor.ts";
+import { emptyUsage } from "./rpc-run.ts";
+import {
+	getResultOutput,
+	isFailedResult,
+	truncateResultOutput,
+	writeResultArtifact,
+	type SingleResult,
+	type UsageStats,
+} from "./spawn.ts";
+
+export function queuedResult(agent: AgentConfig, task: string, thinking?: string): SingleResult {
+	return {
+		agent: agent.name,
+		task,
+		exitCode: -1,
+		messages: [],
+		stderr: "",
+		usage: emptyUsage(),
+		model: agent.model,
+		...(thinking ? { thinking } : {}),
+	};
+}
+
+export function failedStartResult(agentName: string, task: string, errorMessage: string): SingleResult {
+	return {
+		agent: agentName,
+		task,
+		exitCode: 1,
+		messages: [],
+		stderr: errorMessage,
+		usage: emptyUsage(),
+		errorMessage,
+		dispatchFailed: true,
+	};
+}
+
+/** Failed result for a background task that crashed with an exception (spawn
+ * infra, delivery API, ...) instead of returning a normal result. */
+export function dispatchFailedResult(agent: AgentConfig, task: string, error: unknown, thinking?: string): SingleResult {
+	const errorMessage = error instanceof Error ? error.message : String(error);
+	return {
+		...queuedResult(agent, task, thinking),
+		exitCode: 1,
+		stderr: errorMessage,
+		stopReason: "error",
+		errorMessage,
+		dispatchFailed: true,
+	};
+}
+
+function formatTokens(count: number): string {
+	if (count >= 1_000_000) return `${(count / 1_000_000).toFixed(1)}M`;
+	if (count >= 1_000) return `${(count / 1_000).toFixed(1)}k`;
+	return String(count);
+}
+
+export function formatUsage(usage: UsageStats): string {
+	const parts: string[] = [];
+	if (usage.turns) parts.push(`${usage.turns} turn${usage.turns > 1 ? "s" : ""}`);
+	if (usage.input) parts.push(`↑${formatTokens(usage.input)}`);
+	if (usage.output) parts.push(`↓${formatTokens(usage.output)}`);
+	if (usage.cacheRead) parts.push(`R${formatTokens(usage.cacheRead)}`);
+	if (usage.cacheWrite) parts.push(`W${formatTokens(usage.cacheWrite)}`);
+	if (usage.cost) parts.push(`$${usage.cost.toFixed(4)}`);
+	return parts.join(" ");
+}
+
+export interface CompletionFormatOptions {
+	/** Project-scoped directory the full result is written to when the output
+	 * is truncated: <projectRoot>/results. */
+	resultRoot?: string;
+}
+
+/** Display width of the task-derived label folded into the completion heading:
+ * tight enough to stay a hint rather than a second copy of the task. */
+const COMPLETION_LABEL_MAX = 24;
+
+export function formatCompletionBlock(
+	result: SingleResult,
+	maxResultLines: number,
+	options: CompletionFormatOptions = {},
+): string {
+	const failed = isFailedResult(result);
+	const failedTools = result.failedTools ?? [];
+	const status = failed ? "failed" : "completed";
+	const usage = formatUsage(result.usage);
+	const output = getResultOutput(result);
+	const { text, truncated } = truncateResultOutput(output, maxResultLines);
+	const fallbackNote = result.modelFallbackFrom
+		? ` (selected model ${result.modelFallbackFrom} failed → main ${result.model ?? "dynamic default"})`
+		: "";
+	const startupRetryNote = result.startupRetries
+		? ` (recovered after ${result.startupRetries} startup retr${result.startupRetries === 1 ? "y" : "ies"} — concurrent pi startup race)`
+		: "";
+	const runNote = result.runId !== undefined ? ` · run #${result.runId}` : "";
+	// A short task-derived label, not the task itself: the parent authored the
+	// task and still has it in context, but a wide fan-out of same-agent runs
+	// needs more than a run id to tell completions apart.
+	const label = shrinkRunLabel(runLabel(result.task), COMPLETION_LABEL_MAX);
+	const lines = [
+		`### [${result.agent}${label ? `·${label}` : ""}] ${status}${usage ? ` (${usage})` : ""}${fallbackNote}${startupRetryNote}${runNote}`,
+		"",
+	];
+	if (result.isolation === "worktree") {
+		const isolation =
+			result.integrationStatus === "integrated"
+				? "worktree · changes integrated into the original working tree"
+				: result.integrationStatus === "no_changes"
+					? "worktree · no changes; temporary worktree removed"
+					: result.integrationStatus === "retained"
+						? result.integrationApplied
+							? "worktree · changes applied, but cleanup failed; recovery artifacts retained"
+							: "worktree · integration failed; recovery artifacts retained"
+						: "worktree · isolated";
+		lines.push(`Isolation: ${isolation}`);
+		if (result.integrationWorktreePath) lines.push(`Retained worktree: ${result.integrationWorktreePath}`);
+		if (result.integrationPatchPath) lines.push(`Retained patch: ${result.integrationPatchPath}`);
+		if (result.integrationError) lines.push(`Integration error: ${result.integrationError}`);
+		lines.push("");
+	}
+	lines.push(text);
+	// Failed-tool diagnostics ride along only when the run itself failed: they
+	// explain the failure, while on a successful run a transient failed call
+	// (no-match grep, rejected edit) is noise the agent already worked around.
+	if (failed && failedTools.length > 0) {
+		lines.push(
+			"",
+			`⚠ ${failedTools.length} failed tool call${failedTools.length === 1 ? "" : "s"}:`,
+			...failedTools.map((tool) => `- ${tool.toolName}: ${tool.error.trim() || "(no output)"}`),
+		);
+	}
+	if (truncated) {
+		// The full text lives on disk so the main agent can read it on demand.
+		const artifact = options.resultRoot
+			? writeResultArtifact(output, result.agent, options.resultRoot)
+			: "(result root unavailable)";
+		// State the loss (shown vs total) and condition the read: handing the
+		// parent both a summary and a full-text entrance invites the same content
+		// into its context twice.
+		const totalLines = output.split("\n").length;
+		lines.push("", `(${maxResultLines} of ${totalLines} lines shown; full result ${artifact} — read only if these are insufficient)`);
+	}
+	return lines.join("\n");
+}
+
+/** Instruction appended to a model-level failure: the sub-agent's provider never
+ * produced usable output (or the run stalled), so the task is handed back to the
+ * main window instead of being left as a dead failure. When the run preserved a
+ * session with earlier work (and the run id is known), steer the main agent to
+ * RESUME it in-context once a model is available, instead of re-dispatching
+ * fresh (which would re-scan everything). */
+export function modelLevelTakeoverNote(result: SingleResult, opts?: { runId?: number }): string {
+	const retry = result.modelFallbackFrom ? ", and the current main model also failed" : "";
+	const detail = result.errorMessage?.trim();
+	const cause = detail
+		? `its model/provider call failed (${detail})`
+		: "its model was unavailable or failed (or the run stalled)";
+	const sessionPreserved = Boolean(result.sessionDir && result.sessionId) && opts?.runId !== undefined;
+	const recovery = sessionPreserved
+		? ` The sub-agent's earlier work in this run is preserved. Once a model is available again, call subagent_control with { action: "resume", id: ${opts!.runId} } to CONTINUE it in-context (it keeps the same run id and does not re-scan), or execute the task in the main window with your own tools.`
+		: ` Please execute this task in the main window with your own tools; do not re-dispatch it as a sub-agent.`;
+	return `The sub-agent could not complete this task: ${cause}${retry}.${recovery}`;
+}
+
+/** Resolve a run-id request to actual ids: an exact numeric match always wins
+ * (so "1" never fans out to 10, 11, …); only when no exact match exists does a
+ * prefix match run, as a convenience for partial ids. Keeps single-digit lookups
+ * from returning — or, for subagent_stop, acting on — a whole prefix family. */
+export function matchRunIds(ids: number[], requested: string): number[] {
+	const exact = ids.filter((id) => String(id) === requested);
+	if (exact.length > 0) return exact;
+	return ids.filter((id) => String(id).startsWith(requested));
+}

package/src/index.ts

@@ -1,100 +1,102 @@
-/**
- * pi-subagents — focused sub-agent delegation for pi.
- *
- * Assembly point: builds the shared runtime and registers everything.
- * The heavy lifting lives in focused modules:
- *   - dispatch.ts         — tool contract, managed role policy, internal steps
- *   - thread-lifecycle.ts — stable generations, controls, final integration/delivery
- *   - tools.ts            — subagent_control / subagent_stop
- *   - announcements.ts — session-start recovery, notices, and widget install
- *   - widget.ts        — active-only TUI run status
- *   - runtime.ts       — shared per-session state
- *
- * Also registers the `/subagents-setup` command and a `before_agent_start` hook
- * that injects a delegation directive into the parent system prompt so the main
- * model uses the tool proactively.
- *
- * The tool is not registered inside child sub-agent processes, which prevents
- * runaway recursion and keeps child context windows clean.
- */
-
-import { getAgentDir, type ExtensionAPI } from "@earendil-works/pi-coding-agent";
-import { Text } from "@earendil-works/pi-tui";
-import { discoverAgents } from "./agents.ts";
-import { registerAnnouncements } from "./announcements.ts";
-import { getConfigPath, loadConfig } from "./config.ts";
-import { registerSubagentTool } from "./dispatch.ts";
-import { matchRunIds } from "./format.ts";
-import { buildDelegationDirective } from "./prompt.ts";
-import { createRuntime } from "./runtime.ts";
-import { runSetup } from "./setup.ts";
-import { currentSubagentDepth } from "./spawn.ts";
-import { bootstrapDurableState } from "./thread-lifecycle.ts";
-import { registerLookupTools } from "./tools.ts";
-import { clearActiveRunsWidget } from "./widget.ts";
-
-export { matchRunIds };
-
-export default function (pi: ExtensionAPI): void {
-	const configPath = getConfigPath(getAgentDir());
-	const runtime = createRuntime(pi, configPath);
-
-	// Recursion guard: sub-agent children are leaf processes. The `subagent` tool is
-	// excluded from their toolset at spawn (--exclude-tools); this check is defense
-	// in depth so a child can never expose the tool back to its model, even if
-	// another extension ignores the depth marker.
-	if (currentSubagentDepth() >= 1) {
-		pi.registerCommand("subagents-setup", {
-			description: "Configure pi-subagents (unavailable in nested sub-agent processes)",
-			handler: async (_args, ctx) => {
-				ctx.ui.notify("pi-subagents setup is unavailable in nested sub-agent processes.", "warning");
-			},
-		});
-		return;
-	}
-
-	pi.registerMessageRenderer("subagent-result", (message, _options, theme) =>
-		new Text(
-			`${theme.fg("toolTitle", theme.bold("subagent result"))}\n${message.content}`,
-			0,
-			0,
-		),
-	);
-
-	pi.on("session_shutdown", async (_event, ctx) => {
-		clearActiveRunsWidget(ctx);
-		await runtime.shutdown();
-	});
-
-	registerSubagentTool(pi, runtime);
-	registerLookupTools(pi, runtime);
-
-	pi.registerCommand("subagents-setup", {
-		description: "Configure pi-subagents: agents, selected models, capability-aware thinking, and runtime settings",
-		handler: async (_args, ctx) => {
-			await runSetup(ctx, configPath);
-		},
-	});
-
-	registerAnnouncements(pi, runtime);
-
-	// Durable bootstrap: restore parked threads from the manifest so a reload or
-	// restart keeps status and resume working, then age out old records and sweep
-	// leaked temp/state directories. Registration never blocks on it and every
-	// stage is best-effort; the restore pass is published as runtime.durableRestore
-	// so the tools and the session-start notice wait for it instead of racing it.
-	void bootstrapDurableState(runtime);
-
-	// Proactive dispatch: inject the delegation directive into the parent system prompt.
-	pi.on("before_agent_start", async (event, ctx) => {
-		const config = await loadConfig(configPath);
-		const { agents } = discoverAgents(ctx.cwd, {
-			scope: config.agentScope,
-			enabledNames: config.enabledAgents,
-			projectTrusted: ctx.isProjectTrusted?.() === true,
-		});
-		const directive = buildDelegationDirective(agents);
-		if (!directive) return undefined;
-		return { systemPrompt: `${event.systemPrompt}\n${directive}` };
-	});
-}
+/**
+ * pi-subagents — focused sub-agent delegation for pi.
+ *
+ * Assembly point: builds the shared runtime and registers everything.
+ * The heavy lifting lives in focused modules:
+ *   - dispatch.ts         — tool contract, managed role policy, internal steps
+ *   - thread-lifecycle.ts — stable generations, controls, final integration/delivery
+ *   - tools.ts            — subagent_control / subagent_stop
+ *   - announcements.ts — session-start recovery, notices, and widget install
+ *   - widget.ts        — active-only TUI run status
+ *   - runtime.ts       — shared per-session state
+ *
+ * Also registers the `/subagents-setup` command and a `before_agent_start` hook
+ * that injects a delegation directive into the parent system prompt so the main
+ * model uses the tool proactively.
+ *
+ * The tool is not registered inside child sub-agent processes, which prevents
+ * runaway recursion and keeps child context windows clean.
+ */
+
+import { getAgentDir, type ExtensionAPI } from "@earendil-works/pi-coding-agent";
+import { Text } from "@earendil-works/pi-tui";
+import { discoverAgents } from "./agents.ts";
+import { registerAnnouncements } from "./announcements.ts";
+import { getConfigPath, loadConfig } from "./config.ts";
+import { registerSubagentTool } from "./dispatch.ts";
+import { matchRunIds } from "./format.ts";
+import { buildDelegationDirective } from "./prompt.ts";
+import { createRuntime } from "./runtime.ts";
+import { runSetup } from "./setup.ts";
+import { currentSubagentDepth } from "./spawn.ts";
+import { clearActiveRunsStatus } from "./status.ts";
+import { bootstrapDurableState } from "./thread-lifecycle.ts";
+import { registerLookupTools } from "./tools.ts";
+import { clearActiveRunsWidget } from "./widget.ts";
+
+export { matchRunIds };
+
+export default function (pi: ExtensionAPI): void {
+	const configPath = getConfigPath(getAgentDir());
+	const runtime = createRuntime(pi, configPath);
+
+	// Recursion guard: sub-agent children are leaf processes. The `subagent` tool is
+	// excluded from their toolset at spawn (--exclude-tools); this check is defense
+	// in depth so a child can never expose the tool back to its model, even if
+	// another extension ignores the depth marker.
+	if (currentSubagentDepth() >= 1) {
+		pi.registerCommand("subagents-setup", {
+			description: "Configure pi-subagents (unavailable in nested sub-agent processes)",
+			handler: async (_args, ctx) => {
+				ctx.ui.notify("pi-subagents setup is unavailable in nested sub-agent processes.", "warning");
+			},
+		});
+		return;
+	}
+
+	pi.registerMessageRenderer("subagent-result", (message, _options, theme) =>
+		new Text(
+			`${theme.fg("toolTitle", theme.bold("subagent result"))}\n${message.content}`,
+			0,
+			0,
+		),
+	);
+
+	pi.on("session_shutdown", async (_event, ctx) => {
+		clearActiveRunsStatus(ctx);
+		clearActiveRunsWidget(ctx);
+		await runtime.shutdown();
+	});
+
+	registerSubagentTool(pi, runtime);
+	registerLookupTools(pi, runtime);
+
+	pi.registerCommand("subagents-setup", {
+		description: "Configure pi-subagents: agents, selected models, capability-aware thinking, and runtime settings",
+		handler: async (_args, ctx) => {
+			await runSetup(ctx, configPath);
+		},
+	});
+
+	registerAnnouncements(pi, runtime);
+
+	// Durable bootstrap: restore parked threads from the manifest so a reload or
+	// restart keeps status and resume working, then age out old records and sweep
+	// leaked temp/state directories. Registration never blocks on it and every
+	// stage is best-effort; the restore pass is published as runtime.durableRestore
+	// so the tools and the session-start notice wait for it instead of racing it.
+	void bootstrapDurableState(runtime);
+
+	// Proactive dispatch: inject the delegation directive into the parent system prompt.
+	pi.on("before_agent_start", async (event, ctx) => {
+		const config = await loadConfig(configPath);
+		const { agents } = discoverAgents(ctx.cwd, {
+			scope: config.agentScope,
+			enabledNames: config.enabledAgents,
+			projectTrusted: ctx.isProjectTrusted?.() === true,
+		});
+		const directive = buildDelegationDirective(agents);
+		if (!directive) return undefined;
+		return { systemPrompt: `${event.systemPrompt}\n${directive}` };
+	});
+}

package/src/prompt.ts

@@ -40,7 +40,7 @@ export function buildDelegationDirective(
 	];
 
 	const handoffRules = [
-		"Dispatch never blocks or ends your turn — keep working; each completion resumes you automatically. Never sleep or poll for it.",
+		"Dispatch never blocks or ends your turn — keep working, but only on what the children are not: never re-read a scope you just delegated. Each completion resumes you automatically; never sleep or poll for it.",
 		"Results are already shown; add only your conclusion or next action, never a restatement.",
 		"Never declare the overall task done while a dispatched run is still active.",
 	];

package/src/runtime.ts

@@ -131,6 +131,12 @@ export interface SubagentRuntime {
 
 export function createRuntime(pi: ExtensionAPI, configPath: string): SubagentRuntime {
 	const backgroundQueue = new BackgroundTaskQueue(resolveSubagentConcurrency());
+	// Compaction reads the whole history, summarizes it, and replaces it. A
+	// completion injected into that window can be swallowed by the summary,
+	// silently losing a result a child spent minutes producing — so delivery is
+	// held until compaction settles (either outcome) instead.
+	let compactionInFlight = false;
+	let heldCompletions: CompletionMessageItem[] = [];
 
 	const runtime: SubagentRuntime = {
 		configPath,
@@ -142,6 +148,10 @@ export function createRuntime(pi: ExtensionAPI, configPath: string): SubagentRun
 		restoredNotified: false,
 		sendCompletionGroup: (items) => {
 			if (!runtime.sessionActive || items.length === 0) return;
+			if (compactionInFlight) {
+				heldCompletions.push(...items);
+				return;
+			}
 			// A result arriving for one run does not mean sibling runs are done.
 			// Computing this at delivery (emit) time — not when the item was
 			// pushed — reflects the current monitor state, since finishing runs
@@ -242,6 +252,10 @@ export function createRuntime(pi: ExtensionAPI, configPath: string): SubagentRun
 			}
 			const preflights = [...runtime.preflightOperations];
 			runtime.completionBatcher.dispose();
+			// Held items are dropped exactly like the batcher's abandoned ones: the
+			// session is gone, so there is no window left to deliver them into.
+			heldCompletions = [];
+			compactionInFlight = false;
 			runtime.backgroundQueue.cancelAll();
 			// Await live RPC process-tree cleanup and continuation preflight rollback
 			// before persisting records or releasing ownership maps.
@@ -308,6 +322,25 @@ export function createRuntime(pi: ExtensionAPI, configPath: string): SubagentRun
 		},
 	};
 
+	// Hold delivery across compaction. `session_before_compact` opens the
+	// window; BOTH terminal events close it, because a failed or aborted
+	// compaction that never released would strand every held result forever.
+	pi.on("session_before_compact", () => {
+		compactionInFlight = true;
+	});
+	const releaseHeldCompletions = (): void => {
+		compactionInFlight = false;
+		if (heldCompletions.length === 0) return;
+		const items = heldCompletions;
+		heldCompletions = [];
+		// Re-enter the normal path now that the gate is open: the active-runs
+		// footer and the steer/nextTurn choice must reflect delivery time, not
+		// the moment the items were held.
+		runtime.sendCompletionGroup(items);
+	};
+	pi.on("session_compact", releaseHeldCompletions);
+	pi.on("session_compact_failed", releaseHeldCompletions);
+
 	runtime.completionBatcher = createCompletionBatcher<CompletionMessageItem>({
 		emit: runtime.sendCompletionGroup,
 	});