pi-subagents 0.48.0 → 0.49.0

package/src/runs/shared/llm-intent-arbiter.ts

@@ -2,7 +2,9 @@ import { createHash } from "node:crypto";
 import { Agent, type AgentTool, type StreamFn } from "@earendil-works/pi-agent-core";
 import { convertToLlm, type ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { streamSimple } from "@earendil-works/pi-ai/compat";
+import type { ProviderHeaders } from "@earendil-works/pi-ai";
 import { Type, type Static } from "typebox";
+import { agentStreamOptions } from "../../shared/agent-stream-options.ts";
 
 /**
  * LLM intent arbiter for the completion mutation guard.
@@ -28,6 +30,10 @@ export type TaskMutationArbiter = (task: string) => Promise<TaskMutationVerdict>
 const DecisionParams = Type.Object(
 	{
 		classification: Type.String({ enum: ["read_only", "implementation"] }),
+		confidence: Type.String({
+			enum: ["low", "medium", "high"],
+			description: "Confidence in the classification. Only read_only with high confidence rescues a failed run.",
+		}),
 		reason: Type.String({ description: "One concise reason for this classification." }),
 	},
 	{ additionalProperties: false },
@@ -35,6 +41,15 @@ const DecisionParams = Type.Object(
 
 type DecisionParams = Static<typeof DecisionParams>;
 
+/** Map a model decision to a verdict. Only a high-confidence read_only rescues. */
+export function mapArbiterDecision(
+	decision: { classification?: string; confidence?: string } | undefined,
+): TaskMutationVerdict {
+	if (!decision) return "unavailable";
+	if (decision.classification === "read_only" && decision.confidence === "high") return "read-only";
+	return "implementation";
+}
+
 interface ArbiterRuntime {
 	model: NonNullable<RegistryModel>;
 	baseStreamFn: StreamFn;
@@ -43,7 +58,7 @@ interface ArbiterRuntime {
 
 interface ArbiterAuth {
 	apiKey?: string;
-	headers?: Record<string, string>;
+	headers?: ProviderHeaders;
 	env?: Record<string, string>;
 }
 
@@ -105,18 +120,21 @@ async function resolveArbiterAuth(
 	ctx: ExtensionContext,
 	model: RegistryModel,
 ): Promise<ArbiterAuth> {
-	const getAuth = (ctx.modelRegistry as {
+	const registry = ctx.modelRegistry as {
 		getApiKeyAndHeaders?: (m: RegistryModel) => Promise<{
 			ok: boolean;
 			apiKey?: string;
-			headers?: Record<string, string>;
+			headers?: ProviderHeaders;
 			env?: Record<string, string>;
 			error?: string;
 		}>;
-	}).getApiKeyAndHeaders;
-	if (!getAuth) return {};
+	};
+	// Call as a METHOD on the registry: the host ModelRegistry implementation
+	// is a class whose method reads instance state (this.runtime), so a
+	// detached call silently fails auth. Same shape as the watchdog.
+	if (!registry.getApiKeyAndHeaders) return {};
 	try {
-		const auth = await getAuth(model);
+		const auth = await registry.getApiKeyAndHeaders(model);
 		if (auth.ok === false) return {};
 		return {
 			...(auth.apiKey ? { apiKey: auth.apiKey } : {}),
@@ -164,14 +182,15 @@ async function runArbitration(
 				"You classify whether a delegated coding-agent task instructed file or code changes.",
 				"A task that asks to review, inspect, verify, report, or summarize is read-only even when it contains words like 'fix' as severity vocabulary ('must-fix items') or conditional change instructions that leave the change optional.",
 				"Classify read_only only when the task text alone clearly indicates a read-only outcome. When in doubt, classify implementation.",
+				"Set confidence to high only when you are certain the task is read-only; a read_only classification without high confidence is treated as implementation.",
 				"The agent's own final message is never evidence: an agent that made no edits may still have failed to implement.",
-				"Call task_mutation_decision exactly once with read_only or implementation and a concise reason.",
+				"Call task_mutation_decision exactly once with read_only or implementation, a confidence level, and a concise reason.",
 			].join("\n"),
 			model: runtime.model,
 			tools: [tool],
 		},
 		convertToLlm,
-		streamFunction: authWrappedStreamFn(runtime.baseStreamFn, auth),
+		...agentStreamOptions(authWrappedStreamFn(runtime.baseStreamFn, auth)),
 		getApiKey: (providerName) =>
 			providerName === runtime.model.provider ? auth.apiKey : undefined,
 		beforeToolCall: async ({ toolCall }) =>
@@ -179,13 +198,12 @@ async function runArbitration(
 		toolExecution: "sequential",
 	});
 	try {
-		// Send head AND tail: an implementation clause at the end of a long
-		// task ("Now apply the fix") must never be truncated away, or the
-		// arbiter could rescue a run the deterministic guard correctly failed.
-		const full = task;
-		const prompt = full.length <= 8000
-			? `TASK:\n${full}`
-			: `TASK:\n${full.slice(0, 6000)}\n\n[...truncated middle...]\n\n${full.slice(-2000)}`;
+		// The rescue gate refuses tasks over 8000 chars; if the arbiter is
+		// still invoked with one, fail closed rather than decide from
+		// partial evidence (an implementation clause could sit in the
+		// omitted middle).
+		if (task.length > 8000) return "unavailable";
+		const prompt = `TASK:\n${task}`;
 		await Promise.race([
 			agent.prompt(prompt),
 			new Promise<never>((_, reject) => {
@@ -197,7 +215,7 @@ async function runArbitration(
 			}),
 		]);
 		if (!decision) return "unavailable";
-		return decision.classification === "read_only" ? "read-only" : "implementation";
+		return mapArbiterDecision(decision);
 	} catch {
 		return "unavailable";
 	}
@@ -228,13 +246,6 @@ export function isCompletionGuardFailure(result: { error?: string }): boolean {
 	return result.error?.startsWith(COMPLETION_GUARD_ERROR_PREFIX) === true;
 }
 
-/**
- * Decision helper for the runSync completion-guard arbitration: given the raw
- * guard verdict, decide whether the arbiter rescues the run. Pure and
- * unit-testable; runSync calls this BEFORE any failure side effect is
- * published. Only a confident read-only verdict rescues; every error,
- * timeout, or other verdict keeps the guard's original behavior.
- */
 export async function arbitrateCompletionGuardRescue(input: {
 	guardTriggered: boolean;
 	task: string;
@@ -243,6 +254,11 @@ export async function arbitrateCompletionGuardRescue(input: {
 	if (!input.guardTriggered || !input.arbiter) {
 		return { triggered: input.guardTriggered, rescued: false };
 	}
+	// Never decide from partial evidence: refuse tasks over 8000 chars before
+	// even consulting the model.
+	if (input.task.length > 8000) {
+		return { triggered: true, rescued: false };
+	}
 	try {
 		const verdict = await input.arbiter(input.task);
 		if (verdict === "read-only") return { triggered: false, rescued: true };

package/src/shared/agent-stream-options.ts

@@ -0,0 +1,5 @@
+import type { StreamFn } from "@earendil-works/pi-agent-core";
+
+export function agentStreamOptions(streamFn: StreamFn): { streamFunction: StreamFn; streamFn: StreamFn } {
+	return { streamFunction: streamFn, streamFn };
+}

package/src/shared/artifacts.ts

@@ -65,10 +65,6 @@ function patternMatchesArtifactPath(pattern: string, artifactPath: string): bool
 		|| globMatchesPath(normalized, artifactPath);
 }
 
-function patternMatchesProjectArtifacts(pattern: string): boolean {
-	return PROJECT_ARTIFACT_PATHS.some((artifactPath) => patternMatchesArtifactPath(pattern, artifactPath));
-}
-
 function ignoreFileExcludesProjectArtifacts(filePath: string): boolean {
 	if (!fs.existsSync(filePath)) return false;
 	try {
@@ -144,7 +140,7 @@ export function getProjectChainRunsDir(cwd: string): string {
 
 export function getChainRunsDir(
 	projectCwd: string,
-	dirPreference: ArtifactDirPreference = "project",
+	dirPreference: ArtifactDirPreference = "session",
 ): string {
 	switch (dirPreference) {
 		case "project":
@@ -160,7 +156,7 @@ export function getChainRunsDir(
 export function getArtifactsDir(
 	sessionFile: string | null,
 	projectCwd?: string,
-	dirPreference: ArtifactDirPreference = "project",
+	dirPreference: ArtifactDirPreference = "session",
 ): string {
 	switch (dirPreference) {
 		case "session":

package/src/shared/display-text.ts

@@ -98,3 +98,53 @@ export function previewDisplayText(value: string, maxLength: number): string {
 	if (maxLength <= 3) return truncateDisplayText(normalized, maxLength);
 	return `${truncateDisplayText(normalized, maxLength - 3)}...`;
 }
+
+const BINARY_CONTENT_PLACEHOLDER = "[binary content omitted for safe display]";
+
+function isUnsafeTerminalCodePoint(codePoint: number): boolean {
+	const terminalControl = (codePoint <= 0x1f && codePoint !== 0x09 && codePoint !== 0x0a)
+		|| (codePoint >= 0x7f && codePoint <= 0x9f);
+	const bidiControl = codePoint === 0x061c
+		|| codePoint === 0x200e
+		|| codePoint === 0x200f
+		|| (codePoint >= 0x202a && codePoint <= 0x202e)
+		|| (codePoint >= 0x2066 && codePoint <= 0x2069);
+	const privateUse = (codePoint >= 0xe000 && codePoint <= 0xf8ff)
+		|| (codePoint >= 0xf0000 && codePoint <= 0xffffd)
+		|| (codePoint >= 0x100000 && codePoint <= 0x10fffd);
+	const invalidScalar = codePoint >= 0xd800 && codePoint <= 0xdfff;
+	const nonCharacter = (codePoint >= 0xfdd0 && codePoint <= 0xfdef)
+		|| (codePoint & 0xffff) === 0xfffe
+		|| (codePoint & 0xffff) === 0xffff;
+	return terminalControl || bidiControl || privateUse || invalidScalar || nonCharacter;
+}
+
+function looksLikeBinaryContent(text: string): boolean {
+	if (text.includes("\0")) return true;
+	let suspiciousControls = 0;
+	let replacementCharacters = 0;
+	let codePoints = 0;
+	for (const character of text) {
+		codePoints++;
+		const codePoint = character.codePointAt(0) ?? 0;
+		if ((codePoint <= 0x08) || (codePoint >= 0x0e && codePoint <= 0x1f)) suspiciousControls++;
+		if (codePoint === 0xfffd) replacementCharacters++;
+	}
+	if (codePoints === 0) return false;
+	return (suspiciousControls >= 4 && suspiciousControls / codePoints >= 0.1)
+		|| (replacementCharacters >= 3 && replacementCharacters / codePoints >= 0.1);
+}
+
+/** Escapes untrusted text before rendering it in a terminal while preserving safe newlines and tabs. */
+export function safeTerminalText(value: string): string {
+	const normalized = value.replace(/\r\n/g, "\n");
+	if (looksLikeBinaryContent(normalized)) return BINARY_CONTENT_PLACEHOLDER;
+	let safe = "";
+	for (const character of normalized) {
+		const codePoint = character.codePointAt(0) ?? 0;
+		safe += isUnsafeTerminalCodePoint(codePoint)
+			? `[U+${codePoint.toString(16).toUpperCase().padStart(4, "0")}]`
+			: character;
+	}
+	return safe;
+}

package/src/shared/node-executable.ts

@@ -0,0 +1,21 @@
+import * as fs from "node:fs";
+import * as path from "node:path";
+
+function isNodeExecutableName(execPath: string): boolean {
+	const basename = path.basename(execPath).toLowerCase();
+	return basename === "node" || basename === "node.exe" || basename === "nodejs" || basename === "nodejs.exe";
+}
+
+function canUseNodeExecutable(execPath: string): boolean {
+	try {
+		fs.accessSync(execPath, process.platform === "win32" ? fs.constants.F_OK : fs.constants.X_OK);
+		return true;
+	} catch {
+		return false;
+	}
+}
+
+export function resolveNodeExecutable(execPath = process.execPath): string {
+	if (isNodeExecutableName(execPath) && canUseNodeExecutable(execPath)) return execPath;
+	return process.platform === "win32" ? "node.exe" : "node";
+}

package/src/shared/types.ts

@@ -528,6 +528,8 @@ export interface SteeringRecoveryDescriptor {
 	structuredOutputSchema?: JsonSchemaObject;
 	acceptance?: AcceptanceInput;
 	controlConfig?: ResolvedControlConfig;
+	/** Raw per-run bridge override. Omitted descriptors continue to use global config. */
+	intercomBridge?: IntercomBridgeConfig;
 	absoluteDeadlineAt?: number;
 	initialTurnBudget?: ResolvedTurnBudget;
 	initialToolBudget?: ResolvedToolBudget;
@@ -1086,7 +1088,7 @@ export interface Details {
 		trace: Array<{
 			operation: "run" | "status";
 			key: string;
-			state: "started" | "completed" | "failed" | "reused";
+			state: "started" | "completed" | "failed" | "stopped" | "reused";
 			agent?: string;
 			runId?: string;
 			phase?: string;
@@ -1259,6 +1261,8 @@ export interface AsyncStartedEvent {
 	asyncDir?: string;
 	/** Parent-resolved launch directory, used as a trusted artifact root while this session is live. */
 	cwd?: string;
+	/** Parent-resolved child session root, used only while this session owns the live run. */
+	sessionRoot?: string;
 	pid?: number;
 	sessionId?: string;
 	mode?: SubagentRunMode;
@@ -1462,6 +1466,8 @@ export interface AsyncJobState {
 	asyncDir: string;
 	/** Parent-resolved launch directory retained for trusted live artifact lookup. */
 	cwd?: string;
+	/** Parent-resolved child session root retained for trusted live transcript lookup. */
+	sessionRoot?: string;
 	status: "queued" | "running" | "complete" | "failed" | "paused" | "stopped" | "rejected";
 	/** Short caller-facing task/goal shown in fleet surfaces when available. */
 	description?: string;
@@ -1661,6 +1667,10 @@ export interface SubagentState {
 	/** Runtime mission-store snapshot used by optional inspector context. */
 	missionStoreConfig?: MissionStoreConfig;
 	parentSessionFile?: string | null;
+	/** Extension-owned roots trusted for child session transcript reads. */
+	trustedSessionRoots?: string[];
+	/** Live async session roots created by this parent executor, keyed by run id. */
+	liveAsyncSessionRoots?: Map<string, string>;
 	/** Last valid parent session model observed for this session; used when continuation contexts omit ctx.model. */
 	lastParentModel?: { provider: string; id: string };
 	subagentInProgress?: boolean;
@@ -1882,6 +1892,13 @@ export const FLEET_KEYBINDING_ACTIONS = [
 export type FleetKeybindingAction = typeof FLEET_KEYBINDING_ACTIONS[number];
 export type FleetKeybindingsConfig = Partial<Record<FleetKeybindingAction, string[]>>;
 
+export interface MainWindowRendererConfig {
+	/** Unit of horizontal space in main chat subagent call/result rows. Omit to preserve current spacing. Set 0 for no extra padding. */
+	horizontalSpacing?: number;
+	/** Maximum collapsed rich-result rows. Expanded output is not capped. */
+	compactResultMaxLines?: number;
+}
+
 export interface ExtensionConfig {
 	asyncByDefault?: boolean;
 	/** Show the Claude Code-style navigable fleet. Defaults to true. */
@@ -1898,6 +1915,8 @@ export interface ExtensionConfig {
 	legacyChainControls?: boolean;
 	/** Inline chat rendering for the subagent tool. Defaults to rich. */
 	inlineToolDisplay?: InlineToolDisplay;
+	/** Density controls for the main chat subagent call/result renderer. */
+	mainWindowRenderer?: MainWindowRendererConfig;
 	forceTopLevelAsync?: boolean;
 	waitTool?: WaitToolConfig;
 	defaultSessionDir?: string;
@@ -1933,7 +1952,7 @@ export interface ExtensionConfig {
 	worktreeSetupHook?: string;
 	worktreeSetupHookTimeoutMs?: number;
 	worktreeBaseDir?: string;
-	/** Where to store subagent artifact files. Defaults to "project" (cwd/.pi/subagents). Set to "session" for pi session dir, or "temp" for OS temp. */
+	/** Where to store subagent artifact files. Defaults to "session" (the pi session directory, or OS temp when unavailable). Set to "project" for cwd/.pi/subagents. */
 	artifactDir?: ArtifactDirPreference;
 	/** Artifact cleanup retention. Set cleanupDays to 0 to disable cleanup. */
 	artifactConfig?: Pick<ArtifactConfig, "cleanupDays">;
@@ -1957,7 +1976,7 @@ export const DEFAULT_MAX_OUTPUT: Required<MaxOutputConfig> = {
 
 export const DEFAULT_ARTIFACT_CONFIG: ArtifactConfig = {
 	enabled: true,
-	dir: "project",
+	dir: "session",
 	includeInput: true,
 	includeOutput: true,
 	includeJsonl: false,
@@ -2044,7 +2063,7 @@ export const SLASH_SUBAGENT_CANCEL_EVENT = "subagent:slash:cancel";
 export const POLL_INTERVAL_MS = 250;
 export const MAX_WIDGET_JOBS = 4;
 export const DEFAULT_SUBAGENT_MAX_DEPTH = 2;
-export const SUBAGENT_ACTIONS = ["list", "get", "models", "children.list", "guide", "create", "update", "delete", "eject", "disable", "enable", "reset", "mission.create", "mission.list", "mission.show", "mission.update", "mission.resolve-decision", "mission.attach-run", "mission.close", "worktree.discard", "refine", "refine.show", "refine.rollback", "inspector.open", "inspector.status", "inspector.close", "project.open", "project.status", "project.close", "status", "grant-spawn-budget", "interrupt", "resume", "steer", "stop", "dismiss", "append-step", "approve-checkpoint", "reject-checkpoint", "doctor", "watchdog.status", "watchdog.check", "watchdog.configure", "watchdog.recommend-model", "schedule.create", "schedule.list", "schedule.show", "schedule.history", "schedule.pause", "schedule.resume", "schedule.run", "schedule.run-due", "schedule.delete"] as const;
+export const SUBAGENT_ACTIONS = ["list", "get", "models", "children.list", "guide", "create", "update", "delete", "eject", "disable", "enable", "reset", "mission.create", "mission.list", "mission.show", "mission.update", "mission.resolve-decision", "mission.attach-run", "mission.close", "worktree.discard", "refine", "refine.show", "refine.rollback", "inspector.open", "inspector.status", "inspector.close", "project.open", "project.status", "project.close", "status", "debug.run", "grant-spawn-budget", "interrupt", "resume", "steer", "stop", "dismiss", "append-step", "approve-checkpoint", "reject-checkpoint", "doctor", "watchdog.status", "watchdog.check", "watchdog.configure", "watchdog.recommend-model", "schedule.create", "schedule.list", "schedule.show", "schedule.history", "schedule.pause", "schedule.resume", "schedule.run", "schedule.run-due", "schedule.delete"] as const;
 
 export const DEFAULT_FORK_PREAMBLE =
 	"You are a delegated subagent running from a fork of the parent session. " +

package/src/tui/fleet-transcript.ts

@@ -2,59 +2,12 @@ import * as fs from "node:fs";
 import * as path from "node:path";
 import { getLanguageFromPath, highlightCode, type ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { Markdown, truncateToWidth, visibleWidth, wrapTextWithAnsi, type MarkdownTheme } from "@earendil-works/pi-tui";
+import { safeTerminalText as safeDisplayText } from "../shared/display-text.ts";
 
 const DEFAULT_MAX_RECORDS = 240;
 const DEFAULT_MAX_BYTES = 2 * 1024 * 1024;
 const MAX_MESSAGE_CHARS = 64 * 1024;
 const TOOL_PREVIEW_LINES = 7;
-const BINARY_CONTENT_PLACEHOLDER = "[binary content omitted for safe display]";
-
-function isUnsafeDisplayCodePoint(codePoint: number): boolean {
-	const terminalControl = (codePoint <= 0x1f && codePoint !== 0x09 && codePoint !== 0x0a)
-		|| (codePoint >= 0x7f && codePoint <= 0x9f);
-	const bidiControl = codePoint === 0x061c
-		|| codePoint === 0x200e
-		|| codePoint === 0x200f
-		|| (codePoint >= 0x202a && codePoint <= 0x202e)
-		|| (codePoint >= 0x2066 && codePoint <= 0x2069);
-	const privateUse = (codePoint >= 0xe000 && codePoint <= 0xf8ff)
-		|| (codePoint >= 0xf0000 && codePoint <= 0xffffd)
-		|| (codePoint >= 0x100000 && codePoint <= 0x10fffd);
-	const invalidScalar = codePoint >= 0xd800 && codePoint <= 0xdfff;
-	const nonCharacter = (codePoint >= 0xfdd0 && codePoint <= 0xfdef)
-		|| (codePoint & 0xffff) === 0xfffe
-		|| (codePoint & 0xffff) === 0xffff;
-	return terminalControl || bidiControl || privateUse || invalidScalar || nonCharacter;
-}
-
-function looksLikeBinaryContent(text: string): boolean {
-	if (text.includes("\0")) return true;
-	let suspiciousControls = 0;
-	let replacementCharacters = 0;
-	let codePoints = 0;
-	for (const character of text) {
-		codePoints++;
-		const codePoint = character.codePointAt(0) ?? 0;
-		if ((codePoint <= 0x08) || (codePoint >= 0x0e && codePoint <= 0x1f)) suspiciousControls++;
-		if (codePoint === 0xfffd) replacementCharacters++;
-	}
-	if (codePoints === 0) return false;
-	return (suspiciousControls >= 4 && suspiciousControls / codePoints >= 0.1)
-		|| (replacementCharacters >= 3 && replacementCharacters / codePoints >= 0.1);
-}
-
-function safeDisplayText(text: string): string {
-	const normalized = text.replace(/\r\n/g, "\n");
-	if (looksLikeBinaryContent(normalized)) return BINARY_CONTENT_PLACEHOLDER;
-	let safe = "";
-	for (const character of normalized) {
-		const codePoint = character.codePointAt(0) ?? 0;
-		safe += isUnsafeDisplayCodePoint(codePoint)
-			? `[U+${codePoint.toString(16).toUpperCase().padStart(4, "0")}]`
-			: character;
-	}
-	return safe;
-}
 
 function sanitizeJsonDisplayValue(value: unknown): { value: unknown; changed: boolean } {
 	if (typeof value === "string") {

package/src/tui/fleet.ts

@@ -473,7 +473,7 @@ function fleetArtifactsRoot(state: SubagentState, cwd: string): string {
 	return getArtifactsDir(
 		state.parentSessionFile ?? null,
 		cwd,
-		state.artifactDirPreference ?? "project",
+		state.artifactDirPreference,
 	);
 }
 
@@ -510,6 +510,7 @@ function transcriptTarget(item: FleetItem, state: SubagentState): { path: string
 			item.run.asyncDir,
 			fleetArtifactsRoot(state, state.baseCwd),
 			trackedJob?.cwd ? fleetArtifactsRoot(state, trackedJob.cwd) : undefined,
+			item.run.sessionFile ? getArtifactsDir(item.run.sessionFile, item.run.cwd ?? state.baseCwd, state.artifactDirPreference) : undefined,
 		]),
 	};
 }
@@ -1178,6 +1179,7 @@ export async function openSubagentFleet(ctx: ExtensionContext, state: SubagentSt
 			...(input.index !== undefined ? { index: input.index } : {}),
 		}, {
 			state,
+			sessionRoots: state.trustedSessionRoots,
 			cwd: state.baseCwd,
 			...(state.authorityPolicy ? { authorityPolicy: state.authorityPolicy } : {}),
 			...(state.missionStoreConfig ? { missions: state.missionStoreConfig } : {}),