gsd-pi 2.71.0-dev.977c553 → 2.71.0-dev.d4d916a

package/src/resources/extensions/gsd/forensics.ts

@@ -650,19 +650,33 @@ function getDbCompletionCounts(): DbCompletionCounts | null {
  * Exported for testability.
  */
 export function detectStuckLoops(units: UnitMetrics[], anomalies: ForensicAnomaly[]): void {
-  // First, collect unique startedAt values per type/id key
-  const dispatchMap = new Map<string, Set<number>>();
+  // First, collect unique startedAt values per type/id key, bucketed by
+  // autoSessionKey when available so cross-session recovery does not look
+  // like a within-session stuck loop.
+  const dispatchMap = new Map<string, Map<string, Set<number>>>();
   for (const u of units) {
     const key = `${u.type}/${u.id}`;
-    let starts = dispatchMap.get(key);
+    let sessionBuckets = dispatchMap.get(key);
+    if (!sessionBuckets) {
+      sessionBuckets = new Map();
+      dispatchMap.set(key, sessionBuckets);
+    }
+
+    const sessionKey = u.autoSessionKey ?? "__legacy__";
+    let starts = sessionBuckets.get(sessionKey);
     if (!starts) {
       starts = new Set();
-      dispatchMap.set(key, starts);
+      sessionBuckets.set(sessionKey, starts);
     }
     starts.add(u.startedAt);
   }
-  for (const [key, starts] of dispatchMap) {
-    const count = starts.size;
+
+  for (const [key, sessionBuckets] of dispatchMap) {
+    const hasSessionAwareData = Array.from(sessionBuckets.keys()).some((sessionKey) => sessionKey !== "__legacy__");
+    const count = hasSessionAwareData
+      ? Math.max(...Array.from(sessionBuckets.values(), (starts) => starts.size))
+      : (sessionBuckets.get("__legacy__")?.size ?? 0);
+
     if (count > 1) {
       const [unitType, ...idParts] = key.split("/");
       anomalies.push({
@@ -671,7 +685,9 @@ export function detectStuckLoops(units: UnitMetrics[], anomalies: ForensicAnomal
         unitType,
         unitId: idParts.join("/"),
         summary: `Unit ${key} was dispatched ${count} times`,
-        details: `Repeated dispatch suggests the unit completed but its artifacts weren't verified, or the state machine kept returning it.`,
+        details: hasSessionAwareData
+          ? `Repeated dispatch within the same auto session suggests the unit completed but its artifacts were not verified, or the state machine kept returning it. Cross-session recovery runs are ignored.`
+          : `Repeated dispatch suggests the unit completed but its artifacts weren't verified, or the state machine kept returning it.`,
       });
     }
   }

package/src/resources/extensions/gsd/guided-flow.ts

@@ -15,7 +15,7 @@ import { loadPrompt, inlineTemplate } from "./prompt-loader.js";
 import { buildSkillActivationBlock } from "./auto-prompts.js";
 import { deriveState } from "./state.js";
 import { invalidateAllCaches } from "./cache.js";
-import { startAuto } from "./auto.js";
+import { startAutoDetached } from "./auto.js";
 import { clearLock } from "./crash-recovery.js";
 import {
   assessInterruptedSession,
@@ -67,7 +67,6 @@ export {
   showQueue, handleQueueReorder, showQueueAdd,
   buildExistingMilestonesContext,
 } from "./guided-flow-queue.js";
-import { getErrorMessage } from "./error-utils.js";
 import { logWarning } from "./workflow-logger.js";
 
 // ─── ID Generation with Reservation ─────────────────────────────────────────
@@ -244,11 +243,7 @@ export function checkAutoStartAfterDiscuss(): boolean {
 
   pendingAutoStartMap.delete(basePath);
   ctx.ui.notify(`Milestone ${milestoneId} ready.`, "info");
-  startAuto(ctx, pi, basePath, false, { step }).catch((err) => {
-    ctx.ui.notify(`Auto-start failed: ${getErrorMessage(err)}`, "error");
-    logWarning("guided", `auto start error: ${getErrorMessage(err)}`);
-    debugLog("auto-start-failed", { error: getErrorMessage(err) });
-  });
+  startAutoDetached(ctx, pi, basePath, false, { step });
   return true;
 }
 
@@ -1305,7 +1300,7 @@ export async function showSmartEntry(
       ],
     });
     if (resume === "resume") {
-      await startAuto(ctx, pi, basePath, false, {
+      startAutoDetached(ctx, pi, basePath, false, {
         interrupted,
         step: interrupted.pausedSession?.stepMode ?? false,
       });
@@ -1647,7 +1642,7 @@ export async function showSmartEntry(
       });
 
       if (choice === "auto") {
-        await startAuto(ctx, pi, basePath, false);
+        startAutoDetached(ctx, pi, basePath, false);
       } else if (choice === "status") {
         const { fireStatusViaCommand } = await import("./commands.js");
         await fireStatusViaCommand(ctx);
@@ -1859,7 +1854,7 @@ export async function showSmartEntry(
     });
 
     if (choice === "auto") {
-      await startAuto(ctx, pi, basePath, false);
+      startAutoDetached(ctx, pi, basePath, false);
       return;
     }
 

package/src/resources/extensions/gsd/interrupted-session.ts

@@ -34,6 +34,7 @@ export interface PausedSessionMetadata {
   activeEngineId?: string;
   activeRunDir?: string | null;
   autoStartTime?: number;
+  milestoneLock?: string | null;
 }
 
 export interface InterruptedSessionAssessment {

package/src/resources/extensions/gsd/metrics.ts

@@ -41,6 +41,7 @@ export interface UnitMetrics {
   model: string;           // model ID used
   startedAt: number;       // ms timestamp
   finishedAt: number;      // ms timestamp
+  autoSessionKey?: string; // identifies one auto-mode run across pause/resume
   tokens: TokenCounts;
   cost: number;            // total USD cost
   toolCalls: number;
@@ -133,7 +134,16 @@ export function snapshotUnitMetrics(
   unitId: string,
   startedAt: number,
   model: string,
-  opts?: { tier?: string; modelDowngraded?: boolean; contextWindowTokens?: number; truncationSections?: number; continueHereFired?: boolean; promptCharCount?: number; baselineCharCount?: number },
+  opts?: {
+    tier?: string;
+    modelDowngraded?: boolean;
+    contextWindowTokens?: number;
+    truncationSections?: number;
+    continueHereFired?: boolean;
+    promptCharCount?: number;
+    baselineCharCount?: number;
+    autoSessionKey?: string;
+  },
 ): UnitMetrics | null {
   if (!ledger) return null;
 
@@ -181,6 +191,7 @@ export function snapshotUnitMetrics(
     model,
     startedAt,
     finishedAt: Date.now(),
+    ...(opts?.autoSessionKey ? { autoSessionKey: opts.autoSessionKey } : {}),
     tokens,
     cost,
     toolCalls,

package/src/resources/extensions/gsd/milestone-actions.ts

@@ -20,7 +20,7 @@ import {
 } from "./paths.js";
 import { invalidateAllCaches } from "./cache.js";
 import { loadQueueOrder, saveQueueOrder } from "./queue-order.js";
-import { isDbAvailable, updateMilestoneStatus } from "./gsd-db.js";
+import { getMilestone, isDbAvailable, updateMilestoneStatus } from "./gsd-db.js";
 import { logWarning } from "./workflow-logger.js";
 
 // ─── Park ──────────────────────────────────────────────────────────────────
@@ -77,9 +77,16 @@ export function unparkMilestone(basePath: string, milestoneId: string): boolean
   if (!mDir || !existsSync(mDir)) return false;
 
   const parkedPath = join(mDir, buildMilestoneFileName(milestoneId, "PARKED"));
-  if (!existsSync(parkedPath)) return false; // not parked
+  const hadParkedFile = existsSync(parkedPath);
+  const dbThinksParked = isDbAvailable() && getMilestone(milestoneId)?.status === "parked";
 
-  unlinkSync(parkedPath);
+  // Recover the reverse desync too: DB can still say "parked" even when the
+  // PARKED marker was lost on disk, and /gsd unpark should repair that state.
+  if (!hadParkedFile && !dbThinksParked) return false;
+
+  if (hadParkedFile) {
+    unlinkSync(parkedPath);
+  }
   // Sync DB status so deriveStateFromDb picks up the unparked milestone (#2694)
   if (isDbAvailable()) {
     try {

package/src/resources/extensions/gsd/notification-overlay.ts

@@ -1,6 +1,6 @@
 // GSD Extension — Notification History Overlay
 // Scrollable panel showing all persisted notifications with severity filtering.
-// Toggled with Ctrl+Alt+N (⌃⌥N on macOS) or opened from /gsd notifications.
+// Toggled with Ctrl+Alt+N (⌃⌥N on macOS), Ctrl+Shift+N fallback, or /gsd notifications.
 
 import type { Theme } from "@gsd/pi-coding-agent";
 import { truncateToWidth, visibleWidth, matchesKey, Key } from "@gsd/pi-tui";
@@ -9,11 +9,12 @@ import {
   readNotifications,
   markAllRead,
   clearNotifications,
-  getUnreadCount,
+  onNotificationStoreChange,
   type NotificationEntry,
   type NotifySeverity,
 } from "./notification-store.js";
-import { padRight, centerLine, joinColumns, formatDuration } from "../shared/mod.js";
+import { formattedShortcutPair } from "./shortcut-defs.js";
+import { padRight, joinColumns } from "../shared/mod.js";
 
 type FilterMode = "all" | "error" | "warning" | "info";
 const FILTER_CYCLE: FilterMode[] = ["all", "error", "warning", "info"];
@@ -63,6 +64,12 @@ function formatTimestamp(ts: string): string {
   }
 }
 
+function notificationSignature(entries: readonly NotificationEntry[]): string {
+  return entries
+    .map((entry) => `${entry.ts}|${entry.severity}|${entry.read ? 1 : 0}|${entry.message}`)
+    .join("\n");
+}
+
 export class GSDNotificationOverlay {
   private tui: { requestRender: () => void };
   private theme: Theme;
@@ -72,9 +79,11 @@ export class GSDNotificationOverlay {
   private scrollOffset = 0;
   private filterIndex = 0;
   private entries: NotificationEntry[] = [];
+  private entriesSignature = "";
   private refreshTimer: ReturnType<typeof setInterval>;
   private disposed = false;
   private resizeHandler: (() => void) | null = null;
+  private unsubscribeStore: (() => void) | null = null;
 
   constructor(
     tui: { requestRender: () => void },
@@ -88,6 +97,7 @@ export class GSDNotificationOverlay {
     // Mark all as read on open
     markAllRead();
     this.entries = readNotifications();
+    this.entriesSignature = notificationSignature(this.entries);
 
     // Resize handler
     this.resizeHandler = () => {
@@ -97,17 +107,17 @@ export class GSDNotificationOverlay {
     };
     process.stdout.on("resize", this.resizeHandler);
 
-    // Refresh every 3s for new notifications
+    // Subscribe to store mutations for immediate updates
+    this.unsubscribeStore = onNotificationStoreChange(() => {
+      if (this.disposed) return;
+      this._refreshFromDisk();
+    });
+
+    // 30s safety-net for cross-process edits (web subprocess, parallel workers)
     this.refreshTimer = setInterval(() => {
       if (this.disposed) return;
-      const fresh = readNotifications();
-      if (fresh.length !== this.entries.length) {
-        this.entries = fresh;
-        markAllRead();
-        this.invalidate();
-        this.tui.requestRender();
-      }
-    }, 3000);
+      this._refreshFromDisk();
+    }, 30_000);
   }
 
   private get filter(): FilterMode {
@@ -120,7 +130,12 @@ export class GSDNotificationOverlay {
   }
 
   handleInput(data: string): void {
-    if (matchesKey(data, Key.escape) || matchesKey(data, Key.ctrl("c")) || matchesKey(data, Key.ctrlAlt("n"))) {
+    if (
+      matchesKey(data, Key.escape) ||
+      matchesKey(data, Key.ctrl("c")) ||
+      matchesKey(data, Key.ctrlAlt("n")) ||
+      matchesKey(data, Key.ctrlShift("n"))
+    ) {
       this.dispose();
       this.onClose();
       return;
@@ -165,6 +180,7 @@ export class GSDNotificationOverlay {
     if (data === "c") {
       clearNotifications();
       this.entries = [];
+      this.entriesSignature = notificationSignature(this.entries);
       this.scrollOffset = 0;
       this.invalidate();
       this.tui.requestRender();
@@ -199,12 +215,28 @@ export class GSDNotificationOverlay {
   dispose(): void {
     this.disposed = true;
     clearInterval(this.refreshTimer);
+    if (this.unsubscribeStore) {
+      this.unsubscribeStore();
+      this.unsubscribeStore = null;
+    }
     if (this.resizeHandler) {
       process.stdout.removeListener("resize", this.resizeHandler);
       this.resizeHandler = null;
     }
   }
 
+  private _refreshFromDisk(): void {
+    const fresh = readNotifications();
+    const signature = notificationSignature(fresh);
+    if (signature !== this.entriesSignature) {
+      markAllRead();
+      this.entries = readNotifications();
+      this.entriesSignature = notificationSignature(this.entries);
+      this.invalidate();
+      this.tui.requestRender();
+    }
+  }
+
   private wrapInBox(inner: string[], width: number): string[] {
     const th = this.theme;
     const border = (s: string) => th.fg("borderAccent", s);
@@ -250,7 +282,8 @@ export class GSDNotificationOverlay {
     lines.push(hr());
 
     // Controls
-    lines.push(row(th.fg("dim", "↑/↓ scroll  f filter  c clear  Esc close")));
+    const closeShortcut = formattedShortcutPair("notifications");
+    lines.push(row(th.fg("dim", `↑/↓ scroll  f filter  c clear  Esc close  (${closeShortcut})`)));
     lines.push(blank());
 
     // Entries

package/src/resources/extensions/gsd/notification-store.ts

@@ -26,12 +26,16 @@ export interface NotificationEntry {
 const MAX_ENTRIES = 500;
 const FILENAME = "notifications.jsonl";
 const LOCKFILE = "notifications.lock";
+const DEDUP_WINDOW_MS = 30_000;
+const DEDUP_PRUNE_THRESHOLD = 200;
 
 // ─── Module State ───────────────────────────────────────────────────────
 
 let _basePath: string | null = null;
 let _lineCount = 0;  // Hint for rotation — not authoritative for public API
 let _suppressCount = 0;
+let _recentMessageTimestamps = new Map<string, number>();
+const _changeListeners = new Set<() => void>();
 
 // ─── Public API ─────────────────────────────────────────────────────────
 
@@ -40,6 +44,9 @@ let _suppressCount = 0;
  * project root. Seeds in-memory counters from the existing file on disk.
  */
 export function initNotificationStore(basePath: string): void {
+  if (_basePath !== basePath) {
+    _recentMessageTimestamps.clear();
+  }
   _basePath = basePath;
   // Seed line count hint for rotation — public counters read from disk
   _lineCount = _readEntriesFromDisk(basePath).length;
@@ -56,12 +63,23 @@ export function appendNotification(
 ): void {
   if (!_basePath) return;
   if (_suppressCount > 0) return;
+  const persistedMessage = message.length > 500 ? message.slice(0, 500) + "…" : message;
+  const dedupKey = `${_basePath}:${severity}:${source}:${persistedMessage}`;
+  const now = Date.now();
+  const lastSeen = _recentMessageTimestamps.get(dedupKey);
+  if (lastSeen !== undefined && now - lastSeen < DEDUP_WINDOW_MS) return;
+  _recentMessageTimestamps.set(dedupKey, now);
+  if (_recentMessageTimestamps.size > DEDUP_PRUNE_THRESHOLD) {
+    for (const [key, ts] of _recentMessageTimestamps) {
+      if (now - ts > DEDUP_WINDOW_MS) _recentMessageTimestamps.delete(key);
+    }
+  }
 
   const entry: NotificationEntry = {
     id: randomUUID(),
     ts: new Date().toISOString(),
     severity,
-    message: message.length > 500 ? message.slice(0, 500) + "…" : message,
+    message: persistedMessage,
     source,
     read: false,
   };
@@ -76,6 +94,7 @@ export function appendNotification(
     if (_lineCount > MAX_ENTRIES) {
       _rotate();
     }
+    _emitChange();
   } catch {
     // Non-fatal — never let persistence break the caller
   }
@@ -104,6 +123,7 @@ export function markAllRead(basePath?: string): void {
   const hasUnread = entries.some((e) => !e.read);
   if (!hasUnread) return;
 
+  let changed = false;
   try {
     _withLock(bp, () => {
       // Re-read inside lock to get freshest state
@@ -111,10 +131,12 @@ export function markAllRead(basePath?: string): void {
       if (fresh.length === 0 || !fresh.some((e) => !e.read)) return;
       const lines = fresh.map((e) => JSON.stringify({ ...e, read: true }));
       _atomicWrite(bp, lines.join("\n") + "\n");
+      changed = true;
     });
   } catch {
     // Non-fatal
   }
+  if (changed) _emitChange();
 }
 
 /**
@@ -128,6 +150,8 @@ export function clearNotifications(basePath?: string): void {
     _withLock(bp, () => {
       _atomicWrite(bp, "");
     });
+    _lineCount = 0;
+    _emitChange();
   } catch {
     // Non-fatal
   }
@@ -172,6 +196,17 @@ export function unsuppressPersistence(): void {
   _suppressCount = Math.max(0, _suppressCount - 1);
 }
 
+/**
+ * Subscribe to notification-store mutations (append, mark-read, clear).
+ * Returns an unsubscribe function.
+ */
+export function onNotificationStoreChange(listener: () => void): () => void {
+  _changeListeners.add(listener);
+  return () => {
+    _changeListeners.delete(listener);
+  };
+}
+
 // ─── Test Helpers ───────────────────────────────────────────────────────
 
 /**
@@ -181,6 +216,8 @@ export function _resetNotificationStore(): void {
   _basePath = null;
   _lineCount = 0;
   _suppressCount = 0;
+  _recentMessageTimestamps = new Map();
+  _changeListeners.clear();
 }
 
 // ─── Internal ───────────────────────────────────────────────────────────
@@ -216,12 +253,23 @@ function _rotate(): void {
       const trimmed = entries.slice(entries.length - MAX_ENTRIES);
       const lines = trimmed.map((e) => JSON.stringify(e));
       _atomicWrite(_basePath!, lines.join("\n") + "\n");
+      _lineCount = trimmed.length;
     });
   } catch {
     // Non-fatal
   }
 }
 
+function _emitChange(): void {
+  for (const listener of _changeListeners) {
+    try {
+      listener();
+    } catch {
+      // Non-fatal
+    }
+  }
+}
+
 /**
  * Atomic file rewrite via temp-file + rename. Prevents partial reads
  * by other processes (web API subprocess, parallel workers).
@@ -275,10 +323,11 @@ function _withLock<T>(basePath: string, fn: () => T): T {
     }
   }
 
-  // Only run the mutation if we actually own the lock
-  const ownsLock = fd !== null;
+  // Best-effort: mutation runs regardless of lock status (idempotent overwrites).
+  // createdLock gates cleanup only — never skip fn() on lock failure.
+  const createdLock = fd !== null;
   try {
-    if (ownsLock && fd !== null) {
+    if (createdLock && fd !== null) {
       // Write our PID timestamp into the lock for stale detection
       writeFileSync(lockPath, String(Date.now()), "utf-8");
       closeSync(fd);
@@ -286,7 +335,7 @@ function _withLock<T>(basePath: string, fn: () => T): T {
     return fn();
   } finally {
     // Only delete the lock if we created it — never remove another process's lock
-    if (ownsLock) {
+    if (createdLock) {
       try { unlinkSync(lockPath); } catch { /* best-effort cleanup */ }
     }
   }

package/src/resources/extensions/gsd/notification-widget.ts

@@ -5,8 +5,8 @@
 
 import type { ExtensionContext } from "@gsd/pi-coding-agent";
 
-import { getUnreadCount, readNotifications } from "./notification-store.js";
-import { formatShortcut } from "./files.js";
+import { getUnreadCount, onNotificationStoreChange } from "./notification-store.js";
+import { formattedShortcutPair } from "./shortcut-defs.js";
 
 // ─── Pure rendering ──���────────────────────────���─────────────────────────
 
@@ -14,18 +14,7 @@ export function buildNotificationWidgetLines(): string[] {
   const unread = getUnreadCount();
   if (unread === 0) return [];
 
-  const entries = readNotifications();
-  const latest = entries[0]; // newest-first
-  if (!latest) return [];
-
-  const icon = latest.severity === "error" ? "✗" : latest.severity === "warning" ? "⚠" : "●";
-  const badge = `${unread} unread`;
-  const msgMax = 80;
-  const truncated = latest.message.length > msgMax
-    ? latest.message.slice(0, msgMax - 1) + "…"
-    : latest.message;
-
-  return [`  ${icon} [${badge}]  ${truncated}  (${formatShortcut("Ctrl+Alt+N")} to view)`];
+  return [`  🔔 Notifications: ${unread} unread  (${formattedShortcutPair("notifications")})`];
 }
 
 // ─── Widget init ────────────────────────────────────────────────────────
@@ -51,6 +40,7 @@ export function initNotificationWidget(ctx: ExtensionContext): void {
       _tui.requestRender();
     };
 
+    const unsubscribe = onNotificationStoreChange(refresh);
     const refreshTimer = setInterval(refresh, REFRESH_INTERVAL_MS);
 
     return {
@@ -62,6 +52,7 @@ export function initNotificationWidget(ctx: ExtensionContext): void {
         cachedLines = undefined;
       },
       dispose(): void {
+        unsubscribe();
         clearInterval(refreshTimer);
       },
     };

package/src/resources/extensions/gsd/parallel-monitor-overlay.ts

@@ -2,7 +2,8 @@
  * GSD Parallel Monitor Overlay
  *
  * Full-screen TUI overlay showing real-time parallel worker progress.
- * Opened via `/gsd parallel watch` or Ctrl+Alt+P (⌃⌥P on macOS).
+ * Opened via `/gsd parallel watch`, Ctrl+Alt+P (⌃⌥P on macOS),
+ * or Ctrl+Shift+P fallback.
  * Reads the same data sources as `scripts/parallel-monitor.mjs` but
  * renders as a native pi-tui overlay with theme integration.
  */
@@ -15,6 +16,7 @@ import type { Theme } from "@gsd/pi-coding-agent";
 import { truncateToWidth, visibleWidth, matchesKey, Key } from "@gsd/pi-tui";
 
 import { formatDuration, STATUS_GLYPH, STATUS_COLOR } from "../shared/mod.js";
+import { formattedShortcutPair } from "./shortcut-defs.js";
 
 // ─── Types ────────────────────────────────────────────────────────────────
 
@@ -347,7 +349,12 @@ export class ParallelMonitorOverlay {
   }
 
   handleInput(data: string): void {
-    if (matchesKey(data, Key.escape) || data === "q") {
+    if (
+      matchesKey(data, Key.escape) ||
+      matchesKey(data, Key.ctrlAlt("p")) ||
+      matchesKey(data, Key.ctrlShift("p")) ||
+      data === "q"
+    ) {
       this.dispose();
       this.onClose();
       return;
@@ -486,7 +493,7 @@ export class ParallelMonitorOverlay {
       }
       lines.push(`  ${t.bold("Total: $" + this.workers.reduce((s, wk) => s + wk.cost, 0).toFixed(2))}`);
     }
-    lines.push(t.fg("muted", "  ESC/q to close  │  ↑↓ scroll"));
+    lines.push(t.fg("muted", `  ESC/q/${formattedShortcutPair("parallel")} close  │  ↑↓ scroll`));
 
     // Apply scroll — use terminal rows as height estimate
     const termHeight = process.stdout.rows || 40;

package/src/resources/extensions/gsd/pre-execution-checks.ts

@@ -280,6 +280,38 @@ function extractPathFromAnnotation(raw: string): string {
   return trimmed.replace(/`/g, "");
 }
 
+/**
+ * Planning units sometimes use task.inputs for prose like "Current enum shape"
+ * instead of concrete file paths. Those entries should not fail path checks.
+ * Keep validation for anything that still looks like a real file reference:
+ * explicit backticks, globs, separators, dot-paths, or single-token basenames
+ * like Dockerfile.
+ */
+function shouldValidateInputAsPath(raw: string): boolean {
+  const trimmed = raw.trim();
+  if (!trimmed) return false;
+
+  if (/^`+[^`]+`+/.test(trimmed)) {
+    return true;
+  }
+
+  const candidate = extractPathFromAnnotation(trimmed);
+  if (!candidate) return false;
+
+  if (!/\s/.test(candidate)) {
+    return true;
+  }
+
+  return (
+    candidate.startsWith("/") ||
+    candidate.startsWith("./") ||
+    candidate.startsWith("../") ||
+    candidate.startsWith("~/") ||
+    /[\\/]/.test(candidate) ||
+    /[*?[\]{}]/.test(candidate)
+  );
+}
+
 /**
  * Build a set of files that will be created by tasks up to (but not including) taskIndex.
  * All paths are normalized for consistent comparison.
@@ -318,6 +350,7 @@ export function checkFilePathConsistency(
     for (const file of filesToCheck) {
       // Skip empty strings
       if (!file.trim()) continue;
+      if (!shouldValidateInputAsPath(file)) continue;
 
       // Normalize path for consistent comparison
       const normalizedFile = normalizeFilePath(file);
@@ -354,7 +387,7 @@ export function checkFilePathConsistency(
  */
 export function checkTaskOrdering(
   tasks: TaskRow[],
-  _basePath: string
+  basePath: string
 ): PreExecutionCheckJSON[] {
   const results: PreExecutionCheckJSON[] = [];
 
@@ -378,9 +411,13 @@ export function checkTaskOrdering(
     const filesToCheck = [...task.inputs];
 
     for (const file of filesToCheck) {
+      if (!shouldValidateInputAsPath(file)) continue;
+
       const normalizedFile = normalizeFilePath(file);
       const creator = fileCreators.get(normalizedFile);
-      if (creator && creator.index > i) {
+      const absolutePath = resolve(basePath, normalizedFile);
+      const existsOnDisk = existsSync(absolutePath);
+      if (creator && creator.index > i && !existsOnDisk) {
         // Task reads file that is created later — impossible ordering
         results.push({
           category: "file",

package/src/resources/extensions/gsd/prompts/complete-slice.md

@@ -21,7 +21,7 @@ All relevant context has been preloaded below — the slice plan, all task summa
 Then:
 1. Use the **Slice Summary** and **UAT** output templates from the inlined context above
 2. {{skillActivation}}
-3. Run all slice-level verification checks defined in the slice plan. All must pass before marking the slice done. If any fail, fix them first.
+3. Run all slice-level verification checks defined in the slice plan. All must pass before marking the slice done. If any fail, fix them first. Task artifacts use a **flat file layout** directly inside `tasks/` (for example `T01-SUMMARY.md`, `T02-SUMMARY.md`) rather than per-task subdirectories. If you need to count or re-read task summaries during verification, use `find .gsd/milestones/{{milestoneId}}/slices/{{sliceId}}/tasks -name "*-SUMMARY.md"` or `ls .gsd/milestones/{{milestoneId}}/slices/{{sliceId}}/tasks/*-SUMMARY.md`. Never use `tasks/*/SUMMARY.md` — that glob expects subdirectories that do not exist.
 4. If the slice plan includes observability/diagnostic surfaces, confirm they work. Skip this for simple slices that don't have observability sections.
 5. If the slice involved runtime behavior, fill the **Operational Readiness** section (Q8) in the slice summary: health signal, failure signal, recovery procedure, and monitoring gaps. Omit entirely for simple slices with no runtime concerns.
 6. If this slice produced evidence that a requirement changed status (Active → Validated, Active → Deferred, etc.), call `gsd_requirement_update` with the requirement ID, updated `status`, and `validation` evidence. Do NOT write `.gsd/REQUIREMENTS.md` directly — the engine renders it from the database.
@@ -35,7 +35,7 @@ Then:
 
 **Autonomous execution:** Do not call `ask_user_questions` or `secure_env_collect`. You are running in auto-mode — there is no human available to answer questions. Make reasonable assumptions and document them in the slice summary. If a decision genuinely requires human input, note it in the summary and proceed with the best available option.
 
-**File system safety:** Task summaries are preloaded in the inlined context above. If you need to re-read any of them, use `find .gsd/milestones/{{milestoneId}}/slices/{{sliceId}}/tasks -name "*-SUMMARY.md"` to list file paths first — never pass `{{slicePath}}` or any other directory path directly to the `read` tool. The `read` tool only accepts file paths, not directories.
+**File system safety:** Task summaries are preloaded in the inlined context above. Task artifacts use a **flat file layout** — files such as `T01-SUMMARY.md` and `T02-SUMMARY.md` live directly inside the `tasks/` directory, not inside per-task subdirectories like `tasks/T01/SUMMARY.md`. If you need to re-read any of them, use `find .gsd/milestones/{{milestoneId}}/slices/{{sliceId}}/tasks -name "*-SUMMARY.md"` to list file paths first. Never use `tasks/*/SUMMARY.md`, and never pass `{{slicePath}}` or any other directory path directly to the `read` tool. The `read` tool only accepts file paths, not directories.
 
 **You MUST call `gsd_complete_slice` with the slice summary and UAT content before finishing. The tool persists to both DB and disk and renders `{{sliceSummaryPath}}` and `{{sliceUatPath}}` automatically.**
 

package/src/resources/extensions/gsd/prompts/discuss.md

@@ -73,6 +73,8 @@ After each round of answers, decide whether you already have enough depth to wri
 
 You are a thinking partner, not an interviewer.
 
+**Turn-taking contract (non-bypassable).** Never fabricate, simulate, or role-play user responses. Never generate fake transcript markers like `[User]`, `[Human]`, or `User:` to invent input. Ask one question round (1-3 questions) per turn, then stop and wait for the user's actual response before continuing. If you use `ask_user_questions`, call it at most once per turn and treat its returned response as the only valid structured user input for that round.
+
 **Start open, follow energy.** Let the user's enthusiasm guide where you dig deeper. If they light up about a particular aspect, explore it. If they're vague about something, that's where you probe.
 
 **Challenge vagueness, make abstract concrete.** When the user says something abstract ("it should be smart" / "it needs to handle edge cases" / "good UX"), push for specifics. What does "smart" mean in practice? Which edge cases? What does good UX look like for this specific interaction?