gsd-pi 2.78.1-dev.b0759e59b → 2.78.1-dev.d8826a445

package/dist/resources/extensions/gsd/metrics.js

@@ -1,3 +1,4 @@
+// GSD-2 + metrics.ts: token & cost tracking for auto-mode units
 /**
  * GSD Metrics — Token & Cost Tracking
  *
@@ -13,12 +14,14 @@
  *   4. On crash recovery or fresh start, the ledger is loaded from disk
  */
 import { join } from "node:path";
+import { openSync, closeSync, unlinkSync, statSync, writeFileSync } from "node:fs";
 import { gsdRoot } from "./paths.js";
 import { getAndClearSkills } from "./skill-telemetry.js";
 import { loadJsonFile, loadJsonFileOrNull, saveJsonFile } from "./json-persistence.js";
 import { parseUnitId } from "./unit-id.js";
 import { buildAuditEnvelope, emitUokAuditEvent } from "./uok/audit.js";
 import { isUnifiedAuditEnabled } from "./uok/audit-toggle.js";
+import { logWarning } from "./workflow-logger.js";
 // Re-export from shared — import directly from format-utils to avoid pulling
 // in the full barrel (mod.js → ui.js → @gsd/pi-tui) which breaks when loaded
 // outside jiti's alias resolution (e.g. dynamic import in auto-loop reports).
@@ -48,10 +51,15 @@ export function classifyUnitPhase(unitType) {
 // ─── In-memory state ──────────────────────────────────────────────────────────
 let ledger = null;
 let basePath = "";
+// Per-workspace ledger map, keyed by workspace.identityKey.
+// Populated by initMetricsByScope; independent of the module singleton.
+const scopedLedgers = new Map();
 // ─── Public API ───────────────────────────────────────────────────────────────
 /**
  * Initialize the metrics system for a given project.
  * Loads existing ledger from disk if present.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use initMetricsByScope instead.
  */
 export function initMetrics(base) {
     basePath = base;
@@ -59,6 +67,8 @@ export function initMetrics(base) {
 }
 /**
  * Reset in-memory state. Called when auto-mode stops.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use resetMetricsByScope instead.
  */
 export function resetMetrics() {
     ledger = null;
@@ -67,6 +77,8 @@ export function resetMetrics() {
 /**
  * Snapshot usage metrics from the current session before it's wiped.
  * Scans session entries for AssistantMessage usage data.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use snapshotUnitMetricsByScope instead.
  */
 export function snapshotUnitMetrics(ctx, unitType, unitId, startedAt, model, opts) {
     if (!ledger)
@@ -180,6 +192,147 @@ export function snapshotUnitMetrics(ctx, unitType, unitId, startedAt, model, opt
 export function getLedger() {
     return ledger;
 }
+// ─── Scope-aware API (canonical) ─────────────────────────────────────────────
+/**
+ * Initialize the metrics system for a given workspace scope.
+ * Loads existing ledger from disk into the per-scope ledger map.
+ * Does NOT touch the module-level singleton.
+ */
+export function initMetricsByScope(scope) {
+    const base = scope.workspace.projectRoot;
+    const loaded = loadLedger(base);
+    scopedLedgers.set(scope.workspace.identityKey, loaded);
+}
+/**
+ * Get the in-memory ledger for the given scope, or null if not initialized.
+ */
+export function getLedgerByScope(scope) {
+    return scopedLedgers.get(scope.workspace.identityKey) ?? null;
+}
+/**
+ * Reset scoped in-memory state for a workspace. Called when auto-mode stops.
+ */
+export function resetMetricsByScope(scope) {
+    scopedLedgers.delete(scope.workspace.identityKey);
+}
+/**
+ * Snapshot usage metrics using an explicit workspace scope.
+ *
+ * This is the canonical variant. It derives the metrics path from
+ * scope.workspace.projectRoot rather than the module singleton, so it
+ * remains correct across session resume and in multi-workspace processes.
+ *
+ * Preserves the atomic write-merge logic from saveLedger so concurrent
+ * workers cannot silently discard each other's entries.
+ *
+ * If initMetricsByScope has not been called, the ledger is loaded from
+ * disk on first call (lazy init).
+ */
+export function snapshotUnitMetricsByScope(scope, ctx, unitType, unitId, startedAt, model, opts) {
+    const base = scope.workspace.projectRoot;
+    const key = scope.workspace.identityKey;
+    // Lazy init: load from disk if not yet in scoped map.
+    if (!scopedLedgers.has(key)) {
+        scopedLedgers.set(key, loadLedger(base));
+    }
+    const scopedLedger = scopedLedgers.get(key);
+    const entries = ctx.sessionManager.getEntries();
+    if (!entries || entries.length === 0)
+        return null;
+    const tokens = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 };
+    let cost = 0;
+    let toolCalls = 0;
+    let assistantMessages = 0;
+    let userMessages = 0;
+    for (const entry of entries) {
+        if (entry.type !== "message")
+            continue;
+        const msg = entry.message;
+        if (!msg)
+            continue;
+        if (msg.role === "assistant") {
+            assistantMessages++;
+            if (msg.usage) {
+                tokens.input += msg.usage.input ?? 0;
+                tokens.output += msg.usage.output ?? 0;
+                tokens.cacheRead += msg.usage.cacheRead ?? 0;
+                tokens.cacheWrite += msg.usage.cacheWrite ?? 0;
+                tokens.total += msg.usage.totalTokens ?? 0;
+                if (msg.usage.cost != null) {
+                    const c = msg.usage.cost;
+                    cost += typeof c === "number" ? c : (c.total ?? 0);
+                }
+            }
+            if (msg.content && Array.isArray(msg.content)) {
+                for (const block of msg.content) {
+                    if (block.type === "toolCall")
+                        toolCalls++;
+                }
+            }
+        }
+        else if (msg.role === "user") {
+            userMessages++;
+        }
+    }
+    const unit = {
+        type: unitType,
+        id: unitId,
+        model,
+        startedAt,
+        finishedAt: Date.now(),
+        ...(opts?.autoSessionKey ? { autoSessionKey: opts.autoSessionKey } : {}),
+        tokens,
+        cost,
+        toolCalls,
+        assistantMessages,
+        userMessages,
+        apiRequests: assistantMessages,
+        ...(opts?.tier ? { tier: opts.tier } : {}),
+        ...(opts?.modelDowngraded !== undefined ? { modelDowngraded: opts.modelDowngraded } : {}),
+        ...(opts?.contextWindowTokens !== undefined ? { contextWindowTokens: opts.contextWindowTokens } : {}),
+        ...(opts?.truncationSections !== undefined ? { truncationSections: opts.truncationSections } : {}),
+        ...(opts?.continueHereFired !== undefined ? { continueHereFired: opts.continueHereFired } : {}),
+        ...(opts?.promptCharCount != null ? { promptCharCount: opts.promptCharCount } : {}),
+        ...(opts?.baselineCharCount != null ? { baselineCharCount: opts.baselineCharCount } : {}),
+    };
+    // Auto-capture skill telemetry (#599)
+    const skills = getAndClearSkills();
+    if (skills.length > 0) {
+        unit.skills = skills;
+    }
+    // Compute cache hit rate
+    if (tokens.cacheRead > 0 || tokens.input > 0) {
+        const totalInput = tokens.cacheRead + tokens.input;
+        unit.cacheHitRate = totalInput > 0 ? Math.round((tokens.cacheRead / totalInput) * 100) : 0;
+    }
+    // Idempotency guard: update in-place on duplicate, append otherwise.
+    const dupeIdx = scopedLedger.units.findIndex((u) => u.type === unit.type && u.id === unit.id && u.startedAt === unit.startedAt);
+    if (dupeIdx >= 0) {
+        scopedLedger.units[dupeIdx] = unit;
+    }
+    else {
+        scopedLedger.units.push(unit);
+    }
+    saveLedger(base, scopedLedger);
+    if (isUnifiedAuditEnabled()) {
+        emitUokAuditEvent(base, buildAuditEnvelope({
+            traceId: opts?.traceId ?? `metrics:${unitType}:${unitId}`,
+            turnId: opts?.turnId,
+            causedBy: opts?.causedBy,
+            category: "metrics",
+            type: "unit-metrics-snapshot",
+            payload: {
+                unitType,
+                unitId,
+                model,
+                tokens: unit.tokens,
+                cost: unit.cost,
+                toolCalls: unit.toolCalls,
+            },
+        }));
+    }
+    return unit;
+}
 function emptyTokens() {
     return { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 };
 }
@@ -423,6 +576,12 @@ export function pruneMetricsLedger(base, keepCount) {
     if (ledger) {
         ledger.units = ledger.units.slice(-keepCount);
     }
+    // Invalidate all scoped ledger cache entries. Prune is rare; clearing the
+    // entire map is simpler than tracking which entry belongs to `base`. Without
+    // this, scopedLedgers entries for the pruned workspace hold a pre-prune
+    // MetricsLedger that snapshotUnitMetricsByScope would merge back in, causing
+    // pruned units to reappear in subsequent snapshots.
+    scopedLedgers.clear();
     return removed;
 }
 /**
@@ -461,6 +620,133 @@ function deduplicateUnits(units) {
     }
     return Array.from(map.values());
 }
+// How long a lock file must be untouched (in ms) before it is considered
+// orphaned from a crashed process. Set to 2× the acquire timeout.
+export const STALE_LOCK_THRESHOLD_MS = 4000;
+// Retry interval between lock acquire attempts (ms). Caps syscall rate at
+// ~200 attempts over a 2s timeout instead of ~20,000 without any sleep.
+// Exposed for tests.
+export const LOCK_RETRY_INTERVAL_MS = 5;
+// Sync sleep via Atomics.wait — true OS-level sleep, no CPU spin.
+// Int32Array must reference a SharedArrayBuffer; we wait on index 0 which
+// will never be woken by a Atomics.notify, so the wait always times out.
+const _lockSleepBuf = new Int32Array(new SharedArrayBuffer(4));
+function syncSleep(ms) {
+    Atomics.wait(_lockSleepBuf, 0, 0, ms);
+}
+// Counts the number of sleepy retries (non-stale-evicting) made by acquireLock
+// across all calls since the last reset. Exported for test instrumentation only.
+let _lockSleepyRetries = 0;
+export function getLockSleepyRetries() { return _lockSleepyRetries; }
+export function resetLockSleepyRetries() { _lockSleepyRetries = 0; }
+/**
+ * Acquire an exclusive .lock sentinel file via O_EXCL.
+ *
+ * Improvements over the original:
+ *  - No busy spin: the inner `while (Date.now() < waitUntil) {}` spin that
+ *    burned CPU doing nothing useful is removed. Each retry attempt now makes
+ *    one `openSync` syscall and immediately re-checks the deadline, which is
+ *    orders of magnitude cheaper than a tight spin loop.
+ *  - Stale-lock detection: if the existing lock file's mtime is older than
+ *    STALE_LOCK_THRESHOLD_MS, the lock is considered orphaned (e.g. the
+ *    writing process crashed) and is forcibly removed before retrying.
+ *    A warning is logged so operators can detect crash patterns.
+ *  - PID stamp: on success, writes the acquiring process's PID and a
+ *    timestamp into the lock file so external monitors can identify orphans.
+ *  - Retry sleep: after each non-stale-evicting retry, sleeps
+ *    LOCK_RETRY_INTERVAL_MS (5ms) via Atomics.wait so the process yields to
+ *    the OS. This caps syscall rate at ~200–400/s under contention instead of
+ *    the ~20,000/s that would result from a tight openSync loop.
+ *    After a stale-lock eviction (lock already removed), no sleep is injected
+ *    — we retry immediately to close the short race window.
+ *
+ * Returns true on success, false on timeout.
+ */
+function acquireLock(lockPath, timeoutMs = 2000) {
+    const deadline = Date.now() + timeoutMs;
+    while (Date.now() < deadline) {
+        try {
+            const fd = openSync(lockPath, "wx"); // O_WRONLY | O_CREAT | O_EXCL
+            closeSync(fd);
+            // Write PID stamp so external monitors can identify the lock owner.
+            try {
+                writeFileSync(lockPath, `${process.pid}\n${new Date().toISOString()}\n`, "utf-8");
+            }
+            catch { /* non-fatal — stamp is diagnostic only */ }
+            return true;
+        }
+        catch {
+            // Lock held by another process — check for staleness before retrying.
+            try {
+                const st = statSync(lockPath);
+                if (Date.now() - st.mtimeMs > STALE_LOCK_THRESHOLD_MS) {
+                    logWarning("fs", `stale metrics lock at ${lockPath} (age ${Date.now() - st.mtimeMs}ms); forcibly removing and retrying`);
+                    try {
+                        unlinkSync(lockPath);
+                    }
+                    catch { /* already gone */ }
+                    // Do NOT sleep after stale-lock eviction — retry the open
+                    // immediately. The lock file was just removed; a short race window
+                    // exists and sleeping here would unnecessarily delay recovery.
+                    continue;
+                }
+            }
+            catch { /* lock file disappeared between the failed open and stat — retry */ }
+            // Sleep between retries to yield to the OS and cap syscall rate.
+            // Uses Atomics.wait for a true blocking sleep (no CPU spin).
+            _lockSleepyRetries++;
+            syncSleep(LOCK_RETRY_INTERVAL_MS);
+        }
+    }
+    return false;
+}
+function releaseLock(lockPath) {
+    try {
+        unlinkSync(lockPath);
+    }
+    catch { /* ignore */ }
+}
+/**
+ * Save the ledger with cross-process merge semantics.
+ *
+ * Acquires a .lock sentinel file, reads the current on-disk ledger,
+ * merges worker units with existing peer units (worker's entry wins on
+ * type+id+startedAt conflict since it has the latest finishedAt),
+ * then writes atomically. This prevents parallel auto-mode workers from
+ * silently discarding each other's metrics entries.
+ *
+ * Falls back to a direct write (no merge) if the lock cannot be acquired
+ * within the timeout — better to potentially overwrite than to lose data
+ * entirely.
+ */
 function saveLedger(base, data) {
-    saveJsonFile(metricsPath(base), data);
+    const path = metricsPath(base);
+    const lockPath = `${path}.lock`;
+    const acquired = acquireLock(lockPath);
+    if (acquired) {
+        try {
+            // Read current on-disk state and merge with worker's in-memory units.
+            // Worker units take precedence on conflict (by finishedAt in deduplicateUnits).
+            const onDisk = loadJsonFileOrNull(path, isMetricsLedger);
+            if (onDisk && onDisk.units.length > 0) {
+                const merged = deduplicateUnits([...onDisk.units, ...data.units]);
+                saveJsonFile(path, { ...data, units: merged });
+            }
+            else {
+                saveJsonFile(path, data);
+            }
+        }
+        finally {
+            releaseLock(lockPath);
+        }
+    }
+    else {
+        // Lock could not be acquired within the timeout. Fall back to a direct
+        // write (no cross-process merge) to avoid losing this worker's data
+        // entirely. A concurrent writer may overwrite us, but that is preferable
+        // to a torn write caused by two writers simultaneously executing the
+        // read-merge-write sequence without mutual exclusion.
+        logWarning("fs", "saveLedger: lock not acquired — falling back to direct write (no merge)");
+        saveJsonFile(path, data);
+    }
 }

package/dist/resources/extensions/gsd/parallel-merge.js

@@ -8,7 +8,7 @@ import { existsSync, readdirSync } from "node:fs";
 import { join } from "node:path";
 import { spawnSync } from "node:child_process";
 import { loadFile } from "./files.js";
-import { resolveMilestoneFile } from "./paths.js";
+import { resolveGsdPathContract, resolveMilestoneFile } from "./paths.js";
 import { mergeMilestoneToMain } from "./auto-worktree.js";
 import { MergeConflictError } from "./git-service.js";
 import { removeSessionStatus } from "./session-status-io.js";
@@ -16,12 +16,13 @@ import { getErrorMessage } from "./error-utils.js";
 import { logWarning } from "./workflow-logger.js";
 // ─── Merge Queue ───────────────────────────────────────────────────────────
 /**
- * Check whether a milestone is complete by querying its worktree SQLite DB.
+ * Check whether a milestone is complete by querying the canonical project DB.
  * Uses a subprocess to avoid disrupting the global DB singleton.
- * Returns true when milestones.status = 'complete' in the worktree's gsd.db.
+ * Returns true when milestones.status = 'complete' in project gsd.db.
  */
-export function isMilestoneCompleteInWorktreeDb(basePath, mid) {
-    const dbPath = join(basePath, ".gsd", "worktrees", mid, ".gsd", "gsd.db");
+export function isMilestoneCompleteInProjectDb(basePath, mid) {
+    const workRoot = join(basePath, ".gsd", "worktrees", mid);
+    const dbPath = resolveGsdPathContract(workRoot, basePath).projectDb;
     if (!existsSync(dbPath))
         return false;
     try {
@@ -34,15 +35,15 @@ export function isMilestoneCompleteInWorktreeDb(basePath, mid) {
     }
 }
 /**
- * Discover milestone IDs with status='complete' in their worktree DB,
- * scanning .gsd/worktrees/<MID>/.gsd/gsd.db for each worktree directory.
+ * Discover milestone IDs with status='complete' in the canonical DB,
+ * using worktree directories only to enumerate active parallel workers.
  */
 function discoverDbCompletedMilestones(basePath) {
     const completed = new Set();
     const worktreeDir = join(basePath, ".gsd", "worktrees");
     try {
         for (const entry of readdirSync(worktreeDir)) {
-            if (entry.startsWith("M") && isMilestoneCompleteInWorktreeDb(basePath, entry)) {
+            if (entry.startsWith("M") && isMilestoneCompleteInProjectDb(basePath, entry)) {
                 completed.add(entry);
             }
         }
@@ -57,15 +58,15 @@ function discoverDbCompletedMilestones(basePath) {
  * Sequential: merge in milestone ID order (M001 before M002).
  * By-completion: merge in the order milestones finished.
  *
- * When basePath is provided, also checks worktree SQLite DBs as the
- * source of truth — workers with stale orchestrator state (e.g. "error")
- * are included if their worktree DB shows status='complete'.
+ * When basePath is provided, also checks the canonical project DB as the
+ * source of truth. Workers with stale orchestrator state (e.g. "error")
+ * are included if their project DB row shows status='complete'.
  * See: https://github.com/gsd-build/gsd-2/issues/2812
  */
 export function determineMergeOrder(workers, order = "sequential", basePath) {
     // Start with workers the orchestrator already knows are stopped
     const stoppedIds = new Set(workers.filter(w => w.state === "stopped").map(w => w.milestoneId));
-    // When basePath is available, also check worktree DBs for milestones
+    // When basePath is available, also check the project DB for milestones
     // whose orchestrator state is stale but are actually complete (#2812)
     const dbCompleted = basePath ? discoverDbCompletedMilestones(basePath) : new Set();
     // Union: milestone is mergeable if stopped OR DB-complete
@@ -80,7 +81,7 @@ export function determineMergeOrder(workers, order = "sequential", basePath) {
             allMergeable.push(w);
         }
         else {
-            // Milestone discovered from worktree DB but not in workers list
+            // Milestone discovered from project DB but not in workers list
             allMergeable.push({
                 milestoneId: mid,
                 title: mid,

package/dist/resources/extensions/gsd/parallel-monitor-overlay.js

@@ -13,6 +13,7 @@ import { spawnSync } from "node:child_process";
 import { matchesKey, Key } from "@gsd/pi-tui";
 import { formatDuration } from "../shared/mod.js";
 import { formattedShortcutPair } from "./shortcut-defs.js";
+import { resolveGsdPathContract } from "./paths.js";
 // ─── Data Helpers ─────────────────────────────────────────────────────────
 function readJsonSafe(filePath) {
     try {
@@ -74,7 +75,8 @@ function discoverWorkers(basePath) {
     return [...mids].sort();
 }
 function querySliceProgress(basePath, mid) {
-    const dbPath = join(basePath, ".gsd", "worktrees", mid, ".gsd", "gsd.db");
+    const workRoot = join(basePath, ".gsd", "worktrees", mid);
+    const dbPath = resolveGsdPathContract(workRoot, basePath).projectDb;
     if (!existsSync(dbPath))
         return [];
     try {
@@ -119,7 +121,8 @@ function extractCostFromNdjson(basePath, mid) {
     }
 }
 function queryRecentCompletions(basePath, mid) {
-    const dbPath = join(basePath, ".gsd", "worktrees", mid, ".gsd", "gsd.db");
+    const workRoot = join(basePath, ".gsd", "worktrees", mid);
+    const dbPath = resolveGsdPathContract(workRoot, basePath).projectDb;
     if (!existsSync(dbPath))
         return [];
     try {

package/dist/resources/extensions/gsd/paths.js

@@ -1,3 +1,4 @@
+// GSD-2 — ID-based path resolution for GSD project files and directories
 /**
  * GSD Paths — ID-based path resolution
  *
@@ -9,12 +10,13 @@
  * via prefix matching, so existing projects work without migration.
  */
 import { readdirSync, existsSync, realpathSync, Dirent } from "node:fs";
-import { join, dirname, normalize } from "node:path";
+import { join, dirname, normalize, resolve } from "node:path";
 import { homedir } from "node:os";
 import { spawnSync } from "node:child_process";
 import { nativeScanGsdTree } from "./native-parser-bridge.js";
 import { DIR_CACHE_MAX } from "./constants.js";
 import { gsdHome } from "./gsd-home.js";
+import { isGsdWorktreePath, resolveWorktreeProjectRoot } from "./worktree-root.js";
 // ─── Directory Listing Cache ──────────────────────────────────────────────────
 const dirEntryCache = new Map();
 const dirListCache = new Map();
@@ -120,9 +122,15 @@ function cachedReaddir(dirPath) {
     return entries;
 }
 /**
- * Clear the directory listing cache.
+ * Clear the volatile directory listing caches.
  * Call after milestone transitions, file creation in planning directories,
  * or at the start/end of a dispatch cycle.
+ *
+ * NOTE: This does NOT clear gsdRootCache. The project root is stable for
+ * the lifetime of a process; clearing it on every agent turn-end caused a
+ * 250–2500 ms regression per session (git rev-parse + dir walk per turn).
+ * Use _clearGsdRootCache() at session-reset boundaries (workspace switch,
+ * process exit) when the project root may genuinely change.
  */
 export function clearPathCache() {
     dirEntryCache.clear();
@@ -269,11 +277,78 @@ const LEGACY_GSD_ROOT_FILES = {
     CODEBASE: "codebase.md",
 };
 // ─── GSD Root Discovery ───────────────────────────────────────────────────────
+// Process-lifetime cache for gsdRoot() results.
+// Keys are realpath-normalized (via normCacheKey) so /foo and /foo/ share the
+// same entry and so do case-variant paths on case-insensitive volumes. This
+// normalization is the safety net that prevents cache poisoning from the
+// ~/.gsd walk-up bug (fixed in c46cf4786 + b35e070eb), making it safe to
+// hold this cache for the entire process lifetime.
+// Use _clearGsdRootCache() only at session-reset boundaries (workspace switch,
+// process exit) — NOT inside clearPathCache(), which runs on every agent turn.
 const gsdRootCache = new Map();
-/** Exported for tests only — do not call in production code. */
+export function resolveGsdPathContract(workRoot, originalProjectRoot) {
+    const resolvedWorkRoot = resolve(workRoot || process.cwd());
+    const isWorktree = isGsdWorktreePath(resolvedWorkRoot);
+    if (isWorktree && !originalProjectRoot?.trim()) {
+        const externalMatch = /[/\\]\.gsd[/\\]projects[/\\][^/\\]+[/\\]worktrees(?:[/\\]|$)/.exec(resolvedWorkRoot);
+        if (externalMatch) {
+            const worktreesIdx = externalMatch[0].search(/[/\\]worktrees(?:[/\\]|$)/);
+            const projectGsd = resolvedWorkRoot.slice(0, externalMatch.index + worktreesIdx);
+            return {
+                projectRoot: dirname(dirname(projectGsd)),
+                workRoot: resolvedWorkRoot,
+                projectGsd,
+                worktreeGsd: join(resolvedWorkRoot, ".gsd"),
+                projectDb: join(projectGsd, "gsd.db"),
+                isWorktree,
+            };
+        }
+    }
+    const projectRoot = resolve(resolveWorktreeProjectRoot(resolvedWorkRoot, originalProjectRoot));
+    const projectGsd = join(projectRoot, ".gsd");
+    const worktreeGsd = isWorktree ? join(resolvedWorkRoot, ".gsd") : null;
+    return {
+        projectRoot,
+        workRoot: resolvedWorkRoot,
+        projectGsd,
+        worktreeGsd,
+        projectDb: join(projectGsd, "gsd.db"),
+        isWorktree,
+    };
+}
+/**
+ * Invalidate the gsdRoot cache.
+ * Use ONLY at session-reset boundaries: workspace switch, process exit, or
+ * any context where the project root itself may genuinely change.
+ * Do NOT call this on every agent turn — use clearPathCache() for volatile
+ * directory listing invalidation instead.
+ */
 export function _clearGsdRootCache() {
     gsdRootCache.clear();
 }
+/**
+ * Resolve a path to its canonical real path using the native resolver.
+ * On macOS case-insensitive (HFS+/APFS) volumes, realpathSync.native normalizes
+ * case — ensuring that /foo/Bar and /foo/bar resolve to the same string.
+ * Falls back to resolve(p) for non-existent paths.
+ *
+ * Use this helper everywhere a path is used as an identity/cache key so that
+ * all callers agree on the canonical form.
+ */
+export function normalizeRealPath(p) {
+    try {
+        return realpathSync.native(p);
+    }
+    catch {
+        return resolve(p);
+    }
+}
+/** Normalize a path for use as a gsdRootCache key (realpath + trailing-slash strip). */
+function normCacheKey(p) {
+    const r = normalizeRealPath(p);
+    const s = r.replaceAll("\\", "/").replace(/\/+$/, "");
+    return process.platform === "win32" ? s.toLowerCase() : s;
+}
 /**
  * Resolve the `.gsd` directory for a given project base path.
  *
@@ -283,19 +358,25 @@ export function _clearGsdRootCache() {
  *   3. Walk up from basePath — handles moved .gsd in an ancestor (bounded by git root)
  *   4. basePath/.gsd         — creation fallback (init scenario)
  *
- * Result is cached per basePath for the process lifetime.
+ * Result is cached per normalized basePath for the process lifetime.
+ * Keys are realpath-normalized so /foo and /foo/ share the same cache entry.
  */
 export function gsdRoot(basePath) {
-    const cached = gsdRootCache.get(basePath);
+    const cacheKey = normCacheKey(basePath);
+    const cached = gsdRootCache.get(cacheKey);
     if (cached)
         return cached;
-    const result = probeGsdRoot(basePath);
+    // Canonicalize result via realpath before asserting and caching so that
+    // callers always receive a canonical path regardless of whether probeGsdRoot
+    // returned a path through a symlink. Without this, the cached value can
+    // diverge from other realpath-normalized paths (e.g. workspace.identityKey).
+    const result = normalizeRealPath(probeGsdRoot(basePath));
     // Defense-in-depth: if basePath resolves to the user's home directory and
     // the result equals gsdHome(), refuse — project-scoped writes must never
     // land in the global ~/.gsd. Paths under ~/.gsd/projects/<hash>/ are still
     // valid (their basePath does not equal homedir).
     assertNotGlobalGsdHome(basePath, result);
-    gsdRootCache.set(basePath, result);
+    gsdRootCache.set(cacheKey, result);
     return result;
 }
 function assertNotGlobalGsdHome(basePath, result) {
@@ -362,6 +443,9 @@ function isInsideGsdWorktree(p) {
     return false;
 }
 function probeGsdRoot(rawBasePath) {
+    const contract = resolveGsdPathContract(rawBasePath);
+    if (contract.isWorktree)
+        return contract.projectGsd;
     // 1. Fast path — check the input path directly
     const local = join(rawBasePath, ".gsd");
     if (existsSync(local))
@@ -401,9 +485,30 @@ function probeGsdRoot(rawBasePath) {
         }
     }
     catch { /* git not available */ }
+    // Compute gsdHome once for the skip-check used in steps 2 and 3.
+    const normPath = (p) => {
+        let r;
+        try {
+            r = realpathSync.native(p);
+        }
+        catch {
+            r = p;
+        }
+        const s = r.replaceAll("\\", "/").replace(/\/+$/, "");
+        return process.platform === "win32" ? s.toLowerCase() : s;
+    };
+    let gsdHomeNorm;
+    try {
+        gsdHomeNorm = normPath(gsdHome());
+    }
+    catch {
+        gsdHomeNorm = "";
+    }
     if (gitRoot) {
         const candidate = join(gitRoot, ".gsd");
-        if (existsSync(candidate))
+        // Skip if the candidate resolves to the global GSD home — a subdir basePath
+        // must not be anchored to ~/.gsd just because $HOME is a git repo.
+        if (existsSync(candidate) && normPath(candidate) !== gsdHomeNorm)
             return candidate;
     }
     // 3. Walk up from basePath to the git root (only if we are in a subdirectory)
@@ -411,7 +516,7 @@ function probeGsdRoot(rawBasePath) {
         let cur = dirname(basePath);
         while (cur !== basePath) {
             const candidate = join(cur, ".gsd");
-            if (existsSync(candidate))
+            if (existsSync(candidate) && normPath(candidate) !== gsdHomeNorm)
                 return candidate;
             if (cur === gitRoot)
                 break;