gsd-pi 2.78.1-dev.e9d88a536 → 2.78.1-dev.eccf86e27

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

@@ -1,3 +1,4 @@
+// GSD-2 + metrics.ts: token & cost tracking for auto-mode units
 /**
  * GSD Metrics — Token & Cost Tracking
  *
@@ -14,6 +15,7 @@
  */
 
 import { join } from "node:path";
+import { openSync, closeSync, unlinkSync, statSync, writeFileSync } from "node:fs";
 import type { ExtensionContext } from "@gsd/pi-coding-agent";
 import { gsdRoot } from "./paths.js";
 import { getAndClearSkills } from "./skill-telemetry.js";
@@ -21,6 +23,8 @@ import { loadJsonFile, loadJsonFileOrNull, saveJsonFile } from "./json-persisten
 import { parseUnitId } from "./unit-id.js";
 import { buildAuditEnvelope, emitUokAuditEvent } from "./uok/audit.js";
 import { isUnifiedAuditEnabled } from "./uok/audit-toggle.js";
+import type { MilestoneScope } from "./workspace.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
@@ -108,11 +112,17 @@ export function classifyUnitPhase(unitType: string): MetricsPhase {
 let ledger: MetricsLedger | null = null;
 let basePath: string = "";
 
+// Per-workspace ledger map, keyed by workspace.identityKey.
+// Populated by initMetricsByScope; independent of the module singleton.
+const scopedLedgers = new Map<string, MetricsLedger>();
+
 // ─── 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: string): void {
   basePath = base;
@@ -121,6 +131,8 @@ export function initMetrics(base: string): void {
 
 /**
  * Reset in-memory state. Called when auto-mode stops.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use resetMetricsByScope instead.
  */
 export function resetMetrics(): void {
   ledger = null;
@@ -130,6 +142,8 @@ export function resetMetrics(): void {
 /**
  * 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: ExtensionContext,
@@ -272,6 +286,182 @@ export function getLedger(): MetricsLedger | null {
   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: MilestoneScope): void {
+  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: MilestoneScope): MetricsLedger | null {
+  return scopedLedgers.get(scope.workspace.identityKey) ?? null;
+}
+
+/**
+ * Reset scoped in-memory state for a workspace. Called when auto-mode stops.
+ */
+export function resetMetricsByScope(scope: MilestoneScope): void {
+  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: MilestoneScope,
+  ctx: ExtensionContext,
+  unitType: string,
+  unitId: string,
+  startedAt: number,
+  model: string,
+  opts?: {
+    tier?: string;
+    modelDowngraded?: boolean;
+    contextWindowTokens?: number;
+    truncationSections?: number;
+    continueHereFired?: boolean;
+    promptCharCount?: number;
+    baselineCharCount?: number;
+    autoSessionKey?: string;
+    traceId?: string;
+    turnId?: string;
+    causedBy?: string;
+  },
+): UnitMetrics | null {
+  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: TokenCounts = { 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 as any).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: UnitMetrics = {
+    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;
+}
+
 // ─── Aggregation helpers ──────────────────────────────────────────────────────
 
 export interface PhaseAggregate {
@@ -593,6 +783,12 @@ export function pruneMetricsLedger(base: string, keepCount: number): number {
   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;
 }
 
@@ -635,6 +831,130 @@ function deduplicateUnits(units: UnitMetrics[]): UnitMetrics[] {
   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: number): void {
+  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(): number { return _lockSleepyRetries; }
+export function resetLockSleepyRetries(): void { _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: string, timeoutMs = 2000): boolean {
+  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 ${Date.now() - st.mtimeMs}ms); forcibly removing and retrying`,
+          );
+          try { unlinkSync(lockPath); } catch { /* already gone */ }
+          // Do NOT sleep after stale-lock eviction — retry the open
+          // immediately. The lock file was just removed; a short race window
+          // exists and sleeping here would unnecessarily delay recovery.
+          continue;
+        }
+      } catch { /* lock file disappeared between the failed open and stat — retry */ }
+      // Sleep between retries to yield to the OS and cap syscall rate.
+      // Uses Atomics.wait for a true blocking sleep (no CPU spin).
+      _lockSleepyRetries++;
+      syncSleep(LOCK_RETRY_INTERVAL_MS);
+    }
+  }
+  return false;
+}
+
+function releaseLock(lockPath: string): void {
+  try { unlinkSync(lockPath); } catch { /* ignore */ }
+}
+
+/**
+ * Save the ledger with cross-process merge semantics.
+ *
+ * Acquires a .lock sentinel file, reads the current on-disk ledger,
+ * merges worker units with existing peer units (worker's entry wins on
+ * type+id+startedAt conflict since it has the latest finishedAt),
+ * then writes atomically. This prevents parallel auto-mode workers from
+ * silently discarding each other's metrics entries.
+ *
+ * Falls back to a direct write (no merge) if the lock cannot be acquired
+ * within the timeout — better to potentially overwrite than to lose data
+ * entirely.
+ */
 function saveLedger(base: string, data: MetricsLedger): void {
-  saveJsonFile(metricsPath(base), data);
+  const path = metricsPath(base);
+  const lockPath = `${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/src/resources/extensions/gsd/paths.ts

@@ -1,3 +1,4 @@
+// GSD-2 — ID-based path resolution for GSD project files and directories
 /**
  * GSD Paths — ID-based path resolution
  *
@@ -128,9 +129,15 @@ function cachedReaddir(dirPath: string): string[] {
 }
 
 /**
- * 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(): void {
   dirEntryCache.clear();
@@ -285,6 +292,14 @@ const LEGACY_GSD_ROOT_FILES: Record<GSDRootFileKey, string> = {
 
 // ─── 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<string, string>();
 
 export interface GsdPathContract {
@@ -337,11 +352,37 @@ export function resolveGsdPathContract(
   };
 }
 
-/** Exported for tests only — do not call in production code. */
+/**
+ * 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(): void {
   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: string): string {
+  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: string): string {
+  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.
  *
@@ -351,13 +392,19 @@ export function _clearGsdRootCache(): void {
  *   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: string): string {
-  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
@@ -365,7 +412,7 @@ export function gsdRoot(basePath: string): string {
   // valid (their basePath does not equal homedir).
   assertNotGlobalGsdHome(basePath, result);
 
-  gsdRootCache.set(basePath, result);
+  gsdRootCache.set(cacheKey, result);
   return result;
 }
 
@@ -466,9 +513,21 @@ function probeGsdRoot(rawBasePath: string): string {
     }
   } catch { /* git not available */ }
 
+  // Compute gsdHome once for the skip-check used in steps 2 and 3.
+  const normPath = (p: string): string => {
+    let r: string;
+    try { r = realpathSync.native(p); } catch { r = p; }
+    const s = r.replaceAll("\\", "/").replace(/\/+$/, "");
+    return process.platform === "win32" ? s.toLowerCase() : s;
+  };
+  let gsdHomeNorm: string;
+  try { gsdHomeNorm = normPath(gsdHome()); } catch { gsdHomeNorm = ""; }
+
   if (gitRoot) {
     const candidate = join(gitRoot, ".gsd");
-    if (existsSync(candidate)) return 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)
@@ -476,7 +535,7 @@ function probeGsdRoot(rawBasePath: string): string {
     let cur = dirname(basePath);
     while (cur !== basePath) {
       const candidate = join(cur, ".gsd");
-      if (existsSync(candidate)) return candidate;
+      if (existsSync(candidate) && normPath(candidate) !== gsdHomeNorm) return candidate;
       if (cur === gitRoot) break;
       basePath = cur;
       cur = dirname(cur);

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

@@ -28,7 +28,7 @@ This unit runs under the `planning-dispatch` tools-policy: you may use the `suba
 - **Touched auth, network, parsing, file IO, shell exec, or crypto** → dispatch the **security** agent for an OWASP-style audit.
 - **Added or modified tests** → dispatch the **tester** agent to assess coverage gaps relative to the slice plan.
 
-Subagents read the diff and report findings — they do **not** write user source. You remain responsible for acting on their feedback before calling `gsd_complete_slice` with `milestoneId` and `sliceId`.
+Subagents read the diff and report findings — they do **not** write user source. You remain responsible for acting on their feedback before calling `gsd_slice_complete` with `milestoneId` and `sliceId`.
 
 Then:
 1. Use the **Slice Summary** and **UAT** output templates from the inlined context above
@@ -37,11 +37,11 @@ Then:
 4. If the slice plan includes observability/diagnostic surfaces, confirm they work. Skip this for simple slices that don't have observability sections.
 5. Address every gate listed in the **Gates to Close** section above — each gate maps to a specific slice-summary section the handler inspects (for example, Q8 maps to **Operational Readiness**: health signal, failure signal, recovery procedure, and monitoring gaps). Leaving a section empty records the gate as `omitted`.
 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.
-7. Prepare the slice completion content you will pass to `gsd_complete_slice` using the camelCase fields `milestoneId`, `sliceId`, `sliceTitle`, `oneLiner`, `narrative`, `verification`, and `uatContent`. Do **not** manually write `{{sliceSummaryPath}}`. Do **not** manually write `{{sliceUatPath}}` — the DB-backed tool is the canonical write path for both artifacts.
+7. Prepare the slice completion content you will pass to `gsd_slice_complete` using the camelCase fields `milestoneId`, `sliceId`, `sliceTitle`, `oneLiner`, `narrative`, `verification`, and `uatContent`. Do **not** manually write `{{sliceSummaryPath}}`. Do **not** manually write `{{sliceUatPath}}` — the DB-backed tool is the canonical write path for both artifacts.
 8. Draft the UAT content you will pass as `uatContent` — a concrete UAT script with real test cases derived from the slice plan and task summaries. Include preconditions, numbered steps with expected outcomes, and edge cases. This must NOT be a placeholder or generic template — tailor every test case to what this slice actually built. Fill the `UAT Type` and `Not Proven By This UAT` sections explicitly so the artifact states what class of acceptance it covers and what still remains unproven (e.g. live integration paths, performance under load, scenarios deferred to a later slice).
 9. Review task summaries for `key_decisions`. For each significant decision, call `capture_thought` with `category: "architecture"` (or `"pattern"`) and a `structuredFields` payload of `{ scope, decision, choice, rationale, made_by: "agent", revisable }`.
 10. Review task summaries for patterns, gotchas, or non-obvious lessons learned. For each one that would save future agents from repeating investigation, call `capture_thought` with the matching category (`gotcha`, `convention`, `pattern`, `environment`). The memory store is the single source of truth (ADR-013); do not append to `.gsd/DECISIONS.md` or `.gsd/KNOWLEDGE.md` directly.
-11. Call `gsd_complete_slice` with the camelCase fields `milestoneId`, `sliceId`, `sliceTitle`, `oneLiner`, `narrative`, `verification`, and `uatContent`, plus any optional enrichment fields you have. Do NOT manually mark the roadmap checkbox — the tool writes to the DB, renders `{{sliceSummaryPath}}` and `{{sliceUatPath}}`, and updates the ROADMAP.md projection automatically.
+11. Call `gsd_slice_complete` with the camelCase fields `milestoneId`, `sliceId`, `sliceTitle`, `oneLiner`, `narrative`, `verification`, and `uatContent`, plus any optional enrichment fields you have. Do NOT manually mark the roadmap checkbox — the tool writes to the DB, renders `{{sliceSummaryPath}}` and `{{sliceUatPath}}`, and updates the ROADMAP.md projection automatically.
 12. Do not run git commands — the system commits your changes and handles any merge after this unit succeeds.
 13. Update `.gsd/PROJECT.md` if it exists — refresh current state if needed: use the `write` tool with `path: ".gsd/PROJECT.md"` and `content` containing the full updated document reflecting current project state. Do NOT use the `edit` tool for this — PROJECT.md is a full-document refresh.
 
@@ -49,6 +49,6 @@ Then:
 
 **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.**
+**You MUST call `gsd_slice_complete` with the slice summary and UAT content before finishing. The tool persists to both DB and disk and renders `{{sliceSummaryPath}}` and `{{sliceUatPath}}` automatically.**
 
 When done, say: "Slice {{sliceId}} complete."

package/src/resources/extensions/gsd/prompts/execute-task.md

@@ -85,14 +85,14 @@ Then:
 17. If you made an architectural, pattern, library, or observability decision during this task that downstream work should know about, call `capture_thought` with `category: "architecture"` (or `"pattern"`). For decisions, populate `structuredFields` with `{ scope, decision, choice, rationale, made_by: "agent", revisable }` so future projection back to a human-visible decisions register stays lossless. Not every task produces decisions — only capture when a meaningful choice was made.
 18. If you discover a non-obvious rule, recurring gotcha, or useful pattern during execution, call `capture_thought` with `category: "gotcha"`, `"convention"`, `"pattern"`, or `"environment"` as appropriate. Only capture entries that would save future agents from repeating your investigation — don't capture obvious things. The memory store is the single source of truth for cross-session knowledge (ADR-013); do not append to `.gsd/DECISIONS.md` or `.gsd/KNOWLEDGE.md` directly.
 19. Read the template at `~/.gsd/agent/extensions/gsd/templates/task-summary.md`
-20. Use that template to prepare the completion content you will pass to `gsd_complete_task` using the camelCase fields `milestoneId`, `sliceId`, `taskId`, `oneLiner`, `narrative`, `verification`, and `verificationEvidence`. Do **not** manually write `{{taskSummaryPath}}` — the DB-backed tool is the canonical write path and renders the summary file for you.
-21. Call `gsd_complete_task` with milestoneId, sliceId, taskId, and the completion fields derived from the template. This is your final required step — do NOT manually edit PLAN.md checkboxes. The tool marks the task complete, updates the DB, renders `{{taskSummaryPath}}`, and updates PLAN.md automatically.
+20. Use that template to prepare the completion content you will pass to `gsd_task_complete` using the camelCase fields `milestoneId`, `sliceId`, `taskId`, `oneLiner`, `narrative`, `verification`, and `verificationEvidence`. Do **not** manually write `{{taskSummaryPath}}` — the DB-backed tool is the canonical write path and renders the summary file for you.
+21. Call `gsd_task_complete` with milestoneId, sliceId, taskId, and the completion fields derived from the template. This is your final required step — do NOT manually edit PLAN.md checkboxes. The tool marks the task complete, updates the DB, renders `{{taskSummaryPath}}`, and updates PLAN.md automatically.
 22. Do not run git commands — the system reads your task summary after completion and creates a meaningful commit from it (type inferred from title, message from your one-liner, key files from frontmatter). Write a clear, specific one-liner in the summary — it becomes the commit message.
 
 All work stays in your working directory: `{{workingDirectory}}`.
 
 **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 task summary. If a decision genuinely requires human input, note it in the summary and proceed with the best available option.
 
-**You MUST call `gsd_complete_task` before finishing. Do not manually write `{{taskSummaryPath}}`.**
+**You MUST call `gsd_task_complete` before finishing. Do not manually write `{{taskSummaryPath}}`.**
 
 When done, say: "Task {{taskId}} complete."

package/src/resources/extensions/gsd/prompts/guided-discuss-milestone.md

@@ -12,6 +12,13 @@ Discuss milestone {{milestoneId}} ("{{milestoneTitle}}"). Identify gray areas, a
 
 {{fastPathInstruction}}
 
+### Read project shape
+
+Before your first question round, read `.gsd/PROJECT.md` and look for `## Project Shape` → `**Complexity:**`. The verdict is either **`simple`** or **`complex`** (default to `complex` if PROJECT.md is missing the section, predates this convention, or the value is unclear). The verdict scales the rest of this stage:
+
+- **`simple`** — favor 1–2 plain-text question rounds. Skip the parallel-research investigation. Skip `ask_user_questions` unless presenting concrete alternatives.
+- **`complex`** — full investigation, 3–4-option structured questions, multi-round.
+
 ### Before your first question round
 
 Do a lightweight targeted investigation so your questions are grounded in reality:
@@ -36,7 +43,7 @@ Ask **1–3 questions per round**. Keep each question focused on one of:
 
 **Never fabricate or simulate user input.** Never generate fake transcript markers like `[User]`, `[Human]`, or `User:`. Ask one question round, then wait for the user's actual response before continuing.
 
-**If `{{structuredQuestionsAvailable}}` is `true`:** use `ask_user_questions` for each round. 1–3 questions per call, each as a separate question object. Keep option labels short (3–5 words). Always include a freeform "Other / let me explain" option. When the user picks that option or writes a long freeform answer, switch to plain text follow-up for that thread before resuming structured questions. **IMPORTANT: Call `ask_user_questions` exactly once per turn. Never make multiple calls with the same or overlapping questions — wait for the user's response before asking the next round.**
+**If `{{structuredQuestionsAvailable}}` is `true`:** use `ask_user_questions` for each round. 1–3 questions per call, each as a separate question object. Keep option labels short (3–5 words). In **`complex`** mode, each multi-choice question MUST present **3 or 4 concrete, researched options** plus a final **"Other — let me discuss"** option; options must be grounded in the investigation above (codebase signals, library docs, prior `.gsd/` artifacts), not generic placeholders. In **`simple`** mode, 2 options is fine when alternatives are genuinely binary. Binary depth-check / wrap-up gates are exempt from the 3-or-4 rule. When the user picks "Other — let me discuss" or writes a long freeform answer, switch to plain text follow-up for that thread before resuming structured questions. **IMPORTANT: Call `ask_user_questions` exactly once per turn. Never make multiple calls with the same or overlapping questions — wait for the user's response before asking the next round.**
 
 **If `{{structuredQuestionsAvailable}}` is `false`:** ask questions in plain text. Keep each round to 1–3 focused questions. Wait for answers before asking the next round.
 

package/src/resources/extensions/gsd/prompts/guided-discuss-project.md

@@ -26,6 +26,18 @@ Ask the user a single freeform question in plain text, not structured: **"What d
 
 Wait for their response. This grounds every follow-up in their own terminology.
 
+### Classify project shape
+
+After the opening answer, classify the project as **`simple`** or **`complex`** before continuing. Print the verdict in chat as one line: `Project shape: simple` or `Project shape: complex` followed by a one-line rationale.
+
+**`simple`** — most of these apply: single primary user (the user themselves or one immediate team), no external integrations beyond well-known SDKs/libs, greenfield or self-contained, scope describable in 1–2 sentences without ambiguity, no compliance/regulatory needs, ≤5 distinct capabilities.
+
+**`complex`** — any of these apply: multi-user with roles/permissions, non-trivial brownfield codebase, external integrations with auth/data exchange, compliance/security/regulated domain (PII, payments, healthcare), >5 capabilities or unclear scope, cross-team/cross-org coordination, novel domain where assumptions need validation.
+
+**Default to `complex` when uncertain.** The user can override the verdict in plain text; if they do, accept it and proceed.
+
+The verdict drives the rest of this stage and gets persisted to PROJECT.md → `## Project Shape`. Downstream stages (`discuss-requirements`, `discuss-milestone`, `discuss-slice`) read it from there.
+
 ### Before deeper rounds
 
 Do a lightweight targeted investigation so your questions are grounded in reality:
@@ -50,9 +62,11 @@ Ask **1–3 questions per round**. Each round targets one of:
 
 **Never fabricate or simulate user input.** Never generate fake transcript markers like `[User]`, `[Human]`, or `User:`. Ask one question round, then wait for the user's actual response before continuing.
 
-**Plain-text default:** Project discovery is open-ended. Ask question rounds in plain text unless you are presenting 2–3 concrete alternatives with clear tradeoffs.
+**Cadence is shape-dependent:**
+- **`simple`** — favor 1–2 plain-text rounds. Skip `ask_user_questions` unless you are presenting concrete alternatives. Get to the depth checklist fast.
+- **`complex`** — full investigation, multiple rounds, structured questions when meaningful alternatives exist.
 
-**If `{{structuredQuestionsAvailable}}` is `true` and you use `ask_user_questions`:** ask 1–3 questions per call. Every question object MUST include a stable lowercase `id`. Keep option labels short (3–5 words). Do not add a separate "Other" option; the question UI provides a freeform path automatically. Wait for each tool result before asking the next round.
+**If `{{structuredQuestionsAvailable}}` is `true` and you use `ask_user_questions`:** ask 1–3 questions per call. Every question object MUST include a stable lowercase `id`. Keep option labels short (3–5 words). In **`complex`** mode, each multi-choice question MUST present **3 or 4 concrete, researched options** plus a final **"Other — let me discuss"** option; options must be grounded in your investigation (codebase signals, library docs, prior `.gsd/` artifacts), not generic placeholders. In **`simple`** mode, 2 options is fine. Binary depth-check / wrap-up gates are exempt from the 3-or-4 rule. Wait for each tool result before asking the next round.
 
 **If `{{structuredQuestionsAvailable}}` is `false`:** ask questions in plain text. Keep each round to 1–3 focused questions.
 
@@ -126,8 +140,9 @@ Once the user confirms depth:
 
 1. Use the **Project** output template (inlined above).
 2. Call `gsd_summary_save` with `artifact_type: "PROJECT"` and the full project markdown as `content`; omit `milestone_id`. The tool writes `.gsd/PROJECT.md` to disk and persists to DB. Preserve the user's exact terminology, emphasis, and framing.
-3. The `## Capability Contract` section MUST reference `.gsd/REQUIREMENTS.md` — that file does not yet exist; the next stage (`discuss-requirements`) will produce it.
-4. The `## Milestone Sequence` MUST list at least M001 with title and one-liner. Subsequent milestones may be listed as known intents; they will be elaborated in their own discuss-milestone stages.
-5. Do NOT use `artifact_type: "CONTEXT"` and do NOT pass `milestone_id: "PROJECT"`; that creates a fake milestone named PROJECT.
-6. {{commitInstruction}}
-7. Say exactly: `"Project context written."` — nothing else.
+3. The `## Project Shape` section MUST contain `**Complexity:** simple` or `**Complexity:** complex` (matching the verdict you announced) plus a one-line `**Why:**` rationale. Downstream stages read this line.
+4. The `## Capability Contract` section MUST reference `.gsd/REQUIREMENTS.md` — that file does not yet exist; the next stage (`discuss-requirements`) will produce it.
+5. The `## Milestone Sequence` MUST list at least M001 with title and one-liner. Subsequent milestones may be listed as known intents; they will be elaborated in their own discuss-milestone stages.
+6. Do NOT use `artifact_type: "CONTEXT"` and do NOT pass `milestone_id: "PROJECT"`; that creates a fake milestone named PROJECT.
+7. {{commitInstruction}}
+8. Say exactly: `"Project context written."` — nothing else.