@opengsd/gsd-pi 1.3.0-dev.65546769 → 1.3.0-dev.965d1788

package/src/resources/extensions/gsd/layout-policy.ts

@@ -47,8 +47,9 @@ export function dbPath(basePath: string): string {
  * Used by the renderer to derive the phase number from the DB's milestone_id.
  */
 export function milestoneIdToPhaseNum(milestoneId: string): number {
-  // No $ anchor: accepts bare (M012) and team-suffixed (M012-abc123) IDs.
+  // No $ anchor: accepts bare (M012), team-suffixed (M012-abc123), and legacy numeric IDs.
   const m = milestoneId.match(/^M0*(\d+)/i);
+  if (!m && /^\d+$/.test(milestoneId)) return Number.parseInt(milestoneId, 10);
   return m ? Number.parseInt(m[1]!, 10) : 1;
 }
 

package/src/resources/extensions/gsd/markdown-renderer.ts

@@ -25,6 +25,7 @@ import {
   insertArtifact,
   deleteArtifactByPath,
   getGateResults,
+  isDbAvailable,
 } from "./gsd-db.js";
 import type { MilestoneRow, ArtifactRow } from "./db-milestone-artifact-rows.js";
 import type { SliceRow, TaskRow } from "./db-task-slice-rows.js";
@@ -631,16 +632,8 @@ export async function renderRoadmapCheckboxes(
   basePath: string,
   milestoneId: string,
 ): Promise<boolean> {
-  const slices = getMilestoneSlices(milestoneId);
-  if (slices.length === 0) {
-    process.stderr.write(
-      `markdown-renderer: no slices found for milestone ${milestoneId}\n`,
-    );
-    return false;
-  }
-
-  await renderRoadmapFromDb(basePath, milestoneId);
-  return true;
+  const rendered = await renderRoadmapFromDb(basePath, milestoneId);
+  return !("skipped" in rendered);
 }
 
 /**
@@ -737,7 +730,8 @@ function isAutoRecoveryPlaceholderPlan(content: string): boolean {
  * projection (the 4S/0T-vs-5S/13T drift class). The artifacts table is an
  * output sink, never a render input.
  *
- * @returns true if the plan was written, false on skip/error
+ * @returns true if the plan was written, false when the DB slice has no tasks
+ * @throws when the DB connection is unavailable or the render write fails
  */
 export async function renderPlanCheckboxes(
   basePath: string,
@@ -747,6 +741,9 @@ export async function renderPlanCheckboxes(
 ): Promise<boolean> {
   const tasks = getSliceTasks(milestoneId, sliceId);
   if (tasks.length === 0) {
+    if (!isDbAvailable()) {
+      throw new Error(`database unavailable while rendering plan checkboxes for ${milestoneId}/${sliceId}`);
+    }
     process.stderr.write(
       `markdown-renderer: no tasks found for ${milestoneId}/${sliceId}\n`,
     );

package/src/resources/extensions/gsd/memory-backfill.ts

@@ -49,8 +49,8 @@ type DecisionContentFields = Pick<DecisionRow, "decision" | "choice" | "rational
  *   decision is checked individually; only decisions whose id is already
  *   present in the memory store are skipped. A user-authored memory with
  *   their own `sourceDecisionId` does NOT abort the backfill.
- * - Best-effort: never throws. Logs and returns 0 on failure so a broken
- *   backfill cannot block agent startup.
+ * - Best-effort: never throws. Failed memory inserts are logged and skipped;
+ *   setup/read failures log and return 0 so backfill cannot block startup.
  * - Full migration (ADR-013 Stage 2a): both active and superseded rows are
  *   migrated. `structured_fields.superseded_by` preserves the supersedes
  *   chain so the DECISIONS.md projection regen can source from memories
@@ -146,24 +146,30 @@ export function backfillDecisionsToMemories(): number {
       }
 
       const content = synthesizeDecisionMemoryContent(row);
-      const id = createMemory({
-        category: "architecture",
-        content,
-        scope: row.scope || "project",
-        confidence: 0.85,
-        structuredFields: {
-          sourceDecisionId: row.id,
-          when_context: row.when_context,
-          scope: row.scope,
-          decision: row.decision,
-          choice: row.choice,
-          rationale: row.rationale,
-          made_by: row.made_by,
-          revisable: row.revisable,
-          superseded_by: row.superseded_by,
-        },
-      });
-      if (id) written += 1;
+      try {
+        const id = createMemory({
+          category: "architecture",
+          content,
+          scope: row.scope || "project",
+          confidence: 0.85,
+          structuredFields: {
+            sourceDecisionId: row.id,
+            when_context: row.when_context,
+            scope: row.scope,
+            decision: row.decision,
+            choice: row.choice,
+            rationale: row.rationale,
+            made_by: row.made_by,
+            revisable: row.revisable,
+            superseded_by: row.superseded_by,
+          },
+        });
+        if (id) written += 1;
+      } catch (rowErr) {
+        const message = rowErr instanceof Error ? rowErr.message : String(rowErr);
+        logWarning("memory-backfill", `decisions->memories backfill skipped ${row.id}: ${message}`);
+        continue;
+      }
     }
 
     if (healed > 0) {

package/src/resources/extensions/gsd/memory-consolidation-scanner.ts

@@ -142,23 +142,53 @@ function getActiveDecisions(): DecisionRow[] {
 }
 
 /**
- * True when a memory row has a `structured_fields` JSON payload containing
- * the given `markerKey: "value"` pair. Matches the LIKE pattern used by
- * `backfillDecisionsToMemories` so the scanner is consistent with the
- * backfill's idempotency check.
+ * Source IDs already present in memory `structured_fields`. Collected with one
+ * pass over the memories table so the scanner does not perform a non-indexable
+ * LIKE probe for every decision and KNOWLEDGE.md row.
  */
-function memoryHasSourceMarker(markerKey: string, value: string): boolean {
-  if (!isDbAvailable()) return false;
+interface MemorySourceMarkers {
+  decisionIds: Set<string>;
+  knowledgeIds: Set<string>;
+}
+
+function emptyMemorySourceMarkers(): MemorySourceMarkers {
+  return { decisionIds: new Set(), knowledgeIds: new Set() };
+}
+
+function getMemorySourceMarkers(): MemorySourceMarkers {
+  const markers = emptyMemorySourceMarkers();
+  if (!isDbAvailable()) return markers;
   const adapter = _getAdapter();
-  if (!adapter) return false;
+  if (!adapter) return markers;
+  try {
+    const rows = adapter
+      .prepare("SELECT structured_fields FROM memories WHERE structured_fields IS NOT NULL")
+      .all() as Array<Record<string, unknown>>;
+    for (const row of rows) {
+      collectMemorySourceMarker(markers, row["structured_fields"]);
+    }
+  } catch {
+    return markers;
+  }
+  return markers;
+}
+
+function collectMemorySourceMarker(markers: MemorySourceMarkers, raw: unknown): void {
+  if (typeof raw !== "string" || raw.length === 0) return;
   try {
-    const pattern = `%"${markerKey}":"${value}"%`;
-    const row = adapter
-      .prepare("SELECT 1 FROM memories WHERE structured_fields LIKE :pattern LIMIT 1")
-      .get({ ":pattern": pattern });
-    return row !== undefined;
+    const parsed = JSON.parse(raw);
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) return;
+    const fields = parsed as Record<string, unknown>;
+    const decisionId = fields["sourceDecisionId"];
+    if (typeof decisionId === "string" && decisionId.length > 0) {
+      markers.decisionIds.add(decisionId);
+    }
+    const knowledgeId = fields["sourceKnowledgeId"];
+    if (typeof knowledgeId === "string" && knowledgeId.length > 0) {
+      markers.knowledgeIds.add(knowledgeId);
+    }
   } catch {
-    return false;
+    return;
   }
 }
 
@@ -174,10 +204,14 @@ const SAMPLE_LIMIT = 5;
 export function scanConsolidationGaps(basePath: string): ConsolidationGapReport {
   // ── Decisions ────────────────────────────────────────────────────────
   const decisions = getActiveDecisions();
+  const knowledgeRows = parseKnowledgeRows(knowledgeMdContent(basePath));
+  const memorySourceMarkers = decisions.length > 0 || knowledgeRows.length > 0
+    ? getMemorySourceMarkers()
+    : emptyMemorySourceMarkers();
   const decisionSamples: DecisionsSurfaceReport["samples"] = [];
   let decisionMigrated = 0;
   for (const decision of decisions) {
-    if (memoryHasSourceMarker("sourceDecisionId", decision.id)) {
+    if (memorySourceMarkers.decisionIds.has(decision.id)) {
       decisionMigrated += 1;
       continue;
     }
@@ -190,16 +224,14 @@ export function scanConsolidationGaps(basePath: string): ConsolidationGapReport
   }
 
   // ── KNOWLEDGE.md ─────────────────────────────────────────────────────
-  const knowledgeRows = parseKnowledgeRows(knowledgeMdContent(basePath));
   const knowledgeByTable = { rules: 0, patterns: 0, lessons: 0 };
   const knowledgeSamples: KnowledgeSurfaceReport["samples"] = [];
   let knowledgeMigrated = 0;
   for (const row of knowledgeRows) {
     knowledgeByTable[row.table] += 1;
-    // Phase 6 will introduce a `sourceKnowledgeId` marker as part of the
-    // KNOWLEDGE.md backfill. Until that path ships, this check returns
-    // false for every row, which is the honest state of the consolidation.
-    if (memoryHasSourceMarker("sourceKnowledgeId", row.id)) {
+    // KNOWLEDGE.md backfill writes `sourceKnowledgeId`; rows without that
+    // marker are still reported as consolidation gaps.
+    if (memorySourceMarkers.knowledgeIds.has(row.id)) {
       knowledgeMigrated += 1;
       continue;
     }

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

@@ -19,7 +19,7 @@ import { openSync, closeSync, unlinkSync, statSync, writeFileSync } from "node:f
 import type { ExtensionContext } from "@gsd/pi-coding-agent";
 import { gsdRoot } from "./paths.js";
 import { getAndClearSkills } from "./skill-telemetry.js";
-import { loadJsonFile, loadJsonFileOrNull, saveJsonFile } from "./json-persistence.js";
+import { loadJsonFile, loadJsonFileOrNull, saveJsonFileCompact } from "./json-persistence.js";
 import { parseUnitId } from "./unit-id.js";
 import { buildAuditEnvelope, emitUokAuditEvent } from "./uok/audit.js";
 import { isUnifiedAuditEnabled } from "./uok/audit-toggle.js";
@@ -873,7 +873,7 @@ export function pruneMetricsLedger(base: string, keepCount: number): number {
   if (!disk || disk.units.length <= keepCount) return 0;
   const removed = disk.units.length - keepCount;
   disk.units = keepNewestUnits(disk.units, keepCount);
-  saveJsonFile(metricsPath(base), disk);
+  saveJsonFileCompact(metricsPath(base), disk);
   // Keep the in-memory ledger in sync if it is loaded for this session.
   if (ledger) {
     ledger.units = keepNewestUnits(ledger.units, keepCount);
@@ -1041,7 +1041,7 @@ function saveLedger(base: string, data: MetricsLedger): void {
           : dataUnits;
       merged.sort(compareUnitsByTime);
       data.units = keepNewestUnits(merged);
-      saveJsonFile(path, data);
+      saveJsonFileCompact(path, data);
     } finally {
       releaseLock(lockPath);
     }
@@ -1053,6 +1053,6 @@ function saveLedger(base: string, data: MetricsLedger): void {
     // read-merge-write sequence without mutual exclusion.
     logWarning("fs", "saveLedger: lock not acquired — falling back to direct write (no merge)");
     data.units = keepNewestUnits(data.units);
-    saveJsonFile(path, data);
+    saveJsonFileCompact(path, data);
   }
 }

package/src/resources/extensions/gsd/migration-auto-check.ts

@@ -90,6 +90,28 @@ function scanHasExtraIdentities(a: HierarchyScan, b: HierarchyScan): boolean {
   );
 }
 
+function paddedMilestoneId(id: string): string | null {
+  return /^\d+$/.test(id) ? `M${id.padStart(3, "0")}` : null;
+}
+
+function replaceSetPrefix(values: Set<string>, from: string, to: string): void {
+  for (const value of [...values]) {
+    if (value !== from && !value.startsWith(`${from}/`)) continue;
+    values.delete(value);
+    values.add(`${to}${value.slice(from.length)}`);
+  }
+}
+
+function alignNumericMarkdownIdsWithDb(markdownScan: HierarchyScan, dbScan: HierarchyScan): void {
+  for (const dbId of dbScan.milestones) {
+    const paddedId = paddedMilestoneId(dbId);
+    if (!paddedId || dbScan.milestones.has(paddedId) || !markdownScan.milestones.has(paddedId)) continue;
+    replaceSetPrefix(markdownScan.milestones, paddedId, dbId);
+    replaceSetPrefix(markdownScan.slices, paddedId, dbId);
+    replaceSetPrefix(markdownScan.tasks, paddedId, dbId);
+  }
+}
+
 /**
  * True when the DB holds any milestone/slice/task identity the markdown lacks —
  * i.e. a `/gsd recover --confirm` (markdown → DB) would DELETE authoritative DB
@@ -185,6 +207,7 @@ export async function checkMarkdownHierarchyAgainstDb(
 
   const dbScan = scanDbHierarchy();
   const beforeDb = dbScan.counts;
+  alignNumericMarkdownIdsWithDb(markdownScan, dbScan);
 
   // Discussion-phase scratch: a milestone dir with no ROADMAP and no DB row is
   // a pre-registration discussion artifact (CONTEXT/CONTEXT-DRAFT only — the

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

@@ -141,7 +141,8 @@ export function clearReservedMilestoneIds(): void {
 
 // ─── Discovery ──────────────────────────────────────────────────────────────
 
-function scanMilestoneIdsFromDir(dir: string): string[] {
+function scanMilestoneIdsFromDir(basePath: string, dir: string): string[] {
+  const legacyNumericIds = idsWithLegacyNumericDirs(basePath);
   return readdirSync(dir, { withFileTypes: true })
     .filter((d) => d.isDirectory())
     .map((d) => {
@@ -149,6 +150,12 @@ function scanMilestoneIdsFromDir(dir: string): string[] {
       if (MILESTONE_ID_RE.test(d.name)) {
         return d.name;
       }
+      // Legacy recovery/imports may contain bare numeric milestone directories
+      // (for example `milestones/15/15-ROADMAP.md`). The DB can preserve that
+      // id, and rebuild renders it, so drift scans must see it too.
+      if (/^\d+$/.test(d.name)) {
+        return d.name;
+      }
       // Legacy layout: M001-abcdef-slug descriptor directories
       const legacyMatch = d.name.match(/^(M\d{3}(?:-[a-z0-9]{6})?)-/);
       if (legacyMatch) {
@@ -162,6 +169,10 @@ function scanMilestoneIdsFromDir(dir: string): string[] {
         if (MILESTONE_ID_RE.test(fromSlug)) {
           return fromSlug;
         }
+        const numericId = String(phaseNum);
+        if (legacyNumericIds.has(numericId)) {
+          return numericId;
+        }
         return `M${String(phaseNum).padStart(3, "0")}`;
       }
       return null;
@@ -169,6 +180,24 @@ function scanMilestoneIdsFromDir(dir: string): string[] {
     .filter((id): id is string => id !== null);
 }
 
+function idsWithLegacyNumericDirs(basePath: string): Set<string> {
+  const legacyDir = join(gsdProjectionRoot(basePath), "milestones");
+  const ids = new Set<string>();
+  try {
+    for (const entry of readdirSync(legacyDir, { withFileTypes: true })) {
+      if (entry.isDirectory() && /^\d+$/.test(entry.name)) ids.add(entry.name);
+    }
+  } catch (err) {
+    // A missing legacy directory is the common, benign case (nothing to
+    // consult). Only surface a real failure when the directory exists but
+    // could not be read.
+    if (existsSync(legacyDir)) {
+      logWarning("engine", `idsWithLegacyNumericDirs: ${legacyDir} exists but readdirSync failed — ${getErrorMessage(err)}`);
+    }
+  }
+  return ids;
+}
+
 /** Scan the milestones directory and return IDs sorted by queue order (or numeric fallback). */
 export function findMilestoneIds(basePath: string): string[] {
   const root = gsdProjectionRoot(basePath);
@@ -181,7 +210,7 @@ export function findMilestoneIds(basePath: string): string[] {
   const ids = new Set<string>();
   for (const dir of dirs) {
     try {
-      for (const id of scanMilestoneIdsFromDir(dir)) ids.add(id);
+      for (const id of scanMilestoneIdsFromDir(basePath, dir)) ids.add(id);
     } catch (err) {
       if (existsSync(dir)) {
         logWarning("engine", `findMilestoneIds: ${dir} exists but readdirSync failed — ${getErrorMessage(err)}`);

package/src/resources/extensions/gsd/milestone-implementation-evidence.ts

@@ -19,6 +19,16 @@ import { resolveTasksDir } from "./paths.js";
 
 /** Large enough for unbounded milestone-history git log scans in big repos. */
 const GIT_LOG_MAX_BUFFER = 16 * 1024 * 1024;
+const LOG_FIELD_SEPARATOR = "\x1f";
+const LOG_RECORD_SEPARATOR = "\x1e";
+
+type CommitRecord = {
+  hash: string;
+  parents: string;
+  committedAt: string;
+  message: string;
+  files: string[];
+};
 
 /**
  * Check whether a milestone produced implementation artifacts (non-`.gsd/`
@@ -198,7 +208,7 @@ function getChangedFilesFromMilestoneTaggedCommits(
   // Primary: path-scoped log against .gsd/milestones/<id>. Fast and unbounded
   // by depth when .gsd/ is tracked in git.
   const scoped = scanGsdTaggedCommits(basePath, milestoneId, [
-    "log", "--format=%H%x1f%B%x1e", "HEAD", "--", `.gsd/milestones/${milestoneId}`,
+    "log", "--full-diff", "--name-only", "--format=%x1e%H%x1f%B%x1f", "HEAD", "--", `.gsd/milestones/${milestoneId}`,
   ]);
   if (!scoped.ok) return scoped;
   if (scoped.matched && classifyImplementationFiles(scoped.files) === "present") return scoped;
@@ -213,7 +223,7 @@ function getChangedFilesFromMilestoneTaggedCommits(
   // reintroducing the rolling-depth failure class removed in #4699 where
   // milestone evidence aged out behind unrelated activity.
   const unscoped = scanGsdTaggedCommits(basePath, milestoneId, [
-    "log", "--format=%H%x1f%B%x1e", "HEAD",
+    "log", "--name-only", "--format=%x1e%H%x1f%B%x1f", "HEAD",
   ]);
   if (!unscoped.ok) return scoped.matched ? scoped : unscoped;
   if (!unscoped.matched) return scoped;
@@ -299,8 +309,7 @@ function backfillChangedFilesFromUntaggedMilestoneCommits(
       if (record.parents.trim().split(/\s+/).filter(Boolean).length > 1) continue;
       if (commitMessageHasGsdTrailer(record.message)) continue;
 
-      const commitFiles = getChangedFilesForCommit(basePath, record.hash);
-      const implementationFiles = commitFiles.map(normalizeRepoPath).filter(isImplementationPath);
+      const implementationFiles = record.files.map(normalizeRepoPath).filter(isImplementationPath);
       if (implementationFiles.length === 0) continue;
       if (!implementationFiles.some((file) => hintSet.has(file))) continue;
 
@@ -323,25 +332,30 @@ function backfillChangedFilesFromUntaggedMilestoneCommits(
   }
 }
 
-function getCommitRecords(basePath: string): Array<{ hash: string; parents: string; committedAt: string; message: string }> {
-  const logOutput = execFileSync("git", ["log", "--format=%H%x1f%P%x1f%cI%x1f%B%x1e", "HEAD"], {
+function getCommitRecords(basePath: string): CommitRecord[] {
+  const logOutput = execFileSync("git", ["log", "--name-only", "--format=%x1e%H%x1f%P%x1f%cI%x1f%B%x1f", "HEAD"], {
     cwd: basePath,
     stdio: ["ignore", "pipe", "pipe"],
     encoding: "utf-8",
     maxBuffer: GIT_LOG_MAX_BUFFER,
   });
   return logOutput
-    .split("\x1e")
-    .map((record) => record.trim())
+    .split(LOG_RECORD_SEPARATOR)
     .filter(Boolean)
     .flatMap((record) => {
-      const parts = record.split("\x1f");
-      if (parts.length < 4) return [];
-      const [hash, parents, committedAt, ...messageParts] = parts;
-      return [{ hash: hash.trim(), parents: parents.trim(), committedAt: committedAt.trim(), message: messageParts.join("\x1f") }];
+      const parts = record.split(LOG_FIELD_SEPARATOR);
+      if (parts.length < 5) return [];
+      const [hash, parents, committedAt] = parts;
+      const files = parseNameOnlyFiles(parts.at(-1) ?? "");
+      const message = parts.slice(3, -1).join(LOG_FIELD_SEPARATOR);
+      return [{ hash: hash.trim(), parents: parents.trim(), committedAt: committedAt.trim(), message, files }];
     });
 }
 
+function parseNameOnlyFiles(rawFiles: string): string[] {
+  return rawFiles.split(/\r?\n/).map((file) => file.trim()).filter(Boolean);
+}
+
 function isFullCommitSha(value: string): boolean {
   return /^[0-9a-f]{40}$/i.test(value);
 }
@@ -359,23 +373,23 @@ function scanGsdTaggedCommits(
       maxBuffer: GIT_LOG_MAX_BUFFER,
     });
     const records = logOutput
-      .split("\x1e")
-      .map((record) => record.trim())
+      .split(LOG_RECORD_SEPARATOR)
       .filter(Boolean)
       .flatMap((record) => {
-        const sep = record.indexOf("\x1f");
-        if (sep === -1) return [];
-        const hash = record.slice(0, sep).trim();
-        const message = record.slice(sep + 1);
-        return [{ hash, message }];
+        const parts = record.split(LOG_FIELD_SEPARATOR);
+        if (parts.length < 3) return [];
+        const hash = parts[0].trim();
+        if (!hash) return [];
+        const files = parseNameOnlyFiles(parts.at(-1) ?? "");
+        const message = parts.slice(1, -1).join(LOG_FIELD_SEPARATOR);
+        return [{ message, files }];
       });
 
     const files = new Set<string>();
     let matched = false;
-    for (const { hash, message } of records) {
+    for (const { message, files: commitFiles } of records) {
       if (!commitMessageHasGsdTrailer(message)) continue;
 
-      const commitFiles = getChangedFilesForCommit(basePath, hash);
       if (!commitMatchesMilestone(basePath, message, milestoneId, commitFiles)) continue;
 
       matched = true;

package/src/resources/extensions/gsd/orphan-stash-audit.ts

@@ -33,16 +33,87 @@ function _isAlreadyRestoredApplyError(err: unknown): boolean {
 
 export { _isAlreadyRestoredApplyError };
 
+function gitOutput(basePath: string, args: string[]): string {
+  return execFileSync("git", args, {
+    cwd: basePath,
+    stdio: ["ignore", "pipe", "pipe"],
+    encoding: "utf-8",
+    env: GIT_NO_PROMPT_ENV,
+  });
+}
+
+function listStashUntrackedPaths(basePath: string, stashRef: string): string[] | null {
+  try {
+    return gitOutput(basePath, ["ls-tree", "-r", "-z", "--name-only", `${stashRef}^3`]).split("\0").filter(Boolean);
+  } catch {
+    return null;
+  }
+}
+
+function listStashTrackedPaths(basePath: string, stashRef: string): string[] | null {
+  try {
+    return gitOutput(basePath, ["diff", "--name-only", "-z", `${stashRef}^1`, stashRef]).split("\0").filter(Boolean);
+  } catch {
+    return null;
+  }
+}
+
+interface PorcelainEntry {
+  code: string;
+  path: string;
+}
+
+function readWorkingTreeStatus(basePath: string): PorcelainEntry[] | null {
+  try {
+    return gitOutput(basePath, ["status", "--porcelain", "-z", "--untracked-files=all"])
+      .split("\0")
+      .filter(Boolean)
+      .map((entry) => ({ code: entry.slice(0, 2), path: entry.slice(3) }));
+  } catch {
+    return null;
+  }
+}
+
+function isAlreadyRestoredUntrackedStatus(basePath: string, stashRef: string, statusEntries: PorcelainEntry[]): boolean {
+  if (statusEntries.length === 0 || statusEntries.some((entry) => entry.code !== "??")) return false;
+  const stashTrackedPaths = listStashTrackedPaths(basePath, stashRef);
+  if (!stashTrackedPaths || stashTrackedPaths.length > 0) return false;
+  const stashUntrackedPaths = listStashUntrackedPaths(basePath, stashRef);
+  if (!stashUntrackedPaths || stashUntrackedPaths.length === 0) return false;
+  const stashUntrackedPathSet = new Set(stashUntrackedPaths);
+  const workingTreeUntrackedPathSet = new Set(statusEntries.map((entry) => entry.path));
+  // "Already restored" requires the working tree's untracked paths to match the
+  // stash's untracked paths exactly, in BOTH directions:
+  //   1. every current untracked path comes from the stash (no unexplained user
+  //      work that an apply would clobber), AND
+  //   2. every stash untracked path is already present in the working tree.
+  // Without (2), a partial restore (where only some stash files were recovered)
+  // would be misread as fully restored: the audit would no-op without applying
+  // or warning, leaving the remaining pre-merge files missing.
+  return (
+    statusEntries.every((entry) => stashUntrackedPathSet.has(entry.path)) &&
+    stashUntrackedPaths.every((path) => workingTreeUntrackedPathSet.has(path))
+  );
+}
+
+function manualApplyWarning(stashRef: string, milestoneId: string, reason: string): string {
+  return (
+    `Pre-merge stash ${stashRef} for milestone ${milestoneId} not auto-applied: ${reason}; ` +
+    `run \`git stash apply ${stashRef}\` manually to restore your pre-merge changes.`
+  );
+}
+
 /**
  * Audit `git stash list` for orphaned `gsd-preflight-stash:M00x:*` entries.
  *
  * The matching merge code in `phases.ts` previously skipped the postflight
  * pop whenever `mergeAndExit` threw, leaking the user's pre-merge working
- * tree into the stash list. For every preflight-stash entry whose milestone
- * is now complete (per the supplied callback), `git stash apply` is invoked —
+ * tree into the stash list. Completed-milestone stashes are only auto-applied
+ * when the working tree is clean and the stash does not modify tracked files.
+ * Dirty trees and tracked-file stashes are left untouched with a manual
+ * recovery warning. When auto-apply is safe, `git stash apply` is invoked —
  * NOT `pop`. The stash entry stays in the list so the user retains a backup
- * if the apply produces unexpected merge results. Idempotent across repeated
- * startup runs.
+ * if the apply produces unexpected merge results.
  *
  * Failures are best-effort: a list error (no repo, git unavailable) returns
  * an empty result. An apply error becomes a warning the user sees alongside
@@ -56,16 +127,7 @@ export function auditOrphanedPreflightStashes(
 
   let listOutput: string;
   try {
-    listOutput = execFileSync(
-      "git",
-      ["stash", "list", "--format=%gd%x00%s"],
-      {
-        cwd: basePath,
-        stdio: ["ignore", "pipe", "pipe"],
-        encoding: "utf-8",
-        env: GIT_NO_PROMPT_ENV,
-      },
-    );
+    listOutput = gitOutput(basePath, ["stash", "list", "--format=%gd%x00%s"]);
   } catch {
     return result;
   }
@@ -93,13 +155,39 @@ export function auditOrphanedPreflightStashes(
     }
     if (!complete) continue;
 
+    const statusEntries = readWorkingTreeStatus(basePath);
+    if (!statusEntries) {
+      result.warnings.push(
+        manualApplyWarning(ref, milestoneId, "could not verify working tree status"),
+      );
+      continue;
+    }
+    if (statusEntries.length > 0) {
+      if (isAlreadyRestoredUntrackedStatus(basePath, ref, statusEntries)) continue;
+      result.warnings.push(manualApplyWarning(ref, milestoneId, "working tree not clean"));
+      continue;
+    }
+
+    const trackedPaths = listStashTrackedPaths(basePath, ref);
+    if (trackedPaths === null) {
+      result.warnings.push(
+        manualApplyWarning(ref, milestoneId, "could not inspect tracked paths in stash"),
+      );
+      continue;
+    }
+    if (trackedPaths.length > 0) {
+      // A retained tracked-file stash has no durable marker showing whether a
+      // clean tree needs first recovery or would be re-mutated by an old backup.
+      const preview = trackedPaths.slice(0, 5).join(", ");
+      const suffix = trackedPaths.length > 5 ? `, and ${trackedPaths.length - 5} more` : "";
+      result.warnings.push(
+        manualApplyWarning(ref, milestoneId, `stash modifies tracked files (${preview}${suffix})`),
+      );
+      continue;
+    }
+
     try {
-      execFileSync("git", ["stash", "apply", "--quiet", ref], {
-        cwd: basePath,
-        stdio: ["ignore", "pipe", "pipe"],
-        encoding: "utf-8",
-        env: GIT_NO_PROMPT_ENV,
-      });
+      gitOutput(basePath, ["stash", "apply", "--quiet", ref]);
       result.applied.push({ milestoneId, stashRef: ref });
     } catch (err) {
       // Idempotent steady state: stash was already applied in a prior audit

package/src/resources/extensions/gsd/prompts/code-review.md

@@ -16,18 +16,20 @@ You are running the GSD **code-review** workflow — review source files changed
 
 1. **Determine the file set.** Use the explicit `--files` list if provided. Otherwise derive the changed source files from the active slice's recent commits / SUMMARY, or fall back to the git diff against the base branch. Exclude `.gsd/`, lockfiles, generated/vendored code, and docs unless `--files` names them.
 
-2. **Review each file** at the requested depth:
+2. **Gather diff-first context.** Run `git diff` for the file set (or the relevant commit range) and use that as the primary review evidence. Use `read` only when the diff is insufficient — new/untracked files, truncated hunks, or missing surrounding context needed for a finding. Do **not** read every changed file when the diff already covers the edits. If the Tool Call Loop Guard blocks a tool, stop calling tools for the rest of this turn and continue in text.
+
+3. **Review at the requested depth:**
    - **quick** — bugs, security, and correctness only.
    - **standard** (default) — quick + maintainability, error handling, edge cases.
    - **deep** — standard + performance, concurrency, API design, test coverage gaps.
 
    Ground every finding in the actual code with file path + line reference. Categorize each by severity (critical / warning / nit) and type (bug, security, quality, performance).
 
-3. **Write the review.** Produce `.gsd/reviews/{{reviewId}}-REVIEW.md` with the findings table and per-finding detail (location, issue, suggested fix).
+4. **Write the review.** Produce `.gsd/reviews/{{reviewId}}-REVIEW.md` with the findings table and per-finding detail (location, issue, suggested fix).
 
-4. **Present results.** Summarize counts by severity and list the critical findings first.
+5. **Present results.** Summarize counts by severity and list the critical findings first.
 
-5. **Fix mode** (only when `--fix` is set): for each finding that is safely auto-fixable (deterministic, no behavior change beyond the fix), apply it. After each fix batch, run the project's tests. Commit atomically with the finding IDs referenced in the message. Do not auto-apply findings that are ambiguous or that change public behavior — surface those for a human decision instead.
+6. **Fix mode** (only when `--fix` is set): for each finding that is safely auto-fixable (deterministic, no behavior change beyond the fix), apply it. After each fix batch, run the project's tests. Commit atomically with the finding IDs referenced in the message. Do not auto-apply findings that are ambiguous or that change public behavior — surface those for a human decision instead.
 
 ## Success criteria