@opengsd/gsd-pi 1.3.0 → 1.4.0-dev.5fc22c6f

package/src/resources/extensions/gsd/tests/guided-flow-session-isolation.test.ts

@@ -20,6 +20,7 @@ import {
   clearPendingAutoStart,
   checkAutoStartAfterDiscuss,
 } from "../guided-flow.ts";
+import { closeDatabase, openDatabase } from "../gsd-db.ts";
 
 function pendingInput(basePath: string, milestoneId: string) {
   return {
@@ -34,6 +35,7 @@ function pendingInput(basePath: string, milestoneId: string) {
 
 describe("#2985 Bug 3 — concurrent discuss sessions must be independent", () => {
   beforeEach(() => {
+    closeDatabase();
     clearPendingAutoStart();
   });
 
@@ -75,6 +77,7 @@ describe("#2985 Bug 3 — concurrent discuss sessions must be independent", () =
 
 describe("#2985 Bug 4 — getDiscussionMilestoneId must be keyed by basePath", () => {
   beforeEach(() => {
+    closeDatabase();
     clearPendingAutoStart();
   });
 
@@ -110,6 +113,7 @@ describe("#2985 Bug 4 — getDiscussionMilestoneId must be keyed by basePath", (
 test("checkAutoStartAfterDiscuss ignores missing manifest for single-milestone discuss on established project", () => {
   const base = mkdtempSync(join(tmpdir(), "gsd-auto-start-manifest-"));
   try {
+    openDatabase(":memory:");
     const gsdDir = join(base, ".gsd");
     const milestoneDir = join(gsdDir, "milestones", "M001");
     mkdirSync(milestoneDir, { recursive: true });
@@ -132,6 +136,7 @@ test("checkAutoStartAfterDiscuss ignores missing manifest for single-milestone d
     const started = checkAutoStartAfterDiscuss();
     assert.equal(started, true, "project history alone should not require a manifest");
   } finally {
+    closeDatabase();
     clearPendingAutoStart();
     rmSync(base, { recursive: true, force: true });
   }
@@ -151,6 +156,7 @@ test("checkAutoStartAfterDiscuss(basePath) selects the matching pending entry wh
   }
 
   try {
+    openDatabase(":memory:");
     clearPendingAutoStart();
     writeReadyArtifacts(projectA, "M001");
     writeReadyArtifacts(projectB, "M002");
@@ -172,6 +178,7 @@ test("checkAutoStartAfterDiscuss(basePath) selects the matching pending entry wh
     assert.equal(getDiscussionMilestoneId(projectA), "M001", "projectA should remain pending");
     assert.equal(getDiscussionMilestoneId(projectB), null, "projectB should be cleared after start");
   } finally {
+    closeDatabase();
     clearPendingAutoStart();
     rmSync(projectA, { recursive: true, force: true });
     rmSync(projectB, { recursive: true, force: true });
@@ -183,6 +190,7 @@ test("checkAutoStartAfterDiscuss can accept context handoff without scheduling a
   let waitForIdleCalls = 0;
   const notifications: string[] = [];
   try {
+    openDatabase(":memory:");
     const gsdDir = join(base, ".gsd");
     const milestoneDir = join(gsdDir, "milestones", "M001");
     mkdirSync(milestoneDir, { recursive: true });
@@ -215,6 +223,7 @@ test("checkAutoStartAfterDiscuss can accept context handoff without scheduling a
     assert.equal(waitForIdleCalls, 0, "headless-owned auto start must not schedule guided-flow auto");
     assert.equal(getDiscussionMilestoneId(base), null, "accepted handoff should still be cleared");
   } finally {
+    closeDatabase();
     clearPendingAutoStart();
     rmSync(base, { recursive: true, force: true });
   }

package/src/resources/extensions/gsd/tests/health-widget.test.ts

@@ -3,8 +3,10 @@
 
 import test from "node:test";
 import assert from "node:assert/strict";
-import { mkdirSync, rmSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
+import { execFileSync } from "node:child_process";
+import { chmodSync, mkdirSync, rmSync, writeFileSync } from "node:fs";
+import { performance } from "node:perf_hooks";
+import { delimiter, join } from "node:path";
 import { tmpdir } from "node:os";
 import {
   buildHealthLines,
@@ -12,8 +14,9 @@ import {
   formatRelativeTime,
   type HealthWidgetData,
 } from "../health-widget-core.ts";
-import { HEALTH_WIDGET_ACTIVE_HINTS } from "../health-widget.ts";
+import { HEALTH_WIDGET_ACTIVE_HINTS, getCachedProjectState, initHealthWidget } from "../health-widget.ts";
 import { registerHooks } from "../bootstrap/register-hooks.ts";
+import { GIT_NO_PROMPT_ENV } from "../git-constants.ts";
 
 function makeTempDir(prefix: string): string {
   const dir = join(
@@ -32,6 +35,58 @@ function cleanup(dir: string): void {
   }
 }
 
+function runGit(cwd: string, ...args: string[]): string {
+  return execFileSync("git", args, {
+    cwd,
+    encoding: "utf-8",
+    stdio: ["ignore", "pipe", "pipe"],
+  }).trim();
+}
+
+function makeTempRepo(prefix: string): string {
+  const dir = makeTempDir(prefix);
+  runGit(dir, "init");
+  runGit(dir, "config", "user.email", "test@test.com");
+  runGit(dir, "config", "user.name", "Test");
+  writeFileSync(join(dir, "README.md"), "# test\n", "utf-8");
+  runGit(dir, "add", "README.md");
+  runGit(dir, "commit", "-m", "initial commit");
+  return dir;
+}
+
+function installSlowGitLogShim(binDir: string): void {
+  writeFileSync(
+    join(binDir, "git"),
+    [
+      "#!/bin/sh",
+      'if [ "$1" = "log" ]; then sleep 1; fi',
+      'PATH="$GSD_REAL_PATH"',
+      "export PATH",
+      'exec git "$@"',
+      "",
+    ].join("\n"),
+    "utf-8",
+  );
+  chmodSync(join(binDir, "git"), 0o755);
+
+  writeFileSync(
+    join(binDir, "git.cmd"),
+    [
+      "@echo off",
+      'if "%1"=="log" powershell -NoProfile -Command "Start-Sleep -Seconds 1"',
+      'set "PATH=%GSD_REAL_PATH%"',
+      "git %*",
+      "",
+    ].join("\r\n"),
+    "utf-8",
+  );
+}
+
+type HealthWidgetFactory = (
+  tui: { requestRender(): void },
+  theme: { fg(style: string, text: string): string },
+) => { dispose(): void };
+
 function activeData(overrides: Partial<HealthWidgetData> = {}): HealthWidgetData {
   return {
     projectState: "active",
@@ -70,6 +125,86 @@ test("detectHealthWidgetProjectState: milestone without metrics returns active",
   assert.equal(detectHealthWidgetProjectState(dir), "active");
 });
 
+test("getCachedProjectState: reuses project state until the refresh TTL expires", (t) => {
+  const dir = makeTempDir("cached-state");
+  t.after(() => { cleanup(dir); });
+
+  let now = 1_000_000;
+  const dateNow = t.mock.method(Date, "now", () => now);
+  t.after(() => { dateNow.mock.restore(); });
+
+  mkdirSync(join(dir, ".gsd"), { recursive: true });
+  assert.equal(getCachedProjectState(dir), "initialized");
+
+  mkdirSync(join(dir, ".gsd", "milestones", "M001"), { recursive: true });
+  assert.equal(getCachedProjectState(dir), "initialized");
+
+  now += 60_000;
+  assert.equal(getCachedProjectState(dir), "initialized");
+
+  now += 1;
+  assert.equal(getCachedProjectState(dir), "active");
+});
+
+test("getCachedProjectState: force=true bypasses TTL and returns fresh state within TTL window", (t) => {
+  const dir = makeTempDir("forced-state");
+  t.after(() => { cleanup(dir); });
+
+  let now = 2_000_000;
+  const dateNow = t.mock.method(Date, "now", () => now);
+  t.after(() => { dateNow.mock.restore(); });
+
+  mkdirSync(join(dir, ".gsd"), { recursive: true });
+  // Prime the cache with "initialized".
+  assert.equal(getCachedProjectState(dir), "initialized");
+
+  // Disk changes within the TTL window.
+  mkdirSync(join(dir, ".gsd", "milestones", "M001"), { recursive: true });
+  now += 1_000; // well within 60s TTL
+
+  // Normal call still returns stale cached value.
+  assert.equal(getCachedProjectState(dir), "initialized");
+
+  // force=true bypasses TTL and returns the fresh disk state.
+  assert.equal(getCachedProjectState(dir, true), "active");
+
+  // Subsequent non-forced call also reflects the freshened cache.
+  assert.equal(getCachedProjectState(dir), "active");
+});
+
+test("initHealthWidget: re-init paints fresh project state within cache TTL", (t) => {
+  const dir = makeTempDir("reinit-state");
+  t.after(() => { cleanup(dir); });
+
+  let now = 3_000_000;
+  const dateNow = t.mock.method(Date, "now", () => now);
+  t.after(() => { dateNow.mock.restore(); });
+
+  const originalCwd = process.cwd();
+  process.chdir(dir);
+  t.after(() => { process.chdir(originalCwd); });
+
+  const initialLineSets: string[][] = [];
+  const ctx = {
+    hasUI: true,
+    ui: {
+      setWidget: (_key: string, value: unknown) => {
+        if (Array.isArray(value)) initialLineSets.push(value as string[]);
+      },
+    },
+  } as any;
+
+  mkdirSync(join(dir, ".gsd"), { recursive: true });
+  initHealthWidget(ctx);
+  assert.equal(initialLineSets.at(-1)?.[0], "  GSD  Project Initialized");
+
+  mkdirSync(join(dir, ".gsd", "milestones", "M001"), { recursive: true });
+  now += 1_000;
+
+  initHealthWidget(ctx);
+  assert.match(initialLineSets.at(-1)?.[0] ?? "", /System OK/);
+});
+
 test("buildHealthLines: none state shows single onboarding line pointing at /gsd", (t) => {
   const lines = buildHealthLines(activeData({ projectState: "none" }));
   assert.equal(lines.length, 1, "renders exactly one line");
@@ -101,6 +236,136 @@ test("health widget active hints include visualization and notifications", () =>
   assert.match(HEALTH_WIDGET_ACTIVE_HINTS, /\/gsd help/);
 });
 
+test("health widget async refresh does not block timers while git log is slow", async (t) => {
+  const dir = makeTempRepo("slow-git-log");
+  const binDir = makeTempDir("slow-git-log-bin");
+  mkdirSync(join(dir, ".gsd", "milestones", "M001"), { recursive: true });
+  installSlowGitLogShim(binDir);
+
+  const originalCwd = process.cwd();
+  const originalProcessPath = process.env.PATH;
+  const originalEnvPath = GIT_NO_PROMPT_ENV.PATH;
+  const originalEnvRealPath = GIT_NO_PROMPT_ENV.GSD_REAL_PATH;
+  const shimmedPath = `${binDir}${delimiter}${originalProcessPath ?? ""}`;
+
+  process.chdir(dir);
+  process.env.PATH = shimmedPath;
+  GIT_NO_PROMPT_ENV.PATH = shimmedPath;
+  GIT_NO_PROMPT_ENV.GSD_REAL_PATH = originalProcessPath ?? "";
+
+  let factory: HealthWidgetFactory | null = null;
+  let resolveRefresh: (() => void) | undefined;
+  const refreshed = new Promise<void>((resolve) => { resolveRefresh = resolve; });
+  const gaps: number[] = [];
+  let lastTick = performance.now();
+  let heartbeat: NodeJS.Timeout | undefined;
+  let refreshTimeout: NodeJS.Timeout | undefined;
+  let widget: { dispose(): void } | undefined;
+
+  t.after(() => {
+    if (widget) widget.dispose();
+    if (heartbeat) clearInterval(heartbeat);
+    if (refreshTimeout) clearTimeout(refreshTimeout);
+    process.chdir(originalCwd);
+    if (originalProcessPath === undefined) delete process.env.PATH;
+    else process.env.PATH = originalProcessPath;
+    if (originalEnvPath === undefined) delete GIT_NO_PROMPT_ENV.PATH;
+    else GIT_NO_PROMPT_ENV.PATH = originalEnvPath;
+    if (originalEnvRealPath === undefined) delete GIT_NO_PROMPT_ENV.GSD_REAL_PATH;
+    else GIT_NO_PROMPT_ENV.GSD_REAL_PATH = originalEnvRealPath;
+    cleanup(binDir);
+    cleanup(dir);
+  });
+
+  initHealthWidget({
+    hasUI: true,
+    ui: {
+      setWidget: (_key: string, value: unknown) => {
+        if (typeof value === "function") factory = value as HealthWidgetFactory;
+      },
+    },
+  } as any);
+
+  assert.ok(factory, "health widget factory is registered");
+
+  heartbeat = setInterval(() => {
+    const now = performance.now();
+    gaps.push(now - lastTick);
+    lastTick = now;
+  }, 25);
+
+  // assert.ok above guards at runtime; double-cast is needed because TypeScript
+  // cannot track the factory assignment through the `as any` closure call.
+  widget = (factory as unknown as HealthWidgetFactory)(
+    { requestRender: () => { resolveRefresh?.(); } },
+    { fg: (_style: string, text: string) => text },
+  );
+
+  await Promise.race([
+    refreshed,
+    new Promise<never>((_, reject) => {
+      refreshTimeout = setTimeout(() => reject(new Error("health widget refresh did not complete")), 4_000);
+    }),
+  ]);
+  if (refreshTimeout) clearTimeout(refreshTimeout);
+
+  assert.ok(gaps.length > 0, "heartbeat ran while refresh was in flight");
+  const maxGap = Math.max(...gaps);
+  assert.ok(maxGap < 750, `slow git log must not starve timers; max gap was ${Math.round(maxGap)}ms`);
+});
+
+test("initHealthWidget: synchronous first-paint render never contains last-commit info (regression #964)", (t) => {
+  // Before the fix, loadHealthWidgetData with includeChecks:true called
+  // loadLastCommitInfo — synchronous native-git-bridge ops (nativeIsRepo,
+  // nativeGetCurrentBranch, nativeLastCommitEpoch, nativeCommitSubject) — which
+  // froze the TUI on slow repos. The fix removes that synchronous git path
+  // entirely: lastCommitEpoch/lastCommitMessage are now always null from the
+  // synchronous loader; only the async refresh (loadLastCommitInfoAsync) fills
+  // them in. This test guards that contract by verifying that the initial
+  // string-array setWidget call never contains "Last commit:" even on a real
+  // git repo where native git queries would succeed.
+  const dir = makeTempRepo("sync-last-commit-regression");
+  mkdirSync(join(dir, ".gsd", "milestones", "M001"), { recursive: true });
+
+  const originalCwd = process.cwd();
+  process.chdir(dir);
+
+  let widget: { dispose(): void } | undefined;
+  t.after(() => {
+    if (widget) widget.dispose();
+    process.chdir(originalCwd);
+    cleanup(dir);
+  });
+
+  const initialRenders: string[][] = [];
+
+  initHealthWidget({
+    hasUI: true,
+    ui: {
+      setWidget: (_key: string, value: unknown) => {
+        if (Array.isArray(value)) {
+          initialRenders.push(value as string[]);
+        } else if (typeof value === "function") {
+          // Instantiate the factory to satisfy dispose(), but do not await the
+          // async refresh — we are only inspecting the synchronous first-paint.
+          widget = (value as unknown as HealthWidgetFactory)(
+            { requestRender: () => {} },
+            { fg: (_style: string, text: string) => text },
+          );
+        }
+      },
+    },
+  } as any);
+
+  assert.ok(initialRenders.length > 0, "at least one synchronous setWidget call");
+
+  const combined = initialRenders.flat().join("\n");
+  assert.ok(
+    !combined.includes("Last commit:"),
+    "synchronous first-paint render must not contain 'Last commit:' — sync git path removed (regression #964)",
+  );
+});
+
 test("buildHealthLines: active state with budget ceiling shows percent summary", (t) => {
   const lines = buildHealthLines(activeData({ budgetSpent: 2.5, budgetCeiling: 10 }));
   assert.equal(lines.length, 1);

package/src/resources/extensions/gsd/tests/integration/auto-recovery.test.ts

@@ -560,7 +560,12 @@ test("verifyExpectedArtifact execute-task rejects heading-style plan without che
 });
 
 test("verifyExpectedArtifact plan-slice passes for rendered slice/task plan artifacts from DB", async () => {
-  const base = makeTmpBase();
+  // Do NOT use makeTmpBase() here — it creates a legacy milestones/ dir which forces
+  // renderPlanFromDb into the legacy path layout and creates a spurious tasks/ dir that
+  // triggers the per-task file check even when dbPrimary=true. Create only .gsd/ so
+  // resolveMilestonePath returns null and the flat-phase path is used instead.
+  const base = join(tmpdir(), `gsd-test-${randomUUID()}`);
+  mkdirSync(join(base, ".gsd"), { recursive: true });
   const dbPath = join(base, ".gsd", "gsd.db");
   openDatabase(dbPath);
   try {
@@ -614,26 +619,28 @@ test("verifyExpectedArtifact plan-slice passes for rendered slice/task plan arti
 
     const rendered = await renderPlanFromDb(base, "M001", "S01");
     assert.ok(existsSync(rendered.planPath), "renderPlanFromDb should write the slice plan");
-    assert.equal(rendered.taskPlanPaths.length, 2, "renderPlanFromDb should render one task plan per task");
+    // Flat-phase: tasks are checkboxes inside the slice plan — no per-task plan files.
+    assert.equal(rendered.taskPlanPaths.length, 0, "flat-phase: renderPlanFromDb returns no per-task plan paths");
 
     const planContent = readFileSync(rendered.planPath, "utf-8");
     const parsedPlan = parsePlan(planContent);
     assert.equal(parsedPlan.tasks.length, 2, "rendered slice plan should parse into task entries");
 
-    const taskPlanContent = readFileSync(rendered.taskPlanPaths[0], "utf-8");
-    const taskPlan = parseTaskPlanFile(taskPlanContent);
-    assert.deepEqual(taskPlan.frontmatter.skills_used, [], "rendered task plans should use conservative empty skills_used");
-
+    // DB-primary: task rows prove the slice was planned; no per-task file check.
     const result = verifyExpectedArtifact("plan-slice", "M001/S01", base);
-    assert.equal(result, true, "plan-slice verification should pass when rendered task plan files exist");
+    assert.equal(result, true, "plan-slice verification should pass when DB has task rows");
   } finally {
     closeDatabase();
     cleanup(base);
   }
 });
 
-test("verifyExpectedArtifact plan-slice fails after deleting a rendered task plan file", async () => {
-  const base = makeTmpBase();
+test("verifyExpectedArtifact plan-slice fails after deleting the slice plan file", async () => {
+  // Do NOT use makeTmpBase() here — same reason as the passing test above: the legacy
+  // milestones/ dir forces a legacy path and creates a tasks/ dir that incorrectly
+  // triggers the per-task file check. Create only .gsd/ to stay on the flat-phase path.
+  const base = join(tmpdir(), `gsd-test-${randomUUID()}`);
+  mkdirSync(join(base, ".gsd"), { recursive: true });
   const dbPath = join(base, ".gsd", "gsd.db");
   openDatabase(dbPath);
   try {
@@ -686,10 +693,11 @@ test("verifyExpectedArtifact plan-slice fails after deleting a rendered task pla
     });
 
     const rendered = await renderPlanFromDb(base, "M001", "S01");
-    rmSync(rendered.taskPlanPaths[1]);
+    // Flat-phase: per-task plan files do not exist; delete the slice plan itself to fail verification.
+    rmSync(rendered.planPath);
 
     const result = verifyExpectedArtifact("plan-slice", "M001/S01", base);
-    assert.equal(result, false, "plan-slice verification should fail when a rendered task plan file is removed");
+    assert.equal(result, false, "plan-slice verification should fail when the slice plan file is removed");
   } finally {
     closeDatabase();
     cleanup(base);

package/src/resources/extensions/gsd/tests/integration/commands-eval-review.integration.test.ts

@@ -139,8 +139,12 @@ describe("integration: /gsd eval-review helper chain on a real on-disk slice", (
     assert.ok(prompt.includes("src/llm/budget.ts:42"));
     assert.ok(prompt.includes("langfuse"));
 
-    // Output path is the canonical slice file path
-    assert.ok(prompt.includes(`${layout.sliceId}-EVAL-REVIEW.md`));
+    // Output path is the canonical slice file path — buildSliceFileName("S07", ...)
+    // produces "07-EVAL-REVIEW.md" (flat-phase MM-SUFFIX.md format).
+    assert.ok(
+      prompt.includes("07-EVAL-REVIEW.md") || prompt.includes(`${layout.sliceId}-EVAL-REVIEW.md`),
+      `prompt should include eval-review output filename (got: ${ctx.outputPath})`,
+    );
   });
 
   it("falls back to the no-spec audit mode when AI-SPEC.md is absent", async () => {

package/src/resources/extensions/gsd/tests/integration/git-service.test.ts

@@ -3,7 +3,7 @@
 import { describe, test } from 'node:test';
 import assert from 'node:assert/strict';
 import { mkdtempSync, mkdirSync, writeFileSync, rmSync, existsSync, symlinkSync, readFileSync, chmodSync } from "node:fs";
-import { join, dirname } from "node:path";
+import { join, dirname, delimiter } from "node:path";
 import { tmpdir } from "node:os";
 import { execFileSync, execSync } from "node:child_process";
 
@@ -23,6 +23,7 @@ import {
   type PreMergeCheckResult,
   type TaskCommitContext,
 } from "../../git-service.ts";
+import { GIT_NO_PROMPT_ENV } from "../../git-constants.ts";
 import { nativeAddAllWithExclusions, nativeHasChanges, _resetHasChangesCache } from "../../native-git-bridge.ts";
 function run(command: string, cwd: string): string {
   return execSync(command, { cwd, stdio: ["ignore", "pipe", "pipe"], encoding: "utf-8" }).trim();
@@ -293,12 +294,13 @@ describe('git-service', async () => {
 
   assert.deepStrictEqual(
     RUNTIME_EXCLUSION_PATHS.length,
-    17,
-    "exactly 17 runtime exclusion paths"
+    18,
+    "exactly 18 runtime exclusion paths"
   );
 
   const expectedPaths = [
     ".gsd-worktrees/",
+    ".gsd-backups/**",
     ".gsd/activity/",
     ".gsd/audit/",
     ".gsd/forensics/",
@@ -405,6 +407,49 @@ describe('git-service', async () => {
     }).trim();
   }
 
+  function findGitExecutable(): string {
+    const command = process.platform === "win32" ? "where git" : "command -v git";
+    const output = execSync(command, { encoding: "utf-8" });
+    const found = output.split(/\r?\n/).map(line => line.trim()).find(Boolean);
+    assert.ok(found, "git executable is available");
+    return found;
+  }
+
+  function installGitRmCachedCounterShim(shimDir: string): void {
+    const shimScript = join(shimDir, "git-shim.cjs");
+    writeFileSync(shimScript, `
+const { appendFileSync } = require("node:fs");
+const { spawnSync } = require("node:child_process");
+
+const args = process.argv.slice(2);
+if (args[0] === "rm" && args[1] === "--cached") {
+  appendFileSync(process.env.GSD_GIT_SHIM_LOG, args.join("\\0") + "\\n");
+}
+
+const result = spawnSync(process.env.GSD_REAL_GIT || "git", args, {
+  stdio: "inherit",
+  env: process.env,
+});
+if (result.error) {
+  console.error(result.error.message);
+  process.exit(1);
+}
+process.exit(result.status ?? 0);
+`.trimStart(), "utf-8");
+
+    const posixShim = join(shimDir, "git");
+    const quotedScript = `'${shimScript.replace(/'/g, "'\\''")}'`;
+    writeFileSync(posixShim, `#!/bin/sh\nexec node ${quotedScript} "$@"\n`, "utf-8");
+    chmodSync(posixShim, 0o755);
+
+    writeFileSync(join(shimDir, "git.cmd"), "@echo off\r\nnode \"%~dp0git-shim.cjs\" %*\r\n", "utf-8");
+  }
+
+  function countRmCachedInvocations(logFile: string): number {
+    if (!existsSync(logFile)) return 0;
+    return readFileSync(logFile, "utf-8").split(/\r?\n/).filter(Boolean).length;
+  }
+
   // ─── GitServiceImpl: smart staging ─────────────────────────────────────
 
   test('GitServiceImpl: smart staging', () => {
@@ -447,6 +492,80 @@ describe('git-service', async () => {
     rmSync(repo, { recursive: true, force: true });
   });
 
+  test('GitServiceImpl: runtime cleanup runs once per repo across service instances', () => {
+    const repo = initTempRepo();
+    const shimDir = mkdtempSync(join(tmpdir(), "gsd-git-rm-shim-"));
+    const logFile = join(shimDir, "git-rm-cached.log");
+    const originalPath = process.env.PATH;
+    const originalRealGit = process.env.GSD_REAL_GIT;
+    const originalShimLog = process.env.GSD_GIT_SHIM_LOG;
+    const originalSafePath = GIT_NO_PROMPT_ENV.PATH;
+    const originalSafeRealGit = GIT_NO_PROMPT_ENV.GSD_REAL_GIT;
+    const originalSafeShimLog = GIT_NO_PROMPT_ENV.GSD_GIT_SHIM_LOG;
+
+    try {
+      installGitRmCachedCounterShim(shimDir);
+      const realGit = findGitExecutable();
+      process.env.GSD_REAL_GIT = realGit;
+      process.env.GSD_GIT_SHIM_LOG = logFile;
+      process.env.PATH = `${shimDir}${delimiter}${originalPath ?? ""}`;
+      GIT_NO_PROMPT_ENV.GSD_REAL_GIT = realGit;
+      GIT_NO_PROMPT_ENV.GSD_GIT_SHIM_LOG = logFile;
+      GIT_NO_PROMPT_ENV.PATH = process.env.PATH;
+
+      createFile(repo, "src/first.ts", "first");
+      const first = new GitServiceImpl(repo).commit({ message: "test: first cleanup pass" });
+      assert.deepStrictEqual(first, "test: first cleanup pass", "first commit succeeds");
+      assert.deepStrictEqual(
+        countRmCachedInvocations(logFile),
+        RUNTIME_EXCLUSION_PATHS.length,
+        "first service instance checks each runtime exclusion once",
+      );
+
+      createFile(repo, "src/second.ts", "second");
+      const second = new GitServiceImpl(repo).commit({ message: "test: second cleanup pass" });
+      assert.deepStrictEqual(second, "test: second cleanup pass", "second commit succeeds");
+      assert.deepStrictEqual(
+        countRmCachedInvocations(logFile),
+        RUNTIME_EXCLUSION_PATHS.length,
+        "fresh service instance for the same repo does not rerun runtime cleanup",
+      );
+    } finally {
+      if (originalPath === undefined) {
+        delete process.env.PATH;
+      } else {
+        process.env.PATH = originalPath;
+      }
+      if (originalRealGit === undefined) {
+        delete process.env.GSD_REAL_GIT;
+      } else {
+        process.env.GSD_REAL_GIT = originalRealGit;
+      }
+      if (originalShimLog === undefined) {
+        delete process.env.GSD_GIT_SHIM_LOG;
+      } else {
+        process.env.GSD_GIT_SHIM_LOG = originalShimLog;
+      }
+      if (originalSafePath === undefined) {
+        delete GIT_NO_PROMPT_ENV.PATH;
+      } else {
+        GIT_NO_PROMPT_ENV.PATH = originalSafePath;
+      }
+      if (originalSafeRealGit === undefined) {
+        delete GIT_NO_PROMPT_ENV.GSD_REAL_GIT;
+      } else {
+        GIT_NO_PROMPT_ENV.GSD_REAL_GIT = originalSafeRealGit;
+      }
+      if (originalSafeShimLog === undefined) {
+        delete GIT_NO_PROMPT_ENV.GSD_GIT_SHIM_LOG;
+      } else {
+        GIT_NO_PROMPT_ENV.GSD_GIT_SHIM_LOG = originalSafeShimLog;
+      }
+      rmSync(repo, { recursive: true, force: true });
+      rmSync(shimDir, { recursive: true, force: true });
+    }
+  });
+
   test('GitServiceImpl: task autoCommit skips keyFiles inside submodules', () => {
     const repo = initTempRepo();
     const subSrc = mkdtempSync(join(tmpdir(), "gsd-git-submodule-src-"));
@@ -1505,11 +1624,13 @@ describe('git-service', async () => {
     // Create and track runtime files (simulates pre-.gitignore state)
     mkdirSync(join(repo, ".gsd", "activity"), { recursive: true });
     mkdirSync(join(repo, ".gsd", "runtime"), { recursive: true });
+    mkdirSync(join(repo, ".gsd-backups", "migrate-old"), { recursive: true });
     writeFileSync(join(repo, ".gsd", "completed-units.json"), '["u1"]');
     writeFileSync(join(repo, ".gsd", "metrics.json"), '{}');
     writeFileSync(join(repo, ".gsd", "STATE.md"), "# State");
     writeFileSync(join(repo, ".gsd", "activity", "log.jsonl"), "{}");
     writeFileSync(join(repo, ".gsd", "runtime", "data.json"), "{}");
+    writeFileSync(join(repo, ".gsd-backups", "migrate-old", "marker.txt"), "backup");
     writeFileSync(join(repo, "src.ts"), "code");
     runGit(repo, ["add", "-A"]);
     runGit(repo, ["commit", "-m", "init"]);
@@ -1518,6 +1639,8 @@ describe('git-service', async () => {
     const trackedBefore = run("git ls-files .gsd/", repo);
     assert.ok(trackedBefore.includes("completed-units.json"), "untrack: precondition — completed-units tracked");
     assert.ok(trackedBefore.includes("metrics.json"), "untrack: precondition — metrics tracked");
+    const backupTrackedBefore = run("git ls-files .gsd-backups/", repo);
+    assert.ok(backupTrackedBefore.includes("marker.txt"), "untrack: precondition — migration backup tracked");
 
     // Run untrackRuntimeFiles
     untrackRuntimeFiles(repo);
@@ -1525,6 +1648,8 @@ describe('git-service', async () => {
     // Runtime files should be removed from the index
     const trackedAfter = run("git ls-files .gsd/", repo);
     assert.deepStrictEqual(trackedAfter, "", "untrack: all runtime files removed from index");
+    const backupTrackedAfter = run("git ls-files .gsd-backups/", repo);
+    assert.deepStrictEqual(backupTrackedAfter, "", "untrack: migration backups removed from index");
 
     // Non-runtime files remain tracked
     const srcTracked = run("git ls-files src.ts", repo);
@@ -1535,6 +1660,8 @@ describe('git-service', async () => {
       "untrack: completed-units.json still on disk");
     assert.ok(existsSync(join(repo, ".gsd", "metrics.json")),
       "untrack: metrics.json still on disk");
+    assert.ok(existsSync(join(repo, ".gsd-backups", "migrate-old", "marker.txt")),
+      "untrack: migration backup still on disk");
 
     // Idempotent — running again doesn't error
     untrackRuntimeFiles(repo);