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

package/src/resources/extensions/gsd/tests/auto-paused-ui-cleanup.test.ts

@@ -3,13 +3,14 @@
 import { test } from "node:test";
 import assert from "node:assert/strict";
 import { execFileSync } from "node:child_process";
-import { mkdirSync, mkdtempSync, realpathSync, rmSync, writeFileSync } from "node:fs";
+import { existsSync, mkdirSync, mkdtempSync, realpathSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { join, dirname } from "node:path";
 
 import {
   anchorProcessCwdForAutoResume,
   cleanupAfterLoopExit,
+  maybeRerootStepSessionForHighContext,
   pauseAuto,
   rerootCommandSession,
   stopAuto,
@@ -478,8 +479,8 @@ test("cleanupAfterLoopExit preserves step-mode surface and worktree session afte
   }
 });
 
-test("cleanupAfterLoopExit re-roots step-mode session when context exceeds compaction threshold", async (t) => {
-  const base = mkdtempSync(join(tmpdir(), "gsd-step-reroot-"));
+test("cleanupAfterLoopExit warns but does not re-root step-mode session at soft compaction threshold", async (t) => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-step-soft-context-"));
   const worktree = join(base, ".gsd", "worktrees", "M001");
   const previousCwd = process.cwd();
   const newSessionWorkspaces: string[] = [];
@@ -515,7 +516,59 @@ test("cleanupAfterLoopExit re-roots step-mode session when context exceeds compa
       },
     } as any);
 
-    assert.deepEqual(newSessionWorkspaces, [worktree], "hot step-mode cleanup should re-root in the active worktree");
+    assert.deepEqual(newSessionWorkspaces, [], "soft context pressure must not force a new command session");
+    assert.equal(
+      notifyMessages.some((msg) => msg.includes("context at 72.0%") && msg.includes("Use /compact")),
+      true,
+      "soft threshold should warn with a non-disruptive compaction prompt",
+    );
+    assert.equal(autoSession.basePath, worktree);
+    assert.equal(realpathSync(process.cwd()), realpathSync(worktree));
+  } finally {
+    autoSession.reset();
+    process.chdir(previousCwd);
+    rmSync(base, { recursive: true, force: true });
+  }
+});
+
+test("cleanupAfterLoopExit re-roots step-mode session at hard context threshold", async (t) => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-step-reroot-"));
+  const worktree = join(base, ".gsd", "worktrees", "M001");
+  const previousCwd = process.cwd();
+  const newSessionWorkspaces: string[] = [];
+  const notifyMessages: string[] = [];
+
+  t.mock.method(WorktreeLifecycle.prototype, "restoreToProjectRoot", function () {});
+
+  mkdirSync(worktree, { recursive: true });
+  process.chdir(worktree);
+  autoSession.reset();
+  autoSession.active = true;
+  autoSession.paused = false;
+  autoSession.stepMode = true;
+  autoSession.preserveStepSurfaceAfterLoopExit = true;
+  autoSession.basePath = worktree;
+  autoSession.originalBasePath = base;
+  autoSession.cmdCtx = {
+    getContextUsage: () => ({ percent: 92, tokens: 920_000, contextWindow: 1_000_000 }),
+    newSession: async ({ workspaceRoot }: { workspaceRoot: string }) => {
+      newSessionWorkspaces.push(workspaceRoot);
+      return { cancelled: false };
+    },
+  } as any;
+
+  try {
+    await cleanupAfterLoopExit({
+      hasUI: true,
+      ui: {
+        setStatus: () => {},
+        setWidget: () => {},
+        setHeader: () => {},
+        notify: (_msg: string) => notifyMessages.push(_msg),
+      },
+    } as any);
+
+    assert.deepEqual(newSessionWorkspaces, [worktree], "critical context pressure should re-root in the active worktree");
     assert.equal(
       notifyMessages.some((msg) => msg.includes("Fresh session ready for /gsd next")),
       true,
@@ -530,6 +583,131 @@ test("cleanupAfterLoopExit re-roots step-mode session when context exceeds compa
   }
 });
 
+test("maybeRerootStepSessionForHighContext re-roots at the hard threshold even when the soft pref is set above it", async () => {
+  // Regression: the hard 90% re-root must not be gated behind the configurable
+  // soft warn threshold, which is allowed up to 0.95. At 92% usage with a 0.95
+  // soft pref the soft warn predicate is false, but the session must still
+  // re-root because usage crossed the fixed hard boundary.
+  const dir = mkdtempSync(join(tmpdir(), "gsd-hard-reroot-high-soft-"));
+  const project = join(dir, "project");
+  const gsdHomeDir = join(dir, "home");
+  const previousCwd = process.cwd();
+  const previousGsdHome = process.env.GSD_HOME;
+
+  mkdirSync(join(project, ".gsd"), { recursive: true });
+  mkdirSync(gsdHomeDir, { recursive: true });
+  writeFileSync(
+    join(project, ".gsd", "PREFERENCES.md"),
+    ["---", "version: 1", "context_management:", "  compaction_threshold_percent: 0.95", "---", ""].join("\n"),
+    "utf-8",
+  );
+
+  process.env.GSD_HOME = gsdHomeDir;
+  process.chdir(project);
+
+  const newSessionWorkspaces: string[] = [];
+  const notifyMessages: string[] = [];
+  const cmdCtx = {
+    getContextUsage: () => ({ percent: 92, tokens: 920_000, contextWindow: 1_000_000 }),
+    newSession: async ({ workspaceRoot }: { workspaceRoot: string }) => {
+      newSessionWorkspaces.push(workspaceRoot);
+      return { cancelled: false };
+    },
+  } as any;
+  const ctx = { ui: { notify: (msg: string) => notifyMessages.push(msg) } } as any;
+
+  try {
+    const result = await maybeRerootStepSessionForHighContext(ctx, project, cmdCtx, project);
+
+    assert.equal(result.rerooted, true, "92% usage must re-root even when the soft pref is 95%");
+    assert.deepEqual(newSessionWorkspaces, [project], "hard context pressure must re-root the command session");
+    const rerootMessage = notifyMessages.find((msg) => msg.includes("Fresh session ready for /gsd next"));
+    assert.ok(rerootMessage, "hard re-root should notify the user");
+    // The re-root was triggered by the 90% hard limit, so the notification must
+    // cite that limit and not the configured 95% soft compaction threshold.
+    assert.match(
+      rerootMessage,
+      /hard threshold: 90%/,
+      "hard re-root notification must cite the 90% hard threshold",
+    );
+    assert.equal(
+      rerootMessage.includes("95%"),
+      false,
+      "hard re-root notification must not cite the configured soft compaction threshold",
+    );
+  } finally {
+    process.chdir(previousCwd);
+    if (previousGsdHome === undefined) delete process.env.GSD_HOME;
+    else process.env.GSD_HOME = previousGsdHome;
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("maybeRerootStepSessionForHighContext writes a compaction snapshot only at the hard boundary", async () => {
+  // Regression: the snapshot is reserved for the hard re-root. Soft-only
+  // pressure must warn without writing a compaction snapshot.
+  const dir = mkdtempSync(join(tmpdir(), "gsd-soft-no-snapshot-"));
+  const project = join(dir, "project");
+  const gsdHomeDir = join(dir, "home");
+  const previousCwd = process.cwd();
+  const previousGsdHome = process.env.GSD_HOME;
+
+  mkdirSync(join(project, ".gsd"), { recursive: true });
+  mkdirSync(gsdHomeDir, { recursive: true });
+  writeFileSync(
+    join(project, ".gsd", "PREFERENCES.md"),
+    ["---", "version: 1", "context_management:", "  compaction_threshold_percent: 0.6", "---", ""].join("\n"),
+    "utf-8",
+  );
+
+  process.env.GSD_HOME = gsdHomeDir;
+  process.chdir(project);
+
+  const newSessionWorkspaces: string[] = [];
+  const notifyMessages: string[] = [];
+  let contextPercent = 72;
+  const cmdCtx = {
+    getContextUsage: () => ({ percent: contextPercent, tokens: contextPercent * 10_000, contextWindow: 1_000_000 }),
+    newSession: async ({ workspaceRoot }: { workspaceRoot: string }) => {
+      newSessionWorkspaces.push(workspaceRoot);
+      return { cancelled: false };
+    },
+  } as any;
+  const ctx = { ui: { notify: (msg: string) => notifyMessages.push(msg) } } as any;
+  const snapshotPath = join(project, ".gsd", "last-snapshot.md");
+
+  try {
+    const softResult = await maybeRerootStepSessionForHighContext(ctx, project, cmdCtx, project);
+    assert.equal(softResult.rerooted, false, "soft pressure must not re-root");
+    assert.deepEqual(newSessionWorkspaces, [], "soft pressure must not open a new session");
+    assert.equal(
+      existsSync(snapshotPath),
+      false,
+      "soft-only pressure must not write a compaction snapshot",
+    );
+    assert.equal(
+      notifyMessages.some((msg) => msg.includes("context at 72.0%") && msg.includes("Use /compact")),
+      true,
+      "soft threshold should warn with a non-disruptive compaction prompt",
+    );
+
+    contextPercent = 92;
+    const hardResult = await maybeRerootStepSessionForHighContext(ctx, project, cmdCtx, project);
+    assert.equal(hardResult.rerooted, true, "hard pressure must re-root");
+    assert.deepEqual(newSessionWorkspaces, [project], "hard pressure should re-root once");
+    assert.equal(
+      existsSync(snapshotPath),
+      true,
+      "hard boundary must write the compaction snapshot",
+    );
+  } finally {
+    process.chdir(previousCwd);
+    if (previousGsdHome === undefined) delete process.env.GSD_HOME;
+    else process.env.GSD_HOME = previousGsdHome;
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
 test("cleanupAfterLoopExit restores project root through lifecycle and preserves chdir", async (t) => {
   const base = mkdtempSync(join(tmpdir(), "gsd-cleanup-lifecycle-"));
   const worktree = join(base, ".gsd", "worktrees", "M001");

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

@@ -2,7 +2,7 @@
 // File Purpose: Tests auto-mode artifact verification and recovery behavior.
 import test, { afterEach } from "node:test";
 import assert from "node:assert/strict";
-import { mkdtempSync, mkdirSync, rmSync, writeFileSync, existsSync, readFileSync } from "node:fs";
+import { chmodSync, mkdtempSync, mkdirSync, rmSync, writeFileSync, existsSync, readFileSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
 import { randomUUID } from "node:crypto";
@@ -964,6 +964,45 @@ function makeGitBase(): string {
   return base;
 }
 
+function shellQuote(value: string): string {
+  return `'${value.replace(/'/g, "'\\''")}'`;
+}
+
+function withLoggedGitCommands<T>(base: string, action: () => T): { result: T; commands: string[] } {
+  const realGit = execFileSync("which", ["git"], {
+    encoding: "utf-8",
+    stdio: ["ignore", "pipe", "pipe"],
+  }).trim().split(/\r?\n/)[0];
+  if (!realGit) throw new Error("Unable to resolve git executable for invocation logging test");
+  const binDir = join(base, ".git-wrapper-bin");
+  const logFile = join(base, "git-invocations.log");
+  mkdirSync(binDir, { recursive: true });
+  const wrapper = join(binDir, "git");
+  writeFileSync(
+    wrapper,
+    [
+      "#!/usr/bin/env bash",
+      `printf '%s\\n' "$1" >> ${shellQuote(logFile)}`,
+      `exec ${shellQuote(realGit)} "$@"`,
+      "",
+    ].join("\n"),
+  );
+  chmodSync(wrapper, 0o755);
+
+  const originalPath = process.env.PATH;
+  process.env.PATH = originalPath ? `${binDir}:${originalPath}` : binDir;
+  try {
+    const result = action();
+    const commands = existsSync(logFile)
+      ? readFileSync(logFile, "utf-8").split(/\r?\n/).filter(Boolean)
+      : [];
+    return { result, commands };
+  } finally {
+    if (originalPath === undefined) delete process.env.PATH;
+    else process.env.PATH = originalPath;
+  }
+}
+
 test("hasImplementationArtifacts returns false when only .gsd/ files committed (#1703)", () => {
   const base = makeGitBase();
   try {
@@ -1344,6 +1383,77 @@ test("hasImplementationArtifacts finds implementation commits when .gsd/ is giti
   }
 });
 
+test("hasImplementationArtifacts scans GSD-tagged history without per-commit diff-tree forks (#892)", { skip: process.platform === "win32" }, () => {
+  const base = makeGitBase();
+  try {
+    writeFileSync(join(base, ".git", "info", "exclude"), ".gsd/\n");
+    mkdirSync(join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks"), { recursive: true });
+    writeFileSync(
+      join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks", "T01-SUMMARY.md"),
+      "# Summary",
+    );
+
+    mkdirSync(join(base, "src"), { recursive: true });
+    for (let i = 0; i < 3; i++) {
+      writeFileSync(join(base, "src", `feature-${i}.ts`), `export const feature${i} = true;\n`);
+      execFileSync("git", ["add", "src"], { cwd: base, stdio: "ignore" });
+      execFileSync(
+        "git",
+        ["commit", "-m", `feat: materialize M001 evidence ${i}\n\nGSD-Task: S01/T01`],
+        { cwd: base, stdio: "ignore" },
+      );
+    }
+
+    const { result, commands } = withLoggedGitCommands(base, () => hasImplementationArtifacts(base, "M001"));
+    assert.equal(result, "present", "milestone-tagged commits should still prove implementation evidence");
+    assert.ok(commands.includes("log"), "milestone evidence fallback should scan history with git log");
+    assert.equal(commands.filter((command) => command === "diff-tree").length, 0);
+  } finally {
+    cleanup(base);
+  }
+});
+
+test("hasImplementationArtifacts backfill scans commit records without per-commit diff-tree forks (#892)", { skip: process.platform === "win32" }, () => {
+  const base = makeGitBase();
+  try {
+    mkdirSync(join(base, ".gsd"), { recursive: true });
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone One", status: "active" });
+    insertSlice({
+      id: "S01",
+      milestoneId: "M001",
+      title: "Slice One",
+      status: "complete",
+      risk: "low",
+      depends: [],
+    });
+    insertTask({
+      id: "T01",
+      sliceId: "S01",
+      milestoneId: "M001",
+      title: "Task One",
+      status: "complete",
+      keyFiles: ["src/expected-0.ts", "src/expected-1.ts"],
+      planning: { files: ["src/expected-0.ts", "src/expected-1.ts"] },
+    });
+
+    mkdirSync(join(base, "src"), { recursive: true });
+    for (let i = 0; i < 2; i++) {
+      writeFileSync(join(base, "src", `expected-${i}.ts`), `export const expected${i} = true;\n`);
+      execFileSync("git", ["add", "src"], { cwd: base, stdio: "ignore" });
+      execFileSync("git", ["commit", "-m", `feat: untagged implementation ${i}`], { cwd: base, stdio: "ignore" });
+    }
+
+    const { result, commands } = withLoggedGitCommands(base, () => hasImplementationArtifacts(base, "M001"));
+    assert.equal(result, "present", "completed task file hints should still backfill untagged commits");
+    assert.equal(getMilestoneCommitAttributionShas("M001").length, 2);
+    assert.ok(commands.includes("log"), "backfill should scan commit records with git log");
+    assert.equal(commands.filter((command) => command === "diff-tree").length, 0);
+  } finally {
+    cleanup(base);
+  }
+});
+
 test("hasImplementationArtifacts binds GSD-Task trailer to milestone via DB state when .gsd/ is gitignored", () => {
   const base = makeGitBase();
   try {
@@ -1807,6 +1917,41 @@ test("#57: writeReactiveExecuteBlocker reconciles batch task statuses and append
   }
 });
 
+test("#1088: writeReactiveExecuteBlocker preserves deferred batch task statuses", () => {
+  const base = makeTmpBase();
+  try {
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone", status: "active" });
+    insertSlice({ id: "S01", milestoneId: "M001", title: "Slice", status: "pending" });
+    insertTask({ id: "T01", milestoneId: "M001", sliceId: "S01", title: "One", status: "deferred" });
+    insertTask({ id: "T02", milestoneId: "M001", sliceId: "S01", title: "Two", status: "deferred" });
+    writeFileSync(
+      join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks", "T01-SUMMARY.md"),
+      "# T01 Summary\n",
+      "utf-8",
+    );
+
+    const recovery = writeReactiveExecuteBlocker(
+      "M001/S01/reactive+T01,T02",
+      base,
+      "verification retries exhausted",
+    );
+
+    assert.ok(recovery, "recovery should run with DB available");
+    assert.deepEqual(recovery!.completedTaskIds, []);
+    assert.deepEqual(recovery!.skippedTaskIds, []);
+    assert.deepEqual(recovery!.unchangedTaskIds, ["T01", "T02"]);
+    assert.equal(getTask("M001", "S01", "T01")?.status, "deferred");
+    assert.equal(getTask("M001", "S01", "T02")?.status, "deferred");
+
+    const events = readEvents(join(base, ".gsd", "event-log.jsonl"));
+    assert.equal(events.some((e) => e.params.taskId === "T01" || e.params.taskId === "T02"), false);
+  } finally {
+    closeDatabase();
+    cleanup(base);
+  }
+});
+
 test("#4414: verifyExpectedArtifact parallel-research succeeds when all research-ready slices have RESEARCH", () => {
   const base = makeTmpBase();
   try {

package/src/resources/extensions/gsd/tests/auto-remote-session-lock-cleanup.test.ts

@@ -124,3 +124,44 @@ test("forceStopAutoRemote escalates a live remote PID and releases worker state"
   );
   assert.equal(readCrashLock(base), null, "force stop should remove the visible remote lock");
 });
+
+test("forceStopAutoRemote does not SIGKILL a PID that exits during the grace window", (t) => {
+  const base = makeBase();
+  t.after(() => cleanup(base));
+
+  openDatabase(join(base, ".gsd", "gsd.db"));
+  const workerId = registerAutoWorker({ projectRootRealpath: normalizeRealPath(base) });
+  const pid = 525_252;
+  setWorkerPid(workerId, pid);
+  writeLegacyLock(base, pid);
+
+  // The process is alive until it receives SIGTERM, then exits cooperatively
+  // (liveness probe throws ESRCH) — so the grace loop should break before SIGKILL.
+  const signals: Array<NodeJS.Signals | 0> = [];
+  let alive = true;
+  const originalKill = process.kill;
+  process.kill = ((target: number, signal?: NodeJS.Signals | number) => {
+    assert.equal(target, pid);
+    if (signal === 0) {
+      if (!alive) {
+        const err = new Error("no such process") as NodeJS.ErrnoException;
+        err.code = "ESRCH";
+        throw err;
+      }
+      return true;
+    }
+    signals.push((signal ?? 0) as NodeJS.Signals | 0);
+    if (signal === "SIGTERM") alive = false; // cooperative exit on SIGTERM
+    return true;
+  }) as typeof process.kill;
+  t.after(() => {
+    process.kill = originalKill;
+  });
+
+  const result = forceStopAutoRemote(base);
+
+  assert.deepEqual(result, { found: true, pid });
+  assert.ok(signals.includes("SIGTERM"), "force stop should request graceful termination first");
+  assert.ok(!signals.includes("SIGKILL"), "force stop must not escalate when the PID exits during the grace window");
+  assert.equal(readCrashLock(base), null, "force stop should remove the visible remote lock");
+});

package/src/resources/extensions/gsd/tests/auto-retry-mcp-churn-fixes.test.ts

@@ -15,15 +15,21 @@
 
 import { describe, it, beforeEach } from "node:test";
 import assert from "node:assert/strict";
+import { mkdtempSync, rmSync, statSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 
 import {
   resetEvidence,
   getEvidence,
   recordToolCall,
   recordToolResult,
+  saveEvidenceToDisk,
   type BashEvidence,
 } from "../safety/evidence-collector.js";
 
+const waitForMtimeTick = () => new Promise(resolve => setTimeout(resolve, 25));
+
 describe("evidence-collector: toolCallId-based matching (A-3)", () => {
   beforeEach(() => {
     resetEvidence();
@@ -119,4 +125,28 @@ describe("evidence-collector: toolCallId-based matching (A-3)", () => {
     assert.equal(entries[1].kind, "bash");
     assert.equal(entries[1].command, "rg TODO");
   });
+
+  it("skips byte-identical evidence file rewrites", async (t) => {
+    const base = mkdtempSync(join(tmpdir(), "gsd-evidence-dedupe-"));
+    t.after(() => rmSync(base, { recursive: true, force: true }));
+
+    recordToolCall("tc-bash-1", "bash", { command: "pnpm test" });
+    saveEvidenceToDisk(base, "M001", "S01", "T03");
+
+    const evidenceFile = join(base, ".gsd", "safety", "evidence-M001-S01-T03.json");
+    const firstMtime = statSync(evidenceFile, { bigint: true }).mtimeNs;
+
+    await waitForMtimeTick();
+    saveEvidenceToDisk(base, "M001", "S01", "T03");
+    const secondMtime = statSync(evidenceFile, { bigint: true }).mtimeNs;
+
+    assert.equal(secondMtime, firstMtime, "unchanged evidence must not rewrite the evidence file");
+
+    await waitForMtimeTick();
+    recordToolResult("tc-bash-1", "bash", "Command exited with code 0\nok\n", false);
+    saveEvidenceToDisk(base, "M001", "S01", "T03");
+    const thirdMtime = statSync(evidenceFile, { bigint: true }).mtimeNs;
+
+    assert.ok(thirdMtime > secondMtime, "changed evidence must still be persisted");
+  });
 });

package/src/resources/extensions/gsd/tests/auto-runtime-state.test.ts

@@ -3,6 +3,9 @@
 
 import test from "node:test";
 import assert from "node:assert/strict";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 
 import {
   autoSession,
@@ -11,6 +14,10 @@ import {
   getAutoRuntimeSnapshot,
   recordAutoToolSurfaceSnapshot,
 } from "../auto-runtime-state.ts";
+import {
+  readUnitHarnessAbort,
+  recordUnitHarnessAbort,
+} from "../unit-runtime.ts";
 
 test("getAutoRuntimeSnapshot includes orchestration phase when available", () => {
   autoSession.reset();
@@ -81,6 +88,40 @@ test("clearToolInvocationError clears stale tool error state for active auto ses
   autoSession.reset();
 });
 
+test("clearToolInvocationError clears stale tool error even when a harness abort is durable", () => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-auto-runtime-state-"));
+  const startedAt = 123456;
+  try {
+    autoSession.reset();
+    autoSession.active = true;
+    autoSession.basePath = base;
+    autoSession.setCurrentUnit({
+      type: "gate-evaluate",
+      id: "M001/S01/gates+Q3",
+      startedAt,
+      workspaceRoot: base,
+    });
+    autoSession.lastToolInvocationError = "gsd_save_gate_result: simulated stale tool error";
+    recordUnitHarnessAbort(base, "gate-evaluate", "M001/S01/gates+Q3", startedAt, {
+      kind: "tool-error",
+      reason: "Tool execution failed before the unit could complete its gate evaluation.",
+      toolName: "gsd_uat_exec",
+    });
+
+    clearToolInvocationError();
+
+    assert.equal(autoSession.lastToolInvocationError, null);
+    assert.equal(
+      readUnitHarnessAbort(base, "gate-evaluate", "M001/S01/gates+Q3", startedAt)?.kind,
+      "tool-error",
+      "durable harness abort remains available for result-save blocking",
+    );
+  } finally {
+    autoSession.reset();
+    rmSync(base, { recursive: true, force: true });
+  }
+});
+
 test("getAutoRuntimeSnapshot omits orchestration phase when seam not wired", () => {
   autoSession.reset();
 

package/src/resources/extensions/gsd/tests/auto-session-scope.test.ts

@@ -121,8 +121,8 @@ describe("AutoSession.scope — project mode (basePath equals originalBasePath)"
 
     assert.ok(s.scope, "scope should be set");
     const gsd = join(projectDir, ".gsd");
-    assert.equal(s.scope.contextFile(), join(gsd, "milestones", mid, `${mid}-CONTEXT.md`));
-    assert.equal(s.scope.roadmapFile(), join(gsd, "milestones", mid, `${mid}-ROADMAP.md`));
+    assert.equal(s.scope.contextFile(), join(gsd, "phases", "02-m002", "02-CONTEXT.md"));
+    assert.equal(s.scope.roadmapFile(), join(gsd, "phases", "02-m002", "02-ROADMAP.md"));
     assert.equal(s.scope.stateFile(), join(gsd, "STATE.md"));
   });
 });
@@ -236,8 +236,8 @@ describe("AutoSession.scope — milestoneId change rebuilds scope", () => {
     assert.ok(ctxM001, "M001 contextFile should be set");
     assert.ok(ctxM002, "M002 contextFile should be set");
     assert.notEqual(ctxM001, ctxM002, "contextFile must differ between milestone IDs");
-    assert.ok(ctxM001.includes("M001"), "M001 path should contain M001");
-    assert.ok(ctxM002.includes("M002"), "M002 path should contain M002");
+    assert.ok(ctxM001.includes("01-m001"), "M001 path should contain its phase directory");
+    assert.ok(ctxM002.includes("02-m002"), "M002 path should contain its phase directory");
   });
 });
 

package/src/resources/extensions/gsd/tests/auto-start-bootstrap-await-3420.test.ts

@@ -17,14 +17,17 @@ import {
   clearPendingAutoStart,
   setPendingAutoStart,
 } from "../guided-flow.ts";
+import { closeDatabase, openDatabase } from "../gsd-db.ts";
 
 test.afterEach(() => {
+  closeDatabase();
   clearPendingAutoStart();
 });
 
 test("checkAutoStartAfterDiscuss waits until discussion artifacts exist before re-entering auto-mode", (t) => {
   const base = mkdtempSync(join(tmpdir(), "gsd-bootstrap-await-"));
   const notifications: string[] = [];
+  openDatabase(":memory:");
   t.after(() => rmSync(base, { recursive: true, force: true }));
   mkdirSync(join(base, ".gsd", "milestones", "M001"), { recursive: true });
   setPendingAutoStart(base, {

package/src/resources/extensions/gsd/tests/auto-timeout-recovery-harness-abort.test.ts

@@ -0,0 +1,55 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, mkdirSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+import { recoverTimedOutUnit } from "../auto-timeout-recovery.ts";
+import {
+  readUnitHarnessAbort,
+  readUnitRuntimeRecord,
+  recordUnitHarnessAbort,
+} from "../unit-runtime.ts";
+
+test("timeout recovery retry clears stale harness abort for the same unit run", async () => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-timeout-recovery-abort-"));
+  const startedAt = Date.now();
+  mkdirSync(join(base, ".gsd"), { recursive: true });
+  try {
+    recordUnitHarnessAbort(base, "gate-evaluate", "M001/S01/gates+Q3", startedAt, {
+      kind: "turn-abort",
+      reason: "Agent turn aborted before the unit could complete its gate evaluation.",
+    });
+    assert.equal(
+      readUnitHarnessAbort(base, "gate-evaluate", "M001/S01/gates+Q3", startedAt)?.kind,
+      "turn-abort",
+      "test setup should start with an abort marker",
+    );
+
+    const messages: Array<{ message: unknown; options: unknown }> = [];
+    const notifications: Array<{ message: string; level?: string }> = [];
+    const result = await recoverTimedOutUnit(
+      { ui: { notify: (message: string, level?: string) => notifications.push({ message, level }) } } as any,
+      { sendMessage: (message: unknown, options: unknown) => messages.push({ message, options }) } as any,
+      "gate-evaluate",
+      "M001/S01/gates+Q3",
+      "idle",
+      {
+        basePath: base,
+        verbose: false,
+        currentUnitStartedAt: startedAt,
+        unitRecoveryCount: new Map(),
+      },
+    );
+
+    assert.equal(result, "recovered");
+    assert.equal(readUnitHarnessAbort(base, "gate-evaluate", "M001/S01/gates+Q3", startedAt), null);
+    const record = readUnitRuntimeRecord(base, "gate-evaluate", "M001/S01/gates+Q3");
+    assert.equal(record?.phase, "recovered");
+    assert.equal(record?.lastProgressKind, "idle-recovery-retry");
+    assert.equal(messages.length, 1, "retry recovery should send one steering message");
+    assert.equal(notifications.length, 1, "retry recovery should notify once");
+  } finally {
+    rmSync(base, { recursive: true, force: true });
+  }
+});

package/src/resources/extensions/gsd/tests/auto-warning-noise-regression.test.ts

@@ -34,6 +34,11 @@ import {
   clearPendingAutoStart,
   setPendingAutoStart,
 } from "../guided-flow.ts";
+import {
+  closeDatabase,
+  insertMilestone,
+  openDatabase,
+} from "../gsd-db.ts";
 
 // ─── Bug 2: this-binding regression ─────────────────────────────────────
 
@@ -89,6 +94,8 @@ test("checkAutoStartAfterDiscuss completes when discussion manifest is absent",
     mkdirSync(milestoneDir, { recursive: true });
     writeFileSync(join(milestoneDir, "M001-CONTEXT.md"), "# Context\n", "utf-8");
     writeFileSync(join(base, ".gsd", "STATE.md"), "# State\n", "utf-8");
+    openDatabase(":memory:");
+    insertMilestone({ id: "M001", title: "M001", status: "queued" });
     setPendingAutoStart(base, {
       basePath: base,
       milestoneId: "M001",
@@ -110,6 +117,7 @@ test("checkAutoStartAfterDiscuss completes when discussion manifest is absent",
       },
     ]);
   } finally {
+    closeDatabase();
     clearPendingAutoStart(base);
     rmSync(base, { recursive: true, force: true });
   }