gsd-pi 2.76.0-dev.fe143342a → 2.77.0-dev.1594b263e

package/src/resources/extensions/gsd/tests/pre-execution-checks.test.ts

@@ -1599,6 +1599,221 @@ describe("checkTaskOrdering directory inputs (#4446)", () => {
   });
 });
 
+// ─── Regression Tests: checkTaskOrdering false positive for pre-execution refs (#4071) ──
+
+describe("checkTaskOrdering false positive for pre-execution refs (#4071)", () => {
+  test("completed task at higher index does not trigger ordering violation for its outputs", () => {
+    // Scenario: after a replan, a completed task at higher array index has already
+    // created a file. A new earlier-sequence task reads that file. Since the
+    // completed task already ran, its output is available regardless of disk state.
+    // checkTaskOrdering must not flag this as a sequence violation.
+    const tasks = [
+      createTask({
+        id: "T_NEW",
+        sequence: 1,
+        status: "pending",
+        inputs: ["artifacts/setup.json"],
+        expected_output: [],
+      }),
+      createTask({
+        id: "T_SETUP",
+        sequence: 5,
+        status: "completed",
+        inputs: [],
+        expected_output: ["artifacts/setup.json"],
+      }),
+    ];
+
+    const results = checkTaskOrdering(tasks, "/tmp");
+    assert.equal(
+      results.length,
+      0,
+      "completed task outputs must not trigger ordering violations for earlier-sequence tasks that read them",
+    );
+  });
+
+  test("pending task at higher index still triggers ordering violation", () => {
+    // A PENDING task at higher index creating a file is a real violation.
+    // Only completed tasks get the exemption.
+    const tasks = [
+      createTask({
+        id: "T01",
+        sequence: 1,
+        status: "pending",
+        inputs: ["artifacts/output.json"],
+        expected_output: [],
+      }),
+      createTask({
+        id: "T02",
+        sequence: 5,
+        status: "pending",
+        inputs: [],
+        expected_output: ["artifacts/output.json"],
+      }),
+    ];
+
+    const results = checkTaskOrdering(tasks, "/tmp");
+    assert.equal(
+      results.length,
+      1,
+      "pending task at higher index must still be flagged as ordering violation",
+    );
+    assert.equal(results[0].blocking, true);
+    assert.ok(results[0].message.includes("T01"));
+    assert.ok(results[0].message.includes("T02"));
+    assert.ok(results[0].message.includes("sequence violation"));
+  });
+
+  test("pending-first then completed-later: completed replaces pending in fileCreators (#4572)", () => {
+    // Regression for CodeRabbit Major finding on PR #4572:
+    // fileCreators only stored the FIRST task for a given path. If a PENDING task at
+    // array index 1 was registered first, and a COMPLETED task at array index 2 also
+    // declared the same output path, the completed entry was discarded. Line ~529 then
+    // saw a pending creator with index > i and incorrectly fired a sequence violation
+    // for the reader at array index 0.
+    //
+    // Scenario: path first declared by pending task (index 1), then by completed task
+    // (index 2). Reader is at index 0. Without the fix a violation fires; with the fix
+    // the completed entry replaces the pending entry and grants the exemption.
+    const tasks = [
+      // array index 0 — reads the shared path
+      createTask({
+        id: "T_READER",
+        sequence: 1,
+        status: "pending",
+        inputs: ["shared/artifact.json"],
+        expected_output: [],
+      }),
+      // array index 1 — pending producer (visited first during map build)
+      createTask({
+        id: "T_PENDING_PRODUCER",
+        sequence: 5,
+        status: "pending",
+        inputs: [],
+        expected_output: ["shared/artifact.json"],
+      }),
+      // array index 2 — completed producer (visited second; must replace pending entry)
+      createTask({
+        id: "T_COMPLETED_PRODUCER",
+        sequence: 2,
+        status: "completed",
+        inputs: [],
+        expected_output: ["shared/artifact.json"],
+      }),
+    ];
+
+    // Without the fix: creator = T_PENDING_PRODUCER (index 1), !creator.completed && 1 > 0 → violation.
+    // With the fix:    creator = T_COMPLETED_PRODUCER (index 2), creator.completed → no violation.
+    const results = checkTaskOrdering(tasks, "/tmp");
+    assert.equal(
+      results.length,
+      0,
+      "completed producer must replace pending producer in fileCreators and suppress false violation",
+    );
+  });
+
+  test("completed task output exemption applies regardless of whether file exists on disk", (t) => {
+    // The completed-task exemption must work even when the file is not on disk
+    // (e.g., the file was a temporary artifact that was cleaned up after the task ran).
+    const tempDir = join(tmpdir(), `pre-exec-completed-task-${Date.now()}`);
+    mkdirSync(tempDir, { recursive: true });
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    // File deliberately NOT created on disk — completed task ran in a prior session
+    const tasks = [
+      createTask({
+        id: "T_MAIN",
+        sequence: 0,
+        status: "pending",
+        inputs: ["generated/config.json"],
+        expected_output: [],
+      }),
+      createTask({
+        id: "T_INIT",
+        sequence: 10,
+        status: "completed",
+        inputs: [],
+        expected_output: ["generated/config.json"],
+      }),
+    ];
+
+    const results = checkTaskOrdering(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "completed task exemption must apply even when file is absent from disk",
+    );
+  });
+});
+
+describe("checkFilePathConsistency completed-task output exemption (#4071)", () => {
+  test("completed task at higher index does not cause false positive for file it produced", (t) => {
+    // Parallel to the checkTaskOrdering fix: checkFilePathConsistency also uses
+    // getExpectedOutputsUpTo which historically only looked at prior-index tasks.
+    // A completed task at a higher index has already run and its outputs are available.
+    const tempDir = join(tmpdir(), `pre-exec-fc-completed-${Date.now()}`);
+    mkdirSync(tempDir, { recursive: true });
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    // File is NOT on disk — completed task ran in a prior session and file was cleaned
+    const tasks = [
+      createTask({
+        id: "T_MAIN",
+        sequence: 1,
+        status: "pending",
+        inputs: ["artifacts/config.json"],
+        expected_output: [],
+      }),
+      createTask({
+        id: "T_SETUP",
+        sequence: 10,
+        status: "completed",
+        inputs: [],
+        expected_output: ["artifacts/config.json"],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "completed task at higher index should satisfy inputs of pending tasks that read its outputs",
+    );
+  });
+
+  test("pending task at higher index still causes a missing-file error", (t) => {
+    const tempDir = join(tmpdir(), `pre-exec-fc-pending-${Date.now()}`);
+    mkdirSync(tempDir, { recursive: true });
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    const tasks = [
+      createTask({
+        id: "T01",
+        sequence: 1,
+        status: "pending",
+        inputs: ["artifacts/output.json"],
+        expected_output: [],
+      }),
+      createTask({
+        id: "T02",
+        sequence: 10,
+        status: "pending",
+        inputs: [],
+        expected_output: ["artifacts/output.json"],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      1,
+      "pending task at higher index must still be flagged — the file is not available yet",
+    );
+    assert.equal(results[0].blocking, true);
+    assert.equal(results[0].target, "artifacts/output.json");
+  });
+});
+
 describe("checkFilePathConsistency self-referential inputs (#4459)", () => {
   test("input that is also in the same task's expected_output is not blocking when missing on disk", (t) => {
     const tempDir = join(tmpdir(), `pre-exec-self-output-${Date.now()}`);
@@ -1664,3 +1879,125 @@ describe("checkFilePathConsistency self-referential inputs (#4459)", () => {
     );
   });
 });
+
+// ─── Regression Tests: quote-wrapped inputs treated as literal paths (#3747) ──
+
+describe("checkFilePathConsistency quote-wrapped annotation (#3747)", () => {
+  test("double-quoted path annotation is stripped before path check", (t) => {
+    // Plan documents sometimes emit `"src/foo.ts"` (double-quote wrapped) as an
+    // input value. The checker must strip the quotes before checking existence so
+    // it doesn't produce a false-positive "file not found" error.
+    const tempDir = join(tmpdir(), `pre-exec-quote-${Date.now()}`);
+    mkdirSync(join(tempDir, "src"), { recursive: true });
+    writeFileSync(join(tempDir, "src/foo.ts"), "// content");
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    const tasks = [
+      createTask({
+        id: "T01",
+        inputs: ['"src/foo.ts"'],
+        expected_output: [],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "Double-quoted path should be stripped and resolved to the real file",
+    );
+  });
+
+  test("single-quoted path annotation is stripped before path check", (t) => {
+    const tempDir = join(tmpdir(), `pre-exec-squote-${Date.now()}`);
+    mkdirSync(join(tempDir, "src"), { recursive: true });
+    writeFileSync(join(tempDir, "src/bar.ts"), "// content");
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    const tasks = [
+      createTask({
+        id: "T01",
+        inputs: ["'src/bar.ts'"],
+        expected_output: [],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "Single-quoted path should be stripped and resolved to the real file",
+    );
+  });
+
+  test("backtick-only wrapped path without annotation resolves correctly", (t) => {
+    // The bare form `src/foo.ts` (no dash annotation) must also work
+    const tempDir = join(tmpdir(), `pre-exec-backtick-bare-${Date.now()}`);
+    mkdirSync(join(tempDir, "src"), { recursive: true });
+    writeFileSync(join(tempDir, "src/baz.ts"), "// content");
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    const tasks = [
+      createTask({
+        id: "T01",
+        inputs: ["`src/baz.ts`"],
+        expected_output: [],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "Bare backtick-wrapped path should resolve to the real file",
+    );
+  });
+
+  test("prose value with spaces inside quotes is skipped (not a path)", () => {
+    // "some description text" contains spaces — should not be checked as a path
+    const tasks = [
+      createTask({
+        id: "T01",
+        inputs: ['"some description text"'],
+        expected_output: [],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, "/tmp");
+    assert.equal(
+      results.length,
+      0,
+      "Quoted prose with spaces should not be treated as a file path",
+    );
+  });
+
+  test("17-error scenario: mixed annotated inputs produce 0 blocking errors", (t) => {
+    // Reproduces the M004-ej6j88/S07 scenario from issue #3747 where a plan with
+    // multiple backtick- and quote-wrapped input strings causes 17 false blocking errors.
+    const tempDir = join(tmpdir(), `pre-exec-3747-scenario-${Date.now()}`);
+    mkdirSync(join(tempDir, "src"), { recursive: true });
+    writeFileSync(join(tempDir, "src/foo.ts"), "// content");
+    writeFileSync(join(tempDir, "src/bar.ts"), "// content");
+    t.after(() => rmSync(tempDir, { recursive: true, force: true }));
+
+    const tasks = [
+      createTask({
+        id: "T01",
+        inputs: [
+          "`src/foo.ts`",
+          '"src/bar.ts"',
+          "some description text",
+          "Existing enum definition",
+        ],
+        expected_output: [],
+      }),
+    ];
+
+    const results = checkFilePathConsistency(tasks, tempDir);
+    assert.equal(
+      results.length,
+      0,
+      "Annotated file paths and prose inputs should produce zero blocking errors",
+    );
+  });
+});

package/src/resources/extensions/gsd/tests/pre-execution-pause-wiring.test.ts

@@ -191,7 +191,12 @@ function createFailingTasks(): void {
       estimate: "1h",
       files: [],
       verify: "npm test",
-      inputs: ["nonexistent-file-that-does-not-exist.ts"],
+      inputs: [
+        "nonexistent-file-that-does-not-exist.ts",
+        "missing-second-file.ts",
+        "missing-third-file.ts",
+        "missing-fourth-file.ts",
+      ],
       expectedOutput: [],
       observabilityImpact: "",
     },
@@ -308,6 +313,32 @@ describe("Pre-execution checks → pauseAuto wiring", () => {
         String(call.arguments[0]).includes("Pre-execution checks failed")
     );
     assert.ok(errorNotify, "Should show error notification about pre-execution check failure");
+    const errorMessage = String(errorNotify.arguments[0]);
+    assert.match(
+      errorMessage,
+      /Pre-execution checks failed: \d+ blocking issue/,
+      "failure notification should include the blocking issue count",
+    );
+    assert.ok(
+      errorMessage.includes("[file] nonexistent-file-that-does-not-exist.ts: Task T01 references"),
+      "failure notification should include category, target, and message details",
+    );
+    assert.ok(
+      errorMessage.includes("[file] missing-third-file.ts: Task T01 references"),
+      "failure notification should include up to three actionable check details",
+    );
+    assert.ok(
+      !errorMessage.includes("missing-fourth-file.ts"),
+      "failure notification should truncate details beyond the display limit",
+    );
+    assert.ok(
+      errorMessage.includes("...and 1 more"),
+      "failure notification should summarize truncated blocking checks",
+    );
+    assert.ok(
+      errorMessage.includes(join(".gsd", "milestones", "M001", "slices", "S01", "S01-PRE-EXEC-VERIFY.json")),
+      "failure notification should point to the relative pre-exec evidence file path",
+    );
   });
 
   test("pauseAuto is called when enhanced_verification_strict: true and pre-execution returns warn", async () => {

package/src/resources/extensions/gsd/tests/preferences-worktree-sync.test.ts

@@ -12,6 +12,7 @@ import assert from "node:assert/strict";
 import { readFileSync, mkdtempSync, mkdirSync, writeFileSync, existsSync, readdirSync, rmSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
+import { extractSourceRegion } from "./test-helpers.ts";
 
 test("#2684: preferences files are NOT in ROOT_STATE_FILES (forward-only sync)", () => {
   const srcPath = join(import.meta.dirname, "..", "auto-worktree.ts");
@@ -49,7 +50,7 @@ test("copyPlanningArtifacts prefers canonical PREFERENCES.md with lowercase fall
   assert.ok(fnIdx !== -1, "copyPlanningArtifacts function exists");
 
   // Extract function body (up to the next top-level function)
-  const fnBody = src.slice(fnIdx, fnIdx + 2200);
+  const fnBody = extractSourceRegion(src, "function copyPlanningArtifacts");
 
   assert.ok(
     fnBody.includes("PROJECT_PREFERENCES_FILE") && fnBody.includes("LEGACY_PROJECT_PREFERENCES_FILE"),

package/src/resources/extensions/gsd/tests/prefs-wizard-coverage.test.ts

@@ -0,0 +1,44 @@
+// Guard test — every key in KNOWN_PREFERENCE_KEYS must be reachable from the
+// /gsd prefs wizard.  Without this guard, a new preference can be added to the
+// schema without anyone wiring it into the TUI, silently re-creating the gap
+// this test exists to prevent.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+
+import { KNOWN_PREFERENCE_KEYS } from "../preferences-types.ts";
+
+// Keys exposed via a dedicated command rather than the wizard.  They're still
+// reachable by the user, just not inside the category menu flow.  If you add a
+// new key here, add a comment explaining where it lives.
+const EXPOSED_OUTSIDE_WIZARD = new Set<string>([
+  "version",          // auto-managed by writePreferencesFile
+  "modelOverrides",   // advanced routing — edit PREFERENCES.md directly (not in KNOWN_PREFERENCE_KEYS)
+  "context_mode",     // advanced sandbox config (gsd_exec + compaction) — enabled by default; edit PREFERENCES.md directly to tune timeouts/caps. Wizard coverage tracked separately.
+]);
+
+test("every KNOWN_PREFERENCE_KEYS entry is reachable from the wizard source", () => {
+  const src = readFileSync(
+    new URL("../commands-prefs-wizard.ts", import.meta.url),
+    "utf-8",
+  );
+
+  const missing: string[] = [];
+  for (const key of KNOWN_PREFERENCE_KEYS) {
+    if (EXPOSED_OUTSIDE_WIZARD.has(key)) continue;
+    // The key must appear somewhere in the wizard — either as a direct
+    // prefs[...] / pref reference, or in the orderedKeys serialization list.
+    // A plain substring match is enough because all prefs-wizard references
+    // use the exact key name.
+    if (!src.includes(`"${key}"`) && !src.includes(`.${key}`)) {
+      missing.push(key);
+    }
+  }
+
+  assert.deepEqual(
+    missing,
+    [],
+    `These preference keys are in KNOWN_PREFERENCE_KEYS but are not referenced anywhere in the /gsd prefs wizard — they cannot be configured through the UI. Either add wizard coverage or add them to EXPOSED_OUTSIDE_WIZARD with an explanatory comment:\n${missing.join("\n")}`,
+  );
+});

package/src/resources/extensions/gsd/tests/prompt-cache-optimizer.test.ts

@@ -64,6 +64,18 @@ describe("prompt-cache-optimizer: classifySection", () => {
     assert.equal(classifySection("overrides"), "semi-static");
   });
 
+  // Regression: issue #4719 — KNOWLEDGE falls through to dynamic default.
+  // Knowledge content is reused across all tasks within a milestone, so it
+  // must be classified as semi-static to qualify for prefix caching when the
+  // cache optimizer is wired into the prompt path.
+  it("classifies knowledge as semi-static (issue #4719)", () => {
+    assert.equal(classifySection("knowledge"), "semi-static");
+  });
+
+  it("classifies project-knowledge as semi-static (issue #4719)", () => {
+    assert.equal(classifySection("project-knowledge"), "semi-static");
+  });
+
   it("classifies task-plan as dynamic", () => {
     assert.equal(classifySection("task-plan"), "dynamic");
   });

package/src/resources/extensions/gsd/tests/prompt-step-ordering.test.ts

@@ -13,6 +13,7 @@ import assert from 'node:assert/strict';
 import { readFileSync } from 'node:fs';
 import { fileURLToPath } from 'node:url';
 import { dirname, join } from 'node:path';
+import { extractSourceRegion } from "./test-helpers.ts";
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
@@ -69,11 +70,21 @@ describe('register-extension _gsdEpipeGuard (#3696)', () => {
 
 describe('register-hooks session_before_compact (#3696)', () => {
   test('session_before_compact only checks isAutoActive', () => {
-    // Extract the session_before_compact handler
-    const compactIdx = registerHooksSrc.indexOf('session_before_compact');
+    // Anchor on the full registration token rather than the bare event name —
+    // prevents matching unrelated substring occurrences.
+    const compactIdx = registerHooksSrc.indexOf('pi.on("session_before_compact"');
     assert.ok(compactIdx > -1, 'session_before_compact hook should exist');
-    // The first check in the handler should be isAutoActive(), not isAutoPaused()
-    const afterCompact = registerHooksSrc.slice(compactIdx, compactIdx + 300);
+    // The first check in the handler should be isAutoActive(), not isAutoPaused().
+    // Bound the region to this single handler — register-hooks.ts contains
+    // multiple pi.on("session_before_compact") handlers and a later handler
+    // legitimately references isAutoPaused.
+    const afterCompact = extractSourceRegion(
+      registerHooksSrc,
+      'pi.on("session_before_compact"',
+      'pi.on("',
+      // NB: endAnchor search starts AFTER the startAnchor, so the next
+      // pi.on("... matches the subsequent handler rather than this one.
+    );
     assert.match(afterCompact, /isAutoActive\(\)/,
       'session_before_compact should check isAutoActive()');
     // Should NOT block compaction when paused

package/src/resources/extensions/gsd/tests/provider-errors.test.ts

@@ -13,7 +13,12 @@ import { fileURLToPath } from "node:url";
 import { classifyError, isTransient, isTransientNetworkError } from "../error-classifier.ts";
 import { pauseAutoForProviderError } from "../provider-error-pause.ts";
 import { resumeAutoAfterProviderDelay } from "../bootstrap/provider-error-resume.ts";
+import { MAX_TRANSIENT_AUTO_RESUMES } from "../bootstrap/agent-end-recovery.ts";
 import { getNextFallbackModel } from "../preferences.ts";
+// Zero-import module — imported by path rather than through the package
+// barrel to avoid pulling the full AgentSession / @gsd/pi-ai dep graph into
+// this unit test (see #4837).
+import { RETRYABLE_ERROR_RE } from "../../../../../packages/pi-coding-agent/src/core/retryable-error-regex.ts";
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
@@ -404,6 +409,7 @@ test("resumeAutoAfterProviderDelay restarts paused auto-mode from the recorded b
         stepMode: true,
         basePath: "/tmp/project",
       }),
+      resetTransientRetryState: () => {},
       startAuto: async (_ctx, _pi, base, verboseMode, options) => {
         startCalls.push({ base, verboseMode, step: options?.step });
       },
@@ -428,6 +434,7 @@ test("resumeAutoAfterProviderDelay does not double-start when auto-mode is alrea
         stepMode: false,
         basePath: "/tmp/project",
       }),
+      resetTransientRetryState: () => {},
       startAuto: async () => {
         startCalls += 1;
       },
@@ -458,6 +465,7 @@ test("resumeAutoAfterProviderDelay leaves auto paused when no base path is avail
         stepMode: false,
         basePath: "",
       }),
+      resetTransientRetryState: () => {},
       startAuto: async () => {
         startCalls += 1;
       },
@@ -474,6 +482,35 @@ test("resumeAutoAfterProviderDelay leaves auto paused when no base path is avail
   ]);
 });
 
+test("resumeAutoAfterProviderDelay resets provider retry state without clearing session-timeout attempts", async () => {
+  const calls: string[] = [];
+
+  const result = await resumeAutoAfterProviderDelay(
+    {} as any,
+    { ui: { notify() {} } } as any,
+    {
+      getSnapshot: () => ({
+        active: false,
+        paused: true,
+        stepMode: false,
+        basePath: "/tmp/project",
+      }),
+      resetTransientRetryState: () => {
+        calls.push("reset-transient");
+      },
+      startAuto: async () => {
+        calls.push("start-auto");
+      },
+    },
+  );
+
+  assert.equal(result, "resumed");
+  assert.deepEqual(calls, [
+    "reset-transient",
+    "start-auto",
+  ]);
+});
+
 // ── Escalating backoff for transient errors (#1166) ─────────────────────────
 
 test("agent-end-recovery.ts tracks consecutive transient errors for escalating backoff", () => {
@@ -567,31 +604,6 @@ test("resetTransientRetryState is exported from agent-end-recovery.ts", () => {
   );
 });
 
-test("provider-error-resume.ts calls resetTransientRetryState before startAuto", () => {
-  const src = readFileSync(join(__dirname, "..", "bootstrap", "provider-error-resume.ts"), "utf-8");
-  assert.ok(
-    src.includes("resetTransientRetryState"),
-    "provider-error-resume.ts must import and call resetTransientRetryState",
-  );
-  // Ensure reset is called BEFORE startAuto — order matters
-  const resetIdx = src.indexOf("resetTransientRetryState()");
-  const startIdx = src.indexOf("await deps.startAuto(");
-  assert.ok(
-    resetIdx !== -1 && startIdx !== -1 && resetIdx < startIdx,
-    "resetTransientRetryState() must be called before deps.startAuto()",
-  );
-  // Session timeout counter must also be reset before startAuto
-  assert.ok(
-    src.includes("resetSessionTimeoutState"),
-    "provider-error-resume.ts must import and call resetSessionTimeoutState",
-  );
-  const sessionResetIdx = src.indexOf("resetSessionTimeoutState()");
-  assert.ok(
-    sessionResetIdx !== -1 && startIdx !== -1 && sessionResetIdx < startIdx,
-    "resetSessionTimeoutState() must be called before deps.startAuto()",
-  );
-});
-
 // ── Fix 2: Session creation timeout treated as transient in phases.ts ───────
 
 test("phases.ts handles timeout session-creation failures with pause instead of stopAuto", () => {
@@ -670,13 +682,12 @@ test("phases.ts resets session timeout counter on successful unit completion", (
 // ── Fix 3: MAX_TRANSIENT_AUTO_RESUMES raised to 8 ───────────────────────────
 
 test("MAX_TRANSIENT_AUTO_RESUMES is at least 8 for sustained overload resilience", () => {
-  const src = readFileSync(join(__dirname, "..", "bootstrap", "agent-end-recovery.ts"), "utf-8");
-  const match = src.match(/MAX_TRANSIENT_AUTO_RESUMES\s*=\s*(\d+)/);
-  assert.ok(match, "MAX_TRANSIENT_AUTO_RESUMES must be defined");
-  const value = Number(match![1]);
+  // Import the real constant rather than regex-scraping the source literal —
+  // this way the assertion cannot silently drift if the symbol is renamed or
+  // the value is moved. See #4837.
   assert.ok(
-    value >= 8,
-    `MAX_TRANSIENT_AUTO_RESUMES must be >= 8 for sustained overload resilience, got ${value}`,
+    MAX_TRANSIENT_AUTO_RESUMES >= 8,
+    `MAX_TRANSIENT_AUTO_RESUMES must be >= 8 for sustained overload resilience, got ${MAX_TRANSIENT_AUTO_RESUMES}`,
   );
 });
 
@@ -724,20 +735,22 @@ test("classifyError: 'context window' with 'exceed' is transient server", () =>
 // ── agent-session retryable regex handles server_error (#1166) ──────────────
 
 test("agent-session retryable error regex matches server_error (underscore)", () => {
-  // This regex is extracted from _isRetryableError in agent-session.ts.
-  // It must match both "server error" (space) and "server_error" (underscore)
-  // to properly classify Codex streaming errors as retryable.
-  // "temporarily backed off" intentionally excluded — see #3429
-  const retryableRegex = /overloaded|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|connection.?error|connection.?refused|other side closed|fetch failed|upstream.?connect|reset before headers|terminated|retry delay|network.?(?:is\s+)?unavailable|credentials.*expired|extra usage is required/i;
+  // Import the real regex from the retry-handler so this test can never
+  // silently drift from runtime behaviour. The regex must match both
+  // "server error" (space) and "server_error" (underscore) to properly
+  // classify Codex streaming errors as retryable.
+  // "temporarily backed off" is intentionally excluded — see #3429 / #4837.
 
   // server_error (with underscore — Codex streaming error format)
-  assert.ok(retryableRegex.test("Codex server_error: An error occurred"));
+  assert.ok(RETRYABLE_ERROR_RE.test("Codex server_error: An error occurred"));
   // server error (with space — traditional HTTP error format)
-  assert.ok(retryableRegex.test("server error occurred"));
+  assert.ok(RETRYABLE_ERROR_RE.test("server error occurred"));
   // internal_error (with underscore)
-  assert.ok(retryableRegex.test("internal_error: something went wrong"));
+  assert.ok(RETRYABLE_ERROR_RE.test("internal_error: something went wrong"));
   // internal error (with space)
-  assert.ok(retryableRegex.test("internal error"));
+  assert.ok(RETRYABLE_ERROR_RE.test("internal error"));
   // non-retryable errors must not match
-  assert.ok(!retryableRegex.test("model not found"));
+  assert.ok(!RETRYABLE_ERROR_RE.test("model not found"));
+  // "temporarily backed off" must NOT be matched (intentional exclusion #3429)
+  assert.ok(!RETRYABLE_ERROR_RE.test("temporarily backed off"));
 });