gsd-pi 2.76.0-dev.82e249f7b → 2.76.0-dev.97f5583d9

package/src/resources/extensions/gsd/tests/save-gate-result-render.test.ts

@@ -0,0 +1,95 @@
+/**
+ * Regression test suite for save_gate_result renderResult.
+ *
+ * Verifies that renderResult does not print "undefined: undefined" when
+ * `details` is empty, and that the error fallback does not produce a
+ * duplicated `Error: Error:` prefix when `content[0].text` already starts
+ * with `Error:`.
+ */
+
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { registerDbTools } from '../bootstrap/db-tools.ts';
+
+function makeMockPi() {
+  const tools: any[] = [];
+  return {
+    registerTool: (tool: any) => tools.push(tool),
+    tools,
+  } as any;
+}
+
+const fakeTheme = {
+  fg: (_color: string, text: string) => text,
+  bold: (text: string) => text,
+};
+
+function getSaveGateResultTool() {
+  const pi = makeMockPi();
+  registerDbTools(pi);
+  const tool = pi.tools.find((t: any) => t.name === 'gsd_save_gate_result');
+  assert.ok(tool, 'gsd_save_gate_result should be registered');
+  return tool;
+}
+
+test('save_gate_result renderResult falls back to content text when details is empty', () => {
+  const tool = getSaveGateResultTool();
+  const result = {
+    content: [{ type: 'text', text: 'Gate Q3 result saved: verdict=pass' }],
+    details: {},
+    isError: false,
+  };
+  const rendered = tool.renderResult(result, {}, fakeTheme);
+  const text = String(rendered.content ?? rendered.text ?? rendered);
+  assert.ok(!text.includes('undefined'), `got: ${text}`);
+  assert.ok(
+    text.includes('Gate Q3') || text.includes('verdict=pass'),
+    `expected content summary — got: ${text}`,
+  );
+});
+
+test('save_gate_result renderResult uses structured details when present', () => {
+  const tool = getSaveGateResultTool();
+  const result = {
+    content: [{ type: 'text', text: 'Gate Q3 result saved: verdict=flag' }],
+    details: { operation: 'save_gate_result', gateId: 'Q3', verdict: 'flag' },
+    isError: false,
+  };
+  const rendered = tool.renderResult(result, {}, fakeTheme);
+  const text = String(rendered.content ?? rendered.text ?? rendered);
+  assert.ok(text.includes('Q3'), `got: ${text}`);
+  assert.ok(text.includes('flag'), `got: ${text}`);
+  assert.ok(!text.includes('undefined'), `got: ${text}`);
+});
+
+test('save_gate_result renderResult shows error from content when details.error is missing', () => {
+  const tool = getSaveGateResultTool();
+  const result = {
+    content: [{ type: 'text', text: 'Error: Invalid gateId "Z1"' }],
+    details: {},
+    isError: true,
+  };
+  const rendered = tool.renderResult(result, {}, fakeTheme);
+  const text = String(rendered.content ?? rendered.text ?? rendered);
+  assert.ok(
+    text.includes('Invalid gateId') || text.includes('Error'),
+    `got: ${text}`,
+  );
+  assert.ok(!text.includes('undefined'), `got: ${text}`);
+});
+
+test('save_gate_result renderResult does not duplicate Error: prefix', () => {
+  const tool = getSaveGateResultTool();
+  const result = {
+    content: [{ type: 'text', text: 'Error: Invalid gateId "Z1"' }],
+    details: {},
+    isError: true,
+  };
+  const rendered = tool.renderResult(result, {}, fakeTheme);
+  const text = String(rendered.content ?? rendered.text ?? rendered);
+  assert.ok(
+    !/Error:\s*Error:/i.test(text),
+    `expected a single Error: prefix — got: ${text}`,
+  );
+  assert.ok(text.includes('Invalid gateId'), `got: ${text}`);
+});

package/src/resources/extensions/gsd/tests/session-start-footer.test.ts

@@ -1,19 +1,15 @@
 /**
  * session-start-footer.test.ts
  *
- * Verifies that register-hooks.ts suppresses the built-in footer by calling
- * ctx.ui.setFooter(hideFooter) in both session_start and session_switch when
- * isAutoActive() is true.
+ * Verifies that register-hooks.ts suppresses the gsd-health widget (not the
+ * built-in footer) when isAutoActive() is true, and that setFooter is never
+ * called by the extension in either session_start or session_switch.
  *
  * Testing strategy:
- *   Two layers:
- *   1. Source-code regression guard: ensures the guard and setFooter call are
- *      structurally present in register-hooks.ts for both event handlers.
- *      (node:test does not support mock.module without --experimental-test-module-mocks,
- *       so structural analysis is the correct approach here.)
+ *   1. Source-code regression guards: structural checks on register-hooks.ts.
  *   2. Behavioral integration test: fires the live session_start handler with a
- *      fake ctx when isAutoActive() is false (its default at test time) and
- *      confirms setFooter is NOT called — verifying the guard is conditional.
+ *      fake ctx when isAutoActive() is false (default) and confirms neither
+ *      setFooter nor setWidget("gsd-health") is called.
  *
  * Relates to #4314.
  */
@@ -35,16 +31,14 @@ const HOOKS_SOURCE = readFileSync(
 
 // ─── Source-code regression guards ──────────────────────────────────────────
 
-test("register-hooks.ts imports hideFooter from auto-dashboard", () => {
+test("register-hooks.ts does NOT import hideFooter", () => {
   assert.ok(
-    HOOKS_SOURCE.includes('import { hideFooter } from "../auto-dashboard.js"') ||
-    HOOKS_SOURCE.includes("import { hideFooter } from '../auto-dashboard.js'"),
-    "register-hooks.ts must import hideFooter from auto-dashboard.js",
+    !HOOKS_SOURCE.includes("hideFooter"),
+    "register-hooks.ts must not reference hideFooter — footer is no longer swapped in auto mode",
   );
 });
 
-test("session_start handler calls ctx.ui.setFooter(hideFooter) when isAutoActive()", () => {
-  // Locate the session_start handler body (up to the next pi.on call)
+test("session_start handler guards initHealthWidget with !isAutoActive()", () => {
   const sessionStartIdx = HOOKS_SOURCE.indexOf('"session_start"');
   assert.ok(sessionStartIdx > -1, "session_start handler must exist");
 
@@ -58,20 +52,23 @@ test("session_start handler calls ctx.ui.setFooter(hideFooter) when isAutoActive
     "session_start handler must call isAutoActive()",
   );
   assert.ok(
-    sessionStartBody.includes("ctx.ui.setFooter(hideFooter)"),
-    "session_start handler must call ctx.ui.setFooter(hideFooter)",
+    sessionStartBody.includes("initHealthWidget"),
+    "session_start handler must reference initHealthWidget",
+  );
+  assert.ok(
+    !sessionStartBody.includes("setFooter"),
+    "session_start handler must NOT call setFooter",
   );
 
-  // Guard must wrap the setFooter call
   const guardIdx = sessionStartBody.indexOf("isAutoActive()");
-  const setFooterIdx = sessionStartBody.indexOf("ctx.ui.setFooter(hideFooter)");
+  const healthIdx = sessionStartBody.indexOf("initHealthWidget");
   assert.ok(
-    guardIdx < setFooterIdx,
-    "isAutoActive() guard must appear before ctx.ui.setFooter(hideFooter) in session_start",
+    guardIdx < healthIdx,
+    "isAutoActive() guard must appear before initHealthWidget in session_start",
   );
 });
 
-test("session_switch handler calls ctx.ui.setFooter(hideFooter) when isAutoActive()", () => {
+test("session_switch handler suppresses gsd-health when isAutoActive()", () => {
   const sessionSwitchIdx = HOOKS_SOURCE.indexOf('"session_switch"');
   assert.ok(sessionSwitchIdx > -1, "session_switch handler must exist");
 
@@ -85,21 +82,18 @@ test("session_switch handler calls ctx.ui.setFooter(hideFooter) when isAutoActiv
     "session_switch handler must call isAutoActive()",
   );
   assert.ok(
-    sessionSwitchBody.includes("ctx.ui.setFooter(hideFooter)"),
-    "session_switch handler must call ctx.ui.setFooter(hideFooter)",
+    sessionSwitchBody.includes('setWidget("gsd-health", undefined)'),
+    "session_switch handler must call setWidget(\"gsd-health\", undefined) when auto is active",
   );
-
-  const guardIdx = sessionSwitchBody.indexOf("isAutoActive()");
-  const setFooterIdx = sessionSwitchBody.indexOf("ctx.ui.setFooter(hideFooter)");
   assert.ok(
-    guardIdx < setFooterIdx,
-    "isAutoActive() guard must appear before ctx.ui.setFooter(hideFooter) in session_switch",
+    !sessionSwitchBody.includes("setFooter"),
+    "session_switch handler must NOT call setFooter",
   );
 });
 
-// ─── Behavioral test: setFooter NOT called when auto-mode is inactive ────────
+// ─── Behavioral test: neither setFooter nor health suppression when auto inactive ─
 
-test("session_start does NOT call setFooter when isAutoActive() is false (default)", async (t) => {
+test("session_start does NOT call setFooter or suppress gsd-health when isAutoActive() is false", async (t) => {
   const dir = join(
     tmpdir(),
     `gsd-footer-test-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
@@ -114,6 +108,7 @@ test("session_start does NOT call setFooter when isAutoActive() is false (defaul
   });
 
   let setFooterCallCount = 0;
+  let healthWidgetHideCount = 0;
 
   const handlers = new Map<string, (event: unknown, ctx: any) => Promise<void> | void>();
   const pi = {
@@ -137,17 +132,14 @@ test("session_start does NOT call setFooter when isAutoActive() is false (defaul
       },
       setWorkingMessage: () => {},
       onTerminalInput: () => () => {},
-      setWidget: () => {},
+      setWidget: (key: string, value: unknown) => {
+        if (key === "gsd-health" && value === undefined) healthWidgetHideCount++;
+      },
     },
     sessionManager: { getSessionId: () => null },
     model: null,
   } as any);
 
-  // isAutoActive() is false at test time (no auto session started),
-  // so setFooter must not be called.
-  assert.equal(
-    setFooterCallCount,
-    0,
-    "setFooter must NOT be called when isAutoActive() returns false",
-  );
+  assert.equal(setFooterCallCount, 0, "setFooter must NOT be called when isAutoActive() is false");
+  assert.equal(healthWidgetHideCount, 0, "gsd-health must NOT be hidden when isAutoActive() is false");
 });

package/src/resources/extensions/gsd/tests/stash-queued-context-files.test.ts

@@ -378,3 +378,59 @@ test("#2505: back-to-back merges preserve queued CONTEXT files", () => {
     rmSync(repo, { recursive: true, force: true });
   }
 });
+
+// #4573: When `.gsd` is a gitignored symlink (ADR-002 layout) and the project
+// `.gitignore` contains `.gsd`, `git stash push --include-untracked -- <pathspec>`
+// fatals with "The following paths are ignored by one of your .gitignore files".
+// The prior tests used a symlinked `.gsd` but no `.gitignore`, so this failure
+// mode was invisible to CI. Fixture must include BOTH the symlink AND the
+// ignore rule to reproduce the bug on pre-fix code.
+test("#4573: gitignored .gsd symlink does not break pre-merge stash", () => {
+  const repo = realpathSync(mkdtempSync(join(tmpdir(), "wt-4573-ignored-symlink-")));
+  const stateDir = realpathSync(mkdtempSync(join(tmpdir(), "wt-4573-state-")));
+  try {
+    run("git init", repo);
+    run("git config user.email test@test.com", repo);
+    run("git config user.name Test", repo);
+    writeFileSync(join(repo, "README.md"), "# test\n");
+    // Matches what BASELINE_PATTERNS in gitignore.ts writes for real projects.
+    writeFileSync(join(repo, ".gitignore"), ".gsd\n.gsd-id\n");
+    symlinkSync(stateDir, join(repo, ".gsd"));
+    run("git add README.md .gitignore", repo);
+    run("git commit -m init", repo);
+    run("git branch -M main", repo);
+
+    const wtPath = createAutoWorktree(repo, "M001");
+    const worktreeName = wtPath.replaceAll("\\", "/").split("/").pop() || "M001";
+    const sliceBranch = `slice/${worktreeName}/S01`;
+    run(`git checkout -b "${sliceBranch}"`, wtPath);
+    writeFileSync(join(wtPath, "app.ts"), "export const app = true;\n");
+    run("git add app.ts", wtPath);
+    run('git commit -m "add feature"', wtPath);
+    run("git checkout milestone/M001", wtPath);
+    run(`git merge --no-ff "${sliceBranch}" -m "merge S01"`, wtPath);
+
+    // Dirty a tracked file so the pre-merge stash branch actually runs.
+    writeFileSync(join(repo, "README.md"), "# test\n\nDirty.\n");
+
+    const result = mergeMilestoneToMain(
+      repo,
+      "M001",
+      makeRoadmap("M001", "Feature", [{ id: "S01", title: "Feature" }]),
+    );
+
+    assert.ok(
+      result.commitMessage.includes("GSD-Milestone: M001"),
+      "merge must succeed despite gitignored .gsd symlink",
+    );
+    assert.ok(existsSync(join(repo, "app.ts")), "milestone code merged to main");
+    assert.equal(
+      lstatSync(join(repo, ".gsd")).isSymbolicLink(),
+      true,
+      ".gsd symlink remains in place",
+    );
+  } finally {
+    rmSync(repo, { recursive: true, force: true });
+    rmSync(stateDir, { recursive: true, force: true });
+  }
+});

package/src/resources/extensions/gsd/tests/token-counter.test.ts

@@ -102,7 +102,8 @@ describe("token-counter: estimateTokensForProvider", () => {
 
 describe("token-counter: backward compatibility", () => {
   it("countTokensSync returns heuristic estimate when tiktoken is not loaded", () => {
-    // Without tiktoken loaded, countTokensSync falls back to ceil(len/4)
+    // Without tiktoken loaded, countTokensSync falls back to estimateTokensForProvider.
+    // With no provider (defaults to "unknown", ratio 4.0): ceil(100/4) = 25.
     const text = "A".repeat(100);
     const result = countTokensSync(text);
     // Either tiktoken is loaded (exact count) or heuristic (ceil(100/4) = 25)
@@ -127,3 +128,106 @@ describe("token-counter: backward compatibility", () => {
     assert.equal(result, 0);
   });
 });
+
+// ─── provider-aware fallback (issue #4529) ───────────────────────────────────
+// Regression tests: countTokens/countTokensSync must use provider-specific
+// ratios for their heuristic fallback, not a hardcoded GPT-4o / 4 divisor.
+
+describe("token-counter: provider-aware heuristic fallback", () => {
+  // These tests exercise the heuristic path (no tiktoken or before init).
+  // We call estimateTokensForProvider directly to validate expected values,
+  // then verify countTokens/countTokensSync return the same values when
+  // tiktoken is unavailable.
+
+  it("countTokensSync uses anthropic ratio (3.5) when provider is 'anthropic'", () => {
+    const text = "A".repeat(350);
+    // anthropic: ceil(350 / 3.5) = 100
+    // openai/unknown: ceil(350 / 4.0) = 88
+    // These are different — the provider must matter.
+    const anthropicEstimate = estimateTokensForProvider(text, "anthropic");
+    const unknownEstimate = estimateTokensForProvider(text, "unknown");
+    assert.equal(anthropicEstimate, 100, "anthropic ratio should give 100 tokens for 350 chars");
+    assert.equal(unknownEstimate, 88, "unknown ratio should give 88 tokens for 350 chars");
+    assert.notEqual(
+      anthropicEstimate,
+      unknownEstimate,
+      "anthropic and unknown estimates must differ — if they are equal the provider is being ignored",
+    );
+
+    // Actually call countTokensSync with the anthropic provider.
+    // When tiktoken is not loaded, this must return the provider-aware estimate (100).
+    // When tiktoken is loaded, it returns the tiktoken count (which is also > 0 and
+    // will be in the range [88, 120] for 350 "A" characters with cl100k_base).
+    const syncResult = countTokensSync(text, "anthropic");
+    assert.ok(typeof syncResult === "number", "countTokensSync must return a number");
+    assert.ok(syncResult > 0, "countTokensSync must return a positive count");
+    // If tiktoken is unavailable the result must exactly match the anthropic heuristic.
+    // If tiktoken is available we cannot assert the exact value, but we know it will
+    // not equal the unknown-provider heuristic (88) for 350 identical characters.
+    const tiktokenAvailable = syncResult !== anthropicEstimate;
+    if (!tiktokenAvailable) {
+      assert.equal(
+        syncResult,
+        anthropicEstimate,
+        `countTokensSync with 'anthropic' provider must return ${anthropicEstimate} (not the unknown-provider value ${unknownEstimate}) when tiktoken is unavailable`,
+      );
+    }
+  });
+
+  it("countTokens uses anthropic ratio when provider='anthropic' and tiktoken unavailable", async () => {
+    const text = "A".repeat(350);
+    // anthropic heuristic: ceil(350 / 3.5) = 100
+    // unknown/hardcoded-4 heuristic: ceil(350 / 4.0) = 88
+    const anthropicEstimate = estimateTokensForProvider(text, "anthropic"); // 100
+    const unknownEstimate = estimateTokensForProvider(text, "unknown");     // 88
+    const hardcodedFallback = Math.ceil(text.length / 4);                   // 88
+
+    // The anthropic heuristic must produce more tokens than the old /4 fallback.
+    assert.equal(anthropicEstimate, 100, "anthropic heuristic should give 100 tokens for 350 chars");
+    assert.ok(
+      anthropicEstimate > hardcodedFallback,
+      `anthropic estimate (${anthropicEstimate}) must exceed the hardcoded /4 fallback (${hardcodedFallback})`,
+    );
+    assert.notEqual(
+      anthropicEstimate,
+      unknownEstimate,
+      "anthropic and unknown estimates must differ — provider is being ignored if equal",
+    );
+
+    // Call countTokens with the anthropic provider and verify the result.
+    const result = await countTokens(text, "anthropic");
+    assert.ok(typeof result === "number", "should return a number");
+    assert.ok(result > 0, "should return a positive token count");
+
+    // When tiktoken is unavailable: result must equal the anthropic heuristic (100),
+    // NOT the unknown-provider heuristic (88). This is the core regression guard for #4529.
+    // When tiktoken IS available: result is the tiktoken count, which will differ from
+    // the heuristic — but it must never equal the wrong (unknown) heuristic for this text.
+    if (result === anthropicEstimate || result === unknownEstimate) {
+      // We are on the heuristic path — assert the correct provider ratio was used.
+      assert.equal(
+        result,
+        anthropicEstimate,
+        `countTokens with 'anthropic' provider returned ${result} but expected ${anthropicEstimate} (anthropic heuristic) not ${unknownEstimate} (unknown/hardcoded heuristic)`,
+      );
+    } else {
+      // tiktoken is active — result is an exact BPE count.
+      // For 350 identical "A" characters cl100k_base produces a count in [80, 130].
+      assert.ok(
+        result >= 80 && result <= 130,
+        `tiktoken count ${result} for 350 "A" chars is outside expected range [80, 130]`,
+      );
+    }
+  });
+
+  it("countTokens with provider='anthropic' yields more tokens than provider='openai' (heuristic)", () => {
+    const text = "A".repeat(400);
+    // anthropic: ceil(400/3.5) = 115, openai: ceil(400/4.0) = 100
+    const anthropic = estimateTokensForProvider(text, "anthropic");
+    const openai = estimateTokensForProvider(text, "openai");
+    assert.ok(
+      anthropic > openai,
+      `anthropic estimate (${anthropic}) must exceed openai estimate (${openai}) for same text`,
+    );
+  });
+});

package/src/resources/extensions/gsd/tests/tool-compatibility.test.ts

@@ -14,6 +14,7 @@ import {
   isToolCompatibleWithProvider,
   filterToolsForProvider,
   adjustToolSet,
+  GROQ_MAX_TOOLS,
 } from "../model-router.js";
 
 import {
@@ -197,3 +198,109 @@ describe("adjustToolSet", () => {
     assert.deepEqual(removedTools, ["mcp_complex"]);
   });
 });
+
+// ─── GROQ_MAX_TOOLS constant ─────────────────────────────────────────────────
+
+describe("GROQ_MAX_TOOLS", () => {
+  test("equals 128", () => {
+    assert.equal(GROQ_MAX_TOOLS, 128);
+  });
+});
+
+// ─── Groq tool-count cap (#4376) ────────────────────────────────────────────
+
+describe("filterToolsForProvider — Groq 128-tool cap", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("does not cap when provider is not groq", () => {
+    const toolNames = Array.from({ length: 200 }, (_, i) => `tool_${i}`);
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "openai-completions");
+    assert.equal(compatible.length, 200);
+    assert.equal(filtered.length, 0);
+  });
+
+  test("does not cap when <= 128 tools with groq provider", () => {
+    const toolNames = Array.from({ length: 128 }, (_, i) => `tool_${i}`);
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "openai-completions", "groq");
+    assert.equal(compatible.length, 128);
+    assert.equal(filtered.length, 0);
+  });
+
+  test("caps to 128 when >128 tools with groq provider", () => {
+    const toolNames = Array.from({ length: 200 }, (_, i) => `tool_${i}`);
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "openai-completions", "groq");
+    assert.equal(compatible.length, 128);
+    assert.equal(filtered.length, 72);
+  });
+
+  test("keeps the first 128 tools when capping", () => {
+    const toolNames = Array.from({ length: 200 }, (_, i) => `tool_${i}`);
+    const { compatible } = filterToolsForProvider(toolNames, "openai-completions", "groq");
+    assert.equal(compatible[0], "tool_0");
+    assert.equal(compatible[127], "tool_127");
+  });
+
+  test("trimmed tools appear in filtered list", () => {
+    const toolNames = Array.from({ length: 130 }, (_, i) => `tool_${i}`);
+    const { filtered } = filterToolsForProvider(toolNames, "openai-completions", "groq");
+    assert.equal(filtered.length, 2);
+    assert.equal(filtered[0], "tool_128");
+    assert.equal(filtered[1], "tool_129");
+  });
+
+  test("emits a warning when tools are trimmed", () => {
+    const warnings: string[] = [];
+    const original = console.warn;
+    console.warn = (...args: unknown[]) => { warnings.push(String(args[0])); };
+    try {
+      const toolNames = Array.from({ length: 129 }, (_, i) => `tool_${i}`);
+      filterToolsForProvider(toolNames, "openai-completions", "groq");
+      assert.equal(warnings.length, 1);
+      assert.ok(warnings[0].includes("128"), "warning mentions Groq limit");
+    } finally {
+      console.warn = original;
+    }
+  });
+
+  test("does not warn when tools are at the limit", () => {
+    const warnings: string[] = [];
+    const original = console.warn;
+    console.warn = (...args: unknown[]) => { warnings.push(String(args[0])); };
+    try {
+      const toolNames = Array.from({ length: 128 }, (_, i) => `tool_${i}`);
+      filterToolsForProvider(toolNames, "openai-completions", "groq");
+      assert.equal(warnings.length, 0);
+    } finally {
+      console.warn = original;
+    }
+  });
+});
+
+describe("adjustToolSet — Groq 128-tool cap", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("caps to 128 tools when provider is groq and >128 tools active", () => {
+    const toolNames = Array.from({ length: 150 }, (_, i) => `tool_${i}`);
+    const { toolNames: result, removedTools } = adjustToolSet(toolNames, "openai-completions", "groq");
+    assert.equal(result.length, 128);
+    assert.equal(removedTools.length, 22);
+  });
+
+  test("does not cap for non-groq providers even with >128 tools", () => {
+    const toolNames = Array.from({ length: 150 }, (_, i) => `tool_${i}`);
+    const { toolNames: result, removedTools } = adjustToolSet(toolNames, "openai-completions", "openai");
+    assert.equal(result.length, 150);
+    assert.equal(removedTools.length, 0);
+  });
+
+  test("does not cap when provider is omitted", () => {
+    const toolNames = Array.from({ length: 150 }, (_, i) => `tool_${i}`);
+    const { toolNames: result, removedTools } = adjustToolSet(toolNames, "openai-completions");
+    assert.equal(result.length, 150);
+    assert.equal(removedTools.length, 0);
+  });
+});

package/src/resources/extensions/gsd/tests/uok-plan-v2-wiring.test.ts

@@ -109,6 +109,29 @@ test("plan-v2 gate fails closed for execution phase when finalized context is mi
   assert.match(compiled.reason ?? "", /CONTEXT\.md/i);
 });
 
+test("plan-v2 gate accepts finalized context from project-root fallback", () => {
+  const projectRoot = createBasePath();
+  const worktreeBase = createBasePath();
+  seedGraphRows();
+
+  writeMilestoneFile(projectRoot, "CONTEXT", "Finalized context in project root.");
+  writeMilestoneFile(worktreeBase, "CONTEXT-DRAFT", "Draft context in worktree.");
+
+  const prevProjectRoot = process.env.GSD_PROJECT_ROOT;
+  process.env.GSD_PROJECT_ROOT = projectRoot;
+  try {
+    const compiled = ensurePlanV2Graph(worktreeBase, buildState("executing"));
+    assert.equal(compiled.ok, true);
+    assert.equal(compiled.finalizedContextIncluded, true);
+  } finally {
+    if (prevProjectRoot === undefined) {
+      delete process.env.GSD_PROJECT_ROOT;
+    } else {
+      process.env.GSD_PROJECT_ROOT = prevProjectRoot;
+    }
+  }
+});
+
 test("plan-v2 compiler writes pipeline metadata for clarify/research/draft stages", () => {
   const basePath = createBasePath();
   seedGraphRows();

package/src/resources/extensions/gsd/tests/workflow-tool-executors.test.ts

@@ -1,6 +1,6 @@
 import test from "node:test";
 import assert from "node:assert/strict";
-import { mkdirSync, rmSync, readFileSync, existsSync, writeFileSync } from "node:fs";
+import { mkdirSync, rmSync, readFileSync, existsSync, writeFileSync, unlinkSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
 import { randomUUID } from "node:crypto";
@@ -11,7 +11,7 @@ import {
   _getAdapter,
   insertGateRow,
 } from "../gsd-db.ts";
-import { markDepthVerified, clearDiscussionFlowState } from "../bootstrap/write-gate.ts";
+import { markDepthVerified, clearDiscussionFlowState, loadWriteGateSnapshot } from "../bootstrap/write-gate.ts";
 import {
   executeCompleteMilestone,
   executePlanMilestone,
@@ -742,3 +742,66 @@ test("executeSummarySave leaves sibling CONTEXT-DRAFT intact for non-CONTEXT art
     cleanup(base);
   }
 });
+
+test("executeSummarySave CONTEXT HARD BLOCK clears after write-gate state file is deleted (#4343)", async () => {
+  const base = makeTmpBase();
+  const originalEnv = process.env.GSD_PERSIST_WRITE_GATE_STATE;
+  process.env.GSD_PERSIST_WRITE_GATE_STATE = "1";
+  try {
+    openTestDb(base);
+    clearDiscussionFlowState();
+
+    // First call: CONTEXT artifact without depth verification → HARD BLOCK
+    const blocked = await inProjectDir(base, () => executeSummarySave({
+      milestone_id: "M001",
+      artifact_type: "CONTEXT",
+      content: "# Context\n\ncontent",
+    }, base));
+    assert.equal(blocked.isError, true, "should be blocked without depth verification");
+    assert.match(
+      blocked.content[0].text,
+      /HARD BLOCK/,
+      "blocked result should mention HARD BLOCK",
+    );
+
+    // Verify the state file was written (persist mode is active)
+    const stateFilePath = join(base, ".gsd", "runtime", "write-gate-state.json");
+    // The state file may or may not exist at this point (block doesn't write state).
+    // Write a fake state file simulating stale persisted block state.
+    mkdirSync(join(base, ".gsd", "runtime"), { recursive: true });
+    writeFileSync(stateFilePath, JSON.stringify({
+      verifiedDepthMilestones: [],
+      activeQueuePhase: false,
+      pendingGateId: "depth_verification_M001",
+    }));
+
+    // User deletes the state file to reset the block
+    unlinkSync(stateFilePath);
+    assert.ok(!existsSync(stateFilePath), "state file deleted");
+
+    // The snapshot loaded after deletion should be clean (no pending gate, no block)
+    const snapshot = loadWriteGateSnapshot(base);
+    assert.equal(snapshot.pendingGateId, null, "pendingGateId should be null after file deletion");
+    assert.deepEqual(snapshot.verifiedDepthMilestones, [], "verifiedDepthMilestones should be empty after file deletion");
+
+    // Depth-verify and re-attempt: should succeed after deletion clears stale state
+    markDepthVerified("M001", base);
+
+    const unblocked = await inProjectDir(base, () => executeSummarySave({
+      milestone_id: "M001",
+      artifact_type: "CONTEXT",
+      content: "# Context\n\nfinal content",
+    }, base));
+    assert.equal(unblocked.isError, undefined, "should not be blocked after depth verification");
+    assert.equal(unblocked.details.operation, "save_summary");
+  } finally {
+    if (originalEnv === undefined) {
+      delete process.env.GSD_PERSIST_WRITE_GATE_STATE;
+    } else {
+      process.env.GSD_PERSIST_WRITE_GATE_STATE = originalEnv;
+    }
+    clearDiscussionFlowState();
+    closeDatabase();
+    cleanup(base);
+  }
+});