@opengsd/gsd-pi 1.1.1-dev.74e8dd1 → 1.1.1-dev.75048e7

package/src/resources/extensions/gsd/tests/post-unit-retry-on-orchestrator-bridge.test.ts

@@ -8,6 +8,7 @@ import { mock } from "node:test";
 import { postUnitPostVerification, type PostUnitContext } from "../auto-post-unit.ts";
 import { AutoSession } from "../auto/session.ts";
 import { checkPostUnitHooks, resetHookState, resolveHookArtifactPath } from "../post-unit-hooks.ts";
+import { emitJournalEvent } from "../journal.ts";
 import { _clearGsdRootCache } from "../paths.ts";
 import { invalidateAllCaches } from "../cache.ts";
 
@@ -28,6 +29,43 @@ post_unit_hooks:
   writeFileSync(join(basePath, ".gsd", "PREFERENCES.md"), content, "utf-8");
 }
 
+function writeFailingHookPreferences(basePath: string): void {
+  const content = `---
+post_unit_hooks:
+  - name: review-arbiter
+    after:
+      - execute-task
+    prompt: Review {taskId}
+    artifact: REVIEW-DEBATE.md
+    max_cycles: 1
+    enabled: true
+  - name: follow-up-review
+    after:
+      - execute-task
+    prompt: Follow-up review {taskId}
+    enabled: true
+---
+`;
+  writeFileSync(join(basePath, ".gsd", "PREFERENCES.md"), content, "utf-8");
+}
+
+function writeBlockingPreferences(basePath: string): void {
+  const content = `---
+post_unit_hooks:
+  - name: review-arbiter
+    after:
+      - execute-task
+    prompt: Review {taskId}
+    agent: arbiter
+    artifact: REVIEW-DEBATE.md
+    criticality: blocking
+    max_cycles: 2
+    enabled: true
+---
+`;
+  writeFileSync(join(basePath, ".gsd", "PREFERENCES.md"), content, "utf-8");
+}
+
 test("post-unit retry_on marks trigger unit as retry in orchestrator before redispatch", async () => {
   const originalCwd = process.cwd();
   const base = mkdtempSync(join(tmpdir(), "gsd-post-unit-retry-"));
@@ -91,3 +129,144 @@ test("post-unit retry_on marks trigger unit as retry in orchestrator before redi
     rmSync(base, { recursive: true, force: true });
   }
 });
+
+test("failed post-unit hook pauses auto-mode even when its artifact exists", async () => {
+  const originalCwd = process.cwd();
+  const base = mkdtempSync(join(tmpdir(), "gsd-post-unit-hook-failed-"));
+  const taskDir = join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks");
+  mkdirSync(taskDir, { recursive: true });
+
+  try {
+    process.chdir(base);
+    _clearGsdRootCache();
+    invalidateAllCaches();
+    resetHookState();
+    writeFailingHookPreferences(base);
+
+    const hookDispatch = checkPostUnitHooks("execute-task", "M001/S01/T01", base);
+    assert.equal(hookDispatch?.hookName, "review-arbiter");
+
+    const artifactPath = resolveHookArtifactPath(base, "M001/S01/T01", "REVIEW-DEBATE.md");
+    writeFileSync(artifactPath, "partial review", "utf-8");
+    emitJournalEvent(base, {
+      ts: "2026-06-03T12:00:00.000Z",
+      flowId: "flow-hook-failed",
+      seq: 3,
+      eventType: "unit-end",
+      data: {
+        unitType: "hook/review-arbiter",
+        unitId: "M001/S01/T01",
+        status: "cancelled",
+        artifactVerified: false,
+      },
+    });
+
+    const pauseAuto = mock.fn(async () => {});
+    const notifications: string[] = [];
+    const s = new AutoSession();
+    s.basePath = base;
+    s.active = true;
+    s.currentUnit = { type: "hook/review-arbiter", id: "M001/S01/T01", startedAt: Date.now() };
+
+    const pctx: PostUnitContext = {
+      s,
+      ctx: {
+        ui: {
+          notify: (message: string) => { notifications.push(message); },
+          setStatus: () => {},
+          setWidget: () => {},
+          setFooter: () => {},
+        },
+        model: { id: "test-model" },
+      } as any,
+      pi: { sendMessage: async () => {}, setModel: async () => true } as any,
+      buildSnapshotOpts: () => ({}),
+      lockBase: () => base,
+      stopAuto: async () => {},
+      pauseAuto,
+      updateProgressWidget: () => {},
+    };
+
+    const result = await postUnitPostVerification(pctx);
+    assert.equal(result, "stopped");
+    assert.equal(pauseAuto.mock.callCount(), 1);
+    assert.ok(
+      notifications.some(message => message.includes("Post-unit hook review-arbiter failed")),
+      "pause notification should explain the failed hook",
+    );
+  } finally {
+    process.chdir(originalCwd);
+    resetHookState();
+    invalidateAllCaches();
+    _clearGsdRootCache();
+    rmSync(base, { recursive: true, force: true });
+  }
+});
+
+test("post-unit blocking gate pauses auto-mode on needs-attention verdict", async () => {
+  const originalCwd = process.cwd();
+  const base = mkdtempSync(join(tmpdir(), "gsd-post-unit-gate-"));
+  const taskDir = join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks");
+  mkdirSync(taskDir, { recursive: true });
+
+  try {
+    process.chdir(base);
+    _clearGsdRootCache();
+    invalidateAllCaches();
+    resetHookState();
+    writeBlockingPreferences(base);
+
+    const hookDispatch = checkPostUnitHooks("execute-task", "M001/S01/T01", base);
+    assert.ok(hookDispatch, "hook should dispatch for execute-task");
+
+    const artifactPath = resolveHookArtifactPath(base, "M001/S01/T01", "REVIEW-DEBATE.md");
+    writeFileSync(artifactPath, "---\nverdict: needs-attention\n---\n\nManual review required.\n", "utf-8");
+
+    const pauseAuto = mock.fn(async () => {});
+    const s = new AutoSession();
+    s.basePath = base;
+    s.active = true;
+    s.currentUnit = { type: "hook/review-arbiter", id: "M001/S01/T01", startedAt: Date.now() };
+    s.orchestration = {
+      start: async () => ({ kind: "started" }),
+      advance: async () => ({ kind: "stopped", reason: "unused" }),
+      completeActiveUnit: async () => {},
+      retryActiveUnit: async () => {},
+      resume: async () => ({ kind: "resumed" }),
+      stop: async (reason: string) => ({ kind: "stopped", reason }),
+      getStatus: () => ({ phase: "running", transitionCount: 0 }),
+    };
+
+    const notifications: string[] = [];
+    const pctx: PostUnitContext = {
+      s,
+      ctx: {
+        ui: {
+          notify: (message: string) => { notifications.push(message); },
+          setStatus: () => {},
+          setWidget: () => {},
+          setFooter: () => {},
+        },
+        model: { id: "test-model" },
+      } as any,
+      pi: { sendMessage: async () => {}, setModel: async () => true } as any,
+      buildSnapshotOpts: () => ({}),
+      lockBase: () => base,
+      stopAuto: async () => {},
+      pauseAuto,
+      updateProgressWidget: () => {},
+    };
+
+    const result = await postUnitPostVerification(pctx);
+    assert.equal(result, "stopped");
+    assert.equal(pauseAuto.mock.callCount(), 1);
+    assert.match(notifications.join("\n"), /Post-unit gate "review-arbiter" blocked execute-task M001\/S01\/T01/);
+    assert.match(notifications.join("\n"), /\/gsd status/);
+  } finally {
+    process.chdir(originalCwd);
+    resetHookState();
+    invalidateAllCaches();
+    _clearGsdRootCache();
+    rmSync(base, { recursive: true, force: true });
+  }
+});

package/src/resources/extensions/gsd/tests/preferences.test.ts

@@ -561,6 +561,35 @@ test("post-unit hook max_cycles clamping via validatePreferences", () => {
   assert.equal(p4.post_unit_hooks![0].max_cycles, 3, "valid value passes through");
 });
 
+test("post-unit hook criticality and on_block validation", () => {
+  const base = { name: "h", after: ["execute-task"], prompt: "do something", artifact: "REVIEW.md" };
+
+  const { preferences, errors } = validatePreferences({
+    post_unit_hooks: [{
+      ...base,
+      criticality: "blocking",
+      on_block: { action: "retry-unit", artifact: "NEEDS-REWORK.md" },
+    }],
+  } as any);
+  assert.equal(errors.length, 0);
+  assert.equal(preferences.post_unit_hooks![0].criticality, "blocking");
+  assert.deepEqual(preferences.post_unit_hooks![0].on_block, {
+    action: "retry-unit",
+    artifact: "NEEDS-REWORK.md",
+  });
+
+  const missingArtifact = validatePreferences({
+    post_unit_hooks: [{ name: "blocking", after: ["execute-task"], prompt: "do something", criticality: "blocking" }],
+  } as any);
+  assert.match(missingArtifact.errors.join("\n"), /criticality blocking requires artifact/);
+  assert.equal(missingArtifact.preferences.post_unit_hooks, undefined);
+
+  const invalidAction = validatePreferences({
+    post_unit_hooks: [{ ...base, on_block: { action: "teleport" } }],
+  } as any);
+  assert.match(invalidAction.errors.join("\n"), /invalid on_block action/);
+});
+
 test("pre-dispatch hook action validation via validatePreferences", () => {
   const base = { name: "h", before: ["execute-task"] };
 

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

@@ -2,6 +2,14 @@ import test from "node:test";
 import assert from "node:assert/strict";
 import { readFileSync } from "node:fs";
 import { join } from "node:path";
+import {
+  RUN_UAT_BROWSER_TOOL_NAMES,
+  buildRunUatResultPresentation,
+  buildRunUatPresentationForType,
+  RUN_UAT_READ_ONLY_TOOL_NAMES,
+  RUN_UAT_TOOL_PRESENTATION_PLAN_ID,
+  RUN_UAT_WORKFLOW_TOOL_NAMES,
+} from "../tool-presentation-plan.ts";
 
 const promptsDir = join(process.cwd(), "src/resources/extensions/gsd/prompts");
 const templatesDir = join(process.cwd(), "src/resources/extensions/gsd/templates");
@@ -25,11 +33,15 @@ test("reactive-execute prompt keeps task summaries with subagents and avoids bat
 test("run-uat prompt branches on dynamic UAT mode and supports runtime evidence", () => {
   const prompt = readPrompt("run-uat");
   assert.match(prompt, /\*\*Detected UAT mode:\*\*\s*`\{\{uatType\}\}`/);
-  assert.match(prompt, /uatType:\s*\{\{uatType\}\}/);
+  assert.match(prompt, /uatType:\s*"\{\{uatType\}\}"/);
+  assert.match(prompt, /gsd_uat_result_save/);
+  assert.match(prompt, /presentedTools/);
+  assert.match(prompt, /\{\{canonicalPresentation\}\}/);
   assert.match(prompt, /live-runtime/);
   assert.match(prompt, /browser\/runtime\/network/i);
   assert.match(prompt, /NEEDS-HUMAN/);
   assert.doesNotMatch(prompt, /uatType:\s*artifact-driven/);
+  assert.doesNotMatch(prompt, /Call `gsd_summary_save`/);
 });
 
 test("run-uat prompt lists canonical gsd_uat_exec intent values", () => {
@@ -42,6 +54,45 @@ test("run-uat prompt lists canonical gsd_uat_exec intent values", () => {
   assert.match(prompt, /do not use `artifact`, `runtime`, or `human-follow-up` as `intent`/i);
 });
 
+test("run-uat prompt gives the complete UAT result-save presentation contract", () => {
+  const prompt = readPrompt("run-uat");
+  assert.match(prompt, /Call `gsd_uat_result_save` once after all checks are complete/);
+  assert.doesNotMatch(prompt, /Call `gsd_summary_save` with `artifact_type: "ASSESSMENT"`/);
+  assert.match(prompt, /\{\{canonicalPresentation\}\}/);
+  assert.match(prompt, /\{\{toolPresentationPlanId\}\}/);
+
+  const presentation = buildRunUatResultPresentation();
+  assert.equal(presentation.toolPresentationPlanId, RUN_UAT_TOOL_PRESENTATION_PLAN_ID);
+  for (const toolName of RUN_UAT_WORKFLOW_TOOL_NAMES) {
+    assert.ok(presentation.presentedTools.includes(toolName), `presentation should include required tool ${toolName}`);
+  }
+  for (const toolName of RUN_UAT_READ_ONLY_TOOL_NAMES) {
+    assert.ok(presentation.presentedTools.includes(toolName), `presentation should include read-only tool ${toolName}`);
+  }
+
+  for (const toolName of ["gsd_exec", "gsd_summary_save", "gsd_save_gate_result"] as const) {
+    assert.ok(
+      presentation.blockedTools.some((entry) => entry.name === toolName),
+      `presentation should include blocked tool ${toolName}`,
+    );
+  }
+
+  assert.ok(
+    presentation.blockedTools.every((entry) => entry.reason === "forbidden during run-uat"),
+    "presentation should explain blocked run-uat tools",
+  );
+});
+
+test("browser-executable UAT presentation uses direct managed browser tools", () => {
+  const presentation = buildRunUatPresentationForType("browser-executable");
+
+  assert.equal(presentation.surface, "hybrid");
+  for (const toolName of RUN_UAT_BROWSER_TOOL_NAMES) {
+    assert.ok(presentation.presentedTools.includes(toolName), `presentation should include browser tool ${toolName}`);
+  }
+  assert.ok(!presentation.presentedTools.some((toolName) => toolName.startsWith("mcp__gsd-browser__")));
+});
+
 test("workflow-start prompt defaults to autonomy instead of per-phase confirmation", () => {
   const prompt = readPrompt("workflow-start");
   assert.match(prompt, /Keep moving by default/i);
@@ -475,6 +526,27 @@ test("reactive-execute prompt references tool calls instead of checkbox updates"
   assert.match(prompt, /completion tool calls/);
 });
 
+test("parallel subagent prompts forbid serialized tasks arrays", () => {
+  const expectations = [
+    { name: "reactive-execute", agent: "worker" },
+    { name: "parallel-research-slices", agent: "scout" },
+    { name: "gate-evaluate", agent: "tester" },
+  ] as const;
+
+  for (const { name, agent } of expectations) {
+    const prompt = readPrompt(name);
+    assert.match(prompt, /tasks:\s*\[\.\.\.\]/, `${name} must show the native array placeholder`);
+    assert.match(prompt, /native JSON array/i, `${name} must explicitly require a native JSON array`);
+    assert.match(prompt, /Do NOT JSON\.stringify/i, `${name} must forbid JSON.stringify on tasks`);
+    assert.match(prompt, /must be array/i, `${name} must mention the subagent validation failure`);
+    assert.match(
+      prompt,
+      new RegExp(`tasks:\\s*\\[\\{\\s*agent:\\s*"${agent}"`, "i"),
+      `${name} must show the concrete ${agent} task call shape`,
+    );
+  }
+});
+
 // ─── Project-shape classifier + 3-or-4-options-with-Other-hatch contract ──
 
 test("guided-discuss-project classifies project shape and persists the verdict to PROJECT.md", () => {

package/src/resources/extensions/gsd/tests/prompt-loader-extension-dir.test.ts

@@ -22,6 +22,20 @@ test("resolveExtensionDirFromCandidates prefers user-local dir when both trees a
   assert.equal(resolved, agentDir);
 });
 
+test("resolveExtensionDirFromCandidates prefers source checkout dir when both trees are valid", () => {
+  const moduleDir = "/repo/src/resources/extensions/gsd";
+  const agentDir = "/home/user/.gsd/agent/extensions/gsd";
+  const paths = new Set<string>([
+    join(moduleDir, "prompts"),
+    join(moduleDir, "templates", "task-summary.md"),
+    join(agentDir, "prompts"),
+    join(agentDir, "templates", "task-summary.md"),
+  ]);
+
+  const resolved = resolveExtensionDirFromCandidates(moduleDir, agentDir, makeExists(paths));
+  assert.equal(resolved, moduleDir);
+});
+
 test("resolveExtensionDirFromCandidates rejects module dir missing task-summary template", () => {
   const moduleDir = "/npm/global/gsd";
   const agentDir = "/home/user/.gsd/agent/extensions/gsd";

package/src/resources/extensions/gsd/tests/queued-discuss-fast-path.test.ts

@@ -24,25 +24,24 @@ describe("queued-discuss-fast-path", () => {
     );
   });
 
-  test("2. dispatchDiscussForMilestone computes fastPathInstruction and passes it to loadPrompt", () => {
+  test("2. dispatchDiscussForMilestone computes fastPathInstruction and passes it to buildDiscussMilestonePrompt", () => {
     const source = guidedFlowSrc();
     const fnStart = source.indexOf("async function dispatchDiscussForMilestone(");
     assert.ok(fnStart > 0, "dispatchDiscussForMilestone must exist");
-    const fnEnd = source.indexOf("\nasync function ", fnStart + 1);
     const fnBody = extractSourceRegion(source, "async function dispatchDiscussForMilestone(");
     assert.ok(
       fnBody.includes("fastPathInstruction"),
       "dispatchDiscussForMilestone must compute fastPathInstruction",
     );
     assert.ok(
-      fnBody.includes("loadPrompt("),
-      "dispatchDiscussForMilestone must call loadPrompt",
+      fnBody.includes("buildDiscussMilestonePrompt("),
+      "dispatchDiscussForMilestone must call buildDiscussMilestonePrompt",
     );
-    const loadPromptIdx = fnBody.indexOf("loadPrompt(");
-    const fastPathIdx = fnBody.indexOf("fastPathInstruction", loadPromptIdx);
+    const buildPromptIdx = fnBody.indexOf("buildDiscussMilestonePrompt(");
+    const fastPathIdx = fnBody.indexOf("fastPathInstruction", buildPromptIdx);
     assert.ok(
-      fastPathIdx > loadPromptIdx,
-      "fastPathInstruction must be passed to loadPrompt in dispatchDiscussForMilestone",
+      fastPathIdx > buildPromptIdx,
+      "fastPathInstruction must be passed to buildDiscussMilestonePrompt in dispatchDiscussForMilestone",
     );
   });
 

package/src/resources/extensions/gsd/tests/register-hooks-depth-verification.test.ts

@@ -70,6 +70,50 @@ test("register-hooks hard-blocks destructive bash commands outside auto-mode", a
   assert.match(block?.reason ?? "", /IaC apply\/destroy/);
 });
 
+test("register-hooks keeps depth-gate reason model-facing and adds displayReason", async (t) => {
+  const dir = makeTempDir("display-reason");
+  const originalCwd = process.cwd();
+  process.chdir(dir);
+  resetWriteGateState(dir);
+
+  t.after(() => {
+    try {
+      resetWriteGateState(dir);
+    } finally {
+      process.chdir(originalCwd);
+      rmSync(dir, { recursive: true, force: true });
+    }
+  });
+
+  const handlers = new Map<string, Array<(event: any, ctx?: any) => Promise<any> | any>>();
+  const pi = {
+    on(event: string, handler: (event: any, ctx?: any) => Promise<any> | any) {
+      const existing = handlers.get(event) ?? [];
+      existing.push(handler);
+      handlers.set(event, existing);
+    },
+  } as any;
+
+  registerHooks(pi, []);
+
+  let block: any;
+  for (const handler of handlers.get("tool_call") ?? []) {
+    const result = await handler({
+      toolName: "write",
+      input: {
+        path: join(dir, ".gsd", "milestones", "M001", "M001-CONTEXT.md"),
+        content: "# M001 Context\n",
+      },
+    });
+    if (result?.block) block = result;
+  }
+
+  assert.equal(block?.block, true);
+  assert.match(block?.reason ?? "", /HARD BLOCK: Cannot write to milestone CONTEXT\.md/);
+  assert.match(block?.reason ?? "", /ask_user_questions/);
+  assert.equal(block?.displayReason, "Depth check required before writing milestone context.");
+});
+
 test("register-hooks unlocks milestone depth verification from question id without guided-flow state (#4047)", async (t) => {
   const dir = makeTempDir("manual");
   const originalCwd = process.cwd();

package/src/resources/extensions/gsd/tests/rule-registry.test.ts

@@ -8,6 +8,7 @@ import { test, describe, beforeEach } from "node:test";
 import { mkdtempSync, mkdirSync, rmSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
+import { emitJournalEvent } from "../journal.ts";
 import {
   RuleRegistry,
   getRegistry,
@@ -16,6 +17,7 @@ import {
   resetRegistry,
   convertDispatchRules,
   getOrCreateRegistry,
+  resolveHookArtifactPath,
 } from "../rule-registry.ts";
 import type { UnifiedRule } from "../rule-types.ts";
 import type { DispatchAction, DispatchContext } from "../auto-dispatch.ts";
@@ -443,6 +445,79 @@ describe("RuleRegistry", () => {
     }
   });
 
+  test("failed hook completion with an artifact does not dequeue the next hook", () => {
+    const originalGsdHome = process.env.GSD_HOME;
+    const projectRoot = mkdtempSync(join(tmpdir(), "gsd-hook-failed-"));
+    const tempGsdHome = mkdtempSync(join(tmpdir(), "gsd-hook-home-"));
+    const unitId = "M001/S01/T01";
+
+    try {
+      mkdirSync(join(projectRoot, ".gsd", "milestones", "M001", "slices", "S01", "tasks"), { recursive: true });
+      writeFileSync(
+        join(projectRoot, ".gsd", "PREFERENCES.md"),
+        [
+          "---",
+          "version: 1",
+          "post_unit_hooks:",
+          "  - name: review-arbiter",
+          "    after: [execute-task]",
+          "    prompt: Review {taskId}",
+          "    artifact: REVIEW.md",
+          "    max_cycles: 1",
+          "  - name: follow-up-review",
+          "    after: [execute-task]",
+          "    prompt: Follow-up review {taskId}",
+          "---",
+        ].join("\n"),
+        "utf-8",
+      );
+      process.env.GSD_HOME = tempGsdHome;
+
+      const registry = new RuleRegistry([]);
+      const firstHook = registry.evaluatePostUnit("execute-task", unitId, projectRoot);
+      assert.equal(firstHook?.hookName, "review-arbiter");
+
+      writeFileSync(
+        resolveHookArtifactPath(projectRoot, unitId, "REVIEW.md"),
+        "partial review output",
+        "utf-8",
+      );
+      emitJournalEvent(projectRoot, {
+        ts: "2026-06-03T12:00:00.000Z",
+        flowId: "flow-hook-failed",
+        seq: 3,
+        eventType: "unit-end",
+        data: {
+          unitType: "hook/review-arbiter",
+          unitId,
+          status: "cancelled",
+          artifactVerified: false,
+          errorContext: {
+            message: "Provider error: Stream ended without finish_reason",
+            category: "provider",
+          },
+        },
+      });
+
+      const nextHook = registry.evaluatePostUnit("hook/review-arbiter", unitId, projectRoot);
+      assert.equal(nextHook, null, "failed hook must not allow follow-up hook dispatch");
+      const failure = registry.consumeHookFailure();
+      assert.equal(failure?.hookName, "review-arbiter");
+      assert.match(failure?.reason ?? "", /status cancelled/);
+
+      const resumedRegistry = new RuleRegistry([]);
+      resumedRegistry.restoreState(projectRoot);
+      const resumedHook = resumedRegistry.evaluatePostUnit("execute-task", unitId, projectRoot);
+      assert.equal(resumedHook, null, "resumed hook evaluation must not skip failed hook artifact");
+      assert.equal(resumedRegistry.consumeHookFailure()?.hookName, "review-arbiter");
+    } finally {
+      if (originalGsdHome === undefined) delete process.env.GSD_HOME;
+      else process.env.GSD_HOME = originalGsdHome;
+      rmSync(projectRoot, { recursive: true, force: true });
+      rmSync(tempGsdHome, { recursive: true, force: true });
+    }
+  });
+
   // ── matchedRule provenance (S02 journal support) ───────────────────
 
   test("evaluateDispatch result includes matchedRule on dispatch match", async () => {

package/src/resources/extensions/gsd/tests/run-uat-composer.test.ts

@@ -123,6 +123,10 @@ test("#4782 phase 3: buildRunUatPrompt inlines UAT and keeps summary/project con
   // Project path is advertised on-demand; full project body is not inlined.
   assert.match(prompt, /\.gsd\/PROJECT\.md/);
   assert.ok(!prompt.includes("Run-UAT composer fixture project"), "run-uat should not inline full project context");
+
+  assert.match(prompt, /"toolPresentationPlanId": "run-uat\/default-v1"/);
+  assert.match(prompt, /"gsd_uat_result_save"/);
+  assert.match(prompt, /"read"/);
 });
 
 test("#4782 phase 3: buildRunUatPrompt omits optional slice summary when file is missing", async (t) => {

package/src/resources/extensions/gsd/tests/runtime-invariant-modules.test.ts

@@ -7,6 +7,7 @@ import assert from "node:assert/strict";
 import { classifyFailure } from "../recovery-classification.js";
 import { reconcileBeforeDispatch } from "../state-reconciliation.js";
 import { compileUnitToolContract } from "../tool-contract.js";
+import { shouldBlockAutoUnitToolCall } from "../auto-unit-tool-scope.js";
 import type { GSDState } from "../types.js";
 
 function makeState(overrides: Partial<GSDState> = {}): GSDState {
@@ -63,10 +64,35 @@ test("Tool Contract compiles known Unit prompt and tool policy", () => {
   assert.equal(result.ok, true);
   assert.equal(result.ok && result.contract.unitType, "execute-task");
   assert.deepEqual(result.ok && result.contract.requiredWorkflowTools, ["gsd_task_complete"]);
+  assert.deepEqual(result.ok && result.contract.forbiddenWorkflowTools, []);
   assert.equal(result.ok && result.contract.toolsPolicy.mode, "all");
   assert.ok(result.ok && result.contract.validationRules.includes("closeout-tool-present"));
 });
 
+test("Tool Contract records high-risk cross-phase tool boundaries without single-owning every tool", () => {
+  const completeSlice = compileUnitToolContract("complete-slice");
+  const runUat = compileUnitToolContract("run-uat");
+
+  assert.equal(completeSlice.ok, true);
+  assert.ok(
+    completeSlice.ok &&
+      completeSlice.contract.forbiddenWorkflowTools.some((tool) => tool.name === "gsd_uat_result_save"),
+    "complete-slice should explicitly forbid saving UAT Assessments",
+  );
+
+  assert.equal(runUat.ok, true);
+  assert.ok(
+    runUat.ok &&
+      runUat.contract.requiredWorkflowTools.includes("gsd_uat_result_save"),
+    "run-uat should own the UAT result-save tool",
+  );
+  assert.ok(
+    runUat.ok &&
+      runUat.contract.forbiddenWorkflowTools.some((tool) => tool.name === "gsd_exec"),
+    "run-uat should prefer typed UAT execution over generic gsd_exec",
+  );
+});
+
 test("Tool Contract fails closed for unknown Units", () => {
   const result = compileUnitToolContract("custom-step");
 
@@ -74,10 +100,20 @@ test("Tool Contract fails closed for unknown Units", () => {
   assert.equal(!result.ok && result.reason, "unknown-unit-type");
 });
 
+test("auto Unit tool scope blocks complete-slice from saving UAT Assessment", () => {
+  const result = shouldBlockAutoUnitToolCall("complete-slice", "gsd_uat_result_save");
+
+  assert.equal(result.block, true);
+  assert.match(result.reason ?? "", /Tool Contract failure/);
+  assert.match(result.reason ?? "", /Run UAT owns persisted UAT Assessment/);
+});
+
 test("Recovery Classification covers ADR-015 failure families", () => {
   const cases = [
     ["invalid tool schema enum", "tool-schema", "stop"],
+    ["Tool Contract failure: complete-slice cannot use gsd_uat_result_save", "tool-contract", "stop"],
     ["deterministic policy rejection", "deterministic-policy", "stop"],
+    ["cannot legally advance because required UAT Assessment artifact is missing", "lifecycle-progression", "stop"],
     ["stale worker lease", "stale-worker", "stop"],
     ["worktree root missing .git", "worktree-invalid", "stop"],
     ["verification drift in state snapshot", "verification-drift", "escalate"],