@opengsd/gsd-pi 1.2.0-dev.844675c9 → 1.2.0-dev.955e4da0

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

@@ -4,6 +4,7 @@ import { readFileSync } from "node:fs";
 import { join } from "node:path";
 import {
   RUN_UAT_BROWSER_TOOL_NAMES,
+  resolveToolPresentationPlan,
   buildRunUatResultPresentation,
   buildRunUatPresentationForType,
   RUN_UAT_READ_ONLY_TOOL_NAMES,
@@ -179,6 +180,33 @@ test("browser-executable UAT presentation uses direct browser tools", () => {
   assert.ok(!presentation.presentedTools.some((toolName) => toolName.startsWith("mcp__gsd-browser__")));
 });
 
+test("run-uat presentation plans carry typed tool-surface snapshots", () => {
+  const plan = resolveToolPresentationPlan({
+    phase: "run-uat",
+    surface: "mcp",
+    workflowMcpServerName: "gsd-workflow",
+    availableToolNames: [
+      "gsd_uat_exec",
+      "gsd_uat_result_save",
+      "read",
+    ],
+    includeBrowserTools: [],
+  });
+
+  assert.equal(plan.toolSurface.source, "presentation-plan");
+  assert.equal(plan.toolSurface.phase, "run-uat");
+  assert.deepEqual(plan.toolSurface.scopedToolNames, [
+    "gsd_uat_exec",
+    "gsd_uat_result_save",
+    "read",
+  ]);
+  assert.deepEqual(plan.toolSurface.presentedToolNames, [
+    "mcp__gsd-workflow__gsd_uat_exec",
+    "mcp__gsd-workflow__gsd_uat_result_save",
+    "read",
+  ]);
+});
+
 test("live-runtime and mixed UAT presentations also surface browser tools", () => {
   // Regression (M001/S03): the run-uat prompt tells live-runtime and mixed to
   // drive a browser, so the runner must actually receive the browser tools and
@@ -517,6 +545,24 @@ test("plan-slice prompt clarifies gsd_plan_slice handles task persistence", () =
   assert.match(prompt, /gsd_plan_slice` handles task persistence/i);
 });
 
+test("plan-slice prompt references web app UAT guidance when planning browser work", () => {
+  const prompt = readPrompt("plan-slice");
+  assert.match(prompt, /Web App UAT guidance/i);
+  assert.match(prompt, /Playwright smoke scaffolding/i);
+});
+
+test("complete-slice prompt references web app UAT mode rules", () => {
+  const prompt = readPrompt("complete-slice");
+  assert.match(prompt, /Web App UAT guidance/i);
+  assert.match(prompt, /browser-executable|runtime-executable/i);
+});
+
+test("plan-milestone prompt seeds web verification strategy", () => {
+  const prompt = readPrompt("plan-milestone");
+  assert.match(prompt, /Web apps/i);
+  assert.match(prompt, /Playwright|browser_\*/i);
+});
+
 test("replan-slice prompt uses gsd_replan_slice as canonical DB-backed tool", () => {
   const prompt = readPrompt("replan-slice");
   assert.match(prompt, /gsd_replan_slice/);

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

@@ -740,6 +740,15 @@ test("MAX_TRANSIENT_AUTO_RESUMES is at least 8 for sustained overload resilience
   );
 });
 
+// ── OpenAI-completions mid-stream cut (#577) ─────────────────────────────────
+
+test("classifyError: 'Stream ended without finish_reason' is transient network (#577)", () => {
+  const result = classifyError("Stream ended without finish_reason");
+  assert.ok(isTransient(result), "Stream ended without finish_reason must be transient");
+  assert.equal(result.kind, "network");
+  assert.ok("retryAfterMs" in result && result.retryAfterMs > 0);
+});
+
 // ── Stream idle timeout / partial response (#4558) ──────────────────────────
 
 test("classifyError: 'Stream idle timeout - partial response received' is transient network", () => {

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

@@ -0,0 +1,165 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import type { ContextManagementConfig } from "../preferences-types.js";
+import type { ProviderPayloadPolicyDeps } from "../provider-payload-policy.js";
+
+import { applyProviderPayloadPolicy } from "../provider-payload-policy.js";
+
+function createDeps(
+  overrides: Partial<ProviderPayloadPolicyDeps> & {
+    context?: ContextManagementConfig | undefined;
+    autoActive?: boolean;
+    sourceContextBlock?: string | null;
+  } = {},
+): ProviderPayloadPolicyDeps {
+  return {
+    isAutoActive: () => overrides.autoActive ?? false,
+    loadContextManagementConfig: () => overrides.context,
+    renderSourceContextBlock: () => overrides.sourceContextBlock ?? null,
+    getEffectiveServiceTier: () => undefined,
+    supportsServiceTier: () => false,
+    ...overrides,
+  };
+}
+
+function textFromMessage(message: { content?: unknown }): string {
+  const content = message.content;
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+  const text = content.find((block): block is { text: string } => {
+    return Boolean(block && typeof block === "object" && "text" in block && typeof block.text === "string");
+  });
+  return text?.text ?? "";
+}
+
+test("provider payload policy truncates tool results outside auto-mode without masking", () => {
+  const messageText = "m".repeat(50);
+  const responsesOutput = "r".repeat(50);
+  const payload = {
+    messages: [
+      { role: "user", content: [{ type: "text", text: "keep me" }] },
+      { role: "toolResult", content: [{ type: "text", text: messageText }] },
+    ],
+    input: [
+      { role: "user", content: [{ type: "input_text", text: "keep me" }] },
+      { type: "function_call_output", call_id: "call_test", output: responsesOutput },
+    ],
+  };
+
+  applyProviderPayloadPolicy({
+    payload,
+    deps: createDeps({ context: { observation_mask_turns: 1, tool_result_max_chars: 10 } }),
+  });
+
+  const truncatedMessage = textFromMessage(payload.messages[1]);
+  const truncatedResponsesOutput = String(payload.input[1]?.output ?? "");
+  assert.match(truncatedMessage, /\[truncated\]/);
+  assert.match(truncatedResponsesOutput, /\[truncated\]/);
+  assert.doesNotMatch(truncatedMessage, /result masked/);
+  assert.doesNotMatch(truncatedResponsesOutput, /result masked/);
+});
+
+test("provider payload policy appends source context after masking and truncation", () => {
+  const sourceContextBlock = "## Source Context Block\n\n" + "full source text ".repeat(20);
+  const payload = {
+    messages: [
+      { role: "user", content: [{ type: "text", text: "old turn" }] },
+      { role: "toolResult", content: [{ type: "text", text: "old result ".repeat(20) }] },
+      { role: "assistant", content: [{ type: "text", text: "ok" }] },
+      { role: "user", content: [{ type: "text", text: "new turn" }] },
+      { role: "toolResult", content: [{ type: "text", text: "new result ".repeat(20) }] },
+    ],
+  };
+
+  applyProviderPayloadPolicy({
+    payload,
+    deps: createDeps({
+      autoActive: true,
+      context: { observation_mask_turns: 1, tool_result_max_chars: 80 },
+      sourceContextBlock,
+    }),
+  });
+
+  const oldResult = textFromMessage(payload.messages[1]);
+  const newResult = textFromMessage(payload.messages[4]);
+  const appendedContext = textFromMessage(payload.messages[payload.messages.length - 1]);
+
+  assert.match(oldResult, /result masked/);
+  assert.match(newResult, /\[truncated\]/);
+  assert.equal(appendedContext, sourceContextBlock);
+  assert.doesNotMatch(appendedContext, /\[truncated\]/);
+});
+
+test("provider payload policy applies ordering to Responses input payloads", () => {
+  const sourceContextBlock = "## Source Context Block\n\n" + "responses source text ".repeat(20);
+  const payload = {
+    input: [
+      { role: "user", content: [{ type: "input_text", text: "old turn" }] },
+      { type: "function_call_output", call_id: "call_old", output: "old result ".repeat(20) },
+      { type: "message", role: "assistant", content: [{ type: "output_text", text: "ok" }] },
+      { role: "user", content: [{ type: "input_text", text: "new turn" }] },
+      { type: "function_call_output", call_id: "call_new", output: "new result ".repeat(20) },
+    ],
+  };
+
+  applyProviderPayloadPolicy({
+    payload,
+    deps: createDeps({
+      autoActive: true,
+      context: { observation_mask_turns: 1, tool_result_max_chars: 80 },
+      sourceContextBlock,
+    }),
+  });
+
+  assert.match(String(payload.input[1]?.output ?? ""), /result masked/);
+  assert.match(String(payload.input[4]?.output ?? ""), /\[truncated\]/);
+  assert.equal(textFromMessage(payload.input[payload.input.length - 1]), sourceContextBlock);
+});
+
+test("provider payload policy replaces existing source context blocks", () => {
+  const payload = {
+    messages: [
+      { role: "user", content: [{ type: "text", text: "keep me" }] },
+      { role: "user", content: [{ type: "text", text: "## Source Context Block\n\nstale" }] },
+    ],
+  };
+
+  applyProviderPayloadPolicy({
+    payload,
+    deps: createDeps({
+      autoActive: true,
+      sourceContextBlock: "## Source Context Block\n\nfresh",
+    }),
+  });
+
+  const sourceMessages = payload.messages.filter((message) => {
+    return textFromMessage(message).startsWith("## Source Context Block");
+  });
+  assert.equal(sourceMessages.length, 1);
+  assert.equal(textFromMessage(sourceMessages[0]), "## Source Context Block\n\nfresh");
+});
+
+test("provider payload policy sets service tier only for supported models", () => {
+  const unsupported = {};
+  applyProviderPayloadPolicy({
+    payload: unsupported,
+    modelId: "claude-opus-4-6",
+    deps: createDeps({
+      getEffectiveServiceTier: () => "priority",
+      supportsServiceTier: (modelId) => modelId === "gpt-5.4",
+    }),
+  });
+  assert.equal("service_tier" in unsupported, false);
+
+  const supported: Record<string, unknown> = {};
+  applyProviderPayloadPolicy({
+    payload: supported,
+    modelId: "gpt-5.4",
+    deps: createDeps({
+      getEffectiveServiceTier: () => "priority",
+      supportsServiceTier: (modelId) => modelId === "gpt-5.4",
+    }),
+  });
+  assert.equal(supported.service_tier, "priority");
+});

package/src/resources/extensions/gsd/tests/pull-request-process.test.ts

@@ -0,0 +1,47 @@
+// Project/App: gsd-pi
+// File Purpose: Tests for process-level pull request policy.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import {
+  buildPullRequestEvidence,
+  createDraftPullRequestFromEvidence,
+} from "../pull-request-process.js";
+
+test("buildPullRequestEvidence omits AI credit by policy", () => {
+  const evidence = buildPullRequestEvidence({
+    milestoneId: "M001",
+    milestoneTitle: "Git process",
+    aiAssisted: true,
+  });
+
+  assert.equal(evidence.title, "feat: Git process");
+  assert.ok(!evidence.body.includes("## AI Assistance Disclosure"));
+  assert.ok(!evidence.body.includes("AI assistance"));
+});
+
+test("createDraftPullRequestFromEvidence forwards exact evidence and branch options", () => {
+  const calls: unknown[][] = [];
+  const url = createDraftPullRequestFromEvidence(
+    "/repo",
+    "M001",
+    { title: "feat: Git process", body: "body" },
+    { head: "milestone/M001", base: "main" },
+    {
+      createDraftPR: (basePath, milestoneId, title, body, opts) => {
+        calls.push([basePath, milestoneId, title, body, opts]);
+        return "https://github.example/pr/1";
+      },
+    },
+  );
+
+  assert.equal(url, "https://github.example/pr/1");
+  assert.deepEqual(calls, [[
+    "/repo",
+    "M001",
+    "feat: Git process",
+    "body",
+    { head: "milestone/M001", base: "main" },
+  ]]);
+});

package/src/resources/extensions/gsd/tests/recovery-classification-illegal-transition.test.ts

@@ -0,0 +1,30 @@
+// GSD — recovery-classification: illegal-transition kind (ADR-030)
+
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { classifyFailure } from "../recovery-classification.ts";
+import { IllegalPhaseTransitionError } from "../state-transition-matrix.ts";
+
+test("classifyFailure recognizes IllegalPhaseTransitionError by class and escalates", () => {
+  const classification = classifyFailure({
+    error: new IllegalPhaseTransitionError("executing", "complete"),
+    unitType: "execute-task",
+    unitId: "T-1",
+  });
+
+  assert.equal(classification.failureKind, "illegal-transition");
+  assert.equal(classification.action, "escalate");
+  assert.equal(classification.exitReason, "illegal-transition");
+  assert.match(classification.reason, /Illegal phase transition/);
+});
+
+test("classifyFailure routes an explicit illegal-transition failureKind to the same case", () => {
+  const classification = classifyFailure({
+    error: new Error("derived edge rejected"),
+    failureKind: "illegal-transition",
+  });
+
+  assert.equal(classification.failureKind, "illegal-transition");
+  assert.equal(classification.action, "escalate");
+});

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

@@ -1,6 +1,6 @@
 import test from "node:test";
 import assert from "node:assert/strict";
-import { existsSync, mkdirSync, readFileSync, rmSync } from "node:fs";
+import { existsSync, mkdirSync, readFileSync, rmSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
 
@@ -13,11 +13,18 @@ import { closeDatabase, getMilestone } from "../gsd-db.ts";
 import { deriveState, invalidateStateCache } from "../state.ts";
 import {
   getPendingGate,
+  loadWriteGateSnapshot,
   resetWriteGateState,
+  setPendingGate,
   shouldBlockContextArtifactSave,
 } from "../bootstrap/write-gate.ts";
 import { classifyCommand } from "../safety/destructive-guard.ts";
 import { toRoundResultResponse } from "../../remote-questions/manager.ts";
+import {
+  markInteractiveElicitationStart,
+  markInteractiveElicitationEnd,
+  clearInFlightTools,
+} from "../auto-tool-tracking.ts";
 
 function makeTempDir(prefix: string): string {
   const dir = join(
@@ -708,3 +715,180 @@ test("register-hooks gates MCP ask_user_questions cancellation before requiremen
   assert.equal(requirementBlock?.block, true, "requirement save must be blocked while gate is pending");
   assert.match(requirementBlock?.reason ?? "", /has not been confirmed/);
 });
+
+// ─── Foreground self-cancel regression (#cc-elicitation-self-cancel) ───
+// Product-visible symptom: under claude-code-cli + gsd-MCP, ask_user_questions
+// is routed as an SDK elicitation (the human boundary). The message_update hook
+// would arm the approval-gate pause and emit the "waiting for your approval -
+// pausing" notice, tearing down that elicitation and looping a re-ask. The fix
+// makes message_update bail while an interactive elicitation is in flight, while
+// still pausing for prose-only approvals (native-TUI provider, where the marker
+// is always false). This drives the real registered hook end-to-end.
+test("register-hooks message_update does NOT pause while an interactive elicitation is the human boundary, but still pauses otherwise", async (t) => {
+  const dir = makeTempDir("elicitation-pause-guard");
+  const originalCwd = process.cwd();
+  process.chdir(dir);
+  resetWriteGateState(dir);
+  clearPendingAutoStart(dir);
+  clearInFlightTools();
+
+  t.after(() => {
+    try {
+      resetWriteGateState(dir);
+      clearPendingAutoStart(dir);
+      clearInFlightTools();
+    } 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;
+
+  const notices: Array<{ text: string; level: string }> = [];
+  const ctx = {
+    cwd: dir,
+    ui: { notify: (text: string, level: string) => notices.push({ text, level }) },
+  } as any;
+
+  registerHooks(pi, []);
+
+  // A discuss-milestone is the active unit, so the approval text would normally
+  // arm the pause/notice path.
+  setPendingAutoStart(dir, {
+    basePath: dir,
+    milestoneId: "M001",
+    ctx,
+    pi: { sendMessage: () => undefined } as any,
+  });
+
+  // The model's plain-text approval question — identical in both phases.
+  const approvalMessage = {
+    role: "assistant",
+    content: [
+      { type: "text", text: "Here is the milestone plan.\n\nDid I capture the project correctly?" },
+    ],
+  };
+
+  const fireMessageUpdate = async () => {
+    for (const handler of handlers.get("message_update") ?? []) {
+      await handler({ message: approvalMessage }, ctx);
+    }
+  };
+
+  // Phase 1 — FIX: an interactive elicitation is in flight (claude-code-cli
+  // foreground). The pause/notice MUST be suppressed.
+  markInteractiveElicitationStart();
+  await fireMessageUpdate();
+  assert.equal(
+    notices.some((n) => /waiting for your approval - pausing/.test(n.text)),
+    false,
+    "must NOT emit the approval-pause notice while the elicitation is the human boundary",
+  );
+  markInteractiveElicitationEnd();
+
+  // Phase 2 — control: no elicitation in flight (native-TUI provider or a
+  // prose-only approval). The same message MUST still pause.
+  notices.length = 0;
+  await fireMessageUpdate();
+  assert.equal(
+    notices.some((n) => /discuss-milestone M001 is waiting for your approval - pausing/.test(n.text)),
+    true,
+    "prose-only approval with no elicitation in flight must still arm the pause notice",
+  );
+});
+
+test("register-hooks agent_end does not re-arm deferred gate after workflow MCP verified write-gate on disk", async (t) => {
+  const dir = makeTempDir("mcp-disk-sync");
+  const originalCwd = process.cwd();
+  const originalEnv = process.env.GSD_PERSIST_WRITE_GATE_STATE;
+  process.chdir(dir);
+  resetWriteGateState(dir);
+  clearPendingAutoStart(dir);
+  process.env.GSD_PERSIST_WRITE_GATE_STATE = "1";
+
+  const gateId = "depth_verification_M005_confirm";
+  const statePath = join(dir, ".gsd", "runtime", "write-gate-state.json");
+
+  t.after(() => {
+    try {
+      resetWriteGateState(dir);
+      clearPendingAutoStart(dir);
+    } finally {
+      if (originalEnv === undefined) {
+        delete process.env.GSD_PERSIST_WRITE_GATE_STATE;
+      } else {
+        process.env.GSD_PERSIST_WRITE_GATE_STATE = originalEnv;
+      }
+      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;
+
+  const ctx = {
+    cwd: dir,
+    ui: { notify: () => undefined },
+  } as any;
+
+  registerHooks(pi, []);
+
+  setPendingAutoStart(dir, {
+    basePath: dir,
+    milestoneId: "M005",
+    ctx,
+    pi: { sendMessage: () => undefined } as any,
+  });
+
+  const approvalMessage = {
+    role: "assistant",
+    content: [
+      { type: "text", text: "Did I capture the depth right?" },
+    ],
+  };
+
+  for (const handler of handlers.get("message_update") ?? []) {
+    await handler({ message: approvalMessage }, ctx);
+  }
+
+  setPendingGate(gateId, dir);
+  mkdirSync(join(dir, ".gsd", "runtime"), { recursive: true });
+  writeFileSync(statePath, JSON.stringify({
+    verifiedDepthMilestones: ["M005"],
+    verifiedApprovalGates: [gateId],
+    activeQueuePhase: false,
+    pendingGateId: null,
+  }, null, 2), "utf-8");
+
+  for (const handler of handlers.get("agent_end") ?? []) {
+    await handler({ messages: [] }, ctx);
+  }
+
+  assert.equal(getPendingGate(dir), null, "agent_end must not re-arm a gate the MCP subprocess already verified");
+  assert.equal(
+    shouldBlockContextArtifactSave("CONTEXT", "M005", null, dir).block,
+    false,
+    "verified milestone context writes must stay unlocked after agent_end",
+  );
+  assert.deepEqual(loadWriteGateSnapshot(dir), {
+    verifiedDepthMilestones: ["M005"],
+    verifiedApprovalGates: [gateId],
+    activeQueuePhase: false,
+    pendingGateId: null,
+  });
+});

package/src/resources/extensions/gsd/tests/roadmap-parse-regression.test.ts

@@ -396,4 +396,44 @@ test('T. #1736: Checkbox format unchanged', () => {
     assert.deepStrictEqual(slices[1].done, false, '#1736 checkbox compat: S02 not done');
 });
 
+// ═══════════════════════════════════════════════════════════════════════
+// U. Regression #566: double-bracket depends serialization recovery
+// ═══════════════════════════════════════════════════════════════════════
+test('U. #566: double-bracket depends:[[S01]] recovers to ["S01"]', () => {
+    // Simulates a roadmap written by markdown-renderer when the DB had
+    // depends=["[S01]"] — produces `depends:[[S01]]` on disk.
+    const content = [
+      '# M020: Corrupted Depends',
+      '',
+      '## Slices',
+      '',
+      '- [ ] **S01: Foundation** `risk:low` `depends:[]`',
+      '- [ ] **S02: Build** `risk:medium` `depends:[[S01]]`',
+      '- [ ] **S03: Ship** `risk:high` `depends:[[S01, S02]]`',
+      '',
+    ].join('\n');
+
+    const slices = parseRoadmapSlices(content);
+    assert.deepStrictEqual(slices.length, 3, '#566 double-bracket: 3 slices');
+    assert.deepStrictEqual(slices[1].depends, ['S01'], '#566: S02 depends recovered to ["S01"]');
+    assert.deepStrictEqual(slices[2].depends, ['S01', 'S02'], '#566: S03 depends recovered to ["S01","S02"]');
+});
+
+test('U. #566: bracket-wrapped single element depends:[[S01]] recovers to ["S01"]', () => {
+    // Simulates the exact DB corruption seen in the forensic evidence: depends=["[S01]"]
+    const content = [
+      '# M020: Single Bracket Corruption',
+      '',
+      '## Slices',
+      '',
+      '- [ ] **S01: First** `risk:low` `depends:[]`',
+      '- [ ] **S02: Second** `risk:medium` `depends:[[S01]]`',
+      '',
+    ].join('\n');
+
+    const slices = parseRoadmapSlices(content);
+    assert.deepStrictEqual(slices[1].depends.length, 1, '#566 single: S02 has exactly 1 dep');
+    assert.deepStrictEqual(slices[1].depends[0], 'S01', '#566 single: dep is "S01" not "[S01]"');
+});
+
 });

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

@@ -6,7 +6,7 @@ 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 { compileUnitContextContract, compileUnitToolContract } from "../tool-contract.js";
 import { shouldBlockAutoUnitToolCall } from "../auto-unit-tool-scope.js";
 import type { GSDState } from "../types.js";
 
@@ -68,6 +68,16 @@ test("Tool Contract compiles known Unit prompt and tool policy", () => {
   assert.equal(result.ok && result.contract.toolsPolicy.mode, "all");
   assert.ok(result.ok && result.contract.validationRules.includes("closeout-tool-present"));
   assert.ok(result.ok && result.contract.validationRules.includes("source-observation-contract-present"));
+  assert.deepEqual(result.ok && result.contract.promptContext.artifacts.inline, [
+    "task-plan",
+    "slice-plan",
+    "prior-task-summaries",
+    "templates",
+  ]);
+  assert.deepEqual(result.ok && result.contract.promptContext.artifacts.onDemand, ["slice-research"]);
+  assert.ok(result.ok && result.contract.promptObligations.includes("context-inline:task-plan,slice-plan,prior-task-summaries,templates"));
+  assert.ok(result.ok && result.contract.promptObligations.includes("context-on-demand:slice-research"));
+  assert.ok(result.ok && result.contract.promptObligations.includes("source-observations:whole-file-active-unit"));
   assert.deepEqual(result.ok && result.contract.sourceObservations, {
     mode: "whole-file-active-unit",
     seedFields: ["task.files", "task.inputs"],
@@ -77,6 +87,20 @@ test("Tool Contract compiles known Unit prompt and tool policy", () => {
   });
 });
 
+test("Unit Context Contract exposes prompt context without workflow tool surface", () => {
+  const result = compileUnitContextContract("execute-task");
+
+  assert.equal(result.ok, true);
+  assert.equal(result.ok && result.contract.unitType, "execute-task");
+  assert.equal(result.ok && result.contract.contextMode, "execution");
+  assert.equal(result.ok && result.contract.toolsPolicy.mode, "all");
+  assert.deepEqual(result.ok && result.contract.artifacts.excerpt, []);
+  assert.deepEqual(result.ok && result.contract.artifacts.computed, []);
+  assert.deepEqual(result.ok && result.contract.artifacts.prepend, []);
+  assert.equal(result.ok && result.contract.maxSystemPromptChars, 1_500_000);
+  assert.equal(result.ok && result.contract.sourceObservations.mode, "whole-file-active-unit");
+});
+
 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");

package/src/resources/extensions/gsd/tests/safety-harness-false-positives.test.ts

@@ -266,3 +266,41 @@ test("safety-harness: planned changed file avoids unexpected-file warning", (t)
   assert.deepEqual(audit!.unexpectedFiles, [], "planned index.html must not be unexpected");
   assert.deepEqual(audit!.missingFiles, [], "planned index.html must not be missing");
 });
+
+// ─── External engine evidence (claude-code-cli pre-executed tools) ──────────
+// External engines skip beforeToolCall/tool_call, so register-hooks.ts records
+// evidence at tool_execution_start instead. These tests lock the collector
+// semantics that wiring relies on: idempotent recording by toolCallId, and
+// capitalized external tool names (Bash/Write/Edit) being recognized.
+
+test("safety-harness: recordToolCall is idempotent by toolCallId", () => {
+  resetEvidence();
+
+  // Native tools fire tool_execution_start AND tool_call — both record.
+  recordToolCall("tc-dup-1", "bash", { command: "npm test" });
+  recordToolCall("tc-dup-1", "bash", { command: "npm test" });
+
+  assert.equal(getEvidence().length, 1, "same toolCallId must not duplicate");
+});
+
+test("safety-harness: external capitalized Bash call records and resolves evidence", () => {
+  resetEvidence();
+
+  // tool_execution_start with Claude Code's native tool name
+  recordToolCall("ext-bash-1", "Bash", { command: "node /tmp/t01-verify.mjs" });
+  const entries = getEvidence().filter((e): e is BashEvidence => e.kind === "bash");
+  assert.equal(entries.length, 1, "capitalized Bash must be recognized as execution tool");
+  assert.equal(entries[0]!.command, "node /tmp/t01-verify.mjs");
+
+  // tool_execution_end fills the result
+  recordToolResult("ext-bash-1", "Bash", { content: [{ type: "text", text: "13 checks passed" }] }, false);
+  assert.equal(entries[0]!.exitCode, 0, "non-error external result resolves to exitCode 0");
+});
+
+test("safety-harness: external Write call records file evidence", () => {
+  resetEvidence();
+
+  recordToolCall("ext-write-1", "Write", { file_path: "/tmp/app/index.html" });
+  const writes = getEvidence().filter((e) => e.kind === "write");
+  assert.equal(writes.length, 1, "capitalized Write must record file evidence");
+});