@opengsd/gsd-pi 1.2.0 → 1.3.0-dev.72e3af2a

package/src/resources/extensions/gsd/tests/evidence-cross-ref.test.ts

@@ -86,6 +86,101 @@ test("stale verification evidence batches are ignored when a newer completion ba
   assert.deepEqual(mismatches, []);
 });
 
+test("WSL bash-spawn failure is not flagged as a falsified passing verification", () => {
+  // Issue #814: on Windows, `gsd_exec runtime=bash` resolves to a WSL with no
+  // /bin/bash. The bash-runtime verification call exits 1 with a spawn-failure
+  // banner (the command never ran); the LLM re-ran via a node runtime (exit 0)
+  // that findMatches does not capture. The infra failure must not block.
+  const command = "npx playwright test e2e/m039-s05-comparison-legibility.spec.ts";
+  const mismatches = crossReferenceEvidence(
+    [{ command, exitCode: 0, verdict: "passed" }],
+    [
+      {
+        kind: "bash",
+        toolCallId: "call-1",
+        command: `${command} --reporter=line 2>&1 | tail -40`,
+        exitCode: 1,
+        outputSnippet:
+          "<3>WSL (12 - Relay) ERROR: CreateProcessCommon:800: execvpe(/bin/bash) failed: No such file or directory",
+        timestamp: 1,
+      },
+    ] as EvidenceEntry[],
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "warning");
+  assert.match(mismatches[0].reason, /inconclusive/);
+});
+
+test("missing tool failure (command not found: eslint) is a real error, not an infra spawn failure", () => {
+  // Regression for overly-broad spawn signature: `command not found: eslint`
+  // is a genuine verification failure (eslint not installed), not a missing
+  // shell interpreter. It must produce a blocking error, not a warning.
+  const command = "eslint src/";
+  const mismatches = crossReferenceEvidence(
+    [{ command, exitCode: 0, verdict: "passed" }],
+    [
+      {
+        kind: "bash",
+        toolCallId: "call-1",
+        command,
+        exitCode: 1,
+        outputSnippet: "zsh: command not found: eslint",
+        timestamp: Date.now(),
+      },
+    ] as EvidenceEntry[],
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "error");
+  assert.match(mismatches[0].reason, /Claimed exitCode=0/);
+});
+
+test("missing tool failure (command not found: node) is a real error, not an infra spawn failure", () => {
+  // node is not a shell interpreter; its absence is a genuine env problem,
+  // not a shell-spawn infra failure.
+  const command = "node --test tests/verify.test.js";
+  const mismatches = crossReferenceEvidence(
+    [{ command, exitCode: 0, verdict: "passed" }],
+    [
+      {
+        kind: "bash",
+        toolCallId: "call-1",
+        command,
+        exitCode: 127,
+        outputSnippet: "bash: command not found: node",
+        timestamp: Date.now(),
+      },
+    ] as EvidenceEntry[],
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "error");
+  assert.match(mismatches[0].reason, /Claimed exitCode=0/);
+});
+
+test("missing shell interpreter (command not found: bash) is treated as an infra spawn failure", () => {
+  // bash itself missing is the shell-spawn infra case; must remain a warning.
+  const command = "npm test";
+  const mismatches = crossReferenceEvidence(
+    [{ command, exitCode: 0, verdict: "passed" }],
+    [
+      {
+        kind: "bash",
+        toolCallId: "call-1",
+        command,
+        exitCode: 1,
+        outputSnippet: "command not found: bash",
+        timestamp: Date.now(),
+      },
+    ] as EvidenceEntry[],
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "warning");
+  assert.match(mismatches[0].reason, /inconclusive/);
+});
+
 test("missing recorded bash evidence remains a warning", () => {
   const mismatches = crossReferenceEvidence(
     [{ command: "npm test", exitCode: 0, verdict: "passed" }],

package/src/resources/extensions/gsd/tests/evidence-xref-gsd-exec.test.ts

@@ -0,0 +1,157 @@
+// Project/App: gsd-pi
+// File Purpose: Regression tests for evidence cross-referencing of gsd_exec /
+// gsd_uat_exec tool calls. Mirrors the live false-positive where an
+// execute-task agent ran its verification commands through gsd_exec (script
+// body in the `script` argument) and the cross-referencer reported
+// "No bash tool call found" despite successful execution.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+import {
+  resetEvidence,
+  getEvidence,
+  recordToolCall,
+  recordToolResult,
+  isExecutionToolName,
+  type BashEvidence,
+} from "../safety/evidence-collector.ts";
+import { crossReferenceEvidence } from "../safety/evidence-cross-ref.ts";
+
+function gsdExecResult(exitCode: number, id = "4858202d-2ed7-4a0a-9ef7-4e159e65da83"): unknown {
+  return {
+    content: [{
+      type: "text",
+      text: JSON.stringify({
+        operation: "gsd_exec",
+        id,
+        runtime: "bash",
+        exit_code: exitCode,
+        signal: null,
+        timed_out: false,
+        duration_ms: 272,
+        stdout_bytes: 592,
+        stderr_bytes: 0,
+        meta_path: `/tmp/does-not-exist/.gsd/exec/${id}.meta.json`,
+      }),
+    }],
+  };
+}
+
+test("evidence-xref: verification run through gsd_exec script matches the claimed command", () => {
+  resetEvidence();
+
+  // The live false positive: agent runs `node --test tests/verify-s01.test.js`
+  // inside a gsd_exec script with a cd prefix and exit-code echo suffix.
+  recordToolCall("tc-exec-1", "gsd_exec", {
+    script: 'cd /work/.gsd/worktrees/M001 && node --test tests/verify-s01.test.js; echo "EXIT=$?"',
+    purpose: "T02: run node --test contract checks against T01 index.html",
+  });
+  recordToolResult("tc-exec-1", "gsd_exec", gsdExecResult(0), false);
+
+  const mismatches = crossReferenceEvidence(
+    [{ command: "node --test tests/verify-s01.test.js", exitCode: 0, verdict: "passed" }],
+    getEvidence(),
+  );
+
+  assert.deepEqual(mismatches, [], "gsd_exec-executed verification must not be flagged as missing");
+});
+
+test("evidence-xref: multi-line gsd_exec script matches claims for each embedded command", () => {
+  resetEvidence();
+
+  recordToolCall("tc-exec-2", "gsd_exec", {
+    script: [
+      "cd /work/.gsd/worktrees/M001",
+      "sed -i '' \"s/'todos'/'tasks-v1'/\" index.html",
+      "node --test tests/verify-s01.test.js > /dev/null 2>&1",
+      'echo "BROKEN_EXIT=$?"',
+    ].join("\n"),
+    purpose: "T02: deliberate contract break must fail, then restore",
+  });
+  recordToolResult("tc-exec-2", "gsd_exec", gsdExecResult(0), false);
+
+  const mismatches = crossReferenceEvidence(
+    [{ command: "node --test tests/verify-s01.test.js > /dev/null 2>&1", exitCode: 0, verdict: "passed" }],
+    getEvidence(),
+  );
+
+  assert.deepEqual(mismatches, [], "command embedded in a multi-line script must match");
+});
+
+test("evidence-xref: claimed pass with failing gsd_exec exit_code is still an error", () => {
+  resetEvidence();
+
+  recordToolCall("tc-exec-3", "gsd_exec", {
+    script: "node --test tests/verify-s01.test.js",
+    purpose: "verification",
+  });
+  // gsd_exec reports failures via the JSON envelope's exit_code (and isError).
+  recordToolResult("tc-exec-3", "gsd_exec", gsdExecResult(1), true);
+
+  const mismatches = crossReferenceEvidence(
+    [{ command: "node --test tests/verify-s01.test.js", exitCode: 0, verdict: "passed" }],
+    getEvidence(),
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "error");
+  assert.match(mismatches[0].reason, /Claimed exitCode=0 but actual exitCode=1/);
+});
+
+test("evidence-collector: gsd_uat_exec and MCP-namespaced variants are execution tools", () => {
+  assert.equal(isExecutionToolName("gsd_uat_exec"), true);
+  assert.equal(isExecutionToolName("mcp__gsd-workflow__gsd_uat_exec"), true);
+  assert.equal(isExecutionToolName("mcp__gsd-workflow__gsd_exec"), true);
+
+  resetEvidence();
+  recordToolCall("tc-uat-1", "gsd_uat_exec", { script: "curl -fsS http://localhost:3000/health" });
+  const bash = getEvidence().filter((e): e is BashEvidence => e.kind === "bash");
+  assert.equal(bash.length, 1, "gsd_uat_exec must record bash evidence");
+  assert.equal(bash[0].command, "curl -fsS http://localhost:3000/health");
+});
+
+test("evidence-xref: blank-command evidence does not satisfy arbitrary claims", () => {
+  // Before script extraction existed, gsd_exec calls were recorded with
+  // command: "" — and `"x".includes("")` made them match every claim,
+  // masking genuine fabrications. Blank entries must never match.
+  const mismatches = crossReferenceEvidence(
+    [{ command: "node --test tests/verify-s01.test.js", exitCode: 0, verdict: "passed" }],
+    [{
+      kind: "bash",
+      toolCallId: "tc-blank",
+      command: "",
+      exitCode: 0,
+      outputSnippet: "",
+      timestamp: 1,
+    }],
+  );
+
+  assert.equal(mismatches.length, 1);
+  assert.equal(mismatches[0].severity, "warning");
+  assert.match(mismatches[0].reason, /No bash tool call found/);
+});
+
+test("evidence-collector: exit code falls back to .gsd/exec meta.json when result text omits it", (t) => {
+  const dir = mkdtempSync(join(tmpdir(), "gsd-exec-meta-"));
+  t.after(() => rmSync(dir, { recursive: true, force: true }));
+
+  const metaPath = join(dir, "run-1.meta.json");
+  writeFileSync(metaPath, JSON.stringify({ id: "run-1", exit_code: 7 }));
+
+  resetEvidence();
+  recordToolCall("tc-meta-1", "gsd_exec", { script: "exit 7" });
+  // Truncated result: meta_path survives but exit_code was cut off.
+  recordToolResult(
+    "tc-meta-1",
+    "gsd_exec",
+    { content: [{ type: "text", text: `{"operation":"gsd_exec","meta_path":${JSON.stringify(metaPath)}` }] },
+    false,
+  );
+
+  const bash = getEvidence().filter((e): e is BashEvidence => e.kind === "bash");
+  assert.equal(bash[0].exitCode, 7, "exit code must be recovered from meta.json");
+});

package/src/resources/extensions/gsd/tests/exec-graceful-kill.test.ts

@@ -0,0 +1,231 @@
+// Project/App: gsd-pi
+// File Purpose: Deterministic timing tests for the exec-sandbox graceful-kill
+// ladder + hard-deadline force-resolve (S04/T02).
+//
+// T01 extended runExecSandbox's timeout path to a graceful
+// SIGTERM -> grace -> SIGKILL ladder (via killProcessTree) plus a caller-side
+// force-resolve hard deadline so a non-closing ("D-state") child can never hang
+// the awaiting promise. A true D-state child is non-portable, so we simulate the
+// exact symptom deterministically:
+//   1. A SIGTERM-cooperative child (trap 'exit 0' TERM) closes promptly on the
+//      first SIGTERM — proving the ladder resolves via `close` well before the
+//      force-resolve deadline.
+//   2. A SIGTERM-ignoring (trap '' TERM) child combined with a LARGE
+//      kill_grace_ms (so SIGKILL never fires in-window) and a SHORT
+//      force_resolve_delay_ms — the only way the promise can settle is the
+//      hard deadline, proving the no-hang guarantee.
+//
+// Unlike the sync bash path (which throws on force-resolve), runExecSandbox
+// always RESOLVES; force-resolve calls finalize(null, "SIGKILL") and preserves
+// timed_out === true so the agent can still distinguish a graceful timeout exit
+// from a forced kill.
+
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { existsSync, mkdtempSync, readFileSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+
+import { EXEC_DEFAULTS, runExecSandbox, type ExecSandboxOptions } from '../exec-sandbox.ts';
+
+const isWin = process.platform === 'win32';
+
+function freshBase(): string {
+  return mkdtempSync(join(tmpdir(), 'gsd-exec-gracekill-'));
+}
+
+function cleanup(dir: string): void {
+  rmSync(dir, { recursive: true, force: true });
+}
+
+function baseOpts(base: string, overrides: Partial<ExecSandboxOptions> = {}): ExecSandboxOptions {
+  return {
+    baseDir: base,
+    clamp_timeout_ms: EXEC_DEFAULTS.clampTimeoutMs,
+    default_timeout_ms: 10_000,
+    stdout_cap_bytes: 1_024,
+    stderr_cap_bytes: 1_024,
+    digest_chars: 120,
+    env_allowlist: EXEC_DEFAULTS.envAllowlist,
+    ...overrides,
+  };
+}
+
+// Wall-clock guard: reject if the call has not settled by `ms`. Proves no-hang
+// independently of the assertions on timing/markers.
+function wallClockGuard<T>(promise: Promise<T>, ms: number): Promise<T> {
+  return Promise.race([
+    promise,
+    new Promise<never>((_, reject) => {
+      const t = setTimeout(() => reject(new Error(`Call did not settle within ${ms}ms (hang)`)), ms);
+      if (typeof t === 'object' && 'unref' in t) t.unref();
+    }),
+  ]);
+}
+
+test(
+  'runExecSandbox: aborted signal terminates promptly without waiting for timeout',
+  { timeout: 10_000 },
+  async (t) => {
+    const base = freshBase();
+    t.after(() => cleanup(base));
+
+    const controller = new AbortController();
+    controller.abort();
+    const TIMEOUT_MS = 10_000;
+
+    const start = Date.now();
+    const result = await wallClockGuard(
+      runExecSandbox(
+        { runtime: 'node', script: 'setTimeout(() => {}, 30_000);', timeout_ms: TIMEOUT_MS },
+        baseOpts(base, {
+          signal: controller.signal,
+          kill_grace_ms: 500,
+          force_resolve_delay_ms: 2_000,
+        }),
+      ),
+      5_000,
+    );
+    const elapsed = Date.now() - start;
+
+    assert.equal(result.aborted, true, 'aborted signal must be recorded on the sandbox result');
+    assert.equal(result.timed_out, false, 'client abort must not be reported as a sandbox timeout');
+    assert.ok(elapsed < TIMEOUT_MS, `expected abort before ${TIMEOUT_MS}ms timeout, took ${elapsed}ms`);
+    assert.ok(
+      result.duration_ms < TIMEOUT_MS,
+      `expected result duration before timeout, got ${result.duration_ms}ms`,
+    );
+    const meta = JSON.parse(readFileSync(result.meta_path, 'utf-8'));
+    assert.equal(meta.aborted, true, 'meta.json must persist the aborted marker');
+    assert.equal(meta.timed_out, false, 'meta.json must not report timed_out when abort caused the kill');
+  },
+);
+
+// Clean up a detached child (and its process group) that a large kill_grace_ms
+// intentionally leaves alive past the force-resolve deadline.
+function cleanupByPidFile(pidFile: string): void {
+  if (!existsSync(pidFile)) return;
+  const pid = Number.parseInt(readFileSync(pidFile, 'utf-8').trim(), 10);
+  if (!Number.isFinite(pid) || pid <= 0) return;
+  if (isWin) return; // best-effort; test is skipped on win32
+  try {
+    process.kill(-pid, 'SIGKILL');
+  } catch {
+    try {
+      process.kill(pid, 'SIGKILL');
+    } catch {
+      // already gone
+    }
+  }
+}
+
+test(
+  'runExecSandbox: SIGTERM-cooperative child exits promptly within grace on timeout',
+  { skip: isWin ? 'Unix-primary graceful semantics' : false, timeout: 20_000 },
+  async (t) => {
+    const base = freshBase();
+    t.after(() => cleanup(base));
+
+    const KILL_GRACE_MS = 5_000;
+    const FORCE_RESOLVE_DELAY_MS = 8_000; // grace + hard-deadline; must NOT be reached
+    const TIMEOUT_MS = 300;
+
+    // Child traps SIGTERM and exits 0, then sleeps long. The first SIGTERM from
+    // the ladder ends it immediately -> `close` fires -> finalize via close,
+    // well before the force-resolve deadline.
+    const script = `trap 'exit 0' TERM; sleep 30`;
+
+    const result = await wallClockGuard(
+      runExecSandbox(
+        { runtime: 'bash', script, timeout_ms: TIMEOUT_MS },
+        baseOpts(base, {
+          kill_grace_ms: KILL_GRACE_MS,
+          force_resolve_delay_ms: FORCE_RESOLVE_DELAY_MS,
+        }),
+      ),
+      15_000,
+    );
+
+    // timed_out is preserved even though the child exited cooperatively.
+    assert.equal(result.timed_out, true, 'timeout fired so timed_out must be true');
+    // Proves SIGTERM (close) resolved it, NOT the force-resolve deadline:
+    // a force-resolve would land at ~TIMEOUT_MS + FORCE_RESOLVE_DELAY_MS (~8.3s).
+    assert.ok(
+      result.duration_ms < KILL_GRACE_MS,
+      `expected prompt close well under grace (${KILL_GRACE_MS}ms), got ${result.duration_ms}ms`,
+    );
+    // A clean SIGTERM-trapped exit resolves via `close` with an exit code,
+    // not the force-resolve sentinel (exit_code null / signal SIGKILL).
+    assert.notEqual(
+      result.signal,
+      'SIGKILL',
+      'cooperative child must not be force-resolved as SIGKILL',
+    );
+    assert.equal(result.force_resolved, false, 'a cooperative close must not be flagged as force-resolved');
+  },
+);
+
+test(
+  'runExecSandbox: SIGTERM-ignoring (non-closing) child force-resolves via hard deadline without hanging',
+  { skip: isWin ? 'Unix-primary graceful semantics; force-resolve covered on POSIX' : false, timeout: 20_000 },
+  async (t) => {
+    const base = freshBase();
+    const pidFile = join(base, 'child.pid');
+    t.after(() => {
+      cleanupByPidFile(pidFile);
+      cleanup(base);
+    });
+
+    // Timing seams: SIGKILL never fires in-window (grace huge), so the only way
+    // the promise can settle is the hard deadline -> proves the caller-side
+    // force-resolve.
+    const KILL_GRACE_MS = 30_000;
+    const FORCE_RESOLVE_DELAY_MS = 800;
+    // Generous timeout so the SIGTERM trap is reliably installed before the kill
+    // fires, even under heavy parallel test load. A too-tight timeout can race
+    // the shell's startup: the kill lands before `trap '' TERM` runs, SIGTERM
+    // takes its default action, the child closes on SIGTERM, and the
+    // force-resolve path never runs (flaky 'SIGTERM' instead of 'SIGKILL').
+    const TIMEOUT_MS = 1_500;
+
+    // Install the SIGTERM trap FIRST (before any other command) so the shell is
+    // already immune by the time the kill fires; then record the detached
+    // shell's PID (process-group leader) for cleanup. The shell loops forever
+    // holding its stdout pipe open -> `close` never fires in-window -> the
+    // caller's hard deadline must force-resolve.
+    const script = `trap '' TERM; echo $$ > '${pidFile}'; while true; do sleep 1; done`;
+
+    const start = Date.now();
+    const result = await wallClockGuard(
+      runExecSandbox(
+        { runtime: 'bash', script, timeout_ms: TIMEOUT_MS },
+        baseOpts(base, {
+          kill_grace_ms: KILL_GRACE_MS,
+          force_resolve_delay_ms: FORCE_RESOLVE_DELAY_MS,
+        }),
+      ),
+      8_000,
+    );
+    const elapsed = Date.now() - start;
+
+    // (a) Did not hang and settled well before the SIGKILL grace would fire.
+    assert.ok(
+      elapsed < KILL_GRACE_MS,
+      `expected force-resolve before SIGKILL grace (${KILL_GRACE_MS}ms), took ${elapsed}ms`,
+    );
+    // (b) Settled in the force-resolve band: >= timeout (kill initiated) and
+    //     roughly timeout + force_resolve_delay (generous CI slack, but kept below
+    //     the 8000ms wallClockGuard so the upper bound is actually reachable).
+    assert.ok(
+      elapsed >= TIMEOUT_MS && elapsed <= TIMEOUT_MS + FORCE_RESOLVE_DELAY_MS + 4_000,
+      `expected settle near ${TIMEOUT_MS + FORCE_RESOLVE_DELAY_MS}ms, took ${elapsed}ms`,
+    );
+    // (c) Force-resolve preserves timed_out and surfaces the SIGKILL sentinel
+    //     (exit_code null) plus the explicit force_resolved flag, so the agent can
+    //     tell a forced kill (D-state hard deadline) from a clean SIGKILL exit.
+    assert.equal(result.timed_out, true, 'force-resolve must preserve timed_out');
+    assert.equal(result.exit_code, null, 'force-resolve exit_code must be null');
+    assert.equal(result.signal, 'SIGKILL', 'force-resolve must mark signal SIGKILL');
+    assert.equal(result.force_resolved, true, 'force-resolve must set force_resolved=true');
+  },
+);

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

@@ -2,8 +2,8 @@ import test from "node:test";
 import assert from "node:assert/strict";
 
 import { registerExecTools } from "../bootstrap/exec-tools.ts";
-import { executeUatExec } from "../tools/exec-tool.ts";
-import type { ExecSandboxRequest, ExecSandboxResult } from "../exec-sandbox.ts";
+import { executeGsdExec, executeUatExec } from "../tools/exec-tool.ts";
+import type { ExecSandboxOptions, ExecSandboxRequest, ExecSandboxResult } from "../exec-sandbox.ts";
 
 function makeExecResult(request: ExecSandboxRequest): ExecSandboxResult {
   return {
@@ -12,6 +12,8 @@ function makeExecResult(request: ExecSandboxRequest): ExecSandboxResult {
     exit_code: 0,
     signal: null,
     timed_out: false,
+    aborted: false,
+    force_resolved: false,
     duration_ms: 1,
     stdout_bytes: 12,
     stderr_bytes: 0,
@@ -51,6 +53,76 @@ test("executeUatExec accepts evidence-mode aliases for intent", async () => {
   assert.equal(requests[0]?.metadata?.intent, "uat-artifact-check");
 });
 
+test("executeGsdExec passes AbortSignal into sandbox options", async () => {
+  const controller = new AbortController();
+  let capturedSignal: AbortSignal | undefined;
+
+  const result = await executeGsdExec(
+    { runtime: "bash", script: "sleep 60" },
+    {
+      baseDir: "/tmp/gsd-exec-abort-signal-test",
+      preferences: null,
+      signal: controller.signal,
+      run: async (request, opts: ExecSandboxOptions) => {
+        capturedSignal = opts.signal;
+        return makeExecResult(request);
+      },
+    },
+  );
+
+  assert.equal(result.isError, false);
+  assert.equal(capturedSignal, controller.signal);
+});
+
+test("gsd_exec surfaces aborted child termination distinctly from a clean exit", async () => {
+  const result = await executeGsdExec(
+    { runtime: "bash", script: "trap 'exit 0' TERM; sleep 60" },
+    {
+      baseDir: "/tmp/gsd-exec-aborted-test",
+      preferences: null,
+      run: async (request) => ({
+        ...makeExecResult(request),
+        aborted: true,
+        exit_code: 0,
+        signal: null,
+        digest: "[no stdout — aborted]",
+      }),
+    },
+  );
+
+  assert.equal(result.isError, true, "an aborted run must be an error even if the child exits 0");
+  assert.equal(result.details?.aborted, true, "details must expose aborted");
+  const text = result.content.map((c) => c.text ?? "").join("\n");
+  assert.match(text, /exit=aborted/, "summary must distinguish an aborted result");
+});
+
+test("gsd_exec surfaces a force-resolved (D-state) kill distinctly from a clean exit", async () => {
+  // A hard-deadline force-resolve sets force_resolved=true with a synthetic SIGKILL
+  // signal and null exit code. The tool result must carry that flag in details and
+  // render it as exit=timeout(force-killed) so the agent can tell it apart from a
+  // normally-exited or cleanly-timed-out command.
+  const result = await executeGsdExec(
+    { runtime: "bash", script: "sleep 60" },
+    {
+      baseDir: "/tmp/gsd-exec-force-resolved-test",
+      preferences: null,
+      run: async (request) => ({
+        ...makeExecResult(request),
+        exit_code: null,
+        signal: "SIGKILL",
+        timed_out: true,
+        force_resolved: true,
+        digest: "[no stdout \u2014 timed out]",
+      }),
+    },
+  );
+
+  assert.equal(result.isError, true, "a force-resolved kill is an error result");
+  assert.equal(result.details?.force_resolved, true, "details must expose force_resolved");
+  const text = result.content.map((c) => c.text ?? "").join("\n");
+  assert.match(text, /exit=timeout\(force-killed\)/, "summary must distinguish a force-killed result");
+});
+
 test("registerExecTools exposes gsd_uat_exec intent as recoverable string schema", () => {
   const tools: Array<{ name: string; parameters: any }> = [];
   registerExecTools({

package/src/resources/extensions/gsd/tests/execute-task-rendering.test.ts

@@ -38,6 +38,7 @@ test("execute-task prompt renders compact execution and completion gates", async
     planPath: ".gsd/milestones/M001/slices/S01/S01-PLAN.md",
     taskPlanPath: ".gsd/milestones/M001/slices/S01/tasks/T01-PLAN.md",
     priorTaskLines: "- None",
+    onDemandContext: "",
     skillActivation: "Load relevant skills.",
     inlinedTemplates: "### Output Template: Task Summary\nSource: `templates/task-summary.md`",
     templatesDir: join(fixtureRoot, "templates"),

package/src/resources/extensions/gsd/tests/export-html-enhancements.test.ts

@@ -108,6 +108,7 @@ function mockData(overrides: Partial<VisualizerData> = {}): VisualizerData {
     changelog: { entries: [] },
     sliceVerifications: [],
     knowledge: { rules: [], patterns: [], lessons: [], exists: false },
+    memories: { entries: [], totalCount: 0 },
     captures: { entries: [], pendingCount: 0, totalCount: 0 },
     health: {
       budgetCeiling: undefined,
@@ -149,6 +150,30 @@ test("report uses the shared GSD HTML shell", () => {
   assert.ok(html.includes('<main>'), "should render content inside shared shell main");
 });
 
+test("knowledge section renders active memories", () => {
+  const html = generateHtmlReport(mockData({
+    memories: {
+      totalCount: 1,
+      entries: [
+        {
+          id: "MEM001",
+          category: "gotcha",
+          content: "Memory rows must be visible in exported visualizer reports",
+          confidence: 0.9,
+          hitCount: 2,
+          scope: "project",
+          tags: ["visualizer"],
+          updatedAt: "2026-06-21T12:00:00.000Z",
+        },
+      ],
+    },
+  }), mockOpts());
+
+  assert.ok(html.includes("Memories"), "should render memories heading");
+  assert.ok(html.includes("MEM001"), "should render memory ID");
+  assert.ok(html.includes("Memory rows must be visible"), "should render memory content");
+});
+
 test("Feature 1: executive summary includes budget context when set", () => {
   const data = mockData({ health: { ...mockData().health, budgetCeiling: 10.00 } });
   const html = generateHtmlReport(data, mockOpts());