mokosh/.planning/phases/04-harden-clean-up-optional/04-04-PLAN.md

---
phase: 04
slug: harden-clean-up-optional
plan: 04
type: execute
wave: 3
depends_on:
  - 01
  - 02
  - 03
files_modified:
  - tests/uat/extension-page-harness.ts
  - tests/uat/lib/harness-page-driver.ts
  - tests/uat/harness.test.ts
autonomous: true
requirements: []
tags:
  - uat-harness
  - a33
  - sw-state-persistence
  - sw-eviction
  - spike-first
  - cdp-worker-close
  - roadmap-sc-1
  - charter-d-p4-01
user_setup: []
must_haves:
  truths:
    - "Wave 0 spike verifies empirically whether the offscreen document survives 5-min SW idle + worker.close() while MediaRecorder is actively recording — informs whether A33 is verification-only OR needs IndexedDB persistence work"
    - "stopServiceWorker(browser, extensionId) helper exists in tests/uat/lib/harness-page-driver.ts using Puppeteer CDP browser.waitForTarget + worker.close() per Chrome devrel canonical pattern"
    - "assertA33 / driveA33 land per the spike outcome: if PASS (offscreen survives) → verification-only A33 that does a real 5-min wall-clock idle + SW kill + SAVE → asserts archive's video/last_30sec.webm size > 100 KB"
    - "A33 is env-gated by SKIP_LONG_UAT (default: RUN for closure + alpha gate; SKIP_LONG_UAT=1 to skip for per-commit iteration)"
    - "UAT harness count flips from 33 → 34 (A33 added); 34/34 GREEN when SKIP_LONG_UAT unset"
    - "ROADMAP SC #1 (SW state persistence) GREEN — A33 empirical evidence that a real-world 5-min idle + SAVE produces a non-empty video buffer"
    - "SAVE_ARCHIVE dispatch reuses the canonical chrome.runtime.sendMessage({type: 'SAVE_ARCHIVE'}, ...) pattern from the harness page realm (NOT a new __mokoshHarness helper); matches the established assertA5/A11/A12/A13/A26/A28/A29/A30/A31 precedent"
  artifacts:
    - path: "tests/uat/extension-page-harness.ts"
      provides: "assertA33 page-side stub (or thin driver) for SAVE_ARCHIVE dispatch after host-side SW kill"
      contains: "assertA33"
    - path: "tests/uat/lib/harness-page-driver.ts"
      provides: "stopServiceWorker(browser, extensionId) NEW helper + driveA33 host-side CDP-kill + JSZip video-size check"
      contains: "worker.close()"
    - path: "tests/uat/harness.test.ts"
      provides: "driveA33 import + wrapped driver const (passes handles.browser + handles.extensionId + handles.downloadsDir) + drivers-array push + SKIP_LONG_UAT env-gate wrapper"
      contains: "driveA33Wrapped"
  key_links:
    - from: "tests/uat/harness.test.ts driveA33Wrapped"
      to: "tests/uat/lib/harness-page-driver.ts driveA33(page, browser, extensionId, downloadsDir)"
      via: "(page) => driveA33(page, handles.browser, handles.extensionId, handles.downloadsDir)"
      pattern: "handles\\.browser.*handles\\.extensionId"
    - from: "tests/uat/lib/harness-page-driver.ts stopServiceWorker"
      to: "Chrome MV3 SW target via Puppeteer CDP"
      via: "browser.waitForTarget(t => t.type()==='service_worker' && t.url().startsWith(`chrome-extension://${extensionId}`)) + target.worker().close()"
      pattern: "service_worker"
    - from: "tests/uat/lib/harness-page-driver.ts driveA33 video-size check"
      to: "zip.file('video/last_30sec.webm') → byteLength > 100_000"
      via: "JSZip.loadAsync + entry.async('uint8array')"
      pattern: "video/last_30sec\\.webm"
---

<objective>
Ship the A33 harness assertion that empirically verifies ROADMAP SC #1 (SW state persistence across the 30s idle unload edge cases). Per RESEARCH Q2: the current architecture stores segments only in offscreen-document RAM (src/offscreen/recorder.ts:91 `let segments: Blob[] = []`). The SW NEVER stores the buffer. So the actual question becomes: does the offscreen document survive 5 minutes of SW idle? Per Chrome docs, the offscreen has its own lifecycle independent of the SW, with active `MediaRecorder` being the canonical "compelling reason" to keep the offscreen alive.

This plan uses the SPIKE-FIRST approach to avoid over-engineering:

**Wave 0 (spike):** A30-min empirical investigation. Start recording; wait 5 min real wall-clock; force-kill the SW via Puppeteer CDP `worker.close()` (Chrome devrel canonical pattern; Puppeteer ≥22.1.0 supports it — our pin ^25 is comfortably above); dispatch SAVE_ARCHIVE from page.evaluate; check the resulting archive's `video/last_30sec.webm` size.

**Wave 1 (impl):** Based on spike outcome:
- **If spike PASSES** (likely outcome per RESEARCH architecture analysis + Chrome docs): A33 is a VERIFICATION-ONLY harness assertion that wraps the spike methodology into a repeatable test. Ships the spike's exact pattern as `driveA33` + `assertA33` + orchestrator wiring + env-gate. ROADMAP SC #1 is satisfied by the CURRENT architecture; no persistence layer needed.
- **If spike FAILS** (the offscreen dies along with the SW, contrary to Chrome docs): A33 implementation expands per RESEARCH Q2 sub-question (b) recommendation (Option C: IndexedDB persistence in offscreen — Blobs serialize cleanly to IDB; structured-clone supports them natively; per-segment write ~3 MB; ~3 writes per 30s window). This is a wider plan rewrite; the plan-checker should flag for re-planning if it materializes. RESEARCH confidence on offscreen-surviving-SW-kill is MEDIUM; the spike-first approach is the canonical risk hedge per Plan 01-07 precedent.

Purpose: Forms the empirical evidence for ROADMAP SC #1 ("After running the extension idle for >5 minutes, then exporting, the archive still contains a non-empty video buffer"). The spike-first approach hedges against the RESEARCH MEDIUM-confidence assumption (A3 — offscreen survives SW eviction). If the assumption holds, A33 is a verification gate; if not, persistence work is plain-needed.

Output: 1 NEW assertion (A33; harness count 33→34); 1 NEW helper (`stopServiceWorker` CDP wrapper); 3-file lockstep update per the Approach B pattern (extension-page-harness.ts + harness-page-driver.ts + harness.test.ts); env-gate via SKIP_LONG_UAT (default = RUN; set to '1' to skip for per-commit iteration).
</objective>

<execution_context>
@$HOME/.claude/get-shit-done/workflows/execute-plan.md
@$HOME/.claude/get-shit-done/templates/summary.md
</execution_context>

<context>
@.planning/PROJECT.md
@.planning/ROADMAP.md
@.planning/STATE.md
@.planning/phases/04-harden-clean-up-optional/04-CONTEXT.md
@.planning/phases/04-harden-clean-up-optional/04-RESEARCH.md
@.planning/phases/04-harden-clean-up-optional/04-PATTERNS.md

# Source files — locus of the harness extension
@tests/uat/extension-page-harness.ts
@tests/uat/lib/harness-page-driver.ts
@tests/uat/harness.test.ts
@tests/uat/lib/launch.ts
@src/offscreen/recorder.ts
@src/background/index.ts

# Prior plan SUMMARYs to mirror — Approach B harness extension precedent
@.planning/phases/02-stabilize-export-pipeline/02-04-SUMMARY.md
@.planning/phases/03-spec-10-smoke-verification-dom-event-log-verification/03-04-SUMMARY.md

<interfaces>
<!-- Key shapes the executor consumes directly. Extracted from codebase + RESEARCH 2026-05-21. -->

From tests/uat/lib/launch.ts:80-90 (HarnessHandles — already exposes browser + extensionId; no extension needed):
```typescript
export interface HarnessHandles {
  readonly browser: Browser;             // ← already exposed; used by driveA33
  readonly extensionId: string;          // ← already exposed; used by driveA33
  readonly harnessPage: Page;
  readonly victimPage: Page;
  readonly downloadsDir: string;
  readonly swConsole: string[];
  readonly offConsole: string[];
}
```

From RESEARCH Q2 Code Example Pattern 1 (stopServiceWorker — NEW helper; verbatim from Chrome devrel doc):
```typescript
import type { Browser } from 'puppeteer';

/**
 * Force-terminate the MV3 service worker via Puppeteer CDP. Required
 * because Puppeteer's persistent CDP attach keeps SWs alive indefinitely;
 * natural 30s idle eviction does NOT fire under test conditions per Chrome
 * docs (https://developer.chrome.com/docs/extensions/develop/concepts/service-workers/lifecycle).
 *
 * Reference: https://developer.chrome.com/docs/extensions/how-to/test/test-serviceworker-termination-with-puppeteer
 */
async function stopServiceWorker(browser: Browser, extensionId: string): Promise<void> {
  const host = `chrome-extension://${extensionId}`;
  const target = await browser.waitForTarget(
    (t) => t.type() === 'service_worker' && t.url().startsWith(host),
  );
  const worker = await target.worker();
  if (worker !== null) {
    await worker.close();
  }
}
```

SAVE_ARCHIVE dispatch (REVISION iter-2 — Option B per plan-checker BLOCKER 2):
The harness page realm has `chrome.runtime` available (it's an extension-page realm — `chrome-extension://<id>/tests/uat/extension-page-harness.html`). The canonical SAVE_ARCHIVE dispatch is `chrome.runtime.sendMessage({type: 'SAVE_ARCHIVE'}, callback)` — used by 9 existing assertions (A5/A11/A12/A13/A26/A28/A29/A30/A31; verified via `grep "type: 'SAVE_ARCHIVE'" tests/uat/extension-page-harness.ts`). The `__mokoshHarness` surface is `assertA1..A31 + getManifestVersion`; there is NO `dispatchSaveArchive` helper. We do NOT add one — driveA33 dispatches SAVE_ARCHIVE directly via `page.evaluate` + a promise-wrapped `chrome.runtime.sendMessage` callback. This matches Plan 04-05's approach (which dispatches SAVE_ARCHIVE from inside an existing assertA34 method that runs in the harness page realm).

From RESEARCH Q2 Code Example Pattern 4 (driveA33 — host-side body; REVISION iter-2 inline-dispatched SAVE):
```typescript
const A33_IDLE_WAIT_MS = 5 * 60 * 1000;           // 300_000 — real wall-clock
const A33_NEW_SW_BOOT_MS = 500;                   // post-worker.close() settle
const A33_OVERALL_TIMEOUT_MS = A33_IDLE_WAIT_MS + 60_000;  // 360_000
const A33_SAVE_ARCHIVE_TIMEOUT_MS = 15_000;

export async function driveA33(
  page: Page,
  browser: Browser,
  extensionId: string,
  downloadsDir: string,
): Promise<AssertionRecord> {
  const r: AssertionRecord = { name: 'A33', passed: false, checks: [], diagnostics: [] };

  // Step 1: prime recording on the probe tab via the existing harness primitive.
  // setupFreshRecording is a module-internal helper inside extension-page-harness.ts
  // and is reachable from page.evaluate only if exposed; in practice driveA33 calls
  // assertA1 (or an equivalent existing harness method that primes a fresh recording)
  // OR a thin Plan-04-04 page-side wrapper if the prior arts don't suffice. Verify
  // in Task 2 read_first which existing assertA* method delivers a fresh-recording
  // SUT and reuse it directly (no new harness method needed).
  await page.evaluate(async () => {
    // Reuse the same fresh-recording primitive that A5/A26/A30/A31 use as their Step 1.
    // The exact call depends on whether setupFreshRecording is exposed; if not, A33's
    // first step calls __mokoshHarness.assertA1 (which is the canonical "fresh
    // recording bootstrap" assertion in the harness surface).
    const harness = (window as { __mokoshHarness: { assertA1: () => Promise<unknown> } }).__mokoshHarness;
    await harness.assertA1();
  });

  // Step 2: 5-min wall-clock idle
  r.diagnostics.push(`waiting ${A33_IDLE_WAIT_MS}ms for SW idle window`);
  await new Promise((res) => setTimeout(res, A33_IDLE_WAIT_MS));

  // Step 3: force SW termination via CDP
  await stopServiceWorker(browser, extensionId);
  r.diagnostics.push('SW terminated via worker.close()');

  // Step 4: brief settle for SW teardown
  await new Promise((res) => setTimeout(res, A33_NEW_SW_BOOT_MS));

  // Step 5: dispatch SAVE_ARCHIVE via chrome.runtime.sendMessage from the harness
  // page realm — matches the established A5/A11/A12/A13/A26/A28/A29/A30/A31 pattern.
  // Wakes SW back up as an event (event-driven respawn is the canonical MV3 wakeup path).
  const saveResult = await page.evaluate(
    (timeoutMs: number) =>
      new Promise<{ success: boolean; error?: string }>((resolve) => {
        const timer = setTimeout(() => {
          resolve({ success: false, error: `SAVE_ARCHIVE timed out after ${timeoutMs}ms` });
        }, timeoutMs);
        chrome.runtime.sendMessage({ type: 'SAVE_ARCHIVE' }, (response: unknown) => {
          clearTimeout(timer);
          if (chrome.runtime.lastError !== undefined) {
            resolve({ success: false, error: String(chrome.runtime.lastError.message) });
            return;
          }
          resolve(response as { success: boolean; error?: string });
        });
      }),
    A33_SAVE_ARCHIVE_TIMEOUT_MS,
  );

  r.checks.push({
    name: 'A33.1: SAVE_ARCHIVE ack success after 5-min idle + SW kill',
    expected: true,
    actual: saveResult.success,
    passed: saveResult.success === true,
  });

  // Step 6: verify zip contains non-empty video buffer
  const zipPath = findLatestZip(downloadsDir);
  if (zipPath === null) {
    r.checks.push({ name: 'A33.0: zip present', expected: '>=1 zip', actual: 'none', passed: false });
    r.passed = false;
    return r;
  }
  const zip = await JSZip.loadAsync(readFileSync(zipPath));
  const videoEntry = zip.file('video/last_30sec.webm');
  const videoSize = videoEntry !== null
    ? (await videoEntry.async('uint8array')).byteLength
    : 0;
  r.checks.push({
    name: 'A33.2: video/last_30sec.webm size > 0 (buffer survived SW eviction)',
    expected: '>0',
    actual: String(videoSize),
    passed: videoSize > 0,
  });
  r.checks.push({
    name: 'A33.3: video size > 100 KB (sanity floor; real archives 1-3 MB)',
    expected: '>100000',
    actual: String(videoSize),
    passed: videoSize > 100_000,
  });

  r.passed = r.checks.every((c) => c.passed);
  return r;
}
```

From RESEARCH Q2 sub-question (c) env-gate recommendation:
```typescript
// In tests/uat/harness.test.ts drivers-array entry:
{
  name: 'A33',
  drive: process.env.SKIP_LONG_UAT === '1'
    ? async (): Promise<AssertionRecord> => ({
        name: 'A33',
        passed: true,
        checks: [],
        diagnostics: ['A33 SKIPPED (SKIP_LONG_UAT=1; unset to run 5-min idle test)'],
      })
    : driveA33Wrapped,
},
```

Default polarity: SKIP_LONG_UAT unset → RUN A33 (this matches the closure + alpha-gate semantics; per-commit dev iteration uses SKIP_LONG_UAT=1).

From src/offscreen/recorder.ts:91 (architecture invariant — segments only in offscreen RAM):
```typescript
let segments: Blob[] = [];  // module-level state; NO chrome.storage.local persistence; NO IndexedDB
```

The spike's job: verify this RAM-only design survives a 5-min SW idle. If it does, ROADMAP SC #1 is satisfied with ZERO source-code changes — just a new harness assertion that exercises the path.
</interfaces>
</context>

<tasks>

<task type="auto">
  <name>Task 1: Wave 0 SPIKE — empirical verification that offscreen survives 5-min SW idle</name>
  <files>tests/uat/lib/harness-page-driver.ts</files>
  <read_first>tests/uat/lib/harness-page-driver.ts (full; ~2200 lines — read selectively: imports lines 1-40, findLatestZip ~1395, driveA30 host-side filter ~2039-2148), tests/uat/extension-page-harness.ts:3932-4021 (__mokoshHarness global registration block — confirm available surface BEFORE writing spike), tests/uat/extension-page-harness.ts:600-700 (setupFreshRecording helper context), src/offscreen/recorder.ts:80-100 (segments array context), .planning/phases/04-harden-clean-up-optional/04-RESEARCH.md Q2 sub-question (b), .planning/phases/01-stabilize-video-pipeline/01-07-SUMMARY.md (Plan 07 spike precedent)</read_first>
  <action>
    1. Add the `stopServiceWorker(browser, extensionId)` helper to `tests/uat/lib/harness-page-driver.ts` per the Code Example in `<interfaces>` above. Place it near the top of the file (after existing imports + before existing driveA-* functions). Add the `import type { Browser } from 'puppeteer';` if not already present.

    2. Create a one-shot spike script `tests/uat/spike-a33-sw-persistence.ts` (NEW; treat as scratch file for this spike — delete after spike concludes; record outcome in plan SUMMARY). The script:
       - Imports `launchHarnessBrowser` from `./lib/launch.ts`.
       - Imports `stopServiceWorker` + `findLatestZip` from `./lib/harness-page-driver.ts`.
       - Launches the harness browser.
       - **Prime recording (REVISION iter-2 — Option B; no `dispatchSaveArchive` helper exists on `__mokoshHarness`):** call the existing fresh-recording primitive via `await handles.harnessPage.evaluate(() => (window as { __mokoshHarness: { assertA1: () => Promise<unknown> } }).__mokoshHarness.assertA1());`. The Task 1 read_first MUST verify that `__mokoshHarness.assertA1` is the canonical fresh-recording bootstrap (it is per the existing harness — `Harness ready. window.__mokoshHarness.{assertA1..A31, getManifestVersion} available.`); if a different assertA* method is more direct for "prime + leave recording active for 5 min", choose that instead and document in the spike script comment.
       - `console.log('SPIKE: waiting 5 minutes for SW idle window...')`
       - `await new Promise(r => setTimeout(r, 5 * 60 * 1000));`
       - `await stopServiceWorker(handles.browser, handles.extensionId);`
       - `await new Promise(r => setTimeout(r, 500));` (settle)
       - **Dispatch SAVE_ARCHIVE (REVISION iter-2 — Option B; canonical chrome.runtime.sendMessage from harness page realm):**
         ```typescript
         const saveResult = await handles.harnessPage.evaluate(
           (timeoutMs: number) =>
             new Promise<{ success: boolean; error?: string }>((resolve) => {
               const timer = setTimeout(() => {
                 resolve({ success: false, error: `SAVE_ARCHIVE timed out after ${timeoutMs}ms` });
               }, timeoutMs);
               chrome.runtime.sendMessage({ type: 'SAVE_ARCHIVE' }, (response: unknown) => {
                 clearTimeout(timer);
                 if (chrome.runtime.lastError !== undefined) {
                   resolve({ success: false, error: String(chrome.runtime.lastError.message) });
                   return;
                 }
                 resolve(response as { success: boolean; error?: string });
               });
             }),
           15_000,
         );
         console.log(`SPIKE: SAVE_ARCHIVE ack -> ${JSON.stringify(saveResult)}`);
         ```
       - `await new Promise(r => setTimeout(r, 5000));` (let download complete)
       - `const zipPath = findLatestZip(handles.downloadsDir);`
       - `const zip = await JSZip.loadAsync(readFileSync(zipPath));`
       - `const videoEntry = zip.file('video/last_30sec.webm');`
       - `const videoSize = videoEntry ? (await videoEntry.async('uint8array')).byteLength : 0;`
       - `console.log(\`SPIKE RESULT: videoSize=${videoSize} bytes (>0 = OFFSCREEN SURVIVED; =0 = OFFSCREEN DIED)\`);`
       - `await handles.browser.close();`

    3. Run the spike: `tsx tests/uat/spike-a33-sw-persistence.ts` with HEADLESS=1 (so it runs in CI mode; ~5 min wall-clock).

    4. Record the result. If videoSize > 100_000 → SPIKE PASSED (offscreen survives) → proceed to Task 2 with verification-only A33. If videoSize ≤ 100_000 OR throw → SPIKE FAILED → SUMMARY documents the failure mode + flag to plan-checker for re-planning (IndexedDB persistence work would expand Plan 04-04 substantially; that's a planning event, not an execution event).

    5. Commit the `stopServiceWorker` helper (Task 1's persisting artifact). The spike script is OK to delete OR keep committed as `tests/uat/spike-*.ts` for future SW-lifecycle investigations.
  </action>
  <verify>
    <automated>npx tsc --noEmit && HEADLESS=1 tsx tests/uat/spike-a33-sw-persistence.ts 2>&1 | tee /tmp/04-04-spike.log; grep -c 'SPIKE RESULT' /tmp/04-04-spike.log</automated>
  </verify>
  <acceptance_criteria>
    - `stopServiceWorker(browser, extensionId)` helper exists at `tests/uat/lib/harness-page-driver.ts` with the canonical Chrome devrel signature (`Browser` + extensionId args; `target.worker()?.close()` body).
    - Spike script ran to completion (no Puppeteer throw).
    - Spike result logged with explicit `videoSize=<N> bytes` line.
    - Spike SAVE_ARCHIVE dispatch uses `chrome.runtime.sendMessage({type: 'SAVE_ARCHIVE'}, ...)` directly (NOT a non-existent `__mokoshHarness.dispatchSaveArchive()` call); verify by `grep -c 'dispatchSaveArchive' tests/uat/spike-a33-sw-persistence.ts` returns 0 AND `grep -c "type: 'SAVE_ARCHIVE'" tests/uat/spike-a33-sw-persistence.ts` returns ≥ 1.
    - If videoSize > 100_000: spike PASSED; proceed to Task 2 with verification-only path.
    - If videoSize ≤ 100_000: spike FAILED; pause plan + flag to plan-checker for re-planning (out of scope for this task to escalate, but SUMMARY documents).
    - Total spike wall-clock: ~6-7 minutes (5 min idle + ~1-2 min orchestration).
  </acceptance_criteria>
  <done>Spike outcome recorded in plan SUMMARY; stopServiceWorker helper committed. Atomic commit: `feat(04-04): Wave 0 spike — stopServiceWorker helper + 5-min SW idle empirical result`.</done>
</task>

<task type="auto">
  <name>Task 2: Wave 1 — A33 assertion + driveA33 + orchestrator wiring (assumes spike PASSED)</name>
  <files>tests/uat/extension-page-harness.ts, tests/uat/lib/harness-page-driver.ts, tests/uat/harness.test.ts</files>
  <read_first>tests/uat/extension-page-harness.ts:3517-3636 (assertA30 — canonical setupFreshRecording + SAVE pattern), tests/uat/extension-page-harness.ts:3878-3917 (assertA31 — most-recent chrome.runtime.sendMessage SAVE_ARCHIVE pattern; copy this), tests/uat/extension-page-harness.ts:3932-4021 (__mokoshHarness global registration block — confirm NO new method added per REVISION iter-2), tests/uat/lib/harness-page-driver.ts:2039-2148 (driveA30 — host-side filter pattern), tests/uat/harness.test.ts:100-110 (import block), tests/uat/harness.test.ts:340-360 (wrapped-driver block), tests/uat/harness.test.ts:459-486 (drivers-array push block), tests/uat/harness.test.ts:225-240 (SKIP_PROD_REBUILD env-gate pattern)</read_first>
  <action>
    **GATING CONDITION:** Task 1 spike produced videoSize > 100_000. (If FAILED, this task is BLOCKED and the plan must be re-planned to add IndexedDB persistence work.)

    3-file lockstep update per the Approach B harness extension pattern:

    **File 1: tests/uat/extension-page-harness.ts**
    - REVISION iter-2 — Option B per plan-checker BLOCKER 2: the existing `__mokoshHarness` surface is `assertA1..A31 + getManifestVersion`; `dispatchSaveArchive` does NOT exist and we do NOT add it. SAVE_ARCHIVE dispatch happens directly via `chrome.runtime.sendMessage` inside driveA33's `page.evaluate` (matches the established assertA31 pattern at lines 3886-3890).
    - Decision: NO new page-side function. driveA33 (host-side) drives Step 1 (prime) by calling an existing `__mokoshHarness.assertA<N>` method that bootstraps a fresh recording (confirm in read_first which existing assertA* is the canonical "prime fresh recording" entrypoint — `assertA1` is the leading candidate; falling back to `assertA5`/`assertA26` if a more direct method matches the spike's actual call site). Step 5 (SAVE) uses inline `chrome.runtime.sendMessage` per the `<interfaces>` block above.
    - Verify no edits needed to `__mokoshHarness` registration block (lines 3932-4015): the surface stays at 31 assertA* + getManifestVersion. The Tier-1 FORBIDDEN_HOOK_STRINGS inventory stays at 12 entries (no new test-only symbol).
    - If, during read_first, the planner determines that NONE of the existing assertA* methods deliver "prime + leave recording active for ≥5 min", THEN add a thin page-side primer `primeForA33` that calls existing production-surface APIs (REQUEST_PERMISSIONS → START_RECORDING via chrome.runtime.sendMessage); this is a deviation from Option B and must be flagged in the SUMMARY. Per RESEARCH note (FORBIDDEN_HOOK_STRINGS stays at 12): NO new test-only `__MOKOSH_UAT__`-gated symbol; any new page-side helper uses production APIs only.

    **File 2: tests/uat/lib/harness-page-driver.ts**
    - Append `driveA33` function per RESEARCH Code Example Pattern 4 (full body in `<interfaces>` above; REVISION iter-2 inline-dispatched SAVE).
    - Place it after the existing driveA32 (which is the most-recent Phase 3 addition).
    - Verify the `stopServiceWorker` helper from Task 1 is in scope (same file).
    - Filter-pipeline form; no `continue`; typed function signature `(page, browser, extensionId, downloadsDir) => Promise<AssertionRecord>` per the new 4-arg shape.
    - Add `import { readFileSync } from 'node:fs';` + `import JSZip from 'jszip';` if not already present (they should be — these are reused from driveA29/30/31).
    - The Step-5 SAVE_ARCHIVE inline `page.evaluate` block uses `chrome.runtime.sendMessage({type: 'SAVE_ARCHIVE'}, callback)` per the `<interfaces>` Code Example; verify by `grep -c "type: 'SAVE_ARCHIVE'" tests/uat/lib/harness-page-driver.ts` increases by ≥ 1 vs pre-edit baseline (was 0 in driveA29/30/31 because those call sites are inside extension-page-harness.ts assertA* methods; A33 is unique in dispatching from the host-side via page.evaluate).

    **File 3: tests/uat/harness.test.ts**
    - Import: add `driveA33,` to the import block at ~line 101 (alongside `driveA29`-`driveA32`).
    - Wrapped-driver: add at ~line 357 (after `driveA31Wrapped`):
      ```typescript
      // Plan 04-04 — driveA33 needs Browser + extensionId for CDP-based SW kill
      //             AND downloadsDir for host-side JSZip parse of post-restart zip.
      const driveA33Wrapped: (page: import('puppeteer').Page) => Promise<AssertionRecord> =
        (page) => driveA33(page, handles.browser, handles.extensionId, handles.downloadsDir);
      ```
    - Drivers-array push: add at ~line 486 (after the existing A32 entry):
      ```typescript
      // Plan 04-04 A33: SW state persistence 5-min idle (ROADMAP SC #1; RESEARCH Q2).
      // Forces SW eviction via Puppeteer CDP worker.close() per the canonical
      // Chrome devrel pattern (RESEARCH Pattern 1). Verifies offscreen-RAM
      // segments survive SW restart. Env-gated by SKIP_LONG_UAT for fast
      // per-commit iteration; defaults to RUN for Phase 4 closure + alpha gate.
      {
        name: 'A33',
        drive: process.env.SKIP_LONG_UAT === '1'
          ? async (): Promise<AssertionRecord> => ({
              name: 'A33',
              passed: true,
              checks: [],
              diagnostics: ['A33 SKIPPED (SKIP_LONG_UAT=1; unset to run 5-min idle test)'],
            })
          : driveA33Wrapped,
      },
      ```

    Verify:
    - `npx tsc --noEmit` exits 0.
    - `npm run build:test` exits 0.
    - Quick UAT: `HEADLESS=1 SKIP_PROD_REBUILD=1 SKIP_LONG_UAT=1 npm run test:uat` exits 0 with 34/34 GREEN (A33 SKIPPED message visible; preserves baseline + adds A33 skip placeholder).
    - Full UAT: `HEADLESS=1 SKIP_PROD_REBUILD=1 npm run test:uat` exits 0 with 34/34 GREEN (A33 actually runs ~6 min wall-clock; A33.1 SAVE ack + A33.2 size > 0 + A33.3 size > 100 KB all PASS).
    - Tier-1 FORBIDDEN_HOOK_STRINGS check: `grep -c 'FORBIDDEN_HOOK_STRINGS' tests/uat/harness.test.ts tests/background/no-test-hooks-in-prod-bundle.test.ts` — verify the inventory count in both files unchanged (preserves the 12-entry invariant per CONTEXT §"Claude's Discretion").
    - REVISION iter-2 gate: `grep -c 'dispatchSaveArchive' tests/uat/lib/harness-page-driver.ts tests/uat/extension-page-harness.ts tests/uat/harness.test.ts` returns 0 (the non-existent helper is NOT introduced).
  </action>
  <verify>
    <automated>npx tsc --noEmit && npm run build:test && HEADLESS=1 SKIP_PROD_REBUILD=1 SKIP_LONG_UAT=1 npm run test:uat 2>&1 | tail -5 | tee /tmp/04-04-task-2-skip.log; grep -c '34/34' /tmp/04-04-task-2-skip.log; grep -c 'dispatchSaveArchive' tests/uat/lib/harness-page-driver.ts tests/uat/extension-page-harness.ts tests/uat/harness.test.ts</automated>
  </verify>
  <acceptance_criteria>
    - `npx tsc --noEmit` exits 0.
    - `npm run build:test` exits 0.
    - UAT harness count flips 33 → 34 (A33 added).
    - Skip-mode run: `HEADLESS=1 SKIP_PROD_REBUILD=1 SKIP_LONG_UAT=1 npm run test:uat` GREEN 34/34 (A33 SKIPPED placeholder GREEN; total takes ~95s — unchanged).
    - Full-mode run: `HEADLESS=1 SKIP_PROD_REBUILD=1 npm run test:uat` GREEN 34/34 (~6.5 min; A33 actually runs and passes A33.1 + A33.2 + A33.3).
    - `grep -c 'A33' tests/uat/harness.test.ts` returns ≥ 4 (import + wrapped + push + comment banner).
    - `grep -c 'SKIP_LONG_UAT' tests/uat/harness.test.ts` returns ≥ 2 (env-gate + comment).
    - FORBIDDEN_HOOK_STRINGS count unchanged at 12 (no new test-only symbols introduced per CONTEXT §"Claude's Discretion"; verify by `wc -l` of the inventory arrays).
    - REVISION iter-2 gate (Option B): `grep -c 'dispatchSaveArchive' tests/uat/` returns 0 across all harness files; SAVE_ARCHIVE dispatched via `chrome.runtime.sendMessage({type: 'SAVE_ARCHIVE'}, ...)` only.
  </acceptance_criteria>
  <done>A33 lands; UAT 33→34 GREEN; SW persistence empirically verified at 5-min idle scale. Atomic commit: `feat(04-04): Wave 1 — A33 SW state persistence harness assertion (34/34 GREEN)`.</done>
</task>

</tasks>

<threat_model>
## Trust Boundaries

| Boundary | Description |
|----------|-------------|
| Puppeteer CDP → Chrome MV3 SW realm | `worker.close()` invokes the SW's `self.close()` via CDP `ServiceWorker.unregister` — this terminates the SW realm but does NOT touch the offscreen document's WebContents target. Native CDP surface; no untrusted input. |
| Test idle interval (5 min wall-clock) → MediaRecorder active segment buffer | the MediaRecorder is in `state === 'recording'` during the idle; segments rotate every 10s; the offscreen-RAM array accumulates 30 segments (5 min × 60 sec / 10 sec per segment); trim-to-last-3 keeps memory bounded ≤ 30 MB (well under CON-ram-ceiling) |

## STRIDE Threat Register

| Threat ID | Category | Component | Disposition | Mitigation Plan |
|-----------|----------|-----------|-------------|-----------------|
| T-04-04-01 | Tampering | a future architectural change might move the segments array to SW-side state, breaking the offscreen-survives-SW assumption A33 verifies | mitigate | A33 is a regression-catching gate; if a future PR moves segments off-offscreen, A33 fails fast (videoSize=0 after SW kill) |
| T-04-04-02 | DoS (CI) | A33's 5-min idle adds ~5 min to harness wall-clock (95s → 395s); per-commit CI lanes would suffer | mitigate | Env-gated by SKIP_LONG_UAT (default RUN for closure + alpha; documented per-commit SKIP_LONG_UAT=1 for dev iteration) |
| T-04-04-03 | Repudiation | natural 30s idle eviction does NOT fire under Puppeteer's CDP attach per Chrome docs; if a developer naively writes "wait 5 min and hope SW dies" the test silently passes via a SW that never died | mitigate | The CDP `worker.close()` call is explicit + cited in code comment; RESEARCH Pitfall 4 documents the misconception |
| T-04-04-04 | Spoofing | Puppeteer 25.x patch versions could in theory change `Worker.close()` semantics; the canonical Chrome devrel pattern is pinned at Puppeteer ≥22.1.0 | accept | The project pin `puppeteer: ^25.0.2` is comfortably past the 22.1.0 floor; minor patch drift expected to be backward-compatible per Puppeteer's semver discipline. If A33 ever fails post-Puppeteer-upgrade, the SUMMARY's commit ref provides the exact Puppeteer version where it was validated. |
</threat_model>

<verification>
- `npx tsc --noEmit` exits 0.
- `npm run build:test` exits 0.
- UAT harness count 33 → 34.
- Skip-mode: `HEADLESS=1 SKIP_LONG_UAT=1 npm run test:uat` GREEN 34/34 in ~95s.
- Full-mode: `HEADLESS=1 npm run test:uat` GREEN 34/34 in ~6.5 min.
- ROADMAP SC #1 GREEN — A33 empirical evidence: video buffer survived 5-min SW idle + worker.close().
- FORBIDDEN_HOOK_STRINGS count unchanged at 12.
- vitest baseline preserved (≥ 181 GREEN from Plans 04-01 + 04-02).
- A29 + A30 + A31 + A32 unchanged (no regression to existing assertions).
- REVISION iter-2 invariant: `grep -c 'dispatchSaveArchive' tests/uat/` returns 0 across spike script + harness files.
</verification>

<success_criteria>
- Wave 0 spike PASSED — empirical evidence that offscreen survives 5-min SW idle (Task 1).
- assertA33 + driveA33 + stopServiceWorker helper + harness orchestrator wiring landed (Task 2).
- UAT harness count 33 → 34 GREEN.
- ROADMAP SC #1 (SW state persistence) GREEN.
- Env-gated long-test pattern established (SKIP_LONG_UAT) — pattern reused by any future ≥5-min test.
- Pre-checkpoint bundle gates 6/6 PASS unchanged (Plan 04-04 makes no source-code changes).
</success_criteria>

<output>
After completion, create `.planning/phases/04-harden-clean-up-optional/04-04-SUMMARY.md` capturing:
- Spike outcome (videoSize value + interpretation; SPIKE PASSED/FAILED tag)
- stopServiceWorker helper diff (full body)
- driveA33 diff (full body; inline chrome.runtime.sendMessage SAVE per REVISION iter-2 Option B)
- Orchestrator wiring diff (3 sites in harness.test.ts)
- SKIP_LONG_UAT env-gate decision (default RUN; rationale)
- UAT before/after (33/33 → 34/34)
- Full-mode wall-clock benchmark (e.g., ~6.5 min)
- ROADMAP SC #1 closure evidence
- Commit refs (Task 1 spike + Task 2 impl)
- If spike FAILED: detailed failure mode + flag for re-planning (this branch is unlikely per RESEARCH MEDIUM-confidence; document as ALPHA-PATH-NOT-TAKEN)
</output>
</content>