---
phase: 01-stabilize-video-pipeline
plan: 07
type: execute
wave: 3
depends_on: ["05", "06"]
files_modified:
- tests/fixtures/last_30sec.webm
autonomous: false
requirements:
- REQ-video-ring-buffer
requirements_addressed:
- REQ-video-ring-buffer
must_haves:
truths:
- "`dist/` loads unpacked into Chrome ≥ 116 with no errors at chrome://extensions"
- "On extension load, the offscreen prompts the operator with Chrome's native screen-share picker"
- "After selecting a screen/window, Chrome's permanent 'Sharing your screen' indicator is shown (D-04 accepted residual risk)"
- "After ≥ 30 s of recording, clicking SAVE_ARCHIVE produces a `session_report_*.zip` in the user's Downloads folder containing `video/last_30sec.webm`"
- "Extracting that webm and running `ffprobe -v error -f matroska -i last_30sec.webm` exits 0 with EMPTY stderr (D-12 acceptance gate)"
- "The captured webm is preserved at `tests/fixtures/last_30sec.webm` as a known-good sample for regression"
- "If the ffprobe gate FAILS, the developer is presented with the D-13 fallback skeleton (already pre-staged in src/offscreen/recorder.ts) and re-plans Phase 1 closure"
artifacts:
- path: "tests/fixtures/last_30sec.webm"
provides: "Captured + verified ffprobe-clean WebM sample"
contains: ""
key_links:
- from: "dist/manifest.json"
to: "dist/src/offscreen/index.html (or dist/offscreen/index.html — whichever Plan 06 produced)"
via: "chrome.offscreen.createDocument at runtime"
pattern: "DISPLAY_MEDIA"
- from: "ffprobe -v error"
to: "last_30sec.webm playability"
via: "exit code 0 + empty stderr"
pattern: ""
---
Manual smoke test against a real Chrome instance + ffprobe verification of
the assembled `last_30sec.webm`. This is the D-12 acceptance gate that
ROADMAP.md Phase 1 Success Criterion #3 names verbatim ("concatenating
header + buffered chunks yields a byte sequence a browser would play").
Purpose: Code-level verification (vitest, tsc) covers correctness of pure
logic and types. Playability of an actual WebM file requires running the
extension against a real browser, recording, and exporting. Phase 1 is not
"done" until ffprobe accepts the output — that is the criterion the
plan-checker is bound to enforce.
This plan is **`autonomous: false`** — it has a `checkpoint:human-verify`
because driving a Chrome instance unpacked-load + screen-picker click +
manual save click is not a CLI operation. ffprobe IS a CLI tool; the
ffprobe step is automated AFTER the human steps complete.
The plan also handles the BRANCH where ffprobe fails — in that case, the
developer SHALL switch to the D-13 fallback (already pre-staged as a
commented skeleton in `src/offscreen/recorder.ts`) and the orchestrator
re-plans Phase 1 closure as a follow-up.
Output: A committed `tests/fixtures/last_30sec.webm` known-good sample (if
ffprobe passes) OR a clear escalation to D-13 (if ffprobe fails).
@$HOME/.claude/get-shit-done/workflows/execute-plan.md
@$HOME/.claude/get-shit-done/templates/summary.md
@.planning/phases/01-stabilize-video-pipeline/01-CONTEXT.md
@.planning/phases/01-stabilize-video-pipeline/01-RESEARCH.md
@.planning/phases/01-stabilize-video-pipeline/01-VALIDATION.md
@.planning/phases/01-stabilize-video-pipeline/01-PATTERNS.md
@dist/manifest.json
@src/offscreen/recorder.ts
This plan runs against a built `dist/` from Plan 06. No new code interfaces
are introduced. The plan exercises the contracts already implemented:
- SW `ensureOffscreen` → `chrome.offscreen.createDocument({reasons: [DISPLAY_MEDIA]})`
- Offscreen on START_RECORDING: `getDisplayMedia` → MediaRecorder.start(2000)
- Offscreen ondataavailable: addChunk(blob, Date.now())
- Offscreen track.onended: clear buffer + RECORDING_ERROR to SW
- SW on SAVE_ARCHIVE: getVideoBufferFromOffscreen() via port → REQUEST_BUFFER → BUFFER → mergeVideoChunks → JSZip → chrome.downloads.download
Manual interactions:
1. `chrome://extensions` → "Load unpacked" → select `dist/`
2. (Whatever currently triggers `startVideoCapture` — the existing flow in `src/popup/index.ts` opens the popup, which probably calls `REQUEST_PERMISSIONS`. The exact UX wire is Phase 3 territory, so Plan 07 accepts that the trigger may be clunky — the goal is to get a recording running and exporting.)
3. Click through Chrome's native screen picker
4. Wait ≥ 35 seconds (so the buffer has rotated at least once and the trim path executed)
5. Click whatever in the popup triggers SAVE_ARCHIVE (likely "Сохранить отчёт об ошибке")
6. Verify a `session_report_*.zip` lands in Downloads
7. Extract and run ffprobe
## Trust Boundaries
| Boundary | Description |
|----------|-------------|
| developer's local Chrome ↔ unpacked extension | The developer is running the extension on their own machine; loading unpacked is the recommended Phase 1 deployment per audit P2 #22 (no CI). No new trust boundary introduced by this plan. |
## STRIDE Threat Register
| Threat ID | Category | Component | Disposition | Mitigation Plan |
|-----------|----------|-----------|-------------|-----------------|
| T-1-NEW-07-01 | Information Disclosure — committed test fixture | `tests/fixtures/last_30sec.webm` | mitigate | The developer chooses what to share with Chrome's picker when running the smoke test. Recommended share: a fresh test browser window with no sensitive content (e.g., the README rendered in Markdown preview). The committed fixture MUST NOT contain identifying information. Reviewer guidance documented in SUMMARY. |
Task 1: HUMAN — load dist/ unpacked, capture a recording, export, save the archive
tests/fixtures/last_30sec.webm
- dist/manifest.json (verify the build is loadable)
- src/popup/index.html and src/popup/index.ts (understand the existing UX trigger — Phase 3 reworks this; for now, click whatever button exists)
This is a `checkpoint:human-verify` task. The "action" is the manual sequence below — Claude prepares the environment (pre-flight checks), the human drives Chrome, then Claude resumes for the ffprobe gate. Full details in `` and `` below.
Step summary:
1. Run pre-flight automation (manifest check + ffprobe presence check).
2. Pause and hand off to the human: load dist/ unpacked, click through the popup, accept the screen-share picker, wait ≥ 35 s, click save.
3. Resume automation: extract the saved archive, run `ffprobe -v error -f matroska -i last_30sec.webm`, capture the exit code + stderr.
4. If ffprobe exit 0 with no stderr lines: commit the captured WebM as `tests/fixtures/last_30sec.webm` (regression fixture for future phases).
5. If ffprobe fails: escalate to the orchestrator with the D-13 fallback signal.
The full Phase-1 Mokosh build:
- `dist/manifest.json` declares `desktopCapture` (not `tabCapture`)
- The offscreen document calls `getDisplayMedia` once per session
- Captured stream is recorded by a single `MediaRecorder` at vp9 / 400 kbps / 2 s timeslice
- Buffer holds first chunk (WebM header) + every chunk newer than now-30s
- SW receives the buffer over the `'video-keepalive'` port on SAVE_ARCHIVE
- Archive is delivered via `chrome.downloads`
which ffprobe && ffprobe -v error -f matroska -i tests/fixtures/last_30sec.webm; test $? -eq 0
**Pre-flight (automated, do these in a bash shell before opening Chrome):**
```bash
# Confirm the build is fresh
ls -la dist/manifest.json dist/assets/*.js
node -e "const m=require('./dist/manifest.json'); if(!m.permissions.includes('desktopCapture')) process.exit(1); if(m.permissions.includes('tabCapture')) process.exit(2); console.log('manifest OK')"
# Confirm ffprobe is installed
which ffprobe || { echo "ffprobe missing — install with: sudo apt install ffmpeg"; exit 1; }
ffprobe -version | head -1
```
**Manual smoke (in Chrome):**
1. Open `chrome://extensions` in a Chromium-based browser (Chrome / Chromium ≥ 116).
2. Enable "Developer mode" (top right).
3. Click "Load unpacked".
4. Select the project's `dist/` directory.
5. Confirm the extension card shows the name "AI Call Recorder" (from `manifest.json`) with NO errors or warnings (red text in the card == FAIL).
6. Open the extension's popup (click the toolbar icon).
7. Click whatever button triggers permission/recording. The popup is from before this phase — Phase 3 reworks it; for Phase 1 the button may say "Сохранить отчёт" or "Запросить разрешения".
8. Chrome SHOULD show its native screen-share picker. **If it doesn't**: open chrome://extensions, click "service worker" on the extension card, inspect the SW console for the `[SW:Main]` log line that should say "Sending START_RECORDING to offscreen..." — that tells us where the flow stalled.
9. Pick "Entire screen" or a specific window (your choice — recommend an innocuous one like a code editor showing this PLAN.md).
10. Confirm Chrome's permanent "Sharing your screen" indicator appears at the top of the screen. (This is the accepted D-04 trade-off.)
11. **Wait at least 35 seconds.** Move the mouse around or scroll a page so the captured stream has actual visual change (helps vp9 produce useful frames).
12. Open the popup again and click the SAVE_ARCHIVE button (likely "Сохранить отчёт об ошибке").
13. Within ~5 seconds, a `session_report_YYYY-MM-DD_HH-MM-SS.zip` SHOULD appear in your Downloads folder.
**Verification of the exported archive:**
```bash
# Find the latest archive
LATEST=$(ls -t ~/Downloads/session_report_*.zip 2>/dev/null | head -1)
echo "Latest archive: $LATEST"
[ -z "$LATEST" ] && { echo "FAIL — no session_report archive found in Downloads"; exit 1; }
# Extract the video
unzip -p "$LATEST" video/last_30sec.webm > /tmp/last_30sec.webm
ls -la /tmp/last_30sec.webm
# Run the D-12 acceptance gate
ffprobe -v error -f matroska -i /tmp/last_30sec.webm
GATE=$?
echo "ffprobe exit: $GATE"
ffprobe -v error -show_format -show_streams /tmp/last_30sec.webm 2>&1 | head -30
```
**The acceptance criterion is `GATE == 0` AND zero lines of stderr from the `ffprobe -v error` invocation above.** If both hold:
- Copy the verified webm into the project as a regression fixture: `cp /tmp/last_30sec.webm tests/fixtures/last_30sec.webm`
- Report success.
**If `GATE != 0`** (the ffprobe gate fails):
- Capture the stderr output and the structural dump (`ffprobe -v error -show_packets -i /tmp/last_30sec.webm 2>&1 | head -50`).
- This is the D-13 escalation: the simple continuous-recorder + age-trim approach didn't survive vp9's keyframe cadence (RESEARCH.md Pitfall 1).
- DO NOT delete the existing recorder code. The D-13 fallback skeleton is already pre-staged in `src/offscreen/recorder.ts` as a commented block. Surface the failure to the orchestrator with a precise summary so a new plan (08?) can be drafted to activate the fallback.
Report ONE of:
- **`approved` + paste the ffprobe exit code (must be 0) + paste any single stderr line if present + confirm `tests/fixtures/last_30sec.webm` is committed.**
- **`ffprobe-failed` + paste the stderr + paste the first 20 lines of `ffprobe -v error -show_packets`. Orchestrator escalates to D-13 fallback re-plan.**
- **`load-failed` + paste the chrome://extensions error text or the SW console error. Orchestrator escalates to a previous-plan revision (likely Plan 05 or 06).**
- **`picker-rejected` + paste the offscreen console line containing `RECORDING_ERROR`. Surface to user as a one-off; retry by reloading the extension card and clicking the popup again.**
Phase 1 is **DONE** when this checkpoint returns `approved`.
- The build at dist/ loads unpacked with NO red error text on the extension card.
- getDisplayMedia picker shows on a popup-triggered start path.
- Chrome's "Sharing your screen" indicator appears.
- After ≥ 35 s of recording, SAVE_ARCHIVE produces a session_report_*.zip in Downloads within 5 seconds.
- ffprobe -v error -f matroska -i last_30sec.webm exits 0 with NO stderr lines.
- tests/fixtures/last_30sec.webm is committed as the regression-fixture.
Phase 1 acceptance gate met. WebM ring buffer ships playable output.
Task 2: AUTOMATED — commit the fixture and update STATE.md
tests/fixtures/last_30sec.webm, .planning/STATE.md
- tests/fixtures/last_30sec.webm (verify it exists from Task 1)
- .planning/STATE.md (current state — phase 1 progress)
This task ONLY runs if Task 1 returned `approved`. If Task 1 returned `ffprobe-failed`, `load-failed`, or `picker-rejected`, SKIP this task — the orchestrator handles the escalation.
Two edits:
(1) Ensure `tests/fixtures/last_30sec.webm` is committed to git. Check its size:
```bash
ls -lh tests/fixtures/last_30sec.webm
# Expected: between ~300 KB and ~2 MB for a 30-second vp9 capture at 400 kbps
```
If the file is suspiciously small (< 100 KB) or large (> 5 MB), STOP and audit — the file may be empty (only the header chunk) or include way more than 30 s of content. The expected bitrate is ~400 kbps × 30 s = ~1.5 MB.
Also verify it's tracked by git:
```bash
git status tests/fixtures/last_30sec.webm
```
(2) Update `.planning/STATE.md` to mark Phase 1 as complete. Use the Edit tool to update the relevant fields:
(2a) `stopped_at:` — replace `"Phase 1 context gathered"` with `"Phase 1 closure: ffprobe acceptance gate passed; tests/fixtures/last_30sec.webm committed"`.
(2b) `last_activity:` — replace the existing line with `2026-05-15 — Phase 1 closure: D-12 ffprobe gate green; ready for Phase 2`.
(2c) `## Current Position` block:
- `Phase: 1 of 5 (Stabilize video pipeline)` → `Phase: 2 of 5 (next: Stabilize DOM + event capture privacy)`
- `Plan: 0 of TBD in current phase` → `Plan: 7 of 7 complete (Phase 1 closed)`
- `Status: Ready to plan` → `Status: Phase 1 complete; ready to plan Phase 2`
(2d) Update the `## Performance Metrics` Phase 1 row (currently `| 1. Stabilize video pipeline | 0 | — | — |`) to reflect 7 plans completed.
(2e) Append a `## Phase 1 Closure Notes` block at the bottom of STATE.md with:
- ffprobe exit code (must be 0)
- Size of the committed fixture
- "Phase 1 outcome: SPEC §10 acceptance criteria #2, #3, #7 are functionally green pending Phase 4 end-to-end smoke verification."
- Note: "Phase 2 owns the DOM/event-capture privacy slice; Phase 3 owns the popup state machine + base64-URL replacement; Phase 4 runs the full SPEC §10 smoke pass."
Do NOT touch the `## Deferred Items` or `## Accumulated Context > ## Decisions` sections — those are owned by the orchestrator across phases.
After both edits, run:
```bash
git status
```
The output should show `tests/fixtures/last_30sec.webm` (new) and `.planning/STATE.md` (modified).
[ -f tests/fixtures/last_30sec.webm ] && grep -q "Phase 1 closure" .planning/STATE.md
- `tests/fixtures/last_30sec.webm` exists, size between 100 KB and 5 MB
- `grep -c "Phase 1 closure" .planning/STATE.md` returns at least 1
- `grep -c "ffprobe acceptance gate passed" .planning/STATE.md` returns at least 1
- `git status .planning/STATE.md tests/fixtures/last_30sec.webm` shows both files staged or modified
Phase 1 closure recorded in STATE.md; regression fixture committed; the orchestrator can transition to Phase 2 cleanly.
After Task 1 returns `approved` AND Task 2 lands:
1. `git log --oneline -5` shows the Phase 1 commits (Plan 01 through Plan 07).
2. `tests/fixtures/last_30sec.webm` is in the repo, size sensible (~1-2 MB).
3. `ffprobe -v error -f matroska -i tests/fixtures/last_30sec.webm; echo $?` prints `0` with no stderr.
4. `.planning/STATE.md` reflects Phase 1 closure.
5. `npx vitest run` exits 0 (8 tests passing).
6. `npx tsc --noEmit` exits 0.
7. `npm run build` exits 0.
8. `grep -RnE "tabCapture|chrome\.alarms|VideoRecorderDB|copy-offscreen|openIndexedDB" src/ vite.config.ts manifest.json | grep -v '^#'` returns nothing.
If ANY of (1)-(8) fail, Phase 1 is NOT closed; orchestrator escalates.
Commit cadence: ONE commit at the end of Task 2 (`feat(01-07): close phase 1 — ffprobe gate green, fixture committed`). The fixture and STATE.md are committed together.
If Task 1 escalates (D-13 fallback or earlier-plan rework needed), this verification block is SKIPPED — the orchestrator carries the escalation.
- Manual smoke + ffprobe gate green
- `tests/fixtures/last_30sec.webm` committed as regression-fixture
- `.planning/STATE.md` records Phase 1 closure
- All grep gates from Plans 01..06 still pass
- Build clean; vitest green; tsc clean
- Phase 2 can begin (the next milestone)