fix(panel): stop live-dot flapping (coalesce frames) + show viewport size (v0.6.3)

[mabry1985]

Field report: "not navigating live, the live dot goes offline, buggy as hell," plus a narrow viewport that doesn't fill the dock.

Stability — the flapping

• Coalescing frame delivery. on_frame now stashes only the newest frame and a sender task drains it. Under load it drops stale frames instead of letting ws.send_bytes block — the backpressure buildup was stalling the socket (frozen → the fleet proxy idle-closes it → the "live" dot flaps offline → reconnect loop). This is the standard "render latest, drop stale" pattern for live video.
• Race input against CDP death. If Chrome's page target goes away (tab closed/replaced), the route breaks and lets the client reconnect to the live page (~1.2s) instead of freezing until an idle-close. Reconnect backoff 2.5s → 1.2s.

Diagnostics for the size issue

The default viewport is 1280×577 (wide), so a narrow strip means the panel-fit resize is landing small. To stop guessing:

• The "live" pill now shows the real viewport dims — live · <w>×<h>. So we can see what the browser is rendering at.
• Resize self-corrects — re-sent ~500ms after connect, in case the dock hadn't finished laying out when the socket opened.

Validation

47 host-free tests pass; nav re-arm + resize re-validated live against the real browser; ruff clean.

Next

Once this is in, the live · W×H readout tells us whether the fill is the real problem or the site's own layout — I'll target the fix from there.

🤖 Generated with Claude Code

[protoquinn]

👀 Quinn is reviewing — verdict (PASS / WARN / FAIL) + findings to follow.

[protoquinn]

QA Audit — PR #17 | fix(panel): stop live-dot flapping (coalesce frames) + show viewport size (v0.6.3)

VERDICT: WARN (non-blocking — CI test still queued; re-evaluate once terminal)

---

CI Status

• test: queued

Diff Review

• Frame coalescing in browser_panel.py:97-123 — on_frame stashes newest frame into a shared latest dict, signals an asyncio.Event, and a sender() coroutine drains it. Stale frames are dropped when sender is blocked on ws.send_bytes — correct "render latest, drop stale" pattern. No data races (single event loop).
• CDP death race in browser_panel.py:133-139 — asyncio.wait races ws.receive_json() against cdp.reader_task; breaks cleanly if the CDP socket dies first. recv.cancel() + finally cleanup of send_task is proper.
• browser_stream.py:241-246 — new reader_task property exposes self._reader. Safe, no side effects.
• JS: live() function replaces static "live" label with live · W×H viewport dims; sendResize() re-fired 500ms after connect; reconnect backoff 2.5s → 1.2s. All clean.

Observations

• LOW: clawpatch structural review unavailable (HTTP 500 on checkout/mapping). Diff reviewed manually — no blocking issues found in the changed paths.
• LOW: cdp.reader_task depends on _reader being set before __aenter__ returns — not in the diff but consistent with the class's async-context-manager pattern. Low risk.

---

Summary: The coalescing pipeline, CDP death race, and viewport diagnostics are all well-structured. No bugs spotted in the diff. 47 host-free tests pass per PR description. Hold for CI terminal-green before clearing for merge.

— Quinn, QA Engineer

[protoquinn]

Submitted COMMENT review on protoLabsAI/agent-browser-plugin#17.