fix(web): prevent browser crash from timer leaks, DOM growth, SSE buffer

[henrypark133]

Summary

Fixes #2406 — the web UI became unresponsive and triggered Chrome's "Pages Unresponsive" dialog during extended bug bash sessions with heavy bot interactions and parallel chats.

Root cause: The browser accumulated resources that were never cleaned up — leaked setInterval timers across SSE reconnects, unbounded DOM node growth, unbounded in-memory job event storage, and a too-small SSE broadcast buffer causing reconnect cascades.

Changes:

• Timer cleanup (cleanupConnectionState()): Clears _streamDebounceTimer, _streamBuffer, _connectionLostTimer, and jobListRefreshTimer on SSE reconnect, tab-hide, and page unload. Intentionally excludes _doneWithoutResponseTimer which is a turn-level concern (fix: web UI messages stuck until refresh — SSE event ordering bug #2079).
• DOM pruning (pruneOldMessages()): Caps #chat-messages at 200 elements. Streaming-aware — skips data-streaming="true" elements. Called at turn boundaries (after response event) and after loadHistory(), not per-addMessage() to avoid O(N²) during bulk loads.
• Job events cap: Limits jobEvents Map to 50 entries with LRU eviction by last event timestamp. Excludes the current job from the eviction scan to prevent self-eviction.
• SSE broadcast buffer: Increased from 256 to 1024 events (configurable via SSE_BROADCAST_BUFFER env var). Includes .filter(|&n| n > 0) guard to prevent broadcast::channel(0) panic from misconfigured env.

Test plan

• cargo fmt — clean
• cargo clippy --all --all-features — zero warnings
• cargo test --lib — passes (pre-existing file_history SIGABRT unrelated)
• 3 new Playwright E2E tests pass (test_dom_resource_limits.py):
  • test_dom_pruned_after_many_messages — injects 250 messages, asserts DOM <= 200
  • test_no_timer_leak_across_reconnects — 5 reconnect cycles, asserts no interval leak
  • test_prune_preserves_streaming_message — streaming element survives pruning
• Manual browser testing on staging: extended session with heavy tool use, tab switching, parallel chats

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Addresses long-running web UI instability (Chrome “Pages Unresponsive”) by bounding frontend resource growth (timers, DOM nodes, in-memory job events) and reducing SSE reconnect cascades via a larger broadcast buffer.

Changes:

• Add connection-level cleanup for SSE reconnect/tab-hide/unload and prune the chat DOM to a fixed max.
• Cap tracked job event history to bound in-memory growth.
• Increase backend SSE broadcast buffer (env-configurable) and add Playwright E2E coverage for resource limits.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
crates/ironclaw_gateway/static/app.js	Adds connection cleanup, DOM pruning, and jobEvents map eviction logic to prevent long-session resource leaks.
src/channels/web/sse.rs	Makes SSE broadcast buffer size larger and configurable via SSE_BROADCAST_BUFFER.
tests/e2e/scenarios/test_dom_resource_limits.py	Adds E2E tests targeting DOM pruning and timer leak regression scenarios.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[gemini-code-assist]

Code Review

This pull request implements several resource management improvements to prevent memory leaks and DOM bloat in the web UI, alongside server-side SSE buffer optimizations. Key additions include a connection state cleanup routine, a mechanism to prune old chat messages from the DOM while preserving active streams, and a cap on tracked job events. Feedback suggests that the current placement of the DOM pruning logic may interfere with message pagination and that the job eviction logic should be updated to protect the job currently being viewed by the user.

[henrypark133]

Review: Prevent browser crash from timer leaks, DOM growth, SSE buffer (Risk: Medium)

Good defensive fixes for a real browser crash — timer cleanup, DOM pruning, job events LRU eviction, and SSE buffer increase. A few items to address.

Concerning: SSE_BROADCAST_BUFFER env var read in wrong module

File: src/channels/web/sse.rs
CLAUDE.md convention: all gateway env vars flow through src/config/channels.rs via GatewayConfig. This PR reads std::env::var("SSE_BROADCAST_BUFFER") directly inside sse.rs, bypassing config resolution, .env.example documentation, and structured config dumps. Move to GatewayConfig as broadcast_buffer: usize, resolved via parse_optional_env, then passed to SseManager::with_max_connections_and_buffer().

Concerning: Verify MAX_BROADCAST_BUFFER clamp exists

File: src/channels/web/sse.rs
Patch 2 claims to add const MAX_BROADCAST_BUFFER: usize = 65_536 and .min(MAX_BROADCAST_BUFFER). Without this clamp, SSE_BROADCAST_BUFFER=18446744073709551615 causes OOM on startup. Verify the clamp is present in the final commit — one reviewer found it missing from the worktree.

Concerning: Verify Patch 2 fixes landed

File: crates/ironclaw_gateway/static/app.js
Patch 2 claims to add _connectionLostAt = null to cleanupConnectionState() and protect currentJobId in LRU eviction (if (k === jobId || k === currentJobId) continue). One reviewer found discrepancies between the diff and the actual file state. Confirm both fixes are present in the final commit.

Concerning: pruneOldMessages during pagination breaks scroll

File: crates/ironclaw_gateway/static/app.js
The pagination path saves/restores scrollHeight, then pruneOldMessages() invalidates the restore. Patch 2 adds if (!isPaginating) pruneOldMessages() — confirm this guard is applied.

Minor: Timer leak test methodology

File: tests/e2e/scenarios/test_dom_resource_limits.py
The setInterval monkey-patch is installed after initApp() has already run, so timers created during initialization (e.g., gatewayStatusInterval from startGatewayStatusPolling()) are invisible to the counter. Consider using page.add_init_script() to install the patch before app JS runs.

Minor: pruneOldMessages counts heterogeneous elements

The selector .message, .activity-group, .time-separator counts all three toward the 200-item cap. A conversation heavy on activity-groups could cause aggressive message pruning. Consider counting .message nodes only for the budget.

Minor: No Rust-side tests for SSE buffer config

No unit test for the buffer size configuration, the clamp, or slow-consumer behavior. Consider a test asserting the env var is parsed and clamped correctly.

Convention notes:

• E2E test file correctly placed in tests/e2e/scenarios/
• Should be added to tests/e2e/CLAUDE.md scenario table
• debug! logging used correctly throughout

🤖 Generated with Claude Code

[henrypark133]

Review follow-up (faa0f60)

All concerns from the review have been addressed:

#	Concern	Status	Commit
1	SSE_BROADCAST_BUFFER env var read in wrong module	Fixed — moved to GatewayConfig in src/config/channels.rs via parse_optional_env	faa0f60
2	MAX_BROADCAST_BUFFER clamp missing	Fixed — added MAX_BROADCAST_BUFFER = 65_536 with .min() clamp and zero-rejection	faa0f60
3	Timer leak test monkey-patches after initApp()	Fixed — now uses page.add_init_script() before navigation	faa0f60
4	Missing from E2E CLAUDE.md scenario table	Fixed — added entry	faa0f60
5	No Rust-side tests for SSE buffer config	Fixed — added broadcast_buffer_defaults_and_clamps test	faa0f60
6	_connectionLostAt = null missing in cleanupConnectionState	Already fixed	236948a
7	currentJobId not protected in LRU eviction	Already fixed	236948a
8	pruneOldMessages during pagination breaks scroll	Already fixed	236948a
9	pruneOldMessages counts heterogeneous elements	Accepted as-is — simpler, edge case risk is low	—