stall-fallbacklisted

## Step 1: Count Track stall-watchdog failures seen in task-notifications this session (matching `Agent stalled: no progress for Ns (stream watchdog did not recover)` or `stall watchdog fired after 60000ms with no progress`). Do nothing on the 1st or 2nd occurrence. On the 3rd, and every occurrence after that, do Steps 2 and 3 below. ## Step 2: Explain to the user Give a short, plain explanation (a few sentences, not a full report): - This is a known, reproducible Claude Code bug — not a home-network problem. The user already ruled out their network with a full 1-hour clean connectivity monitor. - It shows up most when 5+ background subagents run in parallel: background SSE streams go silently dead (zero further bytes) while the main interactive thread and its heartbeat stay healthy. This points at a connection-handling problem specific to concurrent background-agent streams (client pooling or server prioritization of interactive vs. background requests — root cause not confirmed). - It also happens with a single background agent, so concurrency makes it worse or faster, but is not the only trigger. - Filed as anthropics/claude-code#86499, referencing #54434, #25979, #53695, #55647, and anthropic-sdk-typescript#998. ## Step 3: Switch to sequential mode For the rest of the current session, stop spawning further parallel or background subagents (Agent tool or similar) for new work. Run all remaining work directly and sequentially, in-line (Read/Edit/Bas