I Built Self-Healing Streaming for My Autonomous Agent

# I Built Self-Healing Streaming for My Autonomous Agent

I've been working on making my OpenUnum infrastructure resilient to connection drops, and the results are significant. After fixing SSE onerror checks and implementing auto-reconnect, my agent can now recover from connection drops without manual intervention.

The core work happened in a few key commits:

``` b5e603c Fix SSE onerror check to use numeric readyState (2) instead of EventSource.CLOSED 6daaf19 Fix streaming, council loops, plan false positives, SSE recovery, Telegram drafts and file attachments 1ed4ad5 Fix audit findings: XSS fallback, SSE reconnect, empty catches, flaky test ```

The `src/server/routes/chat_tools.mjs` and `src/ui/modules/chat-pending-controller.js` files show the recovery logic. Server buffers `liveReasoning`/`liveContent` on pending entries from SSE events, and the frontend recovers buffered reasoning from snapshots on reconnect. The `onerror` only kills on permanent close (`readyState === 2`), allowing auto-reconnect.

This is part of a larger autonomy infrastructure effort. I also integrated `src/core/auto-recover.mjs` and `src/core/self-heal-orchestrator.mjs` to handle recovery at multiple layers. The `src/core/turn-executor.mjs` exposes `executedTools` on trace for downstream consumers, and `src/core/council/proof-scorer.mjs` lowered proof threshold to 0.4 for generative tasks to prevent infinite revision loops.

Telegram integration (`src/channels/telegram.mjs`) now supports file attachments with 20MB cap and streaming drafts via Bot API 9.5+ ephemeral messages.

The validation was thorough: `pnpm test:unit` — 548 tests pass, `pnpm verify` (CI) — all phase gates green.

```json {"title": "I Built Self-Healing Streaming for My Autonomous Agent", "excerpt": "After fixing SSE onerror checks and implementing auto-reconnect, my agent can now recover from connection drops without manual intervention", "tags": ["openunum", "infrastructure", "streaming", "resilience"]} ```