fix(decopilot): keep chat streams alive and recoverable across disconnects

[viktormarinho]

What is this contribution about?

Fixes the frequent "network error" banner during chat streaming and the related "thread stuck in_progress forever" symptom on self-hosted deploys, especially during deep research. Three root causes were addressed:

• The AI SDK's JsonToSseTransformStream emits no SSE keepalive bytes during silent gaps (long tool calls, deep-research polling), so reverse proxies (nginx ingress, Istio, ALB — all 15–60s read timeouts) cut the connection. Added a defensive wrapWithSseKeepalive that injects : keepalive\n\n comment lines every 15s, only on event boundaries, with full cleanup on close/error/cancel.
• claimOrphanedRun excluded same-pod ownership, so on single-pod self-hosted deploys a run that left the in-memory registry while run_owner_pod still pointed at the live pod was permanently unrecoverable. Removed the exclusion (callers already verify runRegistry.isRunning); same fix applied to listOrphanedRuns for K8s StatefulSet rolling-restart recovery.
• MAX_RUN_AGE_MS of 30 minutes prematurely reaped legitimate deep-research runs while the provider job kept burning credits — bumped to 90 minutes.
• Plus client-side: chatError is now cleared when silent resume succeeds and at the start of a new send, so transient blips don't leave a sticky "network error" toast.

How to Test

1. With bun run dev, start a long deep-research chat and let it run. Confirm the response continues to flow (heartbeats are invisible to the parser; verify via DevTools → Network → response stream payload includes : keepalive lines).
2. To simulate a proxy disconnect: kill the dev server mid-stream, restart, revisit the thread — the orphan resume path should attach successfully via /decopilot/attach/:threadId instead of leaving the thread in_progress.
3. To verify same-pod recovery: with POD_NAME set, manually mark a thread status=in_progress, run_owner_pod=<current pod> while the registry is empty (e.g. between requests), revisit the thread — orphan resume should now claim it.
4. bun test apps/mesh/src — 1506 passing.

Migration Notes

None. No schema or config changes.

Review Checklist

• PR title is clear and descriptive
• Changes are tested and working (1506 mesh tests pass; new unit tests cover the keepalive boundary, cancel propagation, leak prevention; new test covers same-pod claim)
• Documentation is updated (if needed) — N/A
• No breaking changes

---

Summary by cubic

Keeps chat SSE streams alive and recoverable across proxy disconnects, fixing the “network error” banner and threads stuck in_progress on self‑hosted. Adds 15s SSE keepalives, enables same‑pod orphan recovery, and clears stale error toasts.

• Bug Fixes
  • Injects : keepalive comments every 15s on SSE responses (only at event boundaries) to prevent reverse‑proxy read timeouts.
  • Allows same‑pod claims in orphan recovery and includes same‑pod runs in startup listing to handle StatefulSet restarts and single‑pod deploys.
  • Clears chat error on successful silent resume and when starting a new message to avoid sticky “network error” toasts.

^{Written for commit 6721ccf. Summary will update on new commits.}

[github-actions]

🧪 Benchmark

Should we run the Virtual MCP strategy benchmark for this PR?

React with 👍 to run the benchmark.

Reaction	Action
👍	Run quick benchmark (10 & 128 tools)

Benchmark will run on the next push after you react.

[github-actions]

Release Options

Suggested: Patch (2.310.12) — based on fix: prefix

React with an emoji to override the release type:

Reaction	Type	Next Version
👍	Prerelease	2.310.12-alpha.1
🎉	Patch	2.310.12
❤️	Minor	2.311.0
🚀	Major	3.0.0

Current version: 2.310.11

Note: If multiple reactions exist, the smallest bump wins. If no reactions, the suggested bump is used (default: patch).