fix(chat-widget): retry health check so the widget doesn't vanish on a blip

[web3jenks]

Why

Users report the AI chat widget sometimes doesn't load at all (needs a page refresh). Root cause: both entry points (the floating bubble and the header "Ask Babylon AI" button) are gated on a single, no-retry /health fetch that defaults to hidden and fails closed:

• isApiHealthy starts false; one fetch('/health') with a 5s abort runs on mount.
• On any failure/timeout, isApiHealthy stays false and if (!isApiHealthy) return null renders nothing (and the header button's reveal-CSS is never injected).
• No retry — so a cold pod, a 503 during warmup, a rolling deploy of the single replica, or any transient blip hides the widget until a full reload.

What

• Retry on a front-loaded schedule: attempts at 0, +2s, +7s, +20s, +50s, +110s (6 checks within ~110s, under a 2-minute span), dense early and spreading out, then a slow 60s poll while still unhealthy. Stops the instant /health responds OK.
• Re-check on tab focus (visibilitychange) so a pod that recovers while the tab was backgrounded shows the widget without a reload.
• Per-attempt timeout tightened 5s → 4s; pending retries + listener cancelled on unmount.

Frontend-only; ships independently of the backend keepalive work (devex-ai-bot#63, actions#60), which reduces how often these retries even fire.

Notes

• Commit is unsigned (authored headlessly) — resign with YubiKey before merge if required.
• No API surface change; verified no dangling refs. Recommend letting CI run the Docusaurus typecheck/build.

🤖 Generated with Claude Code

[github-actions]

🔐 Commit Signature Verification

✅ All 1 commit(s) passed verification

Commit	Author	Signature	Key Type	Key Check
9528107e7b0d	Jenks	✅	sk-ssh-ed25519	✅

Summary

• Commits verified: 1
• Signature check: ✅ All passed
• Key type enforcement: ✅ All sk-ssh-ed25519

---

Required key type: sk-ssh-ed25519 (FIDO2 hardware key)

Last verified: 2026-07-09 10:35 UTC

[greptile-apps]

Greptile Summary

This PR makes the chat widget retry its health check instead of hiding after one failed request. The main changes are:

• Front-loaded retry delays followed by steady polling while unhealthy.
• A shorter per-attempt health-check timeout.
• A visibility-change recheck when the tab comes back into focus.
• Cleanup for pending retry timers and the visibility listener.

Confidence Score: 5/5

The changed flow looks safe to merge after a small cleanup to avoid duplicate polling.

• The retry loop fixes the main hidden-widget path.
• A visibility-triggered retry can leave parallel health-check chains running.
• No blocking correctness or security issue was found in the changed code.

src/components/ChatWidget.tsx

Important Files Changed

Filename	Overview
src/components/ChatWidget.tsx	Adds retrying health checks, visibility-triggered rechecks, and timer/listener cleanup for the chat widget.

_{Reviews (1): Last reviewed commit: "fix(chat-widget): retry health check so ..." | Re-trigger Greptile}

[web3jenks]

✅ Retry logic unit-tested (faithful port, delays ÷100 to run fast)

A) recovers after 3 failures:
   attempt #0 fail, #1 fail, #2 fail, #3 HEALTHY
   -> visible=true, totalCalls=4   (stops retrying on success)

B) API stays down (old code = 1 try, hidden forever):
   attempts at 0/2/7/20/50/110s (shown ÷100) -> 6 calls in-window
   -> visible=false, still retrying   (old behaviour: calls=1, hidden until reload)

Confirms: retries on the front-loaded schedule (0/2/7/20/50/110s, then 60s poll), stops immediately when /health recovers, and keeps the widget hidden-but-recovering while the API is down instead of hidden-forever. The visibilitychange re-check additionally un-hides on tab focus without a reload.

This is frontend-only; the backend keepalive (devex-ai-bot#63, verified live: cold query 12.4s → 7.0s warmed) reduces how often these retries even fire.

@kkkk666 could you review? 🙏

[web3jenks]

Addressed Greptile review (commit 9528107, signed):

• Parallel retry chains → added a generation counter. Each (re)start bumps generation; an in-flight/scheduled attempt whose gen is stale bails out after its fetch resolves, so a tab-focus re-check can't leave an orphaned polling chain running alongside the current one.

Re-signed with YubiKey. tsc passes with no ChatWidget errors.

[kkkk666]

test pass.