Stop reporting handled chat-context 4xx as exceptions; upload source maps

[posthog-eu]

Problem

• The team's error tracking fills with issues that are really the backend telling a user "that conversation is already in this chat". The user flow is fine (clear toast, optimistic rollback), so this cost lands on triage, not users.
• useAddChatContextMutation.onError calls posthog.captureException(error) as its first line, then rolls back and shows a toast built from the server's detail. So expected 400s from add_chat_context (already attached, too long, context budget exceeded) file as application exceptions. The delete and clear-context handlers repeat the shape.
• No source-map upload is configured, so frames stay minified (IJ in /assets/index-*.js). The fingerprint churns on every deploy and each handled 400 reopens as a first-seen issue.

Changes

• Report only unexpected failures. A new captureUnexpectedError helper reports 5xx responses, network errors, and non-Axios errors, and skips handled 4xx AxiosErrors. It replaces the four raw captureException calls in the add, delete, and clear chat-context handlers. Toasts and optimistic rollback are unchanged.
• Upload source maps. vite build emits maps when POSTHOG_SOURCEMAPS=1. The two dashboard deploy workflows set that flag, then inject and upload maps to PostHog with @posthog/cli, then strip the maps from the deployed bundle. Frames symbolicate and fingerprints stay stable across deploys.

Error	Before	After
Handled 4xx (already attached, too long, budget)	captured	skipped, toast only
5xx response	captured	captured
Network / non-Axios error	captured	captured

Deploy setup

• Add a POSTHOG_CLI_API_KEY repository secret: a PostHog personal API key with the error_tracking:write scope. Without it the upload step is a no-op (continue-on-error: true) and never blocks a deploy.
• Source maps upload for the dashboard only. Preview and portal builds do not set POSTHOG_SOURCEMAPS, so they emit no maps.

Testing

• tsc --noEmit and biome check pass on the changed files.
• One pre-existing unrelated test failure (AgenticChatPanel "Recording limit reached") reproduces on main and is not touched by this change.
• The source-map workflow steps run only on deploy and were not executed here.

Agent context

• Considered the @posthog/rollup-plugin route (upload during vite build), but the frontend build runs through vercel build and the repo keeps secrets in GitHub Actions, not Vercel env. The CLI-in-CI route keeps the key and project id in the workflow.

---

Created with PostHog Desktop from this inbox report.