BuilderIO/agent-native @agent-native/core@0.134.0

on GitHub

Minor Changes

• 46cd162: Make PostHog a first-class error-reporting and LLM-observability backend, and fix
  the malformed exception events it was already receiving.

  captureException() emitted an event named $exception carrying camelCase
  properties. PostHog ingests anything by that name and renders it as an issue, but
  it groups and symbolicates from $exception_list — so every PostHog-configured
  app was already collecting exceptions that arrived empty and ungroupable, which
  reads as coverage rather than as a failure. The PostHog provider now reshapes
  those into a real $exception_list with parsed stack frames.

  Route errors no longer depend on Sentry. The Nitro error hook lived inside
  sentry-plugin.ts, which returns early when no SENTRY_DSN is set, so an app
  running PostHog alone reported no route errors at all. The hook moved to
  core-routes-plugin.ts and goes through the provider-agnostic captureError()
  registry, so every configured backend receives it. The ~150 lines of
  production-tuned drop rules (expected 4xx, permission rejections, Lambda
  freeze/thaw socket hang up) moved out of Sentry's beforeSend into
  server/error-noise-filter.ts and now apply to every backend — without them a
  second backend receives a firehose. Server exceptions are also attributed to the
  in-flight user instead of landing under anonymous.

  Browser exceptions go to PostHog when POSTHOG_PUBLIC_KEY / VITE_POSTHOG_KEY
  is set, posted directly rather than relayed through /_agent-native/track, which
  requires a session and would drop every signed-out crash. POSTHOG_API_KEY is
  deliberately not a fallback for the public key: that value is inlined into the
  public HTML shell. Note that PostHog does not symbolicate without uploaded source
  maps, so minified browser stacks stay minified.

  LLM observability now emits the full PostHog trace tree. Previously a run
  produced a single $ai_generation labelled agent_run whose $ai_parent_id
  pointed at a span that was never sent, so PostHog wrapped it in a placeholder
  trace with no steps. Runs now emit $ai_trace, one $ai_span per tool call, and
  a generation parented to the trace. Tool calls ship inside $ai_output_choices
  even with content capture off, because that is the only thing PostHog derives
  $ai_tools_called from. The previously dead capturePrompts flag is now wired
  and gates $ai_input and assistant text; disabled fields are omitted rather than
  sent empty, and oversized content is replaced with an explicit truncation marker
  instead of being silently shortened. $ai_error became a structured object with
  the terminal code and retryability, and errors captured during a run carry the
  run's $ai_trace_id so an issue and its trace resolve to each other.

  Feedback previously emitted only for thumbs; category and free-text submissions
  emitted nothing. All four now report, with sentiment still limited to thumbs so
  a category follow-up does not double-count the vote. PostHog surfaces feedback in
  LLM analytics only through a survey sent event, so that is emitted too when
  POSTHOG_AI_FEEDBACK_SURVEY_ID is configured — and not at all when it is unset,
  rather than inventing a survey id.

  Agent traces carry the browser session as $session_id (read from a new
  X-Agent-Native-Session-Id header) so a trace joins its session replay, distinct
  from $ai_session_id, which remains the conversation thread.