github BuilderIO/agent-native @agent-native/core@0.160.0

latest releases: v0.1.150-nightly.458, clips-nightly-v0.1.531-0, v0.1.150-nightly.457...
3 months ago

Minor Changes

  • 167be56: Let a hosted app pay for its own AI with Builder credits. BUILDER_GATEWAY_TOKEN
    plus BUILDER_GATEWAY_SPACE_ID now select the Builder engine and back the
    gateway lane (chat, web search, realtime voice, transcription, scheduled and
    event automations), so an anonymous visitor can use AI on a deployed site
    without connecting anything. An injected gateway token can never move a
    customer's spend onto Builder credits: it steps aside for any other engine whose
    credentials resolve. A Builder key pair the customer configured themselves is
    unchanged and keeps winning, as does AGENT_ENGINE_PREFER_BYO_KEY. Where the
    deployment pays, a rejected or missing credential reads as one line to the
    visitor, with the real reason kept on the error code for the owner; that holds
    for the gateway's own 402/403 on voice and realtime transcription, for the auto
    provider chain rather than only an explicit Builder preference, for a gateway
    whose transport dropped or whose stream stopped early, and at the point the chat
    renders an error — where a message the deployment already chose for a visitor is
    no longer re-expanded from its code back into owner instructions. Owner surfaces
    keep the copy that says what to fix, the workspace/preview runtime included, even
    though it carries the same injected token as the published site. No recovery
    decision depends on what the error message says any more: the Builder engine
    marks a retryable gateway rejection and an over-long prompt structurally, and
    those verdicts reach the chat client too. So an overloaded provider retries the
    same way whoever is paying — without turning a provider throttle into a chain of
    background continuations against the limit that just rejected it — a truncated
    stream is still continued from where it stopped instead of ending the turn, and a
    conversation that outgrew the context window still gets its one automatic trim
    and retry. A background or
    scheduled run keeps the message the server chose for its failure rather than
    restating it from the terminal reason, and an earlier transient error in the same
    run can no longer stand in for the reason the run actually died. Pasted
    ANTHROPIC_API_KEY values are now validated when saved.

Patch Changes

  • ed0666b: Allow token-authenticated remote device relay endpoints to reach their route-level device-token verifier.
  • ed0666b: Distinguish a retryable app load failure from an app that is genuinely gone. The chat-first app pane's error branch fell back to appUnavailable, rendering "This workspace app is no longer available." above a Retry button.
  • ed0666b: Bind workspace embed-session adoption to the existing target identity while allowing app-local organization ids.
  • ed0666b: Prevent email security scanners from consuming Electron magic-link sign-ins before the user confirms them.
  • ed0666b: Keep Electron magic-link verification behind an explicit POST confirmation so link scanners cannot consume the sign-in token.
  • ed0666b: Verify confirmed desktop magic links in the confirmation request so the one-time session cookie reaches the native callback reliably.
  • ed0666b: Add no-secret diagnostics around desktop magic-link issuance and verification so invalid-token failures can be isolated without logging the token.
  • ed0666b: Format the Portal reference table in the shared core documentation.
  • ed0666b: Redact nested callback parameters in desktop magic-link diagnostics.
  • ed0666b: Share the canonical localized authentication copy with native sign-in surfaces
    and allow authenticated packaged callers to mint workspace embed sessions.
  • ed0666b: Add redacted diagnostics for desktop magic-link session-cookie handoff failures.

Don't miss a new agent-native release

NewReleases is sending notifications on new releases.