github BuilderIO/agent-native @agent-native/core@0.202.0

latest releases: v0.1.150-nightly.458, clips-nightly-v0.1.531-0, v0.1.150-nightly.457...
one month ago

Minor Changes

  • 123cf36: Add a session-authenticated Builder Gateway streaming endpoint for Desktop Code Agents.
  • 2f7e590: Agent-loop OpenTelemetry spans now follow the GenAI semantic conventions: agent.run, llm.call, and tool.call become invoke_agent, chat {model}, and execute_tool {tool} with gen_ai.* attributes. A registered meter provider also receives gen_ai.client.operation.duration, gen_ai.client.token.usage, agent_native.agent.runs, and agent_native.tool.calls. Spans and metrics carry the engine as gen_ai.provider.name; on metrics, a model the engine does not list as supported is recorded as _OTHER.
  • 6912f40: Report Core Web Vitals (TTFB, LCP, INP, CLS) once per page view as a web_vitals event keyed by React Router route template, mark them on session replays, and add the page's route template to action.response. Pages no manifest route matches send no route rather than a raw path, web_vitals never carries the page's url or path, a page loaded in a background tab sends no load, and a tab switch that measured nothing sends no page view. Session replay network events now carry pageHidden: true for requests made while the page was hidden. An action.response at or over SLOW_ACTION_RESPONSE_MS (1 s) that someone waited for (isWaitedActionResponse: the page stayed visible and the request wasn't cancelled) is also marked on the session replay as agent-native.slow_request, with its own duration, status, and outcome. Turn the capture off with configureTracking({ webVitals: false }).

Patch Changes

  • b22060c: Report why agent runs fail. agent_run_outcome now carries a cause (one of AGENT_TROUBLE_CAUSES) for a failed or interrupted run whose error code names one. It never carries the run's error message, and a code that is not an identifier, such as a sentence a route error set as its data.code, is sent as unrecognized_code (agentErrorCodeForTelemetry), so Analytics groups every other failure by its code. Stopped runs are reported unsampled under their own per-page cap, thumbs-up or thumbs-down feedback sends agent_feedback_submitted with the browser session, and every pageview carries agent_signals: 1 so Analytics counts cancelled runs, thumbs-down, and quick backs only for sessions whose client reports them. Every pageview also carries a page_load_id that stays the same until the page reloads, so Analytics can tell a return to the previous page from a page another tab opened.

    A turn the server refuses at its start, such as a chat with no model connected (AGENT_CHAT_AI_SETUP_REQUIRED, now named no_model_connected), reports its outcome under the turn id the server recorded. Framework route errors now keep a thrown error's data.code as code in the JSON body; before, the client only saw the HTTP status, so a chat with no model connected read as forbidden. A provider error whose HTTP status sits on a wrapped cause, like the Ollama provider's status_code, is now classified as http_<status>, so it reads as provider_error and gets the same retry handling as any other provider's status. agent_chat_stuck_detected now fires once per run, and only while the stuck banner shows. A session replay's console errors now carry exception: false for a plain console.error and exception: true for an exception Monitoring captured, so Analytics can tell which errors could have become issues. A session replay network event for a request the browser cancelled because the page was navigating or reloading now carries pageLeaving: true, so Analytics doesn't read it as a network failure.

  • c442bae: Record last touch beside first touch. The browser now keeps the latest visit that had a source (a ref, UTM tag or ad click id, a share link, or an outside referrer) in an an_lt cookie, and the signup event carries it as last_touch_source, last_touch_ref, last_touch_via, last_touch_utm_*, last_touch_gclid, last_touch_msclkid, last_touch_vector_source, last_touch_referrer, last_touch_site_referrer, last_touch_path, last_touch_site_path, and last_touch_at. As with first touch, last_touch_path is where the visitor entered the app, and last_touch_site_path is the marketing-site page a touch the site forwarded landed on. last_touch_truncated: "true" marks a last touch the browser had to trim to fit its cookie, or one dropped because both touches would overflow the signup handoff header, so first touch always survives. Magic-link signups carry last touch even when the browser has no first touch.

    First touch is still first-write-wins, except that a visit with no source no longer blocks the first one that has a source. Referrers from *.agent-native.com, local dev servers, and Google sign-in don't count as a source, for capture and for referral_source alike. getLastTouchAttribution() is exported from @agent-native/core/client/analytics. The marketing site forwards its own last touch to the apps as last_* params, including last_landing_path, with last_at, the time of that visit, and an older site visit no longer replaces a newer one the app already recorded.

  • 053539c: The client replays a 401 or 403 for a minute only inside an embed, where it stops an expired embed token from setting off a retry storm. Outside an embed every read reaches the server, so a page shared with someone mid-session loads all of its reads as soon as access arrives instead of failing one of them with the earlier refusal. When a link's status turns allowed, useResourceAccessGate also forgets replayed refusals, so a tab that still holds an embed token after leaving the embed reads the page again instead of replaying the old refusal.

  • 6c0c4ef: Hold sign-up until the org tables exist on a fresh database. Sign-up checks org sign-in policy inside its transaction, and a missing table failed it as "Enter a valid email address".

  • 122d27a: Let a suggestion's author withdraw their own pending suggestion with comment access through decide-resource-suggestion (decision: "withdrawn"). A withdrawn suggestion gets the new withdrawn status, so it stays distinct from a reviewer's rejection. Accepting and rejecting still require edit access, and adapters' coordinateDecision now runs only for acceptance. Withdrawal skips the stale-base check, since withdrawing never applies the suggestion to the resource.

  • 123cf36: Keep Builder setup CTAs on the shared account chooser with one-click activation.

  • af9b97b: Fix chats with your own OpenAI or Anthropic key failing with 401 Unauthorized on hosts that inject a provider gateway URL (Netlify AI Gateway sets OPENAI_BASE_URL and ANTHROPIC_BASE_URL at runtime). A user, org, or workspace key now goes to the endpoint saved with it, or to the provider's official API; a deployment OPENAI_BASE_URL or ANTHROPIC_BASE_URL applies only to the deployment's own key.

  • f649ae6: Agent-readable links are shorter. Scoped agent-access tokens (signScopedAgentAccessToken, createScopedAgentAccessGrant) now use a compact format that no longer embeds the resource id in the payload; the signature covers it instead. A Slides deck link drops from about 290 to about 205 characters and a password-protected Clips link from about 360 to about 185, so both fit under the 250-character URL limit of Anthropic's web fetch tool. viewerEmail, agentLabel, and the expiry are still signed into the token. verifyScopedAgentAccessToken accepts both formats, so links minted before the upgrade keep working until they expire. Failures for a token minted for a different resource now report bad_signature instead of wrong_resource. signShortLivedToken and verifyShortLivedToken are unchanged.

  • 1ba0667: Add a server-plugin-default-export doctor guard that flags server/plugins/ files with no default export, so agent-native build stops with a clear fix instead of failing later in the bundler with [MISSING_EXPORT].

  • dfff955: Drop settled mutations from the query cache as soon as no hook observes them, so they no longer keep a render's data alive for five minutes.

  • 84e173d: Provider readiness preflights start a fresh status request instead of reusing an older in-flight probe.

  • a9879f8: Status checks that get refreshed mid-request now return the fresh answer instead of reporting the server as unreachable, so chat no longer gets stuck on "Couldn't confirm AI is ready" with prompts waiting to send. Retry on that message now clears the stuck state.

  • b22060c: useLabState now returns refetch, so a page can retry a Lab state that failed to load.

  • Release all public npm packages with a patch version bump.

  • b52ed3a: Bind cross-app MCP tokens to their endpoint, preserve verified user identity before organization fallback, and deduplicate organization-principal A2A submissions.

  • 1313fe1: Allow directory MCP profiles to omit MCP App widgets.

  • edc7f35: Shareable resources can take access requests. Set accessRequests: true on registerShareableResource, and a signed-in viewer who gets denied from get-resource-access-status sees canRequest and can call request-resource-access with an optional note. The owner and people shared directly as Admin, while their access still resolves, get an inbox notification and the core.access-requested email, which link to /access-requests/<id>; approve-resource-access-request grants through the same organization and recipient rules as share-resource, never lowers a stronger role, even one granted mid-approval, and emails the requester core.access-granted, returning email as sent, skipped (no email set up, or a test identity), or failed, while decline-resource-access-request records the decision without telling the requester, who can ask again after seven days. Requests are capped per requester and per owner each day, and asking again while a request is open notifies no one twice; a request whose send was cut off before anyone was told can be asked again, which sends it again under the same idempotency keys. useResourceAccessGate now offers requestAccess and checks the status every half minute while a request is open; useAccessRequestReview and useResourceAccessRequests read requests for review, the list returning the newest 50 with hasMore. Toolkit's ResourceAccessScreen takes a request prop for Request access with a note, AccessRequestApprovalPage is the review page an app mounts at /access-requests/:requestId, and the share panels of ShareButton and ShareDialog list pending requests with Allow and Decline for people who manage access.

  • 053539c: Add get-resource-access-status, which tells an app what a link the viewer can't open should say: denied (it exists, but not for them), missing, trashed (only to people who could open it), signed-out, or allowed. It never returns the resource's title, owner, visibility, or workspace, and signed-out callers learn nothing about existence. Shareable resources can declare an availability rule, such as not being in the trash, which joins the lightweight access projection. Apps read the status with useResourceAccessGate from @agent-native/core/client/sharing and render Toolkit's ResourceAccessScreen. A registration's canManageAccess hook now always receives the whole row, including during lightweight access checks. A registration can also declare a fallbackAccessContext: { columns, resolve }, such as the authority a space lends its members, which the status uses when the viewer's own context can't open the resource. Its columns join the lightweight projection, so resolve gets the row already loaded; resolveAccess and the share actions don't read it.

  • 3347175: Allow ChatGPT directory preflights from the documented OpenAI origins.

  • 4522907: Keep a session in one replay recording when the user reloads or navigates while an upload is in flight. Previously the next page resent the last chunk's number, the server rejected it with HTTP 409, and the recorder restarted under a new recording, splitting the session in two. An upload too large to outlive the page is no longer started as the page hides or closes; its events stay queued for the next upload of a page that survives.

  • dfff955: Session replay finds an iframe's closing head tag without copying the whole document, so large previews mount faster.

  • 7109d2c: Compact agent-access tokens keep session-replay links within the API URL limit.

  • c8fa837: Render hosted sign-in pages on the server, blur the signup form panel backdrop, and use the shared Calendar wave across signup pages and the homepage hero.

  • fa322d4: Route explicit local chat handoffs through the ChatGPT MCP App host.

  • a88e431: Recover exactly the automation runs a stopped worker left behind: settle every unfinished run past its liveness window instead of only the newest, leave a queued "Run now" dispatch alone while its claim lease is still valid, and finish a run only while the stored claim is unchanged from the one that was read. A queued run is no longer cancelled or overwritten before its worker starts, the stale run it hid behind is no longer left open, and a failed history write no longer stops the remaining runs from being settled.

  • 8283661: Accepting a suggestion that can no longer be placed now fails with a 409 suggestion_stale error and leaves the suggestion pending, instead of returning success while marking it stale. A decision sent with an outdated observedBase now fails with suggestion_conflict without changing the suggestion, and a stale proposal member reports suggestion_stale too.

  • e74a0f5: Correct widget metadata and resource handling for ChatGPT directory MCP profiles.

  • Updated dependencies

    • @agent-native/agentkit@0.202.0
    • @agent-native/recap-cli@0.5.63

Don't miss a new agent-native release

NewReleases is sending notifications on new releases.