Minor Changes
- 53f0c01: Default direct Anthropic and OpenRouter connections to Claude Sonnet 5.5, align the Builder model picker with the current gateway catalog, and refresh model labels.
- 47c52cf: Add
registerObservabilityProvider()to@agent-native/core/serverso a host can hand core its OpenTelemetry tracer and meter providers. Core now records the unsampledhttp.server.request.durationhistogram, measured at the response boundary, and theagent_native.telemetry.flush_failurescounter, and force-flushes registered providers on the response hook with a 2-second cap so serverless exports are not lost; each provider's flush is awaited and counted on its own.isServerlessRuntime()is now exported from@agent-native/core/server. Scaffolded apps no longer carry a first-party template's workspace-only OpenTelemetry wiring. - ba09465: Pass app client compatibility versions through to action routes and support an explicit cache-busted retry after a stale-client reload is canceled.
Patch Changes
- 1f6a79d: An app action that shares a name with a code-execution tool, such as
list-data-programs, now keeps that name. Previously the framework's version replaced it on every agent and MCP surface, which dropped the app'smcpToolexposure. - 01131c3: Keep first-turn chat persistence reliable, record typed run failures, and expose chat readiness and signup attribution in analytics.
- 563e22a: Creating a Builder.io account from "Create and activate" or "Create Builder.io account" is now one
POST /_agent-native/builder/provisionrequest with no popup window, so popup blockers no longer stop it; only connecting an existing account opens Builder's sign-in window, and composer runtimes without the consent popover never create an account. - 4f52c40: Builder.io asset uploads (signed URL, small-file, and upload-complete requests) now refresh the OAuth access token and retry once when refused with 401, instead of failing on a token that was revoked or rotated before its stated expiry. The retry only happens when the refresh produced a different token.
resolveOAuthCredentialAccessgains aforceRefreshoption that refreshes even a token with no stated expiry, andresolveBuilderApiAuthorization,resolveBuilderRequestAuthorization,getBuilderOAuthSession, andgetMcpOAuthAccessTokenpass it through. - 10aa0e1: Add a core search index that apps opt into with
registerSearchableResource, fed by a general resource change feed that captures every write with database triggers. Search answers from the index only when it is current and never polls, so it never wakes a sleeping database. A term too long for the index throwsSearchTermTooLongError, which apps answer with their previous search. The query parser moves to@agent-native/core/search-query. - 5c41297: Preserve verified request identity when recording completed agent runs.
- 891062d: Label
agent_native.telemetry.flush_failureswithagent_native.telemetry.signal(metricsortraces) so a timed-out or failed flush can be attributed to the export that caused it. - da846cf: Keep the JWKS endpoint publishing the live signing key once an app's
jwkstable holds more than 100 rows, and stop minting a new key on every signature in that state. - 1323e3a: Store the default template's
.env.exampleand_gitignore, and the package's.npmignore, with LF line endings like the rest of the repo. No content change. - 24f73c2: Keep Vite hot updates authenticated when a live Design frame has an opaque origin.
- Release all public npm packages with a patch version bump.
- 1f6a79d: Let apps limit
listVisibleMcpToolsandcallMcpToolto servers whose URL belongs to a named provider, so a server that reuses another provider's tool names cannot receive its calls. - 5113a23: Make follow-up queueing reliable, preserve each prompt's run options, and keep internal context out of user-visible text.
- 9a09590: Keep useful agent activity labels visible while a run is active.
- 8aae00f: Keep CLI commands running when the platform browser opener is unavailable.
- 2254122: Scope workspace navigation per app and preserve agent status lookup failures.
- 73c2373: A prompt refused for missing AI setup now keeps what its retry needs (references, model, engine, effort, request mode) and a refusal marker in the thread, so the setup card finds it after a reload; the server lets only one tab send the after-setup resend of a refused run. The composer's
onBeforeSubmitreceives the draft it is holding and the handle gainsgetDraftSnapshot(), so a host resumes only a draft that was not edited while connecting. An unreadable provider 403 during a key check is retryable instead of a rejected key, and a misconfigured Builder host is no longer reported as a credit-service outage. - 73c2373: A refused prompt's refusal marker and retry context now survive a client thread save and reload, and the references stored with it are validated against the composer's bounded reference shape. The after-setup resend claim is released when the thread's run slot is busy and swept with a compare-and-delete, provider 408 and 425 answers during a key check are retryable, and the composer's draft snapshot tells apart a replaced attachment that has the same name and compares every field of each reference, so a held-back draft is resumed only when nothing it would submit changed.
- 73c2373: Answer chat turns the server refuses before a run starts (AI setup missing, no usable model credential) in the thread itself: the prompt and a typed failed run are persisted server-side, the chat shows the connect card with a retry, and the refused prompt is sent again once after Builder or a provider key is connected. Run lifecycle analytics now carry the canonical user id from the request context and count refused turns as
run_no_reply. - da846cf: Complete a chat reply that continued across several runs when the page was reloaded before the last continuation finished, instead of showing only the first half.
- 73c2373:
onAgentRunCompletenow receives{ turnContinues }so an observer can tell a finished turn from a run that handed off to a continuation run. The retry marker on a recovery message now survives a reload, so a refused prompt is sent again only once across cards, tabs, and reloads, and only AI-setup refusal cards are hidden once a later run starts. - a2d1e33: Redact SQL bind parameters from error telemetry and server logs.
- 393d119: Mark tracked browser analytics events on session replay timelines. Each marker records only the event name; telemetry events such as page views and action responses stay unmarked.
- 4bee69d: Explain oversized agent chat requests and improve actionable failure recovery feedback.
- 29a6eef: Restore chat runs and return typed run-slot conflicts after a thread reload.
- 73c2373: Report Builder credit-service outages and unverifiable provider key checks as typed, retryable failures instead of generic 500s and "rejected key" 400s, and accept OpenAI project keys restricted from listing models.
- 73c2373: Persist first-touch utm_source, utm_medium, utm_campaign, utm_term, gclid, msclkid, vector_source, and referring host on the Better Auth user row at signup (additive nullable columns), for every signup path that carries browser attribution.
- Updated dependencies
- Updated dependencies [5113a23]
- @agent-native/agentkit@0.200.0
- @agent-native/recap-cli@0.5.60