Minor Changes
- 0b57293: Add an internal beta/production environment badge and typed deployment-lane metadata to hosted Agent-Native app shells.
Patch Changes
-
0b57293: Fix
useActionQuery/useActionMutation/callActionsurfacing an opaque405when a caller's HTTP verb doesn't match an action's declaredhttp.method(e.g. adefineAction({ http: { method: "DELETE" } })called without{ method: "DELETE" }). The transport now throws a typedaction_method_mismatcherror naming the action, the method that was sent, and the method it declares, instead of a bare "Method not allowed" the caller had to reverse-engineer — and marks it non-retryable, since resending the same wrong verb never succeeds. -
0b57293: Show Approve/Deny again when a tool approval has to be re-asked.
approval_requirednow carries anaskIdidentifying that specific gate hit, and the chat retains a user's resolution per ask instead of per approval key. Previously, if a resume never consumed the grant (expired TTL, turn-id mismatch), the server re-asked for the same call and the client still showed the quiet "Approved" note — the buttons never came back, so the action silently never ran and there was no way to retry. -
0b57293: Fix a bug where closing one chat tab could close several tabs at once (or make a tab reappear right after closing it). A duplicated thread id made two tab-bar entries share one underlying thread, so closing either removed both. Open-tab ids are now de-duplicated in the tab state itself, which covers both a corrupted list restored from localStorage and duplicates introduced at runtime by a synchronous burst of open requests.
-
f97dad9: Keep mounted framework endpoints reachable in dev and strip React Router HMR imports from transplanted app modules.
-
0b57293: Keep semantic settings URLs under the app's mounted workspace path.
-
0b57293: Fix Connect Builder (and other
agentNativePath/appPathcalls) building the wrong URL in a multi-app workspace dev gateway when the current page's client-side route (e.g./settings) isn't itself a workspace app id. PreviouslyappBasePath()would blindly trust the URL's first path segment as the workspace mount, producing URLs like/settings/_agent-native/builder/connectthat the gateway 404s into its app-picker page instead of Builder's real sign-in screen. The guessed segment is now validated against the deployed workspace app manifest when one is available. -
0b57293: Extend the human-in-the-loop tool approval grant window from 15 minutes to 1 hour. A user who stepped away between seeing an "Approve to run..." prompt and clicking it (e.g. to update their client) could return to a silently expired grant — clicking Approve did nothing because the durable row no longer matched
expires_at > now, with no error shown. -
bcd4c14: Time out session-replay uploads so a hung request releases the flush lock instead of growing the replay queue for the rest of the session, and bound the extension-marker scan by its watermark instead of reading every marker row ever written.
-
0b57293: Add regression coverage proving the magic-link
callbackURL/newUserCallbackURLconstruction survives Better Auth's ownoriginCheckvalidator end-to-end (not just a shape assertion) — this is the exact flow behind the{"message":"Invalid callbackURL","code":"INVALID_CALLBACK_URL"}reports from a UTM-tagged signup link and a retried sign-up after a stale?error=redirect. The existing absolute-URL promotion inbetterAuthCallbackURLalready fixed the underlying behavior; this closes the test gap so a future regression is caught even if the constructed URL still "looks" valid.Also stop silently swallowing a failure in the best-effort
email_verifiedrepair that runs after a successful verify-email redirect. A DB error there was previously indistinguishable from "nothing needed repairing," which is exactly the symptom in the "clicked the verify link, login still says not verified" reports — it's now reported viacaptureAuthError(still non-blocking) so a genuine failure is visible instead of silent. -
0b57293: Fix two bugs where a failure looked like success:
- First-run onboarding's Skip/Continue no longer silently do nothing when the completion save fails.
completeFirstRun()now rejects instead of swallowing a failed fetch or non-ok response, andFirstRunOnboardingsurfaces the failure with a "Try again" affordance instead of bouncing to an unrelated full-screen error. - A workspace file (including binary exports) now renders a download card the moment it's created —
show-workspace-file's binary content-type gate is gone, and any tool result shaped like a workspace-file card (e.g.web-request/provider-api-request'ssaveToFile) renders one automatically, without a second discretionaryshow-workspace-filecall.
- First-run onboarding's Skip/Continue no longer silently do nothing when the completion save fails.
-
0b57293: Keep replayed conversations faithful to what the agent actually did.
- Resuming a run (chained background continuation, agent-teams
continue) now
replays the tool calls and results stored inthread_datainstead of
flattening each turn to its prose, so a resumed chunk can see the output of
work already committed rather than re-running it. Integration turns keep their
existing delivered-text-only replay policy, and each replayed result is bounded
with an in-band truncation notice. - The outbound history window no longer slides by one message per turn. Every
prompt cache matches a byte-identical prefix, so a window that moved every turn
meant no cached prefix ever matched once a thread passed the message cap, and
the whole conversation was re-billed at write price on every turn. The window
start is now quantized to a stride. - Anthropic
redacted_thinkingblocks survive normalization and replay verbatim.
They were silently dropped as an unknown block type, which left the next
iteration of a tool-use turn sending an assistant turn the API rejects.
Unrecognized content block types now warn instead of vanishing. - Reducing a long thread is Observational Memory's job, but its Observer only
engages past 30k unobserved tokens while a 24-message count cap bit long
before that, so turns left the request while compaction still had nothing to
say about them. The count cap is now a backstop well above that threshold; the
two char budgets remain the real bound on what a request carries. - A thinking block with no signature is dropped with a warning instead of being
sent with an empty one, which the native API rejects outright — failing the
whole turn on a provider error that points nowhere near the cause. The Builder
gateway path is unchanged, since its tolerance here is unverified.
- Resuming a run (chained background continuation, agent-teams
-
16cbc53: Stop PR Visual Recap gate skips from creating visible pull request comments.
-
0b57293: Fix
provider-api-requestreporting a failed Slack send as a success. Slack's Web API always answers HTTP 200, even on failure, and encodes the real outcome asok: falsein the JSON body —chat.postMessagecalls that failed (e.g.not_in_channel,channel_not_found,msg_too_long) looked identical to a delivered message to any caller checkingresponse.ok, including the agent, which could then tell a user a Slack message was sent when it never was. Provider configs can now declarebodyOkFieldfor this always-200-with-body-encoded-outcome convention; the Slack provider sets it, and a body-levelfalsenow flips the response'soktofalseso a failed or unconfirmed send can no longer be reported as delivered. -
0b57293: Record a rejected Builder credential on the transcription path so it is not
retried forever. The chat engine already marks a 401/403 and stops reusing that
credential for the auth-failure TTL; transcription threw the raw upstream text
and marked nothing, so one unusable credential re-sent the same doomed request
on every attempt — 24 identical "Missing Authentication header" 401s in a day. -
Updated dependencies [16cbc53]
-
Updated dependencies [0b57293]
- @agent-native/recap-cli@0.5.5
- @agent-native/toolkit@0.16.5