BuilderIO/agent-native @agent-native/core@0.136.0

on GitHub

Minor Changes

• 2b6fea3: Fix the Automations page reporting runs that never happened, and add run history and schedule editing.

  A scheduler or dispatcher tick that declined to run an automation used to stamp
  lastRun with the current time, so a permanently blocked automation reported a
  fresh run every minute while its nextRun stayed frozen in the past. Skipped
  ticks now record lastCheck instead, leave lastRun alone, and only rewrite the
  resource when the failure state actually changes. The failure reason
  (lastError) is surfaced on the row and in the details view instead of being
  swallowed behind a bare skipped chip.

  Schedules are also timezone-aware. Cron expressions used to be read in the
  server's zone, so an automation created as "every day at 8am" ran at 8am UTC.
  A schedule now stores the IANA zone it was written in, taken from a new
  scheduling timezone preference in Account settings and falling back to the
  caller's browser zone. Descriptions name their zone ("Every day at 8 AM
  (America/New_York)"), and both the agent tools and the schedule editor accept a
  timezone. Existing schedules keep their current host-relative meaning until
  edited.

  Also adds:

  • automation_runs history for real executions, exposed through a new
    list-automation-runs action and a Past runs section in the details view.
  • A Details view that shows more than the list row: schedule, next/last run,
    last checked, last status, scope, creator and model.
  • An Edit affordance for changing a scheduled automation's cron expression and
    timezone.
  • A "Manage agent" entry in the sidebar organization switcher.

  Run history is bounded and honest about interrupted runs: a row left running
  past the point a run could still be alive is reported as interrupted rather
  than shown as permanently in-flight, and rows are pruned per automation so a
  frequent schedule cannot grow the table without limit. Recording a run's
  outcome also re-reads the automation first, so a schedule edited while it was
  running is no longer reverted by the completion write.

  Also from review: a completion write no longer recreates an automation deleted
  mid-run, a run-history write failure can no longer reclassify a completed
  automation as failed, deleting an automation forgets its run history so a new
  one reusing the name does not inherit it, an unusable X-User-Timezone header
  is rejected rather than persisted, and a settings read failure surfaces instead
  of silently pinning a schedule to the host zone.

  Run history never blocks the automation it describes: opening the record,
  attaching its thread and closing it out are all non-fatal, so an unwritable
  history table costs the record rather than the run.

Patch Changes

• 2b6fea3: Cache peer agent cards instead of re-probing every sibling app on each lookup.
  describe-workspace-apps ships in the default first-request tool set and the
  <available-apps> prompt block names it, so a single turn could probe every
  peer and the next turn would do it all again. Against a local dev gateway each
  probe also cold-starts the app it touches, so one tool call spawned a dev server
  per sibling at once and the machine stalled behind them. Cards are now cached
  per caller for 30s with concurrent probes collapsed onto one request; failures
  expire after 5s so a peer that was still booting is retried promptly rather than
  being reported skill-less.

• 2b6fea3: Add agent-native clean to reclaim disk from regenerable build caches, and report disk usage in agent-native doctor. clean is a dry run unless --apply, prints the bytes it reclaims per category, and surfaces any delete it could not complete instead of reporting a clean total. It refuses any root it cannot confirm is an Agent Native project — a package.json that parses and depends on @agent-native/core, an agent-native.json, or a workspace with an app under apps/ that has either; a manifest it cannot read or parse is a refusal naming the file and the reason, never a permit. Every target is a real directory entry matched by exact name, so a hand-written Build/ is not selected by the build rule on a case-insensitive filesystem, and each one is re-checked against the identity recorded at scan time immediately before the delete, so a parent swapped mid-run is a reported failure rather than a delete somewhere else. It treats an unknown flag or a valueless --cwd as a usage error rather than guessing, and counts a hard-linked file only when every link to it is inside the delete set — one deploy bundle linked into several function directories counts once, and one still linked from node_modules/ counts nothing, because removing it frees nothing. A byte is credited only where the run observed it removed, so an outcome it could not observe is a typed result rather than a number: a tree another process deleted first — a second clean, or Vite recreating .vite mid re-optimize — credits nothing instead of two runs both claiming it, a target already gone when re-checked is a no-op rather than a failure, a re-measure that cannot read what survived reports the bytes as unknown rather than as zero, and a scan stopped by the walk-depth cap says so instead of returning a short total. Under apps/, each app must confirm for itself that it is Agent Native before anything inside it is selected, so one app no longer licenses deleting a sibling Rust project's build/ or a personal folder's dist/. Protected names — data/, .git, node_modules, .env* — now match case-insensitively while targets stay case-exact, so Data/ is not descended into on a case-insensitive filesystem. The walk and the delete stop at a filesystem mount boundary. One unreadable directory is reported once rather than once per pass, a partial run's headline states the exact shortfall rather than rounding to look complete, and --json --help answers in JSON. doctor now shows free space on the volume holding the project and flags low free space; doctor --disk also measures how much clean could give back.

• 2b6fea3: Keep workspace development lazy by default so unused apps do not build Vite
  dependency caches and consume disk and memory. Background prewarming remains
  available with --prewarm or WORKSPACE_PREWARM=1.

• 2b6fea3: Fail closed when an ai-sdk:* engine has no provider key, instead of sending an
  unauthenticated request. The provider factory was previously built with no
  apiKey, so the SDK omitted the Authorization header and the gateway's 401 came
  back as http_401 "Missing Authentication header" — a transport error naming the
  wrong cause, which a scheduled job then retried on every tick forever. It now
  reports missing_credentials and names the env var it wants, matching what
  builder-engine and anthropic-engine already did.

  Also stop reaping in-process background automations (scheduler and trigger runs)
  at the tight 45s post-claim stale window. That window exists to reach a durable
  successor sooner, but these runs carry no dispatch_payload and have no
  successor to reach, so an early reap killed still-working jobs that nothing could
  recover. They now get the 90s background window, and the recovery path reports
  not_redispatchable rather than payload_missing, which read as data loss for
  the one case where nothing was ever lost.

• 2b6fea3: Stop writing a second full copy of the server bundle for every extra Netlify
  function. The durable-background, integration-recovery, workspace
  <app>-server, and Vercel <app>-server.func emits now share one on-disk copy
  through hard links, so a build no longer doubles (or, in a workspace, multiplies)
  its function output. Each function still ships a complete, independent bundle —
  only the wasted disk goes away. Builds also stop emitting the throwaway
  dist/<app>/<app> client build for presets that already mount publicDir at the
  app base path.

• 2b6fea3: Harden hosted authentication and preview database isolation against account-claim exposure.

• 2b6fea3: Stop writing sourcemaps into the Vite dependency pre-bundle cache, which roughly halves node_modules/.vite/deps (and the transient double-size peak during a re-optimize) so workspaces stop running out of disk. Set AGENT_NATIVE_DEP_SOURCEMAPS=1 to restore them when stepping into third-party code in the debugger.

• 2b6fea3: Stop workspaces from accumulating duplicate physical copies of
  @agent-native/core.

  agent-native upgrade now rewrites the latest specs it installs back to the
  exact resolved versions, so a committed manifest pins one release instead of
  re-resolving on every install. If a version cannot be read after a successful
  install, upgrade reports which specs are still floating rather than claiming
  the upgrade finished. A package.json that cannot be parsed is now named in
  that report and stops the run, instead of being skipped as if it were clean.

  Apps added to an existing workspace inherit the framework versions the
  workspace root already pins, and freshly scaffolded workspaces resolve
  @types/node, esbuild, srvx, and zod once workspace-wide. typescript
  is no longer a peer dependency of core — core never imported it, and the peer
  edge forked a separate ~175 MB copy of core per TypeScript version in use.

• 2b6fea3: Keep the full Agent workspace reachable from chat-only sidebars, including generated migrated apps. The sidebar link now opens the Agent page where context, resources, connections, automations, and access are managed.

• 2b6fea3: Allow organization owners and admins to require Google sign-in, revoke current sessions when enabled, and enforce the policy across all auth entry points.

• 2b6fea3: Keep local package linking fast and deterministic when scaffolding during framework development.

• 2b6fea3: Keep scaffolded workspaces compatible with the current h3 and srvx releases.

• 2b6fea3: Share the browser's realtime sync transport across feature subscribers and reconnect local SSE streams after terminal failures.

• 2b6fea3: Share one sync stream per origin across browser tabs instead of opening one per
  tab. The browser caps HTTP/1.1 connections at roughly six per origin per browser
  process, so every extra tab's EventSource permanently consumed one of them and
  ordinary requests queued behind the held streams — worst in local development,
  where the dev gateway serves every workspace app from a single origin. Tabs now
  elect a stream holder through Web Locks and receive its frames over
  BroadcastChannel; followers relax to the fallback poll cadence, and Web Locks
  promotes a new holder automatically when that tab closes. Browsers without Web
  Locks or BroadcastChannel keep the previous per-tab behavior.

• 2b6fea3: Make browser Demo mode visible in the shared organization switcher, explain what it changes, and avoid showing the redacted anonymous email as the signed-in identity.

• 2b6fea3: Shrink the published package by roughly 47 MB of allocated disk per install.
  Source maps, the corpus/core and corpus/toolkit trees, and the src/ tree
  were three separate copies of source that already ships as dist/, docs/, and
  @agent-native/toolkit's own src/. The tarball now ships only the eject
  entries and scaffolding templates out of src/, and corpus/ carries the
  first-party template source that source-search actually reads. Docs and
  source-search now point at dist/ for framework internals instead of a corpus
  path that no longer exists.

• 2b6fea3: Forward incremental action input to browser chat surfaces so apps can preview generated content while an action is being prepared.