BuilderIO/agent-native @agent-native/core@0.131.6

on GitHub

Patch Changes

• c0e7d64: Clarify workspace add-app feedback when no app is selected, and report when every available app is already installed.

• c0e7d64: Give delegated agent turns their app's actions as native tools in dev, so
  asking a sibling app a question actually works.

  In dev the interactive surface deliberately omits template actions from the tool
  registry and lets the agent call them through bash, which sidesteps the
  degenerate empty-object tool call some models emit for complex schemas. That
  trade is fine for a person — they see the bad call and rephrase. It is the wrong
  trade for a delegated turn. An A2A caller, or an external host calling ask_app
  over MCP, has nobody to intervene: with no native action the receiving agent
  shells out, the call misfires, and it repeats the same command until the
  repetition guard ends the run minutes later with no answer. Observed in a
  workspace as a sibling question that ran 4m45s and returned the wrong window,
  and elsewhere as runs stopped after eight identical bash calls.

  Both delegated surfaces now keep template actions native even in dev. A rejected
  {} call returns a schema error the model can correct on its next step, which
  is strictly better than a shell loop no caller can see or recover from.

• c0e7d64: Prevent timed Neon queries from leaking statement timeouts across pooled sessions and cancel background HTTP queries at their deadline.

• c0e7d64: Make cross-app delegation ask the receiving specialist agent by default, keep
  typed remote terminal states intact, retry idempotent transient transport
  failures, prevent recursive agent cycles, and bound delegated context growth.
  Proven durable-background delegated runs also keep the full bounded
  continuation allowance while sharing one cumulative wall-clock deadline, so a
  slow successful child task cannot strand its caller before the caller finishes
  its own tool work. After a provider exhausts its short in-call 429/529 retry
  budget, a proven background delegation now gets one cooled-down continuation,
  with a hard cap that prevents sustained throttling from becoming a request
  storm.

  Receiving agents keep ownership of source selection, schema interpretation,
  queries, joins, and their local tools. Direct read actions remain available for
  exact bounded contracts, but are no longer advertised as a workaround for an
  unreliable agent call.

  Dispatch now opts into the same durable background run contract it emits at
  deploy time, so delegated control-plane work is not cut off by the foreground
  40-second budget while already running in the 15-minute worker.

  Workspace vault ciphertext now prefers the workspace A2A-derived encryption
  key over each app's independent auth secret. Existing app-auth-encrypted rows
  remain readable by their owning app and are compare-and-swap migrated on read,
  so sibling agents can reliably resolve the same organization credentials
  without exposing or copying their values. Automatic engine selection also
  pairs the chosen provider with that provider's credential instead of reusing
  an unrelated active key.

  Documentation now distinguishes framework Core, optional Toolkit, and optional
  Templates, and makes source editing an explicit workspace/write-tool capability
  rather than assuming every embedded agent has filesystem access.

• c0e7d64: Retry transient provider TLS connection failures before surfacing them in agent chat.

• c0e7d64: Hide unauthenticated connection tests from MCP presets that require provider setup.

• c0e7d64: Recover background agent turns that stay alive without making real progress, retry transient run-event subscription failures with bounded backoff, keep reconnect diagnostics honest, and deduplicate repeated system prompts in thread debug history.

• Updated dependencies [c0e7d64]

• Updated dependencies [c0e7d64]

  • @agent-native/toolkit@0.12.0