BuilderIO/agent-native @agent-native/core@0.42.0

on GitHub

Minor Changes

• 58676e6: app-skill packer now generates auto-updating plugin manifests so installed plugins pick up skill changes without a manual re-pack: Claude Code marketplace entries set autoUpdate: true (with commit-SHA plugin versioning) and Codex plugin versions embed a content hash of the bundled skills plus the MCP endpoint. This backs distributing the Agent-Native Plan app (and any app-backed skill) as a one-install Claude Code / Codex marketplace plugin that bundles the skills and the hosted MCP connector together.

Patch Changes

• 58676e6: Fix fresh-install scaffold builds failing on @assistant-ui/tap/react-shim.
  Newly published @assistant-ui/store@0.2.14 bumped its @assistant-ui/tap peer
  to ^0.6.0 and started importing the @assistant-ui/tap/react-shim subpath,
  which exists only in tap 0.6.0. But @assistant-ui/react@0.12.x (our pin)
  still depends on @assistant-ui/tap@^0.5.x, so the single hoisted tap resolves
  to 0.5.14 — which has no react-shim export. A lockfile-less pnpm install
  (the scaffold E2E CI job and any freshly-created app) floated store to 0.2.14
  and the calendar build then failed under Rolldown/Vite with:

  "./react-shim" is not exported … from package @assistant-ui/tap
  ERR_PNPM_RECURSIVE_RUN_FIRST_FAIL  calendar@ build: `agent-native build`
  

  Pin @assistant-ui/store to 0.2.13 (the latest release that still peers
  @assistant-ui/tap@^0.5.x and does not import react-shim) and add an explicit
  @assistant-ui/tap@^0.5.14 so the whole @assistant-ui/* family resolves to a
  self-consistent, tap-0.5.x generation (react 0.12.28 · store 0.2.13 · core
  0.1.17 · tap 0.5.14) on a clean install. No app-code changes — assistant-ui
  is consumed exactly as before.

• 58676e6: Before/after comparisons in plans and visual recaps now label their states with
  the column header instead of a pill baked into the wireframe, and wide frames
  stack instead of cropping.

  • The columns block renders each column's label (e.g. Before / After) as
    a small <h4 class="plan-columns-label"> heading above that column, in
    both the read and edit surfaces. The state name lives outside the content, so
    authors no longer write a Before/After pill inside a child wireframe's
    mockup (where it read as part of the product UI and landed in a random corner).
  • A comparison whose columns hold a wide wireframe (surface: "desktop" or
    "browser") now lays out as a single vertical stack at full document width
    instead of two half-width cells, so a large frame is never squeezed and
    cropped. Narrow surfaces (mobile, popover, panel) stay side by side.
  • The visual-plan / ui-plan / visual-recap skills' shared Wireframe Quality
    core is updated to teach this: name states with the column header (never inside
    the frame) and let the surface choose side-by-side vs. stacked.

• 58676e6: Rich-markdown drag handle: make container blocks (columns, and any nested-region
  block) draggable to reorder again. Each column inside a columns block mounts its
  own nested editor that tiles the container's full footprint and extends an 8rem
  forgiving hover zone into the container's left-margin gutter. The grip's
  "smallest editor wins" hover arbitration therefore handed the gutter to the inner
  column editor everywhere, so the outer columns block (e.g. a before/after
  wireframe pair in a visual plan) could never be selected or dragged.

  Hover arbitration now splits candidates by cursor position: over a block's body
  the innermost (smallest) editor still wins so nested blocks stay grabbable from
  their content, but in the shared left-margin gutter where the grip lives the
  outermost (largest) editor wins so the container can be picked up and reordered.
  Sibling non-container blocks are unaffected.

• 58676e6: create-extension and update-extension can now host a large pasted file by
  reference. When a user pastes a big HTML/Alpine file and asks to host it as an
  extension, the model passes contentFromAttachment (the pasted attachment's
  name, or the literal "latest") instead of copying the whole file into the
  content tool argument. The server resolves it from the turn's attachments —
  which the agent loop now threads into each action's ActionRunContext — so the
  model never re-emits thousands of tokens of pasted content.

  Re-emitting a large paste as content was the root cause of the
  create-extension continuation loop the repetition guard mitigates; this removes
  the need to regurgitate it at all. The inline content path is unchanged.

• 58676e6: Registry block-node legacy-block editing. Legacy (unregistered) document blocks
  can now either:

  • render their own self-contained edit overlay via a legacyBlockSelfEdits
    side-map predicate — the node view renders the block in edit mode and adds NO
    separate corner edit surface (no pencil/JSON/form popover), so the block owns a
    single control overlay (e.g. an image block with one hover toolbar); or
  • supply a schema-driven form editor via the renderLegacyBlockEditor side-map
    hook, rendered in the block-edit popover instead of the raw-JSON fallback.

  The raw-JSON fallback editor's Save button is also right-aligned instead of
  stretching full-width.

• 58676e6: Fix dragging blocks by the grip when they live inside a container block
  (columns/tabs). A container block's node view runs its onMouseDownCapture
  block-select handler in its own React root, so its stopPropagation() halted the
  native mousedown before it ever reached an inner editor's drag-handle grip —
  nested column blocks could not be picked up at all, so dragging a block OUT of a
  column, BETWEEN columns, or onto another block to make new columns silently did
  nothing. Two parts: clickedInteractiveChild now treats the drag handle and any
  target inside a nested editor region as interactive (so the container spares a
  mousedown meant for an inner grip), and the grip's icon is now
  pointer-events:none so a real mouse-down — which lands on the SVG, not the grip
  div — resolves to the grip itself instead of being swallowed.

  Also adds an optional onEditorReady callback to the shared rich-markdown editor
  so a host can capture the root editor view (used by the plan editor to rebuild
  the document when a column move dissolves its container).

• 58676e6: PR Visual Recap: make the Codex backend actually work in CI. Two fixes:

  1. Auth. The Run agent (Codex) step now pipes OPENAI_API_KEY into
     codex login --with-api-key (writing ~/.codex/auth.json) before
     codex exec, instead of relying on the bare environment variable. On the
     gpt-5.5 WebSocket transport Codex was dropping the Authorization header on
     the wss path and its HTTPS fallback, so every recap failed with 401 Missing bearer or basic authentication in header and the PR comment reported
     "generation failed".

  2. Sandbox + approvals. codex exec now runs with
     --dangerously-bypass-approvals-and-sandbox instead of
     --sandbox workspace-write. Codex's bundled bubblewrap sandbox cannot
     initialize on the GitHub runner, so every shell command failed at startup and
     the agent could not read recap.diff; and under an approval gate the
     write-side plan MCP call (create-visual-recap) was auto-cancelled. The
     runner is itself an ephemeral throwaway sandbox, so this is the documented CI
     invocation.

     Both fixes land in the in-repo workflow and the bundled copy the CLI writes into
     consumer repos (kept byte-identical by the recap sync test).

• 58676e6: PR Visual Recap: restore the inline screenshot in the sticky PR comment. The
  recap's PNG upload to POST /_agent-native/recap-image was silently failing, so
  recap shot returned an empty imageUrl and the comment fell back to a
  link-only body.

  Root cause: h3 v2's readRawBody(event, false) resolves a bare Uint8Array,
  not a Node Buffer. The route then called Buffer-only methods on it — the PNG
  magic-byte check (Buffer#equals) threw (.equals doesn't exist on a
  Uint8Array), surfacing as a 500 so the CLI saw !res.ok; and even past that,
  the store's png.toString("base64") would have silently mis-encoded the bytes
  (a bare Uint8Array ignores the encoding argument). The upload route now
  normalizes the body to a Buffer once before the magic-byte check and storage,
  so both the raw image/png and JSON { pngBase64 } paths persist the exact
  bytes uploaded.

  recap shot's image-upload helper now also logs the HTTP status / response
  snippet to stderr on failure (stdout still carries only the machine-readable
  JSON the workflow parses), so a future upload failure is debuggable from the CI
  log instead of vanishing into a null imageUrl. The route's unit test mock now
  mirrors h3 v2 by handing the handler a real Uint8Array, which would have
  caught the original regression.

• 58676e6: Recap code-diff blocks: fix duplicate/loading tab content, render per-diff
  descriptions, and steer recaps toward a labeled "Key changes" section.

  • Diff summary now renders as a description above the code (previously a
    trailing note below it), so a one-line "what this hunk changes and why" reads
    before the diff — the natural order for a review recap.
  • visual-recap skill now instructs the recap generator to give every
    diff a one-line summary, and to introduce a group of file diffs with a
    ## Key changes rich-text heading rather than relying on a tabs block title.
    All four skill copies stay byte-identical (guarded by skills.sync.spec.ts).

  These pair with a plan-template fix where switching tabs in a vertical tabs
  block (the recap "Key changes" diff rail) reused one nested editor instance, so
  its content reconciler skipped re-applying the newly-selected tab — surfacing as
  "every tab shows the same diff" and a stuck "Loading diff block…". The nested
  editor is now keyed per container region so each tab remounts with its own
  content.

• 58676e6: Agent chat now bails out of a degenerate repetition loop quickly instead of
  burning the entire auto-continuation budget. When the model gets stuck
  re-streaming the same narration every continuation without ever finishing a
  tool (the classic "paste a large HTML file and ask to host it as an extension"
  failure), it previously counted each repeat as progress and ran all 32 transient
  continuations — re-sending the large pasted payload each round — before surfacing
  a generic connection_error. A new repetition guard detects the non-advancing
  loop and stops after a few rounds with a clear, actionable message, and the
  create-extension large-payload nudge now also fires on mid-stream cutoffs
  (stream_ended), not just run timeouts.

• 58676e6: The /visual-plan skill now prescribes a gated, read-only, non-blocking
  adversarial self-review before handoff: surface the plan first, then spawn one
  skeptical reviewer (concurrently, against the written plan only — no
  re-research) for high-stakes architecture/backend/data/multi-file plans, apply
  clear-cut fixes via update-visual-plan patches, and route genuine judgment
  calls back to the user. Plan Discipline also gains a reuse-first instruction
  (name what each step reuses before what it adds) and a "decide the hard-to-
  reverse bets first" instruction (settle wire format, public ids, data-model
  shape, auth, and ownership before scoping to the smallest first cut). The
  guidance is byte-synced across the shipped skill constant, the template copy,
  and the exported mirror.