Skip to content
fold-kit-experiments

Endpoint chooser qualification

  • result: complete
  • claim: reconnection choices and native controls ship at package level, with custom inspector and toolbar usage in the Astro workbench.
  • evidence: focused core/composition tests, packed consumer compilation, TypeDoc, workspace gates and the local browser probe listed below.
  • remaining uncertainty: native popup arrow navigation, screen-reader and physical-device acceptance are not established by this slice.
  • next: extend policy-aware candidate discovery to keyboard creation of new connections, reusing the same validation boundary.

Changes and ownership

resolveEdgeReconnection derives typed choices from the current Model. Shared validation checks direction, changed-port permission, duplicate edges and global connection policy. The core revalidates each EdgeReconnected proposal; preflight does not grant permission for a future state. Edge metadata, selection and atomic history remain owned by the existing reducer.

The FoldKit entry exports native field/select/reason primitives. Controls use stable endpoint values and scoped IDs; reasons remain visible independently of diagnostic settings. The expanded toolbar uses a scrollable band clear of reconnection handles. The existing pointer probe caught handle occlusion before this placement fix. No-alternative feedback leaves the current choice focusable, avoiding focus loss after the last eligible replacement is chosen. Explicit read-only, gesture and endpoint locks disable editing.

The workbench’s local enumeration/policy checks were removed. Its custom cards compose the package select and reasons, and both standard and compact edge toolbars consume the package context. The default toolbar remains unchanged unless edgeReconnection is enabled. No dependency, serialized schema, external write, PR or publication was added.

Local evidence

  • bun typecheck, bun run check and bun run test: all exited 0. The affected package typecheck also passed after the toolbar-placement fix.
  • bun test packages/foldkit-flow/test/edge-endpoint.test.ts: eight focused tests cover preflight/commit agreement, history/metadata, stale policies and candidates, visibility, missing endpoints, identity collisions, native semantics and opt-in toolbar composition and handle-safe toolbar placement. The full Flow suite contains 215 tests.
  • The existing packed-consumer test now imports and renders the new core and FoldKit exports from npm tarballs in an external strict TypeScript consumer.
  • bun run --filter @opsydyn/foldkit-flow docs: TypeDoc builds with warnings treated as errors. README, changelog and ADR 0004 describe the public contract.
  • bun run --filter @opsydyn/flow-sample build: production sample build.
  • From apps/web, run bun - < ../flow-sample/evidence/edge-endpoint-keyboard-browser.mjs. Native type-ahead changes source/target endpoints, retains focus, supports Tab exit and the connection-list toolbar handoff, and exercises Undo/Redo, port/end protection, minimal mode, a custom compact toolbar and the last-alternative focus case. Page errors must remain empty. Fixture setup uses programmatic policy controls; reconnection choices themselves use real keyboard events.
  • Existing reconnection/port permission probes use the scoped package inspector after this slice adds a second set of controls to the canvas toolbar.

The local macOS Chromium headless shell did not change even a plain native select using popup arrow keys. An isolated Chrome session reproduced that behavior. Native type-ahead did change both plain controls and the production chooser, so the durable probe uses it. This is evidence for that keyboard path, not proof of every native menu/OS interaction. Undo/Redo retains the editor’s pre-existing selection-clearing behavior; the probe reselects the restored edge before using its toolbar.

Visual evidence

Self-review

Boundary verdict: safe with noted risk. The pure core has no new browser or FoldKit dependency. Candidate policies run once per proposal and must be pure; large candidate sets may need a later search/filter UI. Default visible scope is intentional and documented, with explicit all-authored-port opt-in. Tests exercise both enabled and rejected commits, including a policy changing after preflight; custom renderers still own their widget semantics and must keep protection text associated with their controls. The expanded toolbar placement was checked against screen-space handle positions so new controls do not consume their drag hit targets; the regression probe explicitly checks the hit target before dragging.

Recommendation: needs follow-up for the next creation-chooser slice; no publication or PR is requested.