Skip to content
fold-kit-experiments

Searchable grouped endpoint qualification

  • result: complete
  • claim: shared package-level search/grouping and native controls work for creation and reconnection, with Model-owned workbench queries.
  • evidence: workspace gates, 226 Flow tests, strict packed consumer, TypeDoc, production sample build, five browser probes and inspected screenshots.
  • remaining uncertainty: large-graph latency, native popup arrow navigation, screen-reader/device, CI and hosted/registry qualification are not established.
  • next: benchmark projection/policy cost and browser typing latency on larger workflows before selecting an optimization.

Package behavior

resolveEndpointChoices is a pure presentation projection shared by creation and reconnection. It matches literal case-insensitive tokens across labels and endpoint IDs, preserves entry/event references and input order, and retains current selections even when they do not match. Node groups use node identity, so equal labels do not merge unrelated nodes. Optional node-label metadata has an ID fallback for older/custom entries.

Native field/select/reasons options accept query, groupByNode and custom groupLabel; flat unfiltered behavior remains the default. No-results feedback explicitly explains retained selections. Matching protected choices remain disabled and operation-wide reasons remain visible. Empty creation results keep their placeholder instead of implicitly selecting another endpoint.

The FoldKit entry provides controlled search input, Clear and feedback primitives. Escape in a search clears only that query; creation-select Escape still cancels a pending source. Clear remains focusable after its action. Query changes do not edit the graph or history. The workbench stores four query strings in its Model and composes package primitives for creation and reconnection; matching/grouping is not duplicated in Astro. Queries remain view preferences across Undo and selection changes, with explicit Clear controls.

Evidence

  • bun typecheck: exit 0 across the workspace; the affected package typecheck also passed after the final feedback-copy refinement.
  • bun run check: exit 0, including formatting and catalog validation.
  • bun run test: final workspace rerun exited 0; Flow reported 226 passed, zero failed, including the strict packed-consumer test.
  • Six new tests cover token/ID/literal matching, protected current selection retention, stable grouping/order, equal node labels, legacy metadata fallback, stale authorization rejection, native groups/custom labels, no-result placeholders, controlled search input and associated feedback.
  • The strict external packed consumer compiles the new pure resolver, grouped native controls and query primitives against installed local npm tarballs. The initial full run caught an inline fixture message whose tag widened to string. The fixture now uses FoldKit’s defineMessageUnion; its isolated packed-consumer rerun passed before the full suite was restarted.
  • bun run --filter @opsydyn/foldkit-flow docs: exit 0 with TypeDoc warnings treated as errors. README, changelog and ADR 0004 document the new contract.
  • bun run --filter @opsydyn/flow-sample build: exit 0.
  • From apps/web, bun - < ../flow-sample/evidence/endpoint-search-browser.mjs: exit 0. It verifies native grouping, literal search, pinned current selections, protected choices, query-only Escape/Clear, focus, keyboard creation and reconnection, Undo/Redo and minimal mode, with no captured page errors.
  • Existing creation, keyboard-reconnection, reconnection-permission and port-permission browser probes: exit 0 against the final sample build. The creation probe now traverses the newly composed search controls in natural Tab order. Historical screenshots remain preserved; this slice adds new evidence.

The keyboard evidence uses native select type-ahead. Native popup arrow navigation, screen-reader and physical-device acceptance are not established by this slice. No CI, hosted deployment or npm publication was performed.

Visual evidence

Both final screenshots were inspected. Group structure is asserted against native optgroup elements; the closed-select screenshots primarily show query, retained selection and protection feedback.

Self-review

Boundary verdict: safe with noted risk. The pure projection imports only a core type; FoldKit/Effect remain in the optional adapter. Existing operation contexts/events and commit-time policies are preserved. There is no package-owned query store, document-schema change, dependency, direct DOM mutation or custom combobox state machine. Native control identity and current option identity remain stable while filters change. Tests cover both permitted and denied commits, not only rendered option counts.

Filtering reduces the displayed choices; it does not virtualize the list or reduce candidate policy evaluation. No response-time or large-graph performance claim is made. Recommend a bounded benchmark of projection cost and browser typing latency on larger workflows before choosing an optimization.

Recommendation: needs follow-up for that measured performance slice. No PR or publication is requested.