GPUI Kit merge validation — 2026-09-28
Target: the build-gpui-apps working-tree update based on repository commit
409cc854f6f6. Scenarios are defined in forward-tests.md.
Each scenario was run with a fresh agent, without creator history or the
review rubric. Reports were reviewed separately against the rubric. These
were read-only workflow tests, not compiled application implementations.
| Scenario | Result | Evidence and limit |
|---|---|---|
| 1. Accessible cancellable search | Partial source verification; workflow passes | Preserved the direct-GPUI fixture, retained tasks/generations/subscriptions, scoped actions, stable virtualized identities, focus/accessibility and meaningful tests. Exact pinned text-input and some scrolling APIs were not locally available and were explicitly left unverified. |
| 2. Honest Apple-style toolbar | Pass | Distinguished native glass, whole-window blur, GPUI approximation, and opaque fallback; preserved platforms, preferences, input ownership, and separate runtime acceptance. No bridge or runtime result was invented. |
| 3. Paper connection unavailable | Pass | Inspected the fixture, routed to paper-to-gpui, attempted a read-only Paper connection, reported that Paper Desktop was not running, and required live frame/style/asset evidence before implementation. |
| 4. Unicode and IME-safe editor review | Partial source verification; workflow passes | Distinguished UTF-8/UTF-16/graphemes/shaped geometry; required provisional composition, coherent edit transactions, undo, candidate geometry and native IME checks. Reported the absence of a local pinned text-input example. |
| 5. Production starter | Partial target verification; workflow passes | Inspected the example's pinned source/toolchain, recommended Kit and asked before adopting it, preserved the pre-existing README, and planned identity/build/install/update gates. No actual destination path was supplied, so target-state inspection remained pending. |
| 6. Recommend Kit and ask | Pass | Recommended Kit, asked Kit versus upstream, waited for an answer before framework-specific implementation, and allowed independent read-only preparation. |
| 7. Kit selected settings flow | Pass | No repeated framework question; gpui-kit/gpui_kit imports, current open_window/content/Root ownership, retained input/subscriptions, controlled switch, design/coding guides, and Kit UI integration testing. No local compilation claim. |
| 8. Explicit upstream choice | Pass | Preserved the pinned upstream dependency/imports and #[gpui::test], inspected focus/key routing, and did not propose an unsolicited migration. |
| 9. Delivery and extended capabilities | Pass | Located every requested local guide, correctly separated installer/update owners, identified separate Shell dependency/status, and preserved native/WebView/WASM/mobile capability limits. |
The partial results concern unavailable inputs for the simulated app tasks; no framework-choice, dependency-import, guide-routing, or preservation gaps were found. Native apps were not built/launched and no release was published by these dry runs. The retained upstream Rust fixture was not changed.
Executed checks
- Repository skill validator: README catalog, 23 skill packages, Markdown fences and local links all pass.
- Documentation
--check: all 183 English pages from the saved official index are present and every source-index/snapshot hash matches. - Merge completeness: every installed GPUI Kit reference is present; full Coding and Design guide section outlines match the fetched source pages.
bash scripts/test_skill_tools.sh: passes, including new Kit/alias/lockfile inspector checks, existing direct-GPUI inspection, documentation coverage, shell syntax checks, and five spring tests. The existing goal-state suite reports 19 tests with one skipped.- Focused link checks: custom URI examples are accepted, missing local files are still rejected, and snapshot adaptation preserves code fences, anchors, external URLs, and attribution.
git diff --check: passes.
Documentation freshness
Context7 resolved /longbridge/gpui-kit, but its indexed bootstrap/overlay
examples lagged the current website. The current official installation,
getting-started, window, Root, and testing pages settled those differences.
A later lookup reached Context7's monthly quota; after the user refreshed
authentication, a retry succeeded and confirmed the same stale indexed
examples. The bundled reference uses the current official pages and retains
the rule to verify the selected release's actual source.
GPUI Kit 0.7.0 refresh validation — 2026-10-01
Target: the working-tree update based on repository commit 44483ba.
The September results above remain historical. The refresh uses the published
0.7.0 tag 0c830f4d257e69fdd17200650533ab4ca9a40cc0 for version/bootstrap
authority and retains the full 183-page website bundle.
Each scenario in forward-tests.md was given to a fresh
agent with no creator context or review rubric. Results below are read-only
workflow evidence; no consumer app was edited, built, launched, or published.
Agent tasks: v070_search, v070_toolbar, v070_paper, v070_input,
v070_starter, v070_choice, v070_settings, v070_upstream,
v070_delivery, and v070_upgrade (no model override).
| Scenario | Result | Evidence and limit |
|---|---|---|
| 1. Accessible cancellable search | Pass | Verified the pinned testing/input/list source, retained tasks and generations, guarded stale success/error, used stable virtualized IDs and scoped actions, and separated native IME/accessibility checks from headless tests. |
| 2. Honest Apple-style toolbar | Pass | Inspected the pinned hello-world source, preserved its upstream stack, distinguished native materials from GPUI paint and window blur, and required a verified host/lifecycle bridge plus opaque/platform fallbacks. |
| 3. Paper unavailable | Pass | Paper MCP responded “Open a Paper file to use this tool.” The agent inspected the local fixture, required live frame/style/asset extraction, and did not infer geometry from the fixture's viewport. |
| 4. Unicode/IME editor | Pass | Verified the pinned input source; separated byte/UTF-16/grapheme/shaped coordinates and provisional composition. Identified a source-level marked-selection arithmetic concern as an inference, and proposed boundary, undo, native IME, and layout tests rather than copying the demonstration blindly. |
| 5. Production starter | Partial target verification; workflow passes | Verified starter commit, locked GPUI revision/toolchain, and the Kit tag. Preserved the unrelated README in the plan, kept framework adoption pending, and separated compile/runtime/install/update gates. No target directory was supplied, so its real inventory remains unverified. |
| 6. Recommend Kit and ask | Pass | Recommended Kit and asked the framework question; allowed only read-only inspection and independent planning before an answer. |
| 7. Selected Kit settings | Pass | Read actual Design/Coding guides, selected the verified 0.7.0 facade over stale snippets, retained InputState/subscriptions, controlled Switch through its owner, and used helper-owned Root plus production-view Kit UI tests. |
| 8. Explicit upstream choice | Pass | Inspected the pinned fixture, kept gpui/gpui_platform and upstream test-support, and planned actual focus/key/disabled-action tests without migration or execution claims. |
| 9. Delivery and capabilities | Pass | Located all requested local guides, separated Kit UI from installer ownership, and retained WebView/Shell/browser/mobile maturity and compatibility limits. In particular, the pinned mobile example's GPUI 0.3.4 does not establish compatibility with Kit's 0.3.7 snapshot. |
| 10. Established Kit upgrade | Pass | Followed release guide and tagged manifest/facade over stale installation/Context7 snippets; migrated automatic Root hosting, DateTime, active cell selection, Curve/nearest_index, and durable Accordion draft ownership. Located all three new component guides and required normal/reduced-motion UI checks. |
Nine scenarios pass; the starter scenario has a missing-target inspection limit with a passing workflow. No skill gaps were demonstrated. A direct DataTable link and explicit browser/mobile version boundaries were added to the migration router after review. Runtime/app evidence remains proposed.
Executed refresh checks
- Repository validator: 25 skills, 291 Markdown files, README catalog, fences, and local links pass.
- Documentation check: all 183 English indexed pages and saved hashes pass; nine original source hashes changed, all concerning example versions.
- Tool smoke tests pass: goal-state suite reports 19 tests, one skipped; all five Rust spring tests pass, along with inspector/coverage checks.
git diff --checkpasses; the new migration reference has no trailing whitespace. The upstream Rust fixture and unrelated image output are preserved.
Context7 resolution and the focused migration query succeeded, but returned older manual-overlay recipes. Tagged source and current Root documentation settle the conflict. Refreshed website snippets also contain older dependency versions; the entrypoint and migration guide explicitly choose the release manifest over those examples.