mirror of
https://github.com/webstudio-is/webstudio.git
synced 2026-11-04 02:58:23 +09:00
Oleg Isonen
e80ca3b54c
Closes #5954 Closes #5951 Closes #5950 Closes #5953 ## Summary - route external screenshot `url` and `baseUrl` targets through the self-closing browser path - keep reusable Chromium sessions limited to Webstudio-managed session previews - validate typed style values at the public runtime boundary while retaining CSS strings for design tokens - build MCP and shell validation messages from compact structured issues - advertise typed style objects with a concrete keyword example - persist `update-asset` metadata patches to the related file record and verify the saved value - store multi-sentence page titles and metadata as fixed text without misclassifying sentence punctuation as member access - capture privacy-safe aggregate token usage and MCP workflow metrics in every high-impact evaluation result - compare each evaluation with a compatible checked-in baseline and fail on correctness, token, tool-use, error/retry, broad-read, or MCP-latency regressions ## Root causes ### External screenshot lifecycle The fast iterative preview change made `session` the default preview source. External captures were then treated as managed session-preview captures and opened a reusable Chromium session. One-shot MCP processes do not call `preview.stop`, so the browser kept the process alive and could be orphaned when the caller killed the hung CLI. ### Validation errors and schemas `update-styles` declared `value` as unconstrained input, and design-token objects were checked only for a string `type`. Invalid values therefore passed the advertised/input schema and failed later in the full `StyleValue` parser, exposing the raw Zod union message. Error serialization already carried flattened issues but did not use them for `error.message`. ### Asset metadata persistence `update-asset` correctly produced an `assets/[id]/meta` patch, but asset persistence compared only Asset-table fields. Metadata lives on the related File row, so the patch was treated as unchanged and the successful mutation never wrote the new value. ### Page text expression disambiguation Page text normalization treated a period followed by whitespace and a word as JavaScript member access. Normal prose such as `Marques. Toutes nos marques` was therefore validated as an expression and rejected instead of being stored as a string literal. ### Evaluation observability The evaluation result only stored duration and two call counts, so token cost, failures, retries, mutation flow, verification coverage, and MCP latency could regress without a durable comparison. Results now use a bounded schema-v2 aggregate and never persist prompts, transcripts, stdout/stderr, credentials, arguments, or response payloads. ## Impact - external screenshots complete without retaining browser processes - repeated managed-preview screenshots retain browser reuse - invalid style inputs fail at the public boundary with paths such as `updates.0.value` - `error.message` is concise and actionable instead of a repeated raw union dump - generated schemas no longer advertise primitive or array values for `update-styles` - supported design-token CSS strings remain accepted - `update-asset` font metadata remains changed after commit, refresh, and reload - `create-page` and `update-page` accept ordinary multi-sentence titles and SEO descriptions while preserving computed expressions - every high-impact run reports input, cached-input, cache-write, output, reasoning-output, total-token, and cache-rate aggregates - MCP results report calls, success/failure, retries, focused/broad reads, dry runs, planned/committed mutations, verifications, total/p50/p95 latency, and first-mutation/verification timing - compatible results are compared with accepted baselines using a 15% relative regression threshold; correctness failures and zero-baseline errors, retries, or broad reads remain zero-tolerance - stochastic wall-clock and milestone timings remain visible as deltas without making a single noisy run fail ## Todo - [x] Reproduce external URL browser-session routing - [x] Use self-closing capture for external screenshots - [x] Preserve browser reuse for managed session previews - [x] Reproduce primitive style-value schema acceptance - [x] Validate typed style values at the public boundary - [x] Preserve CSS-string design-token inputs - [x] Format stdio and shell validation messages from structured issues - [x] Regenerate and verify runtime operation contracts - [x] Reproduce dropped asset metadata at the persistence boundary - [x] Persist and verify related File metadata updates - [x] Reproduce multi-sentence page text failures on create and update - [x] Distinguish sentence punctuation from computed member access - [x] Add privacy-safe token and MCP workflow metrics - [x] Add per-fixture baseline comparisons and regression gates - [x] Refresh accepted schema-v2 baselines for all fixtures - [x] Review the complete diff - [x] Run repository checks and the complete agent evaluation suite ## Verification - `pnpm checks` — passed, including workspace tests, typechecks, lint, package-boundary checks, generated API/docs checks, fixture sync, and fixture builds - builder tests — 189 files, 2,356 tests passed - CLI tests — 53 files, 852 tests passed - project-build tests — 81 files, 2,023 tests passed - asset-uploader tests — 11 files, 98 tests passed - targeted evaluation metrics/regression/trace/runner tests — 5 files, 17 tests passed - all four issue regression tests and the new evaluation harness tests were demonstrated failing without their implementations and passing with them - `pnpm evaluations -- --update-baselines` — 3/3 fixtures passed and schema-v2 baselines were updated - a subsequent real `font-assets-v1` result compared successfully after the regression policy was finalized