mirror of
https://github.com/webstudio-is/webstudio.git
synced 2026-11-03 10:37:48 +09:00
codex/email
20
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Oleg Isonen
|
a4c3b37fc2 |
fix: Unblock reported MCP workflows and improve issue diagnostics (#6438)
## What this fixes
MCP authors can reuse an existing design token in JSX fragments, create
valid border shorthands with CSS variables, and bypass a font that never
settles during responsive screenshots. Common fixed page titles with
spaced separators are also accepted. Issue reports now include the
project ID and safe, bounded IDs from the failed operation, plus clearer
response and browser-startup diagnostics.
## Scope and tradeoffs
- `token("name")` attaches an existing design token; an unknown name
returns a not-found error. Empty token definitions from Builder/UI
remain definitions and follow the normal conflict rules.
- Border shorthands map a single unresolved `var()` to the remaining
width, style, or color field when the other fields are explicit.
- `waitForFonts: false` skips font readiness but still waits for stable
layout frames. The default remains unchanged.
- Fixed titles such as `Pricing - Plans` stay literal. Addition such as
`first + second` remains an expression. Wrap ambiguous subtraction in
parentheses, for example `(first - second)`. Other ambiguous title forms
may still need an exact report.
- Only recognized operation-level IDs are attached to a public GitHub
issue. Authored values, raw arguments, URLs, and credentials are
excluded. IDs from later successful diagnostic reads are not captured.
- This does not fix the underlying response-transform failure; it
improves future reports for that case.
Closes #6424
Closes #6425
Closes #6434
Closes #6439
Ref #6435
Ref #6430
Ref #6417
## Verification
- [x] Reproduced the privacy, page-expression, empty-token, and
content-mode token-reference regressions with failing tests, then
confirmed the fixes pass. A regression test also covers parenthesized
subtraction.
- [x] Full protocol, template, vision, project-build, and CLI test
suites pass; focused Builder issue-report and fragment tests pass. The
protocol, CLI, and project-build suites were rerun after the CI fix.
- [x] Relevant package typechecks, lint, package-boundary check, fixture
pipeline, generated API check, and generated docs check pass. Fixture
output has no diff.
- [x] Reviewed the authoritative MCP documentation and regenerated the
API metadata and MCP docs.
- [x] CSS parser regressions cover border, each side shorthand, and
unresolved variables in width, style, and color positions; css-data
tests, typecheck, and lint pass.
- [x] Fixed the local evaluation runner to pass the task and Webstudio
MCP tools to Luna/low. Each of the six fixtures passed at least once;
focused MCP/evaluator tests, CLI and project-build typechecks, lint, and
formatting pass.
- [x] The complete Luna/low suite passed all six fixtures in two
consecutive sequential runs. Baselines remain incompatible with this
model/effort override and were not updated.
## Follow-up
- If another fixed-title form still fails, use the new project-level
report context and the exact title shape to narrow the parser case.
|
||
|
Oleg Isonen
|
39c290bfa7 |
fix: resolve MCP-reported runtime and component issues (#6362)
## Summary - harden MCP runtime behavior around asset uploads and resources, HTTP responses, component and Content Block application, data reads, and Chromium screenshots - reject shared Slots inside Content Block templates and avoid unsafe unused-asset findings when Content Blocks can hide MDX dependencies - expand named grid-area values according to CSS shorthand fallback rules - add a consent-safe YouTube preconnect control and only warm origins that are actually used - add focused and end-to-end regression coverage for the reported paths ## Verification - project-build: 2,419 tests passed - css-data: 4,623 tests passed, 1 skipped - sdk-components-react: 134 tests passed, 1 skipped - high-impact CLI fixture: 7 tests passed - Builder asset REST service: 22 tests passed Closes #6336 Closes #6339 Closes #6353 Closes #6354 Closes #6355 Closes #6356 Closes #6357 Closes #6358 Closes #6359 |
||
|
Oleg Isonen
|
e69a3d2f8e |
feat: Create structured MDX entries from collection folders (#6317)
## Paragraph editing follow-up — |
||
|
Oleg Isonen
|
05706b5425 |
feat: Make Content Block templates the MDX authoring interface (#6307)
Closes #6306 ## Outcome Content Block templates are now the public component model for MDX authoring. - Standard Markdown resolves through matching Content Block templates. - PascalCase JSX resolves to a named template first, then to a registered component. - Adding or changing a template immediately affects matching MDX, including content authored before that template existed. - Every supported standard MDX element is present in new Content Blocks by default. - Template `name` is a stable PascalCase JSX identifier. `label` remains optional display text only. - Lowercase JSX represents native HTML and SVG without exposing `ws.element` in authored syntax. - Unknown, ambiguous, and unsafe JSX stays in source, produces a source-ranged Builder diagnostic, and is omitted from published output. - Frontmatter images preserve exact `$ref` values and use the Asset description for `alt` text. - Keyboard and mouse insertion from the slash menu remove the trigger slash and preserve the selected template. ## Public MDX contract - Markdown is canonical for standard prose. - Lowercase JSX creates native elements: `<section><p>Text</p></section>`. - PascalCase JSX uses templates and registered components: `<PromotionCard />`, `<Heading tag="h1">Title</Heading>`, and `<YouTube src="https://youtu.be/KI5JHpzBK1s" />`. - Component-name collisions use stable qualified identifiers such as `Checkbox` and `RadixCheckbox`. - A self-closing template keeps its designed default subtree. An explicitly paired empty template clears it. - Static quoted values and bare booleans are supported. - Imports, exports, expressions, spreads, functions, executable JavaScript, unsafe attributes, `<$.*>`, and authored `<ws.element>` are rejected. - `ws` remains only for compiler primitives with no React component. ## Implementation - [x] Add stable, unique template names and keep them separate from labels. - [x] Parse and validate Markdown and JSX through their ASTs with precise source ranges. - [x] Share component identity between template JSX, Content MDX, Builder materialization, reload, copy, and publication. - [x] Resolve templates live after add, edit, rename, reorder, removal, undo/redo, collaboration updates, and reload. - [x] Preserve authored props, children, Asset references, template defaults, and the exact slash-menu template choice. - [x] Make one typed MDX semantic inventory drive default templates and slash-menu eligibility. - [x] Move component-dependent template assembly into the component registry. - [x] Replace production `$`, `animation`, and `radix` JSX namespaces with direct component imports. - [x] Keep only explicit compiler primitives in `ws` template JSX. - [x] Migrate repository templates, fixtures, stories, examples, MCP guidance, CLI guidance, and generated outputs. - [x] Add Builder unit and Playwright coverage for authoring, diagnostics, insertion, synchronization, reload, and publication. - [x] Review and update the authoritative Content Block and generated MCP documentation. ## Compatibility - Persisted project data continues to use namespaced component IDs and `ws:element` for generic HTML elements. No project-data migration is required. - Legacy `<ws.element ws:tag="…">`, `<ws.element ws:name="…">`, `<$.*>`, `<animation.*>`, and `<radix.*>` input remains readable during the compatibility window. - Current serializers and repository-owned sources emit only canonical Markdown, lowercase intrinsic JSX, or PascalCase identifiers. - Untouched user MDX is not rewritten automatically. Legacy syntax is canonicalized when its owning source is edited. - Deferred follow-up: remove legacy namespace readers and proxy support only after the canonical writer is deployed and existing projects are backfilled. This is intentionally outside this PR. ## Verification - [x] CI package tests, typechecks, lint, package boundaries, generated assets, migrations, and size checks pass. - [x] All six Builder E2E shards pass. - [x] Fixture checks pass on Ubuntu and Windows. - [x] Targeted local Content Engine, Template, SDK, registry, Project Build, migration, Builder, component, CLI, and Playwright tests pass. - [x] Runtime contracts, CLI docs, and MCP docs were regenerated from their owning sources. - [x] Local fixture generation completed without an unexpected fixture diff. - [x] Documentation impact reviewed and included in this PR. - [x] Review and approve the intentional List visual snapshot change. - [x] Run the complete agent evaluation suite after explicit approval. - [x] Run the network fixture link/sync pipeline against the matching PR deployment after explicit permission. ## Known visual change The generated List story now targets its actual default `ul` element instead of `ol`, so the default list styling appears in the visual snapshot. The visual change was reviewed and the approved visual check passes. |
||
|
Oleg Isonen
|
8718feb465 |
test: simplify browser testing infrastructure (#6271)
## Summary Unifies frontend testing around three intentional boundaries: - Vitest Node projects for pure and server tests - Vitest Browser Mode with Playwright/Chromium for DOM and component tests - Playwright Test for Builder end-to-end workflows The Builder no longer uses jsdom. Ordinary E2E workflows use an automatically managed isolated `page` fixture; explicit contexts remain only for simultaneous identities and suite-scoped setup. ## Technical choices - Split Builder tests into one Node project and independent browser projects so each Chromium/Vite process has a bounded workload. - Added shared Browser Mode configuration for the affected SDK React packages. - Preserve the Builder's production top-window bridge ownership; Browser Mode uses explicit test-only installation seams. - Centralized Playwright context options and kept action/navigation timeouts bounded. - Validate Playwright filters before Docker, migrations, browser installation, or Builder builds start. - Preserve the first failed E2E attempt's report, traces, screenshots, and videos when CI retries a shard. - Split static checks, typechecks, CLI tests, core package tests, browser package tests, and Builder projects into parallel CI jobs with five-minute hard ceilings. - Distribute Builder E2E tests across ten isolated shards with two workers; parallel suites create unique per-worker setup data. - Reuse color contrast DOM/canvas samples so the exhaustive theme sweep takes about four seconds instead of timing out at 30 seconds. - The removed cost is jsdom behavior drift, oversized serial check jobs, and unnecessary E2E startup for empty selections. Real browser startup and E2E environment setup remain. ## Todo - [x] Split Builder pure/server tests from browser/component tests. - [x] Migrate Builder DOM tests from jsdom to Vitest Browser Mode. - [x] Migrate affected SDK React DOM tests to the shared Browser Mode setup. - [x] Remove jsdom annotations, dependencies, and compatibility setup. - [x] Replace ordinary manual E2E contexts with the isolated `page` fixture. - [x] Keep explicit contexts only for concurrent identities and `beforeAll` setup. - [x] Preserve setup authentication isolation from workflow pages. - [x] Add bounded Playwright action and navigation timeouts. - [x] Preserve first-attempt E2E failure evidence across CI retry. - [x] Reject empty E2E selections before expensive environment setup. - [x] Parallelize repository checks into bounded CI shards. - [x] Isolate parallel E2E setup data across workers. - [x] Remove the color contrast sweep bottleneck. - [x] Update E2E contributor documentation. - [x] Review product documentation impact; no user-facing product docs are affected. - [x] Run the full repository tests and typechecks. - [x] Run formatting, lint, package-boundary, generated-file, and fixture checks. - [x] Complete final correctness and simplification review. ## Verification - `pnpm --filter @webstudio-is/builder test` - `pnpm -r typecheck` - `pnpm lint` - `pnpm check:package-boundaries` - `pnpm check:generated-api` - `pnpm check:generated-docs` - `pnpm fixtures` - Design-system: 222 tests pass locally. - Builder: all seven Vitest projects pass locally. - The formerly slow E2E shard passes locally in 1m55s. - Full database-backed Builder E2E execution passes in CI. - All CI jobs pass; the slowest completed in 4m58s. - No fixture inputs/outputs or generated files changed. Ref #6270 |
||
|
Oleg Isonen
|
9e504037fb |
fix: improve MCP reliability across project workflows (#6222)
## Summary Fixes the reproducible MCP, preview, redirect, audit, mutation, and asset-upload failures from the latest issue batch. Historical and non-reproducible reports were investigated without speculative fixes. Visual verification is now opt-in unless the user explicitly requests it. ## Issue checklist - [x] #6207 — report tested recovery when an explicitly selected browser fails to start. - [x] #6208 — accept generic and registered buttons as valid label targets without requiring a separate form label. - [x] #6209 — reject unknown style coordinates such as `breakpointId` and document the supported `breakpoint` field. - [x] #6210 — traced Node 18 access to CLI 0.293.0 predating the launcher guard; closed as already fixed by #6201. - [x] #6211 — verified the supported launcher blocks the reported Node 18 crypto path; closed as already fixed by #6201. - [x] #6212 — give sequential responsive captures the sum of their per-capture timeout budgets. - [x] #6213 — keep preview install, build, and server subprocesses on the parent CLI Node runtime. - [x] #6214 — add safe confirmation and diagnostics for ambiguous mutation responses; this does not claim to remove upstream latency. - [x] #6215 — confirmed as the same historical Node 18 release path as #6210; closed as already fixed by #6201. - [x] #6216 — verified malformed input returns structured errors on current main; closed as not reproducible and improved future failure diagnostics. - [x] #6217 — add transaction-id confirmation for large fragment mutations after ambiguous 5xx responses; this does not make the operation asynchronous. - [x] #6218 — verified the existing DevTools reconnect fix from #6205; closed as already fixed. - [x] #6219 — preserve the incoming pathname and query when a redirect target supplies only a hash. - [x] #6220 — return ordered `uploaded`, `failed`, and `ambiguous` results so partial asset successes are not lost or retried unsafely. - [x] #6221 — verified parser recovery without a Node crash; closed as not reproducible and improved future JSX failure diagnostics. - [x] #6223 — verified browser-session reuse, reconnect, and cleanup; closed as already fixed by #6205 after the reported CLI version. - [x] #6224 — expose both text-asset sources and encode the exactly-one requirement in the standard MCP schema. - [x] #6225 — could not reproduce the project-specific upstream timeout; closed with safe confirmation coverage and improved structured diagnostics, without claiming a latency fix. - [x] #6226 — could not reproduce the project-specific upstream timeout; closed with safe confirmation coverage for insert-fragment, without adding a speculative fragment-specific fix. - [x] #6229 — reproduced the reported two-process sequence with the current local CLI and a real browser; both captures passed, so the 0.293.0 report was closed without a speculative fix. - [x] Require explicit user consent before preview, screenshots, diffs, OCR, or rendered visual audits unless visual verification was requested. ## Technical notes - Ambiguous mutation responses are confirmed only with the same idempotent transaction id. If confirmation is also ambiguous, the session reports `AMBIGUOUS_COMMIT`, refreshes state, and never replays a fresh mutation. - CLI error mapping retains the original remote error as a non-enumerable cause, so JSON gateway status reaches the ambiguity classifier without exposing raw remote context in serialized diagnostics. - Forced asset uploads are not automatically retried after an unknown/network/5xx outcome. Batch results retain the original input index, including duplicate filenames. - Stable issue codes and constraints are retained for diagnostics. Arbitrary validation paths, messages, argument values, project content, and raw contextual data are excluded. - The MCP handshake remains the source of tool input schemas; goal-specific guides avoid duplicating those contracts and reduce model-context cost. - `update-asset-content` now exposes `path` and `content` with a machine-readable `oneOf` constraint, matching the runtime rule that exactly one source is required. - Redirect targets with their own pathname or query still replace those request components. - Generated runtime contracts, CLI documentation, product documentation, and deployment fixtures were regenerated from their sources. ## Verification - [x] Three independent review/test agents exercised visual consent and preview behavior; mutation, asset, and privacy behavior; and accessibility, styles, and redirects. - [x] `pnpm checks` passed after the final transport follow-up: all workspace tests and typechecks, lint with warnings denied, package boundaries, generated API contracts, generated docs and documentation checks, and the complete fixture link/sync/build pipeline. - [x] Changed-package suites passed, including Project Build (2,125 tests), HTTP client (42 tests), CLI, Protocol, and Builder. - [x] Complete `pnpm evaluations` run on the preceding implementation commit: all five real-world fixtures and every behavioral check passed. Four baseline comparisons passed; one font-assets sample exceeded only its model-output token allowance, and its isolated rerun passed. Evaluations were not rerun after the final transport and schema follow-ups because fresh approval is required. - [x] Design evaluation MCP response bytes improved from the 39.8 KB baseline to 37.6 KB after removing duplicated guide contracts. - [x] Documentation impact reviewed; generated MCP docs and multi-file upload guidance are updated. ## Related issues Closes #6207 Closes #6208 Closes #6209 Ref #6210 Ref #6211 Closes #6212 Closes #6213 Ref #6214 Ref #6215 Ref #6216 Ref #6217 Ref #6218 Closes #6219 Closes #6220 Ref #6221 Ref #6223 Closes #6224 Ref #6225 Ref #6226 Ref #6229 |
||
|
Oleg Isonen
|
5ec3547d36 |
fix: Support JSON content files in Builder (#6168)
## Summary - make the complete asset filename editable in Asset settings - convert existing editable text assets when their extension changes while preserving the asset ID - accept every JSON root type and normalize JSON-compatible syntax into formatted strict JSON - expose the same optional extension change through MCP/API `update-asset-content` - initialize new JSON files and empty text-to-JSON conversions with a valid empty object - distinguish browser URL imports from JSON file uploads so JSON objects with a root `url` field remain file content - update the Assets and Content Engine documentation ## Root cause Asset settings rendered the extension as a fixed suffix outside the name input. The metadata rename mutation could only change the display basename, while content revisions always inherited the current storage extension. As a result, changing `.md` to `.json` was impossible in the UI and unavailable through the shared MCP/API revision command. Asset settings now edits the complete filename. Ordinary renames keep using the metadata mutation. A text-extension change loads the current content, validates it for the target format, and writes a new immutable revision through the existing conflict-safe content endpoint. The existing Acorn-backed `parseJsonExpression` utility accepts JSON-compatible syntax without evaluating code and serializes it as strict JSON. Objects, arrays, strings, numbers, booleans, and `null` are all accepted. ## Checklist - [x] Create JSON files by entering a `.json` filename - [x] Edit the complete filename in Asset settings - [x] Convert editable text-file extensions while preserving the asset ID - [x] Expose extension conversion through `update-asset-content` - [x] Create JSON files and convert empty text files with a valid empty object - [x] Normalize safe JSON-compatible syntax to strict JSON - [x] Accept objects, arrays, and scalar JSON roots - [x] Reject executable or incomplete content - [x] Preserve browser-uploaded JSON object shapes - [x] Review and update authoritative Builder documentation - [x] Add focused UI, revision, HTTP-client, CLI/MCP, protocol, and upload regression tests - [x] Run affected typechecks, lint, documentation checks, and the Builder production build - [ ] Manually verify create, edit, rename `.md` to `.json`, reload persistence, and Content Engine query in a browser - [ ] Run the complete local agent evaluation suite if approved ## Verification - 217 focused tests passed across Builder, asset revisions, HTTP client, CLI/MCP, and protocol - Builder, asset-uploader, HTTP-client, protocol, CLI, and project-build typechecks passed - Builder production build passed - MCP and Content Engine documentation checks passed - `oxlint` on changed TypeScript files passed - `git diff --check` passed - Regression tests were demonstrated failing before the implementation and passing afterward A disposable E2E database and PostgREST stack bootstrapped successfully in the earlier real-path test, but UI automation remained blocked because no controllable browser was available. Fixture inputs and generated fixture outputs are not affected. Agent evaluations were not run because approval is required and was not requested. Closes #6167 |
||
|
Oleg Isonen
|
dbfb0a9849 |
fix: Preserve variable font styles (#6153)
## Summary Preserve font style metadata for variable fonts so upright and italic files from the same family produce distinct `@font-face` rules. Variable fonts with an `ital` or `slnt` axis are also registered for the matching CSS styles. ## Technical choices - Detect `normal`, `italic`, and `oblique` from the font binary, preferring OS/2 `fsSelection`, then `post.italicAngle`, then the existing subfamily parser. - Add `style` to variable-font metadata while defaulting legacy stored records to `normal`. - Keep update metadata partial and strict, but accept and persist `style` and `variationAxes` together so a `get-asset` result can be written back. - Group generated font faces by family, style, and browser-selectable variable descriptors so separate upright and italic files cannot overwrite each other. - Regenerate the runtime operation contracts from the updated schema. ## Checklist - [x] Add regressions using real upright and italic Inter variable-font binaries. - [x] Cover style detection, metadata compatibility, updates, and generated font-face selection. - [x] Verify the final stylesheet contains distinct upright and italic `@font-face` rules. - [x] Verify the public asset update contract accepts a complete variable-font metadata object. - [x] Review documentation impact and update the custom-font guide. - [x] Regenerate and verify runtime API contracts. ## Verification - `@webstudio-is/fonts`: 14 tests and typecheck passed. - `@webstudio-is/asset-uploader`: 198 tests and typecheck passed. - `@webstudio-is/sdk`: 514 tests and typecheck passed. - `@webstudio-is/project-build`: 2,094 tests and typecheck passed. - CLI: 980 tests and typecheck passed. - Builder typecheck passed. - `pnpm lint`, `pnpm check:package-boundaries`, `pnpm check:generated-api`, and `pnpm check:generated-docs` passed. - Regression tests were demonstrated failing against the previous behavior and passing with the fix. - Fixture builds regenerated successfully with no diff. Full remote fixture sync was attempted, but the deployed API rejected the source CLI contract as incompatible. - Agent evaluations were not run because they require separate explicit approval. Closes #6150 |
||
|
Oleg Isonen
|
87968f40ff |
fix: Improve MCP authoring and preview reliability (#6105)
## Summary
Fix MCP reliability failures reported in ten issues against CLI 0.288.0
and 0.289.0:
- keep every advertised array input schema client-valid, including
compact and tuple schemas
- report a text asset update as successful when its persisted revision
and bytes confirm that the server committed despite a failed response
- preserve standard package import resolution in generated React Router
previews
- upload variable WOFF2 fonts when callers provide the documented static
font metadata shape
- explain unmatched JSX as a stateless fragment syntax error
- distinguish the `animation` component namespace from value helpers and
reject unsupported CSS at-rules with an actionable validation error
- forward optional session tools such as `download-asset` through
top-level CLI shortcuts
- isolate published generated-preview dependencies from parent
workspaces
- return structured screenshot failures when React Router production
builds fail
Sensitive visual regression checks also run automatically for
organization-authored fork PRs. External fork PRs remain gated by
`safe-to-deploy`.
## Root causes and implementation
The MCP handshake compactor retained an array's type but discarded its
`items` schema. Unconstrained and tuple arrays could also be advertised
without `items`. Schema normalization now adds `items: {}` whenever an
array omits it, while focused discovery retains complete constraints.
Catalog coverage rejects every missing array item, including tuple
schemas, and the handshake remains below its 200 KB budget.
Asset content updates previously treated any failed response as a failed
mutation, even when the server had already swapped the revision.
Recovery now reads the current revision, compares the persisted bytes,
and confirms the revision remains stable across the read. It returns the
updated asset only when all evidence agrees. Otherwise it returns
`ASSET_UPDATE_COMMIT_UNCERTAIN` with guidance not to retry blindly.
Definitive 4xx responses retain their original errors.
The generated React Router Vite config replaced the client and SSR
resolver condition lists without explicitly retaining the standard
`import` condition. Both lists now preserve `import`. The regression
test generates a dynamic route, stages the built React Router SDK as a
published dependency without the local `webstudio` source condition, and
runs the generated build.
Variable fonts are detected with `variationAxes`. Merging the documented
static `style` and `weight` upload metadata into that detected variable
shape made validation fail after the binary was read. Font override
validation and shape-aware merging now live in `@webstudio-is/fonts`.
The SDK delegates to that domain function, which validates the complete
override before applying compatible fields, retains the detected
static-or-variable shape, and keeps the detected format.
Fragment parsing is stateless. Unmatched opening and closing tags now
explain that every opening element must use the same closing tag or be
self-closing, and that refreshing the session cannot repair malformed
JSX.
The `animation` binding is a component namespace, not a callable helper.
Tool schema text, evaluation errors, and the MCP manual now separate
component namespaces from value helpers. The MCP JSX `css` binding
parses the final interpolated CSS with CSS-tree and recursively rejects
selectors and unsupported at-rules such as `@keyframes`, while
preserving supported declarations and case-insensitive `@media` rules.
Keeping this validation at the external binding avoids changing internal
component templates.
CLI shortcut discovery previously omitted optional session tools even
though the runtime registered their handlers. Shortcut discovery now
includes the same optional tool categories as the CLI MCP host, while
real top-level CLI commands keep precedence.
Published generated previews could reuse a stale `node_modules` symlink
into a parent source workspace. Only development manifests containing
the local placeholder version may now reuse the CLI dependency tree.
Published manifests replace symlinks with a local install and disable
npm workspace discovery.
React Router could report a bundler failure while its build command
exited successfully outside CI, allowing screenshot startup to terminate
without a structured result. Managed production builds now set `CI=1`,
so the existing MCP error boundary receives the nonzero build failure
and prints structured JSON.
Visual regression authorization now distinguishes repository members
from external contributors. Organization members' fork PRs run on open,
synchronize, and reopen events; external fork code runs only from a
`safe-to-deploy` label event, and a later commit does not reuse that
approval.
## Implementation checklist
- [x] Preserve valid array item schemas in compact MCP discovery
- [x] Normalize tuple input schemas with an `items` fallback
- [x] Keep the MCP handshake below its 200 KB size budget
- [x] Cover the real `update-props` schema and oversized array
compaction
- [x] Reconcile failed asset-update responses against stable persisted
content
- [x] Reject recovery when a concurrent asset revision is detected
- [x] Preserve `import` resolution in generated React Router client and
SSR builds
- [x] Build a generated dynamic route against a staged published SDK
package
- [x] Regenerate and verify affected React Router fixtures
- [x] Preserve detected variable-font axes with documented upload
metadata
- [x] Add actionable unmatched-tag and unsupported-CSS validation
- [x] Correct animation namespace guidance in tool discovery and errors
- [x] Forward optional session tools through top-level CLI shortcuts
- [x] Replace parent-workspace dependency links for published previews
- [x] Disable npm workspace discovery for isolated preview installs
- [x] Make React Router build failures return nonzero in managed
previews
- [x] Run visual regression automatically for organization-authored PRs
- [x] Verify each regression test fails without its corresponding fix
- [x] Review documentation impact and update the authoritative MCP
manual plus generated outputs
- [x] Confirm no database migration or compatibility migration is
required
- [ ] Run `pnpm evaluations` after explicit approval; not run because
repository policy requires approval before agent evaluations
## Verification
- `packages/cli/src/prebuild.test.ts`: 57 tests passed
- `packages/cli/src/commands/schema.test.ts` and `mcp.test.ts`: 51 tests
passed
- `packages/project-build/src/runtime/components.test.ts` and
`mcp.test.ts`: 194 tests passed
- `packages/asset-uploader`: 196 tests passed
- focused font schema, SDK asset schema, CSS parser, and template CSS
tests passed
- `packages/http-client/src/index.test.ts`: 38 tests passed
- relevant CLI, project-build, asset-uploader, fonts, SDK, CSS data,
template, HTTP client, and React Router SDK typechecks passed
- MCP documentation tests, typecheck, generation, and Markdown source
tests passed
- React Router SDK package build passed
- full fixture link, sync, and generation pipeline passed
- affected Docker and Vercel fixture typechecks and locally bootstrapped
production builds passed
- complete CLI suite: 61 test files and 966 tests passed
- CLI typecheck passed
- preserved nested-workspace dynamic preview installed local `0.289.0`
dependencies and completed its production build
- visual-regression workflow YAML parsing and authorization truth table
passed
- repository lint passed with warnings denied
- formatting and `git diff --check` passed
## Tradeoffs and limitations
Deferred array schemas advertise unconstrained `items` in the compact
handshake to remain client-valid without duplicating detailed
constraints. Agents can retrieve the complete item schema through
focused `meta.get-more-tools` discovery.
Asset reconciliation performs three read requests only after an
uncertain 5xx, malformed response, or transport failure. This keeps the
normal successful path unchanged and avoids claiming success if the
revision changes during recovery.
For variable fonts, detected axes define the metadata shape. Static
`style` and `weight` hints are ignored because they cannot be
represented alongside `variationAxes`; compatible fields such as
`family` still override detection.
Native CSS keyframe authoring remains unsupported in fragment `css`
templates. The tool now rejects it before rendering and directs callers
to editable Webstudio Animation components instead of advertising a
nonexistent callable helper.
Published previews perform one isolated npm install when no valid local
dependency tree is cached. Local source development keeps the fast
workspace symlink path for placeholder-version manifests.
Closes #6104
Closes #6099
Closes #6098
Closes #6106
Closes #6107
Closes #6108
Closes #6114
Closes #6115
Closes #6116
Closes #6117
|
||
|
Oleg Isonen
|
d5b3f3ab80 |
fix: Improve MCP authoring reliability (#6056)
## Summary This consolidates the reported MCP reliability fixes into one delivery. - validate all inserted fragment expressions with the shared runtime rules and reject text/expression children when component content models require elements - return an actionable error for unsupported existing-resource references in JSX - preserve siblings rendered before expression children - accept unresolved CSS-variable shorthand expansion and eliminate dynamic-alt/`htmlFor` audit false positives - resolve exposed underscore MCP names during focused discovery - omit unavailable stopped-preview fields and restart stale iterative previews for external clients - bound single and responsive screenshot capture, return `SCREENSHOT_TIMEOUT`, and discard failed reusable browser sessions - dispose previews when shared MCP runs fail or stdio clients disconnect, including complete Windows process-tree cleanup - resolve generated route templates from the CLI-managed preview workspace during responsive verification - retry a browser that exits during rendered-audit startup and report sanitized exit diagnostics with a concrete browser-path recovery action - materialize complete current session data before preview generation so route order and project styles cannot collapse from missing lazy namespaces - always include CLI and Node.js version fields in generated issue reports, with explicit legacy-client fallbacks - document the updated preview and screenshot contracts and regenerate the CLI manual The current branch also retains the protections already present on `main` for dry-run isolation, CSS rewrite truncation metadata, stderr-only progress, shortcut argument forwarding, strict MCP schemas, and refreshed conflict diagnostics. ## Technical choices - MCP command orchestration stays in `mcp.ts`; preview port/retry behavior lives with `mcp-preview`, and persisted checkpoint storage lives in `mcp-checkpoint`. Each module owns its matching tests. - Expression validation lives on the shared fragment input schema, reuses `listFragmentExpressions`, and emits the same structured validation issues as prop bindings, so JSX-derived and structured fragments follow one contract across fragment, collection, and runtime UI insertion. - Element-child requirements come from component registry content models during fragment insertion; no component names are hardcoded in the JSX parser. `CollapsibleTrigger` metadata now matches its enforced `asChild` implementation. - Stopped preview status returns only `running: false`; unavailable `url`, `pid`, and `mode` fields are optional and omitted from JSON. - Explicit `preview.start` restarts a stale compatible iterative preview; managed screenshots keep their optimized refresh path. - Screenshot operations share one timeout utility. A single capture or complete responsive batch is bounded after preview readiness; lower-level CDP deadlines use the same stable error code. - Failed managed captures close and forget their browser session. A closed browser session cannot restart itself while cancellation is in progress, which prevents orphaned Chromium processes. - One ownership helper disposes one-shot and multi-project MCP hosts on success or failure. Shared `mcp run` hosts dispose in one `finally` boundary, long-running MCP hosts dispose on connection close, and Windows uses `taskkill /T /F` instead of killing only the npm wrapper. - Framework template readers receive the absolute template directory captured from the build root, so later async work cannot redirect them to the linked project's cwd. The same option drives cleanup for React Router, Remix, and Vike. - Browser startup gets one bounded retry before a rendered capture batch fails. Persistent failures report only the exit code, signal, or spawn error code plus the supported `WEBSTUDIO_BROWSER_PATH` recovery action; local paths and raw browser output are not exposed. - CLI and MCP preview share one session materializer that ensures the canonical Builder namespace set before writing `.webstudio/data.json`. Serialization rejects incomplete snapshots instead of converting missing style/instance namespaces into valid-looking empty arrays, and preview freshness records the post-materialization project version. - The CLI injects its package version and `process.versions.node`; issue formatting now keeps both fields visible even for legacy reports that omit runtime metadata. ## Checklist - [x] Add regression coverage for every changed behavior - [x] Demonstrate new regression tests fail without the corresponding fixes and pass with them - [x] Run affected package test suites and typechecks - [x] Run lint, package-boundary, and generated API checks - [x] Review fixture impact — no fixture inputs or outputs are affected - [x] Review documentation impact and update the authoritative MCP CLI manual source/generated output - [x] Run the complete local agent evaluation suite ## Verification - `@webstudio-is/project-build`: 2,090 tests passed - MCP asset query result modes are preserved across create, read, and update serialization - complete local agent evaluation suite: all 5 fixtures passed, 1,308,276 aggregate tokens - `@webstudio-is/react-sdk`: 140 tests passed - `@webstudio-is/template`: 43 tests passed - `webstudio` CLI: 956 tests passed - issue reporter: 5 targeted tests passed - MCP host disposal regression was demonstrated failing without the shared ownership boundary and passing with it - relevant package typechecks passed - `pnpm lint` passed - `pnpm check:package-boundaries` passed - `pnpm check:generated-api` passed - generated-doc tests, typecheck, sync, and regeneration passed; the aggregate clean-tree command reports the intentionally changed generated CLI manual - stalled single capture, stalled four-viewport capture, startup retry across a four-viewport batch, sanitized browser-exit diagnostics, complete preview materialization, post-refresh version tracking, closed-session restart, Windows process-tree cleanup, and preview-workspace regressions were demonstrated failing without their fixes and passing with them ## Known limitations - External `baseUrl` screenshots remain caller-owned and intentionally do not regenerate, restart, or stop that external server. Closes #6034 Closes #6035 Closes #6038 Closes #6042 Closes #6043 Closes #6044 Closes #6045 Closes #6046 Closes #6047 Closes #6048 Closes #6049 Closes #6050 Closes #6051 Closes #6052 Closes #6059 Closes #6060 |
||
|
Oleg Isonen
|
92084aeea3 |
fix: Remove MCP workflow friction (#5974)
## Summary - Make MCP discoverable from the CLI README and expose direct MCP tool calls as restart-free CLI subcommands. - Preserve required discriminator fields in compact MCP SDK schemas and surface the `set-text-content` contract. - Return instance props by default from `inspect-instance` and improve the `list-instances` input hint. - Add a Builder context-menu **Move** submenu as an accessible alternative to drag-and-drop sibling reordering. Companion generated documentation: https://github.com/webstudio-is/webstudio-community/pull/274 ## Technical choices and compatibility - Compact SDK schemas retain required fields only inside root composition branches, avoiding broad schema expansion. - `inspect-instance` now defaults to `props` only; larger styles, bindings, sources, children, and ancestors remain opt-in. - Existing move commands and shortcuts back the new menu, so command behavior stays centralized. - No migrations, fixture inputs, generated fixture bundles, or generated API contracts are affected. ## Todo - [x] Add direct CLI MCP discovery and restart guidance. - [x] Preserve discriminated-union requirements for SDK clients. - [x] Improve instance inspection and input diagnostics. - [x] Add accessible menu-based instance movement. - [x] Add regression coverage and verify fail/pass behavior. - [x] Run targeted tests, typechecks, lint, package-boundary, and generated-doc checks. - [ ] Deferred: run `pnpm evaluations` after separate explicit approval, as required by repository policy. ## Verification - `@webstudio-is/project-build`: 172 targeted tests passed. - `@webstudio-is/protocol`: 8 targeted tests passed; typecheck passed. - `@webstudio-is/builder`: 8 targeted tests passed; typecheck passed. - CLI and project-build typechecks passed. - Root lint, package-boundary check, generated-doc check, and diff checks passed. - New and modified regression tests were demonstrated failing before their fixes and passing afterward. ## Known limitations Synthetic drag-and-drop behavior is unchanged. The new menu and existing keyboard shortcuts provide the automatable and accessible reordering path. Agent evaluations were not run because explicit approval was not provided. |
||
|
Oleg Isonen
|
f335340a54 |
fix: Correct content database sizing and Markdown references (#5972)
## Summary - Report only serialized content-database bytes that are actually included in the published bundle. - Replace the ambiguous `markdown-body` query mode with the strict `markdown-body-ref` mode. - Require every selected Markdown body to have a document-graph node; compilation now fails instead of embedding source bytes as a fallback. - Apply the same Markdown eligibility rule during compilation and runtime query execution, excluding non-Markdown files from reference-only databases and results. - Highlight database sizes in red only when documents were omitted because of the size limit. - Refresh Assets diagnostics after an Assets resource edit is confirmed by the server. - Update CLI/MCP guidance, examples, schemas, generated contracts, and fixtures to use the reference-only mode. ## Root cause Dynamic Markdown-body queries conservatively treated unrelated files, including remote images, as possible hydration candidates. The old mode could also fall back to embedding a Markdown body when graph resolution was unavailable, despite its intended storage-backed behavior. Compilation and runtime selection did not previously share the same content-mode eligibility check. Diagnostics displayed the unbounded pre-limit estimate instead of the serialized bundled size, did not carry the omission reason into query-preview diagnostics, and an early cached preview could leave the merged measurement stale until reload. ## Impact Content database diagnostics and publish warnings now reflect bundled database bytes rather than remote source-file sizes. Saved query filters trigger a post-sync diagnostics refresh. Markdown detail queries must use `content.mode: "markdown-body-ref"`; the old `markdown-body` value is rejected, missing graph support is an explicit compilation error, and non-Markdown assets cannot be selected or counted by a reference-only query. Ref #5966 ## Verification - Content engine: 39 files / 261 tests passed. - Asset uploader: 22 files / 183 tests passed. - SDK: 29 files / 505 tests passed. - Protocol: 9 files / 86 tests passed. - Project build: 82 files / 2,055 tests passed. - CLI: full suite previously passed (57 files / 933 tests); final prebuild verification passed 48 tests. - Focused Builder diagnostics tests: 4 files / 10 tests passed. - Typechecks passed for Builder, content engine, asset uploader, SDK, protocol, project build, and CLI. - Generated API contracts and docs are current. - Fixture build passed and is reproducible after the final review changes. - Formatting, lint, package-boundary checks, and diff checks passed. - New strict-reference, runtime-selection, compilation-selection, and size-highlight regression tests were demonstrated failing without their implementations and passing with them. - Agent evaluations were not run because this follow-up did not receive separate evaluation approval. |
||
|
Oleg Isonen
|
56a22fde8a |
feat: Add storage-backed document graph resolution (#5966)
## Summary Introduce a document graph architecture for Assets resources so structured documents can reference one another while keeping full document payloads on CDN-backed storage. The content database will retain the searchable index and dependency graph. At runtime, the server will select a root document, determine its complete dependency set before performing I/O, fetch each unique document concurrently, and assemble the final result server-side. ## Implementation todos ### Document graph foundation - [x] Add normalized internal reference contracts for whole documents, JSON paths, Markdown body, and Markdown frontmatter. - [x] Add an immutable graph abstraction with deterministic ordering, dependency traversal, revision validation, and cycle detection. - [x] Add a storage-independent resolver with dependency closure, bounded parallel loading, shared-content deduplication, representation selection, assembly hooks, cancellation, and structured errors. - [x] Add focused unit coverage for the isolated document-graph abstraction. - [x] Define and validate the external remote-document reference syntax. - [x] Add a pure graph compiler that resolves canonical document URLs into IDs and revisions and validates missing, duplicate, and cyclic dependencies. ### Build-time graph - [x] Add a storage-independent source-catalog compiler for bounded parallel reference discovery and deterministic graph compilation. - [x] Add a versioned deterministic document graph artifact with canonical serialization and integrity verification. - [x] Discover and normalize statically declared references while Assets resources are processed. - [x] Resolve document identities and revisions and report missing, stale, or cyclic dependencies with document context. - [x] Persist document nodes and dependency edges alongside searchable resource metadata. - [x] Include persisted graph data in artifact integrity and size validation. ### Document representations - [x] Add a standalone JSON document adapter for exact reference markers, compiler occurrences, JSON path selection, and immutable assembly. - [x] Add a standalone Markdown document adapter for bounded decoding, frontmatter references, immutable assembly, and source, body, and frontmatter selection. - [x] Parse fetched JSON and Markdown documents through format-specific adapters. - [x] Select complete documents, JSON fragments, Markdown body, and Markdown frontmatter through the shared reference model. - [x] Assemble nested and shared references into the final server-side value. ### Runtime resolution - [x] Enforce reachable-document, per-document byte, and aggregate-byte limits before parsing and assembly. - [x] Add a revision-keyed bounded source-cache wrapper with replayable bytes, stale-entry fallback, and in-flight deduplication. - [x] Add a storage-neutral loaded-source contract that rejects revision mismatches before parsing. - [x] Resolve graph-backed property paths before filtering, sorting, pagination, and projection; skip graph loading for unrelated queries. - [x] Add an injected HTTP/CDN document loader with revision metadata validation. - [x] Apply runtime payload limits, bounded concurrency, caching, and structured failure mapping. - [x] Ensure shared documents are fetched and parsed once per resolution. ### Product integration - [x] Integrate equivalent behavior in local preview, SSR, SSG, and published runtime paths. - [x] Preserve existing embedded Assets resource behavior without requiring migration. - [x] Add observability for graph loading, document fetching, caching, and resolution failures. - [x] Document the public reference and resolution behavior. ### Review hardening - [x] Enforce graph, artifact, cache, and published-storage identity consistency. - [x] Compile only graph sources reachable from selected build entries. - [x] Resolve graph-backed fields before query evaluation and avoid unnecessary graph loading. - [x] Consolidate duplicated graph schemas, utilities, and document content-type detection. - [x] Correct runtime documentation and use structural evaluation validation. ### Verification - [x] Keep a deterministic high-impact agent evaluation for Markdown-to-JSON references across the dynamic blog overview and article routes. - [x] Add a separate discoverability evaluation that must learn the reference workflow from MCP guidance without prompt-supplied syntax, queries, bindings, or call order. - [ ] Re-run the complete local agent evaluation suite after the discoverability fixture change and compare every fixture with its accepted baseline. - [x] Add integration coverage for CDN-backed hydration, parallel fetching, caching, and shared dependencies. - [x] Add integration coverage for missing references, cycles, revision changes, limits, and cancellation. - [x] Verify equivalent behavior across preview, SSR, SSG, and published deployments. - [x] Verify that existing embedded Assets resources retain their current behavior. ## Intended outcome - Storage-backed documents do not require their complete payloads to be embedded in the content database. - The server knows the complete statically discoverable dependency set before fetching document content. - Shared dependencies are fetched once per resolution and independent resources can be fetched in parallel. - JSON fragments and Markdown body/frontmatter representations use the same document-resolution system. - Reference and payload revisions cannot be mixed silently. - Existing Assets resources continue to work without migration. ## Verification results - `@webstudio-is/content-engine`: 37 files / 244 tests passed; typecheck passed. - `@webstudio-is/asset-uploader`: 22 files / 182 tests passed; typecheck passed. - CLI: 57 files / 925 tests passed, including generated SSR and SSG build/prerender coverage; typecheck passed. - Repository-wide `pnpm oxfmt --check` passed across 2,757 files. - Package-boundary validation passed. - Previous `pnpm evaluations` run: all 4 original high-impact fixtures passed. The complete 5-fixture rerun is pending explicit approval after adding the discoverability fixture. |
||
|
Oleg Isonen
|
efbc3098c4 |
feat: add queryable Assets resources (#5946)
Closes #5919 Ref #5920 ## Outcome Extend the existing Assets system resource with optional queries for Markdown, JSON, and standard file metadata. Unconfigured Assets resources retain the existing fetch-all behavior and activate none of the content-query infrastructure. Files are the authoring model. A standalone content engine converts an immutable snapshot of only the files required by configured queries into one bounded, deterministic runtime database. Builder preview, SSR, and SSG use the same query semantics without requiring a database service for a published site. ## Final architecture ```text Markdown / JSON files │ ▼ ContentSource adapter (FS or Webstudio Assets) │ immutable snapshot ▼ ContentCompilationPlan (union of configured query requirements) │ ▼ ContentCompiler (parse, normalize, discover fields, enforce budget) │ ▼ ContentArtifact (bounded, deterministic, immutable JSON) │ ▼ ContentDatabase (filter, OR groups, sort, paginate, project, read content) │ ├── Builder preview ├── SSR runtime └── SSG build only ``` The authoring/storage layer owns mutable files. The content engine consumes immutable snapshots and produces immutable, disposable artifacts. Query authoring is a separate OpenAPI-driven path: ```text Assets REST OpenAPI operation + standard request JSON Schema + dynamic field paths │ ▼ Generic OpenAPI adapter │ resolved schemas, controls, paths, and transport mapping ▼ Generic query-builder core and React renderer │ arbitrary provider-defined value ▼ Assets REST request adapter │ ▼ { query: value } ``` The selected standard OpenAPI operation is the only source of query UI semantics. Request parameters and JSON Schema describe the editable data; the generic renderer derives controls and owns layout. ### Activation and compilation - The compiler does not index every asset. It combines all reachable configured Assets-resource queries into one `ContentCompilationPlan`. - Static constraints such as folder, path, and file type select candidate files. Dynamic values such as a route-bound slug cannot reduce the build-time candidate set. - The plan contains the union of fields used by filtering, sorting, routing, projection, excerpts, and requested full content. - Documents shared by multiple resources are stored once in one deployment-wide runtime database. - Markdown/frontmatter, JSON properties, excerpts, and full bodies are prepared only when required. - No configured content query means no parsing, artifact, or shipped runtime dependency. Fresh generated apps need no content-runtime cleanup; incremental generation removes stale integration if a previously configured query is deleted. ### Runtime database limit - `CONTENT_DATABASE_MAX_BYTES` limits the complete serialized runtime database, not source files, images, or unrelated assets. - The default is `500 * 1024` bytes and is presented as 500 KiB. - Counted data includes indexed metadata, frontmatter/JSON properties, excerpts, the field catalog, serialization overhead, and full Markdown bodies only when requested. - Prepublish validation builds and serializes the exact database that will be attached to the immutable build, so it knows the unbounded size, bounded size, included documents, omitted documents, and affected resources. - Overflow never stores partial documents. Candidates are ordered by `createdAt` descending, canonical path ascending, and stable document ID ascending. Missing dates follow dated documents. A document that does not fit is skipped while smaller later documents are still considered. - Exceeding the limit warns but does not block publishing. - Builder preview can report exact bounded versus unbounded results when dynamic expressions have concrete preview values. Prepublish warnings remain conservative for future route values. ### Snapshot consistency The source contract guarantees a consistent compilation, while adapters provide the mechanism: - Webstudio Assets captures asset metadata and immutable content references. - Filesystem sources capture file identity and verify files did not change before accepting the artifact. - A changed source retries once, then fails clearly instead of producing a mixed-revision database. - This requires no universal snapshot service, source-tree copy, or new versioned-asset infrastructure. ### Package ownership - `@webstudio-is/content-engine`: runtime artifact schema, query execution, field catalog, compact runtime statistics. - `@webstudio-is/content-engine/compiler`: source snapshots, compilation plans, Markdown/JSON parsing, deterministic compilation, byte accounting, and compiler diagnostics. This entry exists to keep Remark/YAML/future MDX dependencies out of runtime bundles. - `@webstudio-is/query-builder`: resolves standard OpenAPI operation schemas into an internal provider-neutral editor model, validation, mutations, and source serialization. It owns no Assets, provider-authored UI contract, or fixed REST-query semantics. - `@webstudio-is/query-builder/react`: container-neutral schema renderer whose layout is renderer policy. It remains separate from the content engine. - SDK: Webstudio Assets/Resource domain models and mapping Webstudio expressions into generic content requirements. - Protocol: external Assets REST/OpenAPI transport contracts. - Asset uploader/repository: asset and folder mutations, generic metadata, immutable byte references, and storage access. Its hosted adapter supplies Webstudio Assets snapshots and parsed-metadata cache persistence to the content engine; parsing, artifact creation, and query semantics remain engine-owned. - Builder and generated-app integrations decide when to compile, cache, embed, load, or discard a database. ### Runtime lifecycle - Builder caches immutable `ContentDatabase` instances by source revision plus compilation-plan hash and coalesces concurrent compilation. Matching operations reuse the same instance; field discovery uses a dedicated minimal all-properties plan that excludes excerpts and bodies. - SSR attaches the validated bounded database to the build and includes the small runtime only when required. - SSG uses the database only while generating pages and omits it and the runtime from final output unless a future runtime-query feature requires them. - Local development compiles from the filesystem and uses identical query semantics. - Published pages query their local artifact; they do not call Builder REST endpoints. - Published artifacts need no external index store, invalidation, garbage collection, SaaS content proxy, new bucket, or new environment variable. ## Implemented foundation These existing parts remain useful and should be moved behind the final ownership boundaries rather than reimplemented. - [x] Upgrade the existing Assets system resource in place and preserve unconfigured fetch-all behavior. - [x] Implement typed filtering, nested boolean groups, deterministic sorting, pagination, projections, dynamic values, and content hydration. - [x] Add Markdown frontmatter, excerpt, JSON-property, and field-catalog extraction. - [x] Add reusable query-control implementations and a separately imported React UI as a foundation for the final OpenAPI-driven contract. - [x] Add Assets query authoring and dynamic field discovery as existing behavior to migrate into the Assets OpenAPI operation. - [x] Add synchronized visual and editable expression authoring. - [x] Unify Assets REST operations under `/rest/assets/*`. - [x] Define authenticated collection, item, content, upload, folder, query, field-catalog, and OpenAPI operations. - [x] Support Builder cookie authentication and public API token authentication. - [x] Add Builder preview, MCP/CLI operations, SSR execution, SSG path generation, and conditional activation. - [x] Add revision-keyed parsed metadata persistence for the hosted Builder. - [x] Add focused query, API, authorization, Builder, CLI, SSR, SSG, migration, and fixture coverage. ## Remaining implementation plan ### 1. Establish the content engine - [x] Rename the generic `@webstudio-is/asset-resource` implementation to `@webstudio-is/content-engine`. - [x] Move existing behavior instead of creating a parallel implementation. - [x] Keep the primary entry runtime-only and add only the compiler entry needed for bundle separation. - [x] Remove the old package after all consumers migrate. - [x] Ensure the engine has no Builder, SDK, asset-uploader, PostgreSQL, R2, or deployment-target dependency. ### 2. Define source snapshots and compilation plans - [x] Introduce `ContentSource.openSnapshot()` and a revision-bearing `ContentSourceSnapshot`. - [x] Include canonical path, stable document ID, content type, immutable content reference, size, and optional `createdAt`. - [x] Implement Webstudio Assets and filesystem snapshot guarantees using their adapter-specific mechanisms. - [x] Introduce `ContentCompilationPlan` for the union of reachable configured resource requirements. - [x] Separate statically selectable candidates from dynamic runtime predicates. - [x] Add mutation-during-compilation tests for replacement, deletion, folder movement, and filesystem changes. ### 3. Separate storage from content processing - [x] Keep upload, replacement, folders, byte storage, revisions, and generic metadata in the Assets repository/uploader. - [x] Move parsing, excerpts, field discovery, query execution, and artifact creation into the content engine. - [x] Remove content-index coordination from asset mutation business logic. - [x] Preserve original source `createdAt` across content replacement and expose it as optional queryable metadata. - [x] Keep hosted parsed-document persistence behind the content compiler's optional cache interface. ### 4. Build deterministic compilation - [x] Move the existing Markdown, frontmatter, JSON, canonicalization, hashing, document-building, and tests into the compiler. - [x] Keep frontmatter behind a dedicated module and continue using the existing Remark/Micromark pipeline. - [x] Parse only candidate files and only the tiers required by the compilation plan. - [x] Deduplicate documents and fields required by multiple resources. - [x] Define a versioned, deterministic `ContentArtifact` with integrity information. - [x] Keep verbose parse and omission diagnostics outside the runtime artifact. - [x] Reject incompatible artifact versions cleanly. ### 5. Enforce the deployment-wide database budget - [x] Replace the old asset-index limit with `CONTENT_DATABASE_MAX_BYTES`, defaulting to 500 KiB. - [x] Have applications read configuration and pass the byte limit into the compiler. - [x] Measure the exact serialized database, including selected bodies and overhead. - [x] Implement deterministic skip-and-continue packing of complete documents. - [x] Return compiler diagnostics with unbounded/bounded bytes, included/omitted counts, omitted identities, largest contributors, and affected resources. - [x] Expose only compact artifact statistics at runtime through `ContentDatabase.getStats()`. - [x] Add exact boundary, deterministic ordering, deduplication, and oversized-document tests. ### 6. Finalize the runtime database - [x] Define `ContentDatabase` over an immutable artifact. - [x] Preserve filtering, nested AND/OR groups, sorting, offset/limit, projection, field references, and content retrieval. - [x] Expose `getFieldCatalog()` and compact `getStats()`. - [x] Keep Markdown/YAML/compiler dependencies out of the runtime entry. - [x] Add execution guards and maximum-size runtime benchmarks. ### 7. Make query authoring OpenAPI-driven The selected standard OpenAPI operation is the complete query-authoring contract. The query builder resolves parameters and request-body JSON Schema, derives its internal editor model, and applies renderer-owned layout policy. It must not embed or fetch a second capabilities/UI language through `x-webstudio-query` or another side channel. - [x] Remove `x-webstudio-query` and all vendor query-definition metadata from the Assets OpenAPI operation. - [x] Resolve local OpenAPI references and derive the editable request value, defaults, validation, labels, choices, arrays, objects, and unions solely from standard operation parameters and request-body JSON Schema. - [x] Represent dynamic Assets field paths with standard JSON Schema vocabulary in the `queryAssets` request schema. - [x] Keep the normalized editor representation internal to query-builder; do not expose another provider-authored capabilities contract. - [x] Remove provider-controlled layout metadata such as `group`; make compact scalar packing and all other layout renderer policy. - [x] Keep Webstudio expression binding as a consumer/editor concern layered over schema values, not part of the OpenAPI contract. - [x] Preserve the existing Assets filter, sort, output selection, full-content/range selection, limit/offset, editable source, and configured-field behavior. - [x] Keep the Assets resource adapter responsible for serializing the edited query into its request body. - [x] Remove the obsolete Assets query-capabilities endpoint, URL, and UI-side capability parsing. - [x] Add a non-Assets OpenAPI proof with differently named parameters and no vendor metadata. - [x] Verify the generic packages contain no Assets imports, fixed Assets property contract, or provider-authored layout model. - [x] Update focused core, React, SDK, protocol/OpenAPI, Builder, and source round-trip tests. ### 8. Manage ContentDatabase lifecycle in Builder - [x] Cache databases by source revision plus normalized compilation-plan hash. - [x] Coalesce concurrent compilation requests. - [x] Reuse cached database instances across field discovery, autocomplete, query preview, and response preview when their normalized compilation plans match. - [x] Replace it when source revision or query requirements change. - [x] Preserve the optional revision-keyed hosted parsed-document cache without making it a source of truth. - [x] Show exact preview warnings when concrete values reveal missing bounded results. ### 9. Materialize and verify immutable builds - [x] Compile and serialize the exact deployment-wide database during prepublish validation. - [x] Attach that validated bounded database to the immutable build without rebuilding it from later asset state. - [x] Show non-blocking prepublish warnings with database bytes, omissions, and potentially affected resources. - [x] Avoid claiming exact future route matches when dynamic values are unknown. - [x] Include the database/runtime only when a reachable configured query requires them. - [x] Ensure detail resources requesting Markdown content include all statically eligible bodies needed by dynamic routes. ### 10. Integrate all targets - [x] SSR: load the build's bounded database and conditionally include the small runtime. - [x] SSG: query the same database during generation, then omit database/runtime output when no runtime query is needed. - [x] Local/CLI: compile from the filesystem source with the same semantics. - [x] Preserve overview pagination, alternative route fields, draft-as-user-data behavior, and full article rendering. - [x] Verify ordinary Assets and fresh query-free sites perform no parser, artifact, content-runtime dependency install, or cleanup work. ### 11. Correct API and package ownership - [x] Move external Assets REST/OpenAPI contracts to `@webstudio-is/protocol`. - [x] Keep Webstudio domain models and Assets system-resource configuration in the SDK. - [x] Keep both cookie and token authentication behavior on the unified Assets REST surface. - [x] Remove pass-through route wrappers/reexports that add no transport behavior. - [x] Keep published runtime queries local to the artifact rather than routing them through Builder REST. ### 12. Remove superseded deployment infrastructure - [x] Remove external deployment-index upload, fetch, invalidation, cleanup, and garbage-collection code. - [x] Remove the SaaS asset-origin adapter requirement; close webstudio-is/webstudio-saas#652 as superseded because it has no independent purpose. - [x] Remove index-bucket compatibility configuration and avoid adding Infisical variables. - [x] Remove obsolete external-index paths, package dependencies, routes, scripts, and tests. - [x] Consolidate unreleased intermediate migrations where safe and retain only schema required by the final hosted cache/API design. ### 13. Verification and release readiness - [x] Move and preserve existing tests while changing ownership; add new tests only for changed contracts and behavior. - [x] Run content-engine compiler/runtime tests, query-builder tests, SDK/protocol/API tests, Builder tests, CLI tests, and affected typechecks. - [x] Run formatting, generated-file idempotence, fixtures, clean Supabase-compatible migrations, E2E shards, and complete agent evaluations. - [x] Verify SSR and SSG conditional inclusion and query-free generated package manifests. - [x] Run the five-article MCP blog workflow: authoring, field discovery, overview, dynamic detail routes, full Markdown rendering, publishing, and published-site verification. - [x] Run a maximum-budget database test and a mutation-during-compilation test. - [x] Perform a final architecture, security, API compatibility, migration, bundle, correctness, and DRY review. - [x] Reconcile every checkbox and replace the verification section with results from the final architecture before release. ## Future MDX work MDX is intentionally separate from this release. - The compiler contract should allow a future build-only MDX parser. - MDX compilation must not enter the published runtime. - Embedded JSX must use an explicit compiled representation rather than runtime source evaluation. - Webstudio components, tokens, styles, and component allowlists remain Webstudio integration concerns. - The future MDX work should reuse `ContentArtifact` and `ContentDatabase` where practical. ## Superseded design The following earlier design choices are no longer the target architecture: - GROQ or GraphQL as the Assets query language. - One external or long-lived index object per resource/query. - A deployment index fetched by the Worker from SaaS/CDN infrastructure. - Lazy production reads of source Markdown bodies. - A separate index bucket, asset origin, invalidation service, or garbage collector. - A 4 MiB published index limit. - Rebuilding an old build's database from the current mutable asset set. - Bundling query infrastructure into projects without a configured content query. ## Verification baseline The embedded runtime-database architecture previously passed the checks below. The OpenAPI-driven query-authoring change additionally passed: - Query-builder core: 4 files / 19 tests. - Query-builder React: 1 file / 7 tests. - SDK Assets source/config: 1 file / 7 tests. - Protocol Assets OpenAPI: 1 file / 9 tests, including independent OpenAPI validation. - Builder Assets OpenAPI description: 1 file / 5 tests. - Query-builder, React, SDK, protocol, and Builder typechecks. - Package-boundary validation and formatting of every affected file. Local verification baseline: - Content engine: 15 files / 104 tests. - Asset uploader: 24 files / 175 tests. - Builder: 204 files / 2,423 tests. - CLI: 54 files / 866 tests. - Complete workspace typecheck, formatting, lint, package boundaries, generated fixtures, and fixture builds. - The complete migration history applied successfully to a fresh Supabase-compatible database, including every migration after the consolidated Assets migration. - Builder production build and full local E2E suite passed. - Complete agent evaluations passed: authenticated page (170,859 ms), design input (207,101 ms), and font assets (67,050 ms). - Final review additionally verified correct REST edit/view/build permits, cookie and public-token read paths, immutable snapshot retries, exact database budgeting, query-free dependency omission, and SSR/SSG conditional runtime inclusion. Final staging verification passed against the embedded runtime-database architecture: MCP uploaded and updated five Markdown articles, discovered dynamic frontmatter fields, previewed a five-item sorted overview and one full-body slug query, updated both Assets resources, found zero binding issues, rendered `/blog` and `/blog/dynamic-routes` at HTTP 200, published build version 266, and verified both published routes at HTTP 200 with the expected article metadata and Markdown body. The workflow also found and fixed offset-form asset timestamps failing canonical datetime validation and made MCP `preview.start({ port: 0 })` correctly select an available local port. ## Release constraints - No new SaaS content service, bucket, or environment variable. - No published-site database dependency. - No parser/compiler dependency in query-free sites or runtime-only bundles. - The validated content database is immutable build output. - Publishing warnings do not block deployment. - The selected standard OpenAPI operation and its request JSON Schema are the only source of query UI semantics. |
||
|
Oleg Isonen
|
7bc9942774 |
perf: Improve MCP evaluation reliability and efficiency (#5959)
## Summary - capture bounded MCP failure codes, validation issue paths, per-tool latency, mutation timing, uncached input, catalog size, and per-call/per-tool response bytes in every evaluation result - require every raw fixture outcome to have zero failed calls and zero retries, even when the workflow later recovers - update baselines only when all raw outcomes and all regression comparisons pass - refuse to replace an accepted aggregate token baseline with a noisier, higher-token sample - regression-gate aggregate response bytes alongside tokens, calls, focused reads, mutations, errors, and retries - compact repeated SDK schemas while preserving required arguments, destructive safety controls, and ordinary optional fields for common goal-guided tools - omit low-frequency optional argument names until focused discovery, reducing the initial tool handshake without removing callable tools - stop echoing the caller's goal from `meta.guide` responses - make evaluations copy their canonical objective verbatim into `meta.guide`, preventing vague paraphrases from selecting a generic workflow - normalize grouped `styles` and `declarations` maps into canonical flat `update-styles` operations before validation - normalize relative `create-page` paths through the platform URL parser - bundle the four bounded authentication discovery reads into one `inspect-auth-context` call - bound large MCP style-key mutation results while preserving a representative sample and exact total count - combine fragment insertion with immediate persisted binding verification to remove a sequential model turn ## Root causes and impact Recovered MCP failures could previously be reported as successful raw outcomes, and baseline files could be written before the complete comparison result was known. Raw outcomes now fail on any MCP error or retry, and accepted baselines are written only after every fixture and comparison passes. Rejected samples exposed two concrete error sources. A vague authentication objective selected a generic workflow and produced four validation failures, 20 calls, and 1.07M tokens; evaluations now pass the canonical objective verbatim. A relative `create-page.path` caused an `INVALID_INPUT` failure and retry; the adapter now canonicalizes it before validation. The deterministic catalog in the accepted baseline measures **50,864 bytes**, down from **53,317 bytes** (−4.60%). Input schemas fell from **31,190 to 28,737 bytes** (−7.86%). The two composite workflows add two conditionally exposed tools and 360 catalog bytes (about +0.71%) when their backing operations are available. Broad experiments that removed focused descriptions or all goal-specific examples made guide responses smaller but increased token use or caused retries. Those variants were rejected. The current implementation compacts only recognized goal guides, while retaining the detailed `update-styles` schema and the `upload-assets` batch example proven necessary by the evaluations. ## Workflow-churn iteration Authenticated discovery uses one model-visible `inspect-auth-context` call. Fragment authoring can now use `insert-fragment-verified`, which commits the insertion and immediately runs persisted binding verification for up to 200 bindings. This removes the following sequential verification turn without changing mutation or verification counts. Current deterministic workflow shape: - authenticated workflow calls: **14 → 10** (−28.57%) - design workflow calls: **17 → 16** (−5.88%) - aggregate workflow calls: **39 → 34** (−12.82%) - failures: **0** - retries: **0** - authenticated MCP response bytes: **52,524 → 46,900** (−10.71%) - authenticated guide response: **26,782 → 24,358 bytes** (−9.05%) One complete run measured authenticated tokens at **348,225** (−11.44%) and first verification at **38.3s** (−30.14%); design first verification moved to **28.1s** (−27.27%). The final exact-code run also passed with **10/16/8** calls and zero failures/retries, measuring authenticated tokens at **377,802** (−3.92%) and design first verification at **34.5s** (−10.67%). The composite preserves post-commit safety: if verification infrastructure fails after insertion, the result remains `committed:true`, returns `POST_COMMIT_VERIFICATION_FAILED`, and explicitly instructs the agent not to repeat the insertion. A guarded baseline-update run was rejected because stochastic aggregate tokens did not beat the accepted result, so no result artifacts were rewritten. ## Style response compaction Large style mutations previously returned every generated style key. In the design fixture, that single successful response was **19,121 bytes** and remained in context for all later turns. MCP responses now retain the first 20 keys plus `styleKeyCount` and `styleKeysTruncated`; the typed runtime operation remains unchanged. Focused evaluation results: - `update-styles` response: **19,121 → 2,815 bytes** (−85.28%) - design-workflow MCP responses: **70,941 → 54,633 bytes** (−22.99%) - calls: **17 → 17** - failures/retries: **0 / 0** - tokens: **498,764 → 505,136** (+1.28%, treated as noise and not accepted as a new baseline) The subsequent complete `pnpm evaluations` run passed all three fixtures with exact **11/17/8** call sequences and zero failures/retries. The accepted lower-token baseline remains unchanged. ## Goal-guide context compaction Recognized goal guides no longer repeat every selected tool's description and examples. Current Codex clients defer MCP tools and load each named tool's exact contract before invocation, so the guide keeps the goal-specific workflow while returning compact tool references. Free-form guide searches remain fully descriptive. Two complex inputs remain intentionally hydrated: `update-styles` keeps its detailed schema, and `upload-assets` keeps one batch example. Removing the upload example caused five failed attempts in a complete run; restoring only that example returned the font workflow to its exact 8-call, zero-error path. Final complete-suite results: - calls: **10 / 16 / 8** (**34 aggregate**) - failures/retries: **0 / 0** - goal-guide payload: **58,082 → 37,718 bytes** (−35.06%) - authenticated MCP responses: **46,900 → 38,268 bytes** (−18.41%) - design MCP responses: **51,854 → 44,008 bytes** (−15.13%) - aggregate MCP responses versus the accepted baseline: **147,335 → 101,922 bytes** (−30.82%) - aggregate tokens: **1,146,951**, only **+0.33%** above the stricter accepted 1,143,217-token baseline The complete suite and every per-fixture comparison passed. The guarded updater did not replace the accepted artifacts because the aggregate token sample did not beat the existing baseline. ## Responsive verification batching The MCP now exposes `screenshot.responsive` when reusable page capture is available. It accepts one generated route plus 1–8 viewports, refreshes the session preview once, reuses one browser target, and returns bounded evidence for every viewport. Auth and design evaluation workflows use one responsive capture instead of a separate preview start and two screenshot turns. Automatic screenshots now allocate an available preview port and reuse an already-running managed preview port. This prevents an unrelated service on the historical 5173 default from producing an MCP failure. Explicit ports, local previews, and external URLs remain unchanged. The evaluator records the bounded viewport array, counts the batch as verification and visual evidence, and compares its latency, response size, calls, errors, retries, and token consumption on every run. The new regression tests were demonstrated fail-then-pass. Guarded accepted results: - aggregate tokens: **1,143,217 → 990,190** (−13.39%) - aggregate calls: **39 → 30** (−23.08%) - aggregate MCP response bytes: **147,335 → 98,470** (−33.17%) - failures/retries: **0 / 0** - authenticated calls: **14 → 8** (−42.86%) - design calls: **17 → 14** (−17.65%) ## Design and font workflow batching Design discovery now uses one conditionally exposed `inspect-design-context` call to return bounded pages, breakpoints, design tokens, assets, and variables instead of five model-visible reads. Font persistence now uses one `verify-font-assets` call to refresh the asset namespace and return every changed font asset instead of a refresh plus separate verification reads. The evaluation tasks and validators require these composite paths. Both tools retain the underlying operation limits, namespace metadata, persisted-result checks, and zero-error/retry gates. Review also confirmed that SDK schemas must keep `additionalProperties:false` and that non-default MCP safety hints must remain explicit; those protocol constraints were preserved rather than traded for catalog-token savings. Guarded accepted improvements versus the previous responsive-batching baseline: - aggregate tokens: **990,190 → 914,760** (−7.62%) - aggregate calls: **30 → 24** (−20.00%) - aggregate focused reads: **8 → 2** (−75.00%) - aggregate MCP response bytes: **98,470 → 95,614** (−2.90%) - design calls: **14 → 10** (−28.57%) - font calls: **8 → 6** (−25.00%) - failures/retries: **0 / 0** The strict SDK catalog is **51,362 bytes** with **28,964 input-schema bytes**. The two new conditional workflow tools add 355 catalog bytes (+0.70%) and 130 schema bytes (+0.45%) while removing six model-visible workflow turns. ## Terminal responsive verification Auth and design now finish with one conditionally exposed `verify-page-responsive` call. It captures every requested viewport in one browser session and immediately runs the static page audit, removing one sequential model turn per workflow while retaining both kinds of evaluator evidence. The tool inherits the audit operation's permit and read namespaces; strict SDK schemas and MCP safety annotations remain unchanged. Latest complete comparison against the accepted baseline: - aggregate tokens: **914,760 → 869,274** (−4.97%) - aggregate calls: **24 → 22** (−8.33%) - aggregate MCP response bytes: **95,614 → 95,053** (−0.59%) - authenticated calls: **8 → 7** - design calls: **10 → 9** - font calls: **6 → 6** - failures/retries: **0 / 0** - authenticated focused run: **289,078 tokens** (−10.06%) and first verification **55,552 ms** (−17.90%) - design focused run: p95 MCP latency **9,642 ms** (−10.32%); token and first-verification variance increased in that individual sample, so only the complete comparison is used for the aggregate conclusion The complete comparison passed all fixture and aggregate regression gates. Two subsequent guarded `--update-baselines` attempts could not start any fixture because the external Codex runner had reached its usage limit; both exited before the first MCP call and wrote no accepted result files. Therefore the tracked accepted baseline below remains **914,760 tokens / 24 calls**, while **869,274 tokens / 22 calls** is the latest successful complete comparison. Regression coverage was demonstrated fail then pass. After this iteration, project-build MCP tests passed **105/105**, high-impact evaluation tests passed **41/41**, and project/CLI typechecks, lint, package boundaries, generated docs, and diff checks all passed. ## Accepted evaluation baseline | Fixture | Failed | Retries | Calls | Mutations | Tokens | Uncached input | Response bytes | First mutation | First verification | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | authenticated-page-v1 | 0 | 0 | 8 | 4 | 321,396 | 44,353 | 37,536 | 39,076 ms | 67,660 ms | | design-input-v1 | 0 | 0 | 10 | 6 | 330,580 | 54,441 | 40,036 | 30,419 ms | 36,551 ms | | font-assets-v1 | 0 | 0 | 6 | 3 | 262,784 | 20,704 | 18,042 | 37,609 ms | 47,073 ms | Accepted aggregate: **914,760 tokens**, **24 calls**, **2 focused reads**, and **95,614 MCP response bytes**, with zero failures and zero retries. Compared with the previous accepted baseline, tokens fell **7.62%**, calls fell **20.00%**, focused reads fell **75.00%**, and response bytes fell **2.90%**. ## Verification - [x] Complete `pnpm evaluations -- --update-baselines` - [x] Independent complete `pnpm evaluations` after the composite workflow change - [x] Worse-token and error-containing samples rejected without replacing accepted baselines - [x] Added/changed regression tests demonstrated fail then pass - [x] Project-build MCP tests: 105 passed - [x] Complete high-impact evaluation tests: 41 passed - [x] Project-build and CLI typechecks - [x] `pnpm lint` - [x] `pnpm check:package-boundaries` - [x] `pnpm check:generated-docs` - [x] `git diff --check` - [x] Earlier workspace tests, including 2,356 builder tests - [x] Earlier generated API checks - [ ] Fixture sync/build: sync could not reach `main.development.webstudio.is` and timed out; no local test, evaluation, or code check failed ## Implementation checklist - [x] Error and retry rate enforced at zero for raw and compared outcomes - [x] Relative page paths normalized to prevent a known validation failure and retry - [x] Baseline replacement blocked unless every outcome and comparison passes - [x] Aggregate token baselines cannot ratchet upward - [x] Workflow churn bounded and regression-gated - [x] Authentication discovery collapsed from four model-visible reads to one - [x] Fragment insertion and persisted binding verification combined safely - [x] Large style-key mutation results bounded at the MCP boundary - [x] Token consumption captured, compared, and regression-gated - [x] Response bytes captured per call and tool, summarized, and aggregate regression-gated - [x] Per-tool latency, mutation timing, catalog size, and validation paths captured - [x] Repeated catalog/schema and focused-guide context reduced - [x] Canonical goal selection stabilized for evaluations - [x] Grouped style maps normalized before validation - [x] Responsive verification batches desktop and mobile evidence in one call - [x] Automatic screenshot previews avoid occupied default ports - [x] Responsive viewport evidence is captured and regression-gated - [x] Accepted evaluation artifacts preserve the best measured token baseline |
||
|
Oleg Isonen
|
e80ca3b54c |
fix: Improve MCP authoring and screenshot reliability (#5957)
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 |
||
|
Oleg Isonen
|
c3c0e32204 |
fix: make MCP font assets editable and deterministic (#5944)
Closes #5943 Closes #5942 ## User impact MCP-managed font assets now keep their identity, metadata, and generated CSS behavior predictable across upload, editing, synchronization, and rendering. ## Changes - Preserve explicit font metadata overrides while validating supported fields and rejecting unsupported metadata instead of silently dropping it. - Expose font metadata through `update-asset`, with partial updates merged into the existing metadata and validated before the mutation is committed. - Prefer the typographic family name when reading font files, so Fontsource and Google font files resolve to the same logical family where the file provides that information. - Deduplicate generated font sources by URL and format, choose a deterministic source when legacy assets collide, and derive the CSS format hint from the filename when historical metadata is stale. - Add MCP and CLI examples explaining font metadata upload and correction. - Add regression coverage for metadata persistence, invalid overrides, family-name extraction, duplicate source handling, stale format metadata, and generated runtime contracts. ## Verification - `pnpm checks` - 46 workspace packages tested, including 2,355 Builder tests, 809 CLI tests, 2,019 project-build tests, and 97 asset-uploader tests. - All workspace typechecks, Oxlint, package-boundary checks, generated API/docs checks, and fixture link/sync/build checks passed. |
||
|
Oleg Isonen
|
d440e02990 |
feat: Add safer MCP authoring and draft-page verification (#5905)
- **Restore points for agent work**: Create named restore points before major edits, review changes with a dry run, and safely revert project content if needed. - **Shared Slot authoring**: Agents can turn existing sections into shared Slots and insert linked copies across pages, including valid nested Slots. - **DTCG and Figma token imports**: Import design tokens and CSS variables through CLI, MCP, or direct paste, with aliases, modes, composite values, and mappings supported. - **Consistent token conflict handling**: Every paste workflow now offers the same preserve, rename, or merge choices instead of silently overwriting existing design tokens. - **Draft-page previews**: Agents can generate and inspect previews of draft pages without publishing them or exposing drafts in production builds. - **Safer MCP mutations**: Destructive operations require explicit confirmation, invalid raw patches and invented record IDs are rejected, and restore operations commit atomically. - **More dependable agent contracts**: New tools have complete runtime schemas, clearer validation errors, and consistent behavior across CLI, MCP, Builder, and server APIs. |
||
|
Oleg Isonen
|
98d5c607ee |
feat: Improve MCP safety, context efficiency, and agent workflows (#5895)
## Summary This combines the three reported MCP improvements into one release-ready change: - Safer mutation planning and confirmation for destructive operations. - Much smaller MCP handshake and default discovery responses. - Better agent workflows for checkpoints, multi-project work, and page duplication. ## Safer destructive operations - Exposes `dryRun` in MCP schemas for every local-capable mutation. - Returns planned transactions through `meta.session.transaction`. - Requires a two-step confirmation flow for destructive semantic operations. - Returns a short-lived token bound to the project, build version, input, and exact generated patch plan. - Rejects expired tokens, altered inputs, changed versions, and changed plans. - Sets `destructiveHint: true` only for operations that actually require confirmation. - Explicitly protects destructive operations whose names do not begin with delete or replace, including replacing all redirects and clearing all styles from a style source. ## Smaller, focused MCP context - Defers oversized tool input schemas from the initial MCP handshake. - Keeps complete focused schemas available through `meta.get_more_tools` and CLI schema inspection. - Makes template registry lists compact while preserving complete payloads through focused get calls. - Makes component summaries compact by default and supports paginated component details. - Replaces placeholder operation descriptions with useful command descriptions. - Constrains previously untyped JSON schema fields so clients can validate inputs earlier. - Keeps runtime output schemas compatible with strict MCP clients. ## Better agent workflows - Allows read-only discovery and inspection while a workflow checkpoint is awaiting acknowledgement. - Continues blocking mutations until the checkpoint is acknowledged. - Checks persisted checkpoints before every workflow call rather than only the first call. - Adds `--project` support so saved projects can be selected without relinking the current directory. - Isolates project sessions and checkpoints by selected project. - Adds atomic `duplicate-page` substitutions for copied fixed text and copied variable values. - Prevents substitutions from modifying source content or global root state. - Rejects ambiguous or missing copied variables without leaving a partial duplicate. ## Realistic agent verification Minimal-context tests used a weak agent against the local packaged CLI and stdio MCP: - Planned a page deletion, stopped for explicit user confirmation, committed the unchanged plan, and verified the page was absent. - Tried to reuse a confirmation token with altered input; Webstudio returned `DESTRUCTIVE_CONFIRMATION_INVALID` and preserved both pages. - Verified compact handshake schemas and focused retrieval of the complete oversized schema. - Verified compact template and component discovery responses. - Selected a saved project explicitly without relying on the directory link. - Verified checkpointed workflows permit reads but block mutations. - Duplicated a London page into a Paris page with atomic text and variable substitutions. - Verified source content remained London, copied content became Paris, and temporary pages were removed. ## Validation - Project-build focused suites: 188 tests passed. - CLI focused suites: 183 tests passed. - Protocol focused suites: 21 tests passed. - CLI, project-build, and protocol typechecks passed. - Repository lint passed with zero warnings. - Package-boundary validation passed. - Generated runtime operation contracts were regenerated from source. - Real local CLI/MCP tests passed against a running Builder server. Closes #5887 Closes #5888 Closes #5889 |
||
|
Oleg Isonen
|
12ad74584a |
feat: Improve MCP authoring, auditing, and verification (#5886)
## Summary This improves how coding agents discover, author, verify, and audit Webstudio projects through the CLI and MCP. The work focuses on making agents more reliable with less context, while keeping edits semantic and Builder-editable. ## What changes for users - Agents can verify dynamic bindings after authoring and receive actionable diagnostics when expressions, resources, scopes, or component props do not line up. - Agents can create complete resource-backed runtime UI through a semantic operation instead of coordinating fragile low-level edits. - MCP discovery and command output are compact by default, with pagination and focused verbose detail when an agent needs more context. - Asset, style, token, font, page, and component inspection provides stronger usage context so agents can preserve the existing design system. - Rendered audits provide better evidence for responsive behavior, geometry, contrast, generated-build size, and large-project runs. - Cross-project MCP workflows can run from a manifest with explicit mutation approval, isolated failures, resumable progress, and persistent checkpoints. - Project-specific agent onboarding is clearer and does not add telemetry. - Packaged CLI releases gain stronger smoke verification for install, MCP startup, authoring, preview, screenshots, and audits. ## Reliability and architecture - Runtime operation output schemas are generated from the runtime registry and remain client-compatible. - Project-build exports are consolidated around architectural boundaries instead of feature-by-feature convenience entry points. - HTTP, preview, screenshot, and project-session paths are hardened for longer autonomous workflows. - Craft compatibility work is read-only and provenance-aware; it does not silently opt ordinary projects into Craft behavior. - A minimal-context evaluation harness exercises key workflows with a weaker model and stores bounded, privacy-safe evidence. ## Validation - Rebased on the latest `origin/main`. - `pnpm lint` - Repository-wide `oxfmt --check` - CLI typecheck and full test suite: 773 tests passed. - Project-build typecheck and full test suite: 1,866 tests passed. - Protocol typecheck and full test suite: 83 tests passed. - HTTP client tests: 23 passed. - CSS data tests: 3,620 passed, 1 skipped. - Focused Builder tests: 67 passed. - Package-boundary, generated API, and generated documentation checks passed. - Minimal-context agent evaluations passed for authenticated resource-backed UI and design-system-aware responsive authoring. |