mirror of
https://github.com/webstudio-is/webstudio.git
synced 2026-11-03 18:48:06 +09:00
Oleg Isonen
b2a34f1713
## Summary
This PR documents a measured architecture plan for making the MCP
edit-to-screenshot loop fast enough and reliable enough for iterative
site authoring. It intentionally changes no runtime behavior yet.
The current implementation rebuilds a production preview after small
committed mutations. Testing an editable local clone of the production
Webstudio homepage also exposed a correctness gap: a path screenshot
called directly after a mutation can return a stale, byte-identical
image.
## Measured baseline
Repository fixture:
- committed text edit: **24-25 ms**
- post-edit regeneration, production build, and server restart: **7.9
s**
- first screenshot after restart and browser relaunch: **9.1 s**
- combined first post-edit screenshot: **17.1 s**
- unchanged warm screenshot: **1.1 s**
Production homepage, read-only:
- remote refresh: **3.1 s**
- complete production preview preparation/start: **16.0 s**
- cold homepage screenshot: **10.2 s**
- unchanged warm screenshot: **1.36 s**
- predicted post-edit screenshot from measured phases: **approximately
26.2 s**
Editable local production-homepage clone:
- retained **43 routes** and **266 homepage text nodes**
- committed heading mutation: **0.81-0.87 s**
- explicit post-edit preview generation/build/start: **11.6-12.6 s**
- first screenshot after explicit preview start: **9.8-11.8 s**
- complete mutation-to-visible-screenshot cycle: **22.9-25.2 s**
- complete edit, verify, screenshot, restore, and stop run: **52.7 s**
The changed screenshot visibly rendered the test heading, and the
restored screenshot matched the original byte for byte.
## Correctness finding
Calling a path screenshot directly after the committed mutation took
**20.1 s**, but both post-edit captures were byte-identical to the
baseline and rendered the old heading. Explicitly calling
`preview.start({ source: "session" })` made the edit visible.
The reported generated-build version also remained at 1 while the
project session advanced through versions 4 and 5. Current screenshot
output therefore cannot prove that it rendered the mutation it is meant
to verify.
## Proposed direction
- [ ] Introduce explicit iterative and production preview modes under
one controller contract.
- [ ] Keep a Vite development preview and browser alive for MCP
authoring screenshots.
- [ ] Preserve production preview for rendered audits and final
verification.
- [ ] Regenerate from the authoritative project session, wait for Vite
compilation, then normally reload the retained browser page.
- [ ] Make a direct post-mutation path screenshot refresh the latest
committed session without using HMR.
- [ ] Add a project-version freshness handshake so stale screenshots
fail.
- [ ] Serialize and coalesce regeneration after rapid mutations.
- [ ] Return preview mode, rendered project version, and complete phase
timings.
- [ ] Add lifecycle, integration, and local agent evaluation coverage.
## Performance target
Warm edit-to-screenshot latency should be at most **3 seconds p95**
across ten sequential small edits, without invoking `npm run build`,
`npm run start`, or launching another browser.
## Verification
- [x] Measured the source CLI against the authenticated repository
fixture.
- [x] Measured the production homepage without changing production.
- [x] Synchronized the production bundle into a disposable local Builder
project.
- [x] Committed and visually verified a real heading edit through local
MCP.
- [x] Restored the original heading and verified the restored screenshot
byte for byte.
- [x] Reproduced the stale direct-screenshot behavior.
- [x] Formatted the plan with `oxfmt`.
- [x] Verified the documentation diff with `git diff --check`.
The complete architecture, race handling, test matrix, acceptance
criteria, and non-goals are in
`apps/builder/docs/mcp-iterative-preview-plan.md`.