[MeharPro]

What

Follow-up to #72, which merged with only its first commit; these fixes were pushed 40 minutes after the merge and never reached main. An adversarial review of the fixture (6 reviewers, then 3 skeptics per finding) upheld 54 of 55 findings, none blockers, 16 major. This applies them. All changes are in packages/playwright, plus one CI step, one AGENTS.md cell and the lockfile's engines line.

The ones that matter most:

• Console noise. The fixture logged a console error and a failed request on every page load (the agent link's aborted hello), which fails any suite that checks for those. The agent link now gets a stand-in fetch for its own address and sends nothing, also on strict-CSP pages.
• Chaos. It fetched responses from Node and held their delivery, which bypassed context.route and routeFromHAR mocks, buffered streams, and keyed requests by response arrival, so seeds did not reproduce. It now holds the request and then calls route.fallback(); keys are assigned when the page sends the request. reorder: false sends requests in order, at least 30 ms apart.
• Detection confidence was the action's probability; it is now the model's diagnosis probability, taken from the runtime's decide event.
• Mode. ?genclass=heal or localStorage.genclass could raise the fixture's runtime out of observe; it is now set back.
• Lost or stale files. A retry overwrote the first attempt's file, and files from earlier runs could count as extra seeds. Retries get their own file (-r<n>), skipped tests write none, and the default output is the project's outputDir, which Playwright empties each run.
• Unchecked runs looked clean. The race file gets a fixture record (model option, agent state, requests chaos held, service worker, mode override), and the report notes runs the fixture could not check. Files without that record are reported exactly as before.
• Types and packaging. test is typed from @playwright/test, typesVersions for moduleResolution: node, engines ^20.19.0 || >=22.12.0 (CommonJS projects require the ESM-only fixture), and the README's config uses defineConfig<GenClassFixtures> (without it, TypeScript rejects use.genclass).
• CI. Wider path filters, and the runtime-resolution unit test now runs where the runtime is built. main already has the pipefail fix for the e2e step. run.mjs now fails on other signatures on the buggy page, and the spec fails on any console error or failed request.
• Docs. README sections on request-side chaos, routes and service workers, the globals the page gets, the e2e's script-tag mapping, per-test model download, and the compat numbers with their runtime version and the 270 injected-failure findings.

The commit message lists every change.

Checks

• npm run build -w @genclass/runtime passes (no runtime source is touched; the e2e serves its dist/)
• npm run typecheck -w @genclass/playwright, npm run build -w @genclass/playwright, cd packages/playwright && npx vitest run: 114 passed. The CI smoke run, which loads the reporter by module name, passes
• no packages/runtime/test/review-*.test.ts was edited
• no model-visible text changed (nothing under packages/runtime/src/)
• docs/runtime/API.md, the runtime's src/types.ts and packages/runtime/CHANGELOG.md: not applicable (runtime untouched); the package's README, CHANGELOG and JSDoc are updated
• a correct app is not made worse: the runtime stays in observe; chaos, off unless given, holds matched requests for at most delayMs (10 s cap) and never changes them
• sim/ and demos/ are untouched
• package-lock.json is in sync (one line: the package's engines); npm ci passes

Notes for the reviewer

Measured: packages/playwright/test/e2e/run.mjs, headless Chromium 153, Playwright 1.63.0, model: "skip". Six runs of five seeds on 2026-10-25 before rebasing onto main (30 seeds per row), and one more run after:

page	chaos	seeds with witness:overwrote-newer on GET /api/search	other signatures	console errors or failed requests
buggy	latency 0 to 600 ms, reorder	30 of 30	0	0
buggy	latency only (reorder: false)	0 of 30	0	0
fixed	latency 0 to 600 ms, reorder	0 of 30	0	0

In a separate browser check, a context.route mock still answers under chaos, a default-src 'self' page logs nothing, and ?genclass=heal is set back to observe.

The one finding not upheld: the e2e summary's staleSeeds label (judged a naming choice).

Rebased onto main at 7f96182. Conflicts were in AGENTS.md (the new package rows on main were kept) and in the workflow (the pipefail comment from main was kept). Not run: the runtime's own suites (unchanged), test/smoke/smoke.sh, the install suite, and anything on Azure.

---

Generated by Claude Code