Cancel timed-out Playwright executions

[masnwilliams]

summary

• disconnect the persistent Playwright CDP client when an execution reaches its request timeout
• prevent abandoned scripts from continuing to mutate browser state after the timeout response
• reconnect transparently for the next execution
• extend the timeout regression test to detect delayed page mutations

proof

The reproduction starts a script that alternates scrollIntoViewIfNeeded() between top and bottom targets every 500ms. After ten iterations it sets a marker to late-finished. The request timeout is shorter than the script, so any marker or scroll changes after the timeout response are escaped execution.

before: main

The API returned the timeout after approximately one second, but the script remained live:

{
  "timeout": { "success": false, "elapsed_ms": 1004 },
  "immediate": { "marker": "initial" },
  "after_2_seconds": { "marker": "late" }
}

The post-timeout mutation reproduces consistently with both the production endpoint and a local Chromium daemon built from main.

after: this branch

Built the complete headless image from this branch (sha256:b02c5efcc8a1e06933f38b163326b1f65f6dc144267c2706cee787ea48c01101) and ran the same alternating-scroll request through the image's /playwright/execute endpoint:

{
  "timeout": { "success": false, "elapsed_ms": 2015 },
  "immediate": { "marker": "step-3", "scroll_y": 0 },
  "after_4_seconds": { "marker": "step-3", "scroll_y": 0 }
}

The marker and scroll position remain unchanged after the timeout. The abandoned script does not reach late-finished or perform another scroll. A subsequent Playwright request reconnects and succeeds against the same browser page.

The image-level regression also passes:

=== RUN   TestPlaywrightExecuteTimeoutReturnsPromptlyAndRecovers
playwright timeout regression test passed
--- PASS: TestPlaywrightExecuteTimeoutReturnsPromptlyAndRecovers (8.16s)

validation

• verified the standalone daemon fix with both Playwright and Patchright against local Chromium
• built the full headless image from this branch
• ran TestPlaywrightExecuteTimeoutReturnsPromptlyAndRecovers against the built image
• reran the alternating-scroll reproduction through the built image's real API endpoint
• bun build server/runtime/playwright-daemon.ts --outfile=/tmp/playwright-daemon.js --target=node --external playwright-core --external patchright --external esbuild
• go vet ./...
• go test -v -race $(go list ./... | grep -v /e2e$)

[rgarcia]

reviewed — looks good. the timeout now interrupts the abandoned Playwright execution, reconnects cleanly for the next request, and the regression test covers both recovery and delayed page mutations.