Use session.md to understand what the tester did, then use steps.json or a generated script when you need selectors and runnable actions. Treat these files as a retained sequence, not as a complete record of every interaction or review decision.

Reconstruct the retained sequence

Read the session start, duration, starting page, and numbered Recorded actions in session.md. Action offsets are relative to the recorded session start. Earlier interactions may be missing because Recorder retains a bounded timeline.

Compare the sequence with the tester’s written steps. Repeated input and change events can describe one logical interaction. Navigation, keyboard input, file selection, frame boundaries, and complex widgets can be incomplete or represented differently across exports.

Compare the action formats

steps.json is a Chrome DevTools Recorder flow. Its array indexes and setup steps do not match the numbered actions in session.md. It can include an initial viewport and navigation step that were generated to give the flow a starting point.

playwright.spec.ts and puppeteer.js are conversions of supported actions. Inspect their locators, values, and comments for omitted or redacted steps. They do not include assertions for the tester’s expected result unless someone added them after generation.

Join formats by the action’s label, target, value, and order rather than by array index alone. Treat positional selectors and long generated CSS selectors as candidates that must be checked against the application.

Connect an action to other evidence

Use the relative offset in session.md and the absolute request time in network.har to find nearby traffic. Confirm the same subject with an identifier, payload value, route, or returned record. A nearby request may be polling or unrelated background work.

Use notes.md when it names an action or target. Compare screenshots with the sequence, but do not infer an image’s exact timing unless the summary or note supplies it.

Record the gaps

State which setup occurred before the retained interval, which actions are absent or ambiguous, and which values were redacted. Ask the tester for a focused new recording when the missing interaction changes the conclusion. Use Capture support and limits to explain the boundary without implying that the bundle can be reopened in Recorder.