Playwright

Examining a Playwright trace.zip

A Playwright trace is a zip with NDJSON event streams plus screenshots/snapshots. Never read the .zip directly — extract first, then query.

Contents of a trace.zip

Workflow

1. Extract + summarize

Use the helper. It unzips and prints a structured summary.

./.claude/skills/explore-playwright-trace/scripts/extract-trace.sh ~/Downloads/trace-1.zip
# or specify out dir:
./.claude/skills/explore-playwright-trace/scripts/extract-trace.sh ~/Downloads/trace-1.zip /tmp/trace-1

Output includes: action breakdown, console/errors, screenshot count, top network requests, and the extraction dir.

If the helper is not available (other repo): fallback to

OUT=$(mktemp -d) && unzip -q ~/Downloads/trace-1.zip -d "$OUT" && echo "$OUT"

2. Inspect specific things

After extraction, $OUT holds the files. Use jq to drill in.

Timeline of actions:

jq -c 'select(.type == "before") | {t: .startTime, api: .apiName, params}' "$OUT"/*.trace

Failed actions with error message:

jq -c 'select(.error != null) | {api: .apiName, error: .error.message, stack: .error.stack}' "$OUT"/*.trace

Console messages (browser-side):

jq -c 'select(.type == "console") | {level: .messageType, text}' "$OUT"/*.trace

All page navigations:

jq -c 'select(.apiName == "page.goto" or .method == "navigate") | {url: (.params.url // .url), t: .startTime}' "$OUT"/*.trace

Network failures (status >= 400 or no response):

jq -c 'select(.status != null and .status >= 400) | {method, url, status}' "$OUT"/*.network

Slowest actions:

jq -s 'map(select(.type == "before" and .endTime != null) | {api: .apiName, dur: (.endTime - .startTime)}) | sort_by(-.dur) | .[:20]' "$OUT"/*.trace

3. Look at screenshots

ls "$OUT/resources/"*.jpeg | head
open "$OUT/resources/"   # macOS — opens Finder for visual scan

Each action references a screenshot by sha1 in params.snapshot / pageId. To link an action to its screenshot: grab the sha1 referenced near the action's timestamp and match resources/<sha1>.jpeg.

4. Report

Summarize for the user in this order:

1. Verdict — passed / failed; if failed, which action.
2. Error — exact message + the action that caused it (api, params).
3. Console — relevant web errors/warnings.
4. Network — failed/4xx/5xx requests, if any.
5. Last screenshot — the path; suggest opening it.

Quote error strings exact. Do not paraphrase.

Notes

• Some traces have multiple *.trace files (one per BrowserContext). The helper concatenates them.
• Timestamps are unix epoch seconds (floats). Pretty-print with strftime if needed: (.startTime | strftime("%H:%M:%S")).
• Don't dump full NDJSON to the conversation — it's huge. Always filter with jq.
• The helper uses $TMPDIR by default; clean up with rm -rf when done if the trace was big.