Playwright Trace Analyzer

Turn a Playwright trace.zip into a ranked, evidence-backed report of flakes, slow steps, and root causes.

Index file. Detailed extraction rules, analysis playbooks, and report templates live under rules/, references/, and templates/. Load only what the current phase needs — the body of SKILL.md is a thin orchestrator.

Inputs

The user passes one or more of:

If the user passes a report.json plus a trace.zip, treat the report as a high-level test status map and the trace as the source of truth for timing and network data.

If a test-results/ directory is passed, scan for the most recent trace.zip per failed test and process them in order of failure recency.

See rules/input-detection.md for the precise detection logic and unpack recipe.

Workflow

Six phases. Do not skip a gate.

Required reading by phase

Load on demand — do not preload.

Pass-vs-fail comparison (when given two traces of the same test): scripts/trace-diff.mjs.

Confidence-gated iteration

After the first pass at root-cause analysis, invoke the confidence skill in analysis mode:

Skill(skill="confidence", args="analysis")

Apply this gate:

After two deep-dive iterations without reaching 90%, stop and present findings as a hypothesis with the evidence required to confirm it. See rules/confidence-loop.md.

Core principles

1. Measure before recommending. Every fix must be tied to a number from the trace. "This selector is slow" is not a finding; "page.click('text=Save') waited 4,820ms across 3 attempts before the button became actionable" is.
2. Distinguish symptom from cause. A timeout is the symptom; the reason the wait never resolved is the cause. The trace records the before and after DOM snapshots and every poll attempt — read them.
3. Rank by impact, not by ease. A 50ms selector fix that runs once matters less than a 200ms wait that runs in beforeEach of 40 tests.
4. Auto-detect, do not interrogate. Read the file shape, infer the format, state what you found. Ask the user only if detection genuinely fails.
5. Confidence-gated honesty. If /confidence returns < 90%, dig deeper or admit uncertainty. Do not paper over a weak diagnosis with a confident-sounding fix.
6. One trace at a time, but diff when given two. A passing run vs. a failing run of the same test is the strongest signal available.

Anti-patterns (one-liners — full list in

references/flake-patterns.md)

• Recommending page.waitForTimeout(N) without measuring the underlying race condition.
• Calling a test "flaky" without naming the race — every flake has a cause; "non-deterministic" is not a diagnosis.
• Reporting raw action counts without converting to wall-clock time share.
• Skipping the network log when an action timed out — most action timeouts are blocked on a request that never resolved.
• Swapping text= for getByRole(...) without checking whether the failure was selector resolution or actionability (visibility, pointer-events, animation).
• Fixing one slow action and ignoring the long tail of cumulative auto-wait delays (death-by-a-thousand-cuts is the common case in real suites).

Trace-analysis quickstart

Input is a GitHub Actions run URL

node <skill_dir>/scripts/fetch-gh-run.mjs https://github.com/<owner>/<repo>/actions/runs/<id> [--out <dir>]

The script uses the gh CLI (gh run download) to fetch every artifact whose name matches Playwright conventions (playwright-report*, playwright-traces*, test-results*, *-traces, *-trace), unpacks nested ZIPs, and writes a manifest of all trace.zip files discovered, grouped by failed test where possible. Then continue with the unpacked flow below. If gh is not installed or unauthenticated, ask the user to download the trace artifact manually from the Actions run page and provide the local path — the trace.zip flow below is unaffected.

Input is a trace.zip

1. Unpack and index.

   node <skill_dir>/scripts/trace-extract.mjs <path/to/trace.zip> [--out <dir>]

   Writes a normalised trace.trace.jsonl, trace.network.jsonl, and a manifest of resources/snapshots into <dir> (defaults to a sibling <name>.unpacked/).
2. Run the summary.

   node <skill_dir>/scripts/trace-summary.mjs <dir>

   Prints: total wall-clock, top-N slow actions, top-N slow requests, console errors, page errors, and the failing-action stack trace if present.
3. (Optional) Diff a passing trace against a failing trace.

   node <skill_dir>/scripts/trace-diff.mjs <pass-dir> <fail-dir>

   Surfaces actions that diverge in duration, requests present in one but not the other, and the first action where the two timelines fork.
4. Map suspects to source. Use rules/flake-diagnosis.md Phases 3–4 to go from action callId → test file/line (Playwright trace events embed location: { file, line, column }).

The full extraction methodology (capture protocol, how to interpret the network log, common flake shapes) is in rules/flake-diagnosis.md. Don't preload it — only when an input is detected.

Definition of Done

• Input format detected and stated (zip / dir / single JSONL).
• Failure mode named — timeout, assertion, error, or slow-passing (Phase 1).
• Top-N slow actions listed with measured dur (ms).
• Top-N slow requests listed with responseEnd - requestStart (ms) and status.
• Console / page errors captured verbatim.
• Each hotspot mapped to a test file + line (from location in the trace event), or to an app file when the cause is in product code.
• /confidence analysis reached ≥ 90% (or two deep-dives recorded with the remaining uncertainty surfaced to the user).
• Fix plan written using templates/analysis-report.md, with ranked fixes, expected ms saved, and a re-run verification step.
• User has the next concrete action (apply fix N, re-run with --trace=on, compare).