Allure Agent Mode

Use this skill for feature or bug work that changes tests, for reviewing existing tests, auditing coverage, triaging failing suites, investigating weak evidence, or debugging flaky and environment-sensitive failures.

Why Agent Mode

A test run is an instrument, not a pass/fail to scrape. Running tests through allure agent turns any run into a reviewable account of what actually happened — the steps, HTTP exchanges, SQL, browser sessions, and attachments — plus automated findings about the tests themselves, such as a test that ran with no assertions or an API call with no evidence it was made. The console gives you a status; the agent output gives you understanding and a critique, and a report a human can actually review. Reach for it because it makes the work better.

• Run with intent. allure agent --goal "<what this run should confirm>" -- <test cmd>. Recent allure agent prints a run summary to stdout — counts, findings with severity, expectation result, the report link, and pointers into the output — so the headline reaches you without opening a file (a --silent option is available to quiet passthrough noise; confirm its effect with allure agent --help). Read that summary, then open the agent output for depth: the output dir's AGENTS.md guide, index.md, the findings, and the per-test evidence. The report carries the goal next to the evidence, so you, an upstream agent, or a human can validate the run against what it was meant to prove.
• Make evidence checkable with expectations. When a behavior must be proven — an API call, an attachment, a step — add the matching --expect-* option. Missing expected evidence is a finding, not a pass: it means you cannot be sure the behavior was exercised.
• Triage failures without the hunt. The output names the exact failures with enough context to fix in place. Confirm flakiness by rerunning just the failed tests through agent mode — allure agent --rerun-latest (or --rerun-from <prior-output>, recovering it via allure agent latest) with the matching --rerun-preset, so you need not remember the test name or this framework's filter syntax. Confirm the exact flags with allure agent --help. Don't rerun blind or scrape logs.
• Debug from CI without local repro. Run allure agent inspect on downloaded allure-results to get the same review surface with full context — often enough to fix a failure you cannot reproduce locally (a Windows-only failure from a Mac).
• Use tests as a debugging instrument. A test already sets up the environment, exercises the behavior, and captures the evidence — reach for the relevant test under agent mode instead of ad-hoc bash or driving a browser MCP by hand.
• Learn the real surface. Run allure agent capabilities to see what the local CLI actually supports instead of assuming.

The run already prints a summary digest to stdout — read it, then open the full output. Reducing the run with tail/grep/head or >/dev/null, or stopping at the printed counts, throws the rest away: the findings, the evidence, sometimes whole test binaries. The agent output is the signal.

Read First