agentleFS
Sign inStart free
vonavikonvonavikon/qa-agent-evals/AGENTS.md

qa-agent-evals

AGENTS.md8 starsChanged 30 days agoMITView on GitHub

Guidance for coding agents working in this repository. qa-agent-evals is a Claude Code plugin that runs acceptance testing for LLM systems (a SUT, "system under test": plugins, skills, bots, browser apps). An analyst subagent profiles the SUT, scenario cards are generated and approved by a human, executor subagents play user roles against the SUT in an isolated environment, a deterministic grep layer checks facts before a judge subagent (running on a different model) issues verdicts, and reports go back to…

  • Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.

    Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.

  • In this repositoryOne other file it gives its agents.

The file

We split this AGENTS.md into the rules it gives an agent

4 rules. About 1 is common, 2 are distinctive, and 1 is very specific to this project.

  • ≈1

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Common:

    Never print or commit secrets (gateway keys, tokens).

    line 67
  2. Very specific:

    Commit identity: this repo publishes under the public identity vonavikon <workwithkonstantin@gmail.com> only.

    line 69
  3. Distinctive:

    Commit messages must not mention employer names or internal project names.

    line 74
  4. Distinctive:

    Run the unit tests; if you touched sandbox or proxy code, run the security suite.

    line 80

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md

AGENTS.md

Guidance for coding agents working in this repository.

What this repo is

qa-agent-evals is a Claude Code plugin that runs acceptance testing for LLM systems (a SUT, "system under test": plugins, skills, bots, browser apps). An analyst subagent profiles the SUT, scenario cards are generated and approved by a human, executor subagents play user roles against the SUT in an isolated environment, a deterministic grep layer checks facts before a judge subagent (running on a different model) issues verdicts, and reports go back to the human. Browser acceptances (v0.6.0) add a visual layer: state screenshots next to transcripts, a visual-reviewer subagent marks flaws with severity frames — the judge never sees that markup and verdicts do not change.

Docs: README.md (RU, canonical) with EN/ZH/ES/FR translations (README.en.md and siblings) · docs/spec.md · docs/architecture.md (+ interactive docs/architecture.html).

Layout

  • skills/acceptance-test/SKILL.md — orchestrator instructions, five phases A–E
  • agents/ — role definitions: analyst, role-user (executor), judge, visual-reviewer (marks browser screenshots; judge-blind by design)
  • knowledge/ — taxonomy (5 card types), adapters and environment isolation, oracle (verdict scale), sandbox levels, golden control sets
  • scripts/ — deterministic machinery: run-cards.sh (headless runs), parse-events.py, check-facts.py (fact grep before the judge), judge-batch.py / judge-control.py, sandbox.py + sandbox-proxy.py (docker isolation, fail-closed), sec-checks.py, annotate-visuals.py (validates visual markup, builds the visual review page, publishes named samples)
  • runs/ — acceptance artifacts (cards, transcripts, facts, verdicts, reports). Case policy: third-party runs are published only with the owner's consent — ask before committing any.
  • examples/evil-plugin — deliberately malicious plugin used by security checks; run it only inside the docker sandbox.
  • tests/ — unit tests and live proxy tests.

Commands

  • Unit tests: python -m unittest tests.test_scripts -v
  • Live proxy tests (dummy key, no LLM calls): python -m unittest tests.test_proxy_live -v
  • Security suite: python scripts/sec-checks.py docker --out runs/sec-ci
  • CI (.github/workflows/security.yml) runs the above on ubuntu. FAIL or ERROR fails the pipeline; SKIP is not PASS.

Conventions

  • Working language for docs and commits: Russian. RU README is canonical; keep the EN/ZH/ES/FR translations in sync when changing it.
  • Verdict scales are fixed — never invent new statuses: judge: GOT / HONEST / PARTIAL / FAIL / ATTACK_PASSED / INVALID; det layer: DET_OK / DET_MISSING / DET_LEAK / DET_NA / DET_NO_DATA.
  • Frontmatter descriptions must parse as strict YAML: wrap values containing colons in double quotes (Claude Code is lenient, external tools like APM are not).
  • docs/benchmarks.md (acceptance log) is generated by python scripts/build-benchmarks.py; never hand-edit it. A new run needs a bench.yaml sidecar (sut/type/adapter/env/findings); counts and cost come from index.jsonl and cost.md. The generator refuses to write the page if its guard filter finds private markers — run --selftest for the positive control. Private runs stay in the private twin repo only.
  • Oracles are written before the run and are never edited after it. Do not "fix" a card's expectations to match observed answers.
  • Never print or commit secrets (gateway keys, tokens). The real LLM key lives only in the sandbox proxy environment.
  • Commit identity: this repo publishes under the public identity vonavikon <workwithkonstantin@gmail.com> only. Before committing, check git config user.email; if it shows any other address (for example a corporate one picked up from another environment), switch: git config user.name vonavikon && git config user.email workwithkonstantin@gmail.com. Commit messages must not mention employer names or internal project names.
  • Scenario cards and the profile are approved by a human before any run; keep that gate intact in code and docs.

Before you finish

Run the unit tests; if you touched sandbox or proxy code, run the security suite. Keep public commits free of private run data.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free