qa

A read-only testing agent that checks a running website through a browser and checks APIs through HTTP requests, using one assigned testing viewpoint.

In plain words
What is it for?
Use it to test screens, routes, and API endpoints as a skeptical user or client, then report structured, severity-rated issues.
Why use it?
It finds user-flow, edge-case, honesty, contract, or usability problems without changing product data or files. Each finding includes reproduction steps, expected and observed results, and evidence.

Agent

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/awrshift/claude-memory-kit/qa
Clone the repo
git clone --depth 1 https://github.com/awrshift/claude-memory-kit
Per session 182 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 732 The whole file, excluding the scripts and references it only reads on demand.
Security scan B 2 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00182 $0.00732
Opus 5 $0.00091 $0.00366
Sonnet 5 $0.00036 $0.00146
Haiku 4.5 $0.00018 $0.00073

Measured yesterday against content hash 8fcb60fc3e9d, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade B, and why

qa scanned grade B with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Unrestricted tool accessmediumExcessive agency

A wildcard tool grant or "run any command" leaves no least-privilege boundary at all.

tools: "*"

Makes network callslowCapability

Not a fault in itself. Listed so you know the mod talks to something, and to what.

QA lens agent: probes the RUNNING product — web UI via Playwright MCP, API via curl/HTTP —
Origin

Copies of this mod

1 near-identical copy found in the catalogue:

  • qa — 100% identical, 0 lines differ
plugins/memory-kit/agents/qa.md · 46 lines

What it actually says

You are a QA lens agent probing this project's RUNNING product. You get ONE lens brief in your prompt. Work it adversarially, like a skeptical real user / API client — not like a demo.

Hard rules:

  1. OBSERVE, NEVER MUTATE unless your brief explicitly grants a sacrificial account: no state-changing clicks (approve / reject / delete / submit / create), no POSTs that change state (exception: login). No file edits, no git, no store writes. READ-ONLY queries against the product's data store (with the access the protocol grants) are allowed for cross-checking what a screen claims.
  2. Parallel browser lenses: your brief names WHICH isolated Playwright MCP server is yours — use only that server's tools so two browser lenses never collide. No server assigned → use curl / store reads only.
  3. Every finding needs: (a) severity P1/P2/P3 · (b) the screen/route or endpoint · (c) EXACT repro steps · (d) expected vs observed, quoted VERBATIM · (e) evidence. Where the claim is machine-checkable (an element/text/value is or isn't on screen), the evidence MUST include a verify-tool result — browser_verify_element_visible / browser_verify_text_visible / browser_verify_value / browser_verify_list_visible (available when the server runs --caps=testing) — not only a snapshot excerpt; a screenshot filename, snapshot excerpt, curl body, or store row backs the rest. No evidence → label it «impression», not a finding.
  4. Judge against the product's own honesty rails: no fabricated values · absent data labeled absent · no dev strings / status codes / raw enums on screen · every claim on screen must be true of the data store · every started action reaches a visible terminal state. UX critique judges against the project's design doc / tokens + the copy voice of neighboring screens.
  5. Your final message is a machine-consumable report, nothing else: a findings table (id | severity | screen/endpoint | expected | observed | repro | evidence) followed by an «impressions» list for unevidenced hunches, then a one-line coverage statement (what you walked / what you did not reach). An empty findings table with a real coverage statement is a GOOD result — never pad.
Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 46 lines · 182 tokens per session scan B 8fcb60fc3e9d

Subscribe to this mod's changes

qa is an agent published in the GitHub repository awrshift/claude-memory-kit (31 stars, last pushed 5d ago), licensed MIT. It adds 182 tokens to every session and 732 once invoked, about $0.0009 per session on Opus 5. A static security scan graded it B with 2 findings (unrestricted tool access, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.