/qa — exercise the running system

Static review reads the code; /qa hits the thing that actually runs. Flows first — a broken happy path makes every adversarial result noise — then the inputs real users and hostile users will eventually send.

Adjacent skills: /secure (hunts vulnerabilities; /qa exercises functionality — a /qa auth finding is handed to /secure) · /audit code (reads the code; /qa probes the running system) · /design-audit (how it looks; /qa is how it behaves).

Run before the skill's steps — per invocation, not per session (4.36); failures degrade to markdown:

link="$(readlink "$HOME/.claude/skills/health" 2>/dev/null || true)"   # empty = not symlinked
pack="$(dirname "$(dirname "$link")")"   # NEVER trust this unless $link was non-empty
if [ "${link#/}" != "$link" ] && [ -x "$pack/bin/acstack-config" ] && ! "$pack/bin/acstack-config" runtime | grep -q '=off'; then
  "$pack/bin/acstack-config" || true          # resolved keys, with sources
  "$pack/bin/acstack-update-check" || true    # ≤1 fetch/day; silent ONLY if already checked today
  "$pack/bin/acstack-recall" || true          # LEARNINGS.md + bug-class names, capped 3KB
else
  echo "runtime off — proceeding without recall/update-check"
fi

Operating principles

• Be direct. Push back in writing when the plan or the user is wrong. No sycophancy.
• Never delete a decision. Supersede it: ~~old~~ → **Verdict (YYYY-MM-DD):** new call — reason.
• Never fix, tune, or delete a test or eval case to raise a score. Log the miss honestly and leave the case unchanged.
• Name exact things: regex patterns, function signatures, model names, before → after numbers. Never "fixed bugs".
• Attribution: follow the project's attribution setting (default none) — no AI-tool mentions in generated docs, no attribution trailers in commits or PRs. Commit with explicit -m/-F messages only.
• Config: read .claude/acstack.md at the project root (fall back to ~/.claude/acstack.md) before acting. ## Settings keys override pack defaults; a ## <skill-name> section overrides both. Unknown keys and sections are ignored.
• Docs: BRIEF.md (frozen seed) / PLAN.md (living plan) / JOURNAL.md (rolling journal). If the repo uses legacy names (PLANNING_PROMPT.md / PLANNING.md / STATUS.md), use those instead — never create both.
• Recall: if LEARNINGS.md exists at the project root, read it before starting.
• Conduct: follow the acstack-conduct block in this repo's AGENTS.md — the word is the mode; the user sets the pace.
• Hackathon lane: if the project's AGENTS.md carries the acstack:hackathon-lane block, only /do changes the repository during the event. Any other skill that would write a tracked file, commit or push says what it would have done and stops; a change that is not a task goes through the lane's operator route.

The probe layer

A probe is the only thing that touches the target; this skill's method and report never name a transport. Every probe provides three verbs — reach (is the target up), act (deliver one input or step), observe (status, body, error) — per the contract in references/probe-layer.md.

• probe: http — implemented. curl conventions in the reference file.
• probe: browser — not yet implemented. Asked for it, say exactly: "browser probe deferred until first real need (decision 2026-07-27); http mode is available" — and offer http. Never fake a browser result.

The report skeleton below is identical whichever probe runs. That is the seam: when the browser probe lands, only the reference file grows.

Safety rule (binding)

Mutating probes — anything non-GET, form submits, state-changing steps — run only against targets that are clearly local or dev (localhost, 127.0.0.1, a URL the user names as dev). Target not clearly local → name the risk and get explicit confirmation BEFORE the first mutating probe. Adversarial probing of a production URL never starts uninvited.

Method

First: does this target need the pass at all? If the change under test touches no runtime behaviour — docs, comments, a version bump, a pure rename — say so in one line and stop. Probing an app that could not have changed produces a PASS that means nothing, and a green report nobody earned is worse than no report.

1. Resolve the target. Argument URL beats config base-url (from a ## qa section). Neither → say exactly what's missing and stop: BLOCKED — no target: pass a URL or set base-url in .claude/acstack.md. Never infer a target from source. An app.listen(3000) in the code says a port the app would use if it were running, not an address that is serving now — probing a guessed URL either fails confusingly or, if something unrelated answers, produces results attributed to the wrong process. A target is supplied, never deduced. Same rule, same reason as the credential rule in step 6: the tempting value sitting in the repo is not the one you are entitled to use.
2. Reach first. One reach probe. Target down → BLOCKED, report ends; nothing else is meaningful.
3. Enumerate flows from BRIEF/PLAN (or run only the flow named in the argument). A flow is a named sequence of steps with an expected outcome — "search returns results", "unauthenticated save is rejected". No discoverable flows → probe the endpoints the user names, and say the flow map was thin.
4. Happy path per flow. Expected outcome observed, or a finding.
5. Adversarial pass per input surface, drawing from references/adversarial-inputs.md: garbage tokens, oversized input, regex-special characters, out-of-range numerics, empty and whitespace-only, wrong-type values. Expected: a controlled rejection (4xx, a validation message). A 5xx or a stack trace is a finding even when the input was absurd — absurd input is exactly what arrives in week one.
6. Auth-gate probing. Every gated endpoint, hit unauthenticated, must fail closed: 401/403. A 200 is a security finding — recorded here, handed to /secure. A 5xx is an error-hygiene finding. Credentials come from the user, never from the repo. Probing a gated flow as an authenticated user needs a token or session the user supplies for this run. Do NOT read one out of .env, a fixture, or a config file: a credential committed to the tree is a /secure finding, and using it here would launder a defect into a passing test. None supplied → probe the unauthenticated half (which is the security- relevant half anyway), and report the authenticated flows as not probed — no credentials supplied.

Report shape

First line is the verdict: PASS / FAIL — <n> findings / BLOCKED — <precondition>. Then:

• Per-flow table: step | input class | expected | observed | verdict.
• Findings, each with the exact repro command (the literal curl line) and observed output — a finding that can't be re-run isn't one.
• Scope: target probed, probe mode used, flows and surfaces NOT exercised. /qa reports; fixes are the user's call.