Security Test Planning

Security test planning

You write manual-qa test cases that only observe — a case is admitted by lint when its ## Steps table matches nothing on the forbidden list, or it stays a proposal. There is no reviewable hit and no confirmed state: this bundle never says safe, only "admitted by lint" — a classification about step text, not a claim about the target (references/passive-admission.md).

Where a candidate lives

A candidate case is a manual-qa test case (bundles/manual-qa/knowledge/test-case-format.md) at .agents/security-testing/cases/TC-NNN_<slug>.md — directly under cases/, no per-case subdirectory. The file name must match TC-NNN_<slug>.md and its frontmatter id: must be that same TC-NNN. Use the manual-qa fields as written: id, title, priority (critical|high|medium|low), type, module, size (S|M|L, optional), tags; a ## Steps table with #, Action, Expected Result columns is what gets linted — the Expected Result cell is an observation and is never linted.

A candidate comes from threat-modeler's planned(TC-nnn) dispositions, or you draft one yourself for a finding or a check you want covered.

The admission grammar, in short

Every Action cell must start with an allowed-operation verb (navigate, reload, observe — see the table in references/passive-admission.md) after leading Markdown emphasis is stripped. Anything else is an unknown-operation hit. A forbidden pattern anywhere in the cell — a mutating verb, an injection payload, a security tool name, a volume word, or a literal URL host outside engagement.md targets.browser — is its own hit (references/passive-admission.md has the full rule table). A response header or cookie assertion is written in the audit-step form (references/audit-branch.md) so it reads the way manual-qa's qa-auditor collects headers — browser_network_requests(), then one Inspect the <Header> response header … row per header, never a reload or a dev-tools step.

cases.mjs admit <case.md>

• The candidate must live under .agents/security-testing/cases/, be named TC-NNN_<slug>.md, and its frontmatter id: must be TC-NNN.
• No hits ⇒ the redacted text is copied into tasks/security-<slug>-admitted/ and the case is admitted by lint: ADMITTED <suite>/<file>.
• Any hit ⇒ the redacted text is copied to <st>/proposals/ instead: PROPOSAL <st>/proposals/<file> hits=<n>, then one HIT <rule> step <n>: <action> per hit. Rewrite the step and admit again — there is no review path here that turns a hit into an admission.

cases.mjs verify-suite

Every TC-*.md in the suite directory must be exactly what admit wrote (its current sha256 matches .admitted.json). All match ⇒ SUITE ok=<n>, then the manual-qa hand-off prompt (Run as the active agent (claude --agent test-run-lead): "Run the suite at tasks/security-<slug>-admitted/ against base_url=<url>.") and the test-automation prompt (TA-PROMPT cases=[...] slug=<slug> base=<oid7>). Any mismatch ⇒ one UNADMITTED: <path> per offender, exit 4, no prompts — a file placed by hand or a stale copy has to be admitted (or removed) before either hand-off goes out.

Commands

node scripts/cases.mjs --help:

usage: cases <command> [options]

commands:
  admit
  verify-suite

All: USAGE(<sub>: <why>) (2) · ENGAGEMENT-* (2) · CORRUPT <suite>/.admitted.json (5).

References

references/passive-admission.md — the allowed-operation grammar, the forbidden-pattern table, the classification, the case_sha256 identity. references/audit-branch.md — the audit-step form for header/cookie checks, the mapping to manual-qa's qa-auditor expectations, and what sign-off's UNADMITTED: listing compares against.