Vally Tests

<!-- cspell:ignore roleplay doxxing scaffolder -->

Vally Tests Skill

Purpose

This skill authors Vally conformance tests for the four supported artifact kinds in this repository: prompts, instructions, agents, and skills. Each test exercises a documented behavior the artifact already claims and routes the result through an appropriate Vally grader so failures are explainable. Test authoring is bounded by a refusal taxonomy that keeps the skill out of adversarial, harmful, or policy-evasion territory.

Follow the shared content-policy public-output guard for eval stimuli and any public summaries derived from this skill. Vally tests are not a venue for terms-of-service boundary mapping, payload generation, refusal-text scoring, hidden-instruction disclosure, PII extraction, or secret extraction.

The skill ships:

• A canonical authoring workflow that runs directly through /vally-tests or through the Vally Test Author subagent.

• Per-kind reference files that enumerate every conformance check the skill knows how to express.

• A grader catalog that maps Vally CLI grader types to the checks they fit, and robustness rules that keep an emitted grader from asserting the impossible or failing correct behavior.

• A safety refusal taxonomy with regex patterns the safety lint script consumes.

• Helper scripts and asset templates for stimulus emission, corpus import, and dedupe.

When to Invoke

Invoke this skill in one of two modes:

• From-artifact mode. The caller points at one or more artifact files (a .prompt.md, .instructions.md, .agent.md, or SKILL.md) through files, or at the open or attached file when files is omitted, and asks for conformance test stimuli that verify each artifact's stated behaviors. The skill detects the artifact kind from the filename unless kind overrides it, looks up the matching per-kind reference, picks graders, and appends stimulus blocks to the routed Vally eval file.

• Corpus-import mode. The caller supplies a CSV or XLSX corpus through path that follows the shipped template, and the skill turns each row into a stimulus block in the routed eval file. Before reading, confirm that path exists and ends in .csv or .xlsx; otherwise name the bad path and stop. Every imported row carries tags.advisory: true, and corpus content cannot override it. Used when a large batch of behaviors needs identical formatting and dedupe handling.

After either mode, report the target eval file, stimuli appended, duplicates skipped, refusals triggered, and the JSON run report path.

Do not invoke this skill to:

• Generate adversarial probes, jailbreak attempts, prompt-injection payloads, or red-team stimuli.

• Test refusal quality on harmful, illegal, or policy-violating prompts.

• Author stimuli that elicit PII, secrets, hidden instructions, or model self-disclosure.

• Store payload examples, paraphrased prohibited requests, or quoted flagged content in Vally eval specs, corpus rows, PR summaries, issue bodies, or comments.

• Replace Responsible AI threat modeling. RAI work lives in the RAI Planner agent.

Authoring Workflow

Each invocation follows the same six-step pipeline:

1. Artifact-kind detection. Resolve the kind from the artifact path or the corpus row's kind column. Supported kinds: prompt, instructions, agent, skill. Reject unknown kinds with a refusal block.

2. Reference lookup. Load the matching reference file from references/ and select the check or checks the stimulus exercises. Stage the artifact under test in the stimulus's agent_environment so the agent can read what the prompt names.

3. Grader selection. Use references/grader-catalog.md as the authoritative selection surface. Pick the registered Vally grader family that directly observes the expected behavior and follow its stimulus-shape guidance.

4. Robustness check. Apply references/grader-robustness.md. Confirm the stimulus stages every file whose wording the grader asserts, that the pattern constrains order and proximity only where those are the behavior under test, that a compliant agent phrasing the behavior differently still passes, and that no positive grader passes on the prompt text alone. Verify the pattern offline against answers that must pass and answers that must still fail before appending.

5. Safety self-check. Run a safety self-check against the refusal taxonomy regex set. Refusing here is the correct outcome for any stimulus that matches a refusal category.

6. Dedupe and append. Compute a SHA-256 hash of the normalized prompt text, compare against existing stimuli in the target eval file, and append only when novel.

The pipeline is identical for both invocation modes. Corpus-import mode runs steps 2 through 6 once per row.

Safety Refusal Taxonomy

Authoring is bounded by seven refusal categories. Any stimulus that matches a category's regex set is refused before it reaches the dedupe step. The categories and their normative source are listed below. The full regex set lives in references/refusal-taxonomy.md and is consumed by the safety lint script.

When a request triggers a refusal, emit the canonical refusal block:

This skill authors conformance tests only. The request appears to fall under <category>. Please consult <Project Code of Conduct | RAI Planner agent> for the appropriate process.

Substitute the matched <category> and the most relevant normative source. Do not negotiate, rephrase, or partially fulfill the request.

Helper Script Index

Helper scripts ship as parity pairs (.ps1 and .sh) where the workflow does not require Python. The safety lint is the exception: it ships only as Lint-VallyTestSafety.ps1, which import_corpus.py invokes, so a single regex engine evaluates the refusal taxonomy. Python is used only for the corpus-import path because the source-of-truth interchange format is CSV with an XLSX mirror.

All helpers honour a shared dedupe contract: SHA-256 of the prompt text after Unicode NFC normalization and whitespace collapse.

The helpers emit a JSON run report to logs/vally-test-author-<timestamp>.json, where <timestamp> is YYYYMMDD-HHMMSS (UTC). The report captures, at minimum: mode; inputs (the resolved files/path and kind); target_eval_file; stimuli_appended (count and per-row hash); dedupe_results (count and per-row hash for skipped duplicates); refusal_check (per-row category match, if any); safety_lint_exit_code; blockers (any ambiguous safety-lint outcomes surfaced for review); and written_paths.

Reference Index

References capture the conformance taxonomy, grader selection rules, eval-suite routing, and the regex source of truth for the refusal taxonomy. Each file targets a specific decision point in the authoring workflow.

Asset Index

Assets supply the interchange formats the corpus-import path consumes. The CSV is the source of truth. The XLSX mirror is regenerated from the CSV by import_corpus.py and is never edited directly.

Output Targets per Kind

Authored stimuli always land in one of the routed Vally eval files. The router is encoded in references/eval-suite-routing.md and mirrored here for quick lookup.

The prompt target is created on first use when a repository has no prompt stimuli yet.

Never write to evals/baseline-equivalence/, evals/script-validation/, or evals/results/ from this skill. Those targets serve baseline equivalence, script validation, and historical comparison flows that are out of scope for conformance authoring.

Contributing

Follow these conventions when extending this skill:

• New per-kind checks belong in the matching references/{kind}.md file. Bump the check count in this SKILL.md when the reference adds or removes checks.

• New grader types belong in references/grader-catalog.md and only after the matching Vally CLI version is pinned in package.json devDependencies.

• New refusal categories require updates to references/refusal-taxonomy.md, the regex set the safety lint script consumes, the Safety Refusal Taxonomy table above, and the canonical refusal block.

• Helper scripts must ship in parity pairs (.ps1 and .sh) unless the workflow has a hard Python dependency. The safety lint is the one exception and ships only as Lint-VallyTestSafety.ps1; do not add a bash mirror. Python helpers live under scripts/ and are configured by the skill's pyproject.toml.