Workflow Eval Authoring

Author Vally evals that verify multi-skill, multi-tool, or multi-turn orchestration, state, and outcomes across an entire agent conversation, under evals/workflows/. All the actual guidance — naming convention, placement, per-category requirements, glossary, grading patterns, grader catalog, anti-patterns, and worked examples — lives in one shared place so it never drifts across the three eval-authoring skills: the repository-local eval authoring guide at .github/skills/eval-authoring/README.md. This skill exists to route you there with workflow-eval context already loaded, not to duplicate it.

Triggers

USE FOR: write a workflow eval, add end-to-end eval, test tool sequence, create multi-turn eval, add mock workflow scenario, add live eval WHEN: "write a workflow eval", "add end-to-end eval", "test tool sequence", "create multi-turn eval", "add mock workflow scenario", "add live eval" DO NOT USE FOR: per-skill routing/capability evals (use eval-authoring-skill), isolated prompt-to-tool tests (use eval-authoring-tool)

Steps

1. Read the eval authoring guide (.github/skills/eval-authoring/README.md) Step 0 to find this repo's vallyRoot/evalGlobs for the workflow tier, then the guide's "Workflow" column throughout (naming, requirements, worked example).
2. Use evals/workflows/mock/ by default. Choose live/ only when the behavior cannot be represented by the mock MCP; document writes, authentication, cleanup, and nightly-only execution.
3. Model one user goal per stimulus. Use turns only when conversation state matters; otherwise keep a single prompt. Mount every candidate skill explicitly, provide minimal file/git fixtures, and bound turns, tokens, workers, and timeout.
4. Combine process and outcome graders per the guide's four-layer pattern and grader catalog: required/disallowed skills and tools, ordering where essential, files/commands, response quality. Avoid overfitting to incidental call sequences and the guide's anti-patterns (e.g. boundary anti-triggers with no competing skill mounted).
5. Scope graders to turn only when that turn owns the assertion unambiguously; otherwise grade the full conversation.
6. Materialize generated fixture sources, run strict eval-spec lint, then validate locally per the guide's "Running evals locally" section. Build the matching MCP and prime git fixtures when required. Do not finish or open a PR until both lint and the focused eval pass; inspect every turn and tool call on failure.

Rules

• Keep mock workflows hermetic and repeatable; use fake IDs and canned responses.
• Never run a live write scenario without the documented test area and safety environment.
• Preserve relative path invariants for environment.skills, files, and git.source; every environment.files entry needs both src and dest.
• Weight by grader type, include every used type (use 0 only intentionally), and normalize positive weights to 1.0.

References

• Eval authoring guide: .github/skills/eval-authoring/README.md — naming, placement, requirements, glossary, graders, anti-patterns, worked examples, local commands. Shared with eval-authoring-skill/eval-authoring-tool — update it there, not per-skill.
• Repository-local eval README/configuration discovered in Step 0 (if present)