Create Skill Test

Create Skill Test

Scaffold an evaluation spec (eval.yaml) for a skill or agent so it conforms to the Vally schema, passes skill-validator check and check_eval_quality.py, is powerful enough to return a verdict, and does not overfit to the skill's own wording.

When to Use

• Creating a new eval.yaml for a skill or agent
• Adding stimuli to an existing eval
• Sizing an eval so the pass gate can actually be reached
• Setting up or repairing fixture files alongside an eval
• Reviewing whether rubric items and graders risk overfitting

When Not to Use

• Diagnosing a failing or regressed eval — use improve-skill-quality
• Modifying the skill-validator or the evaluation workflows
• Creating or editing SKILL.md files — use create-skill

Inputs

Workflow

Step 1: Prove the eval belongs to the target

Before writing YAML, state what the stimulus proves. A preference stimulus is necessary when the target should improve the answer, action, restraint, or validation result compared with the same model without the target. A non-voting activation contract or no-op guard is necessary when it protects a meaningful invariant, even if correct behavior is baseline-equivalent.

Do not add a stimulus when it measures:

• generic knowledge the base model already has;
• path recall for a skill that only points to reference files;
• output volume rather than correctness;
• a renamed or lightly reworded copy of an existing case; or
• a disable-model-invocation: true reference skill in isolation.

Map each proposed case to capability, risk, and customer journey tags. Preference cases must add distinct voting value. Activation contracts and no-op guards may share a capability when they protect a separate routing or preservation invariant. If two cases have the same inputs, expected outcome, failure mode, and grader path, keep the stronger one. The five-stimulus floor never justifies padding.

Then locate the target and test directory:

tests/<plugin>/<skill-name>/eval.yaml          # skills
tests/<plugin>/agent.<agent-name>/eval.yaml    # agents (the agent. prefix disambiguates)

Verify the target exists at plugins/<plugin>/skills/<skill-name>/SKILL.md or plugins/<plugin>/agents/<agent-name>.agent.md, and read it.

Agent evals use the native SDK agent lane. Vally 0.14 cannot register custom agents, so agent.* specs do not run through the skill experiment. The evaluation workflow discovers them separately, runs the target agent through skill-validator evaluate, and adapts that evidence into the same schema-versioned result and dashboard pipeline. The distinct-stimulus floor applies to both skill and agent evals.

Be careful with a skill that sets disable-model-invocation: true. The model cannot invoke it, so the skill is absent from the model-facing skilled arm and any direct eval compares two identical arms. Answer-content graders do not create a difference between those arms. The honest coverage for such skills is dependency-level — through the outcome evals of the skills that load them, and through the plugin arm. For example, filter-syntax is covered by the filtered-command scenarios in tests/dotnet-test/run-tests/eval.yaml.

Step 2: Write the spec skeleton

The spec is Vally format. Every eval in this repo uses stimuli: and graders:; scenarios: and assertions: are a pre-Vally format that no longer loads.

name: <skill-name>
description: Evaluates the <plugin>/<skill-name> skill
type: capability
defaults:
  timeout: 5m
  runs: 1
stimuli:
  - name: <what the agent must accomplish>
    prompt: <natural developer request>
    tags:
      capability: <distinct-capability>
      risk: <failure-being-prevented>
      journey: <customer-task>
    environment:
      files:
        - src: fixtures/<case>/Project.csproj
          dest: Project.csproj
    graders:
      - type: output-matches
        config:
          pattern: (root cause|underlying issue)
      - type: exit-success
      - type: prompt
    rubric:
      - <outcome the agent should have reached>

Use defaults: only. config: is a deprecated alias that the repository gate rejects. Vally warns when the alias appears alone and throws when a spec declares both keys. Replace config: with one defaults: block and preserve its settings.

Step 3: Size the eval for power before writing content

The gate gives each distinct stimulus one vote. Repeated runs for one stimulus collapse to one majority-direction vote and remain available as reliability evidence.

1. Distinct stimuli ≥ 5, else the verdict is underpowered — never a pass, never a regression.
2. p ≤ 0.05 on an exact one-sided sign test over discordant (non-tie) stimulus votes. Ties are not discarded; they hold the discordant count down.

At exactly 5 stimuli, one tie is fatal because it leaves 4 discordant votes. At 6 stimuli one tie is survivable; at 7, up to two are. A loss is not. Five is an eligibility floor, not adequate power. For example, 80% power needs 8 discordant votes only for a true 90% conditional win rate; it needs 18 at 80%, 37 at 70%, and 158 at 60%. Size for the effect and tie rate you need to detect.

Use runs for reliability, not task breadth. Vally recommends 3 runs in CI and 5–10 nightly for pass rate, pass@k, pass^k, and flakiness. Extra runs never clear the five-stimulus floor.

Do not set runs in dotnet-skills.experiment.yaml; experiment overrides overwrite every eval's own value rather than defaulting it.

Step 4: Write stimuli

• Name describes what is tested, not how.
• Prompt is a natural developer request. Never mention the skill, the agent, or its vocabulary — cued prompts inflate the overfit score and bias the baseline.
• Each stimulus should discriminate a different property of the skill. Five stimuli covering one property give arithmetic, not evidence.
• Give every capability stimulus non-empty capability, risk, and journey tags. Use stable lowercase kebab-case values. A useful portfolio crosses distinct rows or columns in that matrix; it does not repeat one journey with cosmetic wording changes.
• Give every stimulus a stable, unique name. Vally pairs comparison trajectories by (stimulus name, trial index); duplicate names make slot identity ambiguous.
• Include a boundary / no-op stimulus for any skill that migrates or rewrites code, proving it leaves already-correct input alone.
• Add a dormancy stimulus for each real routing boundary. No-op proves restraint on an on-target, already-correct input; dormancy proves the target stays inactive on an off-target request.

Step 5: Configure the environment

environment:
  files:
    - src: fixtures/broken-build/App.csproj      # path relative to eval.yaml
      dest: App.csproj                           # path in the agent's working directory
    - src: fixtures/broken-build                 # a directory
      dest: .
  commands:
    - dotnet build -bl || exit 0                 # guard intentional failures

Do not set environment.skills in a skill eval. The experiment declares vary: /environment/skills and supplies the value itself — [] for the baseline arm and plugins/<plugin>/skills/<skill> for the skilled arm — so anything the eval declares is replaced, in every arm. It cannot add a skill to one arm only. environment.skills is meaningful in an agent.* eval; the native agent lane loads those entries only in the isolated target run, while the plugin run loads the production plugin's complete skill surface. Copy the shape from an existing agent eval such as tests/dotnet-test/agent.test-quality-auditor/eval.yaml rather than reproducing a remembered form — the specs in this repo are not consistent about how they spell those entries.

Fixture rules — each one has already cost a real result:

• Every referenced fixture must be tracked by git. .gitignore (e.g. coverage*.xml) has silently swallowed a committed fixture: the eval passed locally and failed at setup in CI. Verify with git ls-files, not by looking at the working tree.
• Every fixture must behave as its stimulus assumes. A fixture meant to be healthy must build; a fixture meant to be broken must fail for the exact reason the stimulus is about, and no other. Judges penalize agents for unrelated "pre-existing build issues" that the fixture author introduced.
• Every fixture must reproduce the bug its stimulus is named for. If it does not, the baseline scores well and the skill has nothing to add.
• Coverage fixtures must be internally consistent. A Cobertura report whose declared line-rate, summary totals (lines-covered/lines-valid), and <line> elements disagree lets the two arms read different truths, and the loss is the fixture's fault. Update any rubric item or prompt that quotes a figure in the same change.
• Do not wire duplicate fixtures to raise n; rename leftovers add trials without evidence.
• A setup command that is expected to fail while still producing its artifact must be guarded (|| exit 0), or vally drops the trial.
• A cleanup command that strips sources must skip directories containing SKILL.md — the staged skill lives there, and deleting it aborts only the skilled arm.
• Preserve the complete file set. When the prompt limits edits or requires source/test preservation, snapshot or compare every in-scope file, not one representative file. A grader that checks only the main output can miss deletion, truncation, or edits to sibling files.

Step 6: Write graders

Graders are hard pass/fail checks evaluated on every arm.

Rules:

• A grader whose config is absent or missing its required key parses fine and enforces nothing. The usual cause is an indentation slip during an edit; check_eval_quality.py blocks it.
• Prefer broad patterns that several valid approaches satisfy: (root cause|primary error|underlying issue).
• If the skill mandates an output shape, assert on it. A skill required to emit a decisive Recommendation: line can silently stop doing so while the eval still passes.
• Use file-not-contains / file-not-exists to prove the agent avoided an incorrect action.

Define the deterministic contract before writing the prompt grader:

1. Golden acceptance: materialize the fixture and apply the golden_patch, if any. The golden workspace and final golden_trajectory response must pass every deterministic grader that applies to them.
2. Mutation rejection: make one realistic defect that the eval exists to catch, such as a missing file, zero discovered tests, an out-of-scope edit, or a changed semantic value. The relevant deterministic grader must fail.
3. Complete-state check: cover all files and artifacts named by the request. Do not accept a partial artifact because one positive substring exists.

Use golden_patch for replayable workspace state and golden_trajectory for the expected final response. A narrated edit, build, or test is not proof: completion claims need a patch or a run-command grader that replays the evidence.