Try a skill separately

Try a skill separately

A skill can improve a review, add irrelevant work, or make no difference. A small separate task gives you evidence before you use its recommendations.

The fixture lives in examples/skill-trial in the Ennodia checkout. It contains a short function and a local contract with deliberate mismatches.

Prepare two comparable attempts

1. Select one available agent and model.
2. Check that the requested skill is installed for that agent.
3. Give both attempts the same retry.ts and contract.md files.
4. Keep the evaluator check.ts out of the review materials.
5. Run the plain review and skill review as separate read-only tasks.

Use Ennodia MCP for two separate read-only reviews of examples/skill-trial/retry.ts.
Provide contract.md as the source of truth. Use the same agent and model.
First run: request no additional skill.
Second run: request the installed source-grounded-audit skill.
Do not open check.ts until both reviews finish.
Return each contract mismatch with a concrete failing input.
Compare useful findings, unsupported claims, and time spent.

Your agent can use ennodia_start_compositional with one slice per attempt. Set skillIds: [] on the plain slice and skillIds: ["source-grounded-audit"] on the skill slice. The skill guide describes installation and explicit selection.

A native agent can self-select other available skills. An empty requested skill list does not guarantee a skill-free control. For strict measurement, use separate controlled skill inventories and record the skills actually loaded.

Validate the findings

Run the evaluator after saving both reviews:

bun examples/skill-trial/check.ts

The supplied implementation deliberately fails some cases. Use each reported input to check a proposed fix in a disposable copy. A useful finding needs to explain a contract mismatch, not just suggest a stylistic change. Keep useful partial findings even if the skill adds other irrelevant advice.

This fixture supports a repeatable workflow demonstration. No quality improvement is claimed for it.

Recorded development trial

On September 6, 2026, two Antigravity tasks ran through Ennodia in separate temporary copies. The plain review found the six failing evaluator cases and an additional alternative-number-notation case. The requested-skill attempt produced no answer because native tool permissions blocked a required read.

The agent command exited with code zero despite the permission denial. Ennodia initially labeled that task successful. This trial exposed a classification defect that the next release candidate fixes. The denied attempt cannot establish the quality of the skill. No permission bypass was used to force a successful result.

The fixture includes recorded-trial.json with both observed outcomes and the diagnostic excerpt. Local paths were normalized, and CLI remediation instructions were omitted from the excerpt.