Improve agent-plugin-review skill to pass remaining 3 eval tests

[christso]

Summary

The agent-plugin-review skill passes 6/9 eval tests against pi-cli (mean score 0.722). Three tests fail consistently:

Test	Score	Issue
detect-relative-file-paths	0.500	Partially detected — skill mentions leading / but agent doesn't consistently flag it
detect-repeated-inputs	0.000	Missed — agent doesn't suggest top-level input for repeated file references
detect-missing-hard-gates	0.000	Missed — agent doesn't flag missing artifact existence checks between phases

Approach

Use the agentv-bench eval-driven iteration loop:

1. Analyze the failing test transcripts to understand what the agent does instead
2. Identify which SKILL.md instructions are unclear or missing
3. Make targeted edits to the skill
4. Re-run evals to verify improvement
5. Repeat until all 9 pass

Possible improvements

• Relative file paths: Add an explicit checklist item about checking type: file values in eval YAML
• Repeated inputs: Add guidance about the top-level input field from AgentV eval docs
• Hard gates: Make the workflow-checklist.md more prescriptive about what to look for (artifact existence checks at the start of each phase skill)

Eval command

bun run --filter @agentv/core build && bun apps/cli/src/cli.ts eval evals/agentic-engineering/agent-plugin-review.eval.yaml --target pi-cli

Note: must rebuild @agentv/core dist before running if core source was modified.

Related

• PR feat: add workspace skills for pi-cli eval execution #776 — baseline eval results (6/9 pass)
• PR feat: add agentv-plugin-review skill #772 — original skill creation