Compare HiveSpec vs Superpowers using Claude Code target

[christso]

Objective

Run a head-to-head comparison of HiveSpec vs Superpowers (https://github.com/obra/superpowers/) using Claude Code as the eval target. Measure whether the enriched HiveSpec skills (post EntityProcess/hivespec#1) achieve parity or better on the capabilities that Superpowers covers.

Background

A gap analysis identified three areas where Superpowers had deeper coverage than HiveSpec:

1. TDD rationalization prevention — superpowers:test-driven-development
2. Systematic debugging methodology — superpowers:systematic-debugging
3. Receiving code review discipline — superpowers:receiving-code-review

These gaps were addressed in EntityProcess/hivespec#1 by enriching hs-implement and hs-verify. This eval should verify the enrichments are effective.

Eval Design

Scenarios to cover

1. TDD discipline — Give the agent a feature to implement. Measure whether it writes a failing test first, watches it fail, then implements. Pressure-test with rationalizations ("this is too simple to test", "I already wrote the code").
2. Systematic debugging — Present a failing test or broken feature. Measure whether the agent follows root cause investigation before proposing fixes. Check for the 3-failure escalation pattern.
3. Code review response — Provide review feedback (mix of correct, incorrect, and YAGNI suggestions). Measure whether the agent evaluates technically rather than performatively agreeing.
4. End-to-end lifecycle — Full issue → PR flow. Measure phase discipline, verification evidence, and shipping quality.

Target

• Agent: Claude Code (CLI)
• Plugins: Run each scenario twice — once with HiveSpec installed, once with Superpowers installed
• Metrics: Pass/fail per grading rubric, plus qualitative comparison of agent behavior

Grading approach

Use llm-grader rubrics that check for specific behavioral signals (e.g., "agent wrote test before implementation", "agent traced root cause before fixing", "agent pushed back on incorrect review feedback").

Acceptance criteria

• Eval YAML files created for all 4 scenario categories
• Baseline run with Superpowers
• Comparison run with HiveSpec
• Results published to this repo with agentv compare output
• Summary of parity/gaps documented