skill-eval-comparator

skill-eval-comparator is an agent for Claude Code from QBall-Inc/the-bulwark. It costs 64 tokens per session (1,510 once invoked), scanned A, original, MIT.

A comparison tool for two automated evaluations of coding-agent skills. It reads their grading reports and compares each test result and supporting evidence.

In plain words
What is it for?
Use it to compare skill versions or prompt variants, find which tests each run wins, and produce an overall winner-or-tie decision.
Why use it?
It shows which version performed better without rerunning the evaluations or relying on subjective impressions. It also distinguishes passing tests from stronger or weaker evidence.

Agent for Claude Code

Written for Claude Code: shipped in a Claude Code plugin. Also seen: model in frontmatter; reads .claude/ paths; mentions subagents.

Part of the the-bulwark plugin — 30 skills, 17 agents, 6 hooks shipped together

Good fit Use it to compare skill versions or prompt variants, find which tests each run wins, and produce an overall winner-or-tie decision.

Compare 6 agents from other repositories ↓
Install with agentmods
npx agentmods add agents/qball-inc/the-bulwark/skill-eval-comparator
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Clone the repo
git clone --depth 1 https://github.com/QBall-Inc/the-bulwark

Made for: Claude Code.

Or install the-bulwark, the plugin that ships this one along with the rest of its 30 skills, 17 agents, 6 hooks.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for skill-eval-comparator

README.md
[![agentmods](https://agentmods.dev/badge/agents/qball-inc/the-bulwark/skill-eval-comparator.svg)](https://agentmods.dev/agents/qball-inc/the-bulwark/skill-eval-comparator)
Your own site
<a href="https://agentmods.dev/agents/qball-inc/the-bulwark/skill-eval-comparator"><img src="https://agentmods.dev/badge/agents/qball-inc/the-bulwark/skill-eval-comparator.svg" alt="Measured on agentmods" height="20"></a>
Per session 64 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,510 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00064 $0.01510
Opus 5 $0.00032 $0.00755
Sonnet 5 $0.00013 $0.00302
Haiku 4.5 $0.00006 $0.00151

Measured 8d ago against content hash 55600645e4dc, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-07, from the pricing page.

Security

Grade A, and why

skill-eval-comparator scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/create-skill/agents/skill-eval-comparator.md · 159 lines

How it starts

The opening of the file, as written. The whole thing — 159 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Skill Eval Comparator

You are a blind A/B comparator for Bulwark skill evaluation runs. Your role is to compare two grading reports against the SAME evals.json and produce a structured comparison: per-test winner (or tie), per-assertion delta, and an overall verdict.


Critical Constraint (BINDING)

You MUST compare grading data, not opinions.

Your inputs are two grading-<timestamp>.json files produced by grade.ts. Both were graded against the same evals.json. You are NOT re-grading — you are comparing the verdicts and evidence already produced.

You MUST NOT:

  • Re-run the grading logic.
  • Form opinions about which skill version "feels better" without evidence from the grading files.
  • Conflate "PASS rate" with "quality" — both runs may pass while one has weaker assertion evidence.
  • Read source files unless a tie-break specifically requires inspecting evidence_paths.

Mandatory Execution Checklist (BINDING)

  • Read both grading-<timestamp>.json files.
  • Verify both reference the same evals.json schema version and the same tests[].id set. If they don't, abort with a clear error.
  • For each test ID, build a comparison row: A verdict, B verdict, A pass count, B pass count, A duration, B duration.
  • For each assertion within each test, compare verdict + evidence string across A and B.
  • Identify per-test winners using the precedence rule below.
  • Produce overall summary.
  • Write structured comparison to tmp/eval-comparator/<timestamp>.json and surface a one-paragraph result in the response.

Per-Test Winner Precedence

Apply rules in order. First clear winner stops the chain.

  1. Verdict precedence: PASS > SKIPPED > FAIL. If A=PASS and B=FAIL, A wins.
  2. Assertion pass count: if both verdicts are equal, the run with more PASSed assertions wins.
  3. Evidence quality: if pass counts are equal, compare evidence strings. A win for the run whose evidence cites more specific paths/line numbers/event ids. If indistinguishable, it's a tie.
  4. Duration: if everything else is equal, the faster run wins by tie-break, marked as tie_broken_by: duration.

Read the full file on GitHub · 159 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 8d ago First seen · 159 lines · 64 tokens per session scan A 55600645e4dc

Subscribe to this mod's changes

skill-eval-comparator is an agent published in the GitHub repository QBall-Inc/the-bulwark (8 stars, last pushed today), licensed MIT. It adds 64 tokens to every session and 1,510 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other agents, from other repositories

verification-runner

Run project-aware verification loop. Reads mix.exs to discover tools (credo, dialyzer, sobelow, excheck), test commands, and custom aliases. Use proactively after code changes.

oliver-kriska/claude-elixir-phoenix · 44 tokens

tester

Use this agent after chunk implementation to create comprehensive test suites, or when the user requests test generation. Creates unit, integration, and edge case tests to ensure code works correctly and provide shipping confidence. Context: All chunks are implemented, orchestrator invokes testing phase. user: "All…

drobins25/craft · 200 tokens

symfony-tdd-coach

Guides TDD workflow for Symfony projects using Pest PHP or PHPUnit. Drives strict RED-GREEN-REFACTOR cycles with proper test isolation, Foundry factories, and regression protection. Use when writing tests, adding test coverage, or practicing TDD.

dev-toolings/superpowers-symfony · 58 tokens

mb-test-runner

Structured test runner — detects stack, runs tests, parses per-stack output into strict JSON (pass/fail, counts, per-failure file+name+errorhead). Invoked by /test and plan-verifier Step 3.5. Never silently reports "not-run" as pass.

fockus/skill-memory-bank · 64 tokens

debugger

Systematic debugger using the Iron Law: no fix without confirmed root cause. Reproduces errors, traces execution paths, forms and verifies hypotheses, then implements…

KevinZai/commander · 35 tokens

qa-engineer

Senior QA engineer for test suite creation, coverage analysis, and quality assurance. Runs in an isolated worktree and produces coverage deltas and structured reports…

KevinZai/commander · 34 tokens