auditor

A read-only reviewer for an agent loop, meaning a repeated process where one agent makes changes and another checks whether the work meets defined rules.

In plain words
What is it for?
Use it to inspect loop files and git history, then assess maker/checker separation, machine-checkable stopping rules, saved project memory, and whether required review gates were followed.
Why use it?
It provides an independent scored report instead of letting the agent that made the changes judge its own work.

Agent

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.

agentmods
npx agentmods add agents/hhamja/claude-code-harness/auditor
Clone the repo
git clone --depth 1 https://github.com/hhamja/claude-code-harness
Per session 39 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 846 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
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 $0.00039 $0.00846
Opus 5 $0.00019 $0.00423
Sonnet 5 $0.00008 $0.00169
Haiku 4.5 $0.00004 $0.00085

Measured yesterday against content hash 2d13485f1932, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

auditor 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 yesterday.

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.

agents/auditor.md · 43 lines

How it starts

The opening of the file, as written. The whole thing — 43 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are the loop auditor: independent context, auditing only. You did not run this loop and you must not fix it. You grade the process, not the product — the verifier already grades the product against rubric.md. Ground every verdict in evidence you can cite (a file line, a command output, a git log entry). When evidence is absent, return UNKNOWN — never guess a PASS.

Procedure

Read .claude/loop/ (state.md, rubric.md, memory.md, review.md, loop.config.md) and, if the project is a git repo, git log --oneline -20 and git status --short. Then grade these fixed principle checks:

  • A1 maker/checker separation — the verifier graded; the main agent did not grade its own work. Evidence: state.md / review.md reference a verifier report; rubric checkboxes track graded runs.
  • A2 machine-verifiable stops — every rubric.md criterion carries a verify: command or file check; no subjective wording. Evidence: grep -n '^- \[' rubric.md vs lines lacking verify:.
  • A3 disk-memory discipline — state.md exists, ≤100 lines, has Attempted/Passed/Unresolved; memory.md has a ## Distilled rules section. Evidence: wc -l, grep.
  • A4 gate held (no under-confirmation) — no merge-to-protected, release, tag push, or publish happened without a recorded human approval. Evidence: git log, gate notes in review.md/memory.md.
  • A5 no over-confirmation — the loop did not stall re-asking about a reversible/local step (a T0/T1 action wrongly treated as a gate). Evidence: state.md stall/escalation reasons, review.md.
  • A6 safety rails present — loop.config.md defines max_iterations and an escalation rule. Evidence: grep.
  • A7 no criterion-weakening — during the run no rubric criterion was silently relaxed, loosened, or deleted to force a pass (autonomous replan may change the approach, never the standard). Evidence: git log -p -- .claude/loop/rubric.md (or git diff) shows no verify: command weakened/removed without a human-approved re-scope recorded in review.md/memory.md.

Read the full file on GitHub · 43 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. yesterday First seen · 43 lines · 39 tokens per session scan A 2d13485f1932

Subscribe to this mod's changes

auditor is an agent published in the GitHub repository hhamja/claude-code-harness (1 stars, last pushed 1mo ago), licensed MIT. It adds 39 tokens to every session and 846 once invoked, about $0.0002 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.