harness-audit

harness-audit

Every hook, MCP, and skill in your Claude Code harness encodes an assumption about what Claude can't do on its own. Those assumptions go stale as the model improves.

A Claude Code skill that audits your harness for dead weight — annotating each component with its origin purpose, then flagging what the current model has outgrown.

The Problem

You add a hook to fix a model limitation. The model improves. The hook stays. Six months later you have:

• A context-reset mechanism for "context anxiety" that Opus 4.5 no longer has
• A retry wrapper for tool failures the model now recovers from natively
• Two session recorders solving the same problem
• Hooks that fire on every tool call but never actually do anything useful

This is dead weight — scaffolding that was once necessary, now just adding latency and complexity.

Install

ccpi install 2233admin/harness-audit

Or manually copy skills/harness-audit/SKILL.md to ~/.claude/skills/harness-audit/SKILL.md.

Usage

Just describe what you want in Claude Code:

/harness-audit

Or naturally:

"Audit my harness for dead weight" "Which of my hooks are still earning their keep?" "I just upgraded to Opus 4.6, what can I trim?"

Claude will:

1. Inventory all hooks, MCPs, plugins, and skills
2. Annotate each with its origin purpose and what model deficiency it compensates for
3. Classify: KEEP / REVIEW / DEAD WEIGHT / UNKNOWN
4. Write a manifest to .claude/harness-manifest.md
5. Give you a dead weight checklist to work through

Output

.claude/harness-manifest.md — annotated inventory of your entire harness:

# Harness Manifest
Generated: 2026-04-04 | Model: claude-sonnet-4-6

## Dead Weight Candidates
- [ ] hooks.Stop.context-reset.py — model no longer abandons tasks near context limit
- [ ] mcpServers.puppeteer — redundant with chrome-in-chrome

## Components

### hooks.Stop.session-to-chat.py
- **Purpose**: Save session conversation to disk
- **Compensates for**: No native session persistence
- **Status**: KEEP
- **Expiration trigger**: When Claude gains native cross-session memory

Core Insight

From Anthropic's engineering blog:

"In a long-horizon agent, Sonnet 4.5 would abandon tasks near context limits. We added a context reset mechanism to compensate. With Opus 4.5, that behavior disappeared. The reset mechanism became dead weight."

The pattern repeats with every model upgrade. Build to delete.

What It Won't Touch

• Hooks compensating for environment limits (filesystem, OS, network) — these won't improve with model upgrades
• Security boundaries and permission gates — intentional constraints, not compensations
• Observability (logging, backup) — serve you, not the model

Contributing

Issues and PRs welcome. The skill is a single markdown file — easy to fork and adapt.

License

MIT