HarnessDelta — Review Agent Config Diffs

1---
2name: review-agent-config-diff
3description: Review semantic changes to Claude Code and Codex repository configuration across a working tree, staged diff, or Git range. Use when a user asks whether a PR broadens agent permissions, enables hooks or MCP servers, weakens sandbox or approval policy, removes instruction guardrails, changes skill discovery, or otherwise alters coding-agent behavior. Do not use as proof that a runtime loaded a file or execute changed configuration.
4---
5 
6# Review agent configuration changes with HarnessDelta
7 
8Use HarnessDelta's deterministic local CLI before making claims about an agent-facing diff. The command below downloads the pinned, dependency-free CLI from the public Codeberg npm registry; the analyzer itself reads Git objects and working-tree files locally, makes no network requests, and never executes changed configuration.
9 
10## Workflow
11 
121. Identify the repository and comparison the user placed in scope: working tree, staged changes, or a Git range.
132. Require Node.js 18 or newer. Run one of:
14 
15   ```bash
16   npx --yes --registry=https://codeberg.org/api/packages/automa-tan/npm/ harnessdelta@0.1.1 <repository> --json
17   npx --yes --registry=https://codeberg.org/api/packages/automa-tan/npm/ harnessdelta@0.1.1 <repository> --staged --json
18   npx --yes --registry=https://codeberg.org/api/packages/automa-tan/npm/ harnessdelta@0.1.1 <repository> <git-range> --json
19   ```
20 
213. Summarize critical and high findings first, then medium findings that change instruction scope or discoverable components.
224. Preserve the reported path and finding code. Do not reconstruct or quote redacted commands, endpoints, permission values, credentials, or instruction contents.
235. Explain that the review is static and version-agnostic. Confirm runtime behavior with the harness's own status commands when that distinction matters.
246. For a shareable artifact, replace `--json` with `--markdown`. For CI, add `--check`; exit 2 means a high or critical finding, while exit 1 means the review could not run.
25 
26Source and documentation: https://codeberg.org/automa-tan/harnessdelta
27