ClonePeek v0.2.0 — Agent Hooks & Honest Evaluation

ClonePeek v0.2.0 — Agent hook visibility and explicit evaluation

Know what runs before you open a repo. Read-only, local configuration scanning with no runtime Python dependencies.

What's new

• Claude Code hooks: inspect .claude/settings.json for type: command hooks, show lifecycle events, matchers and accurate source lines.
• Windows command heuristics: recognize PowerShell encoded commands, iwr | iex, and recursive Remove-Item usage.
• Safer static reads: bound actual file bytes, avoid symlink following where supported, refuse named pipes/devices, reject deeply nested JSON, and surface unreadable directories.
• Honest evaluation: 12 reproducible, hand-labeled synthetic fixtures; 6 TP, 1 FP, 1 FN, 4 TN at a high-risk cutoff (85.7% precision/recall on these examples only). A legitimate folder-open task is a false positive and indirect script execution remains a false negative.
• More verification: 27 unit tests, 6-platform/Python CI matrix, demonstration fixtures, and documented limitations.

Install and run

uvx --from git+https://github.com/harshitethic/clonepeek@v0.2.0 clonepeek /path/to/repo

Examples: clonepeek examples/suspicious and clonepeek examples/claude-hooks in a source checkout.

Review before trusting

ClonePeek is a static review aid, not proof that a repository is safe or malware-free. It does not execute repository-defined commands, install dependencies, send telemetry, follow transitive shell-script chains, or comprehensively model every development tool. Outputs may contain source snippets, so do not publish reports without review. See threat model and synthetic evaluation.

Contribute

We welcome safe fixtures, real reproducible false positives and missed execution paths. See issue #1 and issue #2. Do not report private credentials in public issues.