Only the latest commit on master is supported. There is no published PyPI release yet (see ROADMAP.md), so there is no older version line to backport fixes to.

Be precise about what this software touches, because "security" means different things for different kinds of tools:

• orchestrate_kit runs entirely locally. No network calls, no telemetry, no external services. RepoContext.run() explicitly clears known API-key environment variables before running a subprocess, so an audit cannot accidentally depend on (or leak into) live credentials.
• The evaluator executes commands you configure — a plugin's audits() can shell out (e.g. to run python code/main.py). Only run orchestrate evaluate / orchestrate certify / orchestrate release against a repository you trust, the same way you'd only run make or pytest against one. This is not sandboxed, and isn't intended to be — it's a developer tool, not a hosted service accepting untrusted input.
• The judge and mentor accept free-text input but never execute it, call a model, or reach the network. Worst case for malformed input is a wrong score or a crash, not code execution.
• Skills in skills/ are Markdown files an external agent (Claude Code, Cursor, etc.) reads. This repo doesn't control what that agent does with them; review any skill's content the same way you'd review a shell snippet before running it.

If you find something that lets an untrusted input execute code, exfiltrate data, or otherwise misbehave outside the "you're auditing a repo you trust" model above:

1. Do not open a public issue.
2. Use GitHub's private vulnerability reporting: Security → Report a vulnerability on this repository, or contact the maintainer directly through their GitHub profile.
3. Include a reproduction: the exact command, the input, and what happened versus what should have happened.

Expect an acknowledgment within a few days. This is a single-maintainer project without a formal SLA — see SUPPORT.md for what that means in practice.

• A Finding with a wrong severity or a bad heuristic match is a bug, not a security issue — open a normal issue.
• orchestrate evaluate refusing to run, or an audit crashing, is by design reported as a finding (SimpleAudit.run() catches and reports rather than propagating) — that's the intended failure mode, not a vulnerability.