Secrets and limits

Secrets

• vraelis models records file names, sizes and digests only. A file named like a secrets file (.env, keys, id_rsa, .npmrc, ...) is recorded as present only: never read, digested or measured, so its content is not checked.
• vraelis agents never records environment or header values unless you pin them. Pinned values are stored in plain text, by your choice.
• Some agent files are recorded as structure only (keys, booleans, numbers, env and header names, and values you pinned): .claude/settings.local.json, user-level --config files, and any project config or settings file that the project's own ignore rules exclude.
• Ignore rules are read from .git/info/exclude and the .gitignore files inside the project, following git's own matching rules. A file counts as ignored if git would ignore it with either case-sensitive or case-insensitive matching, and any pattern it cannot parse makes the file count as private.
• In shared project files, credentials it recognises are redacted on a best-effort basis. Read a lockfile before you commit or share it.

What a result does not show

• That a model, add-on or server is safe, correct or behaves as intended.
• Which model a device or server actually loads, or which tools a server exposes at run time. Nothing is launched.
• Who published the files, unless you run the optional signature check against a signer you chose.
• Changes inside packages fetched at launch (npx, uvx, docker). Floating or unpinned packages are flagged.
• Changes in files a launched script itself imports or loads.

How it was tested

Testing was self-run, not independent, and no outside user has run it yet. The model checks were tried on three blind sets of real broken releases; the most recent matched 23 of 24 predictions and detected 5 of 22 runnable defects, with no false alarms on 15 healthy models. It is a narrow structural gate: most real breakage (wrong config values, templates, tokenizers) is outside what it checks. The agent checks were tried on four blind sets of real configuration changes from 61 public repositories; on the most recent, 19 of 22 changes were detected with no false alarms.

Other tools do overlapping jobs, several of them well; the README lists them.