Skip to main content

Redaction & secret safety

When cprof snapshots your setup it tries to replace every secret with an ${env:NAME} placeholder, so the resulting claude-profile.json is safe to move. Redaction runs fully offline, in three layers.

Redaction is best-effort. It catches the common, well-shaped secrets — not everything. Always read a generated profile before you share it.

How it works​

A value is redacted if any layer flags it:

  1. Provider keys — every value is checked by secretlint's recommended ruleset (GitHub, Anthropic, OpenAI, Slack, Stripe, GCP, and many more), in-process and offline.
  2. Sensitive key names — values under keys like apiKey, token, password, secret, credential, or authorization. Both camelCase and UPPER_SNAKE_CASE are recognized (dbPassword, AWS_SECRET_ACCESS_KEY).
  3. JWTs and high-entropy values — JSON Web Tokens, plus long random-looking strings (length ≥ 32, mixed character classes, high Shannon entropy).

Each flagged value becomes ${env:NAME}, where NAME is derived from its key, and is added to the profile's secrets.required list.

Existing ${VAR} references are preserved​

If a value already uses shell-style expansion — ${API_KEY}, or Bearer ${TOKEN} — cprof leaves it intact. Those are references, not raw secrets, so they survive the snapshot unchanged.

Secrets inside URLs​

Within a URL, query parameters such as token, access_token, api_key, key, secret, or password are rewritten in place — https://host/path?token=abc123 becomes https://host/path?token=${env:TOKEN} — while the host and path are kept.

The re-scan that guards every write​

Redaction alone isn't trusted. Before init or refresh writes anything, cprof re-scans the generated profile and every bundled file with an independent engine. If that scan finds a secret that slipped through, cprof refuses to write and exits with code 3 rather than emit a leaky profile.

Limits — what cprof does not catch​

These are the cases redaction will miss. Read every profile with them in mind:

  • Low-entropy secrets under ordinary keys. A short password or token stored under a non-sensitive key (say note or value) won't trip any layer.
  • Bare AWS access-key IDs (AKIA…). These are treated as identifiers, not credentials, so they're left as-is. (The AWS secret access key is caught.)
  • Anything below the entropy thresholds, plus values that look like URLs, file paths, content hashes (sha256:…), or UUIDs — deliberately skipped to avoid false positives.

Reports and diffs never print raw secret values — only the path and the chosen placeholder name. But the guarantee is "best-effort, then verified," not "provably secret-free." When in doubt, treat a profile like source code: review it before it leaves your machine, and use .cprofignore to exclude anything cprof should never read.