38 38 Bound learned exact values and restorable placeholder mappings to 8192 each.
39 39 Bound deep traversal by object identity and depth 64.
40 40 On learned-value cap overflow, stop learning new dotenv/env exact values and warn once.
41 -On restorable-mapping cap overflow, still redact provider-bound secrets with non-restorable placeholders, warn once, and keep existing mappings stable.
41 +Phase 1 retains the volatile HMAC12 mapping format.
42 +Durable V1 mappings are deferred to Phase 2.
43 +Until Phase 2 lands, mapping-cap behavior remains unchanged.
43 45 `leaks` must tolerate faulty user input.
44 46 The only planned user input is the `regex` setting.
266 268 URL usernames are metadata, not secrets.
267 269 URL passwords are secrets.
268 270 Query parameter names are metadata.
271 +Sensitive query names use separator-aware complete tokens, never substring containment.
269 272 Sensitive query parameter values are secrets and must be redacted in place.
270 273 HTTP header names are metadata.
271 274 Sensitive header values are secrets and must be redacted in place.
272 275 Cookie names are metadata.
276 +Sensitive cookie names use separator-aware complete tokens, never substring containment.
273 277 Sensitive cookie values are secrets and must be redacted in place.
274 278 Structured config formats are not parsed generically.
275 279 There are too many formats to support cleanly.
278 282 Shell commands are not parsed as shell.
279 283 Header, query-param, URL, and assignment detectors must catch common secret components inside shell text.
280 284 Do not redact whole shell commands just because they contain a secret.
281 -JWTs are redacted as whole tokens.
282 -JWT header/payload decoding is out of scope.
285 +JWTs are redacted as whole tokens only when they have exactly three strict Base64URL segments, header and payload decode as non-array JSON objects, header `alg` is a nonempty string, and signature is nonempty strict Base64URL.
286 +JWT signature verification is out of scope.
283 287 Private key material is redacted as whole logical secrets.
284 288 Covered private-key shapes include PEM/OpenSSH/PGP private key blocks, encrypted cosign key blocks, PuTTY PPK private sections, age secret keys, and minisign secret keys.
289 +PuTTY private sections consume exactly the declared `Private-Lines` payload and never unrelated trailing text.
285 290 Boundary/comment metadata becomes placeholder tags instead of being preserved literally.
286 291 Private key armors are redacted as whole secret values; literal boundaries are not preserved.
287 292 Public keys and public certificates are metadata, not secrets.
288 293 Do not redact SSH public keys or certificate blocks unless a private-key-looking block is present.
289 294 URLs without credentials are not secret just because their key is `DATABASE_URL`, `REDIS_URL`, or `MONGO_URL`.
291 -Sensitive key matching is case-insensitive, treats `_` and `-` as equivalent separators, and allows numeric suffixes such as `_1`.
296 +Sensitive key matching is case-insensitive, treats `_` and `-` as equivalent separators, and uses complete normalized tokens rather than arbitrary substrings.
297 +High-precision names such as `API_KEY`, `ACCESS_TOKEN`, `CLIENT_SECRET`, `PASSWORD`, `PASSPHRASE`, and `PRIVATE_KEY` authorize non-weak exact learning.
298 +Ambiguous names such as `KEY`, `AUTH`, `SESSION`, `COOKIE`, `JWT`, `SEED`, and database URL names require a parsed secret component or provider-specific value shape.
299 +Metadata suffixes such as `_ID`, `_NAME`, `_PATH`, `_FILE`, `_TYPE`, `_HOST`, `_PORT`, `_MANAGER`, `_SOCKET`, and `_ENDPOINT` do not authorize learning.
293 301 Sensitive key examples: