Skip to content

Known limitations

  • One unlock per vault, through the agent daemon. Commands on a registered vault start the agent daemon when none answers (unless VAULTFOLD_NO_AGENT is set or agent.autostart is off) and get their keys from it: the first command after a lock asks for the passphrase once (Passphrase for identity "<name>": , from the terminal or VAULTFOLD_PASSPHRASE_FD), and later commands ask nothing until the vault locks. run and inject have the daemon resolve references, so no key reaches them. In an agent session nothing prompts: a request that needs a human exits 3 saying to run vaultfold agent approve outside the agent session. Without a daemon (machine identity mode, auto-start off, no runtime directory, or a failed start) commands unlock in-process with a passphrase prompt per command, as before (ADR 0020), after a one-time notice per terminal. init, adding a hosted vault and key rotate always unlock in-process.
  • Identity protection. Passphrase and OS keychain identities (see "Identity protection" above). Hardware-resident device keys (a p256 identity key that never leaves the Secure Enclave, TPM or a PIV token) and passkeys are after 1.0 (docs/ROADMAP.md, "Later"); machine identities (VAULTFOLD_IDENTITY) are refused by vault commands until machine-identities (milestone 2). The macOS helper app's release signing, and the daemon's trust in project-signed builds, wait for a Developer ID and an Authenticode certificate: until then the daemon trusts only its own binary by default.
  • Agent daemon. No remote transports yet (agent pair is not available: SSH-forwarded sockets reach the daemon as an unknown program), no MCP server or harness hooks, and the desktop app that would start the daemon at login and offer richer approvals does not exist yet. See docs/threat-model.md, "Same-user limits of the agent daemon".
  • Backends. Git repositories and plain folders only. s3://, dav:// and davs:// are after 1.0 (docs/ROADMAP.md, "Later") and hosted locations are milestone 2; all are recognised and refused as not supported in this build.
  • Clipboard. get --copy uses osascript on macOS (with the concealed pasteboard type) and user32 on Windows (excluded from clipboard history and cloud sync). On Linux it needs wl-copy/wl-paste (Wayland), xclip or xsel on PATH, and no concealed hint is available.
  • Masking. run and inject mask values of at least 8 bytes (--mask-min) in the child's output; encoded forms (base64, URL encoding) are not masked, and a masked child sees pipes rather than a terminal. Use --fd for the most sensitive values.
  • Windows. run --fd is refused (no numbered file descriptors to inherit), run always stays the parent process rather than replacing itself, file modes such as 0600 are advisory (files are created in per-user directories), and inject --to-file files inherit the directory's ACL. The Windows code paths are cross-compiled and vetted by make crossbuild, but while CI is disabled nothing runs the test suite on a Windows host (make check runs natively and make check-in-docker on Linux). This includes the whole Windows side of the agent daemon (Windows Hello, DPAPI, the message box, lock triggers, the named pipe's permission check and terminal identification): it has not yet run on Windows hardware.
  • Other unixes. Paths resolve as on Linux, but the per-vault lock is not implemented on solaris, illumos or aix.
  • Policy chain. Privilege writes need the backend and are refused offline. In a shared-folder (filesystem) vault, a mint or grant that lost its policy slot leaves its wrap files behind in every copy. They block nothing: each wrap lives at epochs/<collection>/<epoch>/<link>/<keyid>.wrap, where <link> is the first 16 hex characters of the id of the policy link that calls for it (ADR 0052), so the next epoch new, compromised device revocation, member revocation or grant, by the losing admin or anyone else, writes under its own directory and succeeds (CLI regression tests cover each). The leftover files are strays, read by nothing except sync's re-apply of the losing device's own data written under its lost mint, and doctor reports each as stray-wrap (see "doctor"). A member with write access to the folder can still stop mints by other means, such as a file planted where an epoch's directory should be (doctor reports it as corrupt-file; remove it), and a mint whose wraps fail to write after its record leaves an epoch nobody can read (an open defect, ticket 01M7FYRJT2KQC73Z8Z003NB5PM). The shared-folder tests have not run on Windows. There is no command yet to register or rotate a recovery key, and the CI verification action is after 1.0 (docs/ROADMAP.md, "Later").
  • Out of scope for now. Everything marked "(not yet available)" above, cross-collection mv (refused with the copy, rotate and remove path), and a mask_min profile setting.