TheMaxMur/RS-Key v0.2.4

RS-Key v0.2.4

on GitHub

[0.2.4] — 2026-06-19

Added

• The rsk CLI can run without Nix. A tools/pyproject.toml packages the
  CLI so it installs from any Python ≥ 3.9 toolchain —
  uvx --from ./tools rsk …, uv tool install ./tools, pipx install ./tools,
  or plain pip. The Nix dev shell stays the primary, pinned path; this mirrors
  its CLI runtime deps (hidapi, cryptography, pyscard, fido2,
  mnemonic, shamir-mnemonic) for hosts without Nix. See
  tools/README.md. Host-tool only; no bcdDevice bump.

Changed

• FIDO2 PIN entry is now uniform across the CLI. Commands disagreed on how
  to take a PIN: most accepted only --pin (and aborted on a PIN-protected
  device when it was omitted), while fido list-passkeys and fido set-pin
  prompted interactively with no flag at all. Every PIN-gated command (backup export/restore, audit log/verify, lock enable/disable, inventory verify, fido list-passkeys/set-pin/attestation import/clear) now
  accepts the PIN either way — --pin flag or an interactive prompt —
  through one chokepoint (rsk.common.resolve_pin) that only prompts when the
  device actually has a PIN, so touch-only devices are never asked. Host-tool
  only; no bcdDevice bump.
• The rsk-tui cockpit now routes PIN entry through one chokepoint too. Its
  four per-action PIN steps collapsed into a single App::gate_pin +
  Step::PinThenRun, so "prompt for the FIDO2 PIN iff the device has one, else
  run" lives in exactly one place (mirroring the CLI's resolve_pin). PIN-vs-
  phrase collection in the modal flow is now explicit instead of a catch-all (a
  stray text input can no longer land in the PIN buffer), and the four
  device requires a PIN strings were unified. No behaviour change for users;
  host-tool only, no bcdDevice bump.

Fixed

• rsk secure-boot no longer refuses provisioning on a chip with a benign
  LOCK_NS. pages_locked() read the whole OTP lock row, so a pre-set
  non-secure-page lock (LOCK_NS=1, 0x040404) looked like a bootloader lock
  and wrongly blocked load-key; it now masks LOCK_BL specifically. Host-tool
  only; a mutation-proven regression test was added.

Security

• Transparency-log monitoring for our release signing identity. A scheduled
  GitHub Action (sigstore/rekor-monitor) watches the Rekor log for entries
  signed under our release workflow's OIDC identity, so illegitimate use of it —
  a signature we did not produce — becomes detectable, complementing the SLSA
  Build L3 provenance. CI only; see docs/supply-chain.md.
• OATH credential secrets are now sealed at rest. Every other applet
  (FIDO, PIV, OpenPGP, rescue) AES-encrypts its keys before they reach flash;
  OATH alone stored its TOTP/HOTP shared secrets — and the SET CODE key — as
  plaintext TLV. They are now AES-256-GCM-sealed under the device kbase
  (HKDF(serial_hash, kbase, "OATH/KEYS")), the same device-seal the PIV slot
  keys use. A one-time boot migration re-seals any credential enrolled before
  this release, so existing accounts keep working. With the OTP MKEK burned, an
  extracted flash image no longer reveals OATH secrets. bcdDevice 0x0765 →
  0x0766.
• The at-rest seal path is now enforced by types, not convention. A slot
  that holds a sealed secret is a KeyFid, distinct from a plaintext u16 file
  id, and the only writer that accepts one is Fs::put_key(KeyFid, Sealed) —
  where Sealed is produced only by a seal routine. A stray
  fs.put(key_fid, raw_secret) no longer compiles (asserted by a compile_fail
  doctest). This is the chokepoint whose absence let OATH ship its secrets in
  the clear; every applet's key FIDs were moved onto it.
• Resident-credential RP domains are now boxed at rest. A discoverable
  credential's EF_RP record stored the relying-party id (the site's domain)
  in cleartext, so a flash dump revealed the list of sites you hold passkeys
  for — a privacy leak, even though the keys themselves were sealed. The domain
  is now ChaCha20-Poly1305-boxed under the device seed (the same seal the
  credential body uses), with the rpId hash kept in cleartext as the O(1)
  lookup key. A boot migration re-boxes records enrolled before this release.
  Honest residual: the rpId hash remains, so a dump can still dictionary-attack
  guessable domains — but the plaintext site list is gone. bcdDevice 0x0766
  → 0x0767.

Verify a download: https://github.com/TheMaxMur/RS-Key/blob/v0.2.4/docs/supply-chain.md