Backlog¶

Open work items beyond the current release. Items are grouped by theme; ordering within a section is rough priority (highest first) but most items are independent. The larger Phase 10 themes are on the roadmap. Section numbers are referenced from code comments, so they stay stable; finished items stay with their outcome.

1. Audit-driven hardening (highest priority)¶

Real defense-in-depth wins deferred because the original threat model didn't make them blocking.

1.1 M&D-backed PIN counters for OpenPGP PW1 / PW3 / RC — ✅ DONE (Phase 8 M4)¶

Phase 8 moved the OpenPGP PIN counters onto MAC-and-Destroy (561e040). Phase 9 M4 replaced that layout: all PINs are now M&D kernel v2 bindings in mirrored record pairs, with no software PIN check (current layout in architecture). Hardware check: tools/validate-pin-md.sh.

1.2 M&D-KEK wrap for X25519 dec priv key — ✅ DONE (Phase 9 M5)¶

Done differently from the sketch below: KEK = HMAC(k, label), k from the PW1 M&D binding; AES-256-GCM; KEK cached while PW1 is verified. See security model. Original note (historical):

The X25519 dec key for gpg --decrypt lives in R-mem slot 1 byte 180-211 (Trezor Safe 7 pattern). Anyone with SH0 + L3 session can read it. Wrap it with a PW1-derived KEK released only via M&D — mirrors Trezor's PIN-encrypted-seed design.

Implementation sketch (~200 LOC, reuses landed M&D framework):

KEK = HMAC(PW1-hash, M&D-derived-secret) — pin_md_kernel already exposes the M&D-derived-secret hook needed for this.
Stored ciphertext = AES-GCM(dec_priv, KEK)
On PSO:DEC: verify PW1 (consumes M&D slot via §1.1's landed wrappers), derive KEK, decrypt priv, perform ECDH, zeroize all on return.

§1.1's framework already landed in Phase 8 M4, so this is unblocked.

1.3 cpp-reviewer audit followups¶

The Phase 6 cpp-reviewer audit caught 2 MEDIUMs that were tagged as deferred. Re-read the audit report (in archive/phases-1-7 branch STATUS.md entry) and address what still applies post-Phase-7.

2. WebAuthn / FIDO2 polish¶

2.1 credProps extension — ✅ DONE (Phase 8 M2)¶

Landed in commit bcff8fb. CTAP2 MakeCredential now parses incoming extensions.credProps request and emits the rk flag in the attestation extension map. webauthn.io shows device-bound credential of unknown discoverability no more.

2.2 hidraw udev rule for FIDO — ✅ DONE (Phase 8 M1)¶

Landed in commit bcff8fb. nixos/tropic.nix now installs a udev rule tagging cafe:4001 with ID_SECURITY_TOKEN=1 + group access to plugdev, so libfido2 / browsers see the dongle without a manual rule.

2.3 ES256 (P-256) credentials — DONE (Phase 9 M3b, 2026-10-15)¶

Done: GetInfo lists EdDSA and ES256, and MakeCredential takes the first the site offers. P-256 keys are generated and used inside TROPIC01 (COSE EC2 keys, DER signatures, the algorithm stored in the slot metadata, one shared 29-credential cap). RS256 stays unsupported.

2.4 Claim FIDO_2_1¶

Needs hmac-secret, credProtect (3.2) and PIN/UV auth protocol 2 (HKDF, AES-CBC with random IV, full HMAC) — CTAP 2.1 §9.

2.5 Store the RP ID text¶

credMgmt shows each RP as the hex of its hash: only rpIdHash is stored.

3. Configurable policies (Yubikey-class)¶

3.1 authenticatorConfig (CTAP2.1 command 0x0D) — ✅ DONE (Phase 9 M3, toggleAlwaysUv)¶

fido2-token -S -u / -D -u. Not implemented: setMinPINLength, enableEnterpriseAttestation (both optional).

3.2 credProtect per-credential UV¶

Per-credential UV requirement (CTAP2.1 §11.1.2). Some RPs (Google, Microsoft) request credProtect: 3 for high-value credentials.

3.3 Token permissions — ✅ DONE (Phase 9 M3)¶

getPinUvAuthTokenUsingPinWithPermissions (0x09), pin_token.c. PIN/UV auth protocol 2 is still open (see 2.4).

3.4 OpenPGP touch policy — ✅ PARTIAL (Phase 8 M3, GLOBAL only)¶

Landed in commit 98486f4 + bcff8fb. Global on/off touch toggle implemented and enforced in PSO:CDS, PSO:DEC, INTERNAL_AUTHENTICATE. Defaults to ON (user-presence required for every signing op).

Still TODO — per-slot policy + cached mode: Yubikey-style per-key granularity (sig / dec / aut independent) and the "cached (1 touch per session)" mode. Maps to OpenPGP DOs D6/D7/D8 properly. Current impl is one boolean for all 3 slots.

4. OpenPGP card polish¶

4.1 gpg --card-status display anomaly verification¶

Post-flash, run gpg-connect-agent "SCD RESTART" then gpg --debug 0x0009 --card-status to capture the trace. Confirm whether "Max. PIN lengths: 0 0 0" and "PIN retry: 0 0 0" actually resolved post-flash (expected: yes, because the schema bump forced clean defaults). If anomalies persist, file an upstream gnupg bug with the captured trace.

4.2 AID manufacturer ID cosmetic¶

gpg --card-status shows "Manufacturer: unknown" because 0x4E58 ("NX") isn't in gnupg's hardcoded table. Options:

Keep 0x4E58 — accept cosmetic display
Swap to 0xFF02 for spec-defined "unmanaged S/N range" label
Patch upstream gnupg to recognize 0x4E58 as "nixtropic"

Option 2 is one-line change but invalidates existing card state and requires TERMINATE+ACTIVATE.

4.3 Default language pref¶

write_activated_defaults zeroes OFF_LANG. Could optionally write 'e' 'n' so gpg --card-status shows "Language: en" out of the box. Trivial.

4.4 PIV applet¶

A second smartcard applet (alongside OpenPGP) adding pkcs11 + Windows smartcard logon use cases. firmware/src/ccid/apdu_dispatch.c is structured to accept a second applet — just needs AID-based routing wired in.

4.5 OpenPGP attestation slot (Yubikey-compat)¶

Yubikey 5.2.3+ adds a vendor extension: an extra "attestation" key slot beyond the standard 3 (sig/dec/aut). Verified at runtime by ykman openpgp keys attest <slot> — the device signs a statement that the user-facing key was generated on-device (not imported), and the resulting X.509 attestation is rooted in a Yubico-issued CA cert. Useful for compliance scenarios that must distinguish HW-generated keys from imported ones.

What it needs on our side: - One more TROPIC01 ECC slot (we have headroom — slot 28 or repurpose one FIDO slot, since 29 FIDO creds is generous). - An on-device-burned attestation CA (Ed25519 key + cert), or per-device unique attestation key signed at provisioning. - Vendor APDU INS 0xF1 (Yubikey-compat) returning the X.509 cert chain. - ykman / gpg --card-edit interop testing.

References: - https://developers.yubico.com/PGP/Attestation.html - Yubikey docs: https://docs.yubico.com/hardware/yubikey/yk-tech-manual/yk5-apps.html

Cost: ~1 KB flash + 1 ECC slot + one-time provisioning step.

5. Publishing / Nixpkgs upstreaming¶

5.1 Nixpkgs upstream PR¶

Package metadata: firmware as fixed-output, lt-util as a library, NixOS module.

5.2 pid.codes VID:PID allocation¶

Currently using cafe:4001 (TinyUSB demo VID:PID). For public release get a pid.codes allocation (free for open hardware), then submit upstream PR to libccid to add the new VID:PID to its known-readers list — eliminates the runtime Info.plist patch.

5.3 TROPIC01 TRNG SP 800-90B compliance — vendor claim found (v0.3.0)¶

The datasheet and ODN_TR01_app_008 state NIST SP 800-90B and AIS 31 compliance for the chip's two TRNGs. That is a vendor claim, not a certification of this device or of nixtropic's entropy path; reopen only for an independent assessment. The RNGTEST_DIS erratum (2026-09-10) is unresolved, see the security model.

5.4 Rust CLI¶

Direct dongle control beyond what lt-util provides — credential enumeration, FIDO reset, OpenPGP card init, validate orchestration.

6. Bug followups + maintenance¶

6.1 picocom termios hang on TinyUSB CDC¶

SET_LINE_CODING race when picocom opens our CDC interface. Workaround documented; underlying fix in TinyUSB upstream pending.

6.2 Validate scripts: substring matching for `--reader`¶

Older opensc-tool versions don't support substring matching on --reader nixtropic; current scripts pin --reader 0 to avoid this. Detect opensc-tool version + fall back to substring when supported.

6.3 Windows compatibility pass¶

Runtime use (FIDO2, OpenPGP, CDC) should auto-bind via in-box Win10/11 class drivers. Needs: scdaemon.conf docs (same disable-ccid + pcsc-shared as Linux, via Gpg4win), DFU flashing recipe (Zadig + dfu-util OR STM32CubeProgrammer), README "Windows setup" section, HW-in-the-loop verification on Win10 22H2 + Win11. No firmware changes expected. Phase 9 M1 changed the release layout (HID + CCID only, device class 0, no IAD). Re-verified 18 Oct 2026 on Windows 11: enumeration and the OpenPGP card work; FIDO2 with a PIN needs 6.7.

The three # banner lines are printed before the host configures the device; TinyUSB clears the CDC TX FIFO when the interface opens, so only the later init log reaches /dev/ttyACM0. Debug build only. Fix if ever needed: print the banner from tud_mount_cb, or after boot.

6.5 README / docs still describe the v0.2 composite layout — ✅ DONE (v0.3.0)¶

README and docs/UBUNTU.md mentioned the CDC + lt-rpc interfaces as if always present, and Force-UV was toggled with lt_rpc.py, which the release build no longer has (M3: fido2-token -S -u / -D -u). Fixed: the documentation was rewritten for v0.3.0 (card setup, flashing).

6.6 Requests interleave during a touch wait — ✅ DONE (Phase 9 M3, design decision 18)¶

user_presence_check() pumps tud_task() while it waits for SW1. CTAPHID refuses other FIDO requests then, but a CCID APDU is processed right away, so OpenPGP chip operations can run between the chip calls of a waiting FIDO request, and FIDO between OpenPGP's (openpgp_applet.c also waits for touch). Not inside a libtropic call, so the L3 session logic (M2) is not affected. Define a dispatch policy covering both directions, nested touch waits, the shared consent state, and the debug lt-rpc. Found by Jan in the M2 review; plan it with the M3 FIDO work.

Fixed: requests run from the main loop, never inside a USB callback; one request and one touch wait at a time.

6.7 CTAPHID KEEPALIVE and reply backpressure — ✅ DONE (v0.3.1)¶

Windows abandoned silent requests (getPinUvAuthToken after 2.6 s, changePIN after 6.6 s; usbmon, 18 Oct 2026). v0.3.1 sends KEEPALIVE from a progress function called before every chip call and in the touch-wait loop, answers INIT resync and CANCEL during a request, queues one-packet replies when the IN endpoint is busy, and holds the debug lt-rpc while FIDO2 runs. Design: docs/history/V0.3.1-KEEPALIVE-DESIGN.md. Verified on Windows 11 with captures (19 Oct 2026; longest silence 240 ms) and in the testing timings. Left over: the PIN operations and Reset reach ~250 ms because the software P-256 steps cannot be interrupted (CTAP's SHOULD is 100 ms). The precomputed P-256 table would speed them up but needs about 74 KB of flash, more than the headroom (§7). The alwaysUv question is 6.10. The Protocols section is in references.

6.8 Retire superseded test suites¶

fido2_test.py validate-phase6-m2/m3 assume lt-rpc Force-UV, legacy-token credMgmt and 32 slots; tools/validate_fido_m3.py replaces them (M6 cleanup).

6.9 Phase 9 M4 follow-ups¶

A host that never reads a stale CCID answer still stalls CCID (Jan, f05e63b); libccid does read it. Revisit if another PC/SC stack shows it.
FIDO boot after a transient chip error: FIDO refuses until power-up (OpenPGP retries its boot up to 3 times per power cycle).
OpenPGP VERIFY ≈ 1.4 s: record-pair writes dominate. Moving the signature counter to a chip monotonic counter removes one O3 write per signature.
validate-openpgp.sh needs pcsc_scan: run it as nix run .#validate-openpgp (the dev shell lacks it).

6.10 alwaysUv after the first PIN — decision open¶

Setting the first FIDO2 PIN after a reset also turns on alwaysUv (fido_pin_set → slots_set_pin_bound(1, 1), Phase 6 design). That is why GetInfo reported alwaysUv on after the 18 Oct Windows session: Windows Settings reset the key and set a new PIN. Options: keep it (documented in Windows), or leave alwaysUv alone on the first PIN so the site decides. A security-behaviour change, so not in a bug-fix release; the owner decides.

6.11 OpenPGP card spec details (review, 20 Oct 2026)¶

Found while checking the PIN and touch policy against OpenPGP card 3.4.1. None is a security hole; deferred until after FOSSCOMM.

Extended capabilities (C0) advertise GET CHALLENGE and key import; neither is implemented (both answer 6D00). Clear the two bits.
A permanent touch policy accepts a repeated write of 02 (switching it off is refused); the spec requires every write to abort once the value is 02.
A refused or timed-out touch answers 6982; the spec recommends 6600 (a SHOULD).
Docs: "until unplug" should read "until the card is reset" (a CCID power cycle also clears it), and the touch-on / PIN-once presets are this project's choice: the spec's baseline is touch off and one signature per PIN.
(23 Oct 2026) The signature counter (DO 93) must reset to 000000 when a new signature key is generated (§7.2.14 p74) or imported (§4.4.3.12 p39). nixtropic resets it only on TERMINATE (openpgp_state.c, the counter is never cleared on GENERATE).

6.12 Check the chip's identity and firmware at boot (23 Oct 2026)¶

Today the firmware assumes a supported chip: it never checks the part number or the chip firmware, so an unsupported chip or old chip firmware shows up only as odd errors later. Two checks, if feasible:

Part number. tropic_chip_id_read() (lt_get_info_chip_id, Application mode) already exists, used only by the debug lt-rpc. At boot, compare the long P/N and silicon revision against the tested list (TR01-C2P-T101, TR01-C2P-T103, ACAB).
Chip firmware. The running App and SPECT versions are cheap Application-mode reads (lt_get_info_riscv_fw_ver / lt_get_info_spect_fw_ver); require 2.0.0 / 1.0.0 (or a range). The four-bank header check (tools/chip-fw-banks-main.c, app note 007 §3 steps 10–11) needs a maintenance reboot, so not on every boot — at most an on-demand command, if at all.

Open questions: what to do on a mismatch (refuse to start, or run and report — e.g. a status the host can read, the LED, or the debug console); boot-time cost of the reads; how to allow a newly tested part without a new build. A chip/SDK change: Jan reviews it.

7. Performance / flash budget¶

v0.3.2 release firmware.bin = 223 460 B / 256 KB (85 %), about 38 KB of headroom (the M5 decryption-key wrap cost about 1.3 KB; v0.3.1 KEEPALIVE about 0.2 KB).

PIV applet (§4.4): ~15-25 KB (also fits). Attestation slot (§4.5): ~1 KB.

If we exceed budget:

Drop USE_PRECOMPUTED_CP=0 (P-256 base point caching) — current setting saves 74 KB at the cost of ECDH speed; could trade some back.
Move trezor_crypto's nist256p1.c precomputed table to flash-stored blob with on-demand access (slower + more complex ECDH).
Strip unused libtropic functions via LTO + -fdata-sections.

Out of scope¶

These come up regularly but belong elsewhere:

Other TROPIC01 dev kits (RPi shield, Arduino shield) — this project is TS1302-specific.
RSA support — locked out by flash budget. Users who need RSA have other devices.
FIDO Alliance AAGUID registration — $25k/year not viable for an open-source project.