Answers the operational question 'can a machine try all three boot
binaries in one go?' The protocol can't carry three NBPs in one cycle
(one boot file per DHCP round, the Secure-Boot refusal happens after
handoff with no error report, and the broken-NIC case specifically needs
the firmware itself to load builtin-driver iPXE — GRUB's network rides
the same broken firmware stack). What we CAN do is make the walk a
once-per-machine-ever event and give operators a way to skip it:
- Learned driver modes persist (<work_dir>/driver_modes.json). A MAC
that reaches the Shim rung, or confirms an iPXE handoff at Builtin,
is pinned to disk: immune to the 30-min TTL, reloaded at startup.
The file only carries exceptions — a healthy fleet never writes it.
Corrupt file starts empty (standard crash-cache policy).
- Boot rules gain an optional driver_mode pin (auto/firmware/builtin/
shim), consulted by the DHCP proxy BEFORE the escalation ladder:
'this OUI is a Secure Boot rack -> serve shim immediately' = zero
failed cycles. Mode-only rules coexist with target rules (a pin
doesn't shadow a later target match). Editor column on Hosts tab.
- grub.cfg now tries to chainload all-drivers iPXE before showing the
signed menu: with SB off the chainload succeeds and the client gets
the full iPXE feature set back in the SAME boot (self-healing for
mis-escalations, and the handoff then pins the working mode); with
SB on, shim's verifier refuses it inline — no reboot — and the
signed menu appears.
DhcpProxyServer now takes the escalation table + rules store from main
(persistence path comes from the configured work dir).
Validation: clippy clean, fmt clean, 299 workspace tests green (+9:
persistence round-trip across restart, Shim pin survives TTL, learned
Builtin survives TTL, corrupt-file recovery, default-mode-never-
persisted, rule-pin matching incl. unknown-mode tolerance and
pin/target coexistence, GRUB chainload-before-menu ordering, API
round-trip of the driver_mode field).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Three features, all zero-toggle and principle-clean (single static musl
binary, container-first, no test certs, no client trust-store changes).
Secure Boot via signed shim+GRUB (automatic):
- The v0.6.1 escalation ladder gains a third rung: Firmware -> Builtin
-> Shim. Secure-Boot firmware downloads our unsigned iPXE but refuses
to execute it — indistinguishable from a failed chainload — so after
two unconfirmed attempts the MAC is offered Fedora's Microsoft-signed
shimx64.efi, which loads the signed GRUB, which fetches a
server-rendered grub.cfg. Fully signed chain, SB stays on.
- scripts/fetch-shim.sh pulls shim-x64/grub2-efi-x64 (+aa64 best-effort)
from the official Fedora 43 packages and ships the EFI binaries
byte-for-byte unmodified; Dockerfile fetch stage gained rpm2cpio/cpio.
- New grub_script renderer (Linux kernel entries only — signed GRUB only
boots signed kernels; sanboot/wimboot have no signed equivalent and
are omitted with an explanatory menu line).
- TFTP server gains a DynamicAsset hook for server-rendered names
(grub.cfg); HTTP serves the same config under /ipxe/grub.cfg for
native UEFI HTTP Boot chains. Arch-aware fallback walks back down the
ladder where no shim exists (BIOS, IA32).
Boot rules + decision webhook (open 'Matrix Boot'):
- Ordered first-match-wins rules over MAC prefix + client arch (the DHCP
proxy now bakes arch into the boot.ipxe chain URL), generalizing
per-MAC pins. Persisted to boot_rules.json; GET/PUT /api/boot-rules;
rules editor + webhook field on the Hosts tab.
- Optional pixiecore-style webhook: unmatched boots GET
<url>?mac=&arch= and 200 {"target":"id"} chains to it. Fail-open
with a 2s budget — a dead endpoint can never block PXE.
- Decision order: exact pin -> rules -> webhook -> menu. Empty config
is byte-for-byte the previous behavior.
Tokenized answer files (the post-WDS/CVE-2026-0386 hardening):
- Every generated unattended URL (inst.ks / preseed url / autoinstall
seed) now carries a 4h boot-scoped token; /unattended/{id} and the
cloud-init seed routes require it (or an operator session) once an
admin exists. Stops answer-file credential harvesting by anything
else on the network. No toggle; setup-mode installs stay open.
Validation: clippy clean, fmt clean, 290 workspace tests green
(+18 new across boot_tokens, boot_rules, arch ladder, escalation,
grub renderer, and four new full-flow integration tests).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
