Files
felhom.eu/documentation/runbooks/RUNBOOK-escrow-ceremony.md
T

133 lines
9.1 KiB
Markdown

# RUNBOOK — Escrow ceremony (customer recovery-code R)
<!--
The zero-knowledge custody ceremony: seals the customer's unrecoverable secrets (PBS key K where
present; the IdentityBundle — restic offsite password, tunnel/PBS tokens, WG key) under a fresh
customer-held recovery code R. The hub stores only the opaque blob + a non-reversible password hash;
neither Felhom nor a hub compromise can open it. Cross-validated live: fork-4 validation (blob →
recover → decrypted the real repo) and the offsite arc (auto-confirm on hash match).
v0.127.0/v0.88.0 (TASK 2026-07-13): the PRIMARY path is the controller wizard — the customer runs
the ceremony themselves; the CLI remains the operator fallback. Mechanics validated by
audits/SPIKE-controller-escrow-2026-07-13.md (all five GO).
-->
## When a ceremony runs
1. **Onboarding** (with the PBS tier): seals K + the identity secrets.
2. **Offsite enable** (fork-4 rule): the restic repo password MUST enter a CURRENT R-blob before any
offsite backup runs — the controller stages the password to the agent at enable and the run-gate holds
`pending` until the hub-verified auto-confirm.
3. **Any new unrecoverable secret / suspected R exposure**: re-ceremony (R cannot be retro-fitted —
it is never retained).
4. **Stale-blob warning** (v0.127.0, Scenario F): the controller re-checks the hub blob's password
hash on every report ACK even in the `escrowed` state. If the CURRENT blob does not cover the
current offsite password (hash mismatch, or a hash-less superseding blob — e.g. a CLI ceremony
run without the staged secret), the Távoli mentés card shows "A letétben lévő helyreállítási
csomag nem fedi a jelenlegi távoli mentési jelszót" and offers a re-ceremony. Runs are NOT
blocked (offsite copies still protect against non-total loss); the fix is one wizard run.
## PRIMARY PATH — the controller wizard (customer-driveable, agent ≥ v0.88.0, controller ≥ v0.127.0)
The customer (or the operator in the customer's session) opens **Távoli mentés →
"Helyreállítási kód létrehozása"** (`/backup/escrow`) and clicks through:
1. **Preflight** — the wizard checks: PBS storage id configured, DR tier applied, `age` present,
hub upload target configured, the FELHOM_ESCROW sudo grant listed (list-mode probe), staged
secret (informational). Red rows block the start.
2. **Warnings** — the R-is-final copy; a re-ceremony additionally shows the supersede warning
("a korábbi kód érvényét veszti…").
3. **Re-auth** — the dashboard password, typed again (rides the login rate-limiter).
4. **Run** — the controller RE-STAGES the current offsite password first (when offsite is
configured; a staging failure ABORTS the start — a ceremony without the staged secret would
mint a hash-less blob), then triggers the agent's local-API job. The agent runs
`sudo -n felhom-agent --config /etc/felhom-agent/agent.json --selftest=escrow-create --upload
--output=json` — ONE fixed argv, byte-matched by sudoers (spike §2.2).
5. **Reveal** — R is delivered by a ONE-SHOT claim: displayed exactly once in the browser,
`Cache-Control: no-store`, never rendered server-side, never logged, wiped agent-side the
moment it is claimed. Unclaimed R expires after **10 minutes** → the ceremony is VOID and a
re-run supersedes it (nothing is lost — the blob is already on the hub; the old R simply never
existed for it).
6. **Typed-back** — the customer re-types two random words of R (client-side check), ticks the
"felírtam" confirmation, finishes.
Then the existing **auto-confirm** flips `pending → escrowed` hands-free within one report cycle
(hash match in the ACK) and wipes the staged copy — same as ever.
**Supersede/void semantics:** every new ceremony's blob REPLACES the hub blob; the previous R keeps
opening only the previous blob (pre-existing offsite history). An unclaimed/void R requires nothing
but a re-run. An agent restart mid-ceremony or pre-claim loses R safely (in-memory by design) — the
wizard reports it honestly; re-run.
### F1 threat model (operator ruling 2026-07-13 — carried verbatim into the product)
R is displayed in the customer's browser, which means it transits the Cloudflare tunnel ONCE at
the reveal moment (TLS end-to-end, `no-store`, one-shot). **This is an accepted, documented risk**
— the same trust class as the claim code and the login password, which already ride the same
channel: an adversary positioned to read this leg already owns the dashboard session and the box.
The agent→controller leg (local API, pinned TLS) never leaves the box. A LAN-direct delivery
variant (à la plex.direct) is PARKED, not forgotten — revisit only if the trust class of the
tunnel changes.
## OPERATOR FALLBACK — the CLI (text mode, unchanged)
```bash
felhom-agent --selftest=escrow-create --upload
```
- On the host, over SSH, as before; output byte-identical to the historical text mode. R is
printed once inside the write-it-down banner.
- `--storage <pbs-storage-id>` only if `escrow.pbs_storage_id` isn't configured.
- **⚠ Staged-secret rule:** when offsite is configured, ensure the staged secret is present
(re-save the offsite config to re-push it) BEFORE a CLI run — a CLI ceremony without it uploads
a hash-less blob, which trips the Scenario-F stale warning and can never auto-confirm.
- `--output=json` exists for the wizard's job runner (machine mode: ONE JSON object with R on
stdout, humans to stderr). Never run json mode by hand into a terminal you are screen-sharing.
## Prerequisites (the wizard preflight checks these for you)
- **Agent ≥ v0.88.0** (wizard endpoints + `--output=json`); ≥ v0.79.0 for `restic_pw_sha256`.
- **⚠ No-PBS hosts: the ceremony CANNOT run — POLICY, not a gap.** Identity-only escrow is
formally PARKED (operator decision 2026-07-12, DR-tier-by-default decision 4); every install
ships the full DR plumbing (host-install ≥ v1.15.0), activation is the hub's per-customer
DR-tier flag (default ON), and offsite backup REQUIRES that flag. A box always reaches the
ceremony THROUGH the tier: WG peer → PBS-DR descriptor applied → ceremony.
- **Host packages:** `age` (installed by host-install since v1.15.0).
- **K gate (PBS hosts):** `escrow.pbs_storage_id` set and the key file present.
- **Sudo grant:** the `FELHOM_ESCROW` alias in `/etc/sudoers.d/felhom-agent` (ships with
host-install ≥ v1.16.0 via the canonical sudoers fetch; on older installs re-ship the sudoers
with the agent deploy).
- The customer is present: R is displayed ONCE and must land in their hands.
## R handling — the rules to say out loud to the customer
1. R is displayed **once**; the system keeps no copy anywhere. Write it down / password manager /
safe — NOT on the server it protects.
2. **A new ceremony's R SUPERSEDES the old one.** Replace the stored code; the old blob remains valid
only for pre-existing history.
3. R + the hub blob = full recovery after total loss (see Recovery below). Losing R does not lose data
while the box lives — but after a disaster, an offsite copy without R is permanently sealed. That is
the zero-knowledge trade, stated honestly.
## After the ceremony — hands off (auto-confirm)
Within one report cycle the hub serves the blob's presence + hash in the ACK; the controller matches
`sha256(local repo password)` and flips `pending → escrowed` on its own, wiping the agent-staged copy.
**Verify remotely:** the customer page / report shows `escrow_state: "escrowed"` with zero manual clicks;
the next scheduled offsite run proceeds and snapshots appear. Hash-mismatch warning instead → the blob
does not cover the CURRENT password (stale blob) → re-run the ceremony (wizard); do NOT manual-confirm
around it. (The manual confirm ENDPOINT exists solely as a deprecated fallback for pre-hash legacy
blobs; its button is gone from the card since v0.127.0.)
## Recovery (the other half — operator-assisted, sketch)
Total loss → the customer's R + the hub blob:
`felhom-agent --selftest=identity-consume --blob <blob> --keydest <0600 scratch>` with R → recovers the
bundle (restic password, tokens; K restored via its own path on PBS hosts). Offsite data: hub re-issues
SFTP credentials (Re-issue button — access is regenerable; the ENCRYPTION key is what R guards) → inject
the recovered password on the fresh controller (`POST /backup/offbox/inject-password`, honored by the
secrets writer) → restore through the real endpoints. Live-proven end-to-end 2026-07-09 (recovered
password opened the production repo).
## Do NOT
- Pipe the CLI ceremony's output to a log or run it unattended (R would land in it). The wizard's
machine mode is exempt BY DESIGN: its pipe terminates in the agent's in-memory one-shot holder,
nowhere else.
- Manual-confirm a hash-mismatch; skip the self-verify; retain or transmit R on Felhom systems.
- Treat the SFTP credential as escrow-worthy (it's regenerable access; only the ENCRYPTION secrets ride R).
- Run a CLI ceremony on an offsite-configured box without the staged secret (hash-less blob →
stale warning; the wizard's re-stage-first makes this impossible from the UI).