Phase 1 of docs/nostr-vpn-integration-plan.md's Phase 0->4 plan (seller-side
rootless feasibility already confirmed there). Two apps, one image:
- apps/nostr-vpn: the daemon. Own network namespace (container.network:
pasta), NET_ADMIN+NET_RAW scoped to that netns, /dev/net/tun, and the
net.ipv4.ip_forward sysctl via the primitive added in e42bd26. UDP 51822
(not upstream's default 51820, which collides with archipelago-wg on
fleet nodes per the Phase 0 log). Seller mode stays off until an operator
explicitly enables it (paid_exit.enabled defaults to false upstream).
- apps/nostr-vpn-web: the control panel, gated behind 127.0.0.1:38080,
talking to the daemon only through the shared /data volume (state-file
status + shelling out to the nvpn CLI) -- no network link between the
two containers, matching upstream's own umbrel/docker-compose.yml.
- docker/nostr-vpn: upstream's umbrel/Dockerfile, unchanged except for how
the pinned commit arrives (shallow git fetch of a verified SHA, since
codeload.github.com archive tarballs 404 from this environment and
GitHub won't fetch an arbitrary SHA directly). Entrypoint seeds a minimal
config.toml with the chosen listen_port on first boot only -- every
AppConfig field is `serde(default = ...)`, confirmed by reading
nostr-vpn-core directly, so this merges with nvpn's own identity/wallet
bootstrap instead of needing a generated_secrets entry or full config
template, and never touches a config that already exists.
Both volumes point at /var/lib/archipelago/nostr-vpn, adopting state from
the old root-mode install. Build and the seed-config path were verified
against the real `nvpn daemon` binary, not just read -- see the plan doc's
Phase 1 log for what that caught (a fabricated commit SHA, the codeload
404, wrong default branch name, and confirming identity/wallet persistence
actually survives container recreation).
Not done here, flagged in the plan doc instead: removing the old root-mode
path (rpc/vpn.rs, rpc/auth.rs's auto-enable-on-login) touches live
onboarding on every node, not just this app -- needs explicit sign-off.
Also missing: a stop-hook/uninstall-guard manifest primitive (doesn't
exist yet -- LifecycleHooks only has post_install/pre_start) for the
collect-due-on-stop and non-zero-wallet uninstall guard, and registry
mirroring + catalog signing (need credentials this pass doesn't have).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Routing apps (a rootless VPN exit) need packet forwarding in their own
network namespace, but /proc/sys is read-only inside a rootless
container, so it can only be set at create time. Add `app.sysctls`,
allow-listed to net.ipv4.ip_forward / net.ipv6.conf.all.forwarding with
values "0"/"1", and rejected under host networking where it would change
the host. Rendered on all three create paths: podman CLI --sysctl, the
libpod spec `sysctl` map, and Quadlet `Sysctl=`. Absent by default and
not serialized when empty, so existing manifests and units are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every paid download logged "filing into filebrowser/Music/... failed
(non-fatal): Permission denied". The purchase played in-app but never
appeared in Files. FileBrowser's folders belong to its rootless container
range (host uid 100000, mode 755). This service is host uid 1000, outside
that range, so it can read them but not create files in them.
New container::filebrowser::save_new_file:
- Writes directly when the folder allows it.
- Otherwise writes through `podman unshare`, where that uid range is
ours: to a temp file, then chowned to the folder's owner, set to 0644,
and hard-linked into place. FileBrowser never sees a partial file and an
existing file is never replaced. A missing folder is created and given
its parent's owner. No sudo.
- Keeps the "name (2).ext" de-duplication the RPC did inline.
Checked the unshare script on amishparadise in a scratch folder owned
like FileBrowser's: new folder + file OK, owner/mode right, no clobber,
no temp file left, and the service can read the result.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- mint_client: a stub mint shows swap() sends the full v2 keyset id when
given a cashuB short id, and leaves complete v1/v2 ids unchanged.
- fips::dial: the single-delivery decisions are now small functions
(fips_answer_is_final, fips_retryable). Tests cover them and, against a
silent local peer, check that a single-delivery request isn't resent
after a timeout while an ordinary one still is.
- content_server: an unreadable paid file returns Unavailable before the
payment gate runs, and a readable one still returns 402. Also covers
ensure_readable's grant/reopen behaviour. The podman grant is replaced
by a refusal under cfg(test) so results don't depend on the host.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After the keyset-id fix, a Minibits paid download still failed and the
buyer lost the sats. What happened, 2026-09-29, amishparadise:
1. The seller redeemed the token, then failed to read the file. It was a
FileBrowser upload owned by the container subuid (100999) with mode
0640. The handler mapped that Err to 404.
2. The buyer's FIPS dial treats 404 as "fall back to Tor" and resent the
request with the same, now spent, token. The seller answered 402, and
the buyer showed "seller doesn't accept your Cashu mint".
Fixes:
- serve_content checks the file is readable before the paid gate. If it
isn't, it grants read with `podman unshare chmod a+r`, which matches
the other shared files. If that also fails it returns Unavailable (503)
without taking payment.
- The content handler returns 500 on internal errors and logs them,
instead of a silent 404.
- New PeerRequest::single_delivery(), used for the paid download: the
FIPS answer is final, FIPS retries only when it never connected, and
there's no Tor replay once the request may have been delivered.
- The buyer shows the seller's error text for non-402 failures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ssmithxandClaude Sonnet 5
chaum