Commit Graph
2 Commits
Author SHA1 Message Date
thePR0M3TH3ANandClaude Fable 5 55879eaa1b security: durable vault writes, path validation, and safe restore semantics
Continues the security-review fixes.

Durability and concurrency:
- Vault, config, fingerprint-registry and parent-seed writes are now atomic:
  temp file in the same directory, fsync, rename, then an fsync of the
  directory. Previously the target was truncated and rewritten in place, so
  a crash could destroy the vault.
- The lock now spans the whole read-modify-write cycle, not just the write.
  Serializing writes alone still lost changes -- two commands each read the
  same index, mutated their copy, and wrote in turn. Every mutating command
  goes through mutateVault(), which holds the lock from open to save. A test
  runs five concurrent writers and asserts all five entries survive.
- 0600 is now enforced on existing files too: writeFile's mode argument only
  applies on creation, so a loosened vault stayed loosened.
- The index is schema-validated before it can reach disk.

Path handling:
- Profile fingerprints are validated (16 uppercase hex) before being used in
  any path, and removeProfile additionally checks the resolved directory
  still sits directly under the app directory. A tampered registry could
  otherwise have steered a recursive delete elsewhere.
- Malformed entries in fingerprints.json are dropped on read.

Restore semantics:
- `nostr restore` merges remote state into the local vault instead of
  replacing it. Replacing silently discarded entries created since the last
  sync; that now requires --replace --yes, and any overwritten vault is
  copied aside first.
- Deltas are bound to the manifest they extend (filtered on the e tag, and
  re-checked locally), so a relay cannot replay a validly-signed delta from
  another snapshot lineage into a restore.
- Manifest selection breaks created_at ties by relay order, so a same-second
  republish cannot leave a restore on a stale snapshot.
- mergeIndexPayloads gained an index0 option. The default still refuses to
  merge populated Python atlas state rather than drop it; restore uses
  preserve-current, keeping the local block verbatim, because Python
  recomputes canonical views on load and the default would block every
  restore into a Python-created profile.

193 core + 69 CLI tests green; 36/36 cross-implementation checks green.

Co-Authored-By: Claude Fable 5 <[email protected]>
2026-08-14 23:10:20 -04:00
thePR0M3TH3ANandClaude Fable 5 e84a7e1c0e TS port slice 9: relay transport + nostr sync/restore — Milestone 5 complete
- core RelayPool over the platform WebSocket (global in Node 22+/browsers,
  no dependency): multi-relay publish with per-relay OK collection, fetch
  aggregation until EOSE with dedupe and signature verification on ingest,
  connect timeouts, dead-relay tolerance.
- Sync flows mirroring nostr/snapshot.py: publishSnapshot (chunk events ->
  manifest with HMAC identity + nonce), fetchLatestSnapshot (newest-first,
  per-chunk hash verify, falls through incomplete manifests), publishDelta,
  fetchDeltasSince (created_at replay order).
- CLI nostr group: get-pubkey (matches app-1237 fixture npub), list/add/
  remove-relay (config-backed, ws:// validation, last-relay guard), sync
  (publish local vault), restore (snapshot + delta merge -> local vault).
- Tested against an in-process NIP-01 mock relay: core round trip incl.
  corrupted-chunk rejection, and CLI end-to-end sync -> destroy local
  vault -> restore with secrets intact. Transport tests are Node-only
  (jsdom WebSocket can't connect; real browsers come with Milestone 6).
- 151 core + 32 CLI tests green. Milestone 5 CLI surface is complete.

Co-Authored-By: Claude Fable 5 <[email protected]>
2026-08-14 14:51:18 -04:00