Skip to content

concealer 0.9.28

Choose a tag to compare

@github-actions github-actions released this 25 Sep 12:37

Rollup release: folds in 0.9.25, 0.9.26 and 0.9.27, which were developed between 0.9.24 and this release and never cut their own tag.

[0.9.28] — 2026-09-23

Added

  • Auto-start service — concealer service install|uninstall|status runs the lan bridge
    automatically at login as a per-user background service (macOS launchd, Linux systemd --user,
    Windows Scheduled Task), so phones can always reach and auto-discover the vault with no
    manual concealer lan. Prefers the globally installed concealer so it opens the real vault.
  • LAN discovery — lan now advertises the vault over mDNS/Bonjour (_concealer._tcp, via
    dns-sd/avahi-publish-service, best-effort) with a stable vault id so a phone can pick
    the vault by identity rather than by ip/host/port (handles DHCP changes and multiple vaults
    on one network). New unauthenticated GET /api/serverinfo returns {id, name, version} (no
    secret) for labeling discovered servers and validating a manually entered address. The id is
    minted once and stored in keys/serverinfo.json (git-ignored).
  • Mobile app (iOS + Android) — a new mobile/ Capacitor app that opens the vault
    offline with the master password, imports the whole vault from the host on first
    connect, lets you add/edit/delete secrets on the phone offline, and syncs bidirectionally
    when online. All crypto runs on‑device via the age-encryption JS library (same age
    passphrase format as concealer bundles — no sops/age binary on the phone). At rest the
    phone stores only the master‑password‑encrypted bundle. See mobile/README.md.
  • Bidirectional device sync — foundation for the mobile app. A new authenticated
    POST /api/sync endpoint exchanges an encrypted vault bundle both ways in one call:
    the peer uploads its bundle (encrypted with the master password), the host does a
    last-write-wins merge (newer updated timestamp wins), saves, and returns the full
    merged set (bundle) so the peer converges to the same state. Secret values never cross
    the wire in plaintext — the sync payload uses the same portable bundle encryption as
    export. Requires the master password (human-owner operation), audited as sync.
  • Deletion sync via tombstones — deletes are now soft: a deleted record keeps its id,
    name and scope (all secret values dropped) and is marked with a deleted timestamp so
    the deletion can propagate to other devices. Tombstones never appear in lists, search,
    get, the web UI, MCP enumeration, or human backups/exports — only in device sync.

Fixed

  • rm --name X could delete the wrong record(s) on a vault with two or more live
    records. The scope selector was rebuilt inside a matching loop and the builder mutates
    its argument list, so after the first row the selector became empty and matched every
    remaining record (e.g. rm --name gamma reporting "2 records matched"). The selector is
    now computed once.

[0.9.27] — 2026-09-23

Added

  • concealer agent install [--all|--mcp-only|--skill-only] [ids] — auto-detects the
    AI agents installed on this machine (Claude Code, Cursor, Windsurf, Gemini CLI, Claude
    Desktop, Codex), lets you pick which to set up interactively (or by id / --all), then
    wires concealer's MCP server into each one's config with a fresh per-agent token and
    installs the concealer skill where supported (Claude Code). Replaces the manual
    agent register + claude mcp add two-step. Config merges are non-destructive (existing
    MCP servers preserved) and idempotent. Tokens stay client-side; revoke with
    concealer agent revoke <id>.

[0.9.26] — 2026-09-21

Added

  • concealer lan [lan_port [web_port]] — reach the web UI from a phone (or any
    device) on the same local network. web binds loopback only and rejects
    non-loopback Host/Origin (anti-DNS-rebinding), so it is unreachable from a phone by
    design. lan runs a thin reverse proxy that accepts only private/loopback/link-local
    clients
    , rewrites Host/Origin back to loopback, and forwards to a local web
    instance (auto-started if not already running, using the installed vault). The master
    password stays the only gate; the proxy holds no vault state or key. Phone URL is
    http://<host>.local:<lan_port> (default 8788 → web 8787) — resolvable over Bonjour/mDNS
    with no extra DNS setup. Intended for a trusted LAN (plain HTTP).

[0.9.25] — 2026-09-15

Changed

  • Default CLI unlock-token lifetime is now 1h (was 8h), overridable with
    CONCEALER_TOKEN_TTL (seconds). Shorter default = smaller window if a shell holding the
    token is exposed. unlock now also exports CONCEALER_TOKEN_EXP (expiry epoch) so a
    shell hook can auto-unset the token string once the server-side TTL lapses (the token is
    already rejected after expiry — a child process cannot unset a var in its parent shell).

Security

  • The TUI now always requires the master password. It previously honored a
    CONCEALER_TOKEN in the environment and would open — and reveal secrets — with no
    prompt, so any shell holding an unlock token could browse the vault. The TUI now ignores
    the token and unlocks from the master password on the terminal, matching the web UI's
    per-session unlock. (Token-driven get/list/MCP are unchanged.)

Fixed

  • run_with_secrets / run now inject usable env vars for secrets whose name isn't a
    valid shell identifier.
    A secret named grafana-rpifx previously emitted
    grafana-rpifx_PASSWORD — not referenceable by the child ($grafana-rpifx_PASSWORD
    doesn't parse, compgen -e won't list it), so website/login and other hyphenated
    secrets looked "not injected". Names are now sanitized to valid identifiers
    (grafana-rpifx → GRAFANA_RPIFX_PASSWORD). Colliding names fail loudly instead of
    silently clobbering.

Added

  • CONCEALER_INJECTED — injected into the child env as a comma-separated list of the
    injected identifier names (names only, never values), and shown on an injected: line
    in the run_with_secrets MCP result, so a script/agent can discover what was injected
    without guessing.
  • env_alias on a record — pin a stable env-injection base (e.g. grafana-rpifx →
    GRAFANA) for scripts that reference fixed variable names. Settable via set_secret.
  • Secure Agentic Autofill (fill-without-reveal). A new MCP tool request_web_login
    lets a browser-driving agent sign in to a site with a stored website/login secret
    without the value ever entering the agent/LLM context. The agent only asks; a human
    approves in the concealer web UI (approve once / for this task / deny, with an
    anti-phishing domain match); the concealer browser extension injects the credentials
    into the page's login form and the value flows app → extension → DOM only. Every step is
    audited (web_autofill_request/approve/deny/fill/exposed) and the agent receives only a
    status string. The extension gains a background service worker + content script (form
    detection, native-setter fill with input/change events, and a post-fill DOM exposure
    scan). Loopback endpoints /api/autofill/{request,pending,jobs,approve,deny,claim,result,status}
    carry the handshake; the MCP process reaches the web server on CONCEALER_WEB_PORT
    (default 8787).