Skip to content

About

Encrypted personal and shared file vaults for humans and coding agents, built on PocketContext.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

VaultContext

Current release controls and platform coverage: common CI and deployment contract.

Encrypted personal and shared file vaults for humans and coding agents, built on PocketContext. Any file type is accepted as exact opaque bytes, initially up to 8 MiB per file. Names and descriptive metadata are encrypted too. A CLI and portable skill provide the file interface. A public single-page website provides product information and a guided onboarding manual; it does not authenticate, unlock vaults or access application data. Clipperz inspired the architecture; no Clipperz code or compatibility is included.

Access and keys

Hosted IdentityContext login authenticates PocketBase's default users identity; explicit direct Google login remains available. It does not unlock encrypted data. Each account initializes an encryption/signing identity protected by a separate passphrase. The encrypted private-key bundle persists only on the server; unlocking downloads and decrypts it in process memory. Application tokens and automatically pinned or independently verified public fingerprints may be stored locally, never the private-key bundle. The token lives in $XDG_STATE_HOME/vaultcontext/ (default ~/.local/state/vaultcontext/, or VAULTCONTEXT_SESSION_DIR), so clearing ~/.cache does not sign you out. There are no recovery keys. A lost passphrase without a usable unlocked session or enrolled macOS Keychain credential means lost access; Google account recovery cannot decrypt the vault. Optional macOS Keychain enrollment stores the passphrase locally, never the private-key bundle.

Personal vaults are private. Shared project vaults have one owner and explicit editor/reader members, initially within the configured Workspace. Share by exact email with an active, Google-verified Workspace account that has initialized encryption. The client automatically pins public keys on first use and rejects changes to known keys, including when accepting invitations, reading another author’s files and rotating keys. This trusts the server’s initial email-to-key mapping; Google authenticates the account, not its encryption key. A compromised server can substitute a key on first contact, and a new machine has no previous pins. Independent fingerprint verification remains optional. New members receive retained history. Revoking a member blocks retrieval immediately and freezes writes until key rotation completes. Downloaded plaintext or keys cannot be recalled, and credentials contained in files may need separate provider rotation. Owners cannot remove themselves; owner transfer and identity-key replacement are not implemented.

The execution host and unlocked agent are trusted. The client retains keys in a same-OS-user Unix socket session, with a 15-minute default lifetime and explicit lock. Linux and macOS are supported client platforms. Python cannot promise perfect memory erasure or protect against a privileged host, swap or an agent already authorized to read local files. No stored content authorizes execution or sharing.

See implementation brief, data and trust model, API contract and deployment preparation.

Onboarding page

The hand-maintained static page lives in pb_public/ and is served at / by the application-owned pb_hooks/frontend.pb.js hook. The container includes these assets. No frontend build is required. Optional production-only analytics are described below. Open Graph and Twitter large-image metadata reference the public pb_public/og-card.png (1730 × 909); its URL and the canonical page URL use the production origin. The built-in image generator created the card; its prompt is recorded in docs/social-card-prompt.txt.

For a page-only local preview, run python3 -m http.server 8769 --bind 127.0.0.1 --directory pb_public. Expose only this public directory when sharing a temporary preview.

Keep onboarding commands aligned with the portable client and skills/vaultcontext/references/workflows.md. After edits, check desktop/mobile layout, keyboard disclosures, installation method switching and clipboard success/failure. Validate routing and security headers with the pinned server in an isolated temporary database before release. The “Bring your files to another computer” guide demonstrates explicit prefix mapping, comparison reports, selected version-pinned restores and verification with synthetic examples. Company-specific onboarding stays in WikiContext.

Public website analytics

Tracking runs only on the exact origin https://vault.pocketcontext.com and root page /, after the visitor chooses Allow analytics. Localhost, forks, previews, alternate ports and self-hosted deployments load no analytics services, even with a remembered choice. Analytics settings in the footer lets visitors withdraw; the browser stores only the choice until analytics is accepted.

The page reuses PocketContext's public GA4 measurement ID G-1X0FZLTWGE and Rybbit site ID d8b36ba4a63f. These are public routing identifiers, not credentials. Rybbit receives explicit /api/track requests, without loading its script or session recorder. GA4 loads in an empty same-origin frame only after acceptance; its automatic capture operates on a blank document instead of the guide. Ads storage, ad personalization and Google signals are disabled. Google analytics cookies use a VaultContext-specific prefix and host. Withdrawal removes the frame and those cookies, and stops new application events; it cannot recall already sent events.

Explicit events are page_view, onboarding_open, onboarding_method and onboarding_copy. The shared GA4 property can also emit automatic scroll or engagement events for the empty frame; filter to the explicit event names when measuring guide use. The frame isolates automatic capture, not untrusted scripts. Custom events contain only an allowlisted target such as compare_prompt or restore_commands. Copy events mean a successful clipboard write, not a completed CLI operation. Requests use a fixed root page URL/title and omit query strings, fragments and referrers. No contents, local paths, clipboard text, identities or vault operations enter analytics. Providers still receive ordinary network metadata; browser blocking can reduce reported counts.

Filter both dashboards by Hostname equals vault.pocketcontext.com. In Rybbit, use the site dashboard filter and save a segment where supported. In GA4, create a Free form exploration, import Hostname, then apply the exact hostname filter. These are reporting filters; the origin guard controls collection. Existing marketing-site tracking settings are not changed by this integration.

Public JS/CSS references include a v query containing the first 12 characters of that asset's SHA-256. Update the HTML reference whenever its asset changes; the browser check verifies these versions. The container smoke gate verifies all public assets, including both analytics-frame files, match source exactly.

Run the offline browser checks (all provider traffic is intercepted):

uv run --with playwright==1.60.0 python -m playwright install chromium
uv run --with playwright==1.60.0 python tests/website_analytics.py
# With an installed Chrome instead: PLAYWRIGHT_CHANNEL=chrome uv run --with playwright==1.60.0 python tests/website_analytics.py

Local server

Build the exact commit in POCKETCONTEXT_VERSION with its go.mod toolchain, CGO and a C compiler. Keep it in a separate checkout if your existing server revision differs:

# In the pinned PocketContext checkout:
make build
# In this repository, with a disposable database for evaluation:
/absolute/path/to/pocketcontext serve --dir /absolute/private/test/pb_data --http 127.0.0.1:8090

The application directory supplies migrations, hooks and SQL configuration. No users or real records are seeded. Operators provision test identities through ordinary maintenance APIs; the CLI never uses superuser credentials.

Production Google settings use a separate Web OAuth client, paired VAULTCONTEXT_GOOGLE_CLIENT_ID / VAULTCONTEXT_GOOGLE_CLIENT_SECRET, and VAULTCONTEXT_GOOGLE_WORKSPACE_DOMAIN. Server-side verified claims enforce the exact Workspace domain. Direct public signup is blocked. Without a configured domain, Google login requires preprovisioned identities. Disable the application account to revoke its tokens; Google suspension alone does not invalidate existing app sessions.

Hosted identity login

Bare vaultcontext login --url https://vault.pocketcontext.com --email you@pocketcontext.com reuses the shared IdentityContext session or opens a hosted Google sign-in request. Over SSH, open its URL on your browser computer and enter the single-use code in the initiating terminal; no callback tunnel is required. Explicit login --google retains direct-app fallback, including its SSH callback forwarding requirement. login --password remains explicit for provisioned accounts. Central rejection never silently selects a fallback or opens a browser during ordinary commands.

Shared identity state is private under $XDG_STATE_HOME/pocketcontext-id/ (default ~/.local/state/pocketcontext-id/), with seven-day idle and thirty-day absolute limits. App tokens remain in VaultContext's own state directory and expire at most fifteen minutes after the assertion was issued. Renewal obtains a new audience-bound assertion instead of extending the old token. Central disable blocks new assertions; complete offboarding also disables independent fallback application accounts.

Login never initializes, downloads or unlocks encryption keys. Verified hosted identity can establish Workspace sharing eligibility, but grants no vault membership. Unlocked operations remain bound to the original immutable account ID, even if saved authentication is replaced or renewed. Lost login state, changed accounts or rejected authentication fail closed. Renewal never extends the memory-session deadline. vaultcontext logout locks that account and removes its app token, retaining the shared identity session. Use the separately installed identitycontext logout --email EMAIL to revoke that shared session; already issued app tokens remain bounded to fifteen minutes. Install the matching IdentityContext launcher from its published skill.

This release requires an explicit client upgrade: lock the old worker, install the new pinned client, then unlock privately. There is no automatic key transfer or unlock.

Saved login settings

Login accepts --url and --email before or after the command. Successful explicit login saves the default URL/email in $XDG_STATE_HOME/vaultcontext/default.json (default ~/.local/state/vaultcontext/default.json, or VAULTCONTEXT_SESSION_DIR/default.json). Each server/account token stays in its separate hashed session file. CLI options override environment variables, which override saved defaults. Files use mode 0600 in a mode-0700 directory. Ordinary commands and token refresh do not change the default. vaultcontext logout removes the selected token but retains settings; logout --forget also removes matching defaults. Old cache sessions are not imported; sign in again. Password login uses login --password with VAULTCONTEXT_USER_PASSWORD; passwords are never saved.

Portable client

The backend accepts exactly one matching client release. See client releases and upgrades. Users upgrade explicitly when the server changes; old clients cannot use protected APIs. Public installation instructions and local lock/logout remain available. The release ID is a compatibility check, not authorization.

Install uv and copy the executable launcher onto your PATH:

install -Dm755 skills/vaultcontext/vaultcontext ~/.local/bin/vaultcontext
export PATH="$HOME/.local/bin:$PATH"
vaultcontext login --url https://vault.pocketcontext.com --email you@pocketcontext.com

The launcher declares its Python requirements and pins the client package to a full Git commit. It can run from any directory without neighboring source files. The first run needs network access to download the package, Python if needed, and dependencies; later runs reuse uv's cache. Updating the copied launcher adopts its new client revision. Before replacing an installed client, run its lock command for each configured server/account with an unlocked session. After updating, have the user run vaultcontext unlock explicitly in their private terminal so the session uses the new client; installation never unlocks automatically. The skill has no script lockfile: transitive dependencies may resolve differently on fresh installations. The repository's uv.lock governs development and tests, not launcher execution.

To migrate from the old vc launcher, run vc lock before installing vaultcontext, then remove only the old VaultContext launcher you installed (for example, ~/.local/bin/vc). The package and skill provide no vc alias or compatibility wrapper. Update scripts to call vaultcontext. Existing authentication configuration and macOS Keychain enrollment remain valid.

The memory session runs in a fresh Python process using the launcher's own interpreter and installed package environment. Unlock waits for worker readiness; verified keys cross only a private inherited socket, never arguments, environment variables or files. Lock sessions before clearing uv's cache. unlock, lock and logout recover an abandoned socket only after checking ownership, permissions and connection refusal; ambiguous or unsafe endpoints remain errors. Repeating lock on an already locked account succeeds.

Alternatively, install the package from a trusted checkout into an external Python 3.11+ virtual environment. This provides the same vaultcontext command without requiring uv at runtime:

python3 -m venv ~/.local/share/vaultcontext-venv
~/.local/share/vaultcontext-venv/bin/pip install /absolute/path/to/vaultcontext
~/.local/share/vaultcontext-venv/bin/vaultcontext login

Run vaultcontext or vaultcontext --help for grouped commands and a first-use example, or vaultcontext COMMAND --help for arguments, prerequisites and examples. Commands return JSON except cat, which writes exact file bytes to stdout. Help does not require configuration or sign-in.

Use the chosen vaultcontext command for whoami, check, init, unlock, and subsequent operations. init and ordinary unlock prompt in an interactive terminal. On macOS, opt in with vaultcontext keychain-enroll, then use vaultcontext unlock --keychain to retrieve the saved passphrase after Touch ID or macOS credential authentication. This requires the signed native helper; see macOS Keychain setup and trust model. Never pass unlock secrets through chat, arguments, environment variables or pipes. Explicit fallback login --google prints the sign-in URL and opens the default browser unless --no-browser is given, an SSH session is detected or no browser is available. For that fallback over SSH, forward Google callback port 8765 as described in the skill workflows; decryption still occurs on the machine running the CLI. Saved login settings work outside the original workspace; environment variables remain optional overrides.

The skill includes the launcher and operation references. It supports vault creation, arbitrary-file saves, local name search, version history, private equality checks, exact restores, invitations, memberships, revocation/rotation, encrypted export and independent archive restore. Agents manage files and metadata but never inspect file contents, including source or restored files. The dedicated compare command may answer a requested equality check without exposing contents or hashes. Content viewing with cat is for users in their own private terminals. This skill rule does not remove the CLI or unlocked session's decryption capability. Saving again with --document creates a new immutable version. The default encrypted display name preserves the exact source-path argument (after shell expansion), including relative or absolute directories; --name overrides it. Names are labels, not unique identifiers or restore destinations. Existing versions keep their names. No format-specific parsing, automatic synchronization or shell activation occurs.

vaultcontext archive DOCUMENT_ID hides a document and all its versions from default list and search; --archived selects archived documents and --all selects both. Owners and editors can archive or unarchive; readers cannot. Repeating either action is harmless. Archived documents retain history, cat and restore access by ID, but reject new versions until unarchived. Archiving changes neither access nor retention: exports, complete backups and whole-vault sharing include archived documents. Archive state is document metadata, distinct from an encrypted export archive.

Listings and name searches fetch selected version metadata in batches and reuse verified keys only for the current command. Every version's signature and document binding are still checked; subsequent commands recheck access and fingerprints. SQL reads that receive HTTP 429 retry at most twice, waiting 10 seconds each time. Writes are never automatically replayed after rate limiting.

vaultcontext compare DOCUMENT_ID LOCAL_PATH compares a local file with the current saved version; add --version VERSION_ID to select history. It requires login and an unlocked session and returns document/version IDs, a same boolean and a comparison method, never contents or hashes. New saves include a SHA-256 checksum of the plaintext inside encrypted, signed version metadata. For those versions, comparison downloads metadata only (encrypted-sha256); it reads local bytes internally to compute their checksum. Older versions use legacy-download, downloading and verifying the stored bytes in memory without plaintext temporary files. No existing version is rewritten. Hash equality has negligible collision risk, but does not prove that stored ciphertext remains downloadable or recoverable. Use backup/restore validation for recovery assurance.

vaultcontext cat DOCUMENT_ID outputs the current file; add --version VERSION_ID for a historical version. It verifies the entire file before output, creates no plaintext temporary files, and adds no formatting or newline. Binary data and terminal control characters pass through unchanged. Output can expose secrets to the terminal, pipes or captured logs. Shell redirection uses ordinary shell permissions and overwrite behavior; use restore for protected disk writes. Neither command executes file contents.

Restores require explicit paths and existing parent directories, reject symlinks and preserve byte contents with mode 0600. Existing destinations require --overwrite. Original executable bits, permissions, ownership and source absolute paths are not restored. An .envrc.private restored into a direnv-enabled directory can execute later when the user's shell loads it; the client never activates it.

Vault owners can use vaultcontext archive-vault VAULT_ID and vaultcontext unarchive-vault VAULT_ID. Archiving applies to all members and hides the vault from default vaults; vaults --archived selects archived vaults and vaults --all selects both. Listings include archived. Existing files, history, protected downloads, comparisons and exports remain accessible by ID. Archived vaults reject saves, document archive changes, new invitations and invitation acceptance. Pending invitations remain pending and can be accepted after unarchiving. Revocation and key rotation remain available. Individual document archive states and any rotation freeze remain unchanged. Archiving neither revokes access nor deletes data or reclaims storage. Repeated requests with current revisions are harmless. Complete backups retain vault archive state; encrypted file exports retain their existing v2 format and do not recreate vault lifecycle.

Storage and backups

Metadata reads use authenticated requester-filtered SQL. Writes use the standard REST vault_actions collection and transactional hooks; action payloads are not persisted. Encrypted file chunks use protected file storage and authenticated downloads so file bytes do not exhaust SQL snapshot budgets. Independent rules cover file downloads, REST and realtime. There is no implicit operator decryption access.

The pinned server reuses each account's private SQL snapshot while the source database is unchanged. Any committed write, including invitations, acceptance, saves, revocation, rotation and account changes, invalidates every retained copy before the next read, so a read after a confirmed write observes it with current membership. At most maxConcurrent (two) copies are retained in private server temporary files. Background snapshots, trackChanges and maxConcurrentPerRequester are deliberately not configured: background reads are eventually consistent and per-account admission would reject concurrent sessions with HTTP 429. tests/snapshot_freshness.py checks reuse, next-read membership grant/revocation and privacy.

Complete recovery requires the database replica and every referenced ciphertext object, verified against its stored checksum. A database-only backup is insufficient. Startup must refuse missing/corrupt originals. An ordinary data export instead decrypts accessible file versions in memory and encrypts an archive under a separately entered archive passphrase; it includes no identity private keys or live vault-key envelopes. Exports restore without the server, but cannot recover a forgotten live-vault passphrase. Export plaintext is capped at 64 MiB including base64 overhead and metadata. New exports use format v2 to retain document archive status and require an updated CLI to inspect or restore. The updated CLI also reads existing v1 exports.

Validation

Use synthetic fixtures and isolated temporary databases only. From this repository, install the locked development environment and run:

On macOS, first run export TMPDIR=/private/tmp in the test shell. The default temporary path traverses the /var symlink, which the client's safe file handling intentionally rejects. The unit suite compiles the native Keychain helper with Apple's Swift command-line tools on macOS; its rejection tests never access Keychain.

uv sync --locked
uv run --locked python -m unittest discover -s tests -p 'test_*.py'
uv run --locked python tests/hosted_identity.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/release_gate.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/identity_rewrap.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/integration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/snapshot_freshness.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/archive_migration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/sharing_migration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/sharing_client_integration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/realtime.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/limits.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/backup_integration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/auth.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/oauth_integration.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/oauth.py
uv run --locked python tests/skill.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/cli_forward.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/deploy.py --binary /absolute/path/to/pinned/pocketcontext
uv run --locked python tests/deploy_workflow.py

Container configuration, smoke and populated complete-restore checks gate image publication; see deployment documentation for commands. Real Google browser/provider configuration and independent security review are separate from synthetic tests. Actual verification results and remaining limitations are recorded in docs/validation.md.

VaultContext is deployed at https://vault.pocketcontext.com; see release evidence. The repository and container image are public. The image workflow deploys production automatically after successful publication through the pocketdeploy-v2 GitHub environment; production moved there from once-pocketcontext-v2 on 24 October 2026 by recovery from the replica. CONTEXT_DEPLOY_PAUSED=true pauses deployment without pausing CI or publication; see deployment controls.

Identity, OAuth and deployment patterns are adapted from RaiseContext; filtered access and complete-original backup patterns follow AccountContext. Their domain schemas and readers are not copied.

Client launcher releases

Follow the release procedure. Publish the tested package implementation without deploying its backend first. Update skills/vaultcontext/vaultcontext to that full commit SHA, then validate the remotely installed client before committing and pushing the launcher:

uv run --locked python tests/skill.py --binary /absolute/path/to/pinned/pocketcontext --client skills/vaultcontext/vaultcontext
uv run --locked python tests/cli_forward.py --binary /absolute/path/to/pinned/pocketcontext --client skills/vaultcontext/vaultcontext

These checks copy only the executable into a temporary directory. The launcher intentionally has no adjacent lockfile; keep the root development lockfile separate. A copied launcher must be replaced to adopt a later client revision.

Runtime maintenance freeze

Superusers use GET /api/context/maintenance and generation-checked PUT /api/context/maintenance with {"readOnly":true,"expectedGeneration":N} to drain and block writes without restarting. Existing authorized reads and protected original downloads remain available. Authentication that creates or updates records is blocked; preserve an existing operator token for thaw. Set readOnly:false with the returned generation to resume writes explicitly.

The private durable pb_data/maintenance.json marker survives restart. Frozen startup requires the existing database, skips restore and superuser/settings provisioning, verifies original files, and refuses pending migrations. Malformed markers fail closed. Litestream supervision remains active; this is a managed database/API freeze, not cross-host writer fencing or byte-immutable disk. Keep the marker with migration snapshots and fence the source before cutover.

Validate with python3 tests/entrypoint.py and python3 tests/maintenance.py --binary /absolute/path/to/pinned/pocketcontext.

Replicated startup waits for a private Litestream IPC synchronization before serving, including fresh Google-only databases. A failed initial sync refuses traffic; clean early shutdown therefore uses an initialized replica.

Container startup and primary object storage

Set all of VAULTCONTEXT_S3_BUCKET, VAULTCONTEXT_S3_ENDPOINT, VAULTCONTEXT_S3_REGION, VAULTCONTEXT_S3_ACCESS_KEY_ID, and VAULTCONTEXT_S3_SECRET_ACCESS_KEY to use a dedicated private S3/R2 bucket for PocketBase uploads. VAULTCONTEXT_S3_FORCE_PATH_STYLE defaults to true. The container requires complete S3 and Litestream configuration; LITESTREAM_DISABLED is rejected. Partial configuration and shared primary/replica buckets or access keys fail closed. This is primary file storage, separate from the LITESTREAM_* SQLite replica bucket and prefix. Protected downloads still require independent application authorization. Vault ciphertext stays encrypted.

Default startup preserves an existing database or strictly restores SQLite through Litestream into a temporary directory. It verifies database integrity and streams every referenced remote file before installing the restored database. Ciphertext chunks must match their stored SHA-256; other file fields, including avatars, are checked for readability and complete responses because they have no stored hash. Every start, including restarts on an existing volume, downloads these objects in parallel (VAULTCONTEXT_VERIFY_CONCURRENCY, 1–64, default 16) so the server becomes ready within ONCE's fixed 120-second deployment wait; the first failure refuses startup. No legacy archive scheduler or restore remains in the container. Direct pinned-server development may still use local synthetic storage. Keep object retention independent of replica retention; SQLite replication alone cannot recover deleted objects. Unexpected crashes can lose database writes since Litestream is asynchronous.

This configuration does not move existing files. Copy and checksum all referenced objects before enabling it. A frozen startup refuses any storage configuration change: prepare the destination settings before establishing its frozen snapshot. Preserve maintenance.json, pause CD, and fence the source before thawing a destination. Never point a second writable process at the live replica.

Run python3 tests/object_storage_settings.py plus the documented backup, maintenance and deployment tests. Use tests/object_storage_integration.py --binary /path/to/pinned/server --synthetic-bucket BUCKET with a disposable local MinIO bucket and the S3 environment above for real uploads, protected downloads and database-only recovery. Container and Litestream fresh-volume recovery validation remain release gates.

Container primary-storage recovery gate

After building the image, run the existing docker/smoke.py config, smoke, and restore checks (use the locked Python environment). restore runs docker/object_storage_smoke.py; either command executes the same populated gate. The restore check builds its pinned local MinIO fixture. The primary-storage check uses separate bucket-scoped synthetic keys, uploads real protected files, rejects unrelated-user downloads, and tests frozen restart plus a new destination volume. It compares every main database table restored by Litestream before destroying the source volume and explicitly thawing the destination. A final upload and clean stop are followed by destruction of that volume and automatic entrypoint recovery into a third empty volume, with authorized and denied protected-file checks. No cloud buckets or live replicas are used.

A frozen handoff bundle must also carry maintenance.json and a consistent SQLite backup of auxiliary.db from the stopped source: frozen startup refuses to create a missing auxiliary database. The main data.db still comes from Litestream. A writable disaster recovery can recreate auxiliary state, but must not be substituted for a deliberately frozen migration. Keep source fencing and this bundle explicit in migration tooling.

A genuinely new installation requires one explicit container init invocation with its intended private configuration and volume, followed by ordinary startup on that same volume. init refuses existing local state or a populated replica; default startup refuses a missing replica. Never initialize an existing deployment or a recovery volume. Interrupted initialization leaves initialization.pending and blocks startup; preserve it for investigation. See deployment commands.

The container uses one Python entrypoint, tini and Litestream; Python execs the supervisor and its server child after preparation. Historical archive recovery uses the previous pinned image in an isolated environment. Test-only legacy archive helpers remain under tests/ and are excluded from the image. No old archives are deleted.

About

Encrypted personal and shared file vaults for humans and coding agents, built on PocketContext.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages