omnifs v0.2.0-dev.0

Added

• omnifs is now distributed on npm as @0xff-ai/omnifs. Install with npm install -g @0xff-ai/omnifs; the CLI binary ships in one of four platform-specific optional-dependency packages (darwin-arm64, darwin-x64, linux-arm64, linux-x64). The Docker image is pulled on omnifs up, not at install time.
• Full omnifs init <provider> / omnifs up / omnifs down mount lifecycle. omnifs init walks through provider auth (device-code or PKCE loopback OAuth), writes a per-mount config under ~/.omnifs/config/mounts/, and stores credentials in the OS keychain (macOS Keychain, Linux libsecret, Windows DPAPI) with a mode-600 file fallback at ~/.omnifs/data/credentials.json. omnifs up pulls the matching runtime image, materialises credentials into a private session directory bind-mounted read-only into the container, then removes the session on stop or failure.
• omnifs auth subcommands: login, logout, status, refresh, scopes, import. OAuth refresh happens automatically with a one-shot 401 retry coordinated by a singleflight plus a cross-process file lock so concurrent CLI invocations do not race on the same refresh token.
• omnifs setup guided first-run walkthrough: detects OS and Docker, helps pick providers, runs init for each, and brings the container up.
• omnifs doctor runs ten ordered probes to diagnose why a mount is not working (Docker availability, FUSE timeout, missing credentials, etc.).
• omnifs reset clears configs and credentials after an explicit confirmation prompt.
• omnifs mounts ls / omnifs mounts rm for listing and removing configured mounts.
• omnifs status readiness card showing runtime state, configured mounts, and auth state.
• omnifs version and omnifs completions commands.
• omnifs dev contributor sandbox: walks up from cwd to find the workspace Cargo.toml, captures gh auth token, downloads the Chinook SQLite fixture into .secrets/db/test.db, builds an omnifs:<short-sha>-dev image, and starts a container with all built-in providers mounted. Replaces the old just dev / docker compose up workflow.
• Three supported OAuth flows in the new omnifs-auth crate: PKCE loopback (browser-redirect), PKCE manual code (paste-back), and device code (visit URL, type short code). GitHub uses device code with no default write scopes; Linear uses PKCE loopback with the read scope.
• Host-managed provider credentials: providers never see tokens; the host attaches them to outgoing HTTP requests after callouts cross the WASM boundary. Provider auth needs are declared in omnifs.provider.json (OAuth endpoints, scopes, injection header, allowed domains); adding a new service does not require patching the host.
• omnifs-creds crate for the keychain + file + in-memory credential store, with CredentialKey::storage_key() (provider:scheme:account) as the public wire form. Stale file-fallback entries are cleaned up after successful keyring writes; durability and permissions are hardened.
• Database provider (omnifs-provider-db) mounted at /db, projecting a read-only SQLite database as a filesystem. Exposes meta/{version.txt,path.txt,info.json} and tables/{name}/{schema.sql,schema.json,indexes.json,count.txt,sample.json}. SQLite runs inside the WASM sandbox via rusqlite with the bundled feature; the host preopens the database file's parent directory through Wasmtime's WASI context with read-only permissions so no bytes cross the WIT boundary. v1 is SQLite-only and read-only.
• Docker provider (omnifs-provider-docker) mounted at /docker, projecting the local Docker daemon over the Unix socket. Exposes /system/{info,version,df}.json, /system/ping, /containers/_listing.json, facets {by-name,by-id,_running,_stopped} each binding to a per-container subtree (inspect.json, summary.json, summary.txt, state), and /compose/{project}/services/{service}/containers/{name} grouping by Compose labels. Container state files are marked volatile so reads bypass the kernel page cache and always reach the provider. Bounded-window /events polling translates container actions into cache-invalidation prefixes.
• Unix-socket HTTP transport in the host. Providers use HttpEndpoint::Unix and build_url(path, query); the host detects unix: URLs, decodes the hex-encoded socket path, builds a per-socket reqwest::Client via ClientBuilder::unix_socket, and rewrites the URL for that client. Unix sockets are gated by a new unix-sockets allowlist on CapabilityGrants.
• Linear provider (omnifs-provider-linear) mounted at /linear, projecting a Linear workspace. Teams appear at /linear/teams/{KEY}, issues at /linear/teams/{KEY}/issues/{_open,_all}/{IDENT}/. Each issue surface has title, state, priority, assignee, and description.md as files. Issue listings preload child files so a cat after ls skips a follow-up round trip. Uses Linear's GraphQL API with hand-written query strings and serde response structs (Linear's endpoint rejects full introspection queries as too complex, ruling out code generation).
• arXiv provider now also exposes a recent-submissions surface per category: /categories/{cat}/recent, /categories/{cat}/recent/_fetched, /categories/{cat}/recent/pages/{n}, and /categories/{cat}/submissions/{YYYYMMDD} directories discovered from fetched pages. Direct paper lookup at /papers/{id} is unchanged.
• Projected file attributes: providers now declare Size (Exact, NonZero, or Unknown), Bytes (inline or deferred), ReadMode, and Stability (Immutable, Mutable, or Volatile) through the Projection API. The host uses these facts to set st_size, FUSE direct-I/O flags, cache behavior, ranged-read handling, and post-read size promotion. The old 256 MiB placeholder is removed. Volatile files return entry_timeout = 0, attr_timeout = 0, and FOPEN_DIRECT_IO; ranged files open a provider handle for snapshot-consistent reads.
• Sandboxed archive extraction via a host-owned omnifs-tool-archive Wasm component. BlobExecutor streams fetch-blob responses into a staged temp file and commits metadata before the body is visible; ArchiveExecutor keys extracted trees by (cache-key, format, strip-prefix) and coalesces concurrent extractions through TreeMaterializer. The extractor runs path sanitization, depth/length/entry-count/per-file/total-byte limits inside the sandbox and publishes completed trees via atomic directory rename so tree refs never observe partial output.
• EffectiveConfig type representing a mount after provider metadata has been merged in. ProviderCatalog::load_mount() returns one; credential targeting, session materialisation, and runtime construction all consume it. The previous InstanceConfig-plus-late-apply_metadata pattern is removed.
• Provider runtime capabilities now come back from init as (State, ProviderInfo, RequestedCapabilities) instead of a separate WIT export. Initialisation runs exactly once in ProviderRuntime::new. Capability entries can be marked dynamic: true when the concrete grant depends on mount config (Docker's socket path is the motivating case).
• omnifs-mount-schema crate is split into typed modules with a checked-in JSON schema at crates/omnifs-mount-schema/schema/omnifs.provider.schema.json (regenerate with just regen-schema).
• Per-crate README files for all published crates.

Changed

• arXiv provider route model is restructured around recent submissions. The calendar/date-query, new, updated, by-author, /authors, and /search surfaces are removed. Category traversal now goes through /categories/{category}/recent and /categories/{category}/submissions/{YYYYMMDD}; direct paper lookup at /papers/{paper} is unchanged. The only live category listing query shape is search_query=cat:{category} sorted by submittedDate descending.
• Provider protocol vocabulary is reorganized into three orthogonal channels: callout (intermediate host work the provider suspends on), return (the completed operation answer), and effect (host-side mutation committed at the return boundary). provider-response is renamed to provider-step with arms suspended(callouts) and returned(provider-return). A return cannot carry callouts; an error return cannot carry effects.
• Dead git callouts (git-list-tree, git-read-blob, git-head-ref, git-list-cached-repos) and the unused reconcile interface are removed from the WIT.
• The sidecar::materialize method is folded into lookup-child and list-children via #[subtree] dispatch from the SDK.
• Host runtime module crates/host/src/runtime/mod.rs is split into focused modules: instance.rs (Wasmtime mechanics), callouts.rs (dispatch and tracing), effects.rs (terminal mutations and ProjectionAccumulator), log_redaction.rs, wit_conversions.rs, op.rs (the Op enum and Validator), and http_stack.rs (shared HTTP transport). RuntimeError construction errors split into RuntimeBuildError.
• Browse cache re-skin enums collapse into their WIT counterparts; cache::SCHEMA_VERSION bumps to 5, invalidating existing L2 records.
• CLI flows are redesigned around mount configs, provider metadata, credential materialisation, and container lifecycle commands. Verbose output is off by default; -v enables INFO logs and -vv adds DEBUG. Common errors surface a Try: block with a concrete next step.
• Provider manifests (omnifs.provider.json) describe auth schemes, token injection policy, capability grants, and config schema. All built-in providers (arxiv, db, dns, docker, github, linear, test) carry an omnifs.provider.json and drop vestigial [package.metadata.component] Cargo sections.
• Docker Compose development entrypoints (compose.yaml, just dev) are replaced by the supported omnifs dev, omnifs shell, omnifs logs, and omnifs down workflow.
• Wasm providers are baked into the runtime image at /root/.omnifs/providers/ with OMNIFS_PROVIDERS_DIR set so the daemon finds them without an entrypoint flag.

Fixed

• Large file content (PR diffs, arXiv papers over the 512 KiB MAX_EAGER_RESPONSE_BYTES cap) no longer returns EIO. The GitHub and arXiv providers route oversized reads through fetch-blob so bytes stay host-side; the SDK gains FileContent::Blob and FileContent::BlobWithAttrs variants for blob-backed file content.
• cd /github/<owner> followed by ls no longer re-fetches the listing. The SDK's projection_exact_lookup was marking dirents non-exhaustive even when the handler returned PageStatus::Exhaustive; a new listing-exhaustive flag on proj-entry propagates the exhaustive bit into the host's projection accumulator.
• lookup_child into a bind site now dispatches correctly when parent_path equals the bind template exactly (not just when it is a strict ancestor). Previously the lookup fell through to the no-handler branch and returned NotFound.
• projection_exact_lookup was packing the looked-up target's children into lookup-entry.siblings instead of the target's actual siblings. The host was caching the listing under the wrong key. The SDK now populates siblings with the target's siblings computed from the parent's static children, with the exhaustive bit derived from StaticChildren::parent_has_dynamic_children.
• Synchronous FUSE invalidation is removed from the provider callout path, eliminating hangs when reading GitHub projected files such as issue bodies.
• DNS and other unknown-size full-read files now return complete content through cat, head, and similar tools instead of appearing empty because the kernel saw a zero or one byte sentinel before provider content was materialised.
• Credential persistence no longer leaves stale file-fallback entries after keyring writes; file-store durability and permissions are hardened.
• Host credential-store setup now logs when keyring access falls back to the file store.
• sdk-macros dev-dependency on omnifs-sdk no longer carries a redundant version specifier, fixing workspace version resolution.