Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Peer administration
Peer administration is receiver-authorized, private federation between paired instances. Installation, pairing, discovery, display names, and planning grants never authorize execution. Development and fixture tests do not create live grants, downloads, updates, or restarts.
Setup and supported actions
On the receiving host, open Instances → the peer → Peer administration. Compare
both instance UUIDs, then review a separate, action-specific execution grant.
execution-v1 requires explicit confirmation; existing planning-v1 grants
remain planning-only. The UI grants one hour; the API permits one minute through
24 hours. The audit identity is the server-derived operator session (including
delegated agent sessions), or a genuine local operator on a password-free host.
Setup and local catalog review retain requireHostControl.
The fixed action vocabulary accepts no arbitrary commands, credentials, paths, URLs, force flags, scripts, import recipes, or license-acceptance flags:
portos.update: the receiver's cleanmaincheckout, pinned to the exact preflight origin commit. Forks must be synchronized locally first. The adapter preserves fork-aware preflight and never forces, resets, or rebases divergent work. Pinned updates use fast-forward only.portos.restart: only the receiver's verified PortOS server and CoS processes. Script identities are verified; no unrelated applications are restarted.catalog.install: only supported, receiver-reviewed single-file LM Studio GGUF catalog artifacts. ConfigureLM_STUDIO_MODELS_DIRlocally to the actual LM Studio model folder; the adapter refuses a guessed default. A local review binds the catalog key, immutable source revision, filename, SHA-256, exact size, license, installed runtime, destination, and runtime memory requirement. Source repositories and actual destinations are derived by the receiver. Unknown, gated, unreviewed, changed, unsupported, or insufficient-resource entries deny. Ollama, MLX, and sharded artifacts are not supported by this execution adapter.
Catalog review is configuration, not a download, model launch, acceptance of new terms, or permission to execute. A paired sender still needs the separate catalog execution grant. Installation verifies the streamed byte count and hash before publication and never replaces an existing destination file.
On the sending host, prepare the desired action, inspect its exact signed preflight, then request execution. The sender stores the request identity and signed preview before sending and refuses dispatch if that recovery record cannot be read back. An uncertain response retains the request and offers Check status; it never starts another attempt automatically. The receiving host owns the durable ledger; browser storage retains the evidence needed to recover its receipt. A proven local pre-send rejection permits a fresh preview. A timeout or unsigned remote refusal never does.
Authority and drain
Execution grants are machine-local settings, bound to both instance UUIDs, current pair credentials, action, grant ID/generation, expiry and the non-rewound execution epoch. Grant changes use compare-and-swap on the previous grant ID and advance the PostgreSQL generation floor before publishing policy. A failed policy write therefore cannot leave old authority active.
Execution shares the instance identity writer lock through the launch handoff. Credential, peer enablement/membership, or receiver identity changes rotate the machine execution epoch before publication. This invalidates all execution grants on that receiver; review new grants afterward. Ordinary telemetry/name changes preserve grants. Restoring instance identity files also rotates the epoch under that lock. Grants and execution receipts are never restored from snapshots.
The receiver consumes a request durably before side effects. Its own maintenance hold drains admitted agents, provider work, renders, saves and cleanup naturally. It never kills or cancels active work. It claims the exact hold ID/revision using a fresh coordinator-owned idle observation, then rechecks identity, grants, version, source and resource evidence. Resume and ordinary admission share this same journal transaction; a separate lock or a ready-status read cannot launch.
A one-use in-process capability binds the exact in-flight claim to a fixed action and evidence digest. Copies, recovered claims, stale revisions and repeated use cannot launch. Revocation during drain prevents dispatch. After launch, revocation does not interrupt work already running; the receiver reconciles that operation.
Protocol and recovery
Planning endpoints retain their original behavior under
/api/federation/admin/v1/{preflight,plans,receipt,execute}. Planning execute
always refuses; upgrading never turns a temporary plan into queued work.
Separate execution POST endpoints are
/api/federation/admin/v1/execution/{preflight,dispatch,status}. Every endpoint
requires the current scoped paired credential, including on password-free hosts.
The sender never forwards Basic auth, instance passwords, sessions or agent tokens.
Exact endpoint admission does not grant general host-control access.
Execution preflights expire within 60 seconds and bind the request, both UUIDs, grant generation, receiver version, epoch, fixed intent and evidence digest. Execution signatures use canonical JSON and a separate HMAC purpose; planning wire signatures remain compatible. Dispatch returns acceptance, not completion. Changed input under a consumed request ID is refused permanently, including after restart, grant renewal or database restore. Identical retries return the existing operation; there is no second launch.
States are queued, draining, in-flight, awaiting-reconnect, succeeded,
failed and uncertain. Active ledger rows are capped at 32; preflights at 128.
Timeouts, disconnected callers, dead PIDs, and general host health never prove
completion. Unknown side effects retain exclusive ownership. Boot never replays a
launch; explicit status reads can reconcile exact completion evidence arriving
after boot. Terminal database persistence, journal settlement and hold release
are independently recoverable, including crashes between those boundaries.
Status requests can include the original signed preview. When the receiver has never consumed that request, it verifies the signature and both instance identities, then atomically records a permanent failed receipt under the same database lock as dispatch. That receipt proves no launch occurred and prevents any delayed dispatch or reused preflight from starting it later. Already accepted requests retain their original state. This recovery works after receiver restart, grant revocation and epoch rotation while the original pair credential remains valid; it never restores authority from the preview. A request without verifiable original evidence remains unresolved instead of treating a bare 404 as proof.
Updates and restarts write per-operation launch evidence before starting their detached adapter. Reconciliation requires its exact successful exit receipt and matching target evidence: restarted fixed process identities, or the expected boot/current commit with verified build, dependencies, submodules and migrations. Catalog installation requires its verified artifact and completed publication. Missing or conflicting evidence remains uncertain. No generic rollback or force Ready button exists. Proven completion releases only the operation's own hold and re-evaluates the already-saved scheduling policies; it never enables a policy.
The durable receiver ledger and restore reconciliation are documented in STORAGE. Fixtures cover authorization refusal, planning compatibility, revocation/rotation during drain, replay, resume fencing, crash gaps, delayed completion and negative resource/hash checks without operating any live peer.