README.md

metadata

title: Alpecca Continuity Core
sdk: docker
app_port: 7860
pinned: false

Alpecca Continuity Core

This public-repository Docker Space is an on-demand fallback for Alpecca. It is not a second active CoreMind and is not an always-on guarantee. Free Hugging Face Spaces may sleep and stop executing; every enabled wake starts from a closed port and must pass verified restore, the configured promotion policy, and continuity fencing again. The image is inert unless ALPECCA_CLOUD_CORE_ENABLED=1 is explicit.

The core uses the existing alpecca.mind Hugging Face backend with exactly Qwen/Qwen3.5-9B. Qwen companion and reflection calls use non-thinking mode. The image does not run Ollama or substitute an older Qwen family.

Startup order

cloud_entrypoint.py and app.py enforce this sequence:

1. Serve a health-only standby identity on port 7860. It is explicitly not CoreMind and the phone launcher will not accept it as active Alpecca.
2. Poll authenticated authority status without loading a model or memory. A promotion attempt begins only when there is no active lease and no fresh local-primary heartbeat.
3. Create a fresh private runtime directory and install the SHA-256-locked V.4 VRM from Alpecca's Hugging Face asset dataset.
4. Fetch /archive/latest from Mindscape Vault. The existing Vault client authenticates AES-256-GCM, checks ciphertext and plaintext digests, and runs SQLite PRAGMA integrity_check before installing alpecca.db.
5. Require either a short-lived CreatorJD approval bound to the restored archive and next fencing epoch, or the explicit deployment-level unattended failover policy described below.
6. Acquire the remote continuity lease as cloud-standby. In manual mode the exact grant must consume the approval's one-use epoch; automatic mode still cannot proceed while a fresh local heartbeat or another lease exists.
7. Publish the Space HTTPS origin under that lease, then and only then spawn uvicorn server:app on port 7860.
8. Renew every 10 seconds. Lease loss sends SIGKILL to the server process group so no shutdown hook or descendant can write after the fence is gone. The supervisor then returns to health-only standby. An unexpected clean server exit also returns to standby; only a container shutdown ends the entrypoint.

There is no pre-lease CoreMind. The credential-free standby response contains only service/state flags and no memory, model output, credential, or private status. A disabled deployment leaves the process inert; missing configuration, Vault failure, promotion-policy failure, lease denial, or endpoint-publication failure returns the enabled process to standby. The local laptop remains preferred while its heartbeat is fresh.

Space settings

The phone-reachable survival endpoint uses a public Docker Space repository with Alpecca's own server authentication still enforced. Hugging Face Space secrets remain private; no credential is stored in repository files or public variables. A private Space adds a Hugging Face login wall that the launcher cannot use for unattended discovery. Required Space secrets:

• HF_TOKEN
• ALPECCA_CONTINUITY_LEASE_TOKEN
• ALPECCA_MINDSCAPE_VAULT_TOKEN
• ALPECCA_MINDSCAPE_VAULT_KEY
• ALPECCA_AUTH_SECRET
• ALPECCA_CREATOR_PASSWORD
• ALPECCA_CLOUD_RESTORE_APPROVAL only when using manual promotion

Required Space variables:

• ALPECCA_CONTINUITY_LEASE_URL
• ALPECCA_MINDSCAPE_VAULT_URL
• ALPECCA_CLOUD_CORE_ENABLED=1 after the reviewed image is deployed and the laptop is confirmed to be renewing its local lease
• ALPECCA_CLOUD_AUTO_FAILOVER=1 to authorize unattended promotion after an authenticated Vault restore; omit it to require one-use approval instead

The deployed Space uses the service-binding gateway rather than direct workers.dev URLs because Hugging Face blocks that DNS suffix in Space containers. Configure:

• ALPECCA_CONTINUITY_LEASE_URL=https://alpecca-continuity-gateway.pages.dev/lease
• ALPECCA_MINDSCAPE_VAULT_URL=https://alpecca-continuity-gateway.pages.dev/vault

Those prefixes forward to the existing Workers without copying continuity state. The final Docker image also removes repository deployment scripts, documentation, and .git; the survival container retains runtime source only.

Hugging Face supplies SPACE_HOST and SPACE_ID. SPACE_HOST becomes the published HTTPS origin. ALPECCA_PUBLIC_URL and ALPECCA_CONTINUITY_NODE_ID are optional explicit overrides. Secrets must stay in Space settings, never image layers, variables, URLs, or repository files.

Camera, screen, microphone, computer use, Discord, local voice workers, legacy Mindscape sync, and cloud Vault writes are forced off. The restored archive is read-only continuity input until those side effects can validate the exact fencing tuple at their own trust boundaries.

Restore approval

A wake without ALPECCA_CLOUD_RESTORE_APPROVAL verifies the latest archive, prints only its sequence and SHA-256 fingerprint, and exits before requesting a lease. CreatorJD or the external creator verifier combines that fingerprint with the authority's expected next epoch and issues a maximum-five-minute approval:

{
  "approvalId": "restore-20260715-001",
  "purpose": "stage-passive-restore",
  "creatorPrincipal": "CreatorJD",
  "snapshotDigest": "sha256:<verified archive fingerprint>",
  "leaseEpoch": 42,
  "issuedAt": "2026-07-15T20:00:00Z",
  "expiresAt": "2026-07-15T20:05:00Z",
  "oneUse": true,
  "verification": {
    "status": "verified",
    "verifier": "external-creator-verifier",
    "evidenceId": "creator-check-20260715-001"
  }
}

Set the compact JSON as the Space secret and restart. The authority's monotonic epoch makes the approval one-use: after acquisition, release, or expiry, a later grant has a different epoch and the old approval cannot publish or start the core. If another actor consumes the predicted epoch, query authority status and issue a new approval.

Automatic survival mode

ALPECCA_CLOUD_AUTO_FAILOVER=1 is a persistent CreatorJD deployment decision, not a bypass of the singleton fence. It removes only the short-lived approval document that cannot be supplied after an unexpected laptop power loss. The core must still authenticate and integrity-check the newest encrypted Vault archive, acquire a newer monotonic fencing epoch, publish under that exact lease, and keep renewing it. A fresh laptop heartbeat blocks acquisition, and lease loss kills the cloud server process group before another host can take over. Discord, sensors, computer use, voice workers, and Vault writes remain disabled in this survival core.

Image and source

The standalone Space Dockerfile clones the reviewed source branch codex/voice-session-audio-normalization, runs the House HQ production build, then copies this supervisor from the Space repository. Override ALPECCA_GIT_REF only with another reviewed ref.

cloud_entrypoint.install_vrm() fetches V.4 from Alpecca's Hugging Face runtime-assets dataset and verifies its locked SHA-256 before restore or lease acquisition. A failed or changed body blocks promotion. Art remains on Hugging Face and is never uploaded to Cloudflare by this scaffold.

Network-free tests

python -m pytest -q tests\test_hf_cloud_core.py

The tests use fake Vault, lease, and child-process adapters. They do not contact Hugging Face, Cloudflare, GitHub, or Mindscape Vault.