infra: deploy co-located Mnemonic MCP server with webapp CORS

infrastructure/ansible/inventory/group_vars/all/features.yml

@@ -0,0 +1,36 @@
+---
+# Feature enable flags, loaded by ansible from group_vars/all (adjacent to the
+# CI-generated inventory hosts.yml). Higher precedence than role defaults,
+# lower than the play's role vars in deploy.yml.
+
+# mnemonik-server: co-located Mnemonic MCP server (Rust GHCR image + local
+# Ollama), fronted by the shared Caddy. Enabled 2026-07-09 once the GHCR
+# package went public and a build passed.
+mnemonik_server_enabled: true
+
+# Pin to the confirmed-good build rather than :latest for a reproducible
+# deploy. Bump when a newer build is verified. Package is public, so the
+# fail-soft GHCR login is skipped/ignored and the pull is anonymous.
+mnemonik_mcp_image_tag: "sha-61fdb3e"
+
+# Hosted MCP is the public participate-mode operator. Default writes should
+# anchor to Irys/Arweave + Solana; local-only writes remain available via
+# explicit mode=local.
+mnemonik_mcp_storage_mode: "full"
+
+# Wallets that have signed hosted anchored uploads. Include both the documented
+# pre-Hetzner server identity and the accidental post-migration identity so the
+# chain-backed public counters recover all indexed history.
+mnemonik_chain_stats_wallets: >-
+  DYVu4Bry3BzGVsR3Hj2iGVT5fNdWFoHw2zRxsdTmrG25,
+  EdFyDHqqWTSPKs7c89U36Uy8QAnX2nX2A7qkYKD2Td21
+
+mnemonik_arweave_url: "https://gateway.irys.xyz"
+mnemonik_chain_stats_gateway_url: "https://gateway.irys.xyz"
+mnemonik_chain_stats_graphql_url: "https://arweave.net/graphql"
+
+# Browser webapp origins. Caddy reflects the request Origin if it matches
+# one of these exact origins.
+mnemonik_mcp_cors_origin:
+  - "https://www.mnemonik.xyz"
+  - "https://mnemonik-webapp.pages.dev"

infrastructure/ansible/playbooks/deploy.yml

@@ -40,6 +40,19 @@
     APT_LISTCHANGES_FRONTEND: none
 
   pre_tasks:
+    # Recovery: a prior failed mnemonik-server deploy may have left a broken
+    # Caddy snippet in the shared conf.d. The vaultwarden role starts the
+    # shared Caddy before mnemonik-server has a chance to rewrite the snippet,
+    # so an invalid snippet blocks the entire deploy. Remove any stale
+    # mnemonik-mcp.conf; mnemonik-server will render a validated one later.
+    - name: Remove stale mnemonik-mcp Caddy snippet if present
+      ansible.builtin.file:
+        path: /opt/vaultwarden/caddy_conf_d/mnemonik-mcp.conf
+        state: absent
+      become: true
+      failed_when: false
+      changed_when: false
+
     - name: Display deployment info
       debug:
         msg: |
@@ -114,6 +127,11 @@
         vault_blogger_repo_ref:              "{{ sops.blogger_repo_ref | default('') }}"
         vault_claude_blog_repo_ref:          "{{ sops.claude_blog_repo_ref | default('') }}"
         vault_publish_callback_hmac_secrets: "{{ sops.publish_callback_hmac_secrets | default('') }}"
+        # Mnemonic MCP server (mnemonik-server role) hosted-mode secrets.
+        # Generate each once (openssl rand -base64 32) and keep stable;
+        # empty/safe default so a deploy with the role disabled never fails.
+        vault_mnemonik_mcp_jwt_secret:       "{{ sops.mnemonik_mcp_jwt_secret | default('') }}"
+        vault_mnemonik_mcp_refresh_salt:     "{{ sops.mnemonik_mcp_refresh_salt | default('') }}"
       no_log: true
 
     - name: Wait for system to stabilize post-boot
@@ -180,6 +198,21 @@
         # at the VM's public IPv4. Empty = tailnet-only (default).
         kaneo_public_domain: "{{ sops.kaneo_public_domain | default('kaneo.mnemonik.xyz') }}"
 
+    # Mnemonic protocol SERVER (Rust MCP image from GHCR + local Ollama),
+    # co-located and fronted by the shared Caddy. Distinct from the
+    # `mnemonic-mcp` role (that installs the CLIENT binary). Disabled by
+    # default: enable once DNS (mcp.mnemonik.xyz → VM IP) + sops secrets are
+    # in place — set mnemonik_server_enabled=true (group_vars or -e).
+    - role: mnemonik-server
+      tags: [mnemonik-server, mcp]
+      when: mnemonik_server_enabled | default(false) | bool
+      vars:
+        mnemonik_mcp_jwt_secret: "{{ vault_mnemonik_mcp_jwt_secret | default('') }}"
+        mnemonik_mcp_refresh_salt: "{{ vault_mnemonik_mcp_refresh_salt | default('') }}"
+        # GHCR pull auth — reuse the sops github PAT for private packages.
+        # Leave the token empty if the GHCR package is public.
+        mnemonik_ghcr_token: "{{ vault_github_deploy_pat | default('') }}"
+
     - role: telegram-init
       tags: [telegram-init, communication]
       vars:
@@ -246,10 +279,12 @@
         molyanov_skills_path: "/opt/molyanov/skills"
         molyanov_pk_guard_enabled: true
 
-    # Phase 6: Attestation + MCP (DESCOPED 2026-05 — disabled by default).
-    # The role still runs every deploy but installs only no-op stub hooks when
-    # `mnemonic_mcp_enabled` is false. Flip to true once the local Mnemonic MCP
-    # server is production-ready (backlog: work/mnemonic-attestation-integration/).
+    # Phase 6: Mnemonik MCP client binary (per-spawn stdio) for
+    # content-publisher attestation. Enabled by default; the discovered binary
+    # path is exposed as fact `mnemonic_mcp_binary` for downstream roles.
+    # (The legacy daemon + hook design was retired 2026-06; this role now
+    # installs only the npm-shipped client binary and runs an MCP handshake
+    # smoke test.)
     - role: mnemonic-mcp
       tags: [mnemonic-mcp, attestation]
       vars:
@@ -391,14 +426,14 @@
     - name: Verify all services are running
       block:
         - name: Check systemd services
-          # mnemonic-mcp only added to the loop when the role is enabled
-          # (descoped 2026-05; backlog: work/mnemonic-attestation-integration/)
+          # mnemonic-mcp is a per-spawn client binary, not a daemon; no
+          # systemd unit to check here. workspace-manager + telegram-ai-agent
+          # are the long-running fabric services.
           systemd:
             name: "{{ item }}"
             state: started
           register: service_check
-          loop: "{{ ['workspace-manager', 'telegram-ai-agent']
-                    + (['mnemonic-mcp'] if (mnemonic_mcp_enabled | default(false) | bool) else []) }}"
+          loop: "{{ ['workspace-manager', 'telegram-ai-agent'] }}"
           failed_when: false
 
         - name: Report service status

infrastructure/ansible/roles/mnemonik-server/.ansible-lint

@@ -0,0 +1,8 @@
+---
+# Role-local ansible-lint config. Mirrors the mnemonic-mcp role.
+#
+# role-name: every role in this repo uses kebab-case (mnemonic-mcp, kaneo,
+# telegram-init, ssh-hardening, ...). Matching that convention trips the
+# lint's snake_case rule; renaming is a repo-wide concern, out of scope here.
+skip_list:
+  - role-name

infrastructure/ansible/roles/mnemonik-server/README.md

@@ -0,0 +1,81 @@
+# Role: mnemonik-server
+
+Deploys the **server side** of the Mnemonic protocol, co-located on the fabric
+VM: the Rust MCP HTTP server (GHCR image `ghcr.io/mnemonik-xyz/mnemonic-mcp`)
+plus a local Ollama for `/chat`. TLS + public ingress are handled by the
+**shared Caddy** (vaultwarden role), not the monorepo's own nginx/certbot.
+
+> **Not** the `mnemonic-mcp` role. That one installs the *client* binary
+> (`@mnemonik-xyz/mcp`, per-spawn stdio) that content-publisher spawns. This
+> role hosts the *server* those clients connect to. Two sides of one protocol.
+
+## What it does
+
+1. **Swap** — idempotent `{{ mnemonik_server_swap_size_mb }}`MB swapfile
+   (ollama + fastembed headroom). Toggle with `mnemonik_server_swap_enabled`.
+2. **State on the persistent volume** — creates `<volume>/mnemonik/{data,keypair,ollama}`
+   and bind-mounts them into the containers so server identity (`id.json` in
+   `/keypair`), the attestation DB (`/data`), and the pulled model survive a
+   VM rebuild.
+3. **GHCR pull** — optional `docker login` (private package), then
+   `docker compose pull` — the box never compiles Rust.
+4. **Ollama** — stock `ollama/ollama` image + a post-up `ollama pull` of
+   `{{ mnemonik_ollama_model }}` (replicates the monorepo's custom image
+   without needing its build context) + a non-fatal warm-up.
+5. **Caddy vhost** — drops `mnemonik-mcp.conf` into the shared conf.d and
+   restarts Caddy; `{{ mnemonik_public_domain }}` gets a real Let's Encrypt
+   cert on first request.
+6. **Health gate** — polls `https://{{ mnemonik_public_domain }}/health` the
+   same way an external client would.
+
+## Boundary with the monorepo pipeline
+
+- **Build stays upstream.** `mnemonik-xyz/monorepo` → `build-mcp-image.yml`
+  builds + pushes the image to GHCR. Untouched by this role.
+- **Deploy is fabric's.** This role does what `deploy-mcp.yml`'s deploy half
+  does (pull + compose up + health gate), inside fabric's existing deploy that
+  already has authenticated access to this box — instead of maintaining the
+  monorepo's separate SSH path into a box whose port 22 is firewalled.
+- **Webapp is elsewhere.** The frontend lives on Cloudflare Pages; only the
+  MCP host is served here.
+
+## Enabling (first deploy)
+
+1. **DNS** — A record `mcp.mnemonik.xyz` → VM public IPv4.
+2. **sops** — `sops edit infrastructure/secrets/secrets.sops.yml`, set
+   `mnemonik_mcp_jwt_secret` + `mnemonik_mcp_refresh_salt`
+   (`openssl rand -base64 32` each, keep stable). Also confirm
+   `caddy_acme_email` is set so Let's Encrypt can issue.
+3. **GHCR** — make the package public, or rely on the sops github PAT wired in
+   `deploy.yml` (`mnemonik_ghcr_token`).
+4. **Enable** — set `mnemonik_server_enabled: true` (group_vars or `-e`) and
+   deploy.
+
+## Key variables
+
+| Variable | Default | Notes |
+|----------|---------|-------|
+| `mnemonik_server_enabled` | `false` | Master switch. |
+| `mnemonik_mcp_image_tag` | `latest` | Pin a real tag for reproducible redeploys. |
+| `mnemonik_ollama_model` | `qwen2.5:3b` | Pulled once, cached on the volume. |
+| `mnemonik_public_domain` | `mcp.mnemonik.xyz` | Caddy public vhost. |
+| `mnemonik_server_swap_size_mb` | `4096` | Swap cushion. |
+
+`mnemonik_mcp_jwt_secret` / `mnemonik_mcp_refresh_salt` have no default — the
+role fails loudly if enabled without them.
+
+## Cross-role contract
+
+Depends on the **vaultwarden** role for the shared Caddy container
+(`{{ vaultwarden_caddy_container_name }}`) and its external docker network
+(`{{ mnemonik_shared_caddy_network }}`). The mcp container joins that network
+with alias `mnemonik-mcp`; Caddy reverse-proxies to it by name. Must run after
+vaultwarden + kaneo in the `deploy.yml` roles list.
+
+## Smoke (manual, post-deploy)
+
+```bash
+curl -fsS https://mcp.mnemonik.xyz/health
+ssh <vm> 'docker compose -f /opt/mnemonik-server/docker-compose.yml ps'
+ssh <vm> 'docker exec mnemonik-ollama-1 ollama list'   # model present
+```

infrastructure/ansible/roles/mnemonik-server/defaults/main.yml

@@ -0,0 +1,110 @@
+---
+# Role: mnemonik-server
+#
+# Deploys the SERVER side of the Mnemonic protocol — the Rust MCP HTTP server
+# (GHCR image ghcr.io/mnemonik-xyz/mnemonic-mcp) plus a co-located Ollama for
+# /chat — co-located on the fabric VM and fronted by the existing shared Caddy
+# (vaultwarden role), NOT by the monorepo's own nginx+certbot.
+#
+# NOTE: distinct from the `mnemonic-mcp` role. That role installs the CLIENT
+# binary (@mnemonik-xyz/mcp, per-spawn stdio) consumed by content-publisher.
+# This role hosts the SERVER the protocol clients connect to.
+#
+# Upstream reference (adapted): mnemonik-xyz/monorepo docker-compose.yml +
+# mcp/deploy/compose.env.example. We keep the `mcp` + `ollama` services and
+# drop `nginx` + `certbot` (Caddy terminates TLS and reverse-proxies instead).
+
+# Master switch. Disabled by default: enable per-environment (group_vars or
+# CLI -e) once DNS + sops secrets are in place. Mirrors the mnemonic-mcp /
+# restic gating pattern.
+mnemonik_server_enabled: false
+
+# --- Deploy dir + compose project ------------------------------------------
+mnemonik_server_install_dir: /opt/mnemonik-server
+mnemonik_server_compose_project: mnemonik
+
+# --- MCP server image ------------------------------------------------------
+# CI (monorepo build-mcp-image.yml) builds + pushes this to GHCR; the fabric
+# box only PULLS — it never compiles Rust. Pin a real tag (e.g. v0.2.8) for
+# reproducible redeploys instead of `latest`.
+mnemonik_mcp_image: "ghcr.io/mnemonik-xyz/mnemonic-mcp"
+mnemonik_mcp_image_tag: "latest"
+
+# GHCR pull auth. If the package is private, the role does `docker login`
+# using this token (reuse the sops github PAT). Leave empty if the GHCR
+# package is public (no login attempted).
+mnemonik_ghcr_username: "mnemonik-dev"
+mnemonik_ghcr_token: ""
+
+# --- Ollama (co-located, local inference) ----------------------------------
+# Stock upstream image + a post-up `ollama pull` (the monorepo's custom
+# `monorepo-ollama` image is just this base + a pre-pulled model + warm-up;
+# we replicate that without needing the monorepo build context on the box).
+mnemonik_ollama_image: "ollama/ollama:0.21.2"
+mnemonik_ollama_model: "qwen2.5:3b"
+
+# --- MCP runtime env (passed into the mcp container) -----------------------
+mnemonik_mcp_storage_mode: "local"
+mnemonik_mcp_embed_provider: "fastembed"
+mnemonik_mcp_turbo_bits: 4
+mnemonik_mcp_payment_mode: "none"
+mnemonik_mcp_rust_log: "info"
+
+# fastembed cache dir inside the container (bind-mounted from persistent
+# volume via docker-compose.yml). Set explicitly so model downloads survive
+# VM rebuilds and land in a writable, pre-created directory.
+mnemonik_mcp_fastembed_cache_dir: "/data/model-cache"
+
+# Chain-backed traction stats (required by server images >= #194).
+# Defaults are safe/public; override in sops/group_vars for real tracking.
+mnemonik_chain_stats_wallets: ""
+mnemonik_solana_rpc_url: "https://api.mainnet-beta.solana.com"
+mnemonik_arweave_url: "https://gateway.irys.xyz"
+mnemonik_chain_stats_graphql_url: "https://arweave.net/graphql"
+mnemonik_chain_stats_gateway_url: "https://gateway.irys.xyz"
+# Origin advertised in OAuth metadata + challenge envelopes. MUST match the
+# public hostname external clients (Cursor/Claude.ai) connect to — RFC 8707
+# origin check is enforced. No trailing slash.
+mnemonik_mcp_public_base_url: "https://mcp.mnemonik.xyz"
+
+# CORS origins for the webapp frontend. Set to the exact origin(s) the browser
+# sends (e.g. "https://mnemonik.xyz" or ["https://www.mnemonik.xyz",
+# "https://mnemonik-webapp.pages.dev"]). Empty string or empty list disables
+# CORS headers at the Caddy layer. Do NOT use "*" with credentials.
+mnemonik_mcp_cors_origin: ""
+
+# Port the mcp server listens on inside the container. Caddy reverse-proxies
+# to this over the shared docker network; NOT published to a host port.
+mnemonik_mcp_container_port: 3000
+
+# --- Caddy vhost (public, real Let's Encrypt) ------------------------------
+# Public hostname served on 443. Requires: DNS A record → VM public IPv4,
+# firewall 80+443 open (already the case), and caddy_acme_email set so LE
+# can issue. The webapp lives on Cloudflare Pages — this is the MCP host only.
+mnemonik_public_domain: "mcp.mnemonik.xyz"
+
+# Shared Caddy — cross-role contract with the vaultwarden role. Snippet is
+# dropped into its conf.d; the container is restarted to pick it up.
+vaultwarden_caddy_container_name: "vaultwarden-caddy-1"
+mnemonik_caddy_conf_d: "/opt/vaultwarden/caddy_conf_d"
+
+# External docker network created by the vaultwarden role's compose stack.
+# The mcp container joins it so Caddy can reverse_proxy by name.
+mnemonik_shared_caddy_network: "vaultwarden_vaultwarden"
+
+# --- Swap (ollama + fastembed headroom) ------------------------------------
+# Safety cushion recommended by the monorepo bootstrap. Idempotent.
+mnemonik_server_swap_enabled: true
+mnemonik_server_swap_size_mb: 4096
+mnemonik_server_swap_path: "/swapfile"
+
+# Health probe budget (post-up). MCP start_period is 180s (fastembed model
+# load), so give it generous retries.
+mnemonik_mcp_health_retries: 120
+mnemonik_mcp_health_delay: 3
+
+# CRITICAL: provided by the playbook via sops decryption (deploy.yml).
+# Not defined here so the role fails loudly if enabled without them.
+#   mnemonik_mcp_jwt_secret    — openssl rand -base64 32, stable across deploys
+#   mnemonik_mcp_refresh_salt  — openssl rand -base64 32, rotating invalidates
+#                                every live refresh token (deliberate event)

infrastructure/ansible/roles/mnemonik-server/handlers/main.yml

@@ -0,0 +1,25 @@
+---
+- name: Restart mnemonik-server
+  # community.docker.docker_compose_v2 lacks `restarted:` on the runner's
+  # collection (see kaneo handler note). Shell out; handler only fires when
+  # a rendered template actually changed, so idempotency isn't needed here.
+  # `--force-recreate` ensures env_file/compose changes (e.g. FASTEMBED_CACHE_DIR)
+  # are picked up; `docker compose restart` alone keeps the old env in the
+  # container.
+  ansible.builtin.command:
+    cmd: docker compose up -d --force-recreate
+    chdir: "{{ mnemonik_server_install_dir }}"
+  environment:
+    COMPOSE_PROJECT_NAME: "{{ mnemonik_server_compose_project }}"
+  become: true
+  become_user: op
+  changed_when: true
+
+- name: Reload caddy for mnemonik vhost
+  # `docker restart` of the shared Caddy — the same reliable hammer the kaneo
+  # role uses (admin-API reload has flaked on this stack). Caddy re-reads
+  # Caddyfile + conf.d/*.conf on boot and issues LE certs for new vhosts.
+  become: true
+  ansible.builtin.command:
+    cmd: "docker restart {{ vaultwarden_caddy_container_name }}"
+  changed_when: true

infrastructure/ansible/roles/mnemonik-server/meta/main.yml

@@ -0,0 +1,17 @@
+---
+galaxy_info:
+  author: "Mnemonic Protocol"
+  description: "Deploy the Mnemonic MCP server (Rust GHCR image) + co-located Ollama, fronted by the shared Caddy"
+  company: "Mnemonic Protocol"
+  license: "MIT"
+  min_ansible_version: "2.15"
+  tags:
+    - docker
+    - mcp
+    - mnemonik
+    - ollama
+    - caddy
+
+dependencies: []
+# Ordering enforced by the playbook (deploy.yml roles list): must run after
+# vaultwarden (owns the shared Caddy + external network) and kaneo.