Pass 2 bring over browser resources and http resource cert gen

Fix #1226

.claude/skills/crud-endpoints/SKILL.md

@@ -0,0 +1,31 @@
+---
+name: crud-endpoints
+description: Use whenever asked to add, create, or scaffold a CRUD endpoint, router, or entity in this repo's server (create/list/get/update/delete handlers, new `server/routers/<entity>/` or `server/private/routers/<entity>/` folder). Points to the established file layout, middleware, ActionsEnum, and route-registration conventions before writing any code.
+---
+
+Before writing any router/handler/middleware code for a new entity, read
+`docs/crud-endpoints.md` in full. It documents, with real examples from
+`server/routers/aiProvider/` (public) and `server/private/routers/alertRule/`
+(enterprise-only), how this repo structures CRUD endpoints:
+
+- Directory/file layout per entity (`index.ts`, `types.ts`, `validation.ts`,
+  one file per operation).
+- The standard handler anatomy (zod parsing, OpenAPI registry, response
+  envelope, error handling).
+- Where access-control middleware (`verify<Entity>Access`) lives and when
+  it's needed vs. plain `verifyOrgAccess`.
+- How to wire up `ActionsEnum` entries, `verifyUserHasAction`, and
+  `logActionAudit`.
+- Which of the four router files (`server/routers/external.ts`,
+  `server/routers/internal.ts`, `server/private/routers/external.ts`,
+  `server/private/routers/internal.ts`) to register routes in, and the
+  middleware chain template per HTTP verb.
+- The repo's non-standard verb convention: **`PUT` = create, `POST` =
+  update** (backwards from typical REST) — don't "fix" this to standard
+  REST verbs, match the existing convention.
+- The `#dynamic` import alias, for the rare case of a hook needing different
+  implementations in OSS vs. enterprise builds.
+
+Follow that doc's checklist (§8) step by step rather than improvising a
+structure. If the doc and the actual code in `aiProvider`/`alertRule` ever
+disagree, trust the code and flag the doc as stale.

.cursor/rules/Button-loading-state.mdc

@@ -0,0 +1,5 @@
+---
+alwaysApply: true
+---
+
+When adding submit buttons, don't change the text of the button during the loading state. Text should stay static and you should use the loading prop on the button.

.cursor/rules/Components.mdc

@@ -0,0 +1,5 @@
+---
+alwaysApply: true
+---
+
+When creating UI for popup dialogs or modals, use the Credenza componennt. This component is mobile responsive and works on desktop and wraps the dialog component and sheet into one.

.cursor/rules/Migrations.mdc

@@ -0,0 +1,5 @@
+---
+alwaysApply: true
+---
+
+Don't write or edit migrations in `server/setup` unless specificall instructed to do so.

.cursor/rules/TypeScript-rules.mdc

@@ -0,0 +1,7 @@
+---
+alwaysApply: true
+---
+
+When writing TypeScript:
+
+Prefer to use types instead of interfaces.

.cursor/rules/Use-React-form-and-Zod-schemas.mdc

@@ -0,0 +1,5 @@
+---
+alwaysApply: true
+---
+
+When creating forms, use React form for validation and use Zod schemas.

.dockerignore

@@ -34,3 +34,5 @@ build.ts
 tsconfig.json
 Dockerfile*
 drizzle.config.ts
+allowedDevOrigins.json
+scratch/

.github/FUNDING.yml

@@ -1,3 +0,0 @@
-# These are supported funding model platforms
-
-github: [fosrl]

.github/ISSUE_TEMPLATE/1.bug_report.yml

@@ -14,12 +14,13 @@ body:
           label: Environment
           description: Please fill out the relevant details below for your environment.
           value: |
-              - OS Type & Version: (e.g., Ubuntu 22.04)
+              - OS Type & Version:
               - Pangolin Version:
+              - Edition (Community or Enterprise):
               - Gerbil Version:
               - Traefik Version:
               - Newt Version:
-              - Olm Version: (if applicable)
+              - Client Version:
       validations:
           required: true
 

.github/dependabot.yml

@@ -1,52 +1,42 @@
 version: 2
+
 updates:
   - package-ecosystem: "npm"
     directory: "/"
     schedule:
       interval: "daily"
+    open-pull-requests-limit: 1
     groups:
-      dev-patch-updates:
-        dependency-type: "development"
-        update-types:
-          - "patch"
-      dev-minor-updates:
-        dependency-type: "development"
-        update-types:
-          - "minor"
-      prod-patch-updates:
-        dependency-type: "production"
-        update-types:
-          - "patch"
-      prod-minor-updates:
-        dependency-type: "production"
-        update-types:
-          - "minor"
-          
+      npm-dependencies:
+        patterns:
+          - "*"
+
   - package-ecosystem: "docker"
     directory: "/"
     schedule:
       interval: "daily"
+    open-pull-requests-limit: 1
     groups:
-      patch-updates:
-        update-types:
-          - "patch"
-      minor-updates:
-        update-types:
-          - "minor"
+      docker-dependencies:
+        patterns:
+          - "*"
 
   - package-ecosystem: "github-actions"
     directory: "/"
     schedule:
       interval: "weekly"
+    open-pull-requests-limit: 1
+    groups:
+      github-actions-dependencies:
+        patterns:
+          - "*"
 
   - package-ecosystem: "gomod"
     directory: "/install"
     schedule:
       interval: "daily"
+    open-pull-requests-limit: 1
     groups:
-      patch-updates:
-        update-types:
-          - "patch"
-      minor-updates:
-        update-types:
-          - "minor"
+      go-install-dependencies:
+        patterns:
+          - "*"

.github/workflows/cicd.yml

@@ -62,7 +62,7 @@ jobs:
 
         steps:
             - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+              uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
             - name: Monitor storage space
               run: |
@@ -77,7 +77,7 @@ jobs:
                   fi
 
             - name: Log in to Docker Hub
-              uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
+              uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4.5.1
               with:
                   registry: docker.io
                   username: ${{ secrets.DOCKER_HUB_USERNAME }}
@@ -134,7 +134,7 @@ jobs:
 
         steps:
             - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+              uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
             - name: Monitor storage space
               run: |
@@ -149,7 +149,7 @@ jobs:
                   fi
 
             - name: Log in to Docker Hub
-              uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
+              uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4.5.1
               with:
                   registry: docker.io
                   username: ${{ secrets.DOCKER_HUB_USERNAME }}
@@ -201,10 +201,10 @@ jobs:
         timeout-minutes: 30
         steps:
             - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+              uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
             - name: Log in to Docker Hub
-              uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
+              uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4.5.1
               with:
                   registry: docker.io
                   username: ${{ secrets.DOCKER_HUB_USERNAME }}
@@ -256,7 +256,7 @@ jobs:
 
         steps:
             - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+              uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
             - name: Extract tag name
               id: get-tag
@@ -264,7 +264,7 @@ jobs:
               shell: bash
 
             - name: Install Go
-              uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6.4.0
+              uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
               with:
                   go-version: 1.25
 
@@ -407,7 +407,7 @@ jobs:
               shell: bash
 
             - name: Login to GitHub Container Registry (for cosign)
-              uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
+              uses: docker/login-action@abd2ef45e78c5afb21d64d4ca52ee8550d9572c7 # v4.5.1
               with:
                 registry: ghcr.io
                 username: ${{ github.actor }}
@@ -415,7 +415,9 @@ jobs:
 
             - name: Install cosign
               # cosign is used to sign container images using keyless (OIDC) signing
-              uses: sigstore/cosign-installer@cad07c2e89fa2edd6e2d7bab4c1aa38e53f76003 # v4.1.1
+              uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
+              with:
+                cosign-release: v3.0.6
 
             - name: Sign (GHCR, keyless)
               # Sign each GHCR image by digest using keyless (OIDC) signing via Sigstore/Rekor.

.github/workflows/linting.yml

@@ -21,10 +21,10 @@ jobs:
         runs-on: ubuntu-latest
         steps:
             - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+              uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
             - name: Set up Node.js
-              uses: actions/setup-node@53b83947a5a98c8d113130e565377fae1a50d02f # v6.3.0
+              uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
               with:
                 node-version: '24'
 

.github/workflows/mirror.yaml

@@ -23,7 +23,7 @@ jobs:
           skopeo --version
 
       - name: Install cosign
-        uses: sigstore/cosign-installer@cad07c2e89fa2edd6e2d7bab4c1aa38e53f76003 # v4.1.1
+        uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
 
       - name: Input check
         run: |

.github/workflows/restart-runners.yml

@@ -1,39 +0,0 @@
-name: Restart Runners
-
-on:
-  schedule:
-    - cron: '0 0 */7 * *'
-
-permissions:
-    id-token: write
-    contents: read
-
-jobs:
-  ec2-maintenance-prod:
-    runs-on: ubuntu-latest
-    permissions: write-all
-    steps:
-      - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@v6
-        with:
-          role-to-assume: arn:aws:iam::${{ secrets.AWS_ACCOUNT_ID }}:role/${{ secrets.AWS_ROLE_NAME }}
-          role-duration-seconds: 3600
-          aws-region: ${{ secrets.AWS_REGION }}
-
-      - name: Verify AWS identity
-        run: aws sts get-caller-identity
-
-      - name: Start EC2 instance
-        run: |
-          aws ec2 start-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_ARM_RUNNER }}
-          aws ec2 start-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_AMD_RUNNER }}
-          echo "EC2 instances started"
-
-      - name: Wait
-        run: sleep 600
-
-      - name: Stop EC2 instance
-        run: |
-          aws ec2 stop-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_ARM_RUNNER }}
-          aws ec2 stop-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_AMD_RUNNER }}
-          echo "EC2 instances stopped"

.github/workflows/saas.yml

@@ -1,160 +0,0 @@
-name: SAAS Pipeline
-
-# CI/CD workflow for building, publishing, mirroring, signing container images and building release binaries.
-# Actions are pinned to specific SHAs to reduce supply-chain risk. This workflow triggers on tag push events.
-
-permissions:
-  contents: read
-  packages: write      # for GHCR push
-  id-token: write      # for Cosign Keyless (OIDC) Signing
-
-on:
-    push:
-        tags:
-            - "[0-9]+.[0-9]+.[0-9]+-s.[0-9]+"
-
-concurrency:
-  group: ${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-    pre-run:
-        runs-on: ubuntu-latest
-        permissions: write-all
-        steps:
-            - name: Configure AWS credentials
-              uses: aws-actions/configure-aws-credentials@v6
-              with:
-                  role-to-assume: arn:aws:iam::${{ secrets.AWS_ACCOUNT_ID }}:role/${{ secrets.AWS_ROLE_NAME }}
-                  role-duration-seconds: 3600
-                  aws-region: ${{ secrets.AWS_REGION }}
-
-            - name: Verify AWS identity
-              run: aws sts get-caller-identity
-
-            - name: Start EC2 instances
-              run: |
-                  aws ec2 start-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_ARM_RUNNER }}
-                  echo "EC2 instances started"
-
-
-    release-arm:
-        name: Build and Release (ARM64)
-        runs-on: [self-hosted, linux, arm64, us-east-1]
-        needs: [pre-run]
-        if: >-
-            ${{
-                needs.pre-run.result == 'success'
-            }}
-        # Job-level timeout to avoid runaway or stuck runs
-        timeout-minutes: 120
-        env:
-          # Target images
-          AWS_IMAGE: ${{ secrets.aws_account_id }}.dkr.ecr.us-east-1.amazonaws.com/${{ github.event.repository.name }}
-
-        steps:
-            - name: Checkout code
-              uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
-
-            - name: Download MaxMind GeoLite2 databases
-              env:
-                MAXMIND_LICENSE_KEY: ${{ secrets.MAXMIND_LICENSE_KEY }}
-              run: |
-                echo "Downloading MaxMind GeoLite2 databases..."
-
-                # Download GeoLite2-Country
-                curl -L "https://download.maxmind.com/app/geoip_download?edition_id=GeoLite2-Country&license_key=${MAXMIND_LICENSE_KEY}&suffix=tar.gz" \
-                  -o GeoLite2-Country.tar.gz
-
-                # Download GeoLite2-ASN
-                curl -L "https://download.maxmind.com/app/geoip_download?edition_id=GeoLite2-ASN&license_key=${MAXMIND_LICENSE_KEY}&suffix=tar.gz" \
-                  -o GeoLite2-ASN.tar.gz
-
-                # Extract the .mmdb files
-                tar -xzf GeoLite2-Country.tar.gz --strip-components=1 --wildcards '*.mmdb'
-                tar -xzf GeoLite2-ASN.tar.gz --strip-components=1 --wildcards '*.mmdb'
-
-                # Verify files exist
-                if [ ! -f "GeoLite2-Country.mmdb" ]; then
-                  echo "ERROR: Failed to download GeoLite2-Country.mmdb"
-                  exit 1
-                fi
-
-                if [ ! -f "GeoLite2-ASN.mmdb" ]; then
-                  echo "ERROR: Failed to download GeoLite2-ASN.mmdb"
-                  exit 1
-                fi
-
-                # Clean up tar files
-                rm -f GeoLite2-Country.tar.gz GeoLite2-ASN.tar.gz
-
-                echo "MaxMind databases downloaded successfully"
-                ls -lh GeoLite2-*.mmdb
-
-            - name: Monitor storage space
-              run: |
-                  THRESHOLD=75
-                  USED_SPACE=$(df / | grep / | awk '{ print $5 }' | sed 's/%//g')
-                  echo "Used space: $USED_SPACE%"
-                  if [ "$USED_SPACE" -ge "$THRESHOLD" ]; then
-                      echo "Used space is below the threshold of 75% free. Running Docker system prune."
-                      echo y | docker system prune -a
-                  else
-                      echo "Storage space is above the threshold. No action needed."
-                  fi
-
-            - name: Configure AWS credentials
-              uses: aws-actions/configure-aws-credentials@v6
-              with:
-                  role-to-assume: arn:aws:iam::${{ secrets.aws_account_id }}:role/${{ secrets.AWS_ROLE_NAME }}
-                  role-duration-seconds: 3600
-                  aws-region: ${{ secrets.AWS_REGION }}
-
-            - name: Login to Amazon ECR
-              id: login-ecr
-              uses: aws-actions/amazon-ecr-login@v2
-
-            - name: Extract tag name
-              id: get-tag
-              run: echo "TAG=${GITHUB_REF#refs/tags/}" >> $GITHUB_ENV
-              shell: bash
-
-            - name: Update version in package.json
-              run: |
-                  TAG=${{ env.TAG }}
-                  sed -i "s/export const APP_VERSION = \".*\";/export const APP_VERSION = \"$TAG\";/" server/lib/consts.ts
-                  cat server/lib/consts.ts
-              shell: bash
-
-            - name: Build and push Docker images (Docker Hub - ARM64)
-              run: |
-                  TAG=${{ env.TAG }}
-                  make build-saas tag=$TAG
-                  echo "Built & pushed ARM64 images to: ${{ env.AWS_IMAGE }}:${TAG}"
-              shell: bash
-
-    post-run:
-        needs: [pre-run, release-arm]
-        if: >-
-            ${{
-                always() &&
-                needs.pre-run.result == 'success' &&
-                (needs.release-arm.result == 'success' || needs.release-arm.result == 'skipped' || needs.release-arm.result == 'failure')
-            }}
-        runs-on: ubuntu-latest
-        permissions: write-all
-        steps:
-            - name: Configure AWS credentials
-              uses: aws-actions/configure-aws-credentials@v6
-              with:
-                  role-to-assume: arn:aws:iam::${{ secrets.AWS_ACCOUNT_ID }}:role/${{ secrets.AWS_ROLE_NAME }}
-                  role-duration-seconds: 3600
-                  aws-region: ${{ secrets.AWS_REGION }}
-
-            - name: Verify AWS identity
-              run: aws sts get-caller-identity
-
-            - name: Stop EC2 instances
-              run: |
-                  aws ec2 stop-instances --instance-ids ${{ secrets.EC2_INSTANCE_ID_ARM_RUNNER }}
-                  echo "EC2 instances stopped"

.github/workflows/stale-bot.yml

@@ -14,7 +14,7 @@ jobs:
   stale:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@b5d41d4e1d5dceea10e7104786b73624c18a190f # v10.2.0
+      - uses: actions/stale@1e223db275d687790206a7acac4d1a11bd6fe629 # v10.4.0
         with:
           days-before-stale: 14
           days-before-close: 14

.github/workflows/test.yml

@@ -14,10 +14,10 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Node
-        uses: actions/setup-node@53b83947a5a98c8d113130e565377fae1a50d02f # v6.3.0
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
 
@@ -62,7 +62,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Build Docker image sqlite
         run: make dev-build-sqlite
@@ -71,7 +71,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repository
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Build Docker image pg
         run: make dev-build-pg

.gitignore

@@ -17,9 +17,9 @@ yarn-error.log*
 *.tsbuildinfo
 next-env.d.ts
 *.db
-*.sqlite
+*.sqlite*
 !Dockerfile.sqlite
-*.sqlite3
+*.sqlite3*
 *.log
 .machinelogs*.json
 *-audit.json
@@ -54,3 +54,5 @@ hydrateSaas.ts
 CLAUDE.md
 drizzle.config.ts
 server/setup/migrations.ts
+solo.yml
+allowedDevOrigins.json

.vscode/settings.json

@@ -18,5 +18,8 @@
     "[json]": {
         "editor.defaultFormatter": "esbenp.prettier-vscode"
     },
-    "editor.formatOnSave": true
+    "editor.formatOnSave": true,
+    "cSpell.words": [
+        "nessicary"
+    ]
 }

Dockerfile

@@ -1,5 +1,5 @@
-# FROM node:24-slim AS base
-FROM public.ecr.aws/docker/library/node:24-slim AS base
+# FROM node:24.18.1-slim AS base
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS base
 
 WORKDIR /app
 
@@ -32,8 +32,8 @@ FROM base AS builder
 
 RUN npm ci --omit=dev
 
-# FROM node:24-slim AS runner
-FROM public.ecr.aws/docker/library/node:24-slim AS runner
+# FROM node:24.18.1-slim AS runner
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS runner
 
 WORKDIR /app
 

Dockerfile.dev

@@ -1,4 +1,4 @@
-FROM node:24-alpine
+FROM node:24.18.1-alpine
 
 WORKDIR /app
 

README.md

@@ -41,7 +41,7 @@
     </strong>
 </p>
 
-Pangolin is an open-source, identity-based remote access platform built on WireGuard® that enables secure, seamless connectivity to private and public resources. Pangolin combines reverse proxy and VPN capabilities into one platform, providing browser-based access to web applications and client-based access to any private resources with NAT traversal, all with granular access controls.
+Pangolin is an open-source, identity-based remote access platform built on WireGuard® that enables secure connectivity to infrastructure anywhere. It combines reverse-proxy and VPN capabilities into one platform, providing browser-based access to web applications and client-based access to private resources with NAT traversal, all with granular access control.
 
 ## Installation
 
@@ -63,11 +63,26 @@ Pangolin is an open-source, identity-based remote access platform built on WireG
 
 Pangolin's site connectors provide gateways into networks so you can access any networked resources. Sites use outbound tunnels and intelligent NAT traversal to make networks behind restrictive firewalls available for authorized access without public IPs or open ports. Easily deploy a site as a binary or container on any platform.
 
+* Lightweight user-space connector runs anywhere
+* Punches through any firewall
+* Doesn't require open ports or a public IP
+* Strict network segmentation
+* WireGuard-based
+* Get alerts when a device or network resource goes down
+
 <img src="public/screenshots/sites.png" alt="Sites" width="100%" />
 
 ### Browser-based reverse proxy access
 
-Expose web applications through identity and context-aware tunneled reverse proxies. Users access applications through any web browser with authentication and granular access control without installing a client. Pangolin handles routing, load balancing, health checking, and automatic SSL certificates without exposing your network directly to the internet.
+Expose HTTPS web applications and connect to VNC, RDP, and SSH entirely in the browser through identity and context-aware tunneled reverse proxies. Users access resources with authentication and granular access control without installing a client. Pangolin handles routing, load balancing, health checking, and automatic SSL certificates without exposing your network directly to the internet.
+
+* Expose a web panel anywhere
+* Access via any web browser
+* Single sign-on across all resources
+* HTTPS resources
+* Remote desktop in the browser with VNC and RDP
+* In-browser SSH terminal with privileged access management (PAM)
+* PIN codes, passcodes, email OTP, geoblocking, allow-lists, and more
 
 <img src="public/clip.gif" alt="Reverse proxy access" width="100%" />
 
@@ -75,14 +90,35 @@ Expose web applications through identity and context-aware tunneled reverse prox
 
 Access private resources like SSH servers, databases, RDP, and entire network ranges through Pangolin clients. Intelligent NAT traversal enables connections even through restrictive firewalls, while DNS aliases provide friendly names and fast connections to resources across all your sites. Add redundancy by routing traffic through multiple connectors in your network.
 
+* Peer-to-peer with intelligent NAT traversal
+* Hosts/IPs and port ranges
+* Network ranges/CIDRs
+* Friendly DNS aliases for network addresses
+* Privileged access management (PAM) with SSH resources
+* Private HTTPS resources only accessible on the private network
+
 <img src="public/screenshots/private-resources.png" alt="Private resources" width="100%" />
 
 ### Give users and roles access to resources
 
-Use Pangolin's built in users or bring your own identity provider and set up role based access control (RBAC). Grant users access to specific resources, not entire networks. Unlike traditional VPNs that expose full network access, Pangolin's zero-trust model ensures users can only reach the applications, services, and routes you explicitly define.
+Use Pangolin's built-in users or bring your own identity provider and set up role-based access control (RBAC). Grant users access to specific resources, not entire networks. Unlike traditional VPNs that expose full network access, Pangolin's zero-trust model ensures users can only reach the applications, services, and routes you explicitly define.
+
+* Bring your existing identity provider (IdP) or use Pangolin identities
+* Sync users and roles from your IdP
+* User- and role-based access control
+* Full network audit and access logs
 
 <img src="public/screenshots/users.png" alt="Users from identity provider with roles" width="100%" />
 
+### Find and launch resources from a personalized home page
+
+Give users a landing page to quickly find and open the resources they can access. Resources are grouped by site or label, searchable, and filterable, with grid or list views. Saved views capture filters, grouping, and layout as personal or organization-wide defaults.
+
+* Single place for admins and non-admins to see accessible resources
+* Create reusable views for common access patterns
+
+<img src="public/screenshots/resource-launcher.png" alt="Resource Launcher" width="100%" />
+
 ## Download Clients
 
 Download the Pangolin client for your platform:
@@ -107,7 +143,7 @@ the docs to illustrate some basic ideas.
 
 ## Licensing
 
-Pangolin is dual licensed under the AGPL-3 and the [Fossorial Commercial License](https://pangolin.net/fcl.html). For inquiries about commercial licensing, please contact us at [contact@pangolin.net](mailto:contact@pangolin.net).
+Pangolin is dual licensed under the AGPL-3 and the [Fossorial Commercial License](https://pangolin.net/fcl). For inquiries about commercial licensing, please contact us at [contact@pangolin.net](mailto:contact@pangolin.net).
 
 ## Contributions
 

cli/commands/rotateServerSecret.ts

@@ -1,5 +1,5 @@
 import { CommandModule } from "yargs";
-import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions } from "@server/db";
+import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions, aiProviders, virtualApiKeys } from "@server/db";
 import { encrypt, decrypt } from "@server/lib/crypto";
 import { configFilePath1, configFilePath2 } from "@server/lib/consts";
 import { eq } from "drizzle-orm";
@@ -132,12 +132,16 @@ export const rotateServerSecret: CommandModule<
             const certs = await db.select().from(certificates);
             const streamingDestinations = await db.select().from(eventStreamingDestinations);
             const webhookActions = await db.select().from(alertWebhookActions);
+            const providers = await db.select().from(aiProviders);
+            const virtualKeys = await db.select().from(virtualApiKeys);
 
             console.log(`Found ${idpConfigs.length} OIDC IdP configuration(s)`);
             console.log(`Found ${licenseKeys.length} license key(s)`);
             console.log(`Found ${certs.length} certificate(s)`);
             console.log(`Found ${streamingDestinations.length} event streaming destination(s)`);
             console.log(`Found ${webhookActions.length} alert webhook action(s)`);
+            console.log(`Found ${providers.length} AI provider(s)`);
+            console.log(`Found ${virtualKeys.length} virtual API key(s)`);
 
             // Prepare all decrypted and re-encrypted values
             console.log("\nDecrypting and re-encrypting values...");
@@ -171,11 +175,24 @@ export const rotateServerSecret: CommandModule<
                 encryptedConfig: string;
             };
 
+            type AiProviderUpdate = {
+                providerId: number;
+                encryptedApiKey: string | null;
+                encryptedHeaders: string | null;
+            };
+
+            type VirtualApiKeyUpdate = {
+                virtualApiKeyId: string;
+                encryptedToken: string;
+            };
+
             const idpUpdates: IdpUpdate[] = [];
             const licenseKeyUpdates: LicenseKeyUpdate[] = [];
             const certUpdates: CertUpdate[] = [];
             const streamingDestinationUpdates: StreamingDestinationUpdate[] = [];
             const webhookActionUpdates: WebhookActionUpdate[] = [];
+            const aiProviderUpdates: AiProviderUpdate[] = [];
+            const virtualApiKeyUpdates: VirtualApiKeyUpdate[] = [];
 
             // Process idpOidcConfig entries
             for (const idpConfig of idpConfigs) {
@@ -306,6 +323,60 @@ export const rotateServerSecret: CommandModule<
                 }
             }
 
+            // Process aiProviders entries (apiKey + headers)
+            for (const provider of providers) {
+                try {
+                    if (!provider.apiKey && !provider.headers) {
+                        continue;
+                    }
+
+                    const encryptedApiKey = provider.apiKey
+                        ? encrypt(decrypt(provider.apiKey, oldSecret), newSecret)
+                        : null;
+                    const encryptedHeaders = provider.headers
+                        ? encrypt(
+                              decrypt(provider.headers, oldSecret),
+                              newSecret
+                          )
+                        : null;
+
+                    aiProviderUpdates.push({
+                        providerId: provider.providerId,
+                        encryptedApiKey,
+                        encryptedHeaders
+                    });
+                } catch (error) {
+                    console.error(
+                        `Error processing AI provider ${provider.providerId}:`,
+                        error
+                    );
+                    throw error;
+                }
+            }
+
+            // Process virtualApiKeys entries (token)
+            for (const key of virtualKeys) {
+                try {
+                    if (!key.token) {
+                        continue;
+                    }
+
+                    virtualApiKeyUpdates.push({
+                        virtualApiKeyId: key.virtualApiKeyId,
+                        encryptedToken: encrypt(
+                            decrypt(key.token, oldSecret),
+                            newSecret
+                        )
+                    });
+                } catch (error) {
+                    console.error(
+                        `Error processing virtual API key ${key.virtualApiKeyId}:`,
+                        error
+                    );
+                    throw error;
+                }
+            }
+
             // Perform all database updates in a single transaction
             console.log("\nUpdating database in transaction...");
             await db.transaction(async (trx) => {
@@ -376,6 +447,32 @@ export const rotateServerSecret: CommandModule<
                             )
                         );
                 }
+
+                // Update AI provider entries
+                for (const update of aiProviderUpdates) {
+                    await trx
+                        .update(aiProviders)
+                        .set({
+                            apiKey: update.encryptedApiKey,
+                            headers: update.encryptedHeaders
+                        })
+                        .where(eq(aiProviders.providerId, update.providerId));
+                }
+
+                // Update virtual API key entries
+                for (const update of virtualApiKeyUpdates) {
+                    await trx
+                        .update(virtualApiKeys)
+                        .set({
+                            token: update.encryptedToken
+                        })
+                        .where(
+                            eq(
+                                virtualApiKeys.virtualApiKeyId,
+                                update.virtualApiKeyId
+                            )
+                        );
+                }
             });
 
             console.log(`Rotated ${idpUpdates.length} OIDC IdP configuration(s)`);
@@ -383,6 +480,8 @@ export const rotateServerSecret: CommandModule<
             console.log(`Rotated ${certUpdates.length} certificate(s)`);
             console.log(`Rotated ${streamingDestinationUpdates.length} event streaming destination(s)`);
             console.log(`Rotated ${webhookActionUpdates.length} alert webhook action(s)`);
+            console.log(`Rotated ${aiProviderUpdates.length} AI provider(s)`);
+            console.log(`Rotated ${virtualApiKeyUpdates.length} virtual API key(s)`);
 
             // Update config file with new secret
             console.log("\nUpdating config file...");
@@ -402,6 +501,7 @@ export const rotateServerSecret: CommandModule<
             console.log(`  - Certificates: ${certUpdates.length}`);
             console.log(`  - Event streaming destinations: ${streamingDestinationUpdates.length}`);
             console.log(`  - Alert webhook actions: ${webhookActionUpdates.length}`);
+            console.log(`  - AI providers: ${aiProviderUpdates.length}`);
             console.log(
                 `\n  IMPORTANT: Restart the server for the new secret to take effect.`
             );

cli/commands/setServerAdmin.ts

@@ -0,0 +1,85 @@
+import { CommandModule } from "yargs";
+import { db, users } from "@server/db";
+import { eq } from "drizzle-orm";
+
+type SetServerAdminArgs = {
+    email: string;
+    remove: boolean;
+};
+
+export const setServerAdmin: CommandModule<{}, SetServerAdminArgs> = {
+    command: "set-server-admin",
+    describe: "Add or remove server admin by email address",
+    builder: (yargs) => {
+        return yargs
+            .option("email", {
+                type: "string",
+                demandOption: true,
+                describe: "User email address"
+            })
+            .option("remove", {
+                type: "boolean",
+                default: false,
+                describe: "Remove server admin status from the user"
+            });
+    },
+    handler: async (argv: SetServerAdminArgs) => {
+        try {
+            const email = argv.email.trim().toLowerCase();
+
+            const [user] = await db
+                .select()
+                .from(users)
+                .where(eq(users.email, email))
+                .limit(1);
+
+            if (!user) {
+                console.error(`User with email '${email}' not found`);
+                process.exit(1);
+            }
+
+            if (argv.remove) {
+                if (!user.serverAdmin) {
+                    console.log(`User '${email}' is not a server admin`);
+                    process.exit(0);
+                }
+
+                const serverAdmins = await db
+                    .select()
+                    .from(users)
+                    .where(eq(users.serverAdmin, true));
+
+                if (serverAdmins.length <= 1) {
+                    console.error(
+                        "Cannot remove server admin: at least one server admin must exist"
+                    );
+                    process.exit(1);
+                }
+
+                await db
+                    .update(users)
+                    .set({ serverAdmin: false })
+                    .where(eq(users.userId, user.userId));
+
+                console.log(`Server admin status removed from user '${email}'`);
+                process.exit(0);
+            }
+
+            if (user.serverAdmin) {
+                console.log(`User '${email}' is already a server admin`);
+                process.exit(0);
+            }
+
+            await db
+                .update(users)
+                .set({ serverAdmin: true })
+                .where(eq(users.userId, user.userId));
+
+            console.log(`User '${email}' has been marked as a server admin`);
+            process.exit(0);
+        } catch (error) {
+            console.error("Error:", error);
+            process.exit(1);
+        }
+    }
+};

cli/index.ts

@@ -11,6 +11,7 @@ import { deleteClient } from "./commands/deleteClient";
 import { generateOrgCaKeys } from "./commands/generateOrgCaKeys";
 import { clearCertificates } from "./commands/clearCertificates";
 import { disableUser2fa } from "./commands/disableUser2fa";
+import { setServerAdmin } from "./commands/setServerAdmin";
 
 yargs(hideBin(process.argv))
     .scriptName("pangctl")
@@ -23,5 +24,6 @@ yargs(hideBin(process.argv))
     .command(generateOrgCaKeys)
     .command(clearCertificates)
     .command(disableUser2fa)
+    .command(setServerAdmin)
     .demandCommand()
     .help().argv;

compose.example.yaml

@@ -41,7 +41,7 @@ services:
       - 80:80 # Port for traefik because of the network_mode
 
   traefik:
-    image: traefik:v3.6
+    image: traefik:v3.7
     container_name: traefik
     restart: unless-stopped
     network_mode: service:gerbil # Ports appear on the gerbil service

config/traefik/traefik_config.yml

@@ -1,54 +1,47 @@
 api:
   insecure: true
   dashboard: true
-
 providers:
   http:
-    endpoint: "http://pangolin:3001/api/v1/traefik-config"
-    pollInterval: "5s"
+    endpoint: http://pangolin:3001/api/v1/traefik-config
+    pollInterval: 5s
   file:
-    filename: "/etc/traefik/dynamic_config.yml"
-
+    filename: /etc/traefik/dynamic_config.yml
 experimental:
   plugins:
     badger:
-      moduleName: "github.com/fosrl/badger"
-      version: "{{.BadgerVersion}}"
-
+      moduleName: github.com/fosrl/badger
+      version: v1.4.1
 log:
-  level: "INFO"
-  format: "common"
+  level: INFO
+  format: common
   maxSize: 100
   maxBackups: 3
   maxAge: 3
   compress: true
-
 certificatesResolvers:
   letsencrypt:
     acme:
       httpChallenge:
         entryPoint: web
-      email: "{{.LetsEncryptEmail}}"
-      storage: "/letsencrypt/acme.json"
-      caServer: "https://acme-v02.api.letsencrypt.org/directory"
-
+      email: '{{.LetsEncryptEmail}}'
+      storage: /letsencrypt/acme.json
+      caServer: https://acme-v02.api.letsencrypt.org/directory
 entryPoints:
   web:
-    address: ":80"
+    address: ':80'
   websecure:
-    address: ":443"
+    address: ':443'
     transport:
       respondingTimeouts:
-        readTimeout: "30m"
+        readTimeout: 30m
     http:
       tls:
-        certResolver: "letsencrypt"
+        certResolver: letsencrypt
       encodedCharacters:
         allowEncodedSlash: true
         allowEncodedQuestionMark: true
-
 serversTransport:
   insecureSkipVerify: true
-
 ping:
-  entryPoint: "web"
+  entryPoint: web

docs/ai-gateway-provider-selection.md

@@ -0,0 +1,285 @@
+# AI Gateway Provider Selection
+
+How the AI gateway picks which attached provider handles a request when an
+inference resource has more than one AI provider.
+
+**Code:**
+
+- Route → capability binding: `server/routers/aiGateway/createAiGatewayRouter.ts`
+- Request pipeline: `server/routers/aiGateway/pipeline.ts` (`selectProvider`)
+- Tie-break scoring: `server/lib/aiProviderSelection.ts`
+- Allow/block matching: `server/lib/aiModelKeyMatch.ts`
+- Model catalog: `server/lib/aiModelCatalog.ts`
+- Default capabilities per provider type: `server/lib/aiProviderDefaults.ts`
+
+Overlapping model allows are permitted at save time. Selection happens at
+request time. If the algorithm cannot confidently pick one provider, the
+gateway returns `403` with an ambiguous-provider error.
+
+## Selection Pipeline
+
+Every gateway request runs through these steps in order. Each step narrows
+the candidate set. Later steps only run when more than one provider remains.
+
+```
+1. Capability filter
+2. Allow / block lists
+3. Most specific allow pattern
+4. Catalog ownership
+5. Provider class preference
+6. Ambiguous → error
+```
+
+### 1. Capability Filter
+
+The incoming path selects a capability before any provider logic runs.
+
+| Path | Capability |
+|------|------------|
+| `POST /v1/chat/completions` | `openai_chat` |
+| `POST /v1/responses` | `openai_responses` |
+| `POST /v1/messages` | `anthropic_messages` |
+| Gemini / Vertex / Bedrock routes | their respective capability ids |
+
+Only attached providers that advertise that capability stay in the candidate
+set. Default capabilities do not overlap for native OpenAI vs Anthropic:
+
+| Provider type | Default capabilities |
+|---------------|----------------------|
+| `openai` | `openai_chat`, `openai_responses` |
+| `anthropic` | `anthropic_messages` |
+| `openRouter` | `openai_chat` |
+| `vercelAiGateway` | `openai_chat`, `openai_responses` |
+| `microsoftFoundry` | `openai_chat`, `openai_responses`, `anthropic_messages` |
+| `custom` | whatever was configured |
+
+### 2. Allow / Block Lists
+
+For each remaining provider, the gateway resolves the effective allow and
+block patterns:
+
+- **`inherit`**: use the provider's own model lists
+- **`select`**: use the resource-selected subset of those lists
+
+A candidate is kept only if `isAllowedByLists(requestedModel, allows, blocks)`
+passes:
+
+1. At least one allow pattern must match
+2. No block pattern may match
+
+Patterns support `*` and `?` globs (`gpt-*`, `claude-3-5-sonnet-?`).
+
+### 3. Most Specific Allow Pattern
+
+Among providers that allow the model, keep those whose matching allow
+pattern is most specific:
+
+1. Exact keys beat patterns
+2. Fewer wildcard characters win
+3. Longer literal length wins
+
+Example: `gpt-4o` beats `gpt-*` beats `*`.
+
+### 4. Catalog Ownership
+
+When specificity is tied (common with multiple `*` allows), score each
+provider against the known model catalog:
+
+| Score | Meaning |
+|------:|---------|
+| 2 | Typed provider whose catalog contains the model (`openai` → openai catalog, `anthropic` → anthropic, etc.) |
+| 1 | Aggregator or custom (`openRouter`, `vercelAiGateway`, `custom`) and the model exists somewhere in the catalog |
+| 0 | No ownership signal (typed catalog miss, or unknown model on aggregator/custom) |
+
+Model id lookup tries the raw id, then a stripped `vendor/model` form
+(e.g. `openai/gpt-4o` → also try `gpt-4o`).
+
+Typed providers map to catalog providers as:
+
+| Provider type | Catalog |
+|---------------|---------|
+| `openai` | `openai` |
+| `anthropic` | `anthropic` |
+| `googleGemini` | `gemini` |
+| `vertexAi` | `vertex` |
+| `bedrock` | `bedrock` |
+| `microsoftFoundry` | `azure` |
+| `openRouter` / `vercelAiGateway` / `custom` | none (aggregator/custom path) |
+
+### 5. Provider Class Preference
+
+If catalog ownership is still tied, prefer:
+
+| Rank | Class |
+|-----:|-------|
+| 2 | Native typed provider (`openai`, `anthropic`, `googleGemini`, …) |
+| 1 | Aggregator (`openRouter`, `vercelAiGateway`) |
+| 0 | `custom` |
+
+### 6. Ambiguous Error
+
+If more than one distinct provider remains after all steps, the gateway
+rejects the request:
+
+```
+Model "<id>" is ambiguous across multiple AI providers on this resource
+```
+
+Typical remaining ties: two OpenAI-type providers both with `*`, or two
+customs advertising the same capability for an unknown model.
+
+## Examples
+
+Assume each provider below is attached and enabled on the same inference
+resource.
+
+### Example A: OpenAI + Anthropic, Both `*`
+
+| Provider | Allow | Capabilities |
+|----------|-------|--------------|
+| OpenAI | `*` | `openai_chat`, `openai_responses` |
+| Anthropic | `*` | `anthropic_messages` |
+
+**Request:** `POST /v1/chat/completions` with `model: "gpt-4o"`
+
+1. Capability → only OpenAI remains
+2. Allow → OpenAI matches `*`
+3. Result → **OpenAI**
+
+Anthropic never reaches pattern or catalog scoring. Capability alone decides.
+
+**Request:** `POST /v1/messages` with `model: "claude-3-5-sonnet-latest"`
+
+1. Capability → only Anthropic remains
+2. Result → **Anthropic**
+
+### Example B: OpenAI + OpenRouter, Both `*`
+
+| Provider | Allow | Capabilities |
+|----------|-------|--------------|
+| OpenAI | `*` | `openai_chat`, … |
+| OpenRouter | `*` | `openai_chat` |
+
+**Request:** `POST /v1/chat/completions` with `model: "gpt-4o"`
+
+1. Capability → both remain (`openai_chat`)
+2. Allow → both match `*`
+3. Specificity → tie (`*` vs `*`)
+4. Catalog → OpenAI scores `2` (owns `gpt-4o`); OpenRouter scores `1`
+5. Result → **OpenAI**
+
+### Example C: OpenRouter Only Serving a Claude Model Over OpenAI Chat
+
+| Provider | Allow | Capabilities |
+|----------|-------|--------------|
+| OpenRouter | `*` | `openai_chat` |
+
+**Request:** `POST /v1/chat/completions` with `model: "anthropic/claude-3.5-sonnet"`
+
+1. Capability → OpenRouter remains
+2. Only one candidate → **OpenRouter**
+
+No tie-breaking needed.
+
+### Example D: OpenAI (`gpt-*`) + OpenRouter (`*`)
+
+| Provider | Allow |
+|----------|-------|
+| OpenAI | `gpt-*` |
+| OpenRouter | `*` |
+
+**Request:** `model: "gpt-4o"` on `openai_chat`
+
+1. Capability → both
+2. Allow → both match
+3. Specificity → OpenAI's `gpt-*` beats OpenRouter's `*`
+4. Result → **OpenAI**
+
+Catalog scoring is not needed because specificity already unique'd the set.
+
+### Example E: OpenAI + Anthropic With Overlapping Custom Capabilities
+
+Someone grants Anthropic `openai_chat` as well (non-default).
+
+| Provider | Allow | Capabilities |
+|----------|-------|--------------|
+| OpenAI | `*` | `openai_chat`, … |
+| Anthropic | `*` | `anthropic_messages`, `openai_chat` |
+
+**Request:** `POST /v1/chat/completions` with `model: "gpt-4o"`
+
+1. Capability → both remain
+2. Allow → both match `*`
+3. Specificity → tie
+4. Catalog → OpenAI `2`, Anthropic `0` (`gpt-4o` is not in the anthropic catalog)
+5. Result → **OpenAI**
+
+### Example F: Two Aggregators, Known Model
+
+| Provider | Allow |
+|----------|-------|
+| OpenRouter | `*` |
+| Vercel AI Gateway | `*` |
+
+**Request:** `model: "gpt-4o"` on `openai_chat`
+
+1. Capability → both
+2. Allow / specificity → tie
+3. Catalog → both score `1` (known model, no typed owner in the set)
+4. Class → both aggregators (rank `1`) → still tied
+5. Result → **ambiguous error**
+
+Attach a native OpenAI provider (or narrow one aggregator's allow list) to
+make this determinable.
+
+### Example G: Two OpenAI Providers, Both `*`
+
+| Provider | Type | Allow |
+|----------|------|-------|
+| OpenAI Prod | `openai` | `*` |
+| OpenAI Staging | `openai` | `*` |
+
+**Request:** `model: "gpt-4o"`
+
+1–5 all leave both candidates (same capability, same specificity, same
+catalog ownership, same class).
+
+Result → **ambiguous error**
+
+Disambiguate with different allow patterns, disable one attachment, or
+split across resources.
+
+### Example H: Unknown Model Across Native + Aggregator
+
+| Provider | Allow |
+|----------|-------|
+| OpenAI | `*` |
+| OpenRouter | `*` |
+
+**Request:** `model: "my-fine-tune-v3"` (not in catalog)
+
+1. Capability → both
+2. Allow / specificity → tie
+3. Catalog → both score `0` (typed miss + unknown aggregator model)
+4. Class → OpenAI (`2`) beats OpenRouter (`1`)
+5. Result → **OpenAI**
+
+## Practical Guidance
+
+- Native OpenAI + Anthropic with `*` is safe. Different default APIs never
+  collide.
+- OpenAI + OpenRouter with `*` is usually fine for catalog-known OpenAI
+  models. Native wins.
+- Prefer specific allow patterns (`gpt-4o`, `gpt-*`) when two providers share
+  a capability.
+- Two providers of the same type both using `*` will stay ambiguous. Narrow
+  at least one allow list.
+- Custom providers only win ties when no stronger native/aggregator signal
+  remains.
+
+## Related Behavior
+
+- **Saving providers on a resource does not reject overlapping allows.**
+  Collisions are resolved (or rejected) per request.
+- Budgets, auth, and upstream URL / target routing run after a single
+  provider has been selected.

docs/crud-endpoints.md

@@ -0,0 +1,347 @@
+# How to build a CRUD endpoint in this repo
+
+Reference for adding a new CRUD entity to the server. Based on two real
+examples already in the codebase — read them side by side with this doc:
+
+- **Public / open-source (Community Edition) pattern**: `server/routers/aiProvider/`
+- **Enterprise-only pattern**: `server/private/routers/alertRule/`
+
+The two are structurally identical. The only difference is *where the files
+live* and *which router they get wired into*.
+
+## 1. Decide: public or private?
+
+- `server/routers/<entity>/` — ships in the open-source Community Edition.
+  Anyone running Pangolin gets this.
+- `server/private/routers/<entity>/` — Enterprise/SaaS only. Gated behind
+  `verifyValidLicense` (and often `verifyValidSubscription(tierMatrix.x)`).
+  Every file here starts with the Fossorial Commercial License header block
+  (copy it verbatim from an existing private file).
+
+Everything below applies to both — swap `@server/...` for `#private/...`
+import paths and add license headers when building the private version.
+
+## 2. Directory layout
+
+One folder per entity, one file per operation, a barrel `index.ts`:
+
+```
+server/routers/<entity>/
+  index.ts          # export * from each operation file + ./types
+  types.ts          # response payload types + row->public mapper
+  validation.ts      # zod schemas/refinements shared by create + update (optional)
+  create<Entity>.ts
+  list<Entities>.ts
+  get<Entity>.ts
+  update<Entity>.ts
+  delete<Entity>.ts
+```
+
+`index.ts` is a flat barrel:
+
+```ts
+export * from "./createAiProvider";
+export * from "./listAiProviders";
+export * from "./getAiProvider";
+export * from "./updateAiProvider";
+export * from "./deleteAiProvider";
+export * from "./types";
+```
+
+## 3. Anatomy of a single handler
+
+Every handler file (`create<Entity>.ts`, etc.) follows the same shape:
+
+```ts
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { <table>, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetXResponse } from "@server/routers/<entity>/types";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()          // or entityId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({ /* ... */ }); // create/update only
+
+registry.registerPath({
+    method: "get", // put | post | delete
+    path: "/org/{orgId}/x",
+    description: "...",
+    tags: [OpenAPITags.<Entity>],
+    request: { params: paramsSchema, /* body: {...} for write ops, query: for list */ },
+    responses: { 200: { description: "Successful response" } }
+});
+
+export async function getX(req: Request, res: Response, next: NextFunction): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, fromError(parsedParams.error).toString()));
+        }
+        // parse body too, if present, same pattern
+
+        // ...business logic against db...
+
+        if (!row) {
+            return next(createHttpError(HttpCode.NOT_FOUND, `X with ID ${id} not found`));
+        }
+
+        return response<GetXResponse>(res, {
+            data: { /* ... */ },
+            success: true,
+            error: false,
+            message: "X retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred"));
+    }
+}
+```
+
+Rules to keep consistent with the rest of the codebase:
+
+- `z.strictObject` for params/body — rejects unknown keys.
+- Params parsed first, then body; each on its own `safeParse` + early
+  `next(createHttpError(...))` — never throw raw errors.
+- Every handler registers itself with the OpenAPI `registry` even if nobody
+  reads the spec directly — it's how `/api/v1/docs` stays accurate.
+- Catch-all `try/catch` at the bottom: `logger.error(error)` +
+  generic `500` message. Never leak internal error details to the client.
+- Use `response<T>(res, { data, success, error, message, status })` from
+  `@server/lib/response` for every response, success or otherwise (errors go
+  through `next(createHttpError(...))` instead, not through `response`).
+- If the route already ran an access-control middleware that fetched the row
+  (see §5), reuse it instead of re-querying:
+  `req.aiProvider && req.aiProvider.providerId === providerId ? [req.aiProvider] : await db.select()...`
+
+### List handler specifics
+
+Pagination is a fixed shape (`page`, `pageSize`, optional `query` for
+search). See `listAiProviders.ts`:
+
+```ts
+const listSchema = z.object({
+    pageSize: z.coerce.number<string>().int().positive().optional().catch(20).default(20),
+    page: z.coerce.number<string>().int().min(0).optional().catch(1).default(1),
+    query: z.string().optional()
+});
+```
+
+Run the count query and the page query in `Promise.all`, and return
+`PaginatedResponse<{ items: T[] }>` (`@server/types/Pagination`) with
+`{ total, pageSize, page }`.
+
+### types.ts specifics
+
+- Define one response type per operation: `List<Entities>Response`,
+  `Get<Entity>Response`, `CreateOrEdit<Entity>Response` (create and update
+  commonly share a response shape).
+- If the raw DB row needs to be shaped for clients (decrypting secrets,
+  parsing a serialized column, hiding a column), put a `toPublic<Entity>()`
+  mapper here — see `toPublicAiProvider` for the pattern of stripping
+  `apiKey`/serialized columns and re-adding decrypted/parsed versions.
+
+### validation.ts specifics
+
+Only needed when create and update share non-trivial zod pieces (enums,
+`superRefine` cross-field rules). Export the raw schemas (`z.enum([...])`)
+and refinement functions, and import them into both `createX.ts` and
+`updateX.ts` — see `aiProvider/validation.ts`'s
+`refineProviderUpstreamFields`.
+
+## 4. Wire up an access-control middleware (for id-scoped routes)
+
+For routes scoped to a single row (`/x/:xId`, as opposed to
+`/org/:orgId/x` create/list), add a `verify<Entity>Access` middleware in
+`server/middlewares/` (or `server/private/middlewares/` for enterprise-only
+entities) and export it from that directory's `index.ts`.
+
+Pattern (`verifyAiProviderAccess.ts`):
+
+1. Read the id param, `Number.parseInt`/validate it.
+2. Load the row by id.
+3. `404` if it doesn't exist.
+4. Resolve the row's `orgId`, then check/attach `req.userOrg` (query
+   `userOrgs` if not already on the request), `403` if the user isn't in
+   that org.
+5. Run `checkOrgAccessPolicy` if `req.orgPolicyAllowed` hasn't been resolved
+   yet.
+6. Set `req.userOrgId`, `req.userOrgRoleIds`, and stash the row on the
+   request (e.g. `req.aiProvider = provider`) so downstream handlers and
+   `verifyUserHasAction` don't have to refetch it.
+
+Org-scoped create/list routes (`/org/:orgId/x`) don't need a bespoke
+middleware — they use the existing generic `verifyOrgAccess` from
+`@server/middlewares`.
+
+## 5. Register an action + permission check
+
+Add one `ActionsEnum` entry per operation in `server/auth/actions.ts`,
+grouped near the entity's other actions, named `create<Entity>`,
+`get<Entity>`, `update<Entity>`, `delete<Entity>`, `list<Entities>`:
+
+```ts
+createAiProvider = "createAiProvider",
+deleteAiProvider = "deleteAiProvider",
+getAiProvider = "getAiProvider",
+listAiProviders = "listAiProviders",
+updateAiProvider = "updateAiProvider",
+```
+
+Every route uses `verifyUserHasAction(ActionsEnum.x)` to check the caller's
+role/permissions for that action, and mutating routes (create/update/delete)
+follow it with `logActionAudit(ActionsEnum.x)` to record the action in the
+audit log.
+
+## 6. Register the routes
+
+There are four router files; which one(s) you touch depends on public vs.
+private and user-facing vs. service-to-service:
+
+| File | Purpose |
+|---|---|
+| `server/routers/external.ts` | Public, user-facing API. Exports `authenticated`, `unauthenticated`, `authRouter` Express routers. |
+| `server/routers/internal.ts` | Public, internal service-to-service API (gerbil, badger, traefik-config) — no user auth, exports `internalRouter`. |
+| `server/private/routers/external.ts` | Enterprise-only, user-facing. Imports `authenticated`/`unauthenticated`/`authRouter` **from the public `external.ts`** and re-exports them, then adds more routes on top. |
+| `server/private/routers/internal.ts` | Enterprise-only, service-to-service. Same re-export trick with `internalRouter`. |
+
+Private router files always start:
+
+```ts
+import {
+    unauthenticated as ua,
+    authenticated as a,
+    authRouter as aa
+} from "@server/routers/external";
+
+export const authenticated = a;
+export const unauthenticated = ua;
+export const authRouter = aa;
+```
+
+...and then call `authenticated.get/put/post/delete(...)` to bolt on
+additional, enterprise-only routes on the *same* router instances the public
+build uses. This is why the private build has strictly more routes than the
+public build, not a divergent copy.
+
+### Route registration order (mutating vs read)
+
+Standard middleware chain per verb, using `alertRule`'s registrations as the
+template:
+
+```ts
+// Create — org-scoped, no row exists yet
+authenticated.put(
+    "/org/:orgId/x",
+    verifyValidLicense,              // private/enterprise routes only
+    verifyOrgAccess,
+    verifyLimits,                    // if the entity counts against a plan limit
+    verifyUserHasAction(ActionsEnum.createX),
+    logActionAudit(ActionsEnum.createX),
+    x.createX
+);
+
+// Update — row-scoped
+authenticated.post(
+    "/org/:orgId/x/:xId",            // or "/x/:xId" if id is globally unique
+    verifyValidLicense,
+    verifyOrgAccess,                 // or verifyXAccess if globally-keyed
+    verifyUserHasAction(ActionsEnum.updateX),
+    logActionAudit(ActionsEnum.updateX),
+    x.updateX
+);
+
+// Delete — row-scoped
+authenticated.delete(
+    "/org/:orgId/x/:xId",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.deleteX),
+    logActionAudit(ActionsEnum.deleteX),
+    x.deleteX
+);
+
+// List — org-scoped, read-only, no audit log
+authenticated.get(
+    "/org/:orgId/xs",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.listXs),
+    x.listXs
+);
+
+// Get one — row-scoped, read-only, no audit log
+authenticated.get(
+    "/org/:orgId/x/:xId",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.getX),
+    x.getX
+);
+```
+
+Notes:
+
+- HTTP verbs: `PUT` = create, `POST` = update, `GET` = read, `DELETE` =
+  delete. This repo does not use `PATCH` for entity updates (site
+  provisioning keys are the one exception, using `PATCH`).
+- `verifyValidLicense` is only needed on private/enterprise routes; public
+  OSS routes skip it.
+- Use `verifyValidSubscription(tierMatrix.someFeature)` right after
+  `verifyValidLicense` when a feature is gated to specific SaaS tiers (see
+  `tierMatrix` usages in `server/private/routers/external.ts`).
+- `verifyLimits` goes on create routes for entities that count against a
+  plan/seat limit.
+- For entities keyed by a globally-unique id (not nested under `/org/:orgId`),
+  use the dedicated `verify<Entity>Access` middleware from §4 instead of
+  `verifyOrgAccess` on the row-scoped routes (see how `/ai-provider/:providerId`
+  uses `verifyAiProviderAccess`, while `/org/:orgId/ai-provider` create/list
+  use plain `verifyOrgAccess`).
+- Read-only routes (`get`, `list`) skip `logActionAudit` — only mutations are
+  audited.
+- `internal*.ts` routes are for trusted internal callers (gerbil/badger
+  sidecars) and generally skip user-facing auth entirely, using
+  `verifySessionUserMiddleware` / `verifyUserFromResourceSessionMiddleware`
+  instead of `verifyOrgAccess`/`verifyUserHasAction`. CRUD entities almost
+  never need internal router entries — only add one if a sidecar process
+  needs direct access to the resource.
+
+## 7. The `#dynamic` alias (advanced — most CRUD work can ignore this)
+
+Some middleware (e.g. `logActionAudit`) needs a real implementation in the
+enterprise/SaaS build but a no-op stub in the open-source build, while
+being imported by identical code in `server/routers/external.ts` in both
+builds. That's done via the `#dynamic/*` import alias, which
+`tsconfig.oss.json` points at `./server/*` and `tsconfig.enterprise.json` /
+`tsconfig.saas.json` point at `./server/private/*`. You only need this
+pattern if you're adding a genuinely dual-implementation hook; a normal
+private-only CRUD entity (like `alertRule`) never touches `#dynamic` — it
+just lives entirely under `server/private/` and is imported with `#private/*`
+directly from `server/private/routers/external.ts`.
+
+## 8. Checklist for a new entity
+
+1. Add the DB table to `server/db/pg/schema/schema.ts` (and sqlite schema if
+   applicable).
+2. Add `ActionsEnum` entries in `server/auth/actions.ts`.
+3. Create `server/routers/<entity>/` (or `server/private/routers/<entity>/`):
+   `types.ts`, optional `validation.ts`, one file per operation, `index.ts`
+   barrel.
+4. If routes are row-scoped by a global id, add
+   `verify<Entity>Access.ts` to `server/middlewares/` or
+   `server/private/middlewares/`, and export it from that directory's
+   `index.ts`.
+5. Wire routes into `external.ts` (public or private) following the verb/
+   middleware table in §6. Add to `internal.ts` only if a sidecar needs
+   direct access.
+6. Add license header block to every new file if it's under `server/private/`.

drizzle.sqlite.config.ts

@@ -1,4 +1,4 @@
-import { APP_PATH } from "@server/lib/consts";
+import { APP_PATH } from "./server/lib/consts";
 import { defineConfig } from "drizzle-kit";
 import path from "path";
 

install/config/config.yml

@@ -22,7 +22,8 @@ server:
         methods: ["GET", "POST", "PUT", "DELETE", "PATCH"]
         allowed_headers: ["X-CSRF-Token", "Content-Type"]
         credentials: false
-    {{if .EnableGeoblocking}}maxmind_db_path: "./config/GeoLite2-Country.mmdb"{{end}}
+    {{if .EnableMaxMind}}maxmind_db_path: "./config/GeoLite2-Country.mmdb"{{end}}
+    {{if .EnableMaxMind}}maxmind_asn_path: "./config/GeoLite2-ASN.mmdb"{{end}}
 {{if .EnableEmail}}
 email:
     smtp_host: "{{.EmailSMTPHost}}"
@@ -36,3 +37,6 @@ flags:
     disable_signup_without_invite: true
     disable_user_create_org: false
     allow_raw_resources: true
+
+{{if .IsPostgreSQL}}postgres:
+    connection_string: postgresql://pangolin:{{.IsPostgreSQLPass}}@postgres:5432/pangolin{{end}}

install/config/docker-compose.yml

@@ -1,15 +1,23 @@
 name: pangolin
 services:
   pangolin:
-    image: docker.io/fosrl/pangolin:{{if .IsEnterprise}}ee-{{end}}{{.PangolinVersion}}
+    image: docker.io/fosrl/pangolin:{{if .IsEnterprise}}ee-{{end}}{{if .IsPostgreSQL}}postgresql-{{end}}{{.PangolinVersion}}
     container_name: pangolin
     restart: unless-stopped
     deploy:
       resources:
         limits:
-          memory: 1g
+          memory: 2g
         reservations:
-          memory: 256m
+          memory: 512m
+    {{if or .IsPostgreSQL .IsRedis}}depends_on:
+      {{if .IsPostgreSQL}}postgres:
+          condition: service_healthy{{end}}
+      {{if .IsRedis}}redis:
+          condition: service_healthy{{end}}
+    networks:
+      - default
+      - backend{{end}}
     volumes:
       - ./config:/app/config
     healthcheck:
@@ -17,8 +25,8 @@ services:
       interval: "10s"
       timeout: "10s"
       retries: 15
-{{if .InstallGerbil}}
-  gerbil:
+
+  {{if .InstallGerbil}}gerbil:
     image: docker.io/fosrl/gerbil:{{.GerbilVersion}}
     container_name: gerbil
     restart: unless-stopped
@@ -39,17 +47,16 @@ services:
       - 21820:21820/udp
       - 443:443
       - 443:443/udp # For http3 QUIC if desired
-      - 80:80
-{{end}}
+      - 80:80{{end}}
+
   traefik:
-    image: docker.io/traefik:v3.6
+    image: docker.io/traefik:v3.7
     container_name: traefik
     restart: unless-stopped
-{{if .InstallGerbil}}    network_mode: service:gerbil # Ports appear on the gerbil service{{end}}{{if not .InstallGerbil}}
+    {{if .InstallGerbil}}network_mode: service:gerbil # Ports appear on the gerbil service{{end}}{{if not .InstallGerbil}}
     ports:
       - 443:443
-      - 80:80
-{{end}}
+      - 80:80{{end}}
     depends_on:
       pangolin:
         condition: service_healthy
@@ -60,8 +67,50 @@ services:
       - ./config/letsencrypt:/letsencrypt # Volume to store the Let's Encrypt certificates
       - ./config/traefik/logs:/var/log/traefik # Volume to store Traefik logs
 
+  {{if .IsPostgreSQL}}postgres:
+    image: postgres:18
+    container_name: postgres
+    restart: unless-stopped
+    environment:
+      POSTGRES_USER: pangolin
+      POSTGRES_PASSWORD: {{.IsPostgreSQLPass}}
+      POSTGRES_DB: pangolin
+    volumes:
+      - ./postgres18:/var/lib/postgresql
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U pangolin"]
+      interval: 10s
+      timeout: 5s
+      retries: 5
+    networks:
+      - backend{{end}}
+
+  {{if .IsRedis}}redis:
+    image: redis:8-trixie
+    container_name: redis
+    restart: unless-stopped
+    command: >
+      redis-server
+      --save 3600 1000
+      --appendonly yes
+      --requirepass {{.IsRedisPass}}
+    volumes:
+      - ./redis8:/data
+    healthcheck:
+      test: ["CMD", "redis-cli", "-a", "{{.IsRedisPass}}", "ping"]
+      interval: 10s
+      timeout: 3s
+      retries: 3
+      start_period: 10s
+    networks:
+      - backend{{end}}
+
 networks:
   default:
     driver: bridge
-    name: pangolin
+    name: pangolin_frontend
 {{if .EnableIPv6}}    enable_ipv6: true{{end}}
+{{if or .IsPostgreSQL .IsRedis}}  backend:
+    driver: bridge
+    name: pangolin_backend
+    internal: true{{end}}

install/config/privateConfig.yml

@@ -0,0 +1,4 @@
+{{if .IsRedis}}redis:
+  host: "redis"
+  port: 6379
+  password: "{{.IsRedisPass}}"{{end}}

install/go.mod

@@ -5,7 +5,7 @@ go 1.25.0
 require (
 	github.com/charmbracelet/huh v1.0.0
 	github.com/charmbracelet/lipgloss v1.1.0
-	golang.org/x/term v0.42.0
+	golang.org/x/term v0.45.0
 	gopkg.in/yaml.v3 v3.0.1
 )
 
@@ -33,6 +33,6 @@ require (
 	github.com/rivo/uniseg v0.4.7 // indirect
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	golang.org/x/sync v0.15.0 // indirect
-	golang.org/x/sys v0.43.0 // indirect
+	golang.org/x/sys v0.47.0 // indirect
 	golang.org/x/text v0.23.0 // indirect
 )

install/go.sum

@@ -69,10 +69,10 @@ golang.org/x/sync v0.15.0 h1:KWH3jNZsfyT6xfAfKiz6MRNmd46ByHDYaZ7KSkCtdW8=
 golang.org/x/sync v0.15.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA=
 golang.org/x/sys v0.0.0-20210809222454-d867a43fc93e/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.43.0 h1:Rlag2XtaFTxp19wS8MXlJwTvoh8ArU6ezoyFsMyCTNI=
-golang.org/x/sys v0.43.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
-golang.org/x/term v0.42.0 h1:UiKe+zDFmJobeJ5ggPwOshJIVt6/Ft0rcfrXZDLWAWY=
-golang.org/x/term v0.42.0/go.mod h1:Dq/D+snpsbazcBG5+F9Q1n2rXV8Ma+71xEjTRufARgY=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/term v0.45.0 h1:NwWyBmoJCbfTHpxrWoZ9C6/VxOf7ic219I8xZZFdrf0=
+golang.org/x/term v0.45.0/go.mod h1:9aqxs0blBcrm/n0L9QW0aRVD+ktan8ssZromtqJC43w=
 golang.org/x/text v0.23.0 h1:D71I7dUrlY+VX0gQShAThNGHFxZ13dGLBHQLVl1mJlY=
 golang.org/x/text v0.23.0/go.mod h1:/BLNzu4aZCJ1+kcD0DNRotWKage4q2rGVAg4o22unh4=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=

install/main.go

@@ -4,6 +4,7 @@ import (
 	"crypto/rand"
 	"embed"
 	"encoding/base64"
+	"flag"
 	"fmt"
 	"io"
 	"io/fs"
@@ -53,9 +54,13 @@ type Config struct {
 	InstallGerbil             bool
 	TraefikBouncerKey         string
 	DoCrowdsecInstall         bool
-	EnableGeoblocking         bool
+	EnableMaxMind             bool
 	Secret                    string
 	IsEnterprise              bool
+    IsPostgreSQL              bool
+	IsPostgreSQLPass          string
+    IsRedis                   bool
+	IsRedisPass               string
 }
 
 type SupportedContainer string
@@ -66,8 +71,14 @@ const (
 	Undefined SupportedContainer = "undefined"
 )
 
+var redisFlag *bool
+
 func main() {
 
+	crowdsecFlag := flag.Bool("crowdsec", false, "Enable the CrowdSec installation prompt")
+	redisFlag = flag.Bool("redis", false, "Install Redis as caching solution. Required for HA. Not required for the Enterprise version.")
+	flag.Parse()
+
 	// print a banner about prerequisites - opening port 80, 443, 51820, and 21820 on the VPS and firewall and pointing your domain to the VPS IP with a records. Docs are at http://localhost:3000/Getting%20Started/dns-networking
 
 	fmt.Println("Welcome to the Pangolin installer!")
@@ -119,11 +130,11 @@ func main() {
 
 		fmt.Println("\nConfiguration files created successfully!")
 
-		// Download MaxMind database if requested
-		if config.EnableGeoblocking {
-			fmt.Println("\n=== Downloading MaxMind Database ===")
+		// Download MaxMind Country / ASN database if requested
+		if config.EnableMaxMind {
+			fmt.Println("\n=== Downloading MaxMind Country and ASN Databases ===")
 			if err := downloadMaxMindDatabase(); err != nil {
-				fmt.Printf("Error downloading MaxMind database: %v\n", err)
+				fmt.Printf("Error downloading MaxMind databases: %v\n", err)
 				fmt.Println("You can download it manually later if needed.")
 			}
 		}
@@ -184,15 +195,15 @@ func main() {
 		fmt.Println("\n=== MaxMind Database Update ===")
 		if _, err := os.Stat("config/GeoLite2-Country.mmdb"); err == nil {
 			fmt.Println("MaxMind GeoLite2 Country database found.")
-			if readBool("Would you like to update the MaxMind database to the latest version?", false) {
+			if readBool("Would you like to update the MaxMind databases (Country and ASN) to the latest version?", false) {
 				if err := downloadMaxMindDatabase(); err != nil {
 					fmt.Printf("Error updating MaxMind database: %v\n", err)
 					fmt.Println("You can try updating it manually later if needed.")
 				}
 			}
 		} else {
-			fmt.Println("MaxMind GeoLite2 Country database not found.")
-			if readBool("Would you like to download the MaxMind GeoLite2 database for geoblocking functionality?", false) {
+			fmt.Println("MaxMind GeoLite2 Country and ASN databases not found.")
+			if readBool("Would you like to download the MaxMind GeoLite2 databases for blocking functionality?", false) {
 				if err := downloadMaxMindDatabase(); err != nil {
 					fmt.Printf("Error downloading MaxMind database: %v\n", err)
 					fmt.Println("You can try downloading it manually later if needed.")
@@ -200,13 +211,15 @@ func main() {
 				// Now you need to update your config file accordingly to enable geoblocking
 				fmt.Print("Please remember to update your config/config.yml file to enable geoblocking! \n\n")
 				// add   maxmind_db_path: "./config/GeoLite2-Country.mmdb" under server
-				fmt.Println("Add the following line under the 'server' section:")
+				// add   maxmind_asn_path: "./config/GeoLite2-ASN.mmdb" under server
+				fmt.Println("Add the following lines under the 'server' section:")
 				fmt.Println("  maxmind_db_path: \"./config/GeoLite2-Country.mmdb\"")
+				fmt.Println("  maxmind_asn_path: \"./config/GeoLite2-ASN.mmdb\"")
 			}
 		}
 	}
 
-	if !checkIsCrowdsecInstalledInCompose() {
+	if *crowdsecFlag && !checkIsCrowdsecInstalledInCompose() {
 		fmt.Println("\n=== CrowdSec Install ===")
 		// check if crowdsec is installed
 		if readBool("Would you like to install CrowdSec?", false) {
@@ -480,6 +493,17 @@ func collectUserInput() Config {
 	fmt.Println("\n=== Basic Configuration ===")
 
 	config.IsEnterprise = readBoolNoDefault("Do you want to install the Enterprise version of Pangolin? The EE is free for personal use or for businesses making less than 100k USD annually.")
+    if config.IsEnterprise {
+        if *redisFlag {
+			config.IsRedis = true
+            config.IsRedisPass = readPassword("Enter a unique password for the Redis service.")
+        }
+    }
+
+    config.IsPostgreSQL = readBool("Do you want to use PostgreSQL (not recommended for most users)?", false)
+	if config.IsPostgreSQL {
+		config.IsPostgreSQLPass = readPassword("Enter a unique password for the PostgreSQL pangolin user.")
+	}
 
 	config.BaseDomain = readString("Enter your base domain (no subdomain e.g. example.com)", "")
 
@@ -523,7 +547,7 @@ func collectUserInput() Config {
 	fmt.Println("\n=== Advanced Configuration ===")
 
 	config.EnableIPv6 = readBool("Is your server IPv6 capable?", true)
-	config.EnableGeoblocking = readBool("Do you want to download the MaxMind GeoLite2 database for geoblocking functionality?", true)
+	config.EnableMaxMind = readBool("Do you want to download the MaxMind GeoLite2 Country and ASN databases for blocking functionality?", true)
 
 	if config.DashboardDomain == "" {
 		fmt.Println("Error: Dashboard Domain name is required")
@@ -776,29 +800,42 @@ func checkPortsAvailable(port int) error {
 }
 
 func downloadMaxMindDatabase() error {
-	fmt.Println("Downloading MaxMind GeoLite2 Country database...")
+	fmt.Println("Downloading MaxMind GeoLite2 Country and ASN databases...")
 
-	// Download the GeoLite2 Country database
+	// Download the GeoLite2 Country databases
 	if err := run("curl", "-L", "-o", "GeoLite2-Country.tar.gz",
 		"https://github.com/GitSquared/node-geolite2-redist/raw/refs/heads/master/redist/GeoLite2-Country.tar.gz"); err != nil {
-		return fmt.Errorf("failed to download GeoLite2 database: %v", err)
+		return fmt.Errorf("failed to download GeoLite2 Country database: %v", err)
+	}
+	if err := run("curl", "-L", "-o", "GeoLite2-ASN.tar.gz",
+		"https://github.com/GitSquared/node-geolite2-redist/raw/refs/heads/master/redist/GeoLite2-ASN.tar.gz"); err != nil {
+		return fmt.Errorf("failed to download GeoLite2 ASN database: %v", err)
 	}
 
-	// Extract the database
+	// Extract the Country database
 	if err := run("tar", "-xzf", "GeoLite2-Country.tar.gz"); err != nil {
-		return fmt.Errorf("failed to extract GeoLite2 database: %v", err)
+		return fmt.Errorf("failed to extract GeoLite2 Country database: %v", err)
+	}
+	if err := run("tar", "-xzf", "GeoLite2-ASN.tar.gz"); err != nil {
+		return fmt.Errorf("failed to extract GeoLite2 ASN database: %v", err)
 	}
 
 	// Find the .mmdb file and move it to the config directory
 	if err := run("bash", "-c", "mv GeoLite2-Country_*/GeoLite2-Country.mmdb config/"); err != nil {
-		return fmt.Errorf("failed to move GeoLite2 database to config directory: %v", err)
+		return fmt.Errorf("failed to move GeoLite2 Country database to config directory: %v", err)
+	}
+	if err := run("bash", "-c", "mv GeoLite2-ASN_*/GeoLite2-ASN.mmdb config/"); err != nil {
+		return fmt.Errorf("failed to move GeoLite2 ASN database to config directory: %v", err)
 	}
 
 	// Clean up the downloaded files
-	if err := run("rm", "-rf", "GeoLite2-Country.tar.gz", "GeoLite2-Country_*"); err != nil {
-		fmt.Printf("Warning: failed to clean up temporary files: %v\n", err)
+	if err := run("sh", "-c", "rm -rf GeoLite2-Country.tar.gz GeoLite2-Country_*"); err != nil {
+		fmt.Printf("Warning: failed to clean up temporary country files: %v\n", err)
+	}
+	if err := run("sh", "-c", "rm -rf GeoLite2-ASN.tar.gz GeoLite2-ASN_*"); err != nil {
+		fmt.Printf("Warning: failed to clean up temporary ASN files: %v\n", err)
 	}
 
-	fmt.Println("MaxMind GeoLite2 Country database downloaded successfully!")
+	fmt.Println("MaxMind GeoLite2 Country and ASN database downloaded successfully!")
 	return nil
 }

messages/zh-TW.json

@@ -152,8 +152,8 @@
     "shareErrorSelectResource": "請選擇一個資源",
     "proxyResourceTitle": "管理公開資源",
     "proxyResourceDescription": "建立和管理可透過網頁瀏覽器公開存取的資源",
-    "proxyResourcesBannerTitle": "基於網頁的公開存取",
-    "proxyResourcesBannerDescription": "公開資源是任何人都可以透過網頁瀏覽器存取的 HTTPS 或 TCP/UDP 代理。與私有資源不同，它們不需要客戶端軟體，並且可以包含基於身份和情境感知的存取策略。",
+    "publicResourcesBannerTitle": "基於網頁的公開存取",
+    "publicResourcesBannerDescription": "公開資源是任何人都可以透過網頁瀏覽器存取的 HTTPS 或 TCP/UDP 代理。與私有資源不同，它們不需要客戶端軟體，並且可以包含基於身份和情境感知的存取策略。",
     "clientResourceTitle": "管理私有資源",
     "clientResourceDescription": "建立和管理只能透過已連接的客戶端存取的資源",
     "privateResourcesBannerTitle": "零信任私有存取",
@@ -1099,6 +1099,7 @@
     "actionGenerateAccessToken": "生成訪問令牌",
     "actionDeleteAccessToken": "刪除訪問令牌",
     "actionListAccessTokens": "訪問令牌",
+    "actionCreateResourceSessionToken": "建立資源工作階段權杖",
     "actionCreateResourceRule": "創建資源規則",
     "actionDeleteResourceRule": "刪除資源規則",
     "actionListResourceRules": "列出資源規則",

next.config.ts

@@ -1,17 +1,42 @@
 import type { NextConfig } from "next";
 import createNextIntlPlugin from "next-intl/plugin";
+import fs from "fs";
+import path from "path";
 
 const withNextIntl = createNextIntlPlugin();
+// read allowedDevOrigins.json if it exists
+let allowedDevOrigins: string[] = [];
+const allowedDevOriginsPath = path.join(
+    process.cwd(),
+    "allowedDevOrigins.json"
+);
+if (fs.existsSync(allowedDevOriginsPath)) {
+    try {
+        const data = fs.readFileSync(allowedDevOriginsPath, "utf-8");
+        allowedDevOrigins = JSON.parse(data);
+    } catch {}
+}
 
 const nextConfig: NextConfig = {
     reactStrictMode: false,
-    eslint: {
-        ignoreDuringBuilds: true
-    },
-    experimental: {
-        reactCompiler: true
-    },
-    output: "standalone"
+    reactCompiler: true,
+    transpilePackages: ["@novnc/novnc"],
+    output: "standalone",
+    allowedDevOrigins,
+    async redirects() {
+        return [
+            {
+                source: "/:orgId/settings/resources/proxy/:path*",
+                destination: "/:orgId/settings/resources/public/:path*",
+                permanent: true
+            },
+            {
+                source: "/:orgId/settings/resources/client/:path*",
+                destination: "/:orgId/settings/resources/private/:path*",
+                permanent: true
+            }
+        ];
+    }
 };
 
 export default withNextIntl(nextConfig);

package.json

@@ -32,13 +32,15 @@
         "format": "prettier --write ."
     },
     "dependencies": {
-        "@asteasolutions/zod-to-openapi": "8.4.1",
-        "@aws-sdk/client-s3": "3.1011.0",
-        "@faker-js/faker": "10.3.0",
-        "@headlessui/react": "2.2.9",
-        "@hookform/resolvers": "5.2.2",
+        "@asteasolutions/zod-to-openapi": "8.5.0",
+        "@aws-sdk/client-s3": "3.1056.0",
+        "@devolutions/iron-remote-desktop": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-0.0.0.tgz",
+        "@devolutions/iron-remote-desktop-rdp": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-rdp-0.0.1.tgz",
+        "@headlessui/react": "2.2.10",
+        "@hookform/resolvers": "5.4.0",
         "@monaco-editor/react": "4.7.0",
         "@node-rs/argon2": "2.0.2",
+        "@novnc/novnc": "^1.7.0",
         "@oslojs/crypto": "1.0.1",
         "@oslojs/encoding": "1.1.0",
         "@radix-ui/react-avatar": "1.1.11",
@@ -59,16 +61,20 @@
         "@radix-ui/react-tabs": "1.1.13",
         "@radix-ui/react-toast": "1.2.15",
         "@radix-ui/react-tooltip": "1.2.8",
-        "@react-email/components": "1.0.8",
-        "@react-email/render": "2.0.4",
-        "@react-email/tailwind": "2.0.5",
+        "@react-email/body": "0.3.0",
+        "@react-email/components": "1.0.12",
+        "@react-email/render": "2.0.8",
+        "@react-email/tailwind": "2.0.7",
         "@simplewebauthn/browser": "13.3.0",
-        "@simplewebauthn/server": "13.3.0",
+        "@simplewebauthn/server": "13.3.1",
         "@tailwindcss/forms": "0.5.11",
-        "@tanstack/react-query": "5.90.21",
+        "@tanstack/react-query": "5.100.14",
         "@tanstack/react-table": "8.21.3",
+        "@xterm/addon-fit": "^0.11.0",
+        "@xterm/addon-web-links": "^0.12.0",
+        "@xterm/xterm": "^6.0.0",
         "arctic": "3.7.0",
-        "axios": "1.15.0",
+        "axios": "1.18.0",
         "better-sqlite3": "11.9.1",
         "canvas-confetti": "1.9.4",
         "class-variance-authority": "0.7.1",
@@ -80,77 +86,77 @@
         "d3": "7.9.0",
         "drizzle-orm": "0.45.2",
         "express": "5.2.1",
-        "express-rate-limit": "8.3.0",
+        "express-rate-limit": "8.5.2",
         "glob": "13.0.6",
-        "helmet": "8.1.0",
+        "gpt-tokenizer": "^3.4.0",
+        "helmet": "8.2.0",
         "http-errors": "2.0.1",
         "input-otp": "1.4.2",
-        "ioredis": "5.10.0",
+        "ioredis": "5.11.0",
         "jmespath": "0.16.0",
-        "js-yaml": "4.1.1",
+        "js-yaml": "4.3.0",
         "jsonwebtoken": "9.0.3",
-        "lucide-react": "0.577.0",
-        "maxmind": "5.0.5",
+        "lucide-react": "1.17.0",
+        "maxmind": "5.0.6",
         "moment": "2.30.1",
-        "next": "15.5.15",
-        "next-intl": "4.8.3",
+        "next": "16.2.11",
+        "next-intl": "4.13.0",
         "next-themes": "0.4.6",
         "nextjs-toploader": "3.9.17",
         "node-cache": "5.1.2",
-        "nodemailer": "8.0.5",
+        "nodemailer": "9.0.1",
         "oslo": "1.2.1",
-        "pg": "8.20.0",
-        "posthog-node": "5.28.0",
+        "pg": "8.21.0",
+        "posthog-node": "5.35.6",
         "qrcode.react": "4.2.0",
-        "react": "19.2.4",
+        "react": "19.2.6",
         "react-day-picker": "9.14.0",
-        "react-dom": "19.2.4",
+        "react-dom": "19.2.6",
         "react-easy-sort": "1.8.0",
-        "react-hook-form": "7.71.2",
+        "react-hook-form": "7.76.1",
         "react-icons": "5.6.0",
-        "recharts": "2.15.4",
+        "recharts": "3.8.1",
         "reodotdev": "1.1.0",
-        "resend": "6.9.2",
-        "semver": "7.7.4",
+        "semver": "7.8.1",
         "sshpk": "1.18.0",
-        "stripe": "20.4.1",
+        "stripe": "22.2.0",
         "swagger-ui-express": "5.0.1",
-        "tailwind-merge": "3.5.0",
+        "tailwind-merge": "3.6.0",
         "topojson-client": "3.1.0",
         "tw-animate-css": "1.4.0",
-        "use-debounce": "10.1.0",
-        "uuid": "13.0.0",
+        "use-debounce": "10.1.1",
+        "uuid": "14.0.0",
         "vaul": "1.1.2",
         "visionscarto-world-atlas": "1.0.0",
         "winston": "3.19.0",
         "winston-daily-rotate-file": "5.0.0",
-        "ws": "8.19.0",
-        "yaml": "2.8.3",
+        "ws": "8.21.0",
+        "yaml": "2.9.0",
         "yargs": "18.0.0",
-        "zod": "4.3.6",
+        "zod": "4.4.3",
         "zod-validation-error": "5.0.0"
     },
     "devDependencies": {
-        "@dotenvx/dotenvx": "1.54.1",
+        "@dotenvx/dotenvx": "1.69.1",
         "@esbuild-plugins/tsconfig-paths": "0.1.2",
-        "@react-email/preview-server": "5.2.10",
-        "@tailwindcss/postcss": "4.2.2",
-        "@tanstack/react-query-devtools": "5.91.3",
+        "@react-email/ui": "^6.5.0",
+        "@tailwindcss/postcss": "4.3.0",
+        "@tanstack/react-query-devtools": "5.100.14",
         "@types/better-sqlite3": "7.6.13",
         "@types/cookie-parser": "1.4.10",
         "@types/cors": "2.8.19",
         "@types/crypto-js": "4.2.2",
         "@types/d3": "7.4.3",
         "@types/express": "5.0.6",
-        "@types/express-session": "1.18.2",
+        "@types/express-session": "1.19.0",
         "@types/jmespath": "0.15.2",
         "@types/js-yaml": "4.0.9",
         "@types/jsonwebtoken": "9.0.10",
-        "@types/node": "25.3.5",
-        "@types/nodemailer": "7.0.11",
+        "@types/node": "25.9.1",
+        "@types/nodemailer": "8.0.0",
         "@types/nprogress": "0.2.3",
-        "@types/pg": "8.18.0",
-        "@types/react": "19.2.14",
+        "@types/pg": "8.20.0",
+        "@types/react": "19.2.15",
         "@types/react-dom": "19.2.3",
         "@types/semver": "7.7.1",
         "@types/sshpk": "1.17.4",
@@ -160,21 +166,22 @@
         "@types/yargs": "17.0.35",
         "babel-plugin-react-compiler": "1.0.0",
         "drizzle-kit": "0.31.10",
-        "esbuild": "0.27.4",
-        "esbuild-node-externals": "1.20.1",
-        "eslint": "10.0.3",
-        "eslint-config-next": "16.1.7",
-        "postcss": "8.5.8",
-        "prettier": "3.8.1",
-        "react-email": "5.2.10",
-        "tailwindcss": "4.2.2",
-        "tsc-alias": "1.8.16",
-        "tsx": "4.21.0",
-        "typescript": "5.9.3",
-        "typescript-eslint": "8.56.1"
+        "esbuild": "0.28.1",
+        "esbuild-node-externals": "1.22.0",
+        "eslint": "10.4.0",
+        "eslint-config-next": "16.2.6",
+        "postcss": "8.5.15",
+        "prettier": "3.8.3",
+        "react-email": "6.5.0",
+        "tailwindcss": "4.3.0",
+        "tsc-alias": "1.8.17",
+        "tsx": "4.22.3",
+        "typescript": "6.0.3",
+        "typescript-eslint": "8.60.0"
     },
     "overrides": {
-        "esbuild": "0.27.4",
-        "dompurify": "3.3.2"
+        "esbuild": "0.28.1",
+        "dompurify": "3.4.0",
+        "postcss": "8.5.15"
     }
 }

server/aiGatewayServer.ts

@@ -0,0 +1,39 @@
+import express from "express";
+import helmet from "helmet";
+import cors from "cors";
+import config from "@server/lib/config";
+import logger from "@server/logger";
+import {
+    errorHandlerMiddleware,
+    notFoundMiddleware
+} from "@server/middlewares";
+import { createAiGatewayRouter } from "@server/routers/aiGateway";
+
+const aiGatewayPort = config.getRawConfig().server.ai_gateway_port;
+
+export function createAiGatewayServer() {
+    const aiGatewayServer = express();
+
+    const trustProxy = config.getRawConfig().server.trust_proxy;
+    if (trustProxy) {
+        aiGatewayServer.set("trust proxy", trustProxy);
+    }
+
+    aiGatewayServer.use(helmet());
+    aiGatewayServer.use(cors());
+    aiGatewayServer.use(express.json());
+
+    aiGatewayServer.use(createAiGatewayRouter());
+
+    aiGatewayServer.use(notFoundMiddleware);
+    aiGatewayServer.use(errorHandlerMiddleware);
+
+    aiGatewayServer.listen(aiGatewayPort, (err?: any) => {
+        if (err) throw err;
+        logger.info(
+            `AI gateway server is running on http://localhost:${aiGatewayPort}`
+        );
+    });
+
+    return aiGatewayServer;
+}

server/auth/actions.ts

@@ -5,6 +5,7 @@ import { and, eq, inArray } from "drizzle-orm";
 import createHttpError from "http-errors";
 import HttpCode from "@server/types/HttpCode";
 import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import logger from "@server/logger";
 
 export enum ActionsEnum {
     createOrgUser = "createOrgUser",
@@ -20,6 +21,8 @@ export enum ActionsEnum {
     getSite = "getSite",
     listSites = "listSites",
     updateSite = "updateSite",
+    updateSiteApprovals = "updateSiteApprovals",
+    restartSite = "restartSite",
     resetSiteBandwidth = "resetSiteBandwidth",
     reGenerateSecret = "reGenerateSecret",
     createResource = "createResource",
@@ -47,6 +50,8 @@ export enum ActionsEnum {
     setResourceUsers = "setResourceUsers",
     setResourceRoles = "setResourceRoles",
     listResourceUsers = "listResourceUsers",
+    listResourceAiModels = "listResourceAiModels",
+    setResourceAiModels = "setResourceAiModels",
     // removeRoleSite = "removeRoleSite",
     // addRoleAction = "addRoleAction",
     // removeRoleAction = "removeRoleAction",
@@ -69,6 +74,7 @@ export enum ActionsEnum {
     setResourceWhitelist = "setResourceWhitelist",
     getResourceWhitelist = "getResourceWhitelist",
     generateAccessToken = "generateAccessToken",
+    createResourceSessionToken = "createResourceSessionToken",
     deleteAcessToken = "deleteAcessToken",
     listAccessTokens = "listAccessTokens",
     createResourceRule = "createResourceRule",
@@ -148,11 +154,57 @@ export enum ActionsEnum {
     updateAlertRule = "updateAlertRule",
     deleteAlertRule = "deleteAlertRule",
     listAlertRules = "listAlertRules",
+    listOrgLabels = "listOrgLabels",
+    createOrgLabel = "createOrgLabel",
+    updateOrgLabel = "updateOrgLabel",
+    deleteOrgLabel = "deleteOrgLabel",
+    attachLabelToItem = "attachLabelToItem",
+    detachLabelFromItem = "detachLabelFromItem",
     getAlertRule = "getAlertRule",
     createHealthCheck = "createHealthCheck",
     updateHealthCheck = "updateHealthCheck",
     deleteHealthCheck = "deleteHealthCheck",
-    listHealthChecks = "listHealthChecks"
+    listHealthChecks = "listHealthChecks",
+    createBrowserGatewayTarget = "createBrowserGatewayTarget",
+    updateBrowserGatewayTarget = "updateBrowserGatewayTarget",
+    deleteBrowserGatewayTarget = "deleteBrowserGatewayTarget",
+    getBrowserGatewayTarget = "getBrowserGatewayTarget",
+    listBrowserGatewayTargets = "listBrowserGatewayTargets",
+    listResourcePolicies = "listResourcePolicies",
+    getResourcePolicy = "getResourcePolicy",
+    createResourcePolicy = "createResourcePolicy",
+    updateResourcePolicy = "updateResourcePolicy",
+    deleteResourcePolicy = "deleteResourcePolicy",
+    listResourcePolicyRoles = "listResourcePolicyRoles",
+    setResourcePolicyRoles = "setResourcePolicyRoles",
+    listResourcePolicyUsers = "listResourcePolicyUsers",
+    setResourcePolicyUsers = "setResourcePolicyUsers",
+    setResourcePolicyPassword = "setResourcePolicyPassword",
+    setResourcePolicyPincode = "setResourcePolicyPincode",
+    setResourcePolicyHeaderAuth = "setResourcePolicyHeaderAuth",
+    setResourcePolicyWhitelist = "setResourcePolicyWhitelist",
+    setResourcePolicyRules = "setResourcePolicyRules",
+    createOrgWideLauncherView = "createOrgWideLauncherView",
+    createAiProvider = "createAiProvider",
+    deleteAiProvider = "deleteAiProvider",
+    getAiProvider = "getAiProvider",
+    listAiProviders = "listAiProviders",
+    updateAiProvider = "updateAiProvider",
+    createAiModel = "createAiModel",
+    deleteAiModel = "deleteAiModel",
+    getAiModel = "getAiModel",
+    listAiModels = "listAiModels",
+    updateAiModel = "updateAiModel",
+    createAiBudget = "createAiBudget",
+    deleteAiBudget = "deleteAiBudget",
+    getAiBudget = "getAiBudget",
+    listAiBudgets = "listAiBudgets",
+    updateAiBudget = "updateAiBudget",
+    createVirtualApiKey = "createVirtualApiKey",
+    deleteVirtualApiKey = "deleteVirtualApiKey",
+    getVirtualApiKey = "getVirtualApiKey",
+    listVirtualApiKeys = "listVirtualApiKeys",
+    updateVirtualApiKey = "updateVirtualApiKey"
 }
 
 export async function checkUserActionPermission(
@@ -185,6 +237,23 @@ export async function checkUserActionPermission(
             }
         }
 
+        // If no direct permission, check role-based permission (any of user's roles)
+        const roleActionPermission = await db
+            .select()
+            .from(roleActions)
+            .where(
+                and(
+                    eq(roleActions.actionId, actionId),
+                    inArray(roleActions.roleId, userOrgRoleIds),
+                    eq(roleActions.orgId, req.userOrgId!)
+                )
+            )
+            .limit(1);
+
+        if (roleActionPermission.length > 0) {
+            return true;
+        }
+
         // Check if the user has direct permission for the action in the current org
         const userActionPermission = await db
             .select()
@@ -202,20 +271,7 @@ export async function checkUserActionPermission(
             return true;
         }
 
-        // If no direct permission, check role-based permission (any of user's roles)
-        const roleActionPermission = await db
-            .select()
-            .from(roleActions)
-            .where(
-                and(
-                    eq(roleActions.actionId, actionId),
-                    inArray(roleActions.roleId, userOrgRoleIds),
-                    eq(roleActions.orgId, req.userOrgId!)
-                )
-            )
-            .limit(1);
-
-        return roleActionPermission.length > 0;
+        return false;
     } catch (error) {
         console.error("Error checking user action permission:", error);
         throw createHttpError(

server/auth/canUserAccessResource.ts

@@ -1,6 +1,12 @@
 import { db } from "@server/db";
-import { and, eq, inArray } from "drizzle-orm";
-import { roleResources, userResources } from "@server/db";
+import { and, eq, inArray, isNull, or } from "drizzle-orm";
+import {
+    rolePolicies,
+    roleResources,
+    resources,
+    userPolicies,
+    userResources
+} from "@server/db";
 
 export async function canUserAccessResource({
     userId,
@@ -11,9 +17,14 @@ export async function canUserAccessResource({
     resourceId: number;
     roleIds: number[];
 }): Promise<boolean> {
-    const roleResourceAccess =
+    const [
+        roleResourceAccess,
+        rolePolicyAccess,
+        userResourceAccess,
+        userPolicyAccess
+    ] = await Promise.all([
         roleIds.length > 0
-            ? await db
+            ? db
                   .select()
                   .from(roleResources)
                   .where(
@@ -23,26 +34,87 @@ export async function canUserAccessResource({
                       )
                   )
                   .limit(1)
-            : [];
-
-    if (roleResourceAccess.length > 0) {
-        return true;
-    }
-
-    const userResourceAccess = await db
-        .select()
-        .from(userResources)
-        .where(
-            and(
-                eq(userResources.userId, userId),
-                eq(userResources.resourceId, resourceId)
+            : [],
+        roleIds.length > 0
+            ? db
+                  .select({
+                      roleId: rolePolicies.roleId,
+                      resourcePolicyId: rolePolicies.resourcePolicyId
+                  })
+                  .from(rolePolicies)
+                  .innerJoin(
+                      resources,
+                      // Shared policy wins; only use default policy when no shared
+                      // policy is assigned to the resource.
+                      or(
+                          eq(
+                              resources.resourcePolicyId,
+                              rolePolicies.resourcePolicyId
+                          ),
+                          and(
+                              isNull(resources.resourcePolicyId),
+                              eq(
+                                  resources.defaultResourcePolicyId,
+                                  rolePolicies.resourcePolicyId
+                              )
+                          )
+                      )
+                  )
+                  .where(
+                      and(
+                          eq(resources.resourceId, resourceId),
+                          inArray(rolePolicies.roleId, roleIds)
+                      )
+                  )
+                  .limit(1)
+            : [],
+        db
+            .select()
+            .from(userResources)
+            .where(
+                and(
+                    eq(userResources.userId, userId),
+                    eq(userResources.resourceId, resourceId)
+                )
             )
-        )
-        .limit(1);
+            .limit(1),
+        db
+            .select({
+                userId: userPolicies.userId,
+                resourcePolicyId: userPolicies.resourcePolicyId
+            })
+            .from(userPolicies)
+            .innerJoin(
+                resources,
+                // Shared policy wins; only use default policy when no shared
+                // policy is assigned to the resource.
+                or(
+                    eq(
+                        resources.resourcePolicyId,
+                        userPolicies.resourcePolicyId
+                    ),
+                    and(
+                        isNull(resources.resourcePolicyId),
+                        eq(
+                            resources.defaultResourcePolicyId,
+                            userPolicies.resourcePolicyId
+                        )
+                    )
+                )
+            )
+            .where(
+                and(
+                    eq(resources.resourceId, resourceId),
+                    eq(userPolicies.userId, userId)
+                )
+            )
+            .limit(1)
+    ]);
 
-    if (userResourceAccess.length > 0) {
-        return true;
-    }
-
-    return false;
+    return (
+        roleResourceAccess.length > 0 ||
+        rolePolicyAccess.length > 0 ||
+        userResourceAccess.length > 0 ||
+        userPolicyAccess.length > 0
+    );
 }

server/auth/sessions/app.ts

@@ -12,7 +12,7 @@ import {
     users
 } from "@server/db";
 import { db } from "@server/db";
-import { eq, inArray } from "drizzle-orm";
+import { and, eq, inArray, ne } from "drizzle-orm";
 import config from "@server/lib/config";
 import type { RandomReader } from "@oslojs/crypto/random";
 import { generateRandomString } from "@oslojs/crypto/random";
@@ -136,6 +136,45 @@ export async function invalidateAllSessions(userId: string): Promise<void> {
     }
 }
 
+export async function invalidateAllSessionsExceptCurrent(
+    userId: string,
+    currentSessionId: string
+): Promise<void> {
+    try {
+        await db.transaction(async (trx) => {
+            const userSessions = await trx
+                .select()
+                .from(sessions)
+                .where(
+                    and(
+                        eq(sessions.userId, userId),
+                        ne(sessions.sessionId, currentSessionId)
+                    )
+                );
+
+            if (userSessions.length > 0) {
+                await trx.delete(resourceSessions).where(
+                    inArray(
+                        resourceSessions.userSessionId,
+                        userSessions.map((s) => s.sessionId)
+                    )
+                );
+            }
+
+            await trx
+                .delete(sessions)
+                .where(
+                    and(
+                        eq(sessions.userId, userId),
+                        ne(sessions.sessionId, currentSessionId)
+                    )
+                );
+        });
+    } catch (e) {
+        logger.error("Failed to invalidate user sessions except current", e);
+    }
+}
+
 export function serializeSessionCookie(
     token: string,
     isSecure: boolean,

server/auth/sessions/resource.ts

@@ -19,6 +19,9 @@ export async function createResourceSession(opts: {
     userSessionId?: string | null;
     whitelistId?: number | null;
     accessTokenId?: string | null;
+    policyPasswordId?: number | null;
+    policyPincodeId?: number | null;
+    policyWhitelistId?: number | null;
     doNotExtend?: boolean;
     expiresAt?: number | null;
     sessionLength?: number | null;
@@ -28,7 +31,10 @@ export async function createResourceSession(opts: {
         !opts.pincodeId &&
         !opts.whitelistId &&
         !opts.accessTokenId &&
-        !opts.userSessionId
+        !opts.userSessionId &&
+        !opts.policyPasswordId &&
+        !opts.policyPincodeId &&
+        !opts.policyWhitelistId
     ) {
         throw new Error("Auth method must be provided");
     }
@@ -49,6 +55,9 @@ export async function createResourceSession(opts: {
         whitelistId: opts.whitelistId || null,
         doNotExtend: opts.doNotExtend || false,
         accessTokenId: opts.accessTokenId || null,
+        policyPasswordId: opts.policyPasswordId || null,
+        policyPincodeId: opts.policyPincodeId || null,
+        policyWhitelistId: opts.policyWhitelistId || null,
         isRequestToken: opts.isRequestToken || false,
         userSessionId: opts.userSessionId || null,
         issuedAt: new Date().getTime()

server/auth/verifyVirtualApiKey.ts

@@ -0,0 +1,311 @@
+import { canUserAccessResource } from "@server/auth/canUserAccessResource";
+import {
+    db,
+    users,
+    virtualApiKeyResources,
+    virtualApiKeys,
+    type VirtualApiKey
+} from "@server/db";
+import config from "@server/lib/config";
+import {
+    decryptVirtualApiKeyToken,
+    VIRTUAL_API_KEY_PREFIX,
+    looksLikeVirtualApiKeyCredential
+} from "@server/lib/virtualApiKey";
+import { getUserOrgRoles } from "@server/lib/userOrgRoles";
+import { and, eq } from "drizzle-orm";
+import { isWithinExpirationDate } from "oslo";
+
+export type VirtualApiKeyCredential = {
+    virtualApiKeyId: string;
+    secret: string;
+};
+
+export type VirtualApiKeyUserData = {
+    userId: string;
+    username: string;
+    email: string | null;
+    name: string | null;
+    role: string | null;
+};
+
+function getHeader(
+    headers: Record<string, string> | undefined,
+    name: string
+): string | undefined {
+    if (!headers) {
+        return undefined;
+    }
+    if (headers[name] !== undefined) {
+        return headers[name];
+    }
+    const lower = name.toLowerCase();
+    for (const [key, value] of Object.entries(headers)) {
+        if (key.toLowerCase() === lower) {
+            return value;
+        }
+    }
+    return undefined;
+}
+
+function parseVkCredential(
+    raw: string | undefined
+): VirtualApiKeyCredential | null {
+    if (!raw || !looksLikeVirtualApiKeyCredential(raw)) {
+        return null;
+    }
+    const withoutPrefix = raw.trim().slice(VIRTUAL_API_KEY_PREFIX.length);
+    const dot = withoutPrefix.indexOf(".");
+    return {
+        virtualApiKeyId: withoutPrefix.slice(0, dot),
+        secret: withoutPrefix.slice(dot + 1)
+    };
+}
+
+export function extractVirtualApiKeyCredential(
+    headers: Record<string, string> | undefined
+): VirtualApiKeyCredential | null {
+    if (!headers) {
+        return null;
+    }
+
+    const authorization = getHeader(headers, "authorization");
+    if (authorization) {
+        const bearerMatch = authorization.match(/^Bearer\s+(.+)$/i);
+        if (bearerMatch) {
+            const credential = parseVkCredential(bearerMatch[1]);
+            if (credential) {
+                return credential;
+            }
+        }
+        const splunkMatch = authorization.match(/^Splunk\s+(.+)$/i);
+        if (splunkMatch) {
+            const credential = parseVkCredential(splunkMatch[1]);
+            if (credential) {
+                return credential;
+            }
+        }
+    }
+
+    const cfAig = getHeader(headers, "cf-aig-authorization");
+    if (cfAig) {
+        const bearerMatch = cfAig.match(/^Bearer\s+(.+)$/i);
+        const credential = parseVkCredential(
+            bearerMatch ? bearerMatch[1] : cfAig
+        );
+        if (credential) {
+            return credential;
+        }
+    }
+
+    for (const name of ["x-api-key", "x-goog-api-key"] as const) {
+        const credential = parseVkCredential(getHeader(headers, name));
+        if (credential) {
+            return credential;
+        }
+    }
+
+    return null;
+}
+
+async function buildUserData(
+    userId: string,
+    orgId: string
+): Promise<VirtualApiKeyUserData | undefined> {
+    const [user] = await db
+        .select()
+        .from(users)
+        .where(eq(users.userId, userId))
+        .limit(1);
+
+    if (!user) {
+        return undefined;
+    }
+
+    if (
+        config.getRawConfig().flags?.require_email_verification &&
+        !user.emailVerified
+    ) {
+        return undefined;
+    }
+
+    const userOrgRoles = await getUserOrgRoles(user.userId, orgId);
+    if (userOrgRoles.length === 0) {
+        return undefined;
+    }
+
+    return {
+        userId: user.userId,
+        username: user.username,
+        email: user.email,
+        name: user.name,
+        role: userOrgRoles.map((r) => r.roleName).join(", ") || null
+    };
+}
+
+async function userHasResourceAccess(
+    userId: string,
+    resourceId: number,
+    orgId: string
+): Promise<{ allowed: boolean; userData?: VirtualApiKeyUserData }> {
+    const [user] = await db
+        .select()
+        .from(users)
+        .where(eq(users.userId, userId))
+        .limit(1);
+
+    if (!user) {
+        return { allowed: false };
+    }
+
+    if (
+        config.getRawConfig().flags?.require_email_verification &&
+        !user.emailVerified
+    ) {
+        return { allowed: false };
+    }
+
+    const userOrgRoles = await getUserOrgRoles(user.userId, orgId);
+    if (userOrgRoles.length === 0) {
+        return { allowed: false };
+    }
+
+    const allowed = await canUserAccessResource({
+        userId,
+        resourceId,
+        roleIds: userOrgRoles.map((r) => r.roleId)
+    });
+
+    if (!allowed) {
+        return { allowed: false };
+    }
+
+    return {
+        allowed: true,
+        userData: {
+            userId: user.userId,
+            username: user.username,
+            email: user.email,
+            name: user.name,
+            role: userOrgRoles.map((r) => r.roleName).join(", ") || null
+        }
+    };
+}
+
+async function manualKeyHasResourceAccess(
+    key: VirtualApiKey,
+    resourceId: number
+): Promise<boolean> {
+    if (key.allResources) {
+        return true;
+    }
+
+    const [row] = await db
+        .select({ resourceId: virtualApiKeyResources.resourceId })
+        .from(virtualApiKeyResources)
+        .where(
+            and(
+                eq(virtualApiKeyResources.virtualApiKeyId, key.virtualApiKeyId),
+                eq(virtualApiKeyResources.resourceId, resourceId)
+            )
+        )
+        .limit(1);
+
+    return Boolean(row);
+}
+
+async function touchLastUsedAt(virtualApiKeyId: string): Promise<void> {
+    try {
+        await db
+            .update(virtualApiKeys)
+            .set({ lastUsedAt: Date.now() })
+            .where(eq(virtualApiKeys.virtualApiKeyId, virtualApiKeyId));
+    } catch {
+        // Best-effort; do not fail auth on audit timestamp updates.
+    }
+}
+
+export async function verifyVirtualApiKey({
+    credential,
+    resourceId,
+    orgId
+}: {
+    credential: VirtualApiKeyCredential;
+    resourceId: number;
+    orgId: string;
+}): Promise<{
+    valid: boolean;
+    error?: string;
+    key?: VirtualApiKey;
+    userData?: VirtualApiKeyUserData;
+}> {
+    const [key] = await db
+        .select()
+        .from(virtualApiKeys)
+        .where(eq(virtualApiKeys.virtualApiKeyId, credential.virtualApiKeyId))
+        .limit(1);
+
+    if (!key) {
+        return { valid: false, error: "Virtual API key not found" };
+    }
+
+    if (key.orgId !== orgId) {
+        return { valid: false, error: "Virtual API key org mismatch" };
+    }
+
+    let plaintext: string;
+    try {
+        plaintext = decryptVirtualApiKeyToken(key.token);
+    } catch {
+        return { valid: false, error: "Virtual API key secret is invalid" };
+    }
+
+    if (plaintext !== credential.secret) {
+        return { valid: false, error: "Invalid virtual API key secret" };
+    }
+
+    if (key.expiresAt && !isWithinExpirationDate(new Date(key.expiresAt))) {
+        return { valid: false, error: "Virtual API key has expired" };
+    }
+
+    if (key.kind === "manual") {
+        const scoped = await manualKeyHasResourceAccess(key, resourceId);
+        if (!scoped) {
+            return {
+                valid: false,
+                error: "Virtual API key is not scoped to this resource"
+            };
+        }
+
+        let userData: VirtualApiKeyUserData | undefined;
+        if (key.userId) {
+            userData = await buildUserData(key.userId, orgId);
+        }
+
+        await touchLastUsedAt(key.virtualApiKeyId);
+        return { valid: true, key, userData };
+    }
+
+    if (key.kind === "user") {
+        if (!key.userId) {
+            return { valid: false, error: "User virtual API key has no user" };
+        }
+
+        const access = await userHasResourceAccess(
+            key.userId,
+            resourceId,
+            orgId
+        );
+        if (!access.allowed || !access.userData) {
+            return {
+                valid: false,
+                error: "User is not allowed to access this resource"
+            };
+        }
+
+        await touchLastUsedAt(key.virtualApiKeyId);
+        return { valid: true, key, userData: access.userData };
+    }
+
+    return { valid: false, error: "Unknown virtual API key kind" };
+}

server/cleanup.ts

@@ -3,12 +3,16 @@ import { flushConnectionLogToDb } from "#dynamic/routers/newt";
 import { flushSiteBandwidthToDb } from "@server/routers/gerbil/receiveBandwidth";
 import { stopPingAccumulator } from "@server/routers/newt/pingAccumulator";
 import { cleanup as wsCleanup } from "#dynamic/routers/ws";
+import { shutdownUsageRecorder } from "@server/lib/aiBudgetEnforcement";
+import { shutdownAiSessionLogger } from "@server/routers/aiGateway/logAiSession";
 
 async function cleanup() {
     await stopPingAccumulator();
     await flushBandwidthToDb();
     await flushConnectionLogToDb();
     await flushSiteBandwidthToDb();
+    await shutdownUsageRecorder();
+    await shutdownAiSessionLogger();
     await wsCleanup();
 
     process.exit(0);

server/db/countries.ts

@@ -795,10 +795,13 @@ export const COUNTRIES = [
         name: "Serbia",
         code: "RS"
     },
-    {
-        name: "Serbia and Montenegro",
-        code: "CS"
-    },
+    // Removed as this is a deprecated ISO country code, not supported anymore
+    // Also the individual flags for Serbia & Montenegro are already included in the list
+    // more details: https://en.wikipedia.org/wiki/ISO_3166-2:CS
+    // {
+    //     name: "Serbia and Montenegro",
+    //     code: "CS"
+    // },
     {
         name: "Seychelles",
         code: "SC"

server/db/names.ts

@@ -1,6 +1,13 @@
 import { join } from "path";
 import { readFileSync } from "fs";
-import { clients, db, resources, siteResources } from "@server/db";
+import {
+    aiProviders,
+    clients,
+    db,
+    resourcePolicies,
+    resources,
+    siteResources
+} from "@server/db";
 import { randomInt } from "crypto";
 import { exitNodes, sites } from "@server/db";
 import { eq, and } from "drizzle-orm";
@@ -107,6 +114,61 @@ export async function getUniqueResourceName(orgId: string): Promise<string> {
     }
 }
 
+export async function getUniqueProviderName(orgId: string): Promise<string> {
+    let loops = 0;
+    while (true) {
+        if (loops > 100) {
+            throw new Error("Could not generate a unique name");
+        }
+
+        const name = generateName();
+
+        const aiProviderCount = await db
+            .select({
+                niceId: aiProviders.niceId,
+                orgId: aiProviders.orgId
+            })
+            .from(aiProviders)
+            .where(
+                and(eq(aiProviders.niceId, name), eq(aiProviders.orgId, orgId))
+            );
+
+        if (aiProviderCount.length === 0) {
+            return name;
+        }
+        loops++;
+    }
+}
+
+export async function getUniqueResourcePolicyName(
+    orgId: string
+): Promise<string> {
+    let loops = 0;
+    while (true) {
+        if (loops > 100) {
+            throw new Error("Could not generate a unique name");
+        }
+
+        const name = generateName();
+        const policyCount = await db
+            .select({
+                niceId: resourcePolicies.niceId,
+                orgId: resourcePolicies.orgId
+            })
+            .from(resourcePolicies)
+            .where(
+                and(
+                    eq(resourcePolicies.niceId, name),
+                    eq(resourcePolicies.orgId, orgId)
+                )
+            );
+        if (policyCount.length === 0) {
+            return name;
+        }
+        loops++;
+    }
+}
+
 export async function getUniqueSiteResourceName(
     orgId: string
 ): Promise<string> {

server/db/pg/driver.ts

@@ -87,7 +87,7 @@ function createDb() {
 
 export const db = createDb();
 export default db;
-export const primaryDb = db.$primary as typeof db; // is this typeof a problem - techincally they are different types
+export const primaryDb = db.$primary as typeof db; // is this typeof a problem - technically they are different types
 export type Transaction = Parameters<
     Parameters<(typeof db)["transaction"]>[0]
 >[0];

server/db/pg/index.ts

@@ -4,3 +4,4 @@ export * from "./safeRead";
 export * from "./schema/schema";
 export * from "./schema/privateSchema";
 export * from "./migrate";
+export { alias } from "drizzle-orm/pg-core";

server/db/pg/logsDriver.ts

@@ -2,7 +2,7 @@ import { drizzle as DrizzlePostgres } from "drizzle-orm/node-postgres";
 import { readConfigFile } from "@server/lib/readConfigFile";
 import { withReplicas } from "drizzle-orm/pg-core";
 import { build } from "@server/build";
-import { db as mainDb, primaryDb as mainPrimaryDb } from "./driver";
+import { db as mainDb } from "./driver";
 import { createPool } from "./poolConfig";
 
 function createLogsDb() {
@@ -63,8 +63,7 @@ function createLogsDb() {
             })
         );
     } else {
-        const maxReplicaConnections =
-            poolConfig?.max_replica_connections || 20;
+        const maxReplicaConnections = poolConfig?.max_replica_connections || 20;
         for (const conn of replicaConnections) {
             const replicaPool = createPool(
                 conn.connection_string,
@@ -91,4 +90,4 @@ function createLogsDb() {
 
 export const logsDb = createLogsDb();
 export default logsDb;
-export const primaryLogsDb = logsDb.$primary;
+export const primaryLogsDb = logsDb.$primary;

server/db/pg/poolConfig.ts

@@ -1,5 +1,5 @@
+import config from "@server/lib/config";
 import { Pool, PoolConfig } from "pg";
-import logger from "@server/logger";
 
 export function createPoolConfig(
     connectionString: string,
@@ -27,7 +27,7 @@ export function attachPoolErrorHandlers(pool: Pool, label: string): void {
     pool.on("error", (err) => {
         // This catches errors on idle clients in the pool. Without this
         // handler an unexpected disconnect would crash the process.
-        logger.error(
+        console.error(
             `Unexpected error on idle ${label} database client: ${err.message}`
         );
     });
@@ -36,10 +36,32 @@ export function attachPoolErrorHandlers(pool: Pool, label: string): void {
         // Set a statement timeout on every new connection so a single slow
         // query can't block the pool forever
         client.query("SET statement_timeout = '30s'").catch((err: Error) => {
-            logger.warn(
+            console.warn(
                 `Failed to set statement_timeout on ${label} client: ${err.message}`
             );
         });
+
+        // Disable JIT compilation for this connection. Our hot-path queries
+        // (e.g. resource-by-domain lookups) join many tables but only ever
+        // return a handful of rows. When planner row estimates drift (e.g.
+        // due to autovacuum lag under write-heavy load), Postgres decides
+        // these plans are expensive enough to JIT-compile, which can add
+        // multiple seconds of pure compilation overhead per query and
+        // saturate the connection pool. JIT never pays off for these
+        // short-lived OLTP queries, so it's disabled outright rather than
+        // relying on statistics staying fresh.
+        //
+        // Set via a runtime SET command rather than the `options: "-c
+        // jit=off"` startup parameter: connections in SaaS mode go through
+        // a pooler (e.g. PgBouncer) that rejects arbitrary startup packet
+        // options with a protocol_violation (08P01) error.
+        if (config.getRawConfig().postgres?.pool.jit_mode == false) {
+            client.query("SET jit = off").catch((err: Error) => {
+                console.warn(
+                    `Failed to set jit=off on ${label} client: ${err.message}`
+                );
+            });
+        }
     });
 }
 
@@ -60,4 +82,4 @@ export function createPool(
     );
     attachPoolErrorHandlers(pool, label);
     return pool;
-}
+}

server/db/pg/schema/privateSchema.ts

@@ -2,6 +2,7 @@ import {
     pgTable,
     serial,
     varchar,
+    unique,
     boolean,
     integer,
     bigint,
@@ -11,7 +12,7 @@ import {
     primaryKey,
     uniqueIndex
 } from "drizzle-orm/pg-core";
-import { InferSelectModel } from "drizzle-orm";
+import { InferSelectModel, sql } from "drizzle-orm";
 import {
     domains,
     orgs,
@@ -19,33 +20,15 @@ import {
     roles,
     users,
     exitNodes,
-    sessions,
-    clients,
     resources,
     siteResources,
     targetHealthCheck,
-    sites
+    sites,
+    clients,
+    sessions,
+    labels
 } from "./schema";
 
-export const certificates = pgTable("certificates", {
-    certId: serial("certId").primaryKey(),
-    domain: varchar("domain", { length: 255 }).notNull().unique(),
-    domainId: varchar("domainId").references(() => domains.domainId, {
-        onDelete: "cascade"
-    }),
-    wildcard: boolean("wildcard").default(false),
-    status: varchar("status", { length: 50 }).notNull().default("pending"), // pending, requested, valid, expired, failed
-    expiresAt: bigint("expiresAt", { mode: "number" }),
-    lastRenewalAttempt: bigint("lastRenewalAttempt", { mode: "number" }),
-    createdAt: bigint("createdAt", { mode: "number" }).notNull(),
-    updatedAt: bigint("updatedAt", { mode: "number" }).notNull(),
-    orderId: varchar("orderId", { length: 500 }),
-    errorMessage: text("errorMessage"),
-    renewalCount: integer("renewalCount").default(0),
-    certFile: text("certFile"),
-    keyFile: text("keyFile")
-});
-
 export const dnsChallenge = pgTable("dnsChallenges", {
     dnsChallengeId: serial("dnsChallengeId").primaryKey(),
     domain: varchar("domain", { length: 255 }).notNull(),
@@ -93,7 +76,8 @@ export const subscriptions = pgTable("subscriptions", {
     billingCycleAnchor: bigint("billingCycleAnchor", { mode: "number" }),
     expiresAt: bigint("expiresAt", { mode: "number" }),
     trial: boolean("trial").default(false),
-    type: varchar("type", { length: 50 }) // tier1, tier2, tier3, or license
+    type: varchar("type", { length: 50 }), // tier1, tier2, tier3, or license
+    override: boolean("override").default(false)
 });
 
 export const subscriptionItems = pgTable("subscriptionItems", {
@@ -197,6 +181,42 @@ export const remoteExitNodes = pgTable("remoteExitNode", {
     })
 });
 
+export const remoteExitNodeResources = pgTable("remoteExitNodeResources", {
+    remoteExitNodeResourceId: serial("remoteExitNodeResourceId").primaryKey(),
+    remoteExitNodeId: varchar("remoteExitNodeId")
+        .notNull()
+        .references(() => remoteExitNodes.remoteExitNodeId, {
+            onDelete: "cascade"
+        }),
+    destination: varchar("destination").notNull() // a cidr range
+});
+
+export const remoteExitNodePreferenceLabels = pgTable(
+    // this controls what sites are enforced to connect to this node
+    "remoteExitNodePreferenceLabels",
+    {
+        remoteExitNodePreferenceLabelId: serial(
+            "remoteExitNodePreferenceLabelId"
+        ).primaryKey(),
+        remoteExitNodeId: varchar("remoteExitNodeId")
+            .references(() => remoteExitNodes.remoteExitNodeId, {
+                onDelete: "cascade"
+            })
+            .notNull(),
+        labelId: integer("labelId")
+            .references(() => labels.labelId, {
+                onDelete: "cascade"
+            })
+            .notNull()
+    },
+    (t) => [
+        unique("remote_exit_node_preference_label_uniq").on(
+            t.remoteExitNodeId,
+            t.labelId
+        )
+    ]
+);
+
 export const remoteExitNodeSessions = pgTable("remoteExitNodeSession", {
     sessionId: varchar("id").primaryKey(),
     remoteExitNodeId: varchar("remoteExitNodeId")
@@ -207,17 +227,28 @@ export const remoteExitNodeSessions = pgTable("remoteExitNodeSession", {
     expiresAt: bigint("expiresAt", { mode: "number" }).notNull()
 });
 
-export const loginPage = pgTable("loginPage", {
-    loginPageId: serial("loginPageId").primaryKey(),
-    subdomain: varchar("subdomain"),
-    fullDomain: varchar("fullDomain"),
-    exitNodeId: integer("exitNodeId").references(() => exitNodes.exitNodeId, {
-        onDelete: "set null"
-    }),
-    domainId: varchar("domainId").references(() => domains.domainId, {
-        onDelete: "set null"
-    })
-});
+export const loginPage = pgTable(
+    "loginPage",
+    {
+        loginPageId: serial("loginPageId").primaryKey(),
+        subdomain: varchar("subdomain"),
+        fullDomain: varchar("fullDomain"),
+        exitNodeId: integer("exitNodeId").references(
+            () => exitNodes.exitNodeId,
+            {
+                onDelete: "set null"
+            }
+        ),
+        domainId: varchar("domainId").references(() => domains.domainId, {
+            onDelete: "set null"
+        })
+    },
+    (t) => [
+        index("idx_loginpage_fulldomain")
+            .on(t.fullDomain)
+            .where(sql`${t.fullDomain} IS NOT NULL`)
+    ]
+);
 
 export const loginPageOrg = pgTable("loginPageOrg", {
     loginPageId: integer("loginPageId")
@@ -583,7 +614,6 @@ export const trialNotifications = pgTable("trialNotifications", {
 export type Approval = InferSelectModel<typeof approvals>;
 export type Limit = InferSelectModel<typeof limits>;
 export type Account = InferSelectModel<typeof account>;
-export type Certificate = InferSelectModel<typeof certificates>;
 export type DnsChallenge = InferSelectModel<typeof dnsChallenge>;
 export type Customer = InferSelectModel<typeof customers>;
 export type Subscription = InferSelectModel<typeof subscriptions>;

server/db/queries/verifySessionQueries.ts

@@ -17,22 +17,39 @@ import {
     resourceHeaderAuth,
     ResourceHeaderAuth,
     resourceRules,
+    resourcePolicyRules,
     resources,
     roleResources,
+    rolePolicies,
     sessions,
     userResources,
+    userPolicies,
     users,
     ResourceHeaderAuthExtendedCompatibility,
-    resourceHeaderAuthExtendedCompatibility
+    resourceHeaderAuthExtendedCompatibility,
+    resourcePolicies,
+    resourcePolicyPincode,
+    ResourcePolicyPincode,
+    resourcePolicyPassword,
+    ResourcePolicyPassword,
+    resourcePolicyHeaderAuth,
+    ResourcePolicyHeaderAuth,
+    resourceWhitelist,
+    resourcePolicyWhiteList
 } from "@server/db";
-import { and, eq, inArray, or, sql } from "drizzle-orm";
+import { alias } from "@server/db";
+import { and, eq, inArray, isNull, or, sql } from "drizzle-orm";
+import logger from "@server/logger";
 
 export type ResourceWithAuth = {
     resource: Resource | null;
-    pincode: ResourcePincode | null;
-    password: ResourcePassword | null;
-    headerAuth: ResourceHeaderAuth | null;
+    pincode: ResourcePincode | ResourcePolicyPincode | null;
+    password: ResourcePassword | ResourcePolicyPassword | null;
+    headerAuth: ResourceHeaderAuth | ResourcePolicyHeaderAuth | null;
     headerAuthExtendedCompatibility: ResourceHeaderAuthExtendedCompatibility | null;
+    applyRules: boolean | null;
+    sso: boolean | null;
+    emailWhitelistEnabled: boolean | null;
     org: Org;
 };
 
@@ -57,6 +74,33 @@ export async function getResourceByDomain(
         wildcardCandidates.push(`*.${parts.slice(i).join(".")}`);
     }
 
+    const sharedPolicy = alias(resourcePolicies, "sharedPolicy");
+    const defaultPolicy = alias(resourcePolicies, "defaultPolicy");
+    const sharedPolicyPincode = alias(
+        resourcePolicyPincode,
+        "sharedPolicyPincode"
+    );
+    const defaultPolicyPincode = alias(
+        resourcePolicyPincode,
+        "defaultPolicyPincode"
+    );
+    const sharedPolicyPassword = alias(
+        resourcePolicyPassword,
+        "sharedPolicyPassword"
+    );
+    const defaultPolicyPassword = alias(
+        resourcePolicyPassword,
+        "defaultPolicyPassword"
+    );
+    const sharedPolicyHeaderAuth = alias(
+        resourcePolicyHeaderAuth,
+        "sharedPolicyHeaderAuth"
+    );
+    const defaultPolicyHeaderAuth = alias(
+        resourcePolicyHeaderAuth,
+        "defaultPolicyHeaderAuth"
+    );
+
     const potentialResults = await db
         .select()
         .from(resources)
@@ -79,6 +123,59 @@ export async function getResourceByDomain(
                 resources.resourceId
             )
         )
+        .leftJoin(
+            sharedPolicy,
+            eq(sharedPolicy.resourcePolicyId, resources.resourcePolicyId)
+        )
+        .leftJoin(
+            sharedPolicyPincode,
+            eq(
+                sharedPolicyPincode.resourcePolicyId,
+                sharedPolicy.resourcePolicyId
+            )
+        )
+        .leftJoin(
+            sharedPolicyPassword,
+            eq(
+                sharedPolicyPassword.resourcePolicyId,
+                sharedPolicy.resourcePolicyId
+            )
+        )
+        .leftJoin(
+            sharedPolicyHeaderAuth,
+            eq(
+                sharedPolicyHeaderAuth.resourcePolicyId,
+                sharedPolicy.resourcePolicyId
+            )
+        )
+        .leftJoin(
+            defaultPolicy,
+            eq(
+                defaultPolicy.resourcePolicyId,
+                resources.defaultResourcePolicyId
+            )
+        )
+        .leftJoin(
+            defaultPolicyPincode,
+            eq(
+                defaultPolicyPincode.resourcePolicyId,
+                defaultPolicy.resourcePolicyId
+            )
+        )
+        .leftJoin(
+            defaultPolicyPassword,
+            eq(
+                defaultPolicyPassword.resourcePolicyId,
+                defaultPolicy.resourcePolicyId
+            )
+        )
+        .leftJoin(
+            defaultPolicyHeaderAuth,
+            eq(
+                defaultPolicyHeaderAuth.resourcePolicyId,
+                defaultPolicy.resourcePolicyId
+            )
+        )
         .innerJoin(orgs, eq(orgs.orgId, resources.orgId))
         .where(
             or(
@@ -108,13 +205,51 @@ export async function getResourceByDomain(
         return null;
     }
 
+    // If a shared (custom) policy is assigned to the resource, use ONLY
+    // its values — do not fall back to the default policy. The default
+    // policy is only consulted when no shared policy is assigned at all.
+    const hasSharedPolicy = result.sharedPolicy !== null;
+
+    const effectivePolicyPincode = hasSharedPolicy
+        ? result.sharedPolicyPincode
+        : (result.defaultPolicyPincode ?? null);
+    const effectivePolicyPassword = hasSharedPolicy
+        ? result.sharedPolicyPassword
+        : (result.defaultPolicyPassword ?? null);
+    const effectivePolicyHeaderAuth = hasSharedPolicy
+        ? result.sharedPolicyHeaderAuth
+        : (result.defaultPolicyHeaderAuth ?? null);
+    const selectedPolicy = hasSharedPolicy
+        ? result.sharedPolicy
+        : result.defaultPolicy;
+    const effectiveApplyRules =
+        selectedPolicy?.applyRules ?? result.resources.applyRules;
+    const effectiveSSO = selectedPolicy?.sso ?? result.resources.sso;
+    const effectiveEmailWhitelistEnabled =
+        selectedPolicy?.emailWhitelistEnabled ??
+        result.resources.emailWhitelistEnabled;
+
     return {
-        resource: result.resources,
-        pincode: result.resourcePincode,
-        password: result.resourcePassword,
-        headerAuth: result.resourceHeaderAuth,
-        headerAuthExtendedCompatibility:
-            result.resourceHeaderAuthExtendedCompatibility,
+        resource: {
+            ...result.resources,
+            applyRules: effectiveApplyRules,
+            sso: effectiveSSO,
+            emailWhitelistEnabled: effectiveEmailWhitelistEnabled
+        }, // doing this for backward compatability so the remote nodes get the value as part of the resource struct
+        pincode: effectivePolicyPincode ?? result.resourcePincode,
+        password: effectivePolicyPassword ?? result.resourcePassword,
+        headerAuth: effectivePolicyHeaderAuth ?? result.resourceHeaderAuth,
+        headerAuthExtendedCompatibility: effectivePolicyHeaderAuth
+            ? ({
+                  headerAuthExtendedCompatibilityId: 0,
+                  resourceId: result.resources.resourceId,
+                  extendedCompatibilityIsActivated:
+                      effectivePolicyHeaderAuth.extendedCompatibility
+              } as ResourceHeaderAuthExtendedCompatibility)
+            : result.resourceHeaderAuthExtendedCompatibility,
+        applyRules: effectiveApplyRules,
+        sso: effectiveSSO,
+        emailWhitelistEnabled: effectiveEmailWhitelistEnabled,
         org: result.orgs
     };
 }
@@ -154,58 +289,195 @@ export async function getRoleName(roleId: number): Promise<string | null> {
 }
 
 /**
- * Check if role has access to resource
+ * Check if role has access to resource (direct or via resource policy)
  */
 export async function getRoleResourceAccess(
     resourceId: number,
     roleIds: number[]
 ) {
-    const roleResourceAccess = await db
-        .select()
-        .from(roleResources)
-        .where(
-            and(
-                eq(roleResources.resourceId, resourceId),
-                inArray(roleResources.roleId, roleIds)
+    const [direct, viaPolicies] = await Promise.all([
+        db
+            .select()
+            .from(roleResources)
+            .where(
+                and(
+                    eq(roleResources.resourceId, resourceId),
+                    inArray(roleResources.roleId, roleIds)
+                )
+            ),
+        db
+            .select({
+                roleId: rolePolicies.roleId,
+                resourcePolicyId: rolePolicies.resourcePolicyId
+            })
+            .from(rolePolicies)
+            .innerJoin(
+                resources,
+                // Shared policy wins; only use default policy when no shared
+                // policy is assigned to the resource.
+                or(
+                    eq(
+                        resources.resourcePolicyId,
+                        rolePolicies.resourcePolicyId
+                    ),
+                    and(
+                        isNull(resources.resourcePolicyId),
+                        eq(
+                            resources.defaultResourcePolicyId,
+                            rolePolicies.resourcePolicyId
+                        )
+                    )
+                )
             )
-        );
+            .where(
+                and(
+                    eq(resources.resourceId, resourceId),
+                    inArray(rolePolicies.roleId, roleIds)
+                )
+            )
+    ]);
 
-    return roleResourceAccess.length > 0 ? roleResourceAccess : null;
+    const combined = [...direct, ...viaPolicies];
+    return combined.length > 0 ? combined : null;
 }
 
 /**
- * Check if user has direct access to resource
+ * Check if user has access to resource (direct or via resource policy)
  */
 export async function getUserResourceAccess(
     userId: string,
     resourceId: number
 ) {
-    const userResourceAccess = await db
-        .select()
-        .from(userResources)
-        .where(
-            and(
-                eq(userResources.userId, userId),
-                eq(userResources.resourceId, resourceId)
+    const [direct, viaPolicies] = await Promise.all([
+        db
+            .select()
+            .from(userResources)
+            .where(
+                and(
+                    eq(userResources.userId, userId),
+                    eq(userResources.resourceId, resourceId)
+                )
             )
-        )
-        .limit(1);
+            .limit(1),
+        db
+            .select({
+                userId: userPolicies.userId,
+                resourcePolicyId: userPolicies.resourcePolicyId
+            })
+            .from(userPolicies)
+            .innerJoin(
+                resources,
+                // Shared policy wins; only use default policy when no shared
+                // policy is assigned to the resource.
+                or(
+                    eq(
+                        resources.resourcePolicyId,
+                        userPolicies.resourcePolicyId
+                    ),
+                    and(
+                        isNull(resources.resourcePolicyId),
+                        eq(
+                            resources.defaultResourcePolicyId,
+                            userPolicies.resourcePolicyId
+                        )
+                    )
+                )
+            )
+            .where(
+                and(
+                    eq(resources.resourceId, resourceId),
+                    eq(userPolicies.userId, userId)
+                )
+            )
+            .limit(1)
+    ]);
 
-    return userResourceAccess.length > 0 ? userResourceAccess[0] : null;
+    return direct[0] ?? viaPolicies[0] ?? null;
 }
 
 /**
- * Get resource rules for a given resource
+ * Get resource rules for a given resource (direct and via resource policy)
  */
 export async function getResourceRules(
     resourceId: number
 ): Promise<ResourceRule[]> {
-    const rules = await db
-        .select()
-        .from(resourceRules)
-        .where(eq(resourceRules.resourceId, resourceId));
+    const [directRules, policyRules] = await Promise.all([
+        db
+            .select()
+            .from(resourceRules)
+            .where(eq(resourceRules.resourceId, resourceId)),
+        db
+            .select({
+                ruleId: resourcePolicyRules.ruleId,
+                resourceId: sql<number>`${resourceId}`,
+                enabled: resourcePolicyRules.enabled,
+                priority: resourcePolicyRules.priority,
+                action: resourcePolicyRules.action,
+                match: resourcePolicyRules.match,
+                value: resourcePolicyRules.value
+            })
+            .from(resourcePolicyRules)
+            .innerJoin(
+                resources,
+                // Shared policy wins; only use default policy when no shared
+                // policy is assigned to the resource.
+                or(
+                    eq(
+                        resources.resourcePolicyId,
+                        resourcePolicyRules.resourcePolicyId
+                    ),
+                    and(
+                        isNull(resources.resourcePolicyId),
+                        eq(
+                            resources.defaultResourcePolicyId,
+                            resourcePolicyRules.resourcePolicyId
+                        )
+                    )
+                )
+            )
+            .where(eq(resources.resourceId, resourceId))
+    ]);
 
-    return rules;
+    const maxDirectPriority = directRules.reduce(
+        (max, r) => Math.max(max, r.priority),
+        0
+    );
+    const offsetPolicyRules = policyRules.map((r) => ({
+        ...r,
+        priority: maxDirectPriority + r.priority
+    }));
+
+    return [...directRules, ...offsetPolicyRules] as ResourceRule[];
+}
+
+/**
+ * Get the whitelisted email associated with a resource session's whitelist
+ * match (either a direct resource whitelist entry or a resource policy
+ * whitelist entry).
+ */
+export async function getWhitelistEmail(
+    whitelistId?: number | null,
+    policyWhitelistId?: number | null
+): Promise<string | null> {
+    if (whitelistId) {
+        const [row] = await db
+            .select({ email: resourceWhitelist.email })
+            .from(resourceWhitelist)
+            .where(eq(resourceWhitelist.whitelistId, whitelistId))
+            .limit(1);
+        return row?.email ?? null;
+    }
+
+    if (policyWhitelistId) {
+        const [row] = await db
+            .select({ email: resourcePolicyWhiteList.email })
+            .from(resourcePolicyWhiteList)
+            .where(eq(resourcePolicyWhiteList.whitelistId, policyWhitelistId))
+            .limit(1);
+        return row?.email ?? null;
+    }
+
+    return null;
 }
 
 /**

server/db/sqlite/driver.ts

@@ -1,79 +1,46 @@
 import { drizzle as DrizzleSqlite } from "drizzle-orm/better-sqlite3";
 import Database from "better-sqlite3";
-import type BetterSqlite3 from "better-sqlite3";
 import * as schema from "./schema/schema";
 import path from "path";
 import fs from "fs";
 import { APP_PATH } from "@server/lib/consts";
 import { existsSync, mkdirSync } from "fs";
+import logger from "@server/logger";
 
 export const location = path.join(APP_PATH, "db", "db.sqlite");
 export const exists = checkFileExists(location);
 
 bootstrapVolume();
 
-/**
- * Wraps better-sqlite3 Statement to call `finalize()` immediately after
- * execution, freeing native sqlite3_stmt memory deterministically instead
- * of waiting for GC. Fixes steady off-heap growth under load (#2120).
- * WARNING: Finalizes after first execution — incompatible with drizzle's
- * reusable .prepare() builders. No such usage exists in this codebase.
- */
-function autoFinalizeStatement(
-    stmt: BetterSqlite3.Statement
-): BetterSqlite3.Statement {
-    const wrapExec = <T extends (...args: any[]) => any>(fn: T): T => {
-        return function (this: any, ...args: any[]) {
-            try {
-                return fn.apply(this, args);
-            } finally {
-                try {
-                    // finalize() exists on the native Statement at runtime but
-                    // is missing from @types/better-sqlite3.
-                    (stmt as any).finalize();
-                } catch {
-                    // Already finalized — harmless
-                }
-            }
-        } as unknown as T;
-    };
-
-    stmt.run = wrapExec(stmt.run);
-    stmt.get = wrapExec(stmt.get);
-    stmt.all = wrapExec(stmt.all);
-
-    return stmt;
-}
-
 function createDb() {
-    const sqlite = new Database(location);
+    const verbose =
+        process.env.QUERY_LOGGING == "true"
+            ? (message: unknown) => logger.debug(String(message))
+            : undefined;
+    const sqlite = new Database(location, { verbose });
 
     if (process.env.ENABLE_SQLITE_WAL_MODE == "true") {
         // Enable WAL mode — allows concurrent readers + single writer, preventing
         // contention across subsystems (verifySession, Traefik, audit, ping).
+        // NOTE: journal_mode persists in the DB file once set; unsetting this
+        // env var does NOT revert an existing WAL database.
         sqlite.pragma("journal_mode = WAL");
         // NORMAL sync mode: safe with WAL, reduces write lock hold time.
         sqlite.pragma("synchronous = NORMAL");
     }
 
-    // Wait up to 5s on SQLITE_BUSY instead of failing — prevents audit log
-    // retry loops that accumulate memory.
-    sqlite.pragma("busy_timeout = 5000");
+    // No busy_timeout pragma: better-sqlite3 already arms
+    // sqlite3_busy_timeout(db, 5000) via its default `timeout` option
+    // (lib/database.js), so an explicit pragma is redundant.
 
-    // 64 MB page cache (default 2 MB) — reduces I/O round-trips on large
-    // TraefikConfigManager JOINs that block the event loop.
-    sqlite.pragma("cache_size = -65536");
+    // Intentionally NOT setting cache_size or mmap_size: a large page cache plus
+    // a multi-hundred-MB mmap region inflate RSS and cause page-cache thrashing
+    // on small (~1 GB) instances. Leave SQLite on its conservative defaults.
 
-    // 256 MB memory-mapped I/O — OS serves reads from page cache directly,
-    // reducing event-loop blocking.
-    sqlite.pragma("mmap_size = 268435456");
-
-    // Wrap prepare() so every drizzle-orm statement is auto-finalized after
-    // first use, preventing sqlite3_stmt accumulation between GC cycles.
-    const originalPrepare = sqlite.prepare.bind(sqlite);
-    (sqlite as any).prepare = function autoFinalizePrepare(source: string) {
-        return autoFinalizeStatement(originalPrepare(source));
-    };
+    // Intentionally NOT wrapping prepare()/statements: better-sqlite3 finalizes
+    // sqlite3_stmt in the Statement destructor at GC, and drizzle-orm prepares a
+    // fresh statement per query (no statement cache), so statements cannot
+    // accumulate. better-sqlite3 11.x exposes no Statement.finalize() at all.
 
     return DrizzleSqlite(sqlite, {
         schema

server/db/sqlite/index.ts

@@ -4,3 +4,4 @@ export * from "./safeRead";
 export * from "./schema/schema";
 export * from "./schema/privateSchema";
 export * from "./migrate";
+export { alias } from "drizzle-orm/sqlite-core";

server/db/sqlite/schema/privateSchema.ts

@@ -12,6 +12,7 @@ import {
     clients,
     domains,
     exitNodes,
+    labels,
     orgs,
     resources,
     roles,
@@ -21,28 +22,6 @@ import {
     targetHealthCheck,
     users
 } from "./schema";
-import { serial, varchar } from "drizzle-orm/mysql-core";
-import { pgTable } from "drizzle-orm/pg-core";
-import { bigint } from "zod";
-
-export const certificates = sqliteTable("certificates", {
-    certId: integer("certId").primaryKey({ autoIncrement: true }),
-    domain: text("domain").notNull().unique(),
-    domainId: text("domainId").references(() => domains.domainId, {
-        onDelete: "cascade"
-    }),
-    wildcard: integer("wildcard", { mode: "boolean" }).default(false),
-    status: text("status").notNull().default("pending"), // pending, requested, valid, expired, failed
-    expiresAt: integer("expiresAt"),
-    lastRenewalAttempt: integer("lastRenewalAttempt"),
-    createdAt: integer("createdAt").notNull(),
-    updatedAt: integer("updatedAt").notNull(),
-    orderId: text("orderId"),
-    errorMessage: text("errorMessage"),
-    renewalCount: integer("renewalCount").default(0),
-    certFile: text("certFile"),
-    keyFile: text("keyFile")
-});
 
 export const dnsChallenge = sqliteTable("dnsChallenges", {
     dnsChallengeId: integer("dnsChallengeId").primaryKey({
@@ -91,7 +70,8 @@ export const subscriptions = sqliteTable("subscriptions", {
     expiresAt: integer("expiresAt"),
     trial: integer("trial", { mode: "boolean" }).default(false),
     billingCycleAnchor: integer("billingCycleAnchor"),
-    type: text("type") // tier1, tier2, tier3, or license
+    type: text("type"), // tier1, tier2, tier3, or license
+    override: integer("override", { mode: "boolean" }).default(false)
 });
 
 export const subscriptionItems = sqliteTable("subscriptionItems", {
@@ -195,6 +175,44 @@ export const remoteExitNodes = sqliteTable("remoteExitNode", {
     })
 });
 
+export const remoteExitNodeResources = sqliteTable("remoteExitNodeResources", {
+    remoteExitNodeResourceId: integer("remoteExitNodeResourceId").primaryKey({
+        autoIncrement: true
+    }),
+    remoteExitNodeId: text("remoteExitNodeId")
+        .notNull()
+        .references(() => remoteExitNodes.remoteExitNodeId, {
+            onDelete: "cascade"
+        }),
+    destination: text("destination").notNull() // a cidr range
+});
+
+export const remoteExitNodePreferenceLabels = sqliteTable(
+    // this controls what sites are enforced to connect to this node
+    "remoteExitNodePreferenceLabels",
+    {
+        remoteExitNodePreferenceLabelId: integer(
+            "remoteExitNodePreferenceLabelId"
+        ).primaryKey({ autoIncrement: true }),
+        remoteExitNodeId: text("remoteExitNodeId")
+            .references(() => remoteExitNodes.remoteExitNodeId, {
+                onDelete: "cascade"
+            })
+            .notNull(),
+        labelId: integer("labelId")
+            .references(() => labels.labelId, {
+                onDelete: "cascade"
+            })
+            .notNull()
+    },
+    (t) => [
+        uniqueIndex("remote_exit_node_preference_label_uniq").on(
+            t.remoteExitNodeId,
+            t.labelId
+        )
+    ]
+);
+
 export const remoteExitNodeSessions = sqliteTable("remoteExitNodeSession", {
     sessionId: text("id").primaryKey(),
     remoteExitNodeId: text("remoteExitNodeId")
@@ -591,7 +609,6 @@ export const trialNotifications = sqliteTable("trialNotifications", {
 export type Approval = InferSelectModel<typeof approvals>;
 export type Limit = InferSelectModel<typeof limits>;
 export type Account = InferSelectModel<typeof account>;
-export type Certificate = InferSelectModel<typeof certificates>;
 export type DnsChallenge = InferSelectModel<typeof dnsChallenge>;
 export type Customer = InferSelectModel<typeof customers>;
 export type Subscription = InferSelectModel<typeof subscriptions>;

server/emails/templates/NotifyTrialExpiring.tsx

@@ -30,14 +30,14 @@ export const NotifyTrialExpiring = ({
     const isLastDay = daysRemaining === 1;
 
     const previewText = hasEnded
-        ? `Your trial for ${orgName} has ended.`
+        ? `Your cloud trial for ${orgName} has ended.`
         : isLastDay
-          ? `Your trial for ${orgName} ends tomorrow.`
-          : `Your trial for ${orgName} ends in ${daysRemaining} days.`;
+          ? `Your cloud trial for ${orgName} ends tomorrow.`
+          : `Your cloud trial for ${orgName} ends in ${daysRemaining} days.`;
 
     const heading = hasEnded
-        ? "Your Trial Ended"
-        : "Your Trial is Ending Soon";
+        ? "Your Cloud Trial Ended"
+        : "Your Cloud Trial is Ending Soon";
 
     return (
         <Html>
@@ -55,7 +55,7 @@ export const NotifyTrialExpiring = ({
                         {hasEnded ? (
                             <>
                                 <EmailText>
-                                    Your free trial for{" "}
+                                    Your cloud free trial for{" "}
                                     <strong>{orgName}</strong> ended on{" "}
                                     <strong>{trialEndsAt}</strong>. Your account
                                     has been moved to the free plan, which
@@ -64,10 +64,11 @@ export const NotifyTrialExpiring = ({
 
                                 <EmailText>
                                     Some features and resources may now be
-                                    restricted. To restore full
-                                    access and continue using all the features
-                                    you had during your trial, please upgrade to
-                                    a paid plan.
+                                    restricted. To restore full access and
+                                    continue using all the features you had
+                                    during your trial, please upgrade to a paid
+                                    plan. This does not effect any self hosted
+                                    licenses.
                                 </EmailText>
 
                                 <EmailText>
@@ -93,7 +94,8 @@ export const NotifyTrialExpiring = ({
                                 <EmailText>
                                     After your trial ends, your account will be
                                     moved to the free plan and some
-                                    functionality may be restricted.
+                                    functionality may be restricted. This does
+                                    not effect any self hosted licenses.
                                 </EmailText>
 
                                 <EmailText>

server/index.ts

@@ -1,19 +1,24 @@
 #! /usr/bin/env node
-import "./extendZod.ts";
+import "./extendZod";
 
 import { runSetupFunctions } from "./setup";
 import { createApiServer } from "./apiServer";
 import { createNextServer } from "./nextServer";
 import { createInternalServer } from "./internalServer";
+import { createAiGatewayServer } from "./aiGatewayServer";
 import { createIntegrationApiServer } from "./integrationApiServer";
 import {
     ApiKey,
     ApiKeyOrg,
+    AiBudget,
+    AiModel,
+    AiProvider,
     RemoteExitNode,
     Session,
     SiteResource,
     User,
-    UserOrg
+    UserOrg,
+    VirtualApiKey
 } from "@server/db";
 import config from "@server/lib/config";
 import { setHostMeta } from "@server/lib/hostMeta";
@@ -22,8 +27,10 @@ import { TraefikConfigManager } from "@server/lib/traefik/TraefikConfigManager";
 import { initCleanup } from "#dynamic/cleanup";
 import license from "#dynamic/license/license";
 import { initLogCleanupInterval } from "@server/lib/cleanupLogs";
-import { initAcmeCertSync } from "#dynamic/lib/acmeCertSync";
+import { initAcmeCertSync } from "@server/lib/acmeCertSync";
 import { fetchServerIp } from "@server/lib/serverIpService";
+import { startRebuildQueueProcessor } from "@server/lib/rebuildClientAssociations";
+import { initAiModelCatalog } from "@server/lib/aiModelCatalog";
 
 async function startServers() {
     await setHostMeta();
@@ -41,10 +48,13 @@ async function startServers() {
 
     initLogCleanupInterval();
     initAcmeCertSync();
+    startRebuildQueueProcessor();
+    await initAiModelCatalog();
 
     // Start all servers
     const apiServer = createApiServer();
     const internalServer = createInternalServer();
+    const aiGatewayServer = createAiGatewayServer();
 
     const nextServer = await createNextServer();
     if (config.getRawConfig().traefik.file_mode) {
@@ -63,6 +73,7 @@ async function startServers() {
         apiServer,
         nextServer,
         internalServer,
+        aiGatewayServer,
         integrationServer
     };
 }
@@ -81,6 +92,10 @@ declare global {
             userOrgIds?: string[];
             remoteExitNode?: RemoteExitNode;
             siteResource?: SiteResource;
+            aiProvider?: AiProvider;
+            aiModel?: AiModel;
+            aiBudget?: AiBudget;
+            virtualApiKey?: VirtualApiKey;
             orgPolicyAllowed?: boolean;
         }
     }

server/integrationApiServer.ts

@@ -12,7 +12,7 @@ import { logIncomingMiddleware } from "./middlewares/logIncoming";
 import helmet from "helmet";
 import swaggerUi from "swagger-ui-express";
 import { OpenApiGeneratorV3 } from "@asteasolutions/zod-to-openapi";
-import { registry } from "./openApi";
+import { registry, openApiTags } from "./openApi";
 import fs from "fs";
 import path from "path";
 import { APP_PATH } from "./lib/consts";
@@ -152,11 +152,19 @@ function getOpenApiDocumentation() {
 
             if (!hasExistingResponses) {
                 def.route.responses = {
-                    "*": {
-                        description: "",
+                    "200": {
+                        description: "Successful response",
                         content: {
                             "application/json": {
-                                schema: z.object({})
+                                schema: z.object({
+                                    data: z
+                                        .record(z.string(), z.any())
+                                        .nullable(),
+                                    success: z.boolean(),
+                                    error: z.boolean(),
+                                    message: z.string(),
+                                    status: z.number()
+                                })
                             }
                         }
                     }
@@ -173,7 +181,8 @@ function getOpenApiDocumentation() {
             version: "v1",
             title: "Pangolin Integration API"
         },
-        servers: [{ url: "/v1" }]
+        servers: [{ url: "/v1" }],
+        tags: openApiTags
     });
 
     if (!process.env.DISABLE_GEN_OPENAPI) {

server/lib/acmeCertSync.ts

@@ -1,3 +1,866 @@
+import fs from "fs";
+import path from "path";
+import crypto from "crypto";
+import {
+    certificates,
+    clients,
+    clientSiteResourcesAssociationsCache,
+    db,
+    domains,
+    newts,
+    siteNetworks,
+    SiteResource,
+    siteResources
+} from "@server/db";
+import { and, eq } from "drizzle-orm";
+import { encrypt, decrypt } from "@server/lib/crypto";
+import logger from "@server/logger";
+import config from "@server/lib/config";
+import {
+    generateSubnetProxyTargetV2,
+    SubnetProxyTargetV2
+} from "@server/lib/ip";
+import { updateTargets } from "@server/routers/client/targets";
+import cache from "#dynamic/lib/cache";
+import { build } from "@server/build";
+
+interface AcmeCert {
+    domain: { main: string; sans?: string[] };
+    certificate: string;
+    key: string;
+    Store: string;
+}
+
+interface AcmeJson {
+    [resolver: string]: {
+        Certificates: AcmeCert[];
+    };
+}
+
+export async function pushCertUpdateToAffectedNewts(
+    domain: string,
+    domainId: string | null,
+    oldCertPem: string | null,
+    oldKeyPem: string | null
+): Promise<void> {
+    // Find all SSL-enabled HTTP site resources that use this cert's domain
+    let affectedResources: SiteResource[] = [];
+
+    if (domainId) {
+        affectedResources = await db
+            .select()
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.domainId, domainId),
+                    eq(siteResources.ssl, true)
+                )
+            );
+    } else {
+        // Fallback: match by exact fullDomain when no domainId is available
+        affectedResources = await db
+            .select()
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.fullDomain, domain),
+                    eq(siteResources.ssl, true)
+                )
+            );
+    }
+
+    if (affectedResources.length === 0) {
+        logger.debug(
+            `acmeCertSync: no affected site resources for cert domain "${domain}"`
+        );
+        return;
+    }
+
+    logger.debug(
+        `acmeCertSync: pushing cert update to ${affectedResources.length} affected site resource(s) for domain "${domain}"`
+    );
+
+    for (const resource of affectedResources) {
+        try {
+            // Get all sites for this resource via siteNetworks
+            const resourceSiteRows = resource.networkId
+                ? await db
+                      .select({ siteId: siteNetworks.siteId })
+                      .from(siteNetworks)
+                      .where(eq(siteNetworks.networkId, resource.networkId))
+                : [];
+
+            if (resourceSiteRows.length === 0) {
+                logger.debug(
+                    `acmeCertSync: no sites for resource ${resource.siteResourceId}, skipping`
+                );
+                continue;
+            }
+
+            // Get all clients with access to this resource
+            const resourceClients = await db
+                .select({
+                    clientId: clients.clientId,
+                    pubKey: clients.pubKey,
+                    subnet: clients.subnet
+                })
+                .from(clients)
+                .innerJoin(
+                    clientSiteResourcesAssociationsCache,
+                    eq(
+                        clients.clientId,
+                        clientSiteResourcesAssociationsCache.clientId
+                    )
+                )
+                .where(
+                    eq(
+                        clientSiteResourcesAssociationsCache.siteResourceId,
+                        resource.siteResourceId
+                    )
+                );
+
+            if (resourceClients.length === 0) {
+                logger.debug(
+                    `acmeCertSync: no clients for resource ${resource.siteResourceId}, skipping`
+                );
+                continue;
+            }
+
+            // Invalidate the cert cache so generateSubnetProxyTargetV2 fetches fresh data
+            if (resource.fullDomain) {
+                await cache.del(`cert:${resource.fullDomain}`);
+            }
+
+            // Generate target once - same cert applies to all sites for this resource
+            const newTargets = await generateSubnetProxyTargetV2(
+                resource,
+                resourceClients
+            );
+
+            if (!newTargets) {
+                logger.debug(
+                    `acmeCertSync: could not generate target for resource ${resource.siteResourceId}, skipping`
+                );
+                continue;
+            }
+
+            // Construct the old targets - same routing shape but with the previous cert/key.
+            // The newt only uses destPrefix/sourcePrefixes for removal, but we keep the
+            // semantics correct so the update message accurately reflects what changed.
+            const oldTargets: SubnetProxyTargetV2[] = newTargets.map((t) => ({
+                ...t,
+                tlsCert: oldCertPem ?? undefined,
+                tlsKey: oldKeyPem ?? undefined
+            }));
+
+            // Push update to each site's newt
+            for (const { siteId } of resourceSiteRows) {
+                const [newt] = await db
+                    .select()
+                    .from(newts)
+                    .where(eq(newts.siteId, siteId))
+                    .limit(1);
+
+                if (!newt) {
+                    logger.debug(
+                        `acmeCertSync: no newt found for site ${siteId}, skipping resource ${resource.siteResourceId}`
+                    );
+                    continue;
+                }
+
+                await updateTargets(
+                    newt.newtId,
+                    { oldTargets: oldTargets, newTargets: newTargets },
+                    newt.version
+                );
+
+                logger.debug(
+                    `acmeCertSync: pushed cert update to newt for site ${siteId}, resource ${resource.siteResourceId}`
+                );
+            }
+        } catch (err) {
+            logger.error(
+                `acmeCertSync: error pushing cert update for resource ${resource?.siteResourceId}: ${err}`
+            );
+        }
+    }
+}
+
+async function findDomainId(certDomain: string): Promise<string | null> {
+    // Strip wildcard prefix before lookup (*.example.com -> example.com)
+    const lookupDomain = certDomain.startsWith("*.")
+        ? certDomain.slice(2)
+        : certDomain;
+
+    // 1. Exact baseDomain match (any domain type)
+    const exactMatch = await db
+        .select({ domainId: domains.domainId })
+        .from(domains)
+        .where(eq(domains.baseDomain, lookupDomain))
+        .limit(1);
+
+    if (exactMatch.length > 0) {
+        return exactMatch[0].domainId;
+    }
+
+    // 2. Walk up the domain hierarchy looking for a wildcard-type domain whose
+    //    baseDomain is a suffix of the cert domain. e.g. cert "sub.example.com"
+    //    matches a wildcard domain with baseDomain "example.com".
+    const parts = lookupDomain.split(".");
+    for (let i = 1; i < parts.length; i++) {
+        const candidate = parts.slice(i).join(".");
+        if (!candidate) continue;
+
+        const wildcardMatch = await db
+            .select({ domainId: domains.domainId })
+            .from(domains)
+            .where(
+                and(
+                    eq(domains.baseDomain, candidate),
+                    eq(domains.type, "wildcard")
+                )
+            )
+            .limit(1);
+
+        if (wildcardMatch.length > 0) {
+            return wildcardMatch[0].domainId;
+        }
+    }
+
+    return null;
+}
+
+function extractFirstCert(pemBundle: string): string | null {
+    const match = pemBundle.match(
+        /-----BEGIN CERTIFICATE-----[\s\S]+?-----END CERTIFICATE-----/
+    );
+    return match ? match[0] : null;
+}
+
+/**
+ * Determine whether an ACME cert entry represents a wildcard cert by checking
+ * both the primary domain (`main`) and the SANs. Some ACME clients (notably
+ * Traefik) store the bare apex in `main` and only put the wildcard form in
+ * `sans` (e.g. main="access.example.com", sans=["*.access.example.com"]).
+ */
+function detectWildcard(
+    main: string,
+    sans: string[] | undefined
+): { wildcard: boolean; wildcardSan: string | null } {
+    if (main.startsWith("*.")) {
+        return { wildcard: true, wildcardSan: null };
+    }
+    if (Array.isArray(sans)) {
+        for (const san of sans) {
+            if (typeof san !== "string") continue;
+            if (san === `*.${main}` || san.startsWith("*.")) {
+                return { wildcard: true, wildcardSan: san };
+            }
+        }
+    }
+    return { wildcard: false, wildcardSan: null };
+}
+
+interface HttpCert {
+    wildcard: boolean;
+    altName: string;
+    certName: string;
+    commonName: string;
+    certFile: string;
+    keyFile: string;
+}
+
+async function syncAcmeCertsFromHttp(endpoint: string): Promise<void> {
+    let response: Response;
+    try {
+        response = await fetch(endpoint);
+    } catch (err) {
+        logger.debug(
+            `acmeCertSync: could not reach HTTP endpoint ${endpoint}: ${err}`
+        );
+        return;
+    }
+
+    if (!response.ok) {
+        logger.debug(
+            `acmeCertSync: HTTP endpoint returned status ${response.status}`
+        );
+        return;
+    }
+
+    let httpCerts: HttpCert[];
+    try {
+        httpCerts = await response.json();
+    } catch (err) {
+        logger.debug(
+            `acmeCertSync: could not parse JSON from HTTP endpoint: ${err}`
+        );
+        return;
+    }
+
+    if (!Array.isArray(httpCerts) || httpCerts.length === 0) {
+        logger.debug(
+            `acmeCertSync: no certificates returned from HTTP endpoint`
+        );
+        return;
+    }
+
+    for (const cert of httpCerts) {
+        const domain = cert?.certName;
+
+        if (!domain || typeof domain !== "string") {
+            logger.debug(
+                `acmeCertSync: skipping HTTP cert with missing certName`
+            );
+            continue;
+        }
+
+        const certPem = cert.certFile;
+        const keyPem = cert.keyFile;
+
+        if (!certPem?.trim() || !keyPem?.trim()) {
+            logger.debug(
+                `acmeCertSync: skipping HTTP cert for ${domain} - empty certFile or keyFile`
+            );
+            continue;
+        }
+
+        const firstCertPemForValidation = extractFirstCert(certPem);
+        if (!firstCertPemForValidation) {
+            logger.debug(
+                `acmeCertSync: skipping HTTP cert for ${domain} - no PEM certificate block found`
+            );
+            continue;
+        }
+
+        let validatedX509: crypto.X509Certificate;
+        try {
+            validatedX509 = new crypto.X509Certificate(
+                firstCertPemForValidation
+            );
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: skipping HTTP cert for ${domain} - invalid X.509 certificate: ${err}`
+            );
+            continue;
+        }
+
+        try {
+            crypto.createPrivateKey(keyPem);
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: skipping HTTP cert for ${domain} - invalid private key: ${err}`
+            );
+            continue;
+        }
+
+        const wildcard = cert.wildcard ?? false;
+
+        const existing = await db
+            .select()
+            .from(certificates)
+            .where(eq(certificates.domain, domain))
+            .limit(1);
+
+        let oldCertPem: string | null = null;
+        let oldKeyPem: string | null = null;
+
+        if (existing.length > 0 && existing[0].certFile) {
+            try {
+                const storedCertPem = decrypt(
+                    existing[0].certFile,
+                    config.getRawConfig().server.secret!
+                );
+                const wildcardUnchanged = existing[0].wildcard === wildcard;
+                if (storedCertPem === certPem && wildcardUnchanged) {
+                    continue;
+                }
+                oldCertPem = storedCertPem;
+                if (existing[0].keyFile) {
+                    try {
+                        oldKeyPem = decrypt(
+                            existing[0].keyFile,
+                            config.getRawConfig().server.secret!
+                        );
+                    } catch (keyErr) {
+                        logger.debug(
+                            `acmeCertSync: could not decrypt stored key for ${domain}: ${keyErr}`
+                        );
+                    }
+                }
+            } catch (err) {
+                logger.debug(
+                    `acmeCertSync: could not decrypt stored cert for ${domain}, will update: ${err}`
+                );
+            }
+        }
+
+        let expiresAt: number | null = null;
+        try {
+            expiresAt = Math.floor(
+                new Date(validatedX509.validTo).getTime() / 1000
+            );
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: could not parse cert expiry for ${domain}: ${err}`
+            );
+        }
+
+        const encryptedCert = encrypt(
+            certPem,
+            config.getRawConfig().server.secret!
+        );
+        const encryptedKey = encrypt(
+            keyPem,
+            config.getRawConfig().server.secret!
+        );
+        const now = Math.floor(Date.now() / 1000);
+
+        const domainId = await findDomainId(domain);
+        if (domainId) {
+            logger.debug(
+                `acmeCertSync: resolved domainId "${domainId}" for HTTP cert domain "${domain}"`
+            );
+        } else {
+            logger.debug(
+                `acmeCertSync: no matching domain record found for HTTP cert domain "${domain}"`
+            );
+        }
+
+        if (existing.length > 0) {
+            logger.debug(
+                `acmeCertSync: updating existing certificate (HTTP) for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+            );
+            await db
+                .update(certificates)
+                .set({
+                    certFile: encryptedCert,
+                    keyFile: encryptedKey,
+                    status: "valid",
+                    expiresAt,
+                    updatedAt: now,
+                    wildcard,
+                    ...(domainId !== null && { domainId })
+                })
+                .where(eq(certificates.domain, domain));
+
+            await pushCertUpdateToAffectedNewts(
+                domain,
+                domainId,
+                oldCertPem,
+                oldKeyPem
+            );
+        } else {
+            logger.debug(
+                `acmeCertSync: inserting new certificate (HTTP) for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+            );
+            await db.insert(certificates).values({
+                domain,
+                domainId,
+                certFile: encryptedCert,
+                keyFile: encryptedKey,
+                status: "valid",
+                expiresAt,
+                createdAt: now,
+                updatedAt: now,
+                wildcard
+            });
+
+            await pushCertUpdateToAffectedNewts(domain, domainId, null, null);
+        }
+    }
+}
+
+async function storeCertForDomain(
+    domain: string,
+    certPem: string,
+    keyPem: string,
+    validatedX509: crypto.X509Certificate
+): Promise<void> {
+    const wildcard = domain.startsWith("*.");
+
+    const existing = await db
+        .select()
+        .from(certificates)
+        .where(eq(certificates.domain, domain))
+        .limit(1);
+
+    let oldCertPem: string | null = null;
+    let oldKeyPem: string | null = null;
+
+    if (existing.length > 0 && existing[0].certFile) {
+        try {
+            const storedCertPem = decrypt(
+                existing[0].certFile,
+                config.getRawConfig().server.secret!
+            );
+            const wildcardUnchanged = existing[0].wildcard === wildcard;
+            if (storedCertPem === certPem && wildcardUnchanged) {
+                return;
+            }
+            oldCertPem = storedCertPem;
+            if (existing[0].keyFile) {
+                try {
+                    oldKeyPem = decrypt(
+                        existing[0].keyFile,
+                        config.getRawConfig().server.secret!
+                    );
+                } catch (keyErr) {
+                    logger.debug(
+                        `acmeCertSync: could not decrypt stored key for ${domain}: ${keyErr}`
+                    );
+                }
+            }
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: could not decrypt stored cert for ${domain}, will update: ${err}`
+            );
+        }
+    }
+
+    let expiresAt: number | null = null;
+    try {
+        expiresAt = Math.floor(
+            new Date(validatedX509.validTo).getTime() / 1000
+        );
+    } catch (err) {
+        logger.debug(
+            `acmeCertSync: could not parse cert expiry for ${domain}: ${err}`
+        );
+    }
+
+    const encryptedCert = encrypt(
+        certPem,
+        config.getRawConfig().server.secret!
+    );
+    const encryptedKey = encrypt(keyPem, config.getRawConfig().server.secret!);
+    const now = Math.floor(Date.now() / 1000);
+
+    const domainId = await findDomainId(domain);
+    if (domainId) {
+        logger.debug(
+            `acmeCertSync: resolved domainId "${domainId}" for cert domain "${domain}"`
+        );
+    } else {
+        logger.debug(
+            `acmeCertSync: no matching domain record found for cert domain "${domain}"`
+        );
+    }
+
+    if (existing.length > 0) {
+        logger.debug(
+            `acmeCertSync: updating existing certificate for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+        );
+        await db
+            .update(certificates)
+            .set({
+                certFile: encryptedCert,
+                keyFile: encryptedKey,
+                status: "valid",
+                expiresAt,
+                updatedAt: now,
+                wildcard,
+                ...(domainId !== null && { domainId })
+            })
+            .where(eq(certificates.domain, domain));
+
+        logger.debug(
+            `acmeCertSync: updated certificate for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+        );
+
+        await pushCertUpdateToAffectedNewts(
+            domain,
+            domainId,
+            oldCertPem,
+            oldKeyPem
+        );
+    } else {
+        logger.debug(
+            `acmeCertSync: inserting new certificate for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+        );
+        await db.insert(certificates).values({
+            domain,
+            domainId,
+            certFile: encryptedCert,
+            keyFile: encryptedKey,
+            status: "valid",
+            expiresAt,
+            createdAt: now,
+            updatedAt: now,
+            wildcard
+        });
+
+        logger.debug(
+            `acmeCertSync: inserted new certificate for ${domain} (expires ${expiresAt ? new Date(expiresAt * 1000).toISOString() : "unknown"})`
+        );
+
+        await pushCertUpdateToAffectedNewts(domain, domainId, null, null);
+    }
+}
+
+function findAcmeJsonFiles(dirPath: string): string[] {
+    const results: string[] = [];
+    let entries: fs.Dirent[];
+    try {
+        entries = fs.readdirSync(dirPath, { withFileTypes: true });
+    } catch (err) {
+        logger.warn(
+            `acmeCertSync: could not read directory "${dirPath}": ${err}`
+        );
+        return results;
+    }
+    for (const entry of entries) {
+        const fullPath = path.join(dirPath, entry.name);
+        if (entry.isDirectory()) {
+            results.push(...findAcmeJsonFiles(fullPath));
+        } else if (entry.isFile()) {
+            // check if it is a json file
+            if (entry.name.endsWith(".json")) {
+                let raw: string;
+                try {
+                    raw = fs.readFileSync(fullPath, "utf8");
+                } catch (err) {
+                    logger.warn(
+                        `acmeCertSync: could not read file "${fullPath}": ${err}`
+                    );
+                    continue;
+                }
+
+                let parsed: any;
+                try {
+                    parsed = JSON.parse(raw);
+                } catch (err) {
+                    logger.warn(
+                        `acmeCertSync: could not parse "${fullPath}" as JSON: ${err}`
+                    );
+                    continue;
+                }
+            }
+
+            results.push(fullPath);
+        }
+    }
+    return results;
+}
+
+async function syncAcmeCerts(acmeJsonPath: string): Promise<void> {
+    let raw: string;
+    try {
+        raw = fs.readFileSync(acmeJsonPath, "utf8");
+    } catch (err) {
+        logger.warn(`acmeCertSync: could not read "${acmeJsonPath}": ${err}`);
+        return;
+    }
+
+    let acmeJson: AcmeJson;
+    try {
+        acmeJson = JSON.parse(raw);
+    } catch (err) {
+        logger.warn(
+            `acmeCertSync: could not parse "${acmeJsonPath}" as JSON: ${err}`
+        );
+        return;
+    }
+
+    const resolvers = Object.keys(acmeJson || {});
+    if (resolvers.length === 0) {
+        logger.debug(`acmeCertSync: no resolvers found in acme.json`);
+        return;
+    }
+
+    // Collect certificates from every resolver. If the same domain appears in
+    // multiple resolvers, the last one wins (resolvers iterated in object order).
+    const allCerts: AcmeCert[] = [];
+    for (const resolver of resolvers) {
+        const resolverData = acmeJson[resolver];
+        if (!resolverData || !Array.isArray(resolverData.Certificates)) {
+            logger.debug(
+                `acmeCertSync: no certificates found for resolver "${resolver}"`
+            );
+            continue;
+        }
+        // logger.debug(
+        //     `acmeCertSync: found ${resolverData.Certificates.length} certificate(s) for resolver "${resolver}"`
+        // );
+        for (const cert of resolverData.Certificates) {
+            allCerts.push(cert);
+        }
+    }
+
+    for (const cert of allCerts) {
+        const mainDomain = cert?.domain?.main;
+
+        if (!mainDomain || typeof mainDomain !== "string") {
+            logger.debug(`acmeCertSync: skipping cert with missing domain`);
+            continue;
+        }
+
+        if (!cert.certificate || !cert.key) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - empty certificate or key field`
+            );
+            continue;
+        }
+
+        let certPem: string;
+        let keyPem: string;
+        try {
+            certPem = Buffer.from(cert.certificate, "base64").toString("utf8");
+            keyPem = Buffer.from(cert.key, "base64").toString("utf8");
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - failed to base64-decode cert/key: ${err}`
+            );
+            continue;
+        }
+
+        if (!certPem.trim() || !keyPem.trim()) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - blank PEM after base64 decode`
+            );
+            continue;
+        }
+
+        // Validate that the decoded data actually parses as a real X.509 cert
+        // before we touch the database. This prevents importing partially-written
+        // or corrupted entries from acme.json.
+        const firstCertPemForValidation = extractFirstCert(certPem);
+        if (!firstCertPemForValidation) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - no PEM certificate block found`
+            );
+            continue;
+        }
+
+        let validatedX509: crypto.X509Certificate;
+        try {
+            validatedX509 = new crypto.X509Certificate(
+                firstCertPemForValidation
+            );
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - invalid X.509 certificate: ${err}`
+            );
+            continue;
+        }
+
+        // Sanity-check the private key parses too
+        try {
+            crypto.createPrivateKey(keyPem);
+        } catch (err) {
+            logger.debug(
+                `acmeCertSync: skipping cert for ${mainDomain} - invalid private key: ${err}`
+            );
+            continue;
+        }
+
+        // Collect all domains covered by this cert: main + every SAN.
+        // Each domain gets its own row in the certificates table so that
+        // lookups by any hostname on the cert succeed independently.
+        const allDomains = new Set<string>([mainDomain]);
+        if (Array.isArray(cert.domain?.sans)) {
+            for (const san of cert.domain.sans) {
+                if (typeof san === "string" && san.trim()) {
+                    allDomains.add(san.trim());
+                }
+            }
+        }
+
+        // logger.debug(
+        //     `acmeCertSync: cert for ${mainDomain} covers ${allDomains.size} domain(s): ${[...allDomains].join(", ")}`
+        // );
+
+        for (const domain of allDomains) {
+            try {
+                await storeCertForDomain(
+                    domain,
+                    certPem,
+                    keyPem,
+                    validatedX509
+                );
+            } catch (err) {
+                logger.error(
+                    `acmeCertSync: error storing cert for domain "${domain}": ${err}`
+                );
+            }
+        }
+    }
+}
+
 export function initAcmeCertSync(): void {
-    // stub
-}
+    if (build == "saas") {
+        logger.debug(`acmeCertSync: skipping ACME cert sync in SaaS build`);
+        return;
+    }
+
+    const configData = config.getRawConfig();
+
+    if (!configData.flags?.enable_acme_cert_sync) {
+        logger.debug(
+            `acmeCertSync: ACME cert sync is disabled by config flag, skipping`
+        );
+        return;
+    }
+
+    const acmeJsonPath =
+        configData.acme?.acme_json_path ?? "config/letsencrypt/acme.json";
+    const intervalMs = configData.acme?.sync_interval_ms ?? 5000;
+    const httpEndpoint = configData.acme?.acme_http_endpoint;
+
+    logger.debug(
+        `acmeCertSync: starting ACME cert sync from "${acmeJsonPath}" across all resolvers every ${intervalMs}ms`
+    );
+    if (httpEndpoint) {
+        logger.debug(
+            `acmeCertSync: also syncing from HTTP endpoint "${httpEndpoint}" every ${intervalMs}ms`
+        );
+    }
+
+    const runSync = () => {
+        if (httpEndpoint) {
+            syncAcmeCertsFromHttp(httpEndpoint).catch((err) => {
+                logger.error(`acmeCertSync: error during HTTP sync: ${err}`);
+            });
+        } else {
+            // only run the file-based sync if the HTTP endpoint is not configured, to avoid doubling up
+            let stat: fs.Stats | null = null;
+            try {
+                stat = fs.statSync(acmeJsonPath);
+            } catch (err) {
+                logger.warn(
+                    `acmeCertSync: cannot stat path "${acmeJsonPath}": ${err}`
+                );
+                return;
+            }
+
+            if (stat.isDirectory()) {
+                const files = findAcmeJsonFiles(acmeJsonPath);
+                if (files.length === 0) {
+                    logger.debug(
+                        `acmeCertSync: no acme.json files found in directory "${acmeJsonPath}"`
+                    );
+                    return;
+                }
+                // logger.debug(
+                //     `acmeCertSync: found ${files.length} acme.json file(s) in directory "${acmeJsonPath}"`
+                // );
+                for (const file of files) {
+                    syncAcmeCerts(file).catch((err) => {
+                        logger.error(
+                            `acmeCertSync: error during sync of "${file}": ${err}`
+                        );
+                    });
+                }
+            } else {
+                syncAcmeCerts(acmeJsonPath).catch((err) => {
+                    logger.error(`acmeCertSync: error during sync: ${err}`);
+                });
+            }
+        }
+    };
+
+    // Run immediately on init, then on the configured interval
+    runSync();
+
+    setInterval(runSync, intervalMs);
+}

server/lib/aiBudgetEnforcement.ts

@@ -0,0 +1,613 @@
+import {
+    and,
+    eq,
+    gte,
+    inArray,
+    isNull,
+    or,
+    sql,
+    SQL,
+    type InferInsertModel
+} from "drizzle-orm";
+import {
+    AiBudget,
+    aiBudgetBreachEvents,
+    aiBudgets,
+    aiModels,
+    aiUsageRecords,
+    db,
+    userOrgRoles
+} from "@server/db";
+import { modelKeyMatches } from "@server/lib/aiModelKeyMatch";
+import type { AiUsage } from "@server/lib/aiUsageExtraction";
+import { regionalCache as cache } from "#dynamic/lib/cache";
+import logger from "@server/logger";
+
+type BudgetPeriod = AiBudget["period"];
+
+const PERIOD_DURATIONS_MS: Record<Exclude<BudgetPeriod, "lifetime">, number> = {
+    hourly: 60 * 60 * 1000,
+    daily: 24 * 60 * 60 * 1000,
+    weekly: 7 * 24 * 60 * 60 * 1000,
+    monthly: 30 * 24 * 60 * 60 * 1000,
+    yearly: 365 * 24 * 60 * 60 * 1000
+};
+
+// Budgets are cheap to be a little stale about (enforcement is already
+// check-then-act, not transactional). Re-derive each budget's usage sum
+// from aiUsageRecords at most this often; in between, completed requests
+// just add their own contribution onto the cached sum instead of
+// re-querying/re-aggregating from scratch.
+const BUDGET_CACHE_REFRESH_MS = 8_000;
+// Redis-level TTL is only a safety net for eviction if a budget stops
+// seeing traffic - the actual staleness check is the computedAt timestamp
+// stored in the cached value, compared against BUDGET_CACHE_REFRESH_MS.
+const BUDGET_CACHE_SAFETY_TTL_SEC = 60;
+
+function applicableBudgetsCacheKey(ctx: BudgetScopeContext): string {
+    const roleKey = [...ctx.roleIds].sort((a, b) => a - b).join(",");
+    return [
+        "aiBudget:applicable",
+        ctx.orgId,
+        ctx.providerId,
+        ctx.requestedModel,
+        ctx.resourceId ?? "",
+        ctx.siteResourceId ?? "",
+        roleKey,
+        ctx.virtualApiKeyId ?? ""
+    ].join(":");
+}
+
+function budgetUsageCacheKey(budgetId: number): string {
+    return `aiBudget:usage:${budgetId}`;
+}
+
+type CachedBudgetUsage = {
+    sum: number;
+    computedAt: number;
+};
+
+// Budget periods are trailing windows from "now", not calendar-aligned
+// (e.g. "daily" = last 24h). "lifetime" has no lower bound.
+function windowStart(period: BudgetPeriod, now: number): number {
+    if (period === "lifetime") {
+        return 0;
+    }
+    return now - PERIOD_DURATIONS_MS[period];
+}
+
+export type BudgetScopeContext = {
+    orgId: string;
+    providerId: number;
+    requestedModel: string;
+    resourceId: number | null;
+    siteResourceId: number | null;
+    roleIds: number[];
+    requestUserId: string | null;
+    virtualApiKeyId: string | null;
+};
+
+/**
+ * Every budget that could apply to this request: the provider itself, any
+ * model on that provider whose (possibly wildcarded) modelKey matches the
+ * requested model, the target resource/site-resource, and any role the
+ * requesting user holds in the org. Cached for BUDGET_CACHE_REFRESH_MS since
+ * budget/model config changes are rare and a request-scoped org/provider/
+ * model/resource/role combination repeats constantly under real traffic.
+ */
+export async function resolveApplicableBudgets(
+    ctx: BudgetScopeContext
+): Promise<AiBudget[]> {
+    const cacheKey = applicableBudgetsCacheKey(ctx);
+    const cached = await cache.get<AiBudget[]>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const budgets = await fetchApplicableBudgets(ctx);
+    await cache.set(cacheKey, budgets, BUDGET_CACHE_REFRESH_MS / 1000);
+    return budgets;
+}
+
+async function fetchApplicableBudgets(
+    ctx: BudgetScopeContext
+): Promise<AiBudget[]> {
+    const providerModels = await db
+        .select({ modelId: aiModels.modelId, modelKey: aiModels.modelKey })
+        .from(aiModels)
+        .where(
+            and(
+                eq(aiModels.providerId, ctx.providerId),
+                eq(aiModels.enabled, true)
+            )
+        );
+
+    const matchingModelIds = providerModels
+        .filter((m) => modelKeyMatches(m.modelKey, ctx.requestedModel))
+        .map((m) => m.modelId);
+
+    const scopeConditions: SQL[] = [
+        and(
+            eq(aiBudgets.providerId, ctx.providerId),
+            isNull(aiBudgets.modelId)
+        )!
+    ];
+    if (matchingModelIds.length > 0) {
+        scopeConditions.push(inArray(aiBudgets.modelId, matchingModelIds));
+    }
+    if (ctx.resourceId != null) {
+        scopeConditions.push(eq(aiBudgets.resourceId, ctx.resourceId));
+    }
+    if (ctx.siteResourceId != null) {
+        scopeConditions.push(eq(aiBudgets.siteResourceId, ctx.siteResourceId));
+    }
+    if (ctx.roleIds.length > 0) {
+        scopeConditions.push(inArray(aiBudgets.roleId, ctx.roleIds));
+    }
+    if (ctx.virtualApiKeyId != null) {
+        scopeConditions.push(
+            eq(aiBudgets.virtualApiKeyId, ctx.virtualApiKeyId)
+        );
+    }
+
+    return db
+        .select()
+        .from(aiBudgets)
+        .where(
+            and(
+                eq(aiBudgets.orgId, ctx.orgId),
+                eq(aiBudgets.enabled, true),
+                or(...scopeConditions)
+            )
+        );
+}
+
+async function sumUsageAmount(
+    where: SQL,
+    unit: AiBudget["unit"]
+): Promise<number> {
+    const column =
+        unit === "usd" ? aiUsageRecords.costUsd : aiUsageRecords.totalTokens;
+    const [row] = await db
+        .select({ total: sql<number>`coalesce(sum(${column}), 0)` })
+        .from(aiUsageRecords)
+        .where(where);
+    return Number(row?.total ?? 0);
+}
+
+/**
+ * Sums recorded usage for a single budget's scope + rolling window. Model
+ * budgets can't be pushed down to SQL because the model's key may itself be
+ * a glob, so those rows are fetched for the provider+window and matched in
+ * JS the same way access-control matching does.
+ */
+export async function sumUsageForBudget(
+    budget: AiBudget,
+    ctx: BudgetScopeContext,
+    now: number
+): Promise<number> {
+    const start = windowStart(budget.period, now);
+
+    if (budget.modelId != null) {
+        const [model] = await db
+            .select({
+                providerId: aiModels.providerId,
+                modelKey: aiModels.modelKey
+            })
+            .from(aiModels)
+            .where(eq(aiModels.modelId, budget.modelId))
+            .limit(1);
+        if (!model) {
+            return 0;
+        }
+        const rows = await db
+            .select({
+                requestedModel: aiUsageRecords.requestedModel,
+                costUsd: aiUsageRecords.costUsd,
+                totalTokens: aiUsageRecords.totalTokens
+            })
+            .from(aiUsageRecords)
+            .where(
+                and(
+                    eq(aiUsageRecords.orgId, ctx.orgId),
+                    eq(aiUsageRecords.providerId, model.providerId),
+                    gte(aiUsageRecords.createdAt, start)
+                )
+            );
+        return rows
+            .filter((r) => modelKeyMatches(model.modelKey, r.requestedModel))
+            .reduce(
+                (sum, r) =>
+                    sum +
+                    (budget.unit === "usd" ? (r.costUsd ?? 0) : r.totalTokens),
+                0
+            );
+    }
+
+    if (budget.providerId != null) {
+        return sumUsageAmount(
+            and(
+                eq(aiUsageRecords.orgId, ctx.orgId),
+                eq(aiUsageRecords.providerId, budget.providerId),
+                gte(aiUsageRecords.createdAt, start)
+            )!,
+            budget.unit
+        );
+    }
+
+    if (budget.resourceId != null) {
+        return sumUsageAmount(
+            and(
+                eq(aiUsageRecords.orgId, ctx.orgId),
+                eq(aiUsageRecords.resourceId, budget.resourceId),
+                gte(aiUsageRecords.createdAt, start)
+            )!,
+            budget.unit
+        );
+    }
+
+    if (budget.siteResourceId != null) {
+        return sumUsageAmount(
+            and(
+                eq(aiUsageRecords.orgId, ctx.orgId),
+                eq(aiUsageRecords.siteResourceId, budget.siteResourceId),
+                gte(aiUsageRecords.createdAt, start)
+            )!,
+            budget.unit
+        );
+    }
+
+    if (budget.roleId != null) {
+        const members = await db
+            .select({ userId: userOrgRoles.userId })
+            .from(userOrgRoles)
+            .where(
+                and(
+                    eq(userOrgRoles.roleId, budget.roleId),
+                    eq(userOrgRoles.orgId, ctx.orgId)
+                )
+            );
+        const userIds = members.map((m) => m.userId);
+        if (userIds.length === 0) {
+            return 0;
+        }
+        return sumUsageAmount(
+            and(
+                eq(aiUsageRecords.orgId, ctx.orgId),
+                inArray(aiUsageRecords.userId, userIds),
+                gte(aiUsageRecords.createdAt, start)
+            )!,
+            budget.unit
+        );
+    }
+
+    if (budget.virtualApiKeyId != null) {
+        return sumUsageAmount(
+            and(
+                eq(aiUsageRecords.orgId, ctx.orgId),
+                eq(aiUsageRecords.virtualApiKeyId, budget.virtualApiKeyId),
+                gte(aiUsageRecords.createdAt, start)
+            )!,
+            budget.unit
+        );
+    }
+
+    return 0;
+}
+
+/**
+ * Cached wrapper around sumUsageForBudget. Reuses a per-budget cached sum
+ * for up to BUDGET_CACHE_REFRESH_MS, and otherwise falls through to the DB
+ * aggregation and reseeds the cache. Completed requests within that window
+ * top the cached sum up via applyUsageToBudgetCache below rather than
+ * forcing a re-aggregation on every request.
+ */
+async function getBudgetUsage(
+    budget: AiBudget,
+    ctx: BudgetScopeContext,
+    now: number
+): Promise<number> {
+    const cacheKey = budgetUsageCacheKey(budget.budgetId);
+    const cached = await cache.get<CachedBudgetUsage>(cacheKey);
+    if (cached && now - cached.computedAt < BUDGET_CACHE_REFRESH_MS) {
+        return cached.sum;
+    }
+
+    const sum = await sumUsageForBudget(budget, ctx, now);
+    await cache.set(
+        cacheKey,
+        { sum, computedAt: now } satisfies CachedBudgetUsage,
+        BUDGET_CACHE_SAFETY_TTL_SEC
+    );
+    return sum;
+}
+
+/**
+ * Called once a request's actual usage is known, for every budget that was
+ * resolved as applicable to it (i.e. checkBudgets' returned `budgets`).
+ * Adds this request's contribution directly onto each budget's cached sum
+ * so the next request in the same refresh window doesn't need to re-query
+ * or re-aggregate. If there's no warm cache entry, or it's already due for
+ * a refresh, this is a no-op - the next reader re-derives from the DB,
+ * which by then already includes this request's row via recordUsage.
+ */
+export async function applyUsageToBudgetCache(
+    budgets: AiBudget[],
+    usage: { usd: number; tokens: number }
+): Promise<void> {
+    await Promise.all(
+        budgets.map(async (budget) => {
+            const delta = budget.unit === "usd" ? usage.usd : usage.tokens;
+            if (!delta) {
+                return;
+            }
+
+            const cacheKey = budgetUsageCacheKey(budget.budgetId);
+            const cached = await cache.get<CachedBudgetUsage>(cacheKey);
+            if (
+                !cached ||
+                Date.now() - cached.computedAt >= BUDGET_CACHE_REFRESH_MS
+            ) {
+                return;
+            }
+
+            await cache.set(
+                cacheKey,
+                {
+                    sum: cached.sum + delta,
+                    computedAt: cached.computedAt
+                } satisfies CachedBudgetUsage,
+                BUDGET_CACHE_SAFETY_TTL_SEC
+            );
+        })
+    );
+}
+
+// Throttled to one durable event per budget per breach window, so a soft
+// budget being exceeded doesn't write a row on every subsequent request
+// while it stays over.
+async function recordBreachEventIfNew(
+    budget: AiBudget,
+    ctx: BudgetScopeContext,
+    usageAmount: number,
+    now: number
+): Promise<void> {
+    try {
+        const start = windowStart(budget.period, now);
+        const [existing] = await db
+            .select({ id: aiBudgetBreachEvents.id })
+            .from(aiBudgetBreachEvents)
+            .where(
+                and(
+                    eq(aiBudgetBreachEvents.budgetId, budget.budgetId),
+                    gte(aiBudgetBreachEvents.createdAt, start)
+                )
+            )
+            .limit(1);
+        if (existing) {
+            return;
+        }
+
+        await db.insert(aiBudgetBreachEvents).values({
+            orgId: ctx.orgId,
+            budgetId: budget.budgetId,
+            enforcement: budget.enforcement,
+            unit: budget.unit,
+            period: budget.period,
+            amount: budget.amount,
+            usageAmount,
+            blocked: budget.enforcement === "hard",
+            requestUserId: ctx.requestUserId,
+            createdAt: now
+        });
+    } catch (error) {
+        logger.error("Failed to record AI budget breach event", {
+            error,
+            budgetId: budget.budgetId
+        });
+    }
+}
+
+export type BudgetCheckResult = {
+    blocked: boolean;
+    blockingBudget?: AiBudget;
+    // Every budget resolved as applicable to this request, regardless of
+    // whether it was breached - pass to applyUsageToBudgetCache once this
+    // request's actual usage is known.
+    budgets: AiBudget[];
+};
+
+export async function checkBudgets(
+    ctx: BudgetScopeContext
+): Promise<BudgetCheckResult> {
+    const budgets = await resolveApplicableBudgets(ctx);
+    if (budgets.length === 0) {
+        return { blocked: false, budgets: [] };
+    }
+
+    const now = Date.now();
+    let blockingBudget: AiBudget | undefined;
+
+    for (const budget of budgets) {
+        const usage = await getBudgetUsage(budget, ctx, now);
+        if (usage < budget.amount) {
+            continue;
+        }
+
+        await recordBreachEventIfNew(budget, ctx, usage, now);
+
+        if (budget.enforcement === "hard" && !blockingBudget) {
+            blockingBudget = budget;
+        }
+    }
+
+    return blockingBudget
+        ? { blocked: true, blockingBudget, budgets }
+        : { blocked: false, budgets };
+}
+
+export type UsageRecordInput = {
+    orgId: string;
+    providerId: number;
+    resourceId: number | null;
+    siteResourceId: number | null;
+    userId: string | null;
+    virtualApiKeyId: string | null;
+    requestedModel: string;
+    usage: AiUsage;
+    costUsd: number | null;
+    createdAt?: number;
+    // Same id as the aiSessionLog row logged for this request, so the two
+    // can be joined to show token/cost usage alongside the session
+    // transcript. Undefined when the session wasn't logged (e.g. session
+    // log retention disabled for the org).
+    sessionId?: string;
+};
+
+type AiUsageRecordInsert = InferInsertModel<typeof aiUsageRecords>;
+
+// In-memory buffer for batching AI usage record inserts, mirroring the
+// approach in server/routers/badger/logRequestAudit.ts. Usage rows are read
+// back on every budget-cache miss (see getBudgetUsage above), which happens
+// at least every BUDGET_CACHE_REFRESH_MS, so this buffer is flushed much
+// more aggressively than the request audit log to keep the table from
+// lagging behind what budget enforcement needs. Unlike the audit log, there
+// is no retention/cleanup job for this table - usage history is kept
+// indefinitely for billing and historical reporting.
+const usageRecordBuffer: AiUsageRecordInsert[] = [];
+
+const USAGE_BATCH_SIZE = 20; // Write to DB every 20 records
+const USAGE_BATCH_INTERVAL_MS = 1000; // Or every 1 second, whichever comes first
+const USAGE_MAX_BUFFER_SIZE = 5000; // Prevent unbounded memory growth
+let usageFlushTimer: NodeJS.Timeout | null = null;
+let isUsageFlushInProgress = false;
+
+async function flushUsageRecords() {
+    if (usageRecordBuffer.length === 0 || isUsageFlushInProgress) {
+        return;
+    }
+
+    isUsageFlushInProgress = true;
+
+    const recordsToWrite = usageRecordBuffer.splice(
+        0,
+        usageRecordBuffer.length
+    );
+
+    try {
+        // Use a transaction to ensure all inserts succeed or fail together
+        await db.transaction(async (tx) => {
+            // Batch insert in groups to avoid overwhelming the database
+            const DB_BATCH_SIZE = 25;
+            for (let i = 0; i < recordsToWrite.length; i += DB_BATCH_SIZE) {
+                const batch = recordsToWrite.slice(i, i + DB_BATCH_SIZE);
+                await tx.insert(aiUsageRecords).values(batch);
+            }
+        });
+        logger.debug(
+            `Flushed ${recordsToWrite.length} AI usage records to database`
+        );
+    } catch (error) {
+        logger.error("Error flushing AI usage records:", error);
+        // On transaction error, put records back at the front of the buffer
+        // to retry, but only if the buffer isn't too large
+        if (
+            usageRecordBuffer.length <
+            USAGE_MAX_BUFFER_SIZE - recordsToWrite.length
+        ) {
+            usageRecordBuffer.unshift(...recordsToWrite);
+            logger.info(
+                `Re-queued ${recordsToWrite.length} AI usage records for retry`
+            );
+        } else {
+            logger.error(
+                `Buffer full, dropped ${recordsToWrite.length} AI usage records`
+            );
+        }
+    } finally {
+        isUsageFlushInProgress = false;
+        // If buffer filled up while we were flushing, flush again
+        if (usageRecordBuffer.length >= USAGE_BATCH_SIZE) {
+            flushUsageRecords().catch((err) =>
+                logger.error("Error in follow-up AI usage flush:", err)
+            );
+        }
+    }
+}
+
+function scheduleUsageFlush() {
+    if (usageFlushTimer === null) {
+        usageFlushTimer = setTimeout(() => {
+            usageFlushTimer = null;
+            flushUsageRecords().catch((err) =>
+                logger.error("Error in scheduled AI usage flush:", err)
+            );
+        }, USAGE_BATCH_INTERVAL_MS);
+    }
+}
+
+/**
+ * Gracefully flush all pending AI usage records (call this on shutdown).
+ */
+export async function shutdownUsageRecorder() {
+    if (usageFlushTimer) {
+        clearTimeout(usageFlushTimer);
+        usageFlushTimer = null;
+    }
+    // Force flush even if one is in progress by waiting and retrying
+    while (isUsageFlushInProgress) {
+        await new Promise((resolve) => setTimeout(resolve, 100));
+    }
+    await flushUsageRecords();
+}
+
+export async function recordUsage(input: UsageRecordInput): Promise<void> {
+    try {
+        const { usage } = input;
+        const totalTokens =
+            usage.promptTokens +
+            usage.cacheReadTokens +
+            usage.cacheWriteTokens +
+            usage.completionTokens +
+            usage.reasoningTokens;
+
+        // Prevent unbounded buffer growth - drop oldest entries if buffer is too large
+        if (usageRecordBuffer.length >= USAGE_MAX_BUFFER_SIZE) {
+            const dropped = usageRecordBuffer.splice(0, USAGE_BATCH_SIZE);
+            logger.warn(
+                `AI usage record buffer exceeded max size (${USAGE_MAX_BUFFER_SIZE}), dropped ${dropped.length} oldest entries`
+            );
+        }
+
+        usageRecordBuffer.push({
+            orgId: input.orgId,
+            providerId: input.providerId,
+            resourceId: input.resourceId,
+            siteResourceId: input.siteResourceId,
+            userId: input.userId,
+            virtualApiKeyId: input.virtualApiKeyId,
+            sessionId: input.sessionId,
+            requestedModel: input.requestedModel,
+            promptTokens: usage.promptTokens,
+            cacheReadTokens: usage.cacheReadTokens,
+            cacheWriteTokens: usage.cacheWriteTokens,
+            completionTokens: usage.completionTokens,
+            reasoningTokens: usage.reasoningTokens,
+            totalTokens,
+            costUsd: input.costUsd,
+            estimated: usage.estimated,
+            createdAt: input.createdAt ?? Date.now()
+        });
+
+        // Flush immediately if buffer is full, otherwise schedule a flush
+        if (usageRecordBuffer.length >= USAGE_BATCH_SIZE) {
+            flushUsageRecords().catch((err) =>
+                logger.error("Error flushing AI usage records:", err)
+            );
+        } else {
+            scheduleUsageFlush();
+        }
+    } catch (error) {
+        logger.error("Failed to record AI usage", { error });
+    }
+}

server/lib/aiCapabilities.ts

@@ -0,0 +1,339 @@
+import type { Request } from "express";
+import { AI_CAPABILITIES, type AiCapability } from "@app/lib/aiCapabilities";
+
+export { AI_CAPABILITIES, type AiCapability };
+
+export type AiCapabilityRoute = {
+    method: "POST";
+    path: string;
+};
+
+export type AiProtocolFamily = "openai" | "anthropic" | "google" | "bedrock";
+
+export type AiCapabilityDefinition = {
+    id: AiCapability;
+    protocolFamily: AiProtocolFamily;
+    routes: AiCapabilityRoute[];
+    extractModel: (req: Request) => string | undefined;
+    resolveUpstreamUrl: (
+        baseUrl: string,
+        req: Request,
+        model: string
+    ) => string;
+    isStreaming: (req: Request, contentType: string) => boolean;
+};
+
+function bodyModel(req: Request): string | undefined {
+    return typeof req.body?.model === "string" ? req.body.model : undefined;
+}
+
+function paramModel(req: Request): string | undefined {
+    const model = req.params?.model;
+    return typeof model === "string" && model.length > 0 ? model : undefined;
+}
+
+export function joinUpstreamUrl(baseUrl: string, path: string): string {
+    const base = baseUrl.replace(/\/+$/, "");
+    let suffix = path.startsWith("/") ? path : `/${path}`;
+
+    let basePathname = "/";
+    try {
+        basePathname = new URL(base).pathname.replace(/\/+$/, "") || "/";
+    } catch {
+        // Fall through with "/" non-absolute bases are not expected in
+        // production, but keep joining usable for malformed input.
+    }
+
+    if (basePathname !== "/") {
+        const baseSegs = basePathname.split("/").filter(Boolean);
+        const pathSegs = suffix.split("/").filter(Boolean);
+        const max = Math.min(baseSegs.length, pathSegs.length);
+        let overlap = 0;
+        for (let n = max; n >= 1; n--) {
+            const baseSuffix = baseSegs.slice(-n);
+            const pathPrefix = pathSegs.slice(0, n);
+            if (baseSuffix.every((seg, i) => seg === pathPrefix[i])) {
+                overlap = n;
+                break;
+            }
+        }
+        if (overlap > 0) {
+            const remaining = pathSegs.slice(overlap);
+            suffix = remaining.length > 0 ? `/${remaining.join("/")}` : "/";
+        }
+    }
+
+    if (suffix === "/") {
+        return base;
+    }
+
+    return `${base}${suffix}`;
+}
+
+function pathFromRequest(req: Request): string {
+    const raw = req.originalUrl || req.url || req.path;
+    return raw.startsWith("/") ? raw : `/${raw}`;
+}
+
+function bodyRequestsStream(req: Request): boolean {
+    return req.body?.stream === true;
+}
+
+function contentTypeIsSse(contentType: string): boolean {
+    return contentType.includes("text/event-stream");
+}
+
+function contentTypeIsAmazonEventStream(contentType: string): boolean {
+    return contentType.includes("application/vnd.amazon.eventstream");
+}
+
+function pathIncludes(req: Request, fragment: string): boolean {
+    return pathFromRequest(req).includes(fragment);
+}
+
+function isBodyOrSseStreaming(req: Request, contentType: string): boolean {
+    return bodyRequestsStream(req) || contentTypeIsSse(contentType);
+}
+
+function isGeminiStyleStreaming(req: Request, contentType: string): boolean {
+    return (
+        pathIncludes(req, "streamGenerateContent") ||
+        pathIncludes(req, "alt=sse") ||
+        contentTypeIsSse(contentType)
+    );
+}
+
+export const AI_CAPABILITY_DEFS: Record<AiCapability, AiCapabilityDefinition> =
+    {
+        openai_chat: {
+            id: "openai_chat",
+            protocolFamily: "openai",
+            routes: [
+                { method: "POST", path: "/v1/chat/completions" },
+                { method: "POST", path: "/chat/completions" }
+            ],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        openai_responses: {
+            id: "openai_responses",
+            protocolFamily: "openai",
+            routes: [{ method: "POST", path: "/v1/responses" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        anthropic_messages: {
+            id: "anthropic_messages",
+            protocolFamily: "anthropic",
+            routes: [{ method: "POST", path: "/v1/messages" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        gemini_generate_content: {
+            id: "gemini_generate_content",
+            protocolFamily: "google",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_generate_content: {
+            id: "google_generate_content",
+            protocolFamily: "google",
+            routes: [
+                {
+                    method: "POST",
+                    // Vertex publisher model generateContent
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_raw_predict: {
+            id: "google_raw_predict",
+            protocolFamily: "google",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:rawPredict"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamRawPredict"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "streamRawPredict") ||
+                pathIncludes(req, "alt=sse") ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_model_invoke: {
+            id: "bedrock_model_invoke",
+            protocolFamily: "bedrock",
+            routes: [
+                { method: "POST", path: "/model/:model/invoke" },
+                {
+                    method: "POST",
+                    path: "/model/:model/invoke-with-response-stream"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "invoke-with-response-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_converse: {
+            id: "bedrock_converse",
+            protocolFamily: "bedrock",
+            routes: [
+                { method: "POST", path: "/model/:model/converse" },
+                { method: "POST", path: "/model/:model/converse-stream" }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "converse-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        }
+    };
+
+export function isAiCapability(value: unknown): value is AiCapability {
+    return (
+        typeof value === "string" &&
+        (AI_CAPABILITIES as readonly string[]).includes(value)
+    );
+}
+
+/**
+ * Convert an Express-style route path from AI_CAPABILITY_DEFS into a RegExp.
+ * Handles `:param` segments and escaped literal colons (`\:`).
+ */
+export function routePatternToRegExp(routePath: string): RegExp {
+    let pattern = "";
+    for (let i = 0; i < routePath.length; i++) {
+        const ch = routePath[i];
+        if (
+            ch === "\\" &&
+            i + 1 < routePath.length &&
+            routePath[i + 1] === ":"
+        ) {
+            pattern += ":";
+            i++;
+            continue;
+        }
+        if (ch === ":") {
+            // Named param: consume until next / or end
+            i++;
+            while (
+                i < routePath.length &&
+                routePath[i] !== "/" &&
+                !(routePath[i] === "\\" && routePath[i + 1] === ":")
+            ) {
+                i++;
+            }
+            i--; // loop will ++
+            pattern += "[^/]+";
+            continue;
+        }
+        // Escape regex special chars
+        if (/[.*+?^${}()|[\]\\]/.test(ch)) {
+            pattern += "\\" + ch;
+        } else {
+            pattern += ch;
+        }
+    }
+    return new RegExp(`^${pattern}$`);
+}
+
+export function resolveAiCapabilityFromPath(path: string): AiCapability | null {
+    const pathname = path.split("?")[0] || "/";
+    const normalized = pathname.startsWith("/") ? pathname : `/${pathname}`;
+
+    for (const def of Object.values(AI_CAPABILITY_DEFS)) {
+        for (const route of def.routes) {
+            if (routePatternToRegExp(route.path).test(normalized)) {
+                return def.id;
+            }
+        }
+    }
+    return null;
+}
+
+export function parseCapabilities(raw: unknown): AiCapability[] {
+    if (raw == null) {
+        return [];
+    }
+
+    let parsed: unknown = raw;
+    if (typeof raw === "string") {
+        const trimmed = raw.trim();
+        if (!trimmed) {
+            return [];
+        }
+        try {
+            parsed = JSON.parse(trimmed);
+        } catch {
+            return [];
+        }
+    }
+
+    if (!Array.isArray(parsed)) {
+        return [];
+    }
+
+    const out: AiCapability[] = [];
+    const seen = new Set<AiCapability>();
+    for (const item of parsed) {
+        if (isAiCapability(item) && !seen.has(item)) {
+            seen.add(item);
+            out.push(item);
+        }
+    }
+    return out;
+}
+
+export function serializeCapabilities(capabilities: AiCapability[]): string {
+    return JSON.stringify(capabilities);
+}
+
+export function providerHasCapability(
+    capabilities: AiCapability[] | string | null | undefined,
+    capability: AiCapability
+): boolean {
+    const list =
+        typeof capabilities === "string" || capabilities == null
+            ? parseCapabilities(capabilities)
+            : capabilities;
+    return list.includes(capability);
+}

server/lib/aiGatewayAuthError.ts

@@ -0,0 +1,139 @@
+import {
+    AI_CAPABILITY_DEFS,
+    type AiCapability,
+    type AiProtocolFamily
+} from "@server/lib/aiCapabilities";
+import HttpCode from "@server/types/HttpCode";
+
+export type ClientErrorResponse = {
+    statusCode?: number;
+    contentType?: string;
+    body: string;
+};
+
+export type AiCapabilityErrorKind =
+    | "authentication"
+    | "invalid_request"
+    | "not_found"
+    | "permission"
+    | "rate_limit"
+    | "internal";
+
+const AUTH_MESSAGE = "Invalid API key provided.";
+
+type KindFields = {
+    openaiType: string;
+    openaiCode: string | null;
+    anthropicType: string;
+    googleStatus: string;
+};
+
+const KIND_FIELDS: Record<AiCapabilityErrorKind, KindFields> = {
+    authentication: {
+        openaiType: "authentication_error",
+        openaiCode: "invalid_api_key",
+        anthropicType: "authentication_error",
+        googleStatus: "UNAUTHENTICATED"
+    },
+    invalid_request: {
+        openaiType: "invalid_request_error",
+        openaiCode: null,
+        anthropicType: "invalid_request_error",
+        googleStatus: "INVALID_ARGUMENT"
+    },
+    not_found: {
+        openaiType: "invalid_request_error",
+        openaiCode: null,
+        anthropicType: "not_found_error",
+        googleStatus: "NOT_FOUND"
+    },
+    permission: {
+        openaiType: "invalid_request_error",
+        openaiCode: null,
+        anthropicType: "permission_error",
+        googleStatus: "PERMISSION_DENIED"
+    },
+    rate_limit: {
+        openaiType: "rate_limit_error",
+        openaiCode: "rate_limit_exceeded",
+        anthropicType: "rate_limit_error",
+        googleStatus: "RESOURCE_EXHAUSTED"
+    },
+    internal: {
+        openaiType: "api_error",
+        openaiCode: null,
+        anthropicType: "api_error",
+        googleStatus: "INTERNAL"
+    }
+};
+
+function resolveProtocolFamily(
+    capability: AiCapability | null
+): AiProtocolFamily {
+    if (capability == null) {
+        return "openai";
+    }
+    return AI_CAPABILITY_DEFS[capability].protocolFamily;
+}
+
+/**
+ * Build a protocol-native error body for the given capability.
+ * Message stays contextual; only the envelope/machine fields follow the
+ * capability's native API shape.
+ */
+export function buildAiCapabilityErrorBody(
+    capability: AiCapability | null,
+    kind: AiCapabilityErrorKind,
+    message: string,
+    httpStatus?: number
+): Record<string, unknown> {
+    const family = resolveProtocolFamily(capability);
+    const fields = KIND_FIELDS[kind];
+
+    switch (family) {
+        case "openai":
+            return {
+                error: {
+                    message,
+                    type: fields.openaiType,
+                    param: null,
+                    code: fields.openaiCode
+                }
+            };
+        case "anthropic":
+            return {
+                type: "error",
+                error: {
+                    type: fields.anthropicType,
+                    message
+                }
+            };
+        case "google":
+            return {
+                error: {
+                    code: httpStatus ?? HttpCode.BAD_REQUEST,
+                    message,
+                    status: fields.googleStatus
+                }
+            };
+        case "bedrock":
+            return { message };
+    }
+}
+
+export function buildInferenceAuthClientError(
+    capability: AiCapability | null
+): ClientErrorResponse {
+    return {
+        statusCode: HttpCode.UNAUTHORIZED,
+        contentType: "application/json",
+        body: JSON.stringify(
+            buildAiCapabilityErrorBody(
+                capability,
+                "authentication",
+                AUTH_MESSAGE,
+                HttpCode.UNAUTHORIZED
+            )
+        )
+    };
+}

server/lib/aiGatewayTrust.ts

@@ -0,0 +1,69 @@
+import { createHash } from "crypto";
+import config from "@server/lib/config";
+
+export const AI_GATEWAY_TRUST_HEADER = "X-Pangolin-Ai-Gateway-Auth";
+
+// Injected by the same Traefik trust middleware as AI_GATEWAY_TRUST_HEADER,
+// but its value differs per router (public inference resource vs. private
+// siteResource) so the gateway can tell which kind of resource a trusted
+// request arrived on without re-deriving it from resourceId/siteResourceId.
+export const AI_GATEWAY_RESOURCE_TYPE_HEADER =
+    "X-Pangolin-Ai-Gateway-Resource-Type";
+
+export type AiGatewayResourceType = "resource" | "site-resource";
+
+// Opt-in (server.enable_ai_gateway_client_ip_header): carries the client IP
+// that Badger resolved at the Traefik hop, so it survives an intermediary
+// proxy between Traefik and the AI gateway that overwrites
+// X-Forwarded-For/X-Real-Ip instead of appending to them. Set by a
+// disableForwardAuth Badger middleware instance (see getTraefikConfig.ts)
+// on the site-resource inference router only, since that's the sole path
+// that resolves request identity from the client IP.
+export const AI_GATEWAY_CLIENT_IP_HEADER = "X-Pangolin-Client-Ip";
+
+/**
+ * Derive a Traefik-injected trust token from the server secret.
+ * Traefik overwrites this header on inference routes so the AI gateway can
+ * trust Badger-injected Remote-* identity without re-validating credentials.
+ */
+export function deriveAiGatewayTrustToken(secret: string): string {
+    return createHash("sha256")
+        .update(`ai-gateway-trust:${secret}`)
+        .digest("hex");
+}
+
+export function getAiGatewayTrustToken(): string {
+    const secret = config.getRawConfig().server.secret;
+    if (!secret) {
+        throw new Error("Server secret is required for AI gateway trust token");
+    }
+    return deriveAiGatewayTrustToken(secret);
+}
+
+export function isAiGatewayTrustHeaderValid(
+    headers: Record<string, string | string[] | undefined> | undefined,
+    expectedToken?: string
+): boolean {
+    if (!headers) {
+        return false;
+    }
+    const expected = expectedToken ?? getAiGatewayTrustToken();
+    const raw =
+        headers[AI_GATEWAY_TRUST_HEADER] ??
+        headers[AI_GATEWAY_TRUST_HEADER.toLowerCase()];
+    const value = Array.isArray(raw) ? raw[0] : raw;
+    return typeof value === "string" && value === expected;
+}
+
+export function getAiGatewayResourceType(
+    headers: Record<string, string | string[] | undefined> | undefined
+): AiGatewayResourceType | null {
+    if (!headers) {
+        return null;
+    }
+    const raw =
+        headers[AI_GATEWAY_RESOURCE_TYPE_HEADER] ??
+        headers[AI_GATEWAY_RESOURCE_TYPE_HEADER.toLowerCase()];
+    const value = Array.isArray(raw) ? raw[0] : raw;
+    return value === "resource" || value === "site-resource" ? value : null;
+}

server/lib/aiGatewayUpstreamFetch.ts

@@ -0,0 +1,82 @@
+import http from "node:http";
+import https from "node:https";
+import { Readable } from "node:stream";
+
+type UpstreamFetchInit = {
+    method: string;
+    headers: Record<string, string>;
+    body?: string;
+    skipTlsVerification?: boolean;
+    signal?: AbortSignal;
+};
+
+const insecureHttpsAgent = new https.Agent({
+    rejectUnauthorized: false,
+    keepAlive: true
+});
+
+export function aiGatewayUpstreamFetch(
+    url: string,
+    init: UpstreamFetchInit
+): Promise<Response> {
+    const parsed = new URL(url);
+    const isHttps = parsed.protocol === "https:";
+    const lib = isHttps ? https : http;
+    const agent =
+        isHttps && init.skipTlsVerification ? insecureHttpsAgent : undefined;
+
+    return new Promise((resolve, reject) => {
+        if (init.signal?.aborted) {
+            reject(init.signal.reason ?? new Error("Request aborted"));
+            return;
+        }
+
+        const req = lib.request(
+            url,
+            {
+                method: init.method,
+                headers: init.headers,
+                agent
+            },
+            (res) => {
+                const headers = new Headers();
+                for (const [key, value] of Object.entries(res.headers)) {
+                    if (value === undefined) {
+                        continue;
+                    }
+                    if (Array.isArray(value)) {
+                        for (const entry of value) {
+                            headers.append(key, entry);
+                        }
+                    } else {
+                        headers.set(key, value);
+                    }
+                }
+
+                const body = Readable.toWeb(res) as ReadableStream<Uint8Array>;
+                resolve(
+                    new Response(body, {
+                        status: res.statusCode ?? 502,
+                        statusText: res.statusMessage,
+                        headers
+                    })
+                );
+            }
+        );
+
+        req.on("error", reject);
+
+        if (init.signal) {
+            const onAbort = () => req.destroy(init.signal!.reason);
+            init.signal.addEventListener("abort", onAbort, { once: true });
+            req.on("close", () =>
+                init.signal!.removeEventListener("abort", onAbort)
+            );
+        }
+
+        if (init.body !== undefined) {
+            req.write(init.body);
+        }
+        req.end();
+    });
+}

server/lib/aiInferenceResource.ts

@@ -0,0 +1,719 @@
+import { and, eq, inArray } from "drizzle-orm";
+import {
+    aiModels,
+    aiProviders,
+    db,
+    resourceAiModels,
+    resourceAiProviders,
+    siteResourceAiModels,
+    siteResourceAiProviders,
+    type Transaction
+} from "@server/db";
+import { z } from "zod";
+
+type DbOrTrx = Transaction | typeof db;
+
+export const modelListTypeSchema = z.enum(["allow", "block"]);
+
+export type ModelListType = z.infer<typeof modelListTypeSchema>;
+
+export const accessModeSchema = z.enum(["inherit", "select"]);
+
+export type AccessMode = z.infer<typeof accessModeSchema>;
+
+export const resourceAiProviderAttachmentSchema = z.strictObject({
+    providerId: z.number().int().positive(),
+    accessMode: accessModeSchema.optional().default("inherit"),
+    enabled: z.boolean().optional().default(true)
+});
+
+export type ResourceAiProviderInput = z.infer<
+    typeof resourceAiProviderAttachmentSchema
+>;
+
+export type ResourceAiProviderAttachment = {
+    providerId: number;
+    accessMode: AccessMode;
+    enabled: boolean;
+};
+
+export const resourceAiModelEntrySchema = z.strictObject({
+    modelId: z.number().int().positive(),
+    listType: modelListTypeSchema
+});
+
+export type ResourceAiModelEntry = z.infer<typeof resourceAiModelEntrySchema>;
+
+export type InferenceFieldsError = {
+    error: string;
+};
+
+export function isInferenceFieldsError(
+    value: { error: string } | object
+): value is InferenceFieldsError {
+    return "error" in value;
+}
+
+/**
+ * Resolve which allow/block patterns apply for an attachment.
+ * inherit → provider lists; select → resource-selected lists (replace).
+ */
+export function resolveEffectiveLists(input: {
+    accessMode: AccessMode;
+    providerAllows: string[];
+    providerBlocks: string[];
+    resourceAllows: string[];
+    resourceBlocks: string[];
+}): { allows: string[]; blocks: string[] } {
+    if (input.accessMode === "select") {
+        return {
+            allows: input.resourceAllows,
+            blocks: input.resourceBlocks
+        };
+    }
+    return {
+        allows: input.providerAllows,
+        blocks: input.providerBlocks
+    };
+}
+
+function normalizeAttachments(
+    inputs: ResourceAiProviderInput[]
+): ResourceAiProviderAttachment[] {
+    const byProviderId = new Map<
+        number,
+        { accessMode: AccessMode; enabled: boolean }
+    >();
+    for (const input of inputs) {
+        byProviderId.set(input.providerId, {
+            accessMode: input.accessMode ?? "inherit",
+            enabled: input.enabled ?? true
+        });
+    }
+    return [...byProviderId.entries()].map(
+        ([providerId, { accessMode, enabled }]) => ({
+            providerId,
+            accessMode,
+            enabled
+        })
+    );
+}
+
+/**
+ * Validate provider attachments for an org.
+ */
+export async function resolveProviderAttachments(input: {
+    orgId: string;
+    attachments: ResourceAiProviderInput[];
+    requireAtLeastOne: boolean;
+}): Promise<ResourceAiProviderAttachment[] | InferenceFieldsError> {
+    const attachments = normalizeAttachments(input.attachments);
+
+    if (input.requireAtLeastOne && attachments.length === 0) {
+        return {
+            error: "At least one AI provider is required for inference-mode resources"
+        };
+    }
+
+    if (attachments.length === 0) {
+        return [];
+    }
+
+    const providerIds = attachments.map((a) => a.providerId);
+    const providers = await db
+        .select({
+            providerId: aiProviders.providerId,
+            orgId: aiProviders.orgId,
+            enabled: aiProviders.enabled
+        })
+        .from(aiProviders)
+        .where(
+            and(
+                inArray(aiProviders.providerId, providerIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (providers.length !== providerIds.length) {
+        return {
+            error: "One or more AI providers were not found in this organization"
+        };
+    }
+
+    const disabled = providers.find((p) => !p.enabled);
+    if (disabled) {
+        return {
+            error: `AI provider with ID ${disabled.providerId} is disabled`
+        };
+    }
+
+    return attachments;
+}
+
+export async function assertInferenceModeAllowsProviderFields(input: {
+    mode: string;
+    hasProviderAttachments: boolean;
+}): Promise<InferenceFieldsError | null> {
+    if (input.mode === "inference") {
+        return null;
+    }
+    if (input.hasProviderAttachments) {
+        return {
+            error: "AI providers can only be attached to inference-mode resources"
+        };
+    }
+    return null;
+}
+
+/**
+ * Attach providers to a resource. Inherit attachments use the provider lists
+ * as-is (resource model rows for those providers are pruned). Select
+ * attachments keep resource-selected allow/block subsets.
+ */
+export async function setPublicResourceAiProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(resourceAiProviders).values(
+            attachments.map((a) => ({
+                resourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await prunePublicResourceModelsToSelectProviders(
+        resourceId,
+        attachments,
+        trx
+    );
+}
+
+export async function setSiteResourceAiProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(siteResourceAiProviders).values(
+            attachments.map((a) => ({
+                siteResourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await pruneSiteResourceModelsToSelectProviders(
+        siteResourceId,
+        attachments,
+        trx
+    );
+}
+
+/**
+ * Keep resource model rows only for providers in select mode.
+ */
+async function prunePublicResourceModelsToSelectProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(eq(resourceAiModels.resourceId, resourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: resourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(resourceAiModels)
+        .innerJoin(aiModels, eq(resourceAiModels.modelId, aiModels.modelId))
+        .where(eq(resourceAiModels.resourceId, resourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    inArray(resourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+async function pruneSiteResourceModelsToSelectProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: siteResourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(siteResourceAiModels)
+        .innerJoin(aiModels, eq(siteResourceAiModels.modelId, aiModels.modelId))
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    inArray(siteResourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+export async function clearPublicResourceAiConfig(
+    resourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiModels)
+        .where(eq(resourceAiModels.resourceId, resourceId));
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function clearSiteResourceAiConfig(
+    siteResourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiModels)
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+export async function listPublicResourceAiProviders(resourceId: number) {
+    return db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: resourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: resourceAiProviders.accessMode
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function listSiteResourceAiProviders(siteResourceId: number) {
+    return db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: siteResourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: siteResourceAiProviders.accessMode
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+export type EffectiveAllowModel = {
+    modelId: number;
+    modelKey: string;
+    name: string;
+    providerId: number;
+    providerName: string;
+};
+
+export async function listEffectiveAllowModels(options: {
+    resourceId?: number;
+    siteResourceId?: number;
+}): Promise<EffectiveAllowModel[]> {
+    if (
+        options.resourceId === undefined &&
+        options.siteResourceId === undefined
+    ) {
+        return [];
+    }
+
+    const attachments =
+        options.resourceId !== undefined
+            ? await listPublicResourceAiProviders(options.resourceId)
+            : await listSiteResourceAiProviders(options.siteResourceId!);
+
+    const activeAttachments = attachments.filter(
+        (a) => a.enabled && a.providerEnabled
+    );
+    if (activeAttachments.length === 0) {
+        return [];
+    }
+
+    const inheritProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "inherit")
+        .map((a) => a.providerId);
+    const selectProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    const providerNameById = new Map(
+        activeAttachments.map((a) => [a.providerId, a.name] as const)
+    );
+
+    const models: EffectiveAllowModel[] = [];
+
+    if (inheritProviderIds.length > 0) {
+        const rows = await db
+            .select({
+                modelId: aiModels.modelId,
+                modelKey: aiModels.modelKey,
+                name: aiModels.name,
+                providerId: aiModels.providerId
+            })
+            .from(aiModels)
+            .where(
+                and(
+                    inArray(aiModels.providerId, inheritProviderIds),
+                    eq(aiModels.enabled, true),
+                    eq(aiModels.listType, "allow")
+                )
+            );
+        for (const row of rows) {
+            models.push({
+                ...row,
+                providerName: providerNameById.get(row.providerId) ?? ""
+            });
+        }
+    }
+
+    if (selectProviderIds.length > 0) {
+        if (options.resourceId !== undefined) {
+            const rows = await db
+                .select({
+                    modelId: aiModels.modelId,
+                    modelKey: aiModels.modelKey,
+                    name: aiModels.name,
+                    providerId: aiModels.providerId
+                })
+                .from(resourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(resourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(resourceAiModels.resourceId, options.resourceId),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(resourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            for (const row of rows) {
+                models.push({
+                    ...row,
+                    providerName: providerNameById.get(row.providerId) ?? ""
+                });
+            }
+        } else if (options.siteResourceId !== undefined) {
+            const rows = await db
+                .select({
+                    modelId: aiModels.modelId,
+                    modelKey: aiModels.modelKey,
+                    name: aiModels.name,
+                    providerId: aiModels.providerId
+                })
+                .from(siteResourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(siteResourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(
+                            siteResourceAiModels.siteResourceId,
+                            options.siteResourceId
+                        ),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(siteResourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            for (const row of rows) {
+                models.push({
+                    ...row,
+                    providerName: providerNameById.get(row.providerId) ?? ""
+                });
+            }
+        }
+    }
+
+    models.sort((a, b) => {
+        const byProvider = a.providerName.localeCompare(
+            b.providerName,
+            undefined,
+            {
+                sensitivity: "base"
+            }
+        );
+        if (byProvider !== 0) {
+            return byProvider;
+        }
+        return a.name.localeCompare(b.name, undefined, { sensitivity: "base" });
+    });
+
+    return models;
+}
+
+/**
+ * Model list APIs require an inference resource with at least one select-mode
+ * attached provider.
+ */
+export async function assertPublicModelListApiEligible(resource: {
+    resourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (resource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: resourceAiProviders.providerId })
+        .from(resourceAiProviders)
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, resource.resourceId),
+                eq(resourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+export async function assertSiteModelListApiEligible(siteResource: {
+    siteResourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (siteResource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: siteResourceAiProviders.providerId })
+        .from(siteResourceAiProviders)
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    siteResource.siteResourceId
+                ),
+                eq(siteResourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+/**
+ * Resource model entries must belong to select-mode attached providers, and
+ * listType must match the provider catalog entry (allow→allow, block→block).
+ */
+export async function assertPublicResourceModelEntriesValid(input: {
+    orgId: string;
+    resourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            accessMode: resourceAiProviders.accessMode,
+            enabled: resourceAiProviders.enabled
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, input.resourceId),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "resource"
+    });
+}
+
+export async function assertSiteResourceModelEntriesValid(input: {
+    orgId: string;
+    siteResourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            accessMode: siteResourceAiProviders.accessMode,
+            enabled: siteResourceAiProviders.enabled
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    input.siteResourceId
+                ),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "site resource"
+    });
+}
+
+function dedupeModelEntries(
+    models: ResourceAiModelEntry[]
+): ResourceAiModelEntry[] {
+    const byModelId = new Map(
+        models.map((m) => [m.modelId, m.listType] as const)
+    );
+    return [...byModelId.entries()].map(([modelId, listType]) => ({
+        modelId,
+        listType
+    }));
+}
+
+async function assertModelEntriesValid(input: {
+    orgId: string;
+    modelEntries: ResourceAiModelEntry[];
+    attachments: ResourceAiProviderAttachment[];
+    resourceLabel: string;
+}): Promise<string | null> {
+    const selectProviderIds = input.attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+
+    const modelIds = input.modelEntries.map((m) => m.modelId);
+    const catalogRows = await db
+        .select({
+            modelId: aiModels.modelId,
+            listType: aiModels.listType,
+            providerId: aiModels.providerId,
+            enabled: aiModels.enabled
+        })
+        .from(aiModels)
+        .innerJoin(aiProviders, eq(aiModels.providerId, aiProviders.providerId))
+        .where(
+            and(
+                inArray(aiModels.modelId, modelIds),
+                inArray(aiModels.providerId, selectProviderIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (catalogRows.length !== modelIds.length) {
+        return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+    }
+
+    const catalogById = new Map(catalogRows.map((row) => [row.modelId, row]));
+    for (const entry of input.modelEntries) {
+        const catalog = catalogById.get(entry.modelId);
+        if (!catalog) {
+            return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+        }
+        if (catalog.listType !== entry.listType) {
+            return `Model ${entry.modelId} must use listType "${catalog.listType}" to match the provider catalog entry`;
+        }
+        if (!catalog.enabled) {
+            return `Model ${entry.modelId} is disabled on its provider`;
+        }
+    }
+
+    return null;
+}

server/lib/aiMessageNormalization.ts

@@ -0,0 +1,538 @@
+import type { AiCapability } from "@server/lib/aiCapabilities";
+import { sseDataFrames, tryParseJson } from "@server/lib/aiUsageExtraction";
+import logger from "@server/logger";
+
+// Uniform, capability-agnostic representation of a chat message, used so
+// the AI session log can be searched/displayed the same way regardless of
+// which provider/capability produced it. Content is flattened to plain text
+// - non-text parts (images, tool calls/results) are rendered as readable
+// placeholders rather than preserved as structured data, which is enough for
+// a transcript-style replay view without a per-capability renderer.
+export type NormalizedRole = "system" | "user" | "assistant" | "tool";
+
+export type NormalizedAiMessage = {
+    role: NormalizedRole;
+    content: string;
+};
+
+function normalizeRole(role: unknown): NormalizedRole {
+    if (
+        role === "system" ||
+        role === "user" ||
+        role === "assistant" ||
+        role === "tool"
+    ) {
+        return role;
+    }
+    if (role === "model") return "assistant"; // Gemini
+    if (role === "function") return "tool"; // OpenAI legacy function role
+    return "user";
+}
+
+function safeJsonStringify(value: unknown): string {
+    try {
+        return JSON.stringify(value ?? {});
+    } catch {
+        return "";
+    }
+}
+
+/**
+ * Flattens one message "part"/"block" (OpenAI content parts, Anthropic
+ * content blocks, Gemini parts, Bedrock converse content blocks - they all
+ * follow the same rough shape) into readable text.
+ */
+function flattenContentPart(part: unknown): string {
+    if (typeof part === "string") return part;
+    if (part == null || typeof part !== "object") return "";
+    const p = part as Record<string, unknown>;
+
+    if (typeof p.text === "string") return p.text;
+
+    if (
+        p.type === "image_url" ||
+        p.type === "image" ||
+        p.type === "input_image" ||
+        p.type === "output_image" ||
+        "inlineData" in p
+    ) {
+        return "[image]";
+    }
+
+    // Anthropic-style tool_use / tool_result blocks
+    if (p.type === "tool_use") {
+        const name = typeof p.name === "string" ? p.name : "tool";
+        return `[tool_call: ${name}(${safeJsonStringify(p.input)})]`;
+    }
+    if (p.type === "tool_result") {
+        const content = p.content;
+        const text =
+            typeof content === "string"
+                ? content
+                : Array.isArray(content)
+                  ? flattenContentParts(content)
+                  : "";
+        return `[tool_result: ${text}]`;
+    }
+
+    // Gemini-style functionCall / functionResponse parts
+    if (p.functionCall && typeof p.functionCall === "object") {
+        const fc = p.functionCall as Record<string, unknown>;
+        return `[tool_call: ${fc.name}(${safeJsonStringify(fc.args)})]`;
+    }
+    if (p.functionResponse && typeof p.functionResponse === "object") {
+        const fr = p.functionResponse as Record<string, unknown>;
+        return `[tool_result: ${fr.name}(${safeJsonStringify(fr.response)})]`;
+    }
+
+    // Bedrock converse-style toolUse / toolResult content blocks
+    if (p.toolUse && typeof p.toolUse === "object") {
+        const tu = p.toolUse as Record<string, unknown>;
+        return `[tool_call: ${tu.name}(${safeJsonStringify(tu.input)})]`;
+    }
+    if (p.toolResult && typeof p.toolResult === "object") {
+        const tr = p.toolResult as Record<string, unknown>;
+        const content = tr.content;
+        const text = Array.isArray(content) ? flattenContentParts(content) : "";
+        return `[tool_result: ${text}]`;
+    }
+
+    return "";
+}
+
+function flattenContentParts(parts: unknown[]): string {
+    return parts.map(flattenContentPart).join("");
+}
+
+function flattenContent(content: unknown): string {
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) return flattenContentParts(content);
+    return "";
+}
+
+/**
+ * Best-effort scan for every `"text":"..."` JSON string value in raw text,
+ * concatenated in order. Fallback for streaming formats we can't fully parse
+ * as JSON/SSE (Gemini's array-JSON stream, Bedrock's binary event-stream
+ * framing) - same spirit as aiUsageExtraction's scanNumericFields.
+ */
+function scanTextFragments(text: string): string {
+    const out: string[] = [];
+    const re = /"text"\s*:\s*"((?:[^"\\]|\\.)*)"/g;
+    let match: RegExpExecArray | null;
+    while ((match = re.exec(text)) !== null) {
+        try {
+            out.push(JSON.parse(`"${match[1]}"`));
+        } catch {
+            out.push(match[1]);
+        }
+    }
+    return out.join("");
+}
+
+// ---------------------------------------------------------------------------
+// Request (input) normalizers - operate on the already-parsed outbound body.
+// ---------------------------------------------------------------------------
+
+function normalizeOpenAiChatRequest(body: any): NormalizedAiMessage[] {
+    const messages = Array.isArray(body?.messages) ? body.messages : [];
+    return messages.map((m: any) => ({
+        role: normalizeRole(m?.role),
+        content: flattenContent(m?.content)
+    }));
+}
+
+function normalizeOpenAiResponsesRequest(body: any): NormalizedAiMessage[] {
+    const out: NormalizedAiMessage[] = [];
+    if (typeof body?.instructions === "string" && body.instructions) {
+        out.push({ role: "system", content: body.instructions });
+    }
+    const input = body?.input;
+    if (typeof input === "string") {
+        out.push({ role: "user", content: input });
+    } else if (Array.isArray(input)) {
+        for (const item of input) {
+            if (item?.role) {
+                out.push({
+                    role: normalizeRole(item.role),
+                    content: flattenContent(item.content)
+                });
+            } else if (typeof item?.type === "string") {
+                out.push({ role: "tool", content: `[${item.type}]` });
+            }
+        }
+    }
+    return out;
+}
+
+function normalizeAnthropicRequest(body: any): NormalizedAiMessage[] {
+    const out: NormalizedAiMessage[] = [];
+    if (body?.system) {
+        const sys = flattenContent(body.system);
+        if (sys) out.push({ role: "system", content: sys });
+    }
+    const messages = Array.isArray(body?.messages) ? body.messages : [];
+    for (const m of messages) {
+        out.push({
+            role: normalizeRole(m?.role),
+            content: flattenContent(m?.content)
+        });
+    }
+    return out;
+}
+
+function normalizeGeminiRequest(body: any): NormalizedAiMessage[] {
+    const out: NormalizedAiMessage[] = [];
+    const sysParts = body?.systemInstruction?.parts;
+    if (Array.isArray(sysParts)) {
+        const text = flattenContentParts(sysParts);
+        if (text) out.push({ role: "system", content: text });
+    }
+    const contents = Array.isArray(body?.contents) ? body.contents : [];
+    for (const c of contents) {
+        out.push({
+            role: normalizeRole(c?.role),
+            content: Array.isArray(c?.parts) ? flattenContentParts(c.parts) : ""
+        });
+    }
+    return out;
+}
+
+function normalizeBedrockConverseRequest(body: any): NormalizedAiMessage[] {
+    const out: NormalizedAiMessage[] = [];
+    if (Array.isArray(body?.system)) {
+        const text = flattenContentParts(body.system);
+        if (text) out.push({ role: "system", content: text });
+    }
+    const messages = Array.isArray(body?.messages) ? body.messages : [];
+    for (const m of messages) {
+        out.push({
+            role: normalizeRole(m?.role),
+            content: Array.isArray(m?.content)
+                ? flattenContentParts(m.content)
+                : ""
+        });
+    }
+    return out;
+}
+
+/**
+ * bedrock_model_invoke and google_raw_predict are passthroughs - the body
+ * shape depends entirely on the underlying model, not the capability. Try
+ * the two shapes we're most likely to see (Anthropic Claude, then plain
+ * OpenAI-style) and give up otherwise, same fallback spirit
+ * aiUsageExtraction.ts uses for these two capabilities' usage extraction.
+ */
+function normalizeBestEffortRequest(body: any): NormalizedAiMessage[] | null {
+    if (!Array.isArray(body?.messages)) return null;
+    const looksAnthropicShaped = body.messages.some((m: any) =>
+        Array.isArray(m?.content)
+    );
+    return looksAnthropicShaped
+        ? normalizeAnthropicRequest(body)
+        : normalizeOpenAiChatRequest(body);
+}
+
+// ---------------------------------------------------------------------------
+// Response (output) normalizers - operate on the raw response text, which
+// may be a single JSON document (non-streaming) or provider-framed streaming
+// text (SSE `data:` frames, a JSON-array stream, or binary event-stream
+// framing with JSON payloads embedded in it).
+// ---------------------------------------------------------------------------
+
+function normalizeOpenAiChatResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    if (isStream) {
+        let role: unknown = "assistant";
+        let content = "";
+        let found = false;
+        for (const frame of sseDataFrames(text)) {
+            const delta = tryParseJson(frame)?.choices?.[0]?.delta;
+            if (!delta) continue;
+            found = true;
+            if (typeof delta.role === "string") role = delta.role;
+            if (typeof delta.content === "string") content += delta.content;
+        }
+        return found ? [{ role: normalizeRole(role), content }] : null;
+    }
+
+    const message = tryParseJson(text)?.choices?.[0]?.message;
+    if (!message) return null;
+    return [
+        {
+            role: normalizeRole(message.role),
+            content: flattenContent(message.content)
+        }
+    ];
+}
+
+function extractOpenAiResponsesOutputText(response: any): string | null {
+    if (typeof response?.output_text === "string") return response.output_text;
+    const output = Array.isArray(response?.output) ? response.output : [];
+    const pieces: string[] = [];
+    for (const item of output) {
+        if (item?.type === "message" && Array.isArray(item.content)) {
+            pieces.push(flattenContentParts(item.content));
+        }
+    }
+    return pieces.length > 0 ? pieces.join("") : null;
+}
+
+function normalizeOpenAiResponsesResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    if (isStream) {
+        let content = "";
+        let found = false;
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (!parsed) continue;
+            if (
+                parsed.type === "response.output_text.delta" &&
+                typeof parsed.delta === "string"
+            ) {
+                content += parsed.delta;
+                found = true;
+            } else if (
+                parsed.type === "response.completed" &&
+                parsed.response
+            ) {
+                const outputText = extractOpenAiResponsesOutputText(
+                    parsed.response
+                );
+                if (outputText != null) {
+                    content = outputText;
+                    found = true;
+                }
+            }
+        }
+        return found ? [{ role: "assistant", content }] : null;
+    }
+
+    const parsed = tryParseJson(text);
+    const outputText = extractOpenAiResponsesOutputText(
+        parsed?.response ?? parsed
+    );
+    return outputText != null
+        ? [{ role: "assistant", content: outputText }]
+        : null;
+}
+
+function normalizeAnthropicResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    if (isStream) {
+        let role: unknown = "assistant";
+        let content = "";
+        let found = false;
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (!parsed) continue;
+            if (parsed.type === "message_start" && parsed.message?.role) {
+                role = parsed.message.role;
+            }
+            if (
+                parsed.type === "content_block_start" &&
+                parsed.content_block?.type === "tool_use"
+            ) {
+                const name = parsed.content_block.name ?? "tool";
+                content += `[tool_call: ${name}]`;
+                found = true;
+            }
+            if (
+                parsed.type === "content_block_delta" &&
+                typeof parsed.delta?.text === "string"
+            ) {
+                content += parsed.delta.text;
+                found = true;
+            }
+        }
+        return found ? [{ role: normalizeRole(role), content }] : null;
+    }
+
+    const parsed = tryParseJson(text);
+    if (!parsed || !Array.isArray(parsed.content)) return null;
+    return [
+        {
+            role: normalizeRole(parsed.role ?? "assistant"),
+            content: flattenContentParts(parsed.content)
+        }
+    ];
+}
+
+function geminiCandidateParts(node: any): string {
+    const parts = node?.candidates?.[0]?.content?.parts;
+    return Array.isArray(parts) ? flattenContentParts(parts) : "";
+}
+
+function normalizeGeminiResponse(
+    text: string,
+    _isStream: boolean
+): NormalizedAiMessage[] | null {
+    const frames = sseDataFrames(text);
+    let content = "";
+    let role: unknown = "model";
+    let found = false;
+
+    if (frames.length > 0) {
+        for (const frame of frames) {
+            const parsed = tryParseJson(frame);
+            const piece = geminiCandidateParts(parsed);
+            if (piece) {
+                content += piece;
+                found = true;
+            }
+            const r = parsed?.candidates?.[0]?.content?.role;
+            if (r) role = r;
+        }
+    } else {
+        const parsed = tryParseJson(text);
+        if (Array.isArray(parsed)) {
+            for (const chunk of parsed) {
+                const piece = geminiCandidateParts(chunk);
+                if (piece) {
+                    content += piece;
+                    found = true;
+                }
+                const r = chunk?.candidates?.[0]?.content?.role;
+                if (r) role = r;
+            }
+        } else if (parsed) {
+            const piece = geminiCandidateParts(parsed);
+            if (piece) {
+                content = piece;
+                found = true;
+            }
+            const r = parsed?.candidates?.[0]?.content?.role;
+            if (r) role = r;
+        }
+    }
+
+    if (!found) {
+        const scanned = scanTextFragments(text);
+        return scanned
+            ? [{ role: normalizeRole(role), content: scanned }]
+            : null;
+    }
+    return [{ role: normalizeRole(role), content }];
+}
+
+function normalizeBedrockConverseResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    if (!isStream) {
+        const message = tryParseJson(text)?.output?.message;
+        if (!message) return null;
+        return [
+            {
+                role: normalizeRole(message.role ?? "assistant"),
+                content: Array.isArray(message.content)
+                    ? flattenContentParts(message.content)
+                    : ""
+            }
+        ];
+    }
+    // converse-stream uses AWS's binary event-stream framing, but the JSON
+    // payload of each event survives intact inside it (same assumption
+    // aiUsageExtraction.ts makes for usage) - scan for the text pieces.
+    const scanned = scanTextFragments(text);
+    return scanned ? [{ role: "assistant", content: scanned }] : null;
+}
+
+function normalizeBedrockModelInvokeResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    const anthropicStyle = normalizeAnthropicResponse(text, isStream);
+    if (anthropicStyle) return anthropicStyle;
+    const scanned = scanTextFragments(text);
+    return scanned ? [{ role: "assistant", content: scanned }] : null;
+}
+
+function normalizeGoogleRawPredictResponse(
+    text: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    const anthropicStyle = normalizeAnthropicResponse(text, isStream);
+    if (anthropicStyle) return anthropicStyle;
+    const scanned = scanTextFragments(text);
+    return scanned ? [{ role: "assistant", content: scanned }] : null;
+}
+
+const REQUEST_NORMALIZERS: Record<
+    AiCapability,
+    (body: any) => NormalizedAiMessage[] | null
+> = {
+    openai_chat: normalizeOpenAiChatRequest,
+    openai_responses: normalizeOpenAiResponsesRequest,
+    anthropic_messages: normalizeAnthropicRequest,
+    gemini_generate_content: normalizeGeminiRequest,
+    google_generate_content: normalizeGeminiRequest,
+    google_raw_predict: normalizeBestEffortRequest,
+    bedrock_model_invoke: normalizeBestEffortRequest,
+    bedrock_converse: normalizeBedrockConverseRequest
+};
+
+const RESPONSE_NORMALIZERS: Record<
+    AiCapability,
+    (text: string, isStream: boolean) => NormalizedAiMessage[] | null
+> = {
+    openai_chat: normalizeOpenAiChatResponse,
+    openai_responses: normalizeOpenAiResponsesResponse,
+    anthropic_messages: normalizeAnthropicResponse,
+    gemini_generate_content: normalizeGeminiResponse,
+    google_generate_content: normalizeGeminiResponse,
+    google_raw_predict: normalizeGoogleRawPredictResponse,
+    bedrock_model_invoke: normalizeBedrockModelInvokeResponse,
+    bedrock_converse: normalizeBedrockConverseResponse
+};
+
+/**
+ * Normalizes an outbound AI gateway request body into a uniform message
+ * transcript, regardless of capability/provider. Returns null if the body
+ * doesn't contain any recognizable messages (or parsing failed) - callers
+ * should fall back to showing the raw request body.
+ */
+export function normalizeAiRequest(
+    capability: AiCapability,
+    body: unknown
+): NormalizedAiMessage[] | null {
+    try {
+        const result = REQUEST_NORMALIZERS[capability](body);
+        return result && result.length > 0 ? result : null;
+    } catch (error) {
+        logger.debug("Failed to normalize AI request messages", {
+            capability,
+            error
+        });
+        return null;
+    }
+}
+
+/**
+ * Normalizes a completed (non-streaming or fully-accumulated streaming) AI
+ * gateway response into a uniform message transcript. Returns null if
+ * nothing recognizable could be extracted - callers should fall back to
+ * showing the raw response body.
+ */
+export function normalizeAiResponse(
+    capability: AiCapability,
+    responseText: string,
+    isStream: boolean
+): NormalizedAiMessage[] | null {
+    try {
+        const result = RESPONSE_NORMALIZERS[capability](responseText, isStream);
+        return result && result.length > 0 ? result : null;
+    } catch (error) {
+        logger.debug("Failed to normalize AI response messages", {
+            capability,
+            error
+        });
+        return null;
+    }
+}

server/lib/aiModelCatalog.ts

@@ -0,0 +1,323 @@
+import fs from "node:fs";
+import axios from "axios";
+import { z } from "zod";
+import config from "@server/lib/config";
+import logger from "@server/logger";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+
+export const CATALOG_PROVIDERS = [
+    "openai",
+    "anthropic",
+    "gemini",
+    "vertex",
+    "azure",
+    "bedrock"
+] as const;
+
+export type CatalogProvider = (typeof CATALOG_PROVIDERS)[number];
+
+const CATALOG_PROVIDER_SET = new Set<string>(CATALOG_PROVIDERS);
+
+// Each of our provider types maps to at most one catalog provider. Provider
+// types that proxy arbitrary underlying models (openRouter, vercelAiGateway,
+// custom) have no mapping.
+const PROVIDER_CATALOG_MAP: Record<
+    Exclude<AiProviderType, "custom">,
+    CatalogProvider | null
+> = {
+    openai: "openai",
+    anthropic: "anthropic",
+    googleGemini: "gemini",
+    vertexAi: "vertex",
+    bedrock: "bedrock",
+    microsoftFoundry: "azure",
+    openRouter: null,
+    vercelAiGateway: null
+};
+
+export function getCatalogProviderForType(
+    type: AiProviderType
+): CatalogProvider | null {
+    if (type === "custom") {
+        return null;
+    }
+    return PROVIDER_CATALOG_MAP[type];
+}
+
+export type AiModelCatalogEntry = {
+    provider: CatalogProvider;
+    model: string;
+    pricing: {
+        in: number | null;
+        out: number | null;
+        cache: number | null;
+        reasoning: number | null;
+    };
+};
+
+const catalogEntrySchema = z.object({
+    model: z.string(),
+    provider: z.string(),
+    pricing: z
+        .object({
+            in: z.number().nullable().optional(),
+            out: z.number().nullable().optional(),
+            cache: z.number().nullable().optional(),
+            reasoning: z.number().nullable().optional()
+        })
+        .optional()
+});
+
+const catalogFileSchema = z.object({
+    data: z.array(catalogEntrySchema).optional().default([])
+});
+
+type RawCatalogEntry = z.infer<typeof catalogEntrySchema>;
+
+function normalizeCatalogProvider(raw: string): CatalogProvider | null {
+    if (CATALOG_PROVIDER_SET.has(raw)) {
+        return raw as CatalogProvider;
+    }
+    if (raw.startsWith("bedrock")) {
+        return "bedrock";
+    }
+    if (raw.startsWith("vertex")) {
+        return "vertex";
+    }
+    if (raw.startsWith("azure")) {
+        return "azure";
+    }
+    return null;
+}
+
+function normalizeEntry(raw: RawCatalogEntry): AiModelCatalogEntry | null {
+    const provider = normalizeCatalogProvider(raw.provider);
+    if (!provider) {
+        return null;
+    }
+
+    if (!raw.model) {
+        return null;
+    }
+
+    return {
+        provider,
+        model: raw.model,
+        pricing: {
+            in: raw.pricing?.in ?? null,
+            out: raw.pricing?.out ?? null,
+            cache: raw.pricing?.cache ?? null,
+            reasoning: raw.pricing?.reasoning ?? null
+        }
+    };
+}
+
+function providerKey(provider: CatalogProvider, key: string): string {
+    return `${provider}\0${key}`;
+}
+
+export class AiModelCatalog {
+    private entries: AiModelCatalogEntry[] = [];
+    private byProvider = new Map<CatalogProvider, AiModelCatalogEntry[]>();
+    private byProviderAndKey = new Map<string, AiModelCatalogEntry>();
+    private byKey = new Map<string, AiModelCatalogEntry[]>();
+    private refreshTimer: NodeJS.Timeout | null = null;
+
+    /**
+     * Loads the catalog into memory and schedules periodic background refreshes.
+     * Call once at server startup.
+     */
+    async init(): Promise<void> {
+        await this.refresh();
+        this.scheduleNextRefresh();
+    }
+
+    /** Exact lookup by catalog provider and model key. */
+    get(
+        provider: CatalogProvider,
+        key: string
+    ): AiModelCatalogEntry | undefined {
+        return this.byProviderAndKey.get(providerKey(provider, key));
+    }
+
+    /** All models for a catalog provider. */
+    list(provider: CatalogProvider): AiModelCatalogEntry[] {
+        return this.byProvider.get(provider) ?? [];
+    }
+
+    /** All catalog entries that share a model key, across providers. */
+    listByKey(key: string): AiModelCatalogEntry[] {
+        return this.byKey.get(key) ?? [];
+    }
+
+    /** Full in-memory catalog. */
+    getAll(): AiModelCatalogEntry[] {
+        return this.entries;
+    }
+
+    private setEntries(entries: AiModelCatalogEntry[]): void {
+        const byProvider = new Map<CatalogProvider, AiModelCatalogEntry[]>();
+        const byProviderAndKey = new Map<string, AiModelCatalogEntry>();
+        const byKey = new Map<string, AiModelCatalogEntry[]>();
+
+        for (const entry of entries) {
+            const list = byProvider.get(entry.provider) ?? [];
+            list.push(entry);
+            byProvider.set(entry.provider, list);
+
+            const mapKey = providerKey(entry.provider, entry.model);
+            if (!byProviderAndKey.has(mapKey)) {
+                byProviderAndKey.set(mapKey, entry);
+            }
+
+            const keyList = byKey.get(entry.model) ?? [];
+            keyList.push(entry);
+            byKey.set(entry.model, keyList);
+        }
+
+        this.entries = entries;
+        this.byProvider = byProvider;
+        this.byProviderAndKey = byProviderAndKey;
+        this.byKey = byKey;
+    }
+
+    private async fetchFromFile(
+        filePath: string
+    ): Promise<AiModelCatalogEntry[] | null> {
+        try {
+            if (!fs.existsSync(filePath)) {
+                logger.warn(
+                    `AI model catalog file not found at ${filePath}; cost calculation will fall back to unknown pricing`
+                );
+                return null;
+            }
+            const raw = fs.readFileSync(filePath, "utf-8");
+            const result = catalogFileSchema.safeParse(JSON.parse(raw));
+            if (!result.success) {
+                logger.warn(
+                    `AI model catalog file at ${filePath} failed validation: ${result.error.message}`
+                );
+                return null;
+            }
+            return result.data.data
+                .map(normalizeEntry)
+                .filter((e): e is AiModelCatalogEntry => e != null);
+        } catch (error) {
+            logger.warn("Failed to read AI model catalog file", { error });
+            return null;
+        }
+    }
+
+    private async fetchFromUpstream(
+        upstreamUrl: string
+    ): Promise<AiModelCatalogEntry[] | null> {
+        try {
+            const res = await axios.get(upstreamUrl, { timeout: 15_000 });
+            const result = catalogFileSchema.safeParse(res.data);
+            if (!result.success) {
+                logger.warn(
+                    `AI model catalog response from ${upstreamUrl} failed validation: ${result.error.message}`
+                );
+                return null;
+            }
+            return result.data.data
+                .map(normalizeEntry)
+                .filter((e): e is AiModelCatalogEntry => e != null);
+        } catch (error: any) {
+            logger.warn(
+                `Failed to fetch AI model catalog from ${upstreamUrl}: ${error.message || error}`
+            );
+            return null;
+        }
+    }
+
+    private async refresh(): Promise<void> {
+        const { file, merge_file, upstream_url } =
+            config.getRawConfig().ai.model_catalog;
+
+        const fetched = file
+            ? await this.fetchFromFile(file)
+            : await this.fetchFromUpstream(upstream_url);
+
+        if (!fetched) {
+            logger.debug(
+                "AI model catalog refresh failed; keeping previously loaded catalog in memory"
+            );
+            return;
+        }
+
+        let merged = fetched;
+        if (merge_file) {
+            const mergeEntries = await this.fetchFromFile(merge_file);
+            if (mergeEntries) {
+                // Entries from the base catalog take precedence; the merge
+                // file only adds models not already present.
+                merged = [...fetched, ...mergeEntries];
+            }
+        }
+
+        this.setEntries(merged);
+        logger.debug(
+            `AI model catalog refreshed: ${this.entries.length} models loaded`
+        );
+    }
+
+    private scheduleNextRefresh(): void {
+        const { refresh_interval_min_hours, refresh_interval_max_hours } =
+            config.getRawConfig().ai.model_catalog;
+
+        // Jittered rather than fixed so that many self-hosted instances don't
+        // all hit the upstream catalog endpoint at the same moment.
+        const minMs = refresh_interval_min_hours * 60 * 60 * 1000;
+        const maxMs = refresh_interval_max_hours * 60 * 60 * 1000;
+        const delayMs = minMs + Math.random() * Math.max(0, maxMs - minMs);
+
+        if (this.refreshTimer) {
+            clearTimeout(this.refreshTimer);
+        }
+        this.refreshTimer = setTimeout(async () => {
+            await this.refresh();
+            this.scheduleNextRefresh();
+        }, delayMs);
+    }
+}
+
+export const aiModelCatalog = new AiModelCatalog();
+
+export function listCatalogModelsForType(
+    type: AiProviderType,
+    query?: string
+): { model: string }[] {
+    const catalogProvider = getCatalogProviderForType(type);
+
+    let models = catalogProvider
+        ? aiModelCatalog.list(catalogProvider).map((entry) => ({
+              model: entry.model
+          }))
+        : [];
+
+    if (query) {
+        const q = query.toLowerCase();
+        models = models.filter((m) => m.model.toLowerCase().includes(q));
+    }
+
+    const seen = new Set<string>();
+    models = models.filter((m) => {
+        if (seen.has(m.model)) {
+            return false;
+        }
+        seen.add(m.model);
+        return true;
+    });
+
+    models.sort((a, b) => a.model.localeCompare(b.model));
+    return models;
+}
+
+/**
+ * Loads the AI model pricing catalog into memory and schedules periodic
+ * background refreshes. Call once at server startup.
+ */
+export async function initAiModelCatalog(): Promise<void> {
+    await aiModelCatalog.init();
+}

server/lib/aiModelKeyMatch.ts

@@ -0,0 +1,101 @@
+const modelKeyRegexCache = new Map<string, RegExp>();
+
+export function isModelKeyPattern(key: string): boolean {
+    return key.includes("*") || key.includes("?");
+}
+
+function getModelKeyRegex(pattern: string): RegExp {
+    let regex = modelKeyRegexCache.get(pattern);
+    if (!regex) {
+        const escaped = pattern.replace(/[.+^${}()|[\]\\]/g, "\\$&");
+        regex = new RegExp(
+            `^${escaped.replace(/\*/g, ".*").replace(/\?/g, ".")}$`
+        );
+        modelKeyRegexCache.set(pattern, regex);
+    }
+    return regex;
+}
+
+export function modelKeyMatches(
+    pattern: string,
+    requestedModel: string
+): boolean {
+    return getModelKeyRegex(pattern).test(requestedModel);
+}
+
+function wildcardCharCount(key: string): number {
+    let count = 0;
+    for (const char of key) {
+        if (char === "*" || char === "?") {
+            count += 1;
+        }
+    }
+    return count;
+}
+
+function literalLength(key: string): number {
+    return key.replace(/[*?]/g, "").length;
+}
+
+/**
+ * Sort comparator: more specific patterns sort before less specific ones
+ * (negative when `a` is more specific than `b`).
+ *
+ * 1. Exact keys beat patterns
+ * 2. Fewer wildcard characters win
+ * 3. Longer literal length wins
+ */
+export function compareModelKeySpecificity(a: string, b: string): number {
+    const aIsPattern = isModelKeyPattern(a);
+    const bIsPattern = isModelKeyPattern(b);
+
+    if (aIsPattern !== bIsPattern) {
+        return aIsPattern ? 1 : -1;
+    }
+
+    const wildcardDiff = wildcardCharCount(a) - wildcardCharCount(b);
+    if (wildcardDiff !== 0) {
+        return wildcardDiff;
+    }
+
+    return literalLength(b) - literalLength(a);
+}
+
+/**
+ * Provider-layer policy: empty allowlist denies all. Blocklist only applies
+ * after an allow match.
+ */
+export function isAllowedByLists(
+    requested: string,
+    allows: string[],
+    blocks: string[]
+): boolean {
+    if (allows.length === 0) {
+        return false;
+    }
+    if (!allows.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    if (blocks.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    return true;
+}
+
+/**
+ * Among allow patterns that match `requested`, return the most specific one,
+ * or null if none match.
+ */
+export function mostSpecificMatchingAllow(
+    requested: string,
+    allows: string[]
+): string | null {
+    const matching = allows.filter((pattern) =>
+        modelKeyMatches(pattern, requested)
+    );
+    if (matching.length === 0) {
+        return null;
+    }
+    matching.sort(compareModelKeySpecificity);
+    return matching[0];
+}

server/lib/aiModelPricing.ts

@@ -0,0 +1,148 @@
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+import type { AiUsage } from "@server/lib/aiUsageExtraction";
+import {
+    aiModelCatalog,
+    getCatalogProviderForType,
+    type AiModelCatalogEntry,
+    type CatalogProvider
+} from "@server/lib/aiModelCatalog";
+
+export type AiModelPricing = {
+    inputCostPerToken: number | null;
+    outputCostPerToken: number | null;
+    cacheReadInputTokenCost: number | null;
+    outputCostPerReasoningToken: number | null;
+    // True when the match came from a different catalog provider than the
+    // one mapped to this provider's type (e.g. an openRouter/custom model
+    // id that only matched a global search across every provider). Costs
+    // found this way are a best-effort approximation, not a guarantee the
+    // upstream provider bills at the same rate.
+    approximate: boolean;
+};
+
+function stripVendorPrefix(modelId: string): string | null {
+    const idx = modelId.indexOf("/");
+    if (idx === -1 || idx === modelId.length - 1) {
+        return null;
+    }
+    return modelId.slice(idx + 1);
+}
+
+function toPricing(
+    entry: AiModelCatalogEntry,
+    approximate: boolean
+): AiModelPricing {
+    return {
+        inputCostPerToken: entry.pricing.in,
+        outputCostPerToken: entry.pricing.out,
+        cacheReadInputTokenCost: entry.pricing.cache,
+        outputCostPerReasoningToken: entry.pricing.reasoning,
+        approximate
+    };
+}
+
+function findEntry(
+    modelId: string,
+    provider: CatalogProvider | null
+): AiModelCatalogEntry | null {
+    const candidates = [modelId, stripVendorPrefix(modelId)].filter(
+        (v): v is string => v != null
+    );
+
+    for (const key of candidates) {
+        if (provider) {
+            const match = aiModelCatalog.get(provider, key);
+            if (match) {
+                return match;
+            }
+            continue;
+        }
+
+        const match = aiModelCatalog.listByKey(key)[0];
+        if (match) {
+            return match;
+        }
+    }
+    return null;
+}
+
+/**
+ * Looks up per-token pricing for a model, scoped first to the catalog
+ * provider that corresponds to our provider type, then falling back to a
+ * global search across every provider (marked `approximate`) for provider
+ * types that proxy arbitrary underlying models.
+ */
+export function getModelPricing(
+    providerType: AiProviderType,
+    modelId: string | undefined
+): AiModelPricing | null {
+    if (!modelId) {
+        return null;
+    }
+
+    const catalogProvider = getCatalogProviderForType(providerType);
+
+    if (catalogProvider) {
+        const scoped = findEntry(modelId, catalogProvider);
+        if (scoped) {
+            return toPricing(scoped, false);
+        }
+    }
+
+    const fallback = findEntry(modelId, null);
+    if (fallback) {
+        return toPricing(fallback, true);
+    }
+
+    return null;
+}
+
+export type AiCostBreakdown = {
+    promptCost: number;
+    cacheReadCost: number;
+    cacheWriteCost: number;
+    completionCost: number;
+    reasoningCost: number;
+    totalCost: number;
+};
+
+/**
+ * Computes a $ cost breakdown for a usage record given a model's pricing.
+ * Cache writes and reasoning tokens fall back to the normal input/output
+ * rate respectively when the catalog has no dedicated rate for them (the
+ * catalog has no cache-write field at all, and only some models report a
+ * distinct reasoning rate).
+ */
+export function calculateAiCost(
+    pricing: AiModelPricing | null,
+    usage: AiUsage
+): AiCostBreakdown | null {
+    if (!pricing) {
+        return null;
+    }
+
+    const inputRate = pricing.inputCostPerToken ?? 0;
+    const outputRate = pricing.outputCostPerToken ?? 0;
+    const cacheReadRate = pricing.cacheReadInputTokenCost ?? inputRate;
+    const reasoningRate = pricing.outputCostPerReasoningToken ?? outputRate;
+
+    const promptCost = usage.promptTokens * inputRate;
+    const cacheReadCost = usage.cacheReadTokens * cacheReadRate;
+    const cacheWriteCost = usage.cacheWriteTokens * inputRate;
+    const completionCost = usage.completionTokens * outputRate;
+    const reasoningCost = usage.reasoningTokens * reasoningRate;
+
+    return {
+        promptCost,
+        cacheReadCost,
+        cacheWriteCost,
+        completionCost,
+        reasoningCost,
+        totalCost:
+            promptCost +
+            cacheReadCost +
+            cacheWriteCost +
+            completionCost +
+            reasoningCost
+    };
+}

server/lib/aiProviderDefaults.ts

@@ -0,0 +1,186 @@
+import { decrypt, encrypt } from "@server/lib/crypto";
+import {
+    parseCapabilities,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { stripVirtualApiKeyAuthHeaders } from "@app/lib/virtualApiKeyFormat";
+import {
+    AI_PROVIDER_AUTH_TYPES,
+    AI_PROVIDER_DEFAULTS,
+    authTypeRequiresApiKey,
+    defaultsForProviderType,
+    providerRequiresUpstreamUrl,
+    type AiBudgetUnit,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@app/lib/aiProviderDefaults";
+
+export {
+    AI_PROVIDER_AUTH_TYPES,
+    AI_PROVIDER_DEFAULTS,
+    authTypeRequiresApiKey,
+    defaultsForProviderType,
+    providerRequiresUpstreamUrl,
+    type AiBudgetUnit,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+};
+
+const CONFLICTING_AUTH_HEADERS = [
+    "authorization",
+    "x-api-key",
+    "x-goog-api-key",
+    "cf-aig-authorization"
+] as const;
+
+export function resolveAiProviderCreateFields(input: {
+    type: AiProviderType;
+    upstreamUrl?: string | null;
+    authType?: AiProviderAuthType | null;
+    routingMode?: AiProviderRoutingMode | null;
+}): {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+    routingMode: AiProviderRoutingMode;
+} {
+    const routingMode =
+        input.type === "custom" ? (input.routingMode ?? "url") : "url";
+
+    if (routingMode === "target") {
+        return {
+            upstreamUrl: null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    if (input.type === "custom") {
+        return {
+            upstreamUrl: input.upstreamUrl ?? null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    const defaults = AI_PROVIDER_DEFAULTS[input.type];
+    return {
+        upstreamUrl: input.upstreamUrl ?? defaults.upstreamUrl,
+        authType: input.authType ?? defaults.authType,
+        routingMode
+    };
+}
+
+export type AiProviderHeader = { name: string; value: string };
+
+export function serializeAiProviderHeaders(
+    headers: AiProviderHeader[] | null | undefined,
+    secret: string
+): string | null {
+    if (!headers || headers.length === 0) {
+        return null;
+    }
+    return encrypt(JSON.stringify(headers), secret);
+}
+
+export function parseAiProviderHeaders(
+    raw: string | null | undefined,
+    secret: string
+): AiProviderHeader[] {
+    if (!raw) {
+        return [];
+    }
+    try {
+        const decrypted = decrypt(raw, secret);
+        const parsed = JSON.parse(decrypted);
+        if (!Array.isArray(parsed)) {
+            return [];
+        }
+        return parsed.filter(
+            (h): h is AiProviderHeader =>
+                h != null &&
+                typeof h === "object" &&
+                typeof h.name === "string" &&
+                typeof h.value === "string"
+        );
+    } catch {
+        return [];
+    }
+}
+
+export function applyAiProviderCustomHeaders(
+    headers: Record<string, string>,
+    raw: string | null | undefined,
+    secret: string
+): void {
+    for (const { name, value } of parseAiProviderHeaders(raw, secret)) {
+        headers[name] = value;
+    }
+}
+
+/**
+ * Apply provider auth to upstream headers.
+ * - Always strips Pangolin virtual API key credentials from client auth headers.
+ * - Injected modes: strip conflicting client auth headers, then set the provider key.
+ * - none: strip conflicting client auth headers, send no auth.
+ * - passthrough: leave remaining client auth headers as-is (after VAK strip).
+ */
+export function applyAiProviderAuthHeaders(
+    headers: Record<string, string>,
+    authType: AiProviderAuthType,
+    apiKey: string | null
+): void {
+    stripVirtualApiKeyAuthHeaders(headers);
+
+    if (authType === "passthrough") {
+        return;
+    }
+
+    for (const name of CONFLICTING_AUTH_HEADERS) {
+        for (const key of Object.keys(headers)) {
+            if (key.toLowerCase() === name) {
+                delete headers[key];
+            }
+        }
+    }
+
+    if (authType === "none") {
+        return;
+    }
+
+    if (!apiKey) {
+        throw new Error(`API key required for authType ${authType}`);
+    }
+
+    switch (authType) {
+        case "bearer":
+            headers["Authorization"] = `Bearer ${apiKey}`;
+            break;
+        case "x-api-key":
+            headers["x-api-key"] = apiKey;
+            break;
+        case "x-goog-api-key":
+            headers["x-goog-api-key"] = apiKey;
+            break;
+        case "hec":
+            headers["Authorization"] = `Splunk ${apiKey}`;
+            break;
+        case "cf-aig-authorization":
+            headers["cf-aig-authorization"] = `Bearer ${apiKey}`;
+            break;
+    }
+}
+
+export function resolveCapabilitiesForCreate(input: {
+    type: AiProviderType;
+    capabilities?: AiCapability[] | null;
+}): AiCapability[] {
+    if (input.capabilities != null) {
+        return parseCapabilities(input.capabilities);
+    }
+    if (input.type === "custom") {
+        return [];
+    }
+    return [...AI_PROVIDER_DEFAULTS[input.type].capabilities];
+}

server/lib/aiProviderSelection.ts

@@ -0,0 +1,97 @@
+import {
+    aiModelCatalog,
+    getCatalogProviderForType,
+    type CatalogProvider
+} from "@server/lib/aiModelCatalog";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+
+function stripVendorPrefix(modelId: string): string | null {
+    const idx = modelId.indexOf("/");
+    if (idx === -1 || idx === modelId.length - 1) {
+        return null;
+    }
+    return modelId.slice(idx + 1);
+}
+
+function modelKeysToTry(modelId: string): string[] {
+    const keys = [modelId];
+    const stripped = stripVendorPrefix(modelId);
+    if (stripped) {
+        keys.push(stripped);
+    }
+    return keys;
+}
+
+function catalogOwnsModel(
+    catalogProvider: CatalogProvider,
+    modelId: string
+): boolean {
+    for (const key of modelKeysToTry(modelId)) {
+        if (aiModelCatalog.get(catalogProvider, key)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+function modelKnownInAnyCatalog(modelId: string): boolean {
+    for (const key of modelKeysToTry(modelId)) {
+        if (aiModelCatalog.listByKey(key).length > 0) {
+            return true;
+        }
+    }
+    return false;
+}
+
+/**
+ * How strongly a provider "owns" a requested model id via the known catalog.
+ *
+ * 2 - Typed provider whose catalog contains the model
+ * 1 - Aggregator/custom that can proxy a catalog-known model
+ * 0 - No ownership signal (typed miss, or unknown model on aggregator/custom)
+ */
+export function catalogOwnershipScore(
+    type: AiProviderType,
+    modelId: string
+): number {
+    const catalogProvider = getCatalogProviderForType(type);
+    if (catalogProvider != null) {
+        return catalogOwnsModel(catalogProvider, modelId) ? 2 : 0;
+    }
+    return modelKnownInAnyCatalog(modelId) ? 1 : 0;
+}
+
+/**
+ * Prefer native vendor providers over aggregators over custom when catalog
+ * ownership is tied.
+ *
+ * 2 - Native typed provider (openai, anthropic, gemini, ...)
+ * 1 - Aggregator gateway (openRouter, vercelAiGateway)
+ * 0 - Custom
+ */
+export function providerClassRank(type: AiProviderType): number {
+    if (type === "custom") {
+        return 0;
+    }
+    if (type === "openRouter" || type === "vercelAiGateway") {
+        return 1;
+    }
+    return 2;
+}
+
+export function keepBestScored<T>(
+    items: T[],
+    scoreFn: (item: T) => number
+): T[] {
+    if (items.length <= 1) {
+        return items;
+    }
+    let best = Number.NEGATIVE_INFINITY;
+    for (const item of items) {
+        const score = scoreFn(item);
+        if (score > best) {
+            best = score;
+        }
+    }
+    return items.filter((item) => scoreFn(item) === best);
+}

server/lib/aiUsageExtraction.ts

@@ -0,0 +1,482 @@
+import { encode } from "gpt-tokenizer";
+import type { AiCapability } from "@server/lib/aiCapabilities";
+import logger from "@server/logger";
+
+export type AiUsage = {
+    // Input tokens billed at the normal input rate (i.e. NOT already
+    // covered by cacheReadTokens/cacheWriteTokens below).
+    promptTokens: number;
+    cacheReadTokens: number;
+    cacheWriteTokens: number;
+    // Output tokens billed at the normal output rate (i.e. NOT already
+    // covered by reasoningTokens below).
+    completionTokens: number;
+    reasoningTokens: number;
+    // True when these numbers are our own best-guess estimate (the upstream
+    // response didn't report usage), rather than provider-reported figures.
+    estimated: boolean;
+};
+
+function emptyUsage(): AiUsage {
+    return {
+        promptTokens: 0,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+/**
+ * Scans raw (possibly binary-framed, e.g. Bedrock's vnd.amazon.eventstream)
+ * text for `"fieldName":123` occurrences and returns the last value seen for
+ * each field. Used as a best-effort fallback for response shapes we can't
+ * fully parse as JSON/SSE (streaming Bedrock, raw predict passthroughs).
+ */
+function scanNumericFields(
+    text: string,
+    fields: string[]
+): Record<string, number> {
+    const out: Record<string, number> = {};
+    for (const field of fields) {
+        const re = new RegExp(`"${field}"\\s*:\\s*(\\d+)`, "g");
+        let match: RegExpExecArray | null;
+        while ((match = re.exec(text)) !== null) {
+            out[field] = Number(match[1]);
+        }
+    }
+    return out;
+}
+
+// Exported for reuse by server/lib/aiMessageNormalization.ts, which needs
+// the same SSE-frame/JSON-parsing groundwork to extract message content
+// instead of usage numbers.
+export function sseDataFrames(text: string): string[] {
+    const frames: string[] = [];
+    for (const rawFrame of text.split(/\r?\n\r?\n/)) {
+        for (const line of rawFrame.split(/\r?\n/)) {
+            if (!line.startsWith("data:")) continue;
+            const data = line.slice("data:".length).trim();
+            if (data && data !== "[DONE]") {
+                frames.push(data);
+            }
+        }
+    }
+    return frames;
+}
+
+export function tryParseJson(text: string): any | null {
+    try {
+        return JSON.parse(text);
+    } catch {
+        return null;
+    }
+}
+
+function extractOpenAiChat(text: string, isStream: boolean): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        usage = tryParseJson(text)?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.prompt_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens =
+        usage.completion_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.prompt_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.completion_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractOpenAiResponses(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (
+                parsed?.type === "response.completed" &&
+                parsed?.response?.usage
+            ) {
+                usage = parsed.response.usage;
+            } else if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        const parsed = tryParseJson(text);
+        usage = parsed?.usage ?? parsed?.response?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens = usage.output_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.input_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.output_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractAnthropicMessages(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let inputTokens = 0;
+    let cacheReadTokens = 0;
+    let cacheWriteTokens = 0;
+    let outputTokens = 0;
+    let found = false;
+
+    const applyUsage = (usage: any) => {
+        if (!usage) return;
+        found = true;
+        if (typeof usage.input_tokens === "number") {
+            inputTokens = usage.input_tokens;
+        }
+        if (typeof usage.cache_read_input_tokens === "number") {
+            cacheReadTokens = usage.cache_read_input_tokens;
+        }
+        if (typeof usage.cache_creation_input_tokens === "number") {
+            cacheWriteTokens = usage.cache_creation_input_tokens;
+        }
+        if (typeof usage.output_tokens === "number") {
+            outputTokens = usage.output_tokens;
+        }
+    };
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (!parsed) continue;
+            applyUsage(parsed.message?.usage);
+            applyUsage(parsed.usage);
+        }
+    } else {
+        applyUsage(tryParseJson(text)?.usage);
+    }
+
+    if (!found) {
+        return null;
+    }
+
+    return {
+        promptTokens: inputTokens,
+        cacheReadTokens,
+        cacheWriteTokens,
+        completionTokens: outputTokens,
+        // Anthropic bills extended-thinking output at the normal output
+        // rate, so there's no separate reasoning bucket to report.
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractGoogleGenerateContent(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Both the plain-JSON-array stream format and the SSE (?alt=sse) format
+    // repeat a cumulative `usageMetadata` object per chunk; the regex scan
+    // below naturally picks up the last (most complete) one either way.
+    const fields = scanNumericFields(text, [
+        "promptTokenCount",
+        "candidatesTokenCount",
+        "cachedContentTokenCount",
+        "thoughtsTokenCount"
+    ]);
+
+    if (fields.promptTokenCount === undefined) {
+        return null;
+    }
+
+    const cacheReadTokens = fields.cachedContentTokenCount ?? 0;
+    const reasoningTokens = fields.thoughtsTokenCount ?? 0;
+
+    return {
+        promptTokens: Math.max(0, fields.promptTokenCount - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: fields.candidatesTokenCount ?? 0,
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractBedrockConverse(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Non-streaming responses are plain JSON; converse-stream frames the
+    // final `metadata` event's usage object inside binary event-stream
+    // framing, but the JSON text survives intact inside that binary
+    // envelope, so the same field scan works for both.
+    const parsed = tryParseJson(text);
+    const usage = parsed?.usage;
+    if (usage) {
+        const cacheReadTokens = usage.cacheReadInputTokens ?? 0;
+        return {
+            promptTokens: Math.max(
+                0,
+                (usage.inputTokens ?? 0) - cacheReadTokens
+            ),
+            cacheReadTokens,
+            cacheWriteTokens: usage.cacheWriteInputTokens ?? 0,
+            completionTokens: usage.outputTokens ?? 0,
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokens",
+        "outputTokens",
+        "cacheReadInputTokens",
+        "cacheWriteInputTokens"
+    ]);
+    if (fields.inputTokens === undefined) {
+        return null;
+    }
+    const cacheReadTokens = fields.cacheReadInputTokens ?? 0;
+    return {
+        promptTokens: Math.max(0, fields.inputTokens - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: fields.cacheWriteInputTokens ?? 0,
+        completionTokens: fields.outputTokens ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractBedrockModelInvoke(
+    text: string,
+    _isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    // Non-streaming invoke reports counts via response headers regardless
+    // of the underlying model's payload format.
+    const headerInput = headers.get("x-amzn-bedrock-input-token-count");
+    const headerOutput = headers.get("x-amzn-bedrock-output-token-count");
+    if (headerInput !== null || headerOutput !== null) {
+        return {
+            promptTokens: Number(headerInput ?? 0),
+            cacheReadTokens: 0,
+            cacheWriteTokens: 0,
+            completionTokens: Number(headerOutput ?? 0),
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    // invoke-with-response-stream has no equivalent headers; the model's
+    // own usage shape (frequently Anthropic-style on Bedrock) is embedded
+    // inside binary event-stream framing, so fall back to a couple of
+    // known field-name shapes via regex.
+    const anthropicStyle = extractAnthropicMessages(text, true);
+    if (anthropicStyle) {
+        return anthropicStyle;
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokenCount",
+        "outputTokenCount"
+    ]);
+    if (fields.inputTokenCount === undefined) {
+        return null;
+    }
+    return {
+        promptTokens: fields.inputTokenCount,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: fields.outputTokenCount ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+const EXTRACTORS: Record<
+    AiCapability,
+    (text: string, isStream: boolean, headers: Headers) => AiUsage | null
+> = {
+    openai_chat: extractOpenAiChat,
+    openai_responses: extractOpenAiResponses,
+    anthropic_messages: extractAnthropicMessages,
+    gemini_generate_content: extractGoogleGenerateContent,
+    google_generate_content: extractGoogleGenerateContent,
+    // rawPredict is a passthrough to whatever the underlying publisher
+    // model speaks (often Anthropic-shaped on Vertex); try that, then give
+    // up to the token-count estimate.
+    google_raw_predict: (text, isStream) =>
+        extractAnthropicMessages(text, isStream),
+    bedrock_model_invoke: extractBedrockModelInvoke,
+    bedrock_converse: extractBedrockConverse
+};
+
+/**
+ * Attempts to pull provider-reported token usage out of an upstream AI
+ * gateway response. Returns null if the response didn't contain (or we
+ * couldn't find) usage data, in which case callers should fall back to
+ * `estimateUsage`.
+ */
+export function extractUsage(
+    capability: AiCapability,
+    responseText: string,
+    isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    try {
+        return EXTRACTORS[capability](responseText, isStream, headers);
+    } catch (error) {
+        logger.debug("Failed to extract AI usage from response", {
+            capability,
+            error
+        });
+        return null;
+    }
+}
+
+/**
+ * Best-guess token estimate for when the provider doesn't report usage.
+ * Uses OpenAI's BPE tokenizer as a stand-in for whatever tokenizer the
+ * actual model uses - close enough for an approximate cost figure, not
+ * exact for non-OpenAI models.
+ */
+export function estimateUsage(
+    promptText: string,
+    completionText: string
+): AiUsage {
+    const usage = emptyUsage();
+    usage.estimated = true;
+    try {
+        usage.promptTokens = promptText ? encode(promptText).length : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate prompt tokens", { error });
+    }
+    try {
+        usage.completionTokens = completionText
+            ? encode(completionText).length
+            : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate completion tokens", { error });
+    }
+    return usage;
+}
+
+/**
+ * OpenAI's Chat Completions API only includes a `usage` field in a
+ * streaming response when the request opts in via `stream_options:
+ * {include_usage: true}` - unlike the Responses API, Anthropic, Gemini and
+ * Bedrock, which report usage in a streaming response by default. Returns
+ * whether we need to inject that option ourselves to be able to track cost.
+ */
+export function needsStreamUsageInjection(
+    capability: AiCapability,
+    body: any
+): boolean {
+    return (
+        capability === "openai_chat" &&
+        body?.stream === true &&
+        body?.stream_options?.include_usage !== true
+    );
+}
+
+/**
+ * Returns a shallow-cloned body with `stream_options.include_usage`
+ * injected, for capabilities/requests where `needsStreamUsageInjection`
+ * is true. Leaves the original body untouched.
+ */
+export function withStreamUsageOption(body: any): any {
+    return {
+        ...body,
+        stream_options: { ...body.stream_options, include_usage: true }
+    };
+}
+
+/**
+ * When we injected stream_options.include_usage ourselves (the caller
+ * didn't ask for it), OpenAI appends an extra terminal SSE frame with an
+ * empty `choices: []` array carrying only the usage data. Callers that
+ * don't expect that shape (most minimal SSE parsers assume a non-empty
+ * choices array) shouldn't see it, so it's stripped back out of the bytes
+ * forwarded to the client.
+ */
+export function stripInjectedUsageFrame(sseText: string): string {
+    const parts = sseText.split(/(\r?\n\r?\n)/);
+    let out = "";
+    for (let i = 0; i < parts.length; i += 2) {
+        const frame = parts[i];
+        const separator = parts[i + 1] ?? "";
+        const dataLine = frame
+            .split(/\r?\n/)
+            .find((line) => line.startsWith("data:"));
+        if (dataLine) {
+            const data = dataLine.slice("data:".length).trim();
+            const parsed = data !== "[DONE]" ? tryParseJson(data) : null;
+            if (
+                parsed &&
+                Array.isArray(parsed.choices) &&
+                parsed.choices.length === 0 &&
+                parsed.usage
+            ) {
+                continue;
+            }
+        }
+        out += frame + separator;
+    }
+    return out;
+}
+
+/**
+ * Best-effort extraction of the model the upstream provider actually
+ * served, which some gateways/routers echo back and which may differ from
+ * the model the caller requested (e.g. an alias resolving to a dated
+ * snapshot). Falls back to the caller's requested model when absent.
+ */
+export function extractResponseModel(responseText: string): string | null {
+    const match = responseText.match(/"model"\s*:\s*"([^"]+)"/);
+    return match ? match[1] : null;
+}
+
+export function isUsageEmpty(usage: AiUsage): boolean {
+    return (
+        usage.promptTokens === 0 &&
+        usage.cacheReadTokens === 0 &&
+        usage.cacheWriteTokens === 0 &&
+        usage.completionTokens === 0 &&
+        usage.reasoningTokens === 0
+    );
+}

server/lib/alerts/events/healthCheckEvents.ts

@@ -202,6 +202,10 @@ async function handleResource(
         return;
     }
 
+    if (!target.resourceId) {
+        return;
+    }
+
     const [resource] = await trx
         .select()
         .from(resources)
@@ -221,10 +225,16 @@ async function handleResource(
         )
         .where(eq(targets.resourceId, resource.resourceId));
 
+    const monitoredTargets = otherTargets.filter(
+        (t) => t.hcHealth !== "unknown"
+    );
+
     let health = "healthy";
-    const allUnknown = otherTargets.every((t) => t.hcHealth === "unknown");
-    const allHealthy = otherTargets.every((t) => t.hcHealth === "healthy");
-    const allUnhealthy = otherTargets.every((t) => t.hcHealth === "unhealthy");
+    const allUnknown = monitoredTargets.length === 0;
+    const allHealthy = monitoredTargets.every((t) => t.hcHealth === "healthy");
+    const allUnhealthy = monitoredTargets.every(
+        (t) => t.hcHealth === "unhealthy"
+    );
 
     if (allUnknown) {
         logger.debug(

server/lib/billing/features.ts

@@ -1,28 +1,39 @@
-export enum FeatureId {
+export enum LimitId {
     USERS = "users",
     SITES = "sites",
     EGRESS_DATA_MB = "egressDataMb",
     DOMAINS = "domains",
     REMOTE_EXIT_NODES = "remoteExitNodes",
-    ORGINIZATIONS = "organizations",
+    ORGANIZATIONS = "organizations",
+    PUBLIC_RESOURCES = "publicResources",
+    PRIVATE_RESOURCES = "privateResources",
+    MACHINE_CLIENTS = "machineClients",
     TIER1 = "tier1"
 }
 
-export async function getFeatureDisplayName(featureId: FeatureId): Promise<string> {
+export async function getFeatureDisplayName(
+    featureId: LimitId
+): Promise<string> {
     switch (featureId) {
-        case FeatureId.USERS:
+        case LimitId.USERS:
             return "Users";
-        case FeatureId.SITES:
+        case LimitId.SITES:
             return "Sites";
-        case FeatureId.EGRESS_DATA_MB:
+        case LimitId.EGRESS_DATA_MB:
             return "Egress Data (MB)";
-        case FeatureId.DOMAINS:
+        case LimitId.DOMAINS:
             return "Domains";
-        case FeatureId.REMOTE_EXIT_NODES:
+        case LimitId.REMOTE_EXIT_NODES:
             return "Remote Exit Nodes";
-        case FeatureId.ORGINIZATIONS:
+        case LimitId.ORGANIZATIONS:
             return "Organizations";
-        case FeatureId.TIER1:
+        case LimitId.PUBLIC_RESOURCES:
+            return "Public Resources";
+        case LimitId.PRIVATE_RESOURCES:
+            return "Private Resources";
+        case LimitId.MACHINE_CLIENTS:
+            return "Machine Clients";
+        case LimitId.TIER1:
             return "Home Lab";
         default:
             return featureId;
@@ -30,15 +41,16 @@ export async function getFeatureDisplayName(featureId: FeatureId): Promise<strin
 }
 
 // this is from the old system
-export const FeatureMeterIds: Partial<Record<FeatureId, string>> = { // right now we are not charging for any data
+export const FeatureMeterIds: Partial<Record<LimitId, string>> = {
+    // right now we are not charging for any data
     // [FeatureId.EGRESS_DATA_MB]: "mtr_61Srreh9eWrExDSCe41D3Ee2Ir7Wm5YW"
 };
 
-export const FeatureMeterIdsSandbox: Partial<Record<FeatureId, string>> = {
+export const FeatureMeterIdsSandbox: Partial<Record<LimitId, string>> = {
     // [FeatureId.EGRESS_DATA_MB]: "mtr_test_61Snh2a2m6qome5Kv41DCpkOb237B3dQ"
 };
 
-export function getFeatureMeterId(featureId: FeatureId): string | undefined {
+export function getFeatureMeterId(featureId: LimitId): string | undefined {
     if (
         process.env.ENVIRONMENT == "prod" &&
         process.env.SANDBOX_MODE !== "true"
@@ -49,22 +61,20 @@ export function getFeatureMeterId(featureId: FeatureId): string | undefined {
     }
 }
 
-export function getFeatureIdByMetricId(
-    metricId: string
-): FeatureId | undefined {
-    return (Object.entries(FeatureMeterIds) as [FeatureId, string][]).find(
+export function getFeatureIdByMetricId(metricId: string): LimitId | undefined {
+    return (Object.entries(FeatureMeterIds) as [LimitId, string][]).find(
         ([_, v]) => v === metricId
     )?.[0];
 }
 
-export type FeaturePriceSet = Partial<Record<FeatureId, string>>;
+export type FeaturePriceSet = Partial<Record<LimitId, string>>;
 
 export const tier1FeaturePriceSet: FeaturePriceSet = {
-    [FeatureId.TIER1]: "price_1SzVE3D3Ee2Ir7Wm6wT5Dl3G"
+    [LimitId.TIER1]: "price_1SzVE3D3Ee2Ir7Wm6wT5Dl3G"
 };
 
 export const tier1FeaturePriceSetSandbox: FeaturePriceSet = {
-    [FeatureId.TIER1]: "price_1SxgpPDCpkOb237Bfo4rIsoT"
+    [LimitId.TIER1]: "price_1SxgpPDCpkOb237Bfo4rIsoT"
 };
 
 export function getTier1FeaturePriceSet(): FeaturePriceSet {
@@ -79,11 +89,11 @@ export function getTier1FeaturePriceSet(): FeaturePriceSet {
 }
 
 export const tier2FeaturePriceSet: FeaturePriceSet = {
-    [FeatureId.USERS]: "price_1SzVCcD3Ee2Ir7Wmn6U3KvPN"
+    [LimitId.USERS]: "price_1SzVCcD3Ee2Ir7Wmn6U3KvPN"
 };
 
 export const tier2FeaturePriceSetSandbox: FeaturePriceSet = {
-    [FeatureId.USERS]: "price_1SxaEHDCpkOb237BD9lBkPiR"
+    [LimitId.USERS]: "price_1SxaEHDCpkOb237BD9lBkPiR"
 };
 
 export function getTier2FeaturePriceSet(): FeaturePriceSet {
@@ -98,11 +108,11 @@ export function getTier2FeaturePriceSet(): FeaturePriceSet {
 }
 
 export const tier3FeaturePriceSet: FeaturePriceSet = {
-    [FeatureId.USERS]: "price_1SzVDKD3Ee2Ir7WmPtOKNusv"
+    [LimitId.USERS]: "price_1SzVDKD3Ee2Ir7WmPtOKNusv"
 };
 
 export const tier3FeaturePriceSetSandbox: FeaturePriceSet = {
-    [FeatureId.USERS]: "price_1SxaEODCpkOb237BiXdCBSfs"
+    [LimitId.USERS]: "price_1SxaEODCpkOb237BiXdCBSfs"
 };
 
 export function getTier3FeaturePriceSet(): FeaturePriceSet {
@@ -116,7 +126,7 @@ export function getTier3FeaturePriceSet(): FeaturePriceSet {
     }
 }
 
-export function getFeatureIdByPriceId(priceId: string): FeatureId | undefined {
+export function getFeatureIdByPriceId(priceId: string): LimitId | undefined {
     // Check all feature price sets
     const allPriceSets = [
         getTier1FeaturePriceSet(),
@@ -125,7 +135,7 @@ export function getFeatureIdByPriceId(priceId: string): FeatureId | undefined {
     ];
 
     for (const priceSet of allPriceSets) {
-        const entry = (Object.entries(priceSet) as [FeatureId, string][]).find(
+        const entry = (Object.entries(priceSet) as [LimitId, string][]).find(
             ([_, price]) => price === priceId
         );
         if (entry) {