@domesystems/templates 0.1.0 → 0.2.0

package/ai-agent-for-documentation/.env.example

@@ -1,12 +0,0 @@
-# The container holds exactly one secret.
-#
-# Tool credentials went into Dome when dome import ran and live behind the gateway.
-# The provider key lives with the Broker. Neither is here, and neither should be.
-
-# Revealed once after import: dome import outputs <job-id>
-DOME_TOKEN=
-
-# The gateway ROOT — not a bare host, and not the /mcp path.
-# Copy the "Gateway URL" from the command above. The container derives
-# /mcp for tools and /v1/messages for models from it.
-DOME_GATEWAY_URL=

package/ai-agent-for-documentation/app/agent.yaml

@@ -1,31 +0,0 @@
-name: docs-answerer
-title: Ask questions about product documentation
-summary: "A chat agent that answers questions from the Dome Platform documentation.\
-  \ Nothing to set up upstream \u2014 the documentation MCP server takes no credential,\
-  \ so this reaches a real system on the first run. The connection carries three tools.\
-  \ Two answer questions. The third returns the whole documentation index, and the\
-  \ agent will reach for it \u2014 Dome refuses, and names the rule that refused it."
-shape: interactive-assistant
-prompt_file: instructions.md
-help_file: help.md
-model: claude-haiku-4-5-20251001
-pool: docs-pool
-max_tokens: 4096
-skills:
-- name: answer-from-documentation
-  description: Search the documentation, read the most relevant pages, answer from
-    them.
-  tools:
-  - domedocs/search
-  - domedocs/get_page
-  - domedocs/list_pages
-invoker:
-  kind: chat
-options:
-  identity:
-    mode: none
-  model:
-    provider: anthropic
-    model: claude-haiku-4-5-20251001
-forbidden:
-- domedocs/list_pages

package/ai-agent-for-documentation/app/help.md

@@ -1,19 +0,0 @@
-Ask a question about the Dome Platform in plain language. The agent searches the
-public documentation at `docs.domesystems.ai`, reads the pages it finds, and
-answers from them.
-
-**Nothing to set up upstream.** The documentation MCP server takes no
-credential, so this template reaches a real system on its first run. The only
-key the workspace needs is a model provider key — every template needs one —
-and it goes to Dome, not into this container.
-
-**Ask something broad.** "What is Dome" or "what does the gateway do". The agent
-will reach for the documentation index to orient itself, that call will come
-back refused, and you will watch it fall back to search and answer anyway. The
-refusal is the rule, not the prompt.
-
-**Show detail** in the header opens the design: the topology, which tools are
-permitted, and the instructions the model was given.
-
-Nothing here is the record. Every call, allowed or refused, is in the audit log
-in Dome.

package/ai-agent-for-documentation/app/instructions.md

@@ -1,21 +0,0 @@
-You answer questions about the Dome Platform from its documentation.
-
-Work like this:
-
-1. If the question is broad — "what is Dome", "how does authorization work" —
-   start from the documentation index to see what exists.
-2. Search the documentation for the terms in the question. Search results carry
-   the path of every page they match.
-3. Read pages rather than searching repeatedly. Two or three searches is
-   plenty; after that, open the most promising pages in full **using the paths
-   the search results gave you**. Do not invent a path and try it.
-4. Answer from what you read, and name the pages you used.
-
-If the documentation does not contain the answer, say so plainly. Do not fill
-the gap from general knowledge — a confident answer that is not in the
-documentation is worse than no answer, because the reader cannot tell the
-difference.
-
-Use whatever tools you have been given, and do not decline a request on the
-assumption that you lack permission. If a call comes back refused, say what came
-back and carry on with the tools that do work.

package/ai-agent-for-documentation/docker-compose.yml

@@ -1,18 +0,0 @@
-#
-# Without DOME_TOKEN this starts in offline mode: the chrome works and the
-# allow-list is still enforced, but tools and model are simulated.
-services:
-  app:
-    # Built and published by dome-systems/template-runtime.
-    image: ghcr.io/dome-systems/runtime-py-chat:dev
-    env_file: .env
-    volumes:
-      - ./app:/config:ro
-    ports:
-      - "3000:3000"
-    healthcheck:
-      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:3000/health', timeout=2)"]
-      interval: 30s
-      timeout: 3s
-      start_period: 10s
-      retries: 3

package/ai-agent-for-documentation/dome.tf

@@ -1,146 +0,0 @@
-terraform {
-  required_providers {
-    dome = { source = "dome-systems/dome", version = ">= 0.1.0" }
-  }
-}
-
-variable "workspace_id" { type = string }
-variable "anthropic_api_key" {
-  type      = string
-  sensitive = true
-}
-
-resource "dome_gateway" "docs" {
-  workspace_id = var.workspace_id
-  name         = "docs"
-  description  = "Read surface over the public Dome Platform documentation"
-  is_default   = true
-}
-
-resource "dome_mcp_connection" "domedocs" {
-  workspace_id = var.workspace_id
-  name         = "domedocs"
-  protocol     = "streamable-http"
-  url          = "https://docs.domesystems.ai/mcp"
-  auth_method  = "none"
-}
-
-resource "dome_llm_connection" "haiku" {
-  workspace_id        = var.workspace_id
-  name                = "haiku"
-  provider_id         = "anthropic"
-  provider_config     = jsonencode({ model = "claude-haiku-4-5-20251001" })
-  auth_method         = "api-key"
-  credential_type     = "shared"
-  managed_header_name = "x-api-key"
-  secret_values       = { "x-api-key" = var.anthropic_api_key }
-}
-
-resource "dome_llm_pool" "docs" {
-  workspace_id = var.workspace_id
-  name         = "docs-pool"
-  is_default   = true
-}
-
-resource "dome_llm_pool_member" "haiku" {
-  workspace_id   = var.workspace_id
-  pool           = dome_llm_pool.docs.name
-  llm_connection = dome_llm_connection.haiku.name
-}
-
-resource "dome_gateway_llm_pool" "docs" {
-  workspace_id = var.workspace_id
-  gateway      = dome_gateway.docs.name
-  llm_pool     = dome_llm_pool.docs.name
-}
-
-resource "dome_gateway_tool" "search" {
-  workspace_id   = var.workspace_id
-  gateway        = dome_gateway.docs.name
-  mcp_connection = dome_mcp_connection.domedocs.name
-  tool           = "search"
-}
-
-resource "dome_gateway_tool" "get_page" {
-  workspace_id   = var.workspace_id
-  gateway        = dome_gateway.docs.name
-  mcp_connection = dome_mcp_connection.domedocs.name
-  tool           = "get_page"
-}
-
-resource "dome_gateway_tool" "list_pages" {
-  workspace_id   = var.workspace_id
-  gateway        = dome_gateway.docs.name
-  mcp_connection = dome_mcp_connection.domedocs.name
-  tool           = "list_pages"
-}
-
-resource "dome_agent" "docs_answerer" {
-  workspace_id     = var.workspace_id
-  name             = "docs-answerer"
-  allowed_gateways = [dome_gateway.docs.name]
-}
-
-resource "dome_agent_rules_bundle" "docs_answerer" {
-  workspace_id = var.workspace_id
-  agent        = dome_agent.docs_answerer.name
-  files = {
-    "answerer.cedar" = <<-EOT
-      permit ( principal, action == Dome::Action::"mcp:discover", resource );
-      permit ( principal, action == Dome::Action::"mcp:call", resource == Dome::MCPTool::"domedocs/search" );
-      permit ( principal, action == Dome::Action::"mcp:call", resource == Dome::MCPTool::"domedocs/get_page" );
-      forbid ( principal, action == Dome::Action::"mcp:call", resource == Dome::MCPTool::"domedocs/list_pages" );
-    EOT
-  }
-}
-
-resource "dome_filter" "keep_ssn_out_of_prompts" {
-  workspace_id = var.workspace_id
-  name         = "keep-ssn-out-of-prompts"
-  description  = "Redact US SSN patterns out of anything sent to the provider"
-  config = jsonencode({
-    text = { components = [{ action = "FILTER_ACTION_REDACT", matchers = [{ ssn = {} }] }] }
-  })
-}
-
-resource "dome_llm_connection_filters" "haiku_requests" {
-  workspace_id   = var.workspace_id
-  llm_connection = dome_llm_connection.haiku.name
-  direction      = "request"
-  filters        = [dome_filter.keep_ssn_out_of_prompts.name]
-}
-
-resource "dome_quota" "agent_daily_spend" {
-  workspace_id = var.workspace_id
-  dimension    = "llm"
-  unit         = "dome_usd"
-  subject_type = "agent"
-  subject      = dome_agent.docs_answerer.name
-  scope        = "total"
-  window       = "daily"
-  limit_amount = 5000000
-  name         = "docs-answerer daily spend"
-}
-
-resource "dome_quota" "search_calls" {
-  workspace_id = var.workspace_id
-  dimension    = "tool"
-  unit         = "calls"
-  subject_type = "tool_connection"
-  subject      = dome_mcp_connection.domedocs.name
-  scope        = "total"
-  window       = "daily"
-  limit_amount = 500
-  name         = "domedocs search calls"
-}
-
-resource "dome_agent_key" "runtime" {
-  workspace_id = var.workspace_id
-  agent        = dome_agent.docs_answerer.name
-  name         = "runtime"
-}
-
-output "docs_answerer_token" {
-  value     = dome_agent_key.runtime.token
-  sensitive = true
-}

package/ai-agent-for-documentation/rules/answerer.cedar

@@ -1,35 +0,0 @@
-// answer-from-dome-docs — authorization rules
-//
-// The documentation MCP server is read-only upstream: there is no write tool to
-// forbid here, so the boundary this template demonstrates is the allow-list
-// itself. `search` and `get_page` are what answering a question needs.
-// `list_pages` returns the entire documentation index — the call an agent makes
-// to enumerate a corpus rather than to answer about it — so it is on the
-// gateway, visible to the model, and denied.
-//
-// The write-boundary version of this story is
-// ai-agent-for-notion, where a Notion page-creation tool is
-// reachable and refused.
-
-permit ( principal, action == Dome::Action::"mcp:discover", resource );
-
-permit (
-  principal,
-  action == Dome::Action::"mcp:call",
-  resource == Dome::MCPTool::"domedocs/search"
-);
-
-permit (
-  principal,
-  action == Dome::Action::"mcp:call",
-  resource == Dome::MCPTool::"domedocs/get_page"
-);
-
-// Registering an agent against a gateway auto-generates an Agent Form policy
-// permitting any tool whose `gateways` attribute contains that gateway. A
-// permit cannot narrow that; only a forbid can. Cedar forbid always wins.
-forbid (
-  principal,
-  action == Dome::Action::"mcp:call",
-  resource == Dome::MCPTool::"domedocs/list_pages"
-);

package/ai-agent-for-documentation/template.yaml

@@ -1,130 +0,0 @@
-apiVersion: templates.domesystems.ai/v1
-kind: Template
-
-metadata:
-  name: answer-from-dome-docs
-  slug: ai-agent-for-documentation
-  title: Ask questions about product documentation
-  target_query: ai chatbot for documentation
-  summary: >-
-    A chat agent that answers questions from the Dome Platform documentation.
-    Nothing to set up upstream — the documentation MCP server takes no
-    credential, so this reaches a real system on the first run. The connection
-    carries three tools. Two answer questions. The third returns the whole
-    documentation index, and the agent will reach for it — Dome refuses, and
-    names the rule that refused it.
-  domain: productivity
-  shape: interactive-assistant
-  ui: chat
-  business_pattern: consumes
-  governance_interest: [read-only, tool-scoped-access]
-
-# Library-only topology and documentation data. `dome.tf` is the deployable source.
-catalog:
-  gateway:
-    name: docs
-    description: Read surface over the public Dome Platform documentation
-    default: true
-
-  connections:
-    - name: domedocs
-      protocol: streamable-http
-      url: https://docs.domesystems.ai/mcp
-      # The whole point of this template. A public MCP server, no token, no
-      # consent screen, no account to create upstream.
-      auth:
-        method: none
-      permit:
-        - search
-        - get_page
-      # On the gateway on purpose, so the model can see it and try. Cedar
-      # forbids it, and a forbid beats the auto-generated Agent Form permit.
-      # Without this there is no denial to observe.
-      forbid:
-        - list_pages
-
-  model:
-    name: haiku
-    provider: anthropic
-    model_id: claude-haiku-4-5-20251001
-    secret:
-      env: ANTHROPIC_API_KEY
-      variable: anthropic_api_key
-      label: an Anthropic API key
-
-  pool:
-    name: docs-pool
-    default: true
-
-  agent:
-    name: docs-answerer
-    # No identity. This template's governance story is the allow-list, and
-    # act-as would add a second secret to the one thing in the library that
-    # needs none. The caller-scoped story is answer-from-knowledge-base.
-    act_as:
-      method: none
-
-  rules:
-    file: rules/answerer.cedar
-
-  # A text filter on the *model* request. Whatever a person types into the box
-  # goes to a provider, and this is the cheapest true statement a template can
-  # make: an SSN never reaches Anthropic. Text-kind, so it binds to the model
-  # connection — a tool connection would refuse it.
-  guards:
-    filters:
-      - name: keep-ssn-out-of-prompts
-        description: Redact US SSN patterns out of anything sent to the provider
-        why: >-
-          Whatever a person types into the box goes to a provider. This is the
-          cheapest true thing a template can claim — an SSN never reaches
-          Anthropic — and it costs one line because the credential and the
-          routing already live in Dome. Text-kind, so it binds to the model
-          connection; a tool connection refuses a text filter outright.
-        redact_ssn: true
-        bind: { model: haiku, direction: request }
-
-  quotas:
-    # What the agent may spend with Dome, per day.
-    - subject: agent
-      agent: docs-answerer
-      unit: dome_usd
-      limit: 5
-      window: daily
-      name: docs-answerer daily spend
-    # And a call ceiling on the one tool that does the work. The docs server is
-    # public and someone else pays for it, so not hammering it is part of using
-    # it politely — a rate limit with a reason, rather than a round number.
-    - subject: tool
-      tool: domedocs/search
-      unit: calls
-      limit: 500
-      window: daily
-      name: domedocs search calls
-
-app:
-  runtime:
-    pattern: chat
-    language: python
-  agent:
-    prompt: app/instructions.md
-    help: app/help.md
-    model: claude-haiku-4-5-20251001
-    # 2048 truncates a documentation answer mid-sentence, and a truncated turn
-    # loses the tool call it was half-way through writing.
-    max_tokens: 4096
-    skills:
-      - name: answer-from-documentation
-        description: Search the documentation, read the most relevant pages, answer from them.
-        tools:
-          - domedocs/search
-          - domedocs/get_page
-          - domedocs/list_pages
-    invoker:
-      kind: chat
-  options:
-    identity:
-      mode: none
-    model:
-      provider: anthropic
-      model: claude-haiku-4-5-20251001

package/ai-agent-for-documentation/verification.json

@@ -1,7 +0,0 @@
-{
-  "verified": "2026-09-12T19:25:26Z",
-  "context": "api-domesystems-ai-dome-systems-core-dome-library",
-  "gateway": "docs",
-  "agent": "docs-answerer",
-  "cli": "1.0.2"
-}

package/ai-agent-for-notion/.env.example

@@ -1,19 +0,0 @@
-# The container holds exactly one secret.
-#
-# Tool credentials went into Dome when dome import ran and live behind the gateway.
-# The provider key lives with the Broker. Neither is here, and neither should be.
-
-# Revealed once after import: dome import outputs <job-id>
-DOME_TOKEN=
-
-# The gateway ROOT — not a bare host, and not the /mcp path.
-# Copy the "Gateway URL" from the command above. The container derives
-# /mcp for tools and /v1/messages for models from it.
-DOME_GATEWAY_URL=
-
-# The second secret, and the only one an identity-enabled blueprint adds.
-# The container signs the act-as header with it; the gateway verifies the
-# signature before it will believe who the request claims to be. Without
-# that, anyone holding the agent token could claim to be anyone.
-# Must match the actas_hmac_secret supplied to dome import.
-ACTAS_HMAC_SECRET=

package/ai-agent-for-notion/app/agent.yaml

@@ -1,32 +0,0 @@
-name: knowledge-answerer
-title: Ask questions across your Notion workspace
-summary: "A chat agent that answers questions from your Notion workspace. The connection\
-  \ carries 43 tools, twenty-odd of which write. Two are permitted. Ask it to create\
-  \ a page and it will try \u2014 and Dome will refuse, naming the rule that refused\
-  \ it."
-shape: interactive-assistant
-prompt_file: instructions.md
-help_file: help.md
-model: claude-sonnet-5
-pool: answerer-pool
-max_tokens: 2048
-skills:
-- name: answer-from-notion
-  description: Search the workspace, read the most relevant pages, answer from them.
-  tools:
-  - notion/notion-search
-  - notion/notion-fetch
-  - notion/notion-create-pages
-invoker:
-  kind: chat
-options:
-  identity:
-    mode: oidc-stub
-    personas:
-    - id: alice
-    - id: bob
-  model:
-    provider: anthropic
-    model: claude-sonnet-5
-forbidden:
-- notion/notion-create-pages

package/ai-agent-for-notion/app/help.md

@@ -1,14 +0,0 @@
-Ask a question in plain language. The agent searches the connected Notion
-workspace, reads the pages it finds, and answers from them.
-
-**Show detail** in the header opens the design: the topology, which tools are
-permitted, and the instructions the model was given.
-
-**Try changing who is asking.** Alice is permitted. Bob is not, and the same
-question comes back refused — the difference is the rule, not the wording.
-
-**Ask it to create a page.** It will try, and the attempt will be blocked. The
-connection carries 43 Notion tools; two are permitted.
-
-Nothing here is the record. Every call, allowed or refused, is in the audit log
-in Dome.

package/ai-agent-for-notion/app/instructions.md

@@ -1,16 +0,0 @@
-You answer questions using the connected Notion workspace.
-
-Work like this:
-
-1. Search Notion for pages relevant to the question.
-2. Fetch the most promising two or three and read them.
-3. Answer from what you found, and name the pages you used.
-
-If the workspace does not contain the answer, say so plainly. Do not fill the
-gap from general knowledge — a confident answer that is not in the
-documentation is worse than no answer, because the reader cannot tell the
-difference.
-
-Use whatever tools you have been given, and do not decline a request on the
-assumption that you lack permission. If an action is not allowed, attempting it
-is how you and the person asking find that out; report what came back.

package/ai-agent-for-notion/docker-compose.yml

@@ -1,18 +0,0 @@
-#
-# Without DOME_TOKEN this starts in offline mode: the chrome works and the
-# allow-list is still enforced, but tools and model are simulated.
-services:
-  app:
-    # Built and published by dome-systems/template-runtime.
-    image: ghcr.io/dome-systems/runtime-py-chat:dev
-    env_file: .env
-    volumes:
-      - ./app:/config:ro
-    ports:
-      - "3000:3000"
-    healthcheck:
-      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:3000/health', timeout=2)"]
-      interval: 30s
-      timeout: 3s
-      start_period: 10s
-      retries: 3