Deployment Assembly Guide

core/deploy assembles one deployment document into shared resources and runnable agents. It knows only the resource protocol; concrete kinds are registered by the application.

Document

version: v1

resources:
  events:
    kind: event.Bus
    impl: memory

  provider:
    kind: inference.Provider
    impl: openai
    settings:
      id: deepseek
      spec:
        endpoint: {base_url: https://api.deepseek.com}
        wire: {reasoning_channel: text}
        models:
          - {name: deepseek-flash, kind: generate,
             capabilities: {inputs: [text, image, data, tool_call, tool_result], outputs: [text]}}
      profiles:
        - secrets:
            api_key: ${env:DEEPSEEK_API_KEY}

  infer:
    kind: inference.Assembly
    impl: unified
    deps:
      provider: provider

## Settings reference expansion

The deploy builder expands inline scalar references in every settings
subtree — resources, agent engines, and agent hooks — before a factory
decodes them. Supported schemes are `${env:NAME}` (missing variable fails the
build), `${base}` / `${base:rel}` (rooted at the deployment document's base
dir), and `${home}` / `${home:rel}` (plus the `~` / `~/...` shorthand). All
schemes are enabled by default; a custom `resource.ReferenceResolver` can be
injected with `deploy.WithResolver` to add application-specific schemes on
top of the defaults (a same-named custom scheme overrides the built-in one).
Expansion is strict: an unknown scheme, a disabled scheme, or a malformed
reference fails the build. A literal `${` is written as `\${...}`, matching
the graph's `${board:*}` escaping convention — **every** literal `${` in any
settings subtree must be escaped, and a leading `~` / `~/...` expands to the
user home directory. Content materialized from `{file: ...}` / `{embed: ...}`
references is expanded the same way. Deployments that previously relied on
literal `${...}` text or `~` paths must escape/quote them after this change.

The agent board namespace is exempt: `${board:<path>}` (and the escaped
`\${board:<path>}` form, backslash preserved) is deferred to execution time,
where it resolves against the live `agent.Board` — graph node configs and the
script bridge's `board.resolve` both understand it. Deploy-time expansion
never touches board references, so existing graph configs need no escaping.

## Secret stores

Credentials can live in declarative `secret.Store` resources instead of
plaintext settings:

```yaml
resources:
  secret.env:
    kind: secret.Store
    impl: env
    settings:
      id: env
      default: true

  provider:
    kind: inference.Provider
    impl: openai
    settings:
      profiles:
        - secrets:
            api_key: ${secret:DEEPSEEK_API_KEY}   # default store
            # api_key: ${secret:env.DEEPSEEK_API_KEY}  # explicit store

• Stores are built ahead of every other resource, then assembled into the ${secret:...} scheme, so any resource can reference them (and may declare them as deps for explicit ordering).
• ${secret:NAME} resolves through the default store — exactly one store may declare default: true. ${secret:ID.NAME} addresses a store by its id (falling back to the resource name).
• The store prefix is everything before the first dot; the rest is the secret name.
• Missing stores or a NAME-only reference with no default store fail the build. Resolution is lazy: a ${secret:...} reference decodes into a typed resource.Secret that looks the value up at use time (cached per store for one minute), so a missing secret surfaces at request time rather than at deploy time. Secret values never appear in error messages, logs, or serialized settings — resource.Secret always renders as <secret>.
• Backends are ordinary resource factories: the built-in env store reads environment variables and the built-in file store reads one file per secret under a configured base directory (escaping the base is rejected, trailing newlines are stripped, and file size is capped). External backends (keychain, vault, ...) register their own secret.Store impls with zero core changes.

agents: assistant: card: name: Assistant engine: kind: agent.Engine impl: graph deps: inference: infer settings: graph: {file: ./graphs/assistant.yaml}

runtime: event_bus: events sessions: idle_timeout: 10m

Top-level fields:

- `version` is required.
- `resources` is a map of named resource entries.
- `agents` is a map of agent definitions.
- `runtime` is opaque to `deploy`; `core/runtime` decodes it strictly.

## Resource entries

```yaml
<name>:
  kind: event.Bus
  impl: memory
  deps: {bus: other}
  settings: {literal: object}

settings may be inline content or a whole-subtree {"file": ...} / {"embed": ...} reference. deps bind declared dependency names to resource refs. A ref is either resource or resource/item.

Build and wire

Builder.Build constructs resources in dependency order. Builder.Wire then:

1. wires resource.Wireable values;
2. builds agent engines and hooks;
3. binds resource.DeploymentBinder values with the completed deployment.

Use Builder.Deploy for the convenience of both phases.

reg := resource.NewRegistry()
reg.MustRegister(event.NewFactory())
reg.MustRegister(inference.Factory{})
reg.MustRegister(graphresource.Factory{})

builder := deploy.NewBuilder(reg)
result, err := builder.Deploy(ctx, doc)
if err != nil {
    return err
}
defer result.Close()

Result.Close closes every built resource value in reverse construction order, then closes each bound agent (engine and lifecycle hooks). Agents can also be assembled individually at runtime via the exported deploy.BindAgent — the same path core/runtime uses for its live agent registry (see runtime.md).

Caller-owned dependencies can be supplied with deploy.WithExternalResources; they are visible to dependency resolution but are never constructed, wired, or closed by deploy. The runtime layer exposes this through its runtime.external_deps document declaration and runtime.Builder.WithExternalResource (see runtime.md).

Layered configuration

deploy.LoadLayers loads multiple Layer values, merges them in ascending priority order, and returns provenance. The first layer must be complete; later layers may be partial. Runtime settings, resource settings, and agent policy are merged.

See resource.md for the factory protocol.