Prompt templates

The ReAct graph injects the following variables when it renders the agent's system prompt for a model call. Reference them with Jinja2 syntax, for example {{ agent_name }}. The set of available variables is graph-dependent, so it is documented here alongside the graph that provides it. See Write templated prompts for usage patterns.

state is only the run state's context object, not the whole run state. The other run state fields, such as model and capability_override, are not in the prompt. A template reads what tools wrote to context through it:

prompt = <<-EOT
  You are a weather assistant.

  {%- if state.last_city %}
  The user last asked about {{ state.last_city }}.
  {%- endif %}
EOT

tools differs between runs, because client tools come with each run request.

Skill

Each entry in skills describes one skill available in the prompt.

ContextVar

Each entry in context_vars describes one context variable available in the prompt.

Tool

Each entry in tools describes one tool the model can call in the run.

Guard any text that describes a tool on that tool being in tools, so the text renders only for runs that carry the tool:

prompt = <<-EOT
  You are a documentation assistant.

  {%- if "open_page" in tools | map(attribute="name") %}
  To take the reader to a page, call `open_page`.
  {%- endif %}
EOT

tools | map(attribute="name") is the list of tool names, and in checks that the tool is there. A client that does not send the tool gets a prompt without the text. Server tools are guarded the same way.

Message

Each entry in messages is one message in the session. type decides which fields it has.

ToolCall

Each entry in a message's tool_calls is one call the model made.

UserContent

Each entry in a user message's content is one content block.

Media

The value of an image, audio, video or document UserContent.

MediaSource

The prompt field itself is declared on the agent block.