All node types, retry policies, conditions, and parallel branches for YAML workflows.

Workflow Authoring (YAML)

YAML is the primary authoring format for JamJet workflows. This page covers every node type, control flow option, and configuration knob.

Workflow header

workflow:
  id: my-workflow          # unique identifier (slug format)
  version: 0.1.0           # semver
  state_schema:            # typed state (all keys optional at start)
    query: str
    results: list[str]
    answer: str
    score: float
  start: first-node        # entry node id

Model node

Calls an LLM and writes the response to a state key.

nodes:
  think:
    type: model
    model: claude-haiku-4-5-20251001       # or gpt-4o, gemini-2.0-flash, etc.
    prompt: |
      You are a helpful assistant.
      Answer this question: {{ state.query }}
    output_key: answer          # writes model response here
    next: end

    # Optional
    system: "You are concise and precise."
    temperature: 0.3
    max_tokens: 1024
    retry:
      max_attempts: 3
      backoff: exponential      # or linear, constant
      delay_ms: 500

Supported models:

Provider	Model IDs
Anthropic	`claude-haiku-4-5-20251001`, `claude-sonnet-4-6`, `claude-opus-4-6`
OpenAI	`gpt-4o`, `gpt-4o-mini`, `o3-mini`
Google	`gemini-2.0-flash`, `gemini-2.5-pro`

Tool node (MCP)

Calls a tool from a connected MCP server.

nodes:
  search:
    type: tool
    server: brave-search        # MCP server name (from jamjet.toml)
    tool: web_search
    arguments:
      query: "{{ state.query }}"
      count: 5
    output_key: search_results
    next: summarize

MCP servers are configured in jamjet.toml:

[[mcp.servers]]
name = "brave-search"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-brave-search"]
env = { BRAVE_API_KEY = "${BRAVE_API_KEY}" }

HTTP node

Makes an HTTP request and writes the response body to state.

nodes:
  fetch-data:
    type: http
    method: GET
    url: "https://api.example.com/data/{{ state.item_id }}"
    headers:
      Authorization: "Bearer {{ env.API_KEY }}"
    output_key: raw_data
    next: process

For POST/PUT, add a body field:

    method: POST
    body:
      query: "{{ state.query }}"
      format: json

Branch node

Routes execution based on a condition.

nodes:
  route:
    type: branch
    conditions:
      - if: "state.confidence >= 0.9"
        next: done
      - if: "state.confidence >= 0.5"
        next: refine
    default: escalate        # fallback if no condition matches

Conditions are evaluated in order — first match wins. Expressions use a simple subset of Python syntax with state.* and env.* available.

Parallel node

Fans out to multiple branches, waits for all to complete, then continues.

nodes:
  gather:
    type: parallel
    branches:
      - search
      - fetch-docs
      - check-cache
    join: synthesize         # node to run after all branches complete

Each branch runs concurrently. State writes from parallel branches are merged; if two branches write the same key, the last writer wins.

Wait node

Pauses execution until an external event is received.

nodes:
  await-approval:
    type: wait
    event: human_approved    # event name to wait for
    timeout_hours: 24
    on_timeout: escalate     # node to run if timeout fires
    next: continue

Resume a waiting execution with:

jamjet resume exec_01JM4X8NKWP2 --event human_approved --data '{"approved": true}'

Eval node

Scores output quality inline. Useful for self-healing loops.

nodes:
  check-quality:
    type: eval
    scorers:
      - type: llm_judge
        rubric: "Is the answer accurate, complete, and under 200 words?"
        min_score: 4          # 1-5 scale
        model: claude-haiku-4-5-20251001
      - type: assertion
        check: "len(output.answer) > 0"
      - type: latency
        max_ms: 5000
    on_fail: retry_with_feedback   # or escalate, halt, log_and_continue
    max_retries: 2
    next: end

on_fail options:

retry_with_feedback — re-runs the previous model node with scorer feedback injected into the prompt
escalate — routes to an escalation node (define escalate: in the workflow header)
halt — stops the execution with an error
log_and_continue — records the failure but continues anyway

End node

Terminates the execution. Every workflow needs at least one.

nodes:
  end:
    type: end

You can have multiple named end nodes (e.g., success, failure, escalated) — they all terminate execution cleanly.

Retry policies

Any node can have a retry policy:

    retry:
      max_attempts: 3
      backoff: exponential    # linear | constant | exponential
      delay_ms: 200           # initial delay
      max_delay_ms: 10000     # cap for exponential backoff
      retry_on:               # optional: only retry on these errors
        - model_overloaded
        - rate_limited
        - timeout

Template expressions

Prompts, URLs, headers, and arguments support {{ }} template expressions:

Expression	Value
`{{ state.key }}`	A state value
`{{ env.KEY }}`	An environment variable
`{{ execution.id }}`	Current execution ID
`{{ execution.workflow_id }}`	Workflow ID
`{{ node.id }}`	Current node ID

Full example

workflow:
  id: research-agent
  version: 0.2.0
  state_schema:
    query: str
    search_results: list[str]
    draft: str
    answer: str
  start: search

nodes:
  search:
    type: tool
    server: brave-search
    tool: web_search
    arguments:
      query: "{{ state.query }}"
      count: 10
    output_key: search_results
    next: draft

  draft:
    type: model
    model: claude-sonnet-4-6
    prompt: |
      Based on these search results:
      {{ state.search_results | join('\n') }}

      Write a comprehensive answer to: {{ state.query }}
    output_key: draft
    next: evaluate

  evaluate:
    type: eval
    scorers:
      - type: llm_judge
        rubric: "Is this answer factually grounded in the search results, complete, and well-structured?"
        min_score: 4
        model: claude-haiku-4-5-20251001
      - type: assertion
        check: "len(output.draft) >= 100"
    on_fail: retry_with_feedback
    max_retries: 2
    next: done

  done:
    type: end

Workflow Authoring (YAML)

On this page