Node Types Reference

This auto-generated reference documents the Input and Output schemas for workflow node types. Descriptions are extracted from proto annotations in workflow_v2.proto.

Common Node Fields

These fields are available on all node types. They provide shared functionality for node identification, conditional execution, machine placement, and output handling. Everything else a node takes is specific to its type, listed in its section below.

Field	Type	Required	Description
`id`	string	Yes	Unique within the workflow; other nodes read this node’s outputs as nodes.`<id>`.*
`type`	string	Yes	The node type, such as workflow, call_llm, run or loop; each section below names its own
`condition`	string	No	Raw CEL (no {`{ }`}): the node runs only when it is true, and is skipped otherwise. On a join node, all or any
`timeout`	string	No	Overrides the node’s default timeout, as a duration such as 30s, 5m or 1h30m
`save_message`	SaveMessageConfig	No	A message to save to the thread when the node completes, as a template over output.*
`daemon`	CelDaemonSelector	No	Which machine runs this node’s tools, overriding the workflow’s daemon setting
`outcome`	string	No	The verdict reaching this node stamps on the run: success or failure. The last node that runs with one decides the run’s outcome (enum: success\|failure)

Thread Configuration

Workflow, loop and router nodes take a thread field that controls conversation context:

- id: analyze
  type: workflow
  ref: builtin://agent
  thread:
    mode: fork      # fork, new, or inherit (default)
    memo: true      # reuse across loop iterations
    inject:
      role: user
      content: "Analyze: {{inputs.data}}"

Conditional Execution

Use condition to skip nodes based on runtime state:

- id: cleanup
  type: run
  condition: "nodes.process.exit_code != 0"
  command: "rm -rf temp/"

Automatic Message Saving

Use save_message to capture output without a separate save_message node:

- id: summarize
  type: workflow
  ref: builtin://agent
  thread:
    mode: fork
  save_message:
    role: assistant
    content: "{{output.response_text}}"

Action

type: action Run an integration action (for example an HTTP request)

Inputs

Field	Type	Required	Default	Description
`uses`	string	No	-	Which integration action to run, as integration/action@major
`with`	map	No	-	Action parameters, keyed as in the action’s schema. Values may be literals or {`{ }`} templates
`connection`	string	No	-	Id of the saved connection to authenticate with. Leave it empty to use your default connection
`hosted_use`	string	No	-	Hosted HTTP use-grant reference for this fixed action

Outputs

Field	Type	Description
`content`	string	-
`is_error`	boolean	-
`status_code`	integer	-
`data`	object	-
`uses`	string	-
`retryable`	boolean	-
`connection_id`	string	-
`error_code`	string	-

Agent

type: workflow Invoke an agent or sub-workflow

Inputs

Field	Type	Required	Default	Description
`ref`	string	No	-	The workflow to run: builtin://`<name>` for a built-in, project://`<name>` for one in this project
`args`	map	No	-	Values for the sub-workflow’s inputs, keyed by input name
`presets`	map	No	-	A preset to apply per input group, as group: preset. Use default for ungrouped inputs
`passthrough`	string	No	-	Inputs of this workflow to hand to the sub-workflow under the same name

Outputs

Field	Type	Description
`outputs`	object	-

Each output the sub-workflow declares is also on the node, by name (nodes.<id>.<output>).

Approval

type: approval Pause workflow execution for user approval

Inputs

Field	Type	Required	Default	Description
`title`	string	No	-	What the approver is asked to decide, shown as the approval’s title
`timeout`	string	No	-	How long to wait for a decision before the step times out

Outputs

Field	Type	Description
`approval_id`	string	-
`status`	string	-
`action_taken`	string	-
`data`	object	-

Ask Question

type: ask_question Pause workflow execution to ask the user a question

Inputs

Field	Type	Required	Default	Description
`metadata`	string	No	-	The questions to ask, as JSON: a questions list, each with question text and its options

Outputs

Field	Type	Description
`has_feedback`	boolean	-
`response`	string	-

Call LLM

type: call_llm Send a prompt to a language model and get a response

Inputs

Field	Type	Required	Default	Description
`model`	model	No	-	A model tag (flagship, fast, cheap) or an explicit model ID
`temperature`	number	No	0.7	How varied replies are: 0 is the most repeatable, higher is more creative (0 to 2) (range: 0-2)
`thinking_level`	string	No	none	Extended thinking level for complex reasoning (enum: none\|low\|medium\|high\|xhigh)
`system_prompt`	string	No	-	Optional instructions for how the model should behave on this call
`skills`	array	No	-	Skills to preload into this call, by path, so the model starts with their instructions

Outputs

Field	Type	Description
`message`	object	-
`message.id`	string
`message.role`	string
`message.text`	string
`message.phase`	string
`response_text`	string	-
`tool_calls`	array	-
`tool_calls[].id`	string
`tool_calls[].name`	string
`tool_calls[].input`	string
`token_count`	integer	-
`upstream_request_id`	string	-
`upstream_proxyman_id`	string	-
`response_data`	object	-
`cost`	number	-
`compaction_threshold`	integer	-
`model`	string	-
`pending_inbox`	boolean	-
`message_id`	string	-
`last_stream_seq`	integer	-
`stop_reason`	string	-
`finish_reason`	string	-
`capabilities`	object	-
`provider_account_id`	string	-

Compact

type: compact Compact conversation context to reduce token usage

Inputs

Field	Type	Required	Default	Description
`model`	model	No	-	Model used to summarize the conversation (defaults to the built-in summarization tier)
`force`	boolean	No	-	Compact even when it would remove little of the context (a compaction the user asked for)

Outputs

Field	Type	Description
`message`	object	-
`message.id`	string
`message.role`	string
`message.text`	string
`message.phase`	string
`skipped_reason`	string	-

Compute

type: compute Compute named values from CEL expressions: no model, no shell, no worker round trip. Each output is {{ expr }} (or a literal); a value that is exactly one {{ }} keeps its type. Read them as nodes.<id>.<name>. Outputs cannot read each other: chain two compute nodes when one feeds the next. A failing expression fails the node. Use it for deterministic logic (labels, sizes, expiry, parsing a node’s JSON stdout) instead of a script

Inputs

Field	Type	Required	Default	Description
`outputs`	map	No	-	Named values, each a {`{ }`} expression (or a literal). Read them as nodes.`<id>`.`<name>`

Create Worktree

type: create_worktree Create a git worktree for isolated development

Inputs

Field	Type	Required	Default	Description
`name`	string	No	-	A short name for the worktree; it becomes part of the worktree’s folder path
`base_branch`	string	No	-	The branch to start from. Empty uses the repository’s default branch
`copy_files`	string	No	-	Exact paths from the project root to copy into the new worktree, such as untracked config
`force`	boolean	No	-	Force creation by deleting existing worktree

Outputs

Field	Type	Description
`id`	string	-
`name`	string	-
`path`	string	-
`branch`	string	-
`base_branch`	string	-
`repo_id`	string	-
`status`	string	-

Join

type: join Wait for parallel branches to complete before continuing

Outputs

Field	Type	Description
`_sources`	array	-

Loop

type: loop Execute a sub-workflow in a loop with conditions

Inputs

Field	Type	Required	Default	Description
`while`	boolean	No	-	CEL condition: loop continues while true
`parallel`	boolean	No	-	Execute all iterations concurrently
`items`	string	No	-	What to iterate over: an expression yielding a list or map. Each iteration sees it as iter.item
`key`	string	No	-	For parallel loops, each iteration’s key in _results. Defaults to the iteration’s index
`on_failure`	string	No	-	Failure policy for parallel iterations (enum: continue\|fail_fast\|fail_all)
`passthrough`	string	No	-	Inputs of this workflow to hand to the loop body under the same name

Outputs

Field	Type	Description
`_iterations`	integer	-
`_results`	map	-
`_completed`	integer	-
`_failed`	integer	-
`_parallel`	boolean	-

A sequential loop also exposes each output its body declares, by name, from the last iteration (nodes.<id>.<output>). _results, _completed, _failed and _parallel are set by parallel loops.

Publish Branch

type: git_publish Commit a worktree’s changes and push them to a feature branch. Refuses the default branch, protected branches, main, master and release/*, and any branch outside the allow list; refuses a fork head; never force-pushes. Nothing to commit is not an error (pushed is false). The credential is used for this one push and never reaches a shell, a file or the environment. Pair it with shell credentials: none

Inputs

Field	Type	Required	Default	Description
`worktree`	string	No	-	Absolute path of the worktree whose changes to commit and push
`repo`	string	No	-	The GitHub repository, as owner/name
`branch`	string	No	-	The branch to push to. Leave it empty when pr_head_of is set
`pr_head_of`	integer	No	-	A pull request number: push to that pull request’s own head branch. The head must be in this repository, not a fork
`allow`	array	No	-	Glob patterns the branch must match, such as autopilot/*. Empty allows any branch the built-in checks accept
`message`	string	No	-	The commit message. Reliant-Run and Reliant-Workflow trailers are appended

Outputs

Field	Type	Description
`pushed`	boolean	-
`sha`	string	-
`branch`	string	-
`refused`	boolean	-
`refusal`	string	-
`files_changed`	integer	-
`insertions`	integer	-
`deletions`	integer	-
`base_sha`	string	-

Router

type: router Route to a workflow or node based on LLM classification

Inputs

Field	Type	Required	Default	Description
`system_prompt`	string	No	-	Instructions for the routing model, replacing the default routing prompt
`model`	model	No	-	The model that makes the routing decision (defaults to a fast model)
`fallback`	string	No	-	Where to go if the routing decision fails: a preset name (workflow routing) or a node id (node routing)
`outputs`	map	No	-	Named outputs for this step, each an expression over the chosen workflow’s outputs

Outputs

Field	Type	Description
`selected_workflow`	string	-
`selected_preset`	string	-
`prompt`	string	-
`reasoning`	string	-
`outputs`	object	-
`selected_node`	string	-

Run Command

type: run Execute a shell command

Inputs

Field	Type	Required	Default	Description
`command`	string	No	-	The shell command to run in the project directory
`credentials`	string	No	-	Set to none to run the command with no GitHub, git, ssh or cloud credentials. Not a sandbox: files the daemon’s user can read stay readable

Outputs

Field	Type	Description
`exit_code`	integer	-
`stdout`	string	-
`stderr`	string	-
`working_dir`	string	-
`log_file`	string	-

Run LLM Tool Calls

type: execute_tools Run the tool calls an upstream Call LLM step returned

Inputs

Field	Type	Required	Default	Description
`tool_calls`	string	No	-	The tool calls to run: wire in the tool_calls output of the Call LLM step that requested them

Outputs

Field	Type	Description
`message`	object	-
`message.id`	string
`message.role`	string
`message.text`	string
`message.phase`	string
`tool_results`	array	-
`tool_results[].tool_call_id`	string
`tool_results[].name`	string
`tool_results[].content`	string
`tool_results[].is_error`	bool
`tool_results[].attachment_ids`	string
`thread_token_count`	integer	-
`total_result_chars`	integer	-
`response_data`	object	-
`granted_tools`	array	-

Run Tool

type: invoke_tool Run one tool you pick, with parameters you set; no LLM involved

Inputs

Field	Type	Required	Default	Description
`tool`	string	No	-	The tool to run, by name. Only tools that allow direct use from a workflow can be named
`params`	map	No	-	The tool’s parameters, keyed as in its schema. Values may be literals or {`{ }`} templates

Outputs

Field	Type	Description
`content`	string	-
`is_error`	boolean	-
`attachment_ids`	array	-
`data`	object	-
`tool`	string	-

Save Message

type: save_message Save a message to the conversation thread

Inputs

Field	Type	Required	Default	Description
`role`	string	No	-	Message role (enum: user\|assistant\|system\|tool)
`content`	string	No	-	The message text, usually a template over an earlier step’s output
`tool_calls`	string	No	-	Tool calls to save with an assistant message: the tool_calls a Call LLM step returned
`display_style`	string	No	-	UI display style hint (enum: info\|warning\|success\|hidden)

Outputs

Field	Type	Description
`message`	object	-
`message.id`	string
`message.role`	string
`message.text`	string
`message.phase`	string
`message_id`	string	-
`thread`	string	-
`tool_calls`	array	-
`tool_calls[].id`	string
`tool_calls[].name`	string
`tool_calls[].input`	string
`tool_results`	array	-
`tool_results[].tool_call_id`	string
`tool_results[].name`	string
`tool_results[].content`	string
`tool_results[].is_error`	bool
`tool_results[].attachment_ids`	string
`thread_token_count`	integer	-
`message_count`	integer	-

​Common Node Fields

​Thread Configuration

​Conditional Execution

​Automatic Message Saving

​Action

​Inputs

​Outputs

​Agent

​Inputs

​Outputs

​Approval

​Inputs

​Outputs

​Ask Question

​Inputs

​Outputs

​Call LLM

​Inputs

​Outputs

​Compact

​Inputs

​Outputs

​Compute

​Inputs

​Create Worktree

​Inputs

​Outputs

​Join

​Outputs

​Loop

​Inputs

​Outputs

​Publish Branch

​Inputs

​Outputs

​Router

​Inputs

​Outputs

​Run Command

​Inputs

​Outputs

​Run LLM Tool Calls

​Inputs

​Outputs

​Run Tool

​Inputs

​Outputs

​Save Message

​Inputs

​Outputs

Common Node Fields

Thread Configuration

Conditional Execution

Automatic Message Saving

Action

Inputs

Outputs

Agent

Inputs

Outputs

Approval

Inputs

Outputs

Ask Question

Inputs

Outputs

Call LLM

Inputs

Outputs

Compact

Inputs

Outputs

Compute

Inputs

Create Worktree

Inputs

Outputs

Join

Outputs

Loop

Inputs

Outputs

Publish Branch

Inputs

Outputs

Router

Inputs

Outputs

Run Command

Inputs

Outputs

Run LLM Tool Calls

Inputs

Outputs

Run Tool

Inputs

Outputs

Save Message

Inputs

Outputs