Roder
Capabilities

Dynamic workflows

Dynamic workflows let Roder coordinate larger jobs without stuffing every child-agent detail into the lead conversation. A workflow is a generated, built-in, user, workspace, or extension-provided script that drafts phases, estimates child-agent cost, requests approval, and then runs bounded agents through the normal runtime.

Plural namespace. Dynamic workflow runs use workflows/*. The older singular workflow/* methods still describe imported workflow files and should not be treated as the same protocol surface.

How workflows start

  • A substantive prompt that includes workflow can trigger a planning hint in the TUI.
  • Saved .workflow.js files can live in .agents/workflows/ or ~/.roder/workflows/.
  • The built-in /deep-research command runs through the same dynamic workflow runner.
  • Clients can call workflows/plan and then workflows/approve.
  • /effort ultracode can opt into automatic workflow planning when configured.

Approval and safety

First runs show the planned phases, child-agent estimate, scope, limits, and script preview before execution. Approval decisions are runOnce, alwaysForScriptAndWorkspace, or deny.

Workflow scripts run inside a constrained JavaScript host. They can use workflow APIs such as ctx.agents, ctx.phase, ctx.checkpoint, and ctx.report, but direct filesystem, shell, network, secret, module, and MCP access stays outside the script. Child agents perform actual work through the same Roder policy and tool gates as any other turn.

Runtime controls

The TUI command /workflows and the app-server protocol expose list, inspect, pause, resume, stop, restart-agent, save, and script-management operations. Run summaries carry status, phase counts, child-agent counts, failures, concurrency peak, elapsed time, token usage when available, and report previews. 

{"jsonrpc":"2.0","id":1,"method":"workflows/plan","params":{"threadId":"THREAD_ID","prompt":"workflow: audit auth changes"}}
{"jsonrpc":"2.0","id":2,"method":"workflows/approve","params":{"runId":"RUN_ID","decision":"runOnce"}}
{"jsonrpc":"2.0","id":3,"method":"workflows/list","params":{"limit":10}}
{"jsonrpc":"2.0","id":4,"method":"workflows/get","params":{"runId":"RUN_ID","includeAgents":true}}

Configuration

[dynamic_workflows]
enabled = true
trigger_word_enabled = true
auto_with_ultracode = true
max_concurrent_agents = 16
max_agents_per_run = 1000
workspace_workflows_dir = ".agents/workflows"
user_workflows_dir = "~/.roder/workflows"

[dynamic_workflows.approval]
require_approval = true
allow_always_for_script_and_workspace = true

Live provider and live web-search checks remain explicit. Normal workflow tests and fixtures are offline unless RODER_DYNAMIC_WORKFLOWS_LIVE or RODER_DEEP_RESEARCH_LIVE is set.