| name | rd-agent |
| description | Start or continue the V3 orchestration loop — intent routing, paused-run detection, stage dispatch. |
| argument-hint | <task description> | --continue | --status |
| allowed-tools | ["Read","Bash","Glob"] |
<codex_skill_adapter>
Invocation
- Claude Code:
rd-agent skill
- Codex:
$rd-agent
- Treat text after skill mention as arguments.
Tool mapping
- AskUserQuestion → request_user_input
</codex_skill_adapter>
Primary orchestration entry for RD-Agent. Routes plain-language intent to
the correct stage skill (rd-propose, rd-code, rd-execute, rd-evaluate) or
detects and continues paused runs. Never executes stage logic directly.
For degraded or blocked rounds, inspect round truth first, apply one recovery action, and continue only after truth is repaired.
Maps to rd_agent.entry.rd_agent.rd_agent.
Required fields
Start path:
titletask_summaryscenario_labelstage_inputs.framing.summarystage_inputs.framing.artifact_ids
Continue path:
- none; the paused-run path is derived from persisted state
Optional fields
initial_branch_labelexecution_modemax_stage_iterationsbranch_hypotheseshypothesis_specsholdout_evaluation_portholdout_split_port
Tool execution context
Use uv run rdagent-tool ... only from the installed standalone V3 runtime bundle root or a checked-out standalone V3 repo root; do not search other repos or HOME for a plausible state directory.
When to route to rd-tool-catalog
Use rd-tool-catalog only as an agent-side escalation path when the agent needs a concrete direct tool in the background. Do not push the operator into manual tool selection for common start or continue flows.
Success contract
rd-agent starts the run or advances the high-level loop, then either keeps ownership on the orchestration surface or tells the operator to route to a stage skill or to rd-tool-catalog.
<execution_context>
@skills/rd-agent/workflows/intent-routing.md
@skills/rd-agent/AUTONOMOUS-WORKFLOW.md
@skills/rd-agent/workflows/start-contract.md
@skills/_shared/references/failure-routing.md
@skills/_shared/references/tool-execution-context.md
@skills/_shared/references/output-conventions.md
</execution_context>
<required_fields>
Start path:
- title (string)
- task_summary (string)
- scenario_label (string)
- stage_inputs.framing.summary (string)
- stage_inputs.framing.artifact_ids (list)
Continue path:
- (none — detected from persisted state)
Optional:
- initial_branch_label
- execution_mode
- max_stage_iterations
- branch_hypotheses
- hypothesis_specs (advanced structured multi-branch path)
- holdout_evaluation_port (required when hypothesis_specs is provided)
- holdout_split_port (optional override for structured holdout splitting)
</required_fields>
Execute intent-routing workflow from @skills/rd-agent/workflows/intent-routing.md.
If start contract needed, also load @skills/rd-agent/workflows/start-contract.md.
<routing_boundary>
- Route to rd-tool-catalog only when the high-level rd-agent boundary is
insufficient and the agent needs a concrete direct tool.
- Route to rd-tool-catalog for inspection-only or primitive-only work that
should not stay on the orchestration surface.
- Do not push the operator into manual tool selection for common start or
continue flows.
</routing_boundary>
