Skip to content

rai pipeline

Pipeline orchestration in RaiSE runs through the rai-workspace MCP server. There is no rai pipeline CLI command — your AI assistant (Claude Code, Codex, or any MCP-capable agent) calls pipeline MCP tools directly.

See Pipelines concept for how pipelines work, or Pipeline Quickstart for a hands-on tutorial.

MCP Tools

Pipelines are controlled via 8 MCP tools exposed by the rai-workspace server:

Tool Parameters Description
pipeline_start pipeline_name, issue_id Start a new pipeline run (story, epic, bugfix, hotfix)
pipeline_advance run_id, approve, cwd Advance to the next phase after gate approval
pipeline_status run_id Check current phase and run state
pipeline_pause run_id Pause a running pipeline
pipeline_cancel run_id Cancel a pipeline run
pipeline_list List available pipeline definitions
pipeline_runs List active and recent pipeline runs
pipeline_restore run_id Restore a pipeline run after context loss

Claude Code reads MCP configuration from .mcp.json. Codex reads .codex/config.toml, generated by rai init --agent codex. Run rai doctor to verify your agent exposes the pipeline MCP tools.

Available Pipelines

Pipeline Phases Use case
story 8 Feature development (start → design → plan → implement → review → close)
epic 6 Multi-story initiatives
bugfix 10 Tracked bug fixes with root cause analysis
hotfix 4 Minimal emergency fixes

Pipeline sources: built-in (shipped with RaiSE), project (.raise/pipelines/), user (~/.rai/pipelines/).

Typical Flow

Your AI assistant handles pipeline orchestration conversationally:

AI calls pipeline_start("story", "RAISE-1234")

  [1/8] start.......................... DONE
  [2/8] design......................... DONE
        → HITL gate: "Approve design?"

  You: "approved"

  AI calls pipeline_advance(run_id, approve=true)

  [3/8] plan........................... DONE
  [4/8] implement...................... DONE
  ...

You never call MCP tools directly — the AI chains the skills and manages state. Your interaction is conversational: review output, approve at gates, provide guidance when asked.

Delegation Levels

Your delegation level is resolved from your developer profile (~/.rai/developer.yaml) and controls HITL gate behavior:

Level Gate behavior ShuHaRi
review Pauses at every HITL gate Shu
notify Proceeds, notifies you Ha
auto Proceeds silently Ri

Gates marked mandatory: true always pause, regardless of delegation level.

State Persistence

Pipeline state is stored in SQLite (~/.rai/raise.db). Runs survive:

  • Terminal restarts
  • Context compaction
  • MCP server restarts

Use the pipeline_restore MCP tool to recover state after context loss.