Skip to content
CKS Documentation

Pipeline Agent (multi‑agent orchestration)

cks-pipeline-agent runs a configurable pipeline of AgentStep implementations coordinated by CKSAgentOrchestrator (ADR-007). Each step writes its result as a knowledge object with provenance and a semantic edge from the previous step, while the orchestrator publishes AgentStepStarted / AgentStepCompleted events.

Quick start

Section titled “Quick start”
Terminal window
CKS_MCP_DB_PATH=~/.cks-mcp/cks_mcp.db cks-pipeline-agent

Built‑in steps

Section titled “Built‑in steps”
  • ResearcherStep — researches a topic and produces a ResearchFinding node.
  • ReviewerStep — reviews another step’s finding and produces a Review node (with reviews edge). Idempotent by content hash.

Configuration

Section titled “Configuration”
Variable Default Description
CKS_MCP_DB_PATH ~/.cks-mcp/cks_mcp.db Shared database path
CKS_PIPELINE_POLL_INTERVAL 5 Outbox poll interval (seconds)
CKS_PIPELINE_MAX_RETRIES 5 Max retries before dead‑lettering
CKS_PIPELINE_RESEARCHER_MAX_TOKENS 512 Token limit for ResearcherStep LLM calls
CKS_PIPELINE_REVIEWER_MAX_TOKENS 256 Token limit for ReviewerStep LLM calls

Architecture

Section titled “Architecture”

The orchestrator runs steps sequentially or concurrently, using the persistent outbox (cks_outbox_tasks) for claiming tasks and the CRDT layer for knowledge object sharing. Each step is an AgentStep implementation that:

  • Checks idempotency via a content hash stored in a transition_log.
  • Calls an LLM (via the existing auto/ollama/anthropic dispatch).
  • Commits its finding/verdict + a semantic edge in one atomic evolve_knowledge call.

Extending

Section titled “Extending”

Implement the AgentStep protocol (orchestrator.py) to add custom steps. Place them in src/cks_mcp/pipeline/ and register them in the pipeline configuration.

See ADR-007 CKSAgentOrchestrator for the full design rationale.