How it works
Pipeline Kit is two kits. The diagram below is kit mode (the default). Orchestrator mode uses the same workflow names and the same features/{slug}/ folder, but a Python engine instead of the markdown loader. Compare them: Two kits.
Portable pack, thin IDE
All process files sit in .pipeline/. The IDE folder holds only run-workflow (and optional hooks). Cursor, Claude Code, and GitHub each get the same pack; only the adapter path changes (--ide cursor | claude-code | github | none).
Workflow as the unit of scale
A workflow is a named procedure (ask, feature-development, jira-bug, …) with a chain and a file allowlist. Adding a customer-specific process means adding a workflow — not cloning the kit.
Allowlist per step
The loader writes allowed_reads for the current specialist. BA does not ingest the deploy runbook; the developer does not ingest Jira intake. That is what keeps a large pack cheap enough to reuse across accounts.
Config is the engagement overlay
Chains, tracker on/off, path-based verify reminders, and deploy target names live in .pipeline/config.json. Skills stay free of customer folder names and site URLs. The next account changes config (and the local-deploy script), not the orchestration skill.
Install scopes
--project (init) materializes <repo>/.pipeline. --user (setup) materializes ~/.pipeline. At runtime the project pack wins; otherwise the user pack is used. Artifacts always land in the current repo’s features/.
Architects can later layer org template → user → project without changing the runtime model:
org template → team defaults → ~/.pipeline → <repo>/.pipeline (wins)
Two kits
Kit mode (--mode kit, the default) copies the markdown pack to .pipeline/ and runs specialists through the loader.
Orchestrator mode (--mode orchestrator) keeps the chain in the Python engine. init does not copy agents/, skills/, loader/, or workflows/. You pass the user ask with --request. Chat is not inherited.
Adding a process is an extension in the matching kit. Full comparison: Two kits — pack and orchestrator.
Receptionist, not a monolith
AGENTS.md stays a routing table. The agent does not load the pack until run-workflow runs the loader. Trivia and off-repo asks skip the pipeline entirely.