-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Cool Workflow, or CW, is an independent Agent Workflow SDK by COOLWHITE LLC.
Its goal is to turn agent work from one-off prompt improvisation into an inspectable workflow runtime:
define -> dispatch -> observe -> verify -> commit -> report
CW is a small workflow kernel for coding-agent-era work. It is CLI-backed, MCP-first, skill-packaged, and runtime-compatible with Node-compatible TypeScript output.
- Origin Story
- Architecture Principles
- Runtime Contract
- CLI and MCP Surface
- Workflow Apps
- Multi-Agent Topologies
- Distribution Strategy
- CLI MCP Parity
- Run Registry Control Plane
- Execution Backends
- Web Desktop Workbench
- Observability Cost Accounting
- Team Collaboration
- Release Tooling
- Roadmap
CW records the agent loop explicitly:
interpret -> act -> observe -> adjust -> checkpoint
That loop is reflected in durable run state, task records, dispatch manifests, worker results, verifier decisions, commit snapshots, reports, and audit records.
- Workflow App SDK for versioned, validated, reusable workflow apps.
- CLI runtime for planning, dispatching, result recording, scoring, committing, reporting, scheduling, and routine triggers.
- MCP JSON-RPC surface for host integrations.
- Sandbox Profiles for named worker policy contracts.
- Worker isolation and candidate scoring.
- Verifier-gated commits and explicit checkpoints.
- Operator UX for human-readable status, graph, report, worker, candidate, feedback, commit, topology, multi-agent, blackboard, coordinator, and trust summaries.
- Coordinator / Blackboard for shared topics, messages, context frames, artifact refs, snapshots, decisions, conflicts, and fanin evidence.
- Multi-Agent Runtime Core for runs, roles, groups, memberships, fanout, fanin, and lifecycle state.
- Multi-Agent Topologies for official
map-reduce,debate, andjudge-panelrecipes. - Multi-Agent trust / policy / audit, operator UX, and a deterministic eval / replay harness with regression gates.
- State Explosion Management: derived, fingerprinted summaries and compact graph views for large runs.
- Evidence Adoption Reasoning Chain: records why a result was adopted or rejected (basis, authority, rationale, counterfactual), fail-closed.
- CLI ↔ MCP Parity: one capability registry as the single source of truth for both surfaces, enforced fail-closed.
- Run Registry / Control Plane: cross-repo search, resume, queue, archive, history, and failed-run rerun over a rebuildable, fingerprinted index.
- Execution Backends: a pluggable driver layer (
node/bun/shell/container/remote/ci) with identical evidence envelopes and fail-closed sandbox attestation. - Cross-vendor distribution: one source-of-truth manifest generates every vendor plugin manifest (Claude/Codex/…) with a fail-closed drift gate.
- Release and migration discipline with fixture compatibility, version sync, and dry-run release checks.
- Main repo: https://github.com/coo1white/cool-workflow
- Package root:
plugins/cool-workflow/ - Runtime source:
plugins/cool-workflow/src/ - Committed runtime output:
plugins/cool-workflow/dist/ - License: BSD-2-Clause
This wiki was organized from local Obsidian notes under:
/Users/lukebai/Documents/Nick/Cool Workflow/
It was first organized on 2026-06-07 and last reconciled with the current repository docs and the v0.1.29 code surface on 2026-06-08 (38 source modules · 34 docs · 29 smoke tests · 6 workflow apps · 30 releases).
Organized from local Obsidian notes and reconciled with the current
coo1white/cool-workflow repository state.
Start here
Go deeper
- Workflow Apps
- Architecture
- Trust And Audit
- Recovery And Restore
- Commands or API
- MCP And Manifests
- Operations
- FAQ
Source docs