Glossary

Glossary

The core vocabulary of Cool Workflow, grouped by theme. New to CW? Read the Mental Model first, then use this page as a reference.

The run model

Control-plane — What CW is. It plans, records, verifies, and reports, but delegates execution to your agent. It embeds no model SDK and holds no API key. The model is fuel; CW keeps the books.

Run — A single execution of a workflow, stored durably under .cw/runs/<id>/. Resumable, diffable, replayable. Nothing about a run lives in a hidden database.

Workflow app — A versioned manifest that defines a job: its inputs, phases, tasks, artifacts, and validation. Apps are userland; the runtime is the base system. (e.g. architecture-review, pr-review-fix-ci, research-synthesis, release-cut.)

Phase — A stage within an app. Phases are static by design — fixed phases are what make a run deterministically replayable.

Task — A unit of work inside a phase, packaged for execution.

Dispatch — Turning tasks into manifests for your agent (or an operator) to execute. CW records the state and evidence; the host actually runs the worker.

Worker — One isolated execution of a task. Produces a result envelope plus a transcript, with scoped outputs and failure records.

Result envelope (result.md) — A worker's output together with its provenance — the structured thing CW verifies, not just free text.

Evidence & trust

Evidence Adoption reasoning chain — The record of why a result was adopted or rejected, capturing four things: basis (the references), authority (who decided), rationale (the stated reason), and counterfactual (the alternative that lost, and why).

unexplained — The explicit, fail-closed state a result lands in when its evidence or rationale is missing. CW never fabricates a reason or silently passes.

Verifier gate / verifier-gated commit — The rule that only verified state becomes committed state. Unverified work cannot advance.

Commit / checkpoint — A verified state checkpoint written under .cw/runs/<id>/commits/.

Provenance — Where a result came from: its sources, authority, and decision trail.

Telemetry ledger (telemetry.json) — A hash-chained record of the run's usage. Tamper-evident, and ed25519-signable by the executing agent. cw telemetry verify re-proves it offline.

Audit — Provenance, policy, and decision records under audit/, re-checkable with cw audit verify. See Trust And Audit.

Bundle (*.cwrun.json) — A run sealed into one portable file. cw report verify-bundle re-checks it offline with nothing but the file (and --require-signatures to insist findings are signed).

Multi-agent

Coordinator / Blackboard — The shared coordination store for multi-agent work (topics, messages, artifacts, context, snapshots, decisions), kept under .cw/runs/<id>/blackboard/.

Role / group / membership — Multi-agent work modeled as a process table: who is who, which group they belong to, and their lifecycle state.

Topology — A reusable multi-agent recipe:

Topology	Shape
map-reduce	Fan out mapper roles, collect evidence, then reduce.
debate	Record opposing claims, rebuttals, conflicts, and a synthesis.
judge-panel	Gather independent judge outputs and select with provenance.

Fanout / fanin — Spreading work across workers and folding their outputs back together.

Surfaces & portability

CLI — The cw command: the lowest-common-denominator interface every host has words for.

MCP — The same runtime exposed as JSON-RPC 2.0 tools, so editors and agent hosts (Claude Desktop, Cursor, VS Code) call CW as a tool. CLI and MCP share one registry and are parity-checked.

Manifest — The single source of truth that generates every vendor plugin adapter (Claude, Codex, …). A fail-closed drift check in CI keeps adapters from forking the logic.

Execution backend — The swappable driver that actually runs a worker (node, shell, container, remote). The kernel never learns which one ran a task — that's what keeps CW vendor-neutral.

Sandbox profile — A named read / write / command / network / env policy contract applied to worker execution (e.g. --sandbox readonly).

Operations

Replay / eval harness — Deterministic snapshot → replay → compare → score → gate, run without live agents. Used to gate releases on regression evidence.

--drive / --incremental — The high-level host loop. --incremental reuses every step whose inputs didn't change, so re-runs are fast.

subWorkflow / loop() — Flow composition: a task can run a whole child workflow, and a loop() phase can iterate until a predicate or token budget says stop.

release:check — A non-destructive dry-run gate. It builds, type-checks, tests, validates canonical apps and the golden path, checks fixture and manifest compatibility, and self-dogfoods on this repo — without tagging, pushing, or publishing.

---

See also: Mental Model · Getting Started · Architecture · Commands or API