Dispatch is written in Kujo. This tour uses
its bytecode VM and local fixture model, so it needs no provider key or network
request. Run each command from the Dispatch repository root after installing
Kujo (minimum version in kennel.toml).
export DISPATCH_OFFLINE_FIXTURE=true
kujo run dispatch.kujo validate --workflow-file examples/workflows/routed-review.json --json
kujo run dispatch.kujo explain-route --workflow-file examples/workflows/routed-review.json --step-id plan --json
kujo run dispatch.kujo demo "What makes a reliable workflow?" \
--workflow-file examples/workflows/routed-review.json \
--yes --non-interactive --output-root tests/tmp/first-workflow
kujo run dispatch.kujo runs --output-root tests/tmp/first-workflow --jsonThe demo prints a Run ID. Substitute that ID in the next command:
kujo run dispatch.kujo inspect <run-id> --output-root tests/tmp/first-workflow --jsonInspect tests/tmp/first-workflow/<run-id>/trace.md for the step-by-step
timeline, trace.json for machine-readable events, state.json for the
resumable state, and report.md for the output. The plan step uses a fixture
model from a versioned catalog; the finalize step writes the report. Neither
this demo nor its green checks establish live-provider readiness.
The second workflow is intentionally small: examples/workflows/plugin-echo.json
declares echo_tool, the built-in sample plugin provides its handler, and
the CLI allowlist grants only that named tool. Plugins are trusted executable
code; review one before loading it from another source.
kujo run dispatch.kujo demo "Kujo extension example" \
--workflow-file examples/workflows/plugin-echo.json \
--plugin sample --allow-tools echo_tool \
--yes --non-interactive --output-root tests/tmp/plugin-echoThe trace shows the echo tool step completing before the report. Removing
--plugin sample leaves the declared tool unavailable; changing the allowlist
to exclude echo_tool denies it. Both are visible failures, not silent
fallbacks. The handler and payload adapter are in
src/plugins/builtin_plugins.kujo; the policy builder is in
src/core/tool_policy.kujo.
flowchart LR
CLI[Kujo VM runs dispatch.kujo] --> Spec[JSON workflow and model catalog]
Spec --> Runner[src/core/runner.kujo]
Runner --> Agent[src/agents/agent.kujo]
Agent --> Bridge[src/bridge/sdk_adapter.kujo]
Bridge --> SDK[External Kujo AI SDK when live]
Runner --> Policy[Tool policy and optional plugin]
Policy --> Evidence[state, trace, report artifacts]
Agent --> Evidence
Next: author a workflow, inspect architecture and extension points, and read enterprise deployment boundaries before enabling live providers.
For the experimental typed-output, review, Eval and budget path, see the operator rehearsal.