examples

Working, copyable Boardwalk workflow templates. Each one is a standalone project: copy it out (or boardwalk init <dir> --template <name>), set its secrets, run.

Every template is a parity test. It must run unmodified under the self-hosted Boardwalk engine and the hosted Boardwalk platform — the CI harness in this repo enforces it. A template that needs per-engine edits is rejected.

Templates

Start with the primitives — one feature each — then compose them into the multi-agent patterns.

Primitives

Template	What it demonstrates
hello-routine	The minimal workflow: manual trigger, one agent() call, one returned output
claude-code-cron	Running an existing claude -p command-line script on a schedule, unchanged
morning-digest	Cron + secrets.get + agent summarization
webhook-responder	Webhook trigger + a typed run(input) + conditional triage
pipeline	workflows.call composition: a parent durably fanning work through a child
long-watch	sleep({ until }) + budget caps over a long-running watch

Patterns

The orchestration shapes behind big multi-agent jobs — bug hunts, migrations, deep research, triage. Each is one composable idea you can drop into a larger program: spawn separate agents with their own context windows so the work gets finished, verified, and kept on-goal.

Template	What it demonstrates
classify-and-act	A classifier agent labels the task; code routes it to a tailored handler
fan-out-judge	Fan out + synthesize: parallel() drafts + a schema-validated judge
adversarial-verify	One skeptical agent per claim — verify by trying to refute
generate-and-filter	Brainstorm wide in parallel, dedupe, keep the best by a rubric
tournament	Rank items by agent-judged pairwise comparison (merge sort holds the bracket)
loop-until-done	Loop spawning finders until nothing new turns up
loop-with-verify	Loop until finders run dry, then a separate checker keeps only the verified findings
quarantine-triage	Read untrusted content with no-privilege agents; act in trusted code
code-review	Review a diff with a reusable reviewer skill loaded on demand (progressive disclosure)

Conventions (every template)

• Two files are the workflow. workflow.jsonc is the deployment descriptor (slug, triggers, permissions, budget — the policy the control plane reads as data), and src/index.ts exports the program: export default async function run(input, context). The trigger's payload is input, and whatever run returns is the run's output.
• Native types are the I/O contract. Templates type input and the return value where the workflow has a shape — the deploy derives their schemas for the dashboard's run form and for callers. A bare run(input) (no annotation) is the untyped floor.
• agent() calls name no model by default — the default boardwalk provider routes automatically on every engine; a comment shows the explicit-model form. BYO keys are an explicit provider, never a fallback.
• Every secret a template needs is declared in the descriptor's permissions.secrets and in the registry secrets list; you set the value with boardwalk secrets set NAME --org <your-org>. No real secrets, ever.
• Minimal dependencies — @boardwalk-labs/workflow and the platform, nothing else unless the template is about a dependency.

Running the harness

node harness/run.mjs                 # uses `boardwalk` from PATH
BOARDWALK_CLI="node ../cli/bin/boardwalk.js" node harness/run.mjs

It validates every template (boardwalk check), checks registry/filesystem consistency, and executes the agent-free templates end-to-end through the self-hosted server engine (set BOARDWALK_SERVER to the boardwalk-server binary). (Agent templates execute in the scheduled platform battery.)

Contributing a template

A new template needs a primitive or pattern not already covered, the conventions above, and a green harness. Keep each template to one idea; don't bundle unrelated features.

The Boardwalk repos

• boardwalk — the open-source single-node engine: cron scheduling, webhooks, durable runs, run history
• sdk-typescript — @boardwalk-labs/workflow, the TypeScript API a workflow program imports
• cli — boardwalk: scaffold, validate, deploy, run
• plugins — coding-agent skills (Claude Code, Codex, Cursor, OpenClaw, OpenCode) + a control-plane MCP server
• runner — self-hosted runner: your machines execute hosted-scheduled runs

Hosted platform and docs: boardwalk.sh.

License

MIT