workflow-orchestration

# ADK Workflow Orchestration Use this sub-skill when the task is to build or debug ADK 2.0 `Workflow` graphs, custom `BaseNode`/`FunctionNode` steps, static and dynamic graph routing, joins, parallel workers, retries, human-in-the-loop interruptions, workflow event paths, or resume/checkpoint behavior. ## Route requests - **Create a graph**: Use [Workflow patterns](references/workflow-patterns.md) for `Workflow(edges=...)`, `START`, chain tuples, explicit `Edge`, route maps, `JoinNode`, parallel worker, dynamic fan-out/fan-in, state, and output patterns. - **Check APIs**: Use [API reference](references/api-reference.md) for constructor signatures, node fields, `Context.run_node`, `Event`, `NodeInfo`, `RequestInput`, and `RetryConfig` names. - **Debug execution**: Use [Troubleshooting](references/troubleshooting.md) for deadlocks, runner vs node-runner confusion, rerun/resume surprises, dynamic node dedup, join stalls, route misses, timeouts, and retry expectations. - **Inspect installed package**: Run [inspect_workflow_api.py](scripts/inspect_workflow_api.py) to print the local `google.adk.workflow` signatures and construct a tiny no-LLM function-node workflow object. ## Boundaries - Use this sub-skill for `Workflow`, `BaseNode`, `node`, `FunctionNode`, `JoinNode`, `RetryConfig`, `RequestInput`, `Event`, `NodeInfo`, graph edges, dynamic nodes through `ctx.run_node`, checkpoint/resume semantics, and workflow state/output routing. - Route LLM `Agent`/`LlmAgent` construction, model configuration, instructions, tools, callbacks, and schemas to `agent-construction`. - Route `Runner`, `App`, session service, memory, artifacts, plugins, and persistence lifecycle details to `runtime-services`. - Route repository maintainer test policy and source-code change validation to `repo-development`. ## Safe assumptions - The installed distribution is `google-adk` 2.3.0 with import root `google.adk` and Python 3.10+. - Base install may omit optional extras; workflow construction with pure function nodes does not require cloud credentials, network access, or LLM calls. - `adk` CLI includes `api_server`, `conformance`, `create`, `deploy`, `eval`, `eval_set`, `migrate`, `optimize`, `run`, `test`, and `web`, but CLI/server/deploy usage belongs in the CLI sub-skill. - Future guidance should not depend on source checkout files; copy or recreate needed snippets inside the agent’s working files. ## Quick validation ```bash python scripts/inspect_workflow_api.py --help python scripts/inspect_workflow_api.py ``` The script performs only local imports and object construction; it does not start a runner, call an LLM, access credentials, open a network connection, or write files.