---

name: framework-runtime description: "Routes DI-engine task/middleware, Parallel, Supervisor, and modern example workflows." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0 disable-model-invocation: true

Framework runtime

Use this sub-skill for the task/middleware runtime and the modern example recipes that build on it. These workflows are more composable than the legacy serial pipeline family and are the best fit for multi-process message routing.

Owns

• ding.framework.task, Task, Context, OnlineRLContext, OfflineRLContext, and the task event helpers
• Parallel.runner and the Parallel message-routing API
• Supervisor, ChildType, and env/process supervision helpers
• middleware-based example scripts in ding/example/
• runtime helpers such as task.use, task.forward, task.backward, task.wait_for, task.emit, and task.run

Does not own

• CLI/config parsing and config compilation; those live in cli-config
• Legacy serial_pipeline and data-collection helpers; those live in serial-pipelines
• Env-wrapper shape debugging; those live in env-integration

Read this first when the user asks

• how to write or debug a task.start(...) workflow
• how the repo's modern ding/example/*.py scripts are structured
• how to use Parallel.runner or recover from multi-process routing failures
• how Supervisor or EnvSupervisor style child management works
• how to build a custom middleware chain for off-policy, on-policy, offline, or self-play-style workflows

Workflow

1. Start with references/workflows.md to choose the right runtime shape.
2. Read references/api-reference.md for the relevant signatures and routing parameters.
3. Use the bundled smoke scripts when you need a small reproducible check that the installed environment still supports the task or parallel runtime.
4. Send env-shape or config-shape problems to the neighboring sub-skill rather than bloating the runtime section.

Common decision points

• Use task.start(async_mode=False) for the simplest single-process workflow.
• Use task.start(async_mode=True) when the runtime must yield and resume between middleware steps.
• Use Parallel.runner when the recipe needs multiple workers or router-aware cross-process messaging.
• Use Supervisor when the workflow needs to manage child envs/processes with explicit restart/timeout behavior.

Helpful bundle links

• references/workflows.md for the example-family map.
• references/api-reference.md for signatures and parameter meanings.
• references/troubleshooting.md for process, routing, and message-queue failures.
• scripts/task_smoke.py and scripts/parallel_smoke.py for safe runtime checks.