[KaiEureka]

Adds manage_workflow for Team members that write their own Workflow source and execute it in their current environment. The tool provides launch, status, wait, stop and result operations, owns child lifetime and reconciles usage with the parent's available budget. Registered run_workflow calls retain their existing interface.

A private owned bridge connects the member's Python process to the configured host model and environment providers. This allows a member in a network-isolated container to run a new Workflow with independently isolated workers, while the host retains Docker control. Model configuration travels through stdin. Native child session records and results are saved before temporary control files and environments are reclaimed.

The implementation reuses the existing Environment, Workflow runtime, model client, request admission, cancellation and environment-pool operations. Independent review repaired internal structured-output permissions, shared budget reservations, native model metadata forwarding, concurrent file-edit ownership and cancellation accounting. Child process cancellation is separate from result and artifact persistence. Pending requests use a small index and separate payload files so several ordinary prompts do not exhaust the environment's per-file read limit.

Functional validation uses real Docker containers and loopback model servers, including completion, cancellation, late usage, model identity and retained artifacts. The final production tree passed all 13 native Docker and bridge checks, and the 69 applicable compatibility and authored-run checks passed. The default invocation skips one explicitly opt-in Docker case, which was exercised with the Docker environment enabled. The offline example also ran from a built wheel outside the checkout. Ruff, import-linter, deptry, translation checks and the proposed-commit secret scan pass. The final merge review also exercised native completion and cancellation while an unrelated container was removed. Cleanup assertions follow this run's owned container IDs, preserving leak detection alongside other Docker consumers. All final-head GitHub checks pass across Python 3.10–3.14. The Python 3.12 suite completed 9,960 passing tests with 53 skips.

An executable offline example is included at examples/member-authored-workflow/. This is the fourth part of the runtime feature series, building on #243.