Summary

• Finish the workflow CLI task/resource surfaces for agents, including async-safe workflow tool execution and manager-capable graph edits.
• Run and document the overnight ResearchRubrics RQ1 CLI-specialism hillclimb, with best artifact tests/real_llm/.rollouts/20260426T234424Z-7fc055f5-03c3-4cab-8117-04e844696482/.
• Fix real-LLM harness/runtime blockers uncovered during the experiment: vanilla ResearchRubrics prompt loading, provider env propagation, context import cycles, failed-run artifact dumping, and dynamic task/ready scheduling for added subtasks.

Experiment Result

• Best run: Variant 1f, run ID 7fc055f5-03c3-4cab-8117-04e844696482.
• Evidence: 5/5 root tasks completed, 5/5 evaluations landed, 5 roots created exactly 15 specialist children, and no recursive child creation after the current-task-level prompt fix.
• Caveat: run-level status is still failed because advisory specialist child failures propagate; this is documented as the next design choice.

Test plan

• uv run pytest tests/unit/runtime/test_workflow_service.py tests/unit/cli/test_workflow_cli.py tests/unit/state/test_workflow_cli_tool.py tests/unit/state/test_research_rubrics_workers.py tests/unit/state/test_research_rubrics_benchmark.py tests/unit/runtime/test_import_boundaries.py -q
• Real-LLM artifact runs documented in docs/experiments/rq1-cli-specialism/changelog.md
• Backend harness check for best run: GET /api/test/read/run/7fc055f5-03c3-4cab-8117-04e844696482/state returned 20 graph nodes, 70 mutations, 5 evaluations, and 5 resources