Workflow Scoped Execution

Workflow Scoped Execution

Role Boundary

workflow-scoped-execution is the scoped execution workflow owner.

It owns:

• governed execution for one concrete task,
• submission hygiene,
• artifact validation,
• post-run closeout.

It does not own:

• global entry policy,
• non-execution workflow ownership,
• cross-turn recursion policy,
• global self-update policy for the full skill system.

workflow-scoped-execution should be used only when:

1. it is the selected workflow owner for one bounded execution task, or
2. another workflow owner explicitly delegates one governed execution stage to it.

Use This Skill When

Use workflow-scoped-execution for experiment, debug, flow, baseline, handoff, monitoring, or infrastructure tasks that already have a scoped execution brief and need explicit gates and artifacts.

Inputs

Provide or derive:

1. scoped objective,
2. target design/config set,
3. required output artifacts,
4. comparison-policy lock if a claim/comparison is requested.
5. selected skill set or explicit statement that no additional skill routing is needed.

Knowledge And Tool Interaction

1. If the task needs shared KB retrieval or shared tool-reuse evidence, delegate that work to eda-context-accessor instead of re-specifying retrieval logic locally.
2. If eda-context-accessor returns kb_feedback_decision != none, record that feedback in the execution summary and route KB/infrastructure follow-up to eda-infra-maintainer when the change is not purely local.
3. Keep workflow-scoped-execution focused on execution hygiene:

• run scripts/common/knowledge_gate.sh for governed execution startup,
• update slurm_logs/00_meta/knowledge_tool_maintenance_log.md,
• update docs/knowledge_base/90_HYPOTHESIS_VALIDATION_LOG.md when a hypothesis was actually tested.

4. When a batch completes and the outcome should be reused beyond the immediate turn, delegate log -> result -> conclusion -> experience extraction to eda-experiment-phenomenology-analyst before handing off to control-postrun-retro or a future control-theory-veto.

Execution Rules

1. Announce selected skills and first action.
2. Execute fully unless the user explicitly asks for plan-only behavior.
3. Keep artifact paths explicit.
4. Every experiment submission must create a canonical monitor file immediately; do not leave monitor generation as manual follow-up.
5. Prefer the reusable monitor stack (scripts/debug/progress_monitor.py or scripts/debug/progress_monitor.sh) over ad hoc markdown.
6. Immediately report monitor/manifest/job ids after submission.
7. Do not close a submission task before monitoring visibility exists.
8. For algorithmic claims, enforce vanilla_replace as primary baseline.

Outputs

Return or update:

• canonical monitor path,
• canonical manifest path,
• key artifact paths,
• structured evidence-lift artifact paths when experiment phenomenology extraction was required,
• execution verdict,
• next constrained action.

Operational References

Load only the parts needed for the current task:

1. Load references/workflow-scoped-execution-checklist.md when starting or closing a normal workflow-scoped-execution execution and you need the full bootstrap/pre-submit/post-run checklist.
2. Load references/execution-gates.md when the task involves submission gating, route/CTS preflight, testcase-backed DC/Innovus checks, or root-cause mode for non-convergence.
3. Load references/output-and-stop-policy.md when deciding whether the task is complete, whether reporting is sufficient, or whether a hard-stop condition blocks closeout.
4. Load references/postrun-improvement.md when the current loop exposed a repeated execution gap and you need to decide the smallest acceptable improvement action.