eda-loop
Execute one scoped EDA task under governed gates: run bootstrap checks, route to minimal specialist skills, validate artifacts, and return execution evidence.
EDA Loop
Role Boundary
eda-loop is the execution orchestrator for a scoped task.
Global entry, cross-task governance, and recursion policy belong to AGENTS.md (repo root).
Use this skill when you need disciplined execution with explicit artifacts.
Inputs
Provide or derive:
- scoped objective,
- target design/config set,
- required output artifacts,
- comparison-policy lock (if claim/comparison is requested).
Step 1: Bootstrap gates
- Enter repo root and ensure task brief exists.
- Run gate before substantive actions:
bash scripts/common/knowledge_gate.sh --scope <scope_name> --task-brief <task_brief_md>
- Query tool catalog before adding scripts:
python3 scripts/common/tool_catalog.py query <keyword1> <keyword2>
- Enforce tri-source evidence before non-trivial conclusions:
- local logs/data,
- local KB history,
- internet/primary-source evidence when assumptions are unstable.
- For infrastructure maintain/develop tasks, run stack guard:
python3 scripts/common/infra_stack_guard.py --out-prefix slurm_logs/00_meta/infra_stack_guard_<tag>
Step 2: Skill routing (minimal set)
Select the smallest set of specialist skills that covers the scoped objective.
Load references/skill-routing.md only as needed.
Primary local skills:
bspdn-goal-drivereda-research-chaineda-knowledge-explorereda-idea-debate-labeda-hypothesis-experiment-designereda-method-implementereda-knowledge-gate-maintainereda-infra-maintainereda-preflight-reflecteda-retroeda-stage-checkpoint-goldengt3-backside-route-policydelay-model-gate-evaluatorbscost-netbscost-theory-opteda-paper-fetcheda-pdf-local-summary
System skills:
skill-creator(create/update skills)skill-installer(install external skills)
For full-chain research requests, prefer eda-research-chain as the primary workflow skill.
Step 3: Execute under hard gates
- Announce selected skills and first action.
- Execute fully (do not stop at plan-only unless asked).
- For any new experiment submission/comparison request, run
eda-preflight-reflectfirst and produce a reflection artifact before launching jobs. - For route/CTS submissions, run unified PDK/flow preflight (
scripts/debug/pdk_flow_preflight.py) and block submission on FAIL. - Keep artifact paths explicit (
monitor.md,summary.md,manifest.tsv, etc.). - After batch completion, run execution-contract validation (
scripts/debug/validate_execution_contract.py) and mark FAIL rows as non-promotable evidence. - Ingest manifest rows into experiment memory (
scripts/common/experiment_memory.py) and produce constrained follow-up candidates (scripts/debug/propose_constrained_experiments.py) when optimization continues. - If a gate fails, fix cause before proceeding.
- Comparison policy lock:
- For algorithmic claims, enforce primary baseline as
vanilla_replace. - If Innovus is used, keep it as secondary realizability validation only.
- If only Innovus baseline exists, mark algorithmic conclusion as invalid and stop promotion.
Step 4: Closeout artifacts
- Update maintenance log (
slurm_logs/00_meta/knowledge_tool_maintenance_log.md) for substantive changes. - If a skill is modified in this interaction, validate it:
python3 /home/grads/d/donghao/.codex/skills/.system/skill-creator/scripts/quick_validate.py <skill_path>
- Append hypothesis-validation-conclusion entry when a hypothesis was tested:
docs/knowledge_base/90_HYPOTHESIS_VALIDATION_LOG.md
- For infrastructure maintain/develop tasks, include both artifacts:
infra_stack_guardsummaryskill_system_auditsummary
Hard-stop conditions
Do not close task if any condition is true:
- No maintenance-log update for this interaction.
- No tool-catalog query evidence for non-trivial changes.
- Artifacts created but not linked in final summary.
- Skill modified but not validated.
- Algorithmic claim made without
vanilla_replaceprimary baseline. - Route/CTS batch referenced without unified preflight artifact.
- Route-level conclusion given while execution-contract artifact contains unresolved FAIL rows.
Out of scope for this skill
- Defining global entry policy.
- Defining recursion policy across interactions.
- Declaring global self-update policy for the whole skill system.