mr-fang-vlsi/eda-infra-maintainer

EDA Infra Maintainer

When to use

Use this skill when the user asks to:

1. improve/repair infrastructure (knowledge_base, tool_registry, skills, agent policy),
2. harden governance and reliability checks,
3. add new infrastructure capability for future research loops,
4. maintain full-chain research workflow infrastructure (init_research_chain, research_chain_guard, chain templates/skills).

Use eda-artifact-hygiene-maintainer instead when the task is primarily artifact cleanup:

• merge duplicated KB/tool/log artifacts,
• archive or delete stale artifacts,
• normalize inaccurate naming.

Scope Boundary

This skill owns infrastructure governance and integrity. It does not own domain experiment conclusions (timing/power/route/model physics).

This skill is the direct workflow owner for infrastructure maintenance and development tasks. workflow-scoped-execution is not the default wrapper here; use it only when infrastructure work explicitly includes one bounded governed execution stage.

Expected Downstream Consumers

Typical consumers include:

• workflow-router when routing exposes a structural gap,
• eda-context-accessor when KB feedback requires infra maintenance,
• execution and theory skills when governance, manifest, or policy drift must be repaired,
• eda-artifact-hygiene-maintainer when cleanup surfaces a policy-level inconsistency.

Knowledge And Tool Interaction

1. Treat AGENTS.md, the knowledge base, the tool registry, and the skill manifest as governed infrastructure state that must stay mutually consistent.
2. Use eda-context-accessor when infrastructure work needs shared KB/tool retrieval artifacts that downstream skills should reuse.
3. Treat non-none kb_feedback_decision from utility or execution skills as an actionable maintenance input and classify it as maintain, update, overturn, or add work.
4. Run guard/audit tools before and after changes, then write the resulting artifact paths into the maintenance report.
5. When repeated script-writing lessons appear across wrappers, parsers, or helpers, capture them through eda-script-pattern-curator instead of leaving them as one-off maintenance notes.
6. When introducing or restructuring a skill, update the skill-adoption monitor artifacts so the later validation window has concrete usage evidence instead of relying on memory.
7. When append-only logs become long and low-density, produce a rollup note that extracts candidate experience entries and candidate principles instead of relying on future readers to scan the raw log.

Inputs

Provide or derive:

1. maintenance mode (use, maintain, or develop),
2. touched infra surfaces (AGENTS.md, skills, docs/knowledge_base, docs/tool_registry, scripts, manifests, or routing),
3. required guard/audit scope,
4. expected artifact outputs,
5. whether any bounded execution stage must be delegated to workflow-scoped-execution.

Outputs

Return or update:

• guard/audit artifacts under slurm_logs/00_meta/,
• touched governance/skill/KB/tool files,
• maintenance report with rollback trigger,
• explicit reference-topology decision for every touched skill,
• explicit statement of whether workflow-scoped-execution was not_used or delegated_stage.

Hard rules

1. Do not mix infrastructure governance with domain result claims.
2. Do not change comparison policy locks silently.
3. If critical guard checks fail, block promotion and fix first.
4. If utility logic is duplicated across multiple skills, centralize it instead of patching the same interaction pattern repeatedly.
5. Do not route routine infrastructure maintenance through workflow-scoped-execution unless the task truly contains a governed execution substep.
6. Treat workflow-owner skills as stability-sensitive control-plane surfaces; when users ask for new behavior, first try to realize it through lower-layer specialist skills, utility skills, or reusable tools before editing the owner skill itself.

Operational References

1. Load references/maintenance-checklist.md when executing any infra maintenance change and you need the mandatory pre/post sequence.
2. Load references/workflow-owner-usage.md when deciding whether this skill should stay the direct workflow owner or delegate one bounded execution stage to workflow-scoped-execution.
3. Load references/baseline-guard-and-audit.md when selecting which guard, audit, and tool-query commands to run before or after an infra update.
4. Load references/skill-reference-topology-policy.md when a change touches SKILL.md or references/*.md layout and you need a merge-vs-split decision.
5. Load references/skill-type-patterns.md when classifying a touched skill as theory-analysis, execution, or utility.
6. Load references/utility-skill-maintenance-pattern.md when the task is specifically reorganizing utility skills or horizontal capability boundaries.
7. Load references/architecture-change-validation.md when the task changes skill architecture, routing, or workflow ownership and you need to record expected benefits plus a later validation plan.
8. Load references/infrastructure-reporting.md when writing the change summary, rollback trigger, and artifact-path report.
9. Use scripts/common/skill_adoption_monitor.py plus the skill-adoption ledger/report when structural skill changes need later usage review.
10. Use scripts/common/maintenance_log_rollup.py and docs/knowledge_base/108_LONG_LOG_ROLLUP_AND_PRINCIPLE_EXTRACTION_WORKFLOW_20260307.md when maintenance logs or similar append-only ledgers become too long to serve as efficient retrieval artifacts.