plan-grill

name	plan-grill
description	Stress-test vague, high-impact, research-facing, advisor-facing, or cross-surface ARIA-NBV decisions before implementation, including theory-rich or elaborate Plan Mode option analysis when requested.
metadata	{"mode":"router","not_when":["a concrete failing command, traceback, or metric owns the task","the edit is already localized and low impact","the user asks for code review of concrete diffs"],"handoff_to":["diagnose-aria for concrete failures","aria-nbv-context when the affected surface is unknown","aria-litkg-memory for KG-backed retrieval or claim checks","code-review-aria-nbv for concrete diff review","docs-curator for public narrative edits after the decision"],"evidence_required":["source-order owner for the decision","success criteria, in/out of scope, and deferred decisions","claim-check output for advisor-facing claims","source ladder and explicit claim strength for theory-rich mode"],"applies_to":["**"],"triggers":["advisor-facing decision","thesis scope","high-impact refactor","scaffold ownership","theory-rich plan","conceptual plan","--conceptual","elaborate plan","rich context","option tradeoffs","literature-grounded plan"],"must_read":[".agents/references/source_order.md","docs/contents/thesis/roadmap.qmd","docs/contents/thesis/questions.qmd",".agents/memory/state/PROJECT_STATE.md","references/plan-mode-theory-patterns.md when using theory-rich or elaborate modifiers"],"canonical_sources":[".agents/references/source_order.md#role-split","docs/contents/thesis/roadmap.qmd","docs/contents/thesis/questions.qmd",".agents/memory/state/PROJECT_STATE.md",".agents/skills/plan-grill/references/plan-mode-theory-patterns.md"],"context7_refs":["/pytorch/pytorch","/facebookresearch/pytorch3d","/websites/typst_app","/websites/quarto"],"literature_refs":["docs/contents/literature/index.qmd","finite-candidate-rl","quality-driven-rri","egocentric-aria-substrate"],"tool_refs":["mcp__MCP_DOCKER.resolve_library_id","mcp__MCP_DOCKER.get_library_docs","mcp__code_index.search_code_advanced"],"verification":["decision-complete plan with assumptions and deferred decisions"]}

Plan Grill

When To Use

Use this skill before implementing ambiguous or high-impact work:

• thesis scope, advisor-facing claims, or roadmap changes
• entity-aware RRI, target-conditioned VIN, rollout/RL, or simulator choices
• docs/public-internal partitioning, scaffold changes, or KG architecture
• broad cleanup where ownership, compatibility, or evidence bar is unclear

Grounding

Before asking the user, resolve discoverable facts from .agents/references/source_order.md and the owning source for the decision. Use docs/typst/shared/glossary.typ for overloaded terms and the nearest AGENTS.md for touched code or docs.

For optional upstream questioning patterns, see references/upstream-mattpocock.md; keep ARIA source-order owners canonical.

Plan-Mode Modifiers

Keep the default path concise unless one of these modifiers applies.

• elaborate: when the user asks for elaboration or option context, explain the practical meaning, pros, cons, and recommended default for each material answer option before asking the next question. Do not perform a literature sweep unless a claim is research-facing or high impact.
• theory-rich: when the user explicitly asks for theory, rich context, literature/API grounding, diagrams, equations, or advisor-facing rationale, read references/plan-mode-theory-patterns.md. Ground theory in the source ladder, state claim strength, and include option tradeoffs before questions.
• conceptual: when the user asks for --conceptual, conceptual planning, systems thinking, or architectural explanation, start with the system boundary and source-owner model before implementation detail. Name vertical truth owners and horizontal evidence sources, include a Mermaid diagram for non-trivial plans, link implementation-facing Python plans to .agents/references/python_conventions.md and the nearest aria_nbv/**/AGENTS.md, use local literature owners before web search for thesis claims, and use Context7 only for external library/API behavior after local owner inspection. Keep $plan, $ralplan, and $prometheus-strict as workflow owners; this skill is the ARIA sidecar that teaches why the routing matters.
• Use Codex-app-safe Markdown/KaTeX equations when they clarify the decision: write inline equations as \(...\), display equations as $$...$$ with blank lines before and after, avoid $...$ inline math because it may not render reliably in Codex chat, and never put equations intended to render in fenced code blocks. Use fenced mermaid blocks for diagrams, and keep math out of Mermaid labels unless it is plain text. For committed .mmd assets, hand off to aria-nbv-mermaid and validate locally.
• Treat Wikipedia as orientation only. Do not use it as advisor-facing, proposal-critical, or thesis-claim evidence.

Interview Rules

• Ask one material decision at a time.
• State the recommended answer with the tradeoff.
• Under elaborate or theory-rich, explain answer-option tradeoffs before calling request_user_input.
• Challenge overloaded terms against docs/typst/shared/glossary.typ.
• For fuzzy thesis or planning terms, test the plan with three concrete scenarios: one normal case, one boundary case, and one failure case.
• Cross-check claims against code, paper, memory state, and roadmap before accepting them.
• Resolved terminology updates docs/typst/shared/glossary.typ or .agents/memory/state/DECISIONS.md. Do not add a parallel root context file or ADR tree as a second source of truth.
• Distinguish current, planned, scratch, and archive docs.
• Capture durable outcomes through the root AGENTS.md Instruction Capture lanes.

Output

End with a decision-complete plan that names:

• goal and success criteria
• in/out of scope
• public interfaces or docs surfaces affected
• implementation packages
• verification commands
• assumptions and deferred decisions