---
name: deep-research-plan
description: Use deep-research-plan for repo-change plans needing downloaded projects, articles, or papers, per-reference analysis, graphify JSON, and stronger-than-huge planning; exclude direct implementation-only, factual/explanation, trivial, or explicit no-plan requests.
---

# Deep Research Plan

Research first; planning and execution still follow `../optim-plans/SKILL.md`.

## Required Flow

1. Read and follow `../optim-plans/SKILL.md`, treating the request as if it named `deep-research-plan`.
2. Inspect the target Git repo read-only and initialize or resume the optim-plans controller before external research.
3. Ensure the root `.gitignore` ignores `refs/`; add that ignore entry only during execution approval if it is missing.
4. Search proactively for related projects, articles, and papers. Prefer `agent-reach` when available. If `agent-reach` is missing, ask exactly one yes/no sentence about installing it locally; on yes, install it first and use it, and on no, use websearch fallback and record the fallback reason.
5. Download at least 3 relevant refs into `./refs/<topic>/` before writing `PLAN_v1.md`. Do not satisfy this with `curl` alone, README-only snapshots, or abstract-only paper notes; keep full local project, article, paper, docs, or artifact content sufficient for deep analysis.
6. Record every downloaded ref through the controller with `deep-record-ref` before relying on it. Record graph output with `deep-record-graph`, analysis artifacts with `deep-record-analysis`, and graphify install refusals with `deep-waiver-question` plus `deep-record-waiver`.
7. For every downloaded ref, generate graphify JSON beside the ref before user adoption questions. If `graphify` is missing, ask exactly one yes/no sentence about installing it locally; on yes, install it first and generate the JSON, and on no, use read tools to deeply inspect the downloaded files and record the nonce-bound graphify waiver.
8. For every ref after analysis, ask at least 3 ref-specific controller-backed choice questions with `deep-adoption-question` about whether or how to use that ref. Base each question on the downloaded content, put the recommended option first, `Other` second-last, and `Auto-complete` last, then record the answer before including the idea in any plan.
9. Before accepting `PLAN_v1.md` as ready for refinement or execution, run `register-plan --path docs/optim-plans/.../PLAN_v1.md --version 1`; treat any deep-research readiness blocker from `status`, `register-plan`, or `prepare-execution` as a hard blocker.
10. Continue as stronger-than-huge planning: at least 10 planning questions, required websearch in brainstorming and refinement, no refinement round or timeout limit, and at most five high-priority comments or questions per refinement round.

## Evidence Rules

- Keep attempted queries, selected refs, download commands/tools, analysis coverage, graphify output paths, backend failures, install refusals, and evidence gaps in the run artifacts.
- Block rather than pad when fewer than 3 credible refs exist, unless the user explicitly narrows or waives the minimum.
- Do not rely only on README files, landing pages, abstracts, summaries, or package metadata when deeper source material is available.