Imported from seonghann/research-pipeline (codex/skills/source-command-prereport/SKILL.md). Install upstream with npx skills add seonghann/research-pipeline --skill source-command-prereport. Copyright stays with the author.

Codex execution conventions

Use the available native subagent tools for delegation. Role names below identify custom agents installed in ~/.codex/agents/ (or $CODEX_HOME/agents/). Pass the role's instructions, research context, file paths, previous results and constraints. Delegation brief(...) examples are explanatory prompt text, not a tool or command. If this client cannot select a custom role, load the matching playbook from references/agents/ in this skill and pass it to an available subagent. If delegation is unavailable, perform the role inline and disclose that fallback. Keep role scope and checkpoints. Do not claim a subagent ran when it did not. Independent subtasks may run concurrently; dependent phases remain sequential. Inherit the user's model. Read/Write/Edit/Bash/Grep/Glob are capability descriptions: use the actual available file, shell and search tools. Never invoke a nonexistent Codex tool.

Use $source-command-NAME for these skills. Legacy /NAME labels below describe the research workflow, not registered CLI slash commands. /status in Codex is a built-in usage command; use $source-command-status for research status. Honor existing user authorization; do not request the same approval again. Do not undo work or delete files on a rejected checkpoint. Preserve and report state. Stage only explicit task-owned paths. Push only when the user explicitly requests it.

Role playbooks

Read only the role needed for the current phase.

• prereport

/prereport — Generate Pre-report

Formalize a discussed problem into a structured pre-report.

Usage

/prereport <task description>

Prerequisite

This command should come AFTER the discuss phase — the user and Codex have a shared understanding of the problem, approach, and success criteria.

If the conversation hasn't reached this point yet, say: "아직 논의가 충분하지 않은 것 같습니다. [구체적으로 무엇이 불명확한지] 먼저 정리할까요?"

Behavior

Delegate to prereport agent. The agent has no conversation context, so you MUST enrich the Task prompt with a summary of the discuss phase.

Task prompt must include:

• Task description (what to implement/investigate)
• Discussion summary: problem definition agreed upon, why it matters
• Relevant file paths discussed (src/, analyze/, obsidian notes)
• Hypotheses considered and their status (채택/기각/검증중)
• Success criteria discussed
• Any constraints or concerns raised during discussion

Example:

Delegation brief("Generate pre-report: attention layer의 NaN 문제 해결.
  논의 요약: training 3 epoch 후 loss NaN 발생. padding token의
  all-zero input → softmax uniform → downstream zero-division이 원인으로 의심.
  관련 파일: src/model/attention.py, analyze/260208_softmax_zero.py.
  접근 방향: attention mask를 -inf로 적용하는 방식.
  성공 기준: 10 epoch 학습 완료 without NaN, 기존 test 통과.
  주의: inference pipeline의 mask format 호환성 확인 필요.")

The agent will:

1. Read docs/obsidian/ for research-level context
2. Read recent docs/progress/*.md (last 3-5)
3. Read relevant analyze/*.py scripts
4. Scan affected src/ modules
5. Combine file-based context with the discussion summary from the prompt
6. Write pre-report to docs/progress/YYMMDD_<slug>.md

After the agent finishes, display the pre-report summary and ask for approval.

After Approval

Once the researcher approves the pre-report, the natural next step is analyze phase — write prototype scripts to validate the approach.

Suggest: "Pre-report 승인 완료. Analyze phase로 진행할까요? 먼저 검증할 항목: [list from plan]"

Notes

• This is the standalone version of Phase 1 in /go.
• Use this when discuss has concluded and you want to formalize the plan.
• The pre-report is NOT the implementation. No code changes should happen.
• After prereport, the user may proceed with /analyze commands individually or use /go to run the remaining phases automatically.