菜单
Document codebase as-is via parallel sub-agents. Use when researching how code works, mapping architecture, or answering questions about existing systems. Auto-detects thoughts/ directory — if present, integrates historical context and syncs; if absent, skips those steps.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | research-codebase |
| description | Document codebase as-is via parallel sub-agents. Use when researching how code works, mapping architecture, or answering questions about existing systems. Auto-detects thoughts/ directory — if present, integrates historical context and syncs; if absent, skips those steps. |
| allowed-tools | Read, Grep, Glob, Bash, Agent, mcp__plugin_claude-mem_mcp-search__smart_search, mcp__plugin_claude-mem_mcp-search__smart_outline, mcp__plugin_claude-mem_mcp-search__smart_unfold |
Comprehensive codebase research via parallel sub-agents. Document what EXISTS — no improvements, no critique.
When research-codebase is invoked from a planning flow such as
ask-codebase-questions, design-concept, structure-outline,
create-plan, or research-and-plan, keep the research context as neutral as
possible:
This prevents the research step from confirming the requested design too early.
For code file exploration, prefer smart-explore tools (4-18x token savings):
| Tool | Use instead of |
|---|---|
smart_search(query, path) — discover files + symbols in one call | Glob + Grep discovery loop |
smart_outline(file_path) — structural skeleton (~1-2k tokens) | Read on files > 100 lines |
smart_unfold(file_path, symbol_name) — one symbol's full source | Read on large files for one function |
Rule: Use Read only for files < 100 lines or non-code files (markdown, JSON, config).
I'm ready to research the codebase. Please provide your research question or area of interest.
For code files > 100 lines, prefer smart_outline + smart_unfold over Read.
For small files, config, or markdown: use Read directly.
test -d thoughts && echo "has_thoughts" || echo "no_thoughts"
This determines which agents to spawn and whether to sync.
Always spawn:
If thoughts/ exists, also spawn:
Wait for ALL sub-agents to complete before synthesizing.
Before spawning each sub-agent, strip the prompt down to neutral research questions. Do not paste the ticket, PRD, user request, proposed design, target behavior, or implementation idea into the sub-agent prompt.
In Hidden-Requirement Mode, phrase each sub-agent prompt like this:
Research these codebase questions and document current behavior only.
You are intentionally not being given the product requirement or proposed
solution. Do not infer or recommend a change.
- [question]
Output:
- Relevant files and symbols
- Current behavior and data/control flow
- Existing patterns and tests
- Open questions answerable only by the caller
Bad sub-agent prompt:
We need to add [feature/change]. Research how to implement it.
If shared/scripts/spec_metadata.sh is accessible:
bash ../../shared/scripts/spec_metadata.sh
Otherwise:
date && git log --oneline -1 && git branch --show-current
If thoughts/ exists: thoughts/shared/research/YYYY-MM-DD[-ENG-XXXX]-description.md
Otherwise: choose an appropriate path in the project
---
date: [ISO datetime with timezone]
git_commit: [hash]
branch: [branch]
repository: [repo]
topic: "[research topic]"
tags: [research, codebase, component-names]
status: complete
last_updated: [YYYY-MM-DD]
---
# Research: [Topic]
## Research Question
## Summary
## Detailed Findings
### [Component 1]
- What exists (file:line)
## Code References
- `path/to/file:123` - Description
## Requirement Visibility
[If Hidden-Requirement Mode was used, note that research was question-driven and did not include the proposed solution.]
## Architecture Documentation
## Historical Context (from thoughts/)
[Only if thoughts/ exists]
- `thoughts/shared/something.md` - Historical decision about X
## Open Questions
If humanlayer is available:
humanlayer thoughts sync
Present concise summary with key file references. Ask for follow-up questions.
thoughts/searchable/allison/... → thoughts/allison/... (remove only "searchable/")Create the Questions artifact before codebase research, design, planning, or implementation. Use for feature tickets, architecture changes, unclear code work, CRISPY/QRSPI Questions phase, or when the agent needs to define code-answerable unknowns and human questions before researching or planning.
Create a concise Design Concept artifact after questions and codebase research are available. Use for CRISPY/QRSPI Design phase, architecture alignment, feature design discussion, or turning current-state research into a small human-reviewable proposed shape before detailed planning.
Create an interface-first Structure Outline before detailed planning or implementation. Use for CRISPY/QRSPI Structure Outline phase, vertical planning, method signatures, type contracts, API/component boundaries, or converting an accepted design concept into end-to-end slices with verification checkpoints.
Create comprehensive handoff documents for transferring work to another agent session. Use when the user requests to create a handoff, end a session, or document work-in-progress for future continuation.
Create detailed implementation plans through interactive research and iteration. Use when starting a new feature or ticket that needs a comprehensive technical plan. Saves plans to the repo-appropriate plans location and emits Personal Knowledge events when applicable.
個人腦力激盪教練。用結構化技巧逼出你自己的想法，而非替你產生點子。 僅在使用者明確說出「腦力激盪」或「brainstorm」時觸發，其他模糊說法（幫我想想、 想不出辦法、有什麼可能）不觸發。 角色：Claude 是 coach，用 probing question 引導，不是 idea generator。