understand-codebase

name	understand-codebase
description	Map a codebase for learning — glossary, architecture, flashcards, and quiz. Produces a PDF guide and local HTML study pack.

Understand Codebase

Consumers: /kon:understand-codebase

Core principles (always): follow skills/core-principles — teach from what the repo actually contains; simplest path through the concepts.

Turn a repository into study materials: key concepts, architecture, flashcards, and a quiz.

Deliverables (after build step)

Output	Contents
understand-guide.pdf	Key concepts + reference code + architecture + FAQ (links in PDF when pandoc)
understand-study.html	Flashcards + quiz with snippets + clickable source links
understand-guide.html	Same as PDF; mermaid + vscode:// / cursor:// links open files in IDE

All files live under sessions/<session-id>/. Orchestrator runs build_understand_codebase.py after Jun finishes.

Pipeline

1. 🎸 Azusa (explore) — read-only codebase map → hook saves understand-explore.md
2. 📚 Jun (author) — reads explore artifact, writes understand-guide.md + understand-study.json
3. Orchestrator — python3 $KON_ROOT/scripts/build_understand_codebase.py --id "$SID"

Azusa — explore mode

Read agents/Azusa.md. Read-only — no Write/Edit on application source.

Pass SESSION_DIR, EXPLORE_FILE: understand-explore.md.

Required sections (hook persists full output)

## Loaded memory entries
…

## Scope
<what was explored; optional user focus from command>

## Code map
| Area | Key paths | One-line role |
|------|-----------|---------------|

## Concepts spotted (raw)
For each term you see in code/docs — name, where found (`path:line`), rough meaning from evidence:

### <Term>
- **Seen in**: `path:line`, …
- **Notes**: …

## Architecture notes (raw)
- **Deployment shape**: single-node / distributed / hybrid — evidence
- **Major components**: …
- **Data flow**: …
- **External deps**: …

## Implementation anchors
Concrete `path:line` pairs worth flashcard/quiz treatment later.

## Gaps
What could not be determined from the repo alone.

Do not invent behaviour — cite paths. Unknown → ## Gaps.

Jun — author mode

Read agents/Jun.md. Write only to SESSION_DIR:

• understand-guide.md — PDF/HTML source
• understand-study.json — flashcards + quiz

Read understand-explore.md first. Do not web-search unless explore gaps require external docs.

understand-guide.md schema

Follow templates/understand-guide.md.

Must include:

• ## Key concepts — each entry: Definition, Usage, Source (path:line link), Reference code (fenced block — copy real source, ≤ 30 lines)
• Use citation fences exactly: ```startLine:endLine:relative/path ` (content pasted from Read output)
• Mermaid diagram(s) when relationships are non-obvious (component map, request flow, state machine)
• ## Architecture — explicit Topology: single-node | distributed | hybrid with evidence; include Source + snippet for key entry points where helpful
• Subsections: Components, Data flow, Boundaries, Operational notes
• ## FAQ — 5–10 Q&A pairs; optional Source + snippet per answer

Reference code rules: Read the file, paste the actual lines — do not paraphrase code. Snippet must match the cited path:line range.

understand-study.json schema

Follow templates/understand-study.json.

Field	Rules
flashcards	≥ 8 cards; mix concept and implementation tags
flashcards[].refs	Required for implementation cards; optional for concept — {path, line, endLine?} repo-relative
flashcards[].code	Required for implementation cards when a short snippet helps (≤ 20 lines)
quiz	≥ 6 questions; mix concept + code; 4 choices each; answer is 0-based index
quiz[].refs	At least one ref on implementation-tagged questions
quiz[].code	Optional snippet shown in quiz review

Build script injects project_path from session JSON and turns refs into clickable IDE links (vscode:// / cursor://).

Orchestrator rules

• Spawn Azusa then Jun — do not self-analyze
• After Jun: run build script; report paths to PDF + HTML
• --yolo has no effect (no auto-skip)
• Skip skills/teammate-flow
• Read-only on application source; session artifacts only

Quality bar

• Concepts: precise definitions a new contributor can use in conversation
• Architecture: answer "single laptop CLI" vs "multi-service" explicitly
• FAQ: practical questions with short, grounded answers — not a repeat of the glossary verbatim
• Flashcards: front = prompt; back = concise answer; refs + code for implementation cards
• Quiz: plausible distractors; explanation + optional refs / code