session-management（会话管理）

Session Management

Purpose

Maintain continuity across AI agent sessions by providing structured work tracking, tech debt management, and report archival. This skill ensures no context is lost between sessions — the agent always knows what was done, what's pending, and what needs attention.

The Report Registry

The report registry (.claude/reports/_registry.md) is the central index of all work performed across sessions.

When to Add Entries

• After completing a meaningful unit of work (bug fix, feature, refactor, investigation)
• After an investigation that produced findings worth preserving
• When identifying an issue that needs future attention

How to Use

1. Read the registry at session start to understand recent work
2. Add entries as work is completed during the session
3. Update entry status when revisiting previous work (Active → Resolved)
4. Follow the 15-category structure for organization

Entry Format

### [Category] Short descriptive title
- **Date:** YYYY-MM-DD
- **Status:** Active | Resolved | Archived
- **File:** `.claude/reports/<category>/<filename>.md`
- **Summary:** One-line description of findings or changes

See references/registry-template.md for the full template with all 15 categories.

The Tech Debt Tracker

The tech debt tracker (.claude/reports/_tech-debt.md) captures technical debt discovered during development.

When to Add Debt

• When you notice code that should be improved but isn't the current focus
• When a workaround is used instead of a proper solution
• When tests are skipped or coverage gaps are identified
• When dependency updates are deferred
• When architectural concerns are discovered

Priority Levels

Item Format

### [P{0-3}] Short title
- **Added:** YYYY-MM-DD
- **Category:** code-quality | performance | security | architecture | testing | dependency | documentation
- **Location:** `path/to/file.ts:123`
- **Impact:** What happens if this isn't addressed
- **Suggested Fix:** Brief approach description
- **Status:** Open | In Progress | Resolved

See references/tech-debt-template.md for the full template with examples.

Archive Strategy

Keep the registry and debt tracker lean:

• Reports: Archive resolved items older than 7 days to .claude/reports/archive/
• Tech Debt: Keep resolved items 90 days for reference, then archive
• Registry Size: Don't exceed 50 active entries — archive aggressively
• Monthly Cleanup: Remove archived items that no longer provide value

See references/archive-strategy.md for detailed archival guidelines.

Self-Improvement via MEMORY.md

MEMORY.md is Claude Code's auto-memory file — it's loaded into every system prompt automatically. Use it as a self-improvement loop that accumulates project-specific lessons across sessions.

Location

~/.claude/projects/<project-path-with-dashes>/memory/MEMORY.md

When to Update MEMORY.md

• After any correction from the user — the most important trigger. If the user corrects your approach, save the lesson immediately.
• After discovering a non-obvious pattern or pitfall in the codebase
• After learning a user preference (e.g., "always use X library", "never auto-commit")
• After resolving a tricky bug whose root cause wasn't obvious

MEMORY.md Structure

Organize by topic, not chronologically:

# Project — Lessons Learned

## Build & Tooling
- [lesson]

## Architecture Patterns
- [lesson]

## Testing
- [lesson]

## Common Pitfalls
- [lesson]

## Owner Preferences
- [lesson]

Size Budget

Keep under 200 lines — this is loaded into every system prompt. If it grows too large:

1. Merge related entries
2. Remove lessons that are now obvious from context
3. Create separate topic files (e.g., debugging.md) and link from MEMORY.md

What NOT to Save

• Session-specific context (current task, in-progress work)
• Information already in CLAUDE.md or project docs
• Speculative or unverified conclusions
• Temporary workarounds that have been properly fixed

Session Start Workflow

When resuming work on a project with session management set up:

1. Read the registry: cat.claude/reports/_registry.md
2. Read tech debt: cat.claude/reports/_tech-debt.md
3. Check git state: git log --oneline -10 and git status
4. Review MEMORY.md lessons relevant to the area being worked on (MEMORY.md is auto-loaded in system prompt)
5. Summarize: Tell the user what's active, what's pending, and suggest what to work on
6. Check for P0 debt: If any P0 items exist, flag them for immediate attention

Session End Workflow

Before ending a session:

1. Update registry with any work completed this session
2. Add tech debt for anything discovered but not addressed
3. Update MEMORY.md with any corrections or lessons learned this session
4. Archive any resolved items past their retention period
5. Summarize what was accomplished and what's next

Cross-References

• references/registry-template.md — Full registry template with 15 categories
• references/tech-debt-template.md — Tech debt tracker template with examples
• references/archive-strategy.md — Archival rules and lifecycle management