Save Session State

Guidance for saving session state before context compression or handoff.

When Claude Should Use This Skill

• User mentions "save session", "running out of context", "before we lose context"
• User is about to run /compact
• User says they're ending the session or switching tasks
• Context usage is high and important work needs to be preserved

Usage

/save-session                           # Save state
/save-session "check auth.ts line 145"  # Save with handoff message

Execution Steps

1. Gather State

Run this to get current git and project state:

echo "PROJECT_PATH=$(pwd)"
echo "---"
git branch --show-current
git status --short
bd list --status=in_progress 2>/dev/null || echo "No in-progress issues"

2. Create Pieces Memory

Use mcp__pieces__create_pieces_memory with:

summary_description: Short title with keywords for retrieval

[project-name] session: [main topic] - [branch] - [YYYY-MM-DD]

Example: dssk-multichannel-gui session: comment-form extraction - feat/kuendigungsdetails - 2026-01-12

summary: Markdown-formatted narrative with searchable keywords:

# Session Summary - {DATE}

## What Was Accomplished
- Completed task 1
- Implemented feature X
- Fixed bug Y

## Current State
- **Branch:** {branch name}
- **Uncommitted changes:** {summary of git status}
- **In-progress issues:** {bd list output}

## Handoff Message
{user's message if provided, otherwise omit}

## Next Steps
1. Continue with X
2. Need to finish Y
3. Review Z

## Key Decisions Made
- Decision 1: rationale
- Decision 2: rationale

## Important Context for Continuation
Any context the next session needs to understand the work.
File paths, architectural decisions, blockers encountered.

project: Absolute path to project root (from step 1)

files: List of key files modified this session (absolute paths)

connected_client: "Claude"

3. Sync Beads

bd sync

4. Provide Continuation Instructions

Always end with:

**Session saved to Pieces.** To continue in a new session:
1. Start fresh Claude Code session
2. Run `bd ready` to see available work
3. Use `ask_pieces_ltm` to retrieve session context

Serena Memory Updates

If significant code understanding was gained, consider updating:

• project_overview - architecture changes
• coding_standards - new patterns established
• types-reference - new types documented

Retrieval Best Practices

Pieces indexes content semantically. To improve retrieval:

In summary_description (acts as title/tag):

• Start with project name: dssk-multichannel-gui session:
• Include main topic: form validation, component extraction
• Add branch name for filtering: feat/kuendigungsdetails
• End with date: 2026-01-12

In summary content:

• Use consistent section headers (What Was Accomplished, Next Steps, etc.)
• Include technology keywords: Angular, signals, reactive forms
• Mention specific components/files by name
• Add a ## Keywords section at the end:

  ## Keywords
  Angular, component extraction, comment-form, kuendigungsdetails, refactoring, Serena, model injection
  

File paths matter:

• Pieces links memories to files - include all modified files
• Use absolute paths for accurate linking
• This enables "what did I do in this file?" queries

Retrieval queries:

ask_pieces_ltm: "What was I working on in dssk-multichannel-gui last week?"
ask_pieces_ltm: "Show me sessions about form validation"
ask_pieces_ltm: "What decisions were made about component extraction?"

Why Pieces Over Basic Memory

• Cross-tool: Accessible from Cursor, Windsurf, other AI tools
• Searchable: ask_pieces_ltm can find context by topic
• Persistent: Survives session restarts and tool switches
• Indexed: Automatically linked to files and project