regression-io/claude-config-plugins/flush

Flush Context to Resumable Doc

Save all current session context to a per-session file .claude/session-context.<session-id>.md for seamless continuation. Per-session naming means parallel sessions never overwrite each other's flush.

When to Use

• Ending a work session
• Switching to a different task
• Before a long break
• When context is getting complex and you want a checkpoint
• User says "save context", "flush", "checkpoint", or similar

Process

Step 1: Pick the .claude Directory (anchor to the project root)

1. Run git rev-parse --show-toplevel from the cwd. If it succeeds, use <repo-root>/.claude/.
2. If cwd is not in a git repo, use <cwd>/.claude/ (the "working directly in ~" case → ~/.claude).
3. Create the dir if it doesn't exist — no prompt, the location is deterministic. Do not offer /init. Print the resolved absolute path so the location is never a surprise.

Do not walk up the tree looking for an ancestor .claude/: that made flush drift into the nearest ancestor (including the global ~/.claude config dir) instead of the project you're working in.

Step 2: Create Session Summary

Summarize the session:

Step 3: Write to File

Resolve the session id (printenv CLAUDE_CODE_SESSION_ID) and write to .claude/session-context.<session-id>.md in the chosen directory. If the env var is empty, fall back to the legacy global name .claude/session-context.md.

Then reap stale resume prompts (per-session files older than 14 days; the glob skips the legacy global file and won't touch the one just written):

find "<chosen-dir>/.claude" -maxdepth 1 -name 'session-context.*.md' -type f -mtime +14 -delete

Output Format

# Session Context - YYYY-MM-DD
 
## Task Summary
[What the user wanted to accomplish]
 
## Current State
**[status]**
[Brief description]
 
## Key Decisions
-**[Decision]**: [Rationale]
 
## Files Modified
-`path/to/file` - [what changed]
 
## Pending Work
-[ ] Task 1
-[ ] Task 2
 
## Important Context
[Critical information for continuation]

Notes

• Resume a saved context with /unflush, which finds the right per-session file (preferring the current session id, else the most recent) and acts on it
• Keep summaries concise but complete enough to resume without re-reading code
• Focus on the "why" not just the "what"