By cosgroveb
Generates running session summaries on stop, viewable via tmux popup
npx claudepluginhub cosgroveb/claude-session-summary --plugin claude-session-summaryA Claude Code plugin that automatically generates session summaries whenever Claude Code stops.
On every Claude Code stop event, a Haiku agent reads the conversation context via --continue and generates a structured summary with a title, objective, completed items, and artifacts. Summaries are stored per-project under ~/.local/share/claude-sessions/.
--continueUsing Claude Code slash commands:
/plugin marketplace add git@github.com:cosgroveb/claude-session-summary.git
/plugin install claude-session-summary@claude-session-summary
Or clone and install manually:
git clone git@github.com:cosgroveb/claude-session-summary.git
cd claude-session-summary
make install
/plugin uninstall claude-session-summary
/plugin marketplace remove claude-session-summary
Or manually:
cd claude-session-summary
make uninstall
Add to ~/.tmux.conf for prefix + S:
bind-key S display-popup -w 80% -h 60% -E "glow -p ~/.local/share/claude-sessions/latest.md 2>/dev/null || cat ~/.local/share/claude-sessions/latest.md 2>/dev/null || { echo 'No session summary yet.'; read; }"
## <2-4 word title>
**Objective:** <1 sentence>
**Completed:**
- <bullet points>
**Artifacts:** <commits, files, or "None">
Each summary uses Claude Haiku, the cheapest Claude model. Costs depend on whether Claude Code's context is cached:
| Scenario | Cost | When it happens |
|---|---|---|
| Cached | ~$0.001-0.003 | Most calls—when you're actively working |
| Not cached | ~$0.03-0.05 | First call in a session, or after ~5 min idle |
Why the difference? Claude Code sends ~30K tokens of system context with each API call. The API caches this context for ~5 minutes. Cached reads cost 1/10th as much. Since you're typically making multiple Claude requests while working, most summary calls hit the cache.
Typical monthly cost:
Costs are logged to ~/.local/share/claude-sessions/<project>/summary.log:
2026-01-26T12:00:00+00:00 session=abc123 status=success cost=$0.003 input_tokens=100 output_tokens=50
MIT
Ultra-compressed communication mode. Cuts ~75% of tokens while keeping full technical accuracy by speaking like a caveman.
Share bugs, ideas, or general feedback.
Persistent memory system for Claude Code - seamlessly preserve context across sessions
Intelligent prompt optimization using skill-based architecture. Enriches vague prompts with research-based clarifying questions before Claude Code executes them