Compress Session

compress_session is an MCP prompt (not a tool) that guides your AI assistant through summarizing accumulated observations for a session. This is how client-mode compression works — the client's own LLM does the summarization instead of the server.

What it does

When invoked, it asks the AI to:

1. Read uncompressed observations via get_uncompressed_observations
2. Summarize them in a structured format:
   • Request — what the user asked to accomplish
   • Completed — what was actually done (specific files, functions, patterns)
   • Learned — key decisions, constraints, or patterns discovered
   • Next Steps — unfinished work or open questions
3. Store the summary via compress_observations, which marks those observations as compressed and saves the summary as a searchable memory

Parameters

When it triggers

You don't normally invoke this manually. Both plugins handle it automatically:

• Claude Code — the observation.sh hook returns additionalContext when uncompressed_count reaches the batch threshold (default 20), which tells the LLM to run this prompt
• OpenCode — the experimental.chat.system.transform hook injects a system prompt when the threshold is hit

Manual usage

If you want to trigger it yourself:

Use the compress_session prompt for session <session_id>

You can find your session ID via session_context.

Why client-side?

Server-side compression requires the server to have an LLM (Anthropic API key or Ollama). Client-side compression uses the LLM that's already running in your editor — already paid for, already available. The server stays dumb storage.

See Compression modes for the full comparison.