full-shake

Marathon thread context reset for long-running sessions (30+ turns). Prunes tool outputs, windows scratchpad thoughts to the last 20 turns, and applies the Milestone Horizon (preserving Turn 1 Genesis

SKILL.md

GitHub

⚡ /full-shake: Marathon Thread Context Reset

Run this skill when you are working on marathon conversations (30+ turns) and need to restore "Turn 1" speed and agility while staying in the exact same chat window.


🌟 The 2-Tier Marathon Reset

  1. 💭 Scratchpad Thought Windowing:
    • Drops ancient internal monologues older than 20 assistant turns while preserving final answers and code verbatim (saves ~100k+ tokens on reasoning models).
  2. 🏛️ The Milestone Horizon (For Sessions > 30 User Turns):
    • Turn 1 (Genesis): Preserved 100% verbatim (project origin, guidelines, and constraints).
    • Middle Horizon (Turns 2 to N-25): Compacted into a structured Milestone Checkpoint block with exact line-indexed backup links to the permanent disk archive.
    • Active Working Window (Last 25 turns): Preserved 100% verbatim (with the last 10 turns of tool outputs completely unpruned).
  3. Natural Fallback:
    • If a session has $\le 30$ user turns, all conversation turns are preserved verbatim, applying only thought windowing and tool pruning.

🚀 Execution Instructions

When the user invokes /full-shake:

  1. Locate the active session's transcript.jsonl log file:

    • Primary path: <appDataDir>/brain/<conversation-id>/.system_generated/logs/transcript.jsonl
    • Artifact directory: <appDataDir>/brain/<conversation-id>/
  2. Execute the native compiled binary with --full:

    ~/.gemini/bin/shake-prune "<transcript_path>" "<artifact_directory>" --full
    
  3. Present the formatted output report directly to the user.