More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Session Bloat Watch — Catch Oversized Agent Sessions Early

    1

    Each turn of a Claude Code session re-reads the entire conversation from the prompt cache.

    $12

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Session Bloat Watch — Catch Oversized Agent Sessions Early

    Session Bloat Watch — Catch Oversized Agent Sessions Early

    Example session with this skill installed

    My usage is burning fast this week. Which of my sessions are re-reading the most?

    • Read your context and instructions
    • Compiled the session bloat watch

    $ python3 scripts/session_bloat_watch.py --root ./demo-logs --top 5
    session size turns imgs cache_read per_turn modified flags
    long-rev 10.6MB 220 55 61.3M 278.6k 2026-11-09 16:49 size>=8MB,images>=40,reread/turn>=150k
    quick-fi 0.0MB 12 0 360.0k 30.0k 2026-11-09 16:49
    scanned 2 session(s), 1 over a threshold

    The long-review session re-reads about 279k tokens per turn and holds 55 images. Save its open notes to a file, then start a fresh session for the next task. quick-fix is fine. (Run on synthetic demo logs.)

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Each turn of a Claude Code session re-reads the entire conversation from the prompt cache. The longer a session lives, the more every later turn re-reads, so two or three long-lived sessions can account for most of your usage, and nothing in the terminal tells you which.

    Session Bloat Watch reads the session logs Claude Code already keeps on disk and prints one row per session, largest cache-read first: size, assistant turns, image blocks, total cache-read tokens, and the average re-read per turn. Rows that cross a threshold (defaults: 8 MB, 40 images, 500M cache-read tokens, 150k re-read per turn) are flagged.

    What you get:

    • Metadata only: it never prints message text, never writes and never deletes.
    • Counts assistant turns once per message id (the log repeats ids per content block), and counts images nested inside tool results.
    • --since-hours to look only at live sessions, --json for machines, --fail-on-warn (exit code 10) for cron or a hook.
    • A SKILL file that teaches Claude Code to report, recommend a restart with notes saved first, and never touch the session files.
    • 12 tests on synthetic logs.

    Requirements and safety: Python 3.8+. Standard library only. No network calls, no installers, no background services, no credentials. It depends on the current JSONL log layout; if a future version changes it, counts can read 0. Numbers describe usage in the logs, not your bill.

    Reads metadata only; no message content access. Cannot delete or move session files. Provides usage estimates, not actual billing data.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 8 days ago

    Requires

    Python 3.8+

    Permissions

    Terminal / Shell
    Read Files
    Listed8 days ago
    Updated5 days ago

    What's inside

    Frequently Asked Questions

    This skill analyzes your local Claude Code metadata to identify sessions that have grown too large. By flagging sessions with excessive image blocks or high 're-read per turn' token counts, it helps you know exactly when to restart a session to avoid wasting your prompt cache budget.

    The skill is designed for developers using Claude Code in their terminal. It requires Python 3 and read access to the local directory where your session logs are stored.

    No, for privacy and security, the script only parses metadata like file size, turn counts, and image block headers. It does not read or report the actual text content of your conversations.

    Yes, while it comes with optimized defaults, you can define your own thresholds for disk size, turn counts, and token limits when running the script or via CI/CD integration.

    You receive the core Python script, a comprehensive README with setup instructions, and documentation on how to integrate the JSON output into your automated monitoring or CI/CD pipelines.

    This is a monitoring tool only; it will never modify or delete your session files. Once it identifies a bloated session, it is up to you to manually save relevant context and restart the session in Claude Code.

    New in LLM Ops