More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Agent Error Tracker

    1

    Records every Claude Code tool call via hooks, groups recurring failures by root cause and writes a Markdown + HTML diagnostic report. Local only, no dependencies, secrets redacted.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Agent Error Tracker

    Agent Error Tracker

    Example session with this skill installed

    /error-tracker examples/sample-session.jsonl

    Why do my tool calls keep failing in this session?

    • Read your context and instructions
    • Compiled the agent error tracker
    • Generated the document

    Ran the analyzer on examples/sample-session.jsonl (1 session).

    6 operations · 3 certain failures · 50% failure rate · 0 suspected failures

    Recurring errors

    ErrorOccurrencesTools
    ModuleNotFoundError2Bash
    MCP Timeout1mcp__docs__search

    ModuleNotFoundError (2)

    • Message: E ModuleNotFoundError: No module named 'requests'
    • Failed operations: pytest tests/; pytest tests/ -x
    • First seen 10:00:01Z, last seen 10:00:40Z
    • Typical cause: a Python module is not importable — missing dependency, wrong virtualenv, or wrong working directory / PYTHONPATH.

    MCP Timeout (1)

    • Message: MCP error -32001: Request timed out
    • Failed operation: mcp__docs__search
    • Typical cause: the operation exceeded its time limit — slow or unreachable service, or a command waiting for input.

    Tools involved

    ToolOperationsFailures
    Bash42
    mcp__docs__search11
    Read10

    Suspected failures

    None: no successful call printed error text.

    Reports

    • Markdown: examples/sample-report.md
    • HTML: examples/sample-report.html (open it in your browser for the visual version)

    Next steps

    • The pytest failures share one root cause: requests is missing from the active environment. Installing it (for example pip install requests) or activating the right virtualenv should fix both. I haven't run anything; tell me if you want me to.
    • The mcp__docs__search timeout happened once: check that the docs MCP server is reachable before retrying.

    sample-report.html

    PDF · document

    Generated

    Example file from a real run - the skill writes it into your workspace.

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

    About this skill

    Tool calls fail silently all the time in Claude Code sessions: a missing module, an MCP server that times out, a command that exits 0 but prints a traceback. You see them scroll by, but nobody keeps count. Agent Error Tracker records every tool call through the official Claude Code hooks, groups recurring failures by root cause, and gives you a clear diagnostic report with one command: /error-tracker.

    What it does

    • Captures every tool call via the PostToolUse and PostToolUseFailure hooks into a local JSONL log (.claude/error-tracker/events.jsonl).
    • Separates certain from suspected failures: certain = reported by the PostToolUseFailure hook; suspected = a "successful" call whose output contains a traceback, an …Error/…Exception, FAILED, fatal: or a timeout. Suspected failures are listed apart and never inflate the failure rate.
    • Groups errors by root cause: interrupts, the last exception in a traceback (e.g. ModuleNotFoundError), timeouts (MCP Timeout for mcp__* tools), Exit code N, or the normalized first line.
    • Ranks the tools involved by operations and failures.
    • Writes two reports: a Markdown report for the chat and a self-contained HTML page (offline, light/dark theme) to open in the browser.
    • Explains well-known errors only: typical causes are shown for known error types; for anything else the report says there is not enough information instead of guessing.

    How it works

    1. The capture.py hook writes one compact, redacted record per tool call. It always exits 0, so it never blocks or changes Claude Code's behaviour.
    2. /error-tracker runs analyzer.py --html, which deduplicates calls by tool_use_id, classifies and groups failures, and renders the reports.
    3. Claude reads the Markdown report and summarizes it: totals, recurring errors with examples, the most unstable tools, suspected failures (labelled as unconfirmed) and the report paths.

    Privacy and safety

    • Local only: no cloud services, no database, no network calls.
    • Secrets redacted before writing: Bearer tokens, token/secret/password/api_key/authorization key-value pairs, --password flags, credentials in URLs, and token formats such as sk-…, ghp_…, github_pat_…, AKIA…, xox…-.
    • File contents are never stored: only the target (command, path, pattern, URL without query string) and a trimmed error text (max 2,000 characters).
    • The skill is read-only on your project: it may suggest a fix, but never runs it unless you ask.

    Use cases

    • Finding out why pytest or a build keeps failing across a long agent session.
    • Spotting a flaky MCP server that times out intermittently.
    • Catching "silent" failures: commands that exit 0 but print an error.
    • Checking tool reliability after changing dependencies or the environment.

    What's inside

    • SKILL.md: the /error-tracker skill.
    • scripts/capture.py: the hook.
    • scripts/analyzer.py: analyzer and CLI (also usable without Claude).
    • scripts/html_report.py: the HTML report.

    Pure Python standard library, Python 3.9+. Open source (MIT), with tests and a reproducible sample session on GitHub.

    Known limitations

    • Hooks must be added to .claude/settings.json once (snippet in the README); only sessions recorded after that are analyzed.
    • The hook command uses python; on macOS and some Linux distributions change it to python3.
    • Secret redaction is pattern-based, so it is best effort. The log stays on your machine and should be git-ignored.

    Built for Claude Code hooks. Only sessions recorded after the hooks are installed in .claude/settings.json (and Claude Code restarted) are analyzed. The hook uses "python"; change it to "python3" where needed. Secret redaction is pattern-based (best effort).

    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 3 days ago

    Requires

    Python 3.9+
    Claude Code
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Environment Variables

    File Scopes

    .claude/error-tracker/**

    Shell: runs the bundled Python analyzer (allowed-tools: Bash(python *), Bash(python3 *)). Read/Write: the hook appends to .claude/error-tracker/events.jsonl and the analyzer writes report.md and report.html there; the skill then reads the report. No network access, no project files modified.

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    Claude Code shows each failure once, in the flow of the session. This skill keeps a log of every tool call, groups recurring failures by root cause, computes the failure rate per tool, and flags "suspected" failures: calls that exited 0 but printed a traceback, FAILED, fatal: or a timeout.

    Claude Code and Python 3.9+. No third-party packages: only the standard library. You add the hooks block to .claude/settings.json once (snippet in the README).

    The /error-tracker SKILL.md, the capture hook (capture.py), the analyzer and CLI (analyzer.py) and the HTML report generator (html_report.py). The full source, tests and a sample session are on GitHub under the MIT license.

    No. It only analyzes tool calls recorded by the hooks after you install them. You can, however, point it at any JSONL log in the same format: /error-tracker path/to/events.jsonl.

    Nothing leaves your machine: no network calls, no cloud, no database. Tokens, passwords, API keys and credentials in URLs are redacted before anything is written, and file contents are never stored. Redaction is pattern-based, so keep the log git-ignored.

    New in Testing & Debugging