Error2fix

error2fix turns noisy terminal failures into compact, structured debugging context.

The project has two complementary entry points:

• A post-failure CLI that records failed shell commands and summarizes the latest failure for humans.
• An MCP server that lets coding agents request compressed failure context instead of reading full raw logs directly.

The core product goal is to reduce the amount of irrelevant log text that a developer or LLM has to inspect before reaching the useful failure signal.

Why it exists

Terminal failures are usually noisy, repetitive, and hard to hand over to an AI assistant cleanly. Developers often have to copy a long log, explain what command ran, mention the working directory, and then ask the model to find the real signal.

error2fix keeps that workflow smaller:

1. The CLI captures normal command failures after they happen.
2. The diagnosis pipeline extracts the useful error signal from raw output.
3. The MCP server gives agents compact failure context before they ask for more evidence.

This is especially important for LLM workflows, where reading an entire raw log can waste tokens before the model reaches the actual error.

CLI Installation

Install the CLI package:

npm install -g @error2fix/cli

Initialize shell integration once:

e2f init

After that, keep using your terminal normally. When a command fails, run:

e2f

The CLI stores local failure data under ~/.e2f.

MCP Installation

Install the MCP server package:

npm install -g @error2fix/mcp

The package exposes:

e2f-mcp

Configure your MCP client to start e2f-mcp as a stdio server.

VS Code style configuration:

{
  "servers": {
    "error2fix": {
      "type": "stdio",
      "command": "e2f-mcp"
    }
  }
}

Cursor, Claude Desktop, Cline, and other mcpServers-style clients usually use:

{
  "mcpServers": {
    "error2fix": {
      "command": "e2f-mcp"
    }
  }
}

Product Direction

error2fix is built around a post-failure workflow. Users should not have to wrap every command with a special runner. A normal command fails, then error2fix helps inspect what happened.

For agent workflows, the goal is stricter: the model should not need to ingest an entire terminal log before it can reason about the failure. The MCP server is designed to expose a small loop:

1. Return the most valuable compressed failure brief.
2. Cache the failure as a session.
3. Let the agent query focused evidence only when the brief is not enough.
4. Return client-provided runtime context on demand.

This keeps raw logs outside the model context as much as the host client allows.

MCP Workflow

The MCP server currently exposes:

• e2f_get_latest_failure_brief: accepts stdout, stderr, and optional command metadata, then returns the compact diagnosis, evidence IDs, token policy metadata, and a sessionId.
• e2f_query_failure_evidence: expands specific evidence from the cached session without returning the raw log.
• e2f_get_runtime_context: returns client-provided command, workspace, shell, OS, git, or safe environment context from the cached session.

The intended agent flow is:

failed command output
  -> e2f_get_latest_failure_brief
  -> optional history search with the returned summary if available
  -> answer if sufficient
  -> e2f_query_failure_evidence only if more detail is needed
  -> e2f_get_runtime_context only if environment context matters

If the agent environment provides a history-search capability, use the brief summary and other compact diagnostic fields as the query seed. Do not search with raw stdout or stderr. When no history search is available, follow the same MCP flow without any extra dependency.

Benchmark

The repository includes an early MCP benchmark dataset under benchmarks/failures. Each case stores the original downloaded failure log as raw.log, so anyone can rerun the same compression workflow locally.

Current frontend benchmark snapshot:

• Cases: 7
• Signal matches: 21/26
• Average reduction: 60.9%
• Average total MCP ratio: 39.1%
• Tool calls: 1 per case in the current run
• Report: benchmarks/reports/report.md

This dataset is intentionally small and currently includes several short logs. Short-log cases expose response-shape overhead, so the MCP brief uses a smaller compact output shape for small raw logs.

Run the benchmark with:

pnpm benchmark:mcp