mcpField
F
FailTrace

FailTrace

Local
github.comDeveloper Tools0Updated 8w ago
RepositoryVisit website

About FailTrace

FailTrace is a local CLI and MCP server for coding agents investigating intermittent test failures. It preserves repeatable trial evidence to compare a captured failure with a proposed fix.

Key features

  1. 01Repeat tests and capture failure baselines
  2. 02Compare healthy and failing trials
  3. 03Search Git history with bisect
  4. 04Minimize reproducing inputs
  5. 05Verify patches against saved baselines
  6. 06Save replayable investigation bundles

Use cases

  1. 01Capture a flaky test before editing
  2. 02Check a proposed fix against the original failure
  3. 03Find a Git change associated with a regression
  4. 04Reduce an input to a smaller reproducer

MCP support

Its local stdio MCP server provides seven tools that use the same engine as the CLI for repetition, comparison, regression search, minimization, verification, and replay. MCP clients can also retrieve saved trial and log pages.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
  "mcpServers": {
    "lbarimi-failtrace": {
      "command": "npx",
      "args": [
        "-y",
        "failtrace"
      ]
    }
  }
}
Run commandnpx -y failtrace

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npmfailtrace1.5.0

Tools

7

Saved tool list · updated 3w ago

ParameterTypeRequiredDescription
cwdstringNoWorking directory; relative paths resolve from the server working directory.
envobjectNoExplicit environment overrides. null unsets an inherited variable.
argsarrayNoLiteral executable arguments, including an empty array, opt out of shell parsing. Minimize replaces only entire {input} arguments with each candidate path. Windows .cmd/.bat shims require shell mode.
repeatintegerNo
commandstringYesShell command when args is absent; executable when args is present. Command and arguments together are limited to 64 KiB. Executed with your local permissions.
predicateanyNo
timeoutMsintegerNo
captureEnvarrayNoOnly these selected environment variables are recorded in metadata.
concurrencyintegerNoMaximum active run trials; default 1, at most 64. Shared ports, files, databases and resource contention can change failure probability.
artifactsDirstringNo
captureContextobjectNoCapture declared regular input/setup/source file identities for verification. Source files select a files-only scope; without them, capture Git identity. Use captureEnv separately for relevant variable names.
captureContext.inputFilesarrayNo
captureContext.setupFilesarrayNo
captureContext.sourceFilesarrayNo
maxOutputBytesintegerNoCombined stdout/stderr byte cap per trial; default 16 MiB. Limit outcomes are inconclusive.
maxTotalOutputBytesintegerNoCombined output byte cap for this whole experiment, including all candidates; default 256 MiB.
executionRequirementobjectNoOptional checkpoint emitted after the intended check ran. Missing evidence makes classification/verification inconclusive; Verify inherits it from the baseline.
executionRequirement.streamenumYesstdout | stderr
executionRequirement.containsstringYes

Details

From the project README

FailTrace

Reproduce the failure. Check the fix.

A test fails intermittently. Your coding agent changes the code. One passing retry leaves you guessing.

FailTrace repeats the same check, saves a failing baseline, and checks the proposed fix against it. Use the CLI or MCP tools to get trial results, a smaller reproducer, and evidence your agent can inspect before accepting a patch.

CI License: MIT

Local execution. No AI API, account, or telemetry required. Keep your existing test runner and assertions.

Quick start

With Node.js 22.12+ and npm, run this in any working directory:

npx --yes failtrace demo

FailTrace demo: capture a failure, reduce its input, reject an unrelated crash, and check a patch

Static walkthrough · Static poster · Demo guide

The demo reduces six input items to ["BUG"], rejects a patch that crashes for another reason, and checks a working patch. It saves the evidence in .failtrace/ and prints a replay command.

For coding agents

Connect the local stdio MCP server through your client's configuration:

npx --yes [email protected] mcp --cwd "/absolute/path/to/your/project"

Copy the MCP configuration and check the connection →

Your client launches this command. Running it alone in a terminal waits for MCP requests. The guide includes Windows setup.

Then ask your agent:

Use FailTrace to capture this test failure before editing. Choose the exact test or failure message, save a bounded baseline, and inspect the matching trial. After the change, verify against that baseline and explain any unrelated errors or incomplete evidence.

The seven MCP tools share the CLI's Core engine. They retain the failure signature and investigation evidence across repetition, comparison, regression search, minimization, verification and replay. Agents can retrieve saved trial and log pages without rerunning the command. Shell-capable agents can also use the CLI with --json.

MCP workflow: capture the target, inspect saved stderr, and verify a declared patch

The agent can inspect the saved failure before checking the patch.

Recheck an existing unit test

Follow an exact NUnit or Unity test through a fix. Each attempt gets a fresh NUnit 3 report. Missing or skipped tests and unrelated failures stay inconclusive, so an agent cannot accept them as evidence that the selected test passed.

Connect your test or try the original EditMode example →

NUnit support is included in 1.3.0 through CLI and MCP. The documented Unity validation covers the Windows EditMode example.

NUnit evidence: a failing baseline, a passing candidate, and a skipped test kept inconclusive

The selected test stays the same; a skipped report is not accepted as a passing test.

Use it on your own failure

From your project, replace the command and message with your own:

npx --yes [email protected] run "npm test -- checkout" --repeat 20 --stderr-contains "checkout failed" --capture-context

Run this before editing in a Git project. --capture-context records source identity for Verify; outside Git, select files with --context-source. Each trial saves its output, and exit 1 can mean the target was captured successfully. Then edit and verify the patch →

Your next questionCommand
How often does this failure appear?run
What differs between a healthy and failing trial?compare
Which Git change introduced it?bisect
What input is enough to reproduce it?minimize
What happened after the proposed fix?verify
How can I replay this investigation?bundle

Command reference · Literal executable arguments · Reusable project scripts

A reduced input packaged with a replay entry point reproduces the target failure

The demo's bundle reproduces its original failure with exit 1. Keep the source, input and replay together; supply the target's prerequisites when packaging your own investigation.

What the results establish

FailTrace reports observations under the chosen settings. Verify separates a target observed, a healthy sample without that target, and inconclusive evidence. Bisect reports a sampled first-parent boundary; minimization rechecks its result without promising the smallest possible input.

The demo shows controlled example outcomes, not performance measurements. A passing sample does not prove a bug is gone.

Commands run with your permissions, and process cleanup is best effort. Retained stdout/stderr is capped by default at 16 MiB per trial and 256 MiB per run or bisect/minimization. Previous investigations accumulate separately. Review logs, commands and selected files before sharing; bundles still require the target's dependencies and setup.

Result and exit-code reference · Resource limits · Storage inventory · Bundle guide

Availability and contributing

The quick start uses npm's latest release; the verified version is 1.5.0. MCP configuration and repeatable installation examples keep that exact version pinned: installation options, GitHub release, and MCP Registry entry.

Version 1.5.0 adds: short run references, Verify readiness and next-step guidance, and intermittent-minimization guidance. See the changelog.

Documentation: choose your next task →

Development instructions · Compatibility · Roadmap · Third-party notices

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.