diagnose-failure

Diagnose stalled or failed training runs and propose targeted fixes.

3|Updated Apr 20, 2026
One-click install
npx skills add https://github.com/qualit527/qec-ai-decoder --skill diagnose-failure
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: diagnose-failure
Source: https://github.com/qualit527/qec-ai-decoder/tree/main/.claude/skills/diagnose-failure
Command: npx skills add https://github.com/qualit527/qec-ai-decoder --skill diagnose-failure

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Quickly identify why a training run stalled or failed and provide targeted, non-invasive guidance to fix the issue without applying changes automatically.

Core Features & Use Cases

  • Root-cause analysis: Examines logs and configurations to identify the most probable failure mode.
  • Actionable recommendations: Proposes concrete fixes or patches to recover runs.
  • Non-destructive guidance: Delivers steps for manual review and execution by the user.

Quick Start

Provide the run_dir or round_dir of the failed run to the skill and request a diagnosis.

Frequently Asked Questions about diagnose-failure

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I diagnose the root cause of a stalled or failed ML training run?

To diagnose a failed ML training run, provide the run_dir or round_dir path so the skill can read train.log, config.yaml, metrics.json, and history.jsonl to identify the root cause and generate a diagnosis.md file with suggested fixes.

What types of training run failures can I analyze from logs and configs?

You can analyze failures such as killed_by_safety, compile_error, and train_error by examining training history and configuration files to pinpoint the most probable failure mode and propose targeted DSL patches.

How do I get actionable fix suggestions for a training run that ended in a compile_error?

Actionable fix suggestions for a compile_error are generated by analyzing the run's logs and configuration files, producing a diagnosis.md with concrete evidence and a suggested DSL patch for manual review.

Does the failure diagnosis process apply patches to my training config automatically?

Failure diagnosis does not apply patches automatically; it delivers non-destructive guidance and targeted fixes in a diagnosis.md file for you to manually review and execute.

What files do I need to provide to identify why my training run was killed_by_safety?

You need to provide the run_dir or round_dir path so the analysis can access train.log, config.yaml, metrics.json, and recent history.jsonl files to identify the root cause of the safety termination.