# Reproduce Failure

> Use when a bug-fix run must first demonstrate the failure it claims to fix: reproducing the failure with a real command/test run, or — when reproduction is genuinely impossible — inspecting and recording the concrete failure evidence. This is the reproduce_failure action of the debugging stretch — one Agent-tool dispatch with subagent_type qa-engineer — the mandatory first step before any cause is hypothesized: no reproduction or inspection, no fix.

- Skill: `bhangeef16/reproduce-failure` (Agent Skill)
- Install (CLI): `npx skillmds@latest add bhangeef16/reproduce-failure`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bhangeef16/reproduce-failure/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: BhangeeF16 (https://skillmd.com/u/bhangeef16)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bhangeef16/reproduce-failure

---


# Reproduce Failure

Demonstrate the failure before anyone reasons about it. A bug that cannot be reproduced or inspected is a report, not yet a defect anyone may fix.

## Dispatch

One synchronous `Agent`-tool call, `subagent_type: qa-engineer`.

## Required inputs

- The failing command output — the actual observed failure, or the report pointing at it.

## Required outputs (all mandatory)

- Reproduced failure (the command run and its failing output, observed now) or, when reproduction is genuinely impossible, inspected failure evidence — the concrete artifacts (logs, test output, data state) that demonstrate the failure actually occurred, with the reason reproduction was not possible stated.

## Rules

- Reproduce first, always — a failure that cannot be demonstrated at all is a stop condition for the whole run, not a license to fix by guess.

## Result

Report each required output produced (with its location), the result (pass / pass_with_risks / blocked / failure), and blocking questions as an explicit list (empty list stated explicitly).