DEV Community

skyestrela
skyestrela

Posted on

A Transparent Codex Bug-Reproduction Comparison: 6/14 vs 14/14

I make an open-source bug-reproduction workflow for coding agents. I wanted to test whether it changed the completeness of a Codex investigation, rather than relying on the workflow sounding sensible.

The result was 6/14 for the concise control and 14/14 for the corrected workflow-assisted run.

That headline needs context. This was one synthetic Python fixture, one model and a self-authored rubric. The workflow used more tokens and produced a much longer answer. The first workflow run did not score 14/14 either: it scored 12/14, exposed two weaknesses, and led to a revision before the final run.

Everything needed to inspect the comparison is public:

https://github.com/skyestrela/ai-agent-skill-preview/tree/main/evidence/bug-reproduction-benchmark

The question

Does supplying a structured Bug Reproduction Brief make a vague-bug investigation more complete and auditable than a concise task instruction alone?

The test was not designed to measure implementation speed or whether Codex could patch the bug. Both prompts explicitly prohibited modifying files or proposing a fix. The boundary was reproduction and evidence collection.

Controls

Both runs used:

  • the same authenticated Codex model: gpt-5.4-mini;
  • the same committed Python fixture: a4b9eab9dcc7b2ebbfe5f5d0502d4866cefd36ce;
  • read-only Codex sandboxes;
  • ephemeral sessions with user configuration and repository rules ignored;
  • PYTHONDONTWRITEBYTECODE=1;
  • clean working trees before and after;
  • the same instruction not to modify files or propose or implement a fix.

The workflow-assisted run additionally received Bug Reproduction Brief v1.0.1.

The fixture contains an intentionally faulty invoice calculation. Its existing tests pass because they do not cover the explicit-zero edge case described in the support report.

What both runs found

Both the control and workflow-assisted run:

  • identified the explicit-zero input that fails;
  • stated the expected and observed output;
  • ran the existing tests;
  • provided a runnable reproduction;
  • left the fixture unchanged;
  • avoided proposing a fix.

This matters because the 6/14 control was not useless. It found the central failure quickly.

What the workflow added

The workflow-assisted answer also recorded:

  • the immutable commit;
  • the inspected runtime environment;
  • that the intake report was second-hand and unverified;
  • an explicit minimal fixture;
  • two separate reproductions;
  • unresolved unknowns;
  • a safe next hypothesis;
  • a strict boundary between reproduced observations and diagnosis.

The concise control jumped from the reproduction to a root-cause statement. The final workflow answer deliberately stopped at observable evidence and a testable next hypothesis.

A deterministic 14-point rubric scored those criteria. The criterion-level result is committed as score.json, alongside the exact prompts and final outputs.

The failed first workflow run

The first workflow-assisted run scored 12/14, not 14/14.

It exposed two weaknesses in the workflow:

  1. the answer did not explicitly preserve the report's second-hand, unverified provenance;
  2. it stated a causal explanation despite the reproduction-only boundary.

I tightened the public skill to v1.0.1 and reran it against the unchanged fixture. The 14/14 result came from that corrected workflow.

That makes this an iterative product-development comparison, not a blinded academic study. I am disclosing the iteration because hiding the 12/14 run would make the final result look cleaner than the process actually was.

Token and verbosity cost

The additional completeness was not free.

Run Input tokens Cached input Output tokens Reasoning output
Control 68,604 62,720 1,400 255
Workflow 74,547 67,328 4,471 2,661

The workflow used 8.7% more input tokens and produced a substantially longer answer.

That trade-off will not suit every task. For a low-risk bug where speed matters more than auditability, the concise instruction may be sufficient. Teams adopting a structured brief should shorten its output contract rather than assuming maximum detail is always better.

What this result does not prove

It does not prove that:

  • every coding agent improves with this workflow;
  • the result repeats across repositories or models;
  • a 14/14 reproduction brief leads to a better patch;
  • the extra tokens provide positive economic value;
  • the workflow is universally better than a concise prompt.

Outputs can vary between repeated model runs. The rubric rewards reproduction-brief completeness, not bug-fix quality or developer productivity.

A stronger follow-up would run multiple trials across several repositories and models using a rubric fixed before execution.

Reproduce or criticise it

The repository contains:

  • the exact control and workflow prompts;
  • the intentionally faulty fixture and passing test suite;
  • the final control and workflow outputs;
  • criterion-level scoring and token usage;
  • the complete MIT-licensed workflow.

Benchmark and reproduction files:

https://github.com/skyestrela/ai-agent-skill-preview/tree/main/evidence/bug-reproduction-benchmark?utm_source=devto&utm_medium=article&utm_campaign=bug-reproduction-benchmark

Workflow source:

https://github.com/skyestrela/ai-agent-skill-preview/blob/main/bug-reproduction-brief/SKILL.md?utm_source=devto&utm_medium=article&utm_campaign=bug-reproduction-benchmark

The most useful feedback would be criticism of the controls, rubric or reproduction boundary—not generic agreement with the score.

Disclosure

I created this workflow and sell an optional wider engineering pack containing ten editable workflows. The complete Bug Reproduction Brief used here is already public and MIT licensed; buying anything is not required to inspect or reproduce this comparison.

Optional wider pack:

https://ai-agent-skills-pack.vercel.app/?utm_source=devto&utm_medium=article&utm_campaign=bug-reproduction-benchmark

Top comments (0)