skills.rest
sandeco
sandecoCOMMUNITY
Explore all skills by this author →
Website𝕏@sandecoGitHub

reversa-debugger-debate

Runs multi-agent debates with isolated judge to decide bug diagnosis, repair, or spec verdicts.

1.5k|405|Updated Apr 26, 2026
Search Tags:#Debugging#Root Cause Analysis#bug triage#legacy systems#llm judge#multi-agent debate#Reversa
One-click install
npx skills add https://github.com/sandeco/reversa --skill reversa-debugger-debate
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: reversa-debugger-debate
Source: https://github.com/sandeco/reversa/tree/main/agents/reversa-debugger-debate
Command: npx skills add https://github.com/sandeco/reversa --skill reversa-debugger-debate

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Debugging legacy systems often relies on a single AI pass that can lock onto the wrong root cause or repair strategy. This Skill orchestrates a structured multi-agent debate where N independent solvers argue over R rounds and an isolated judge synthesizes one audited recommendation for a registered bug.

Core Features & Use Cases

  • Three Debate Modes: diagnosis (competing causal hypotheses), repair (competing fix strategies), and spec (recommendation on whether code, tests, or spec is correct).
  • Isolated Judge with Frozen Rubric: the judge sees only anonymized, shuffled final proposals and applies a rubric frozen before the debate, preventing echo chambers and reward hacking.
  • Auditable State and Cost Control: all rounds, convergence metrics, and the final answer are written to the bug's debate folder, with explicit cost estimates and user opt-in before execution, including optional external CLI debaters (Codex, Gemini CLI).
  • Use Case: A bug is registered in the Reversa Bugs workflow but the root cause is ambiguous. Run this Skill to have three solvers debate hypotheses over two rounds, then receive the judge's synthesized diagnosis with evidence before applying any fix.

Quick Start

Ask the agent to run a multi-agent debate on bug BUG-20260715-A7K3 in diagnosis mode with the default 3 solvers and 2 rounds.

Frequently Asked Questions about reversa-debugger-debate

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

FAQPage Schema
How do I run a multi-agent debate to diagnose a bug?▼

Invoke the debate skill with a registered bug ID and choose diagnosis mode. It runs N solvers (default 3) for R rounds (default 2), then an isolated judge synthesizes one recommendation written to resposta-final.md in the bug's debate folder.

What is the difference between diagnosis, repair, and spec debate modes?▼

Diagnosis debates competing causal hypotheses against the evidence. Repair debates fix strategies once the cause is confirmed, favoring minimal, reversible changes. Spec debates whether code, tests, or spec is correct and ends in a recommendation, with the final verdict left to a human.

Can I use external CLI agents like Codex or Gemini in the debate?▼

Yes, external CLI harnesses detected on the PATH can join as solvers or critics, but only after explicit user approval and a probe confirming non-interactive authenticated operation. Bugs marked visibility: restricted never allow external harnesses.

What happens when a debate participant fails or times out?▼

Each call has a hard 10-minute timeout with one retry only for transport failures. Failed debaters produce a file with status timeout, error, or invalid-output and are never silently replaced; the debate continues automatically only if quorum of max(2, ceil(2N/3)) valid solvers is met.

Does the debate skill modify my project source code?▼

No. It writes only inside the bug's debate directory under _reversa_bugs and never modifies, deletes, or overwrites existing project files. It decides strategy only; applying the fix happens separately in the reversa-debugger-fix workflow.

Related Skills

Explore More Software Engineering Skills→
jlissnerjlissner

grimoire-bug-report

Generate structured bug reports from Playwright and Cypress test outputs.

Community
Advanced
g2i-aig2i-ai
51

code

Implement production code changes from a provided spec.

Official
Intermediate
LisztVanLisztVan
2

verification-before-completion

Require fresh command output and exit codes before reporting task completion.

Community
Basic
SirTheOracleSirTheOracle

adversarial-lite

Generate adversarial plans for small tasks using isolated Codex subagents.

Official
Intermediate