Source note

SetupX: Can LLM Agents Learn from Past Failures in Functionality-Correct Code Repository Setup?

LLM AgentsRepository SetupCode IntelligenceAutomated Software EngineeringVerification

SETUPX is an LLM-agent system for setting up code repositories so their documented commands and tests run in a reproducible container. It learns reusable setup fixes from past runs, tests them with Docker rollback, and checks results with a separate prosecutor-judge audit.

  • Repository setup often fails because dependencies, build tools, package installs, and verification steps differ by project.
  • Current agents tend to treat each repository as a new case, so they repeat failures instead of reusing fixes from similar repositories.
  • Bad setup can look successful when a build or install command exits cleanly, which matters because later software tasks depend on a working execution environment.
  • SETUPX stores past setup fixes as eXPerience Units (XPUs). Each XPU contains error signals, plain-language advice, executable repair commands, and success/failure telemetry.
  • A retriever searches XPUs using the current error state, vector similarity, historical success rate, and an LLM reranker that selects the top 3 entries from 10 candidates.
  • The agent tries retrieved fixes inside Docker containers with a LIFO snapshot stack, so it can roll back after failed or later-harmful install attempts.
  • An in-loop verifier runs read-only checks during setup and separates setup-induced failures from project-intrinsic bugs.
  • A post-hoc Prosecutor-Judge protocol has one agent gather concrete failure charges and another agent independently verify each charge before the final pass/fail verdict.
  • On a curated 100-repository Python benchmark from EnvBench, SETUPX+XPU reports a 92% pass rate.
  • It beats Claude Code by 19 percentage points, ExecutionAgent by 33 points, and other specialized tools by 47–54 points under the same prosecutor-judge adjudication.
  • SETUPX without XPU reaches 82%, so the XPU memory adds 10 percentage points over the base agent loop.
  • On high-difficulty repositories, SETUPX+XPU reaches 79%, compared with 63% for Claude Code and 72% for SETUPX without XPU.
  • In domain failure counts, SETUPX+XPU reduces failures from 27 to 8 versus the strongest CLI baseline; in Toolchain repositories, it has 0/20 failures versus 5/20 for the baseline.
  • On 22 multi-repository scenarios, SETUPX and Qwen Code both get Full or Mostly on 17/22, but SETUPX gets 6 Full verdicts versus 1 for Qwen Code.