An AI coding agent rips out a core invariant across 189 files of a 717,000-line TypeScript app in three days, with no human code review
A single-developer case study on arXiv. The author's own VS Code AI-agent extension (717,725 lines of TypeScript, 3,648 files) needed streaming generations to survive the closing of their UI panel, a change he judged infeasible through incremental refactoring. Under a specification-first protocol (agent-written spec, 14 refinement audits, atomic implementation, 17 verification audits), the agent fixed 201 defects before any human ran the program. No human reviewed the code, and no bugs were seen across about thirty later sessions.
The numbers (as reported)
| Codebase | 717,725 lines of TypeScript, 3,648 files |
|---|---|
| Change | 189 files touched (31 new); 288 files, +34,770 / -16,422 across both commits |
| Time | Three days |
| Cost | USD 2,430 in model inference |
| Audit passes | 31 (14 spec refinement + 17 verification) |
| Defects fixed before first human run | 201 |
Facts
- Company / subject
- Joel Abenhaim (AI Sovereign Labs)
- Tools
- AICode
- Models
- OpenAI ChatGPT 5.6 Sol
- Codebase
- brownfield
- Work type
- Refactor / maintenance
- Stack
- TypeScript, React, VS Code extension
- Published
- 2026-08-12
- Last checked
- 2026-10-20
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- 189 files changed in a 717,725-line codebase in three days for USD 2,430 in inference 3 days; USD 2,430
“Elapsed: three days; cost: USD 2,430.”
- 201 defects corrected across 31 audit passes before any human ran the program 201 defects
“Across 31 audit passes, 201 defects were corrected before any human executed the program.”
Caveats
Self-reported: the author built the codebase, ran the operation and wrote the paper, and his company (AI Sovereign Labs) designs and distributes the agent used, so this is partly vendor. Single case; the codebase is closed source, though the build and raw logs (in French) are published.
Related topics: Refactoring and maintenanceBrownfield codebases