Model swap silently broke agent cancellations, so a developer built a behavior diff tool
Impossible-Alarm-738 · reddit · 2026-07-22
A model swap quietly broke an agent’s cancelsubscription tool calls, so the author built whatbroke, a local diff tool for agent behavior.
- It compares JSONL traces of real agent trajectories, not just text outputs.
- It tracks tool calls, arguments, order, cost, latency, and final output, then exits non-zero on breaking changes so it can run in CI.
- A proxy mode (whatbroke record) captures traces from any language by pointing the base URL at it.
- Because agent runs are nondeterministic, it supports multiple runs per scenario and reports a flap rate so pre-existing randomness gets downweighted.
- It is deterministic, fully offline, requires no API keys or accounts, and is MIT licensed.
Related event: New Tool Detects Silent Agent Regressions After Model Switches(2 posts)→
More from coding & agent
- Letta pitches OS-style long-term memory for local LLM agents — thisguyknowsai · 2026-07-22
- Agent builders should map task complexity before copying Claude Code-style harnesses — hugobowne · 2026-07-22
- Tencent’s agent products split apart as WorkBuddy rebounds to 7.31 million — FinanceYF5 · 2026-07-22
- OpenClaw slid to 4.66 million after Claude cut third-party access, while Hermes rose to 6.16 million — FinanceYF5 · 2026-07-22
- Sensetime’s free public beta lifted Code Xiaoyuanxiong to 3.57 million and Office Xiaoyuanxiong to 3.16 million — FinanceYF5 · 2026-07-22
- Tencent CodeBuddy reached 7.31 million monthly visits six months after launch — FinanceYF5 · 2026-07-22