@evgheni_D 100%. ADIM treats debugging like a differential diagnosis you formulate competing hypotheses first, then pick the discriminating probe that produces divergent outcomes, effectively cutting the hypothesis tree in half.
Why do AI coding agents keep failing at complex, non-trivial bugs?
Because they have zero memory of what already failed. They loop through the same guesses until your context window explodes.
Today, Iโm open-sourcing Adaptive Debugging Intelligence MCP (ADIM) ๐งต๐
@evgheni_D Yep, exactly. It forces the agent to pre-commit to what output would prove the hypothesis right.
If the result doesn't match, that hypothesis gets killed/weakened and it triggers a pivot to a different branch.
@evgheni_D Spot on โ thatโs why ADIM isn't just a memory log, but an Adaptive Engine.
After each failure, `debug_next_experiment_prepare` scores open hypotheses via explainable heuristics (partial wins, info value, research backing) & penalizes duplicate paths to recommend ONE next action.
6/ 138 hermetic tests passing. 22 MCP tools. Built with Bun, TypeScript, and SQLite.
100% open-source under MIT.
Check it out, star the repo, and let me know your thoughts! โญ
https://t.co/8nwFLTzq8S
5/ Features an interactive standalone HTML dashboard (`bun run db:dashboard`) to explore incident timelines, telemetry, and candidate trade-offs.
Works with ANY MCP client:
โข Claude Desktop
โข Cursor & Windsurf
โข Cline & Roo Code
โข OpenCode CLI & Antigravity IDE
@cmfbyNothing Hi, I received a message claiming I can earn โน200 by filling a survey at a link: https://t.co/COMYoxOfYm. Is this genuine or spam? Please confirm. #CMF#NothingPhone#ScamAlert
Love that #Brave lets me support creators! Just tipped https://t.co/Ks07uoqQmu with #BAT. If they verify at https://t.co/b7GqSORGGd before Jun 20, 2023, the BAT is theirs! #TipWithBrave