AgentChanti
An autonomous coding-agent CLI and Python library that plans, codes, reviews and tests a change, then reports separately whether the run completed and whether anything the agent did not write actually verified the result.
AgentChanti distinguishes "completed", the plan ran and every step's own check passed, from "verified", something the agent did not write in that run independently agreed. When the two differ, it says so plainly instead of reporting overall success.
Evidence only counts from a user's own acceptance commands, a pre-existing test the run left byte-identical, or a contract written before any code existed; if the agent edits that contract, it stops counting. A check the agent authored itself can earn a green result but cannot fail the code, since a generated contract is not treated as independent proof. The project is copied before the agent reads a file, and a restore command reverts it, with no git dependency required.
On an eight-task benchmark using external probes the agent never sees, the project reports ground truth of 7 out of 8 with zero false-positive successes. It works with local LLMs through Ollama or LM Studio, or cloud providers including OpenAI, Gemini and Claude, and installs with pipx or pip.