When I give an AI coding agent a task, I never actually know if the context I sent it was enough. So I built the check.
The stack is four small tools, all Python standard library, all MIT, all local:
- GhostSlice cuts a project down to a task bundle and reports the token cut. On my own bench repo: 74,692 tokens full, 29,926 sliced.
- GhostVerify checks the bundle against the task before an agent runs, and says gaps_found when something reachable is missing.
- GhostBench runs the same task full vs sliced and, when you give it a test command, executes your tests in a contained runner (scrubbed environment, your secrets never reach the test process) and labels the result honestly: static analysis is not called a live run.
- GhostConsole renders one board from all of it, and its Verified column now names the strongest evidence behind it and counts the checks that are missing.
A fifth tool, ghostworkflow, runs the chain in one command, and a small local web app puts a face on it.
I ran the whole thing on its own repo and published the result warts and all: verdict attention, 2 verify gaps, and a test failure that turned out to be a path assumption in one of my own tests. The first dogfood run also caught a real bug in the contained runner (it hid the user site-packages, so pytest vanished), which is fixed and shipped.
I am looking for 5 to 10 developers who use AI coding agents daily to run it on one of their own repos and tell me where it breaks or bores them. Quickstart is 6 clones and one command. Links and the case study: https://gist.github.com/littlestjames82-sys/7e8309315c89706aefa97de6a4cc5399
What I want to know most: did the verify verdict match what your agent actually struggled with?
Top comments (0)