Check a coding agent's claim that the work is done
Before a coding agent may stop, compare what it says it did with a ledger of what it actually ran, with Jev judging the gap.
The use case
Developers using Claude Code or Codex find agents sometimes report a task done without running tests or touching the right files. These hooks log the commands and edits the agent really made and, when it tries to finish, ask whether the claim is backed by a passing check; one also scores the work on quality measures as it goes. An unsupported finish is refused and the agent is told why. The developer still reviews the diff and test output before merging, since a check passing is not the same as the code being right.
Example workflow
- Install a Stop or PostToolUse hook in Claude Code or Codex CLI
- Log commands run and files changed during the session
- At finish, ask Jev whether the done claim is supported by the log
- Refuse the finish and return the reason if unsupported
- The developer reviews diff and tests before merging
Tools
- Jev (TypeSafe AI)
- Claude Code
- Codex CLI
Risks and guardrails
- Hooks can loop if the agent cannot satisfy the check
- Passing tests can still be the wrong tests
- Extra calls add latency to every finish
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Canny by Qkal on Made with Jev: madewithjev.com/builds/canny
- Canny: original post or repo by Qkal: github.com/qkal/Canny
- jev-review by Niaz Morshed on Made with Jev: madewithjev.com/builds/jev-review
- jev-review: original post or repo by Niaz Morshed: x.com/niazmorshed_/status/2100465662867218857