[Project] I built a Codex tool that compares proposed file edits with what actually changed — now with reconciliation cards

Ever asked Codex to refactor a couple files, then ended up squinting at git diff afterwards just to make sure it didn't sneak in a change to a production config, .env , or something totally off-task? I built CLIM — a local, open-source, hook-driven observer for Codex. Most AI code review tools are just another LLM checking your LLM-generated code. CLIM takes a different angle: it cross-references what the agent said it would edit against what actually changed on disk, and clearly flags what it couldn't verify. New in v0.3.0-alpha.1: the Execution Reconciliation Card . It pulls the important stuff into one terminal report: Proposed vs. Observed : compare a tool call's declared file targets against real changes on disk Review-first signals : protected-file changes, out-of-proposal edits, and missing Hook callbacks all surface first Less noise : routine events get collapsed, but full counts and underlying evidence stay available Explicit UNKNOWNs : clear markers where Shell effects, task authorization, or attribution can't be reliably confirmed It uses Codex's native PreToolUse / PostToolUse hooks and local filesystem observations. No extra LLM calls, no external AI review service. 👇 Visual demo (synthetic wrapper around real CLIM reconciliation logic; actual CLI output is plain text): preview.redd.it/nars9bmasmwh1.png 👇 Actual plain-text output from a real, controlled Codex run: preview.redd.it/2yvkwtrhsmwh1.png Quick context on that run: we asked Codex to fix two Python files and run tests, plus explicitly had it run a local script that modifies config.prod.json as a controlled test. CLIM surfaced the protected config change, collapsed routine windows, and clearly marked the limits of what it could verify. Proposed: UNKNOWN Observed: config.prod.json Authority: UNKNOWN Attribution: UNCERTAIN After setup and trusting the project hooks through Codex /hooks , generate your own with: clim report --project /path/to/project --format card --limit 20 A few important boundaries: CLIM Shadow only observes; it never blocks tool calls It doesn't replace git diff , prove code correctness, or guarantee complete filesystem coverage Changes outside a tool's proposal aren't automatically unauthorized changes This is still a Developer Alpha. We're looking for real-world feedback, and we haven't measured alert precision or review-time savings yet. GitHub: github.com/LioEinaudi/CLIM Operator Review Feedback Template: .github/ISSUE_TEMPLATE/operator_review.md Curious what everyone's most commonly missed thing is when reviewing AI-generated code? And which warnings would just end up as noise for you?