ChatGPT as your junior code reviewer, with all the reliability that implies

What it does

This action grabs a pull request’s diff, ships it to ChatGPT via a session token, and posts the model’s response as a PR comment. You configure it in a workflow file, pick “YOLO mode” (everything at once) or an experimental split-by-file mode, and let the LLM do the talking.

The interesting bit

The author is admirably upfront about the whole thing being experimental and not production-ready — a rarity in the “slap GPT on it and ship” ecosystem. The split feature, which chunks large diffs per-file to dodge 403 errors, exists but the README warns it “produces very bad results” with too many changed files. There’s even a linked example of the meltdown.

Key highlights

• Requires a CHATGPT_SESSION_TOKEN secret — you’re essentially handing your ChatGPT session to CI
• Two modes: “yolo” (monolithic prompt) or split: true (per-file, unstable)
• Will 403 if prompts get too long or trigger some opaque “attack” heuristic
• Author explicitly flags misleading comments as a risk to contributor sanity
• Future ideas include VS Code extension and suggested changes as commits — neither implemented

Caveats

• Code quality is self-described as “experimental”
• ChatGPT’s reviews can be confidently wrong, which is worse than no review
• The split feature’s instability is documented with a real failure example

Verdict

Worth a look if you want to prototype LLM-assisted review in a low-stakes repo. Skip it if you need reliable feedback, care about contributor trust, or lack patience for debugging why your CI just gaslit a junior dev about a nonexistent bug.