gemini-review

The model writes code, commits it, and nobody reads the diff before it lands: a hardcoded key, a bug the change brings in, or a change that is not what you asked for goes into history. This mod has Gemini review every commit the model makes. Before a Bash git commit runs, Gemini reads the change the commit records and the conversation. A blocking finding stops the commit and the model reads why; a minor finding lets the commit run and the model reads it after the result.

What it does

1. The mod hooks the Bash tool. A command with a git commit in it (the commit skill included) is reviewed before it runs; any other command runs untouched.
2. It reads the command without a shell: cd <dir> and git -C <dir> set the directory, git add before the commit in the same command says what gets staged, and commit -a, add -u and add -A say that every tracked or untracked change goes in. Paths given to the commit itself (git commit -m x -- a.ts) mean git records those paths from the working tree and nothing else, and the review reads just those. A commit message in a heredoc or in quotes is not read as a command. git commit --help, -h and --dry-run record nothing and are not reviewed. A command that runs anything else before the commit (echo x >> f && git commit -am ...) is stopped before it runs, and the model reads that it must commit in a Bash call of its own. The review reads the change before the command runs, so what those steps change would not be in it: such a commit went through with an empty diff and no review (measured on 2.1.278). Commands after the commit (&& git push) are allowed.
3. It collects the change with git, run by argv in the Bash tool's directory ($.session.cwd() follows a Bash cd, measured on 2.1.278): the index, a path the command stages from the working tree (the index still holds its older content until git add runs), and each new file whole, at most 200 of them. Each git call has 20 seconds. When the change is empty, the commit runs with no review and no line.
4. It sends the diff and the conversation in one generateContent request with a schema: a list of findings, each blocker or minor, with a file, a line and a message. The diff always goes whole. When the conversation is over maxInputChars (2,000,000 characters by default), its longest tool outputs are cut to one common length, each keeping its head and tail. gemini-core builds the request with the key, the model and the thinking level it holds for gemini-review, and reads the answer.
5. blocker means a bug the change introduces, data loss, a security hole, a secret or credential in the diff, or a change that contradicts what you asked for. Everything else is minor.
6. The verdict:
   • a blocker: the command does not run; the model reads each blocker and the minor notes, with the instruction to fix and commit again, or, when a finding is wrong, to tell you why and run the same command with GEMINI_REVIEW_SKIP=1 in front;
   • only minor findings: the commit runs and the model reads the notes after the result;
   • no finding: the commit runs and the model reads one line saying the review found nothing.
7. When the review cannot answer (no key, an HTTP error, a malformed answer or one cut at the output limit, a git error, a diff over 1,500,000 characters), the commit runs, a transcript line says why, and the model reads the reason.
8. Gemini answers HTTP 503 ("high demand") now and then, often after 10 seconds or more. gemini-core then has the mod ask again after 1 s, 2 s and 3 s, at most four attempts in all, and no attempt starts whose wait would end past 60 s. A hook's 10-second budget counts its $.clock waits but not its requests, so the waits stay short. After a 429 or a key error, gemini-core hands over the request with its next key, when it holds one.

In a live check on 2.1.278 with gemini-3.8-flash, a commit of a file with sk_live_... was stopped with sub/pay.ts:1: Hardcoded live Stripe secret key committed in source code, a git add pay.ts sub.ts && git commit after the fix ran, a GEMINI_REVIEW_SKIP=1 commit ran unreviewed, and an invalid key let the commit run with Gemini HTTP 400: API key not valid. Of eight reviews in that session, four got an answer (one after two 503s, 42.5 seconds in all) and four got only 503 answers and let the commit run.

What it shows

After each review a toast stays for 10 seconds, and /gemini-review shows the last one:

gemini-review: reviewed 1 file(s) · 1 blocker, 0 minor · 2k in, 515 out · sent to Gemini free tier

The sent to Gemini free tier part appears only on the free tier. A transcript line follows every review, so you read what the model was told. The line holds the findings alone, without the instruction:

gemini-review: commit reviewed: 2 file(s), nothing to report
gemini-review: commit reviewed with 1 minor note(s): pay.ts:12: Name the constant.
gemini-review: commit stopped: reviewed 1 file(s) · 1 blocker, 0 minor · 2k in, 515 out
gemini-review: commit ran without a review: the model used GEMINI_REVIEW_SKIP=1

The context and the line are separate channels: the model never reads the line, and you never read the context.

Command

/gemini-review              on or off, the model, thinking level and tier gemini-core holds, whether a key is set, the last review
/gemini-review on | off     off: commits run without a review; on is refused while gemini-core has no key
/gemini-review reset        off again, the default

The review is off after an install, so nothing goes to Gemini before you set a key and turn it on.

The key, the tier, the model (default gemini-3.8-flash) and the thinking level belong to gemini-core:

/gemini-core model review gemini-3.7-flash
/gemini-core thinking review low
/gemini-core paid

Free tier or paid tier

Every review sends the diff and the conversation: your prompts, the commands the model ran and the contents of the files it read. On the free tier Google may use them and human reviewers may read them; the gemini-core README quotes the Gemini API Additional Terms. On a project you would not show to Google, use a key with billing enabled and set /gemini-core paid.

Install

claude plugin marketplace add KilimcininKorOglu/claude-code-mods
claude plugin install gemini-review@kilimcininkoroglu-mods

It depends on gemini-core, which claude plugin install adds. Function hooks are early access. Claude Code 2.1.288 and later load them by default, so there is nothing to switch on.

After installing

1. Set the Gemini key and the tier in gemini-core, as its After installing section says, then restart Claude Code.
2. Run /gemini-review on. Without a key it answers still off: gemini-core has no Gemini key and stays off.
3. Run /gemini-review. The first line reads on · <model> · thinking ... · <tier> tier · key set.
4. When a commit runs with commit ran without a review: Gemini HTTP 429, the model has no quota on your key. Pick another with /gemini-core model review.

After an update from 0.1.x: claude plugin update does not add gemini-core (measured on 2.1.278), so run claude plugin install gemini-core@kilimcininkoroglu-mods once. Version 0.2.0 moved the key, tier and model to gemini-core; the apiKey, tier and model options and the settings /gemini-review free|paid|model stored before are no longer read, so set them again in gemini-core. Version 0.3.0 made the review off by default: after an update from an earlier version it is off unless you ran /gemini-review on before, so run /gemini-review on once.

Options

| Option | Default | What it sets | |---|---|---| | maxInputChars | 2000000 | Characters of conversation sent at most, 10,000 to 4,000,000; the diff always goes whole |

A value outside the range, or one that is not a whole number, falls back to the default.

What it can reach

Validated with claude plugin validate on Claude Code 2.1.283:

❯ ./register.ts hooks: session.start, command.run{command=gemini-review}, tool.call{tool=Bash}
❯ ./register.ts calls: $.clock.now (via askGemini), $.clock.sleep (via askGemini), $.command.register, $.gemini.enroll, $.gemini.read (via askGemini), $.gemini.request (via askGemini), $.gemini.settings (via review, runCommand, storeEnabled), $.http.fetch (via askGemini), $.process.run (via collectDiff, git), $.session.cwd (via review), $.session.messages (via review), $.store.delete (via runCommand), $.store.get (via isEnabled), $.store.set (via storeEnabled), $.ui.log, $.ui.toast (via verdictOf)

Reach L3, reaches the network.

1. Reads:    each Bash command; at a commit, the repository's diff and new files through git, and the conversation (messages, tool inputs and outputs); its own $.store; from gemini-core, the request with the key
2. Runs:     read-only git by argv, no shell: rev-parse, diff, ls-files; at most 200 new files are read
3. Sends:    the diff and the conversation, one request per commit (up to four after a 503, and once more per extra key after a 429 or a key error), to the URL gemini-core builds (generativelanguage.googleapis.com) with the key in the x-goog-api-key header, never in the URL
4. Persists: in $.store, the on/off setting; the last review line lives in memory
5. Hostile input: a diff or a conversation can steer Gemini's findings, so a hostile change can pass or a sound one be stopped; the model reads the findings as a denial or a note and can skip a wrong one; paths from the command reach git as argv after --, never through a shell

Limits

• A review is a model's opinion. It can miss a problem, and a wrong blocker is passed with the skip prefix, which the model uses by its own decision.
• Only the model's Bash commits are reviewed. A commit from your terminal, from a script the model runs, from an alias or through another tool is not.
• The command reading covers the usual forms. A cd inside a subshell, variables in paths and globs that git expands are not resolved.
• A subagent's commit sends only the diff, because which transcript $.session.messages() answers inside a subagent was not verified.
• Every review sends the whole conversation, so a long session makes each review larger and slower. $.session.messages() answers the newest 4096 messages.
• The Bash hook ran 42.5 seconds without a limit in the live check. A longer limit was not measured.

Development

make install     # eslint, typescript-eslint, typescript
make lint        # complexity limit 10, the build fails above it
make typecheck   # needs .claude/types/ from /plugin-types
make validate
make test        # claude plugin test