PR Pre-flight

A Claude Code plugin that reviews your changes for what a reviewer would flag — before you open the PR. Not a full code review: it only reports the things a reasonable reviewer would actually comment on.

It catches:

• Convention drift — naming or error-handling that doesn't match the file's own neighborhood (read from the surrounding unchanged code and sibling files, not an abstract style guide).
• Missing tests — non-trivial new logic in a file that has a sibling test file which wasn't touched.
• Leftover debug output — console.log, print(, debugger, commented-out code.
• Unexplained TODO/FIXME — added with no ticket or explanation.
• Hardcoded secrets — API keys, tokens, .env values.
• Unused imports/variables introduced by the diff.

Install

/plugin install pr-preflight@github Samanyu-dev/pr-preflight

Or for local development:

claude --plugin-dir /path/to/pr-preflight

Usage

/pr-preflight:preflight

Diffs your current branch against the repo's default branch (auto-detected from origin/HEAD, falling back to a local main/master), plus anything uncommitted. To diff against a specific ref instead:

/pr-preflight:preflight some-branch

If there's nothing to review (clean tree, no commits ahead of default), it says so and stops — it won't invent findings.

On a large diff (~15+ files or ~800+ changed lines), it states the size up front and drops generated files, lockfiles, vendored code, and pure renames from consideration, naming what it dropped. For what's left, if the diff is genuinely too large/varied to review directly at full depth, it fans out — one file-group-reviewer subagent per group of files — instead of reviewing everything shallowly in one pass. In practice this rarely triggers: Claude's own context window comfortably handles direct review well past that size, so the thresholds are a safety valve for the genuinely huge case, not a hard rule.

Repo-specific config

Drop a .preflight.md at your repo root to extend or override the defaults. It's plain prose — read and applied by the command itself, not parsed by a schema:

# Repo preflight config

- Ignore everything under `vendor/` — third-party code we don't own.
- We use `snake_case` for function names in `src/`, not camelCase.
- Treat diffs under 1500 lines as "small" — our files are just long.

One rule never bends: a path exclusion silences style review for that path, but a hardcoded secret found in newly-added code is always flagged regardless of .preflight.md.

Don't want to write it by hand? Generate a starting point instead:

/pr-preflight:init

Scans the repo for its dominant naming convention, where tests actually live (sibling files vs. a separate test/ directory), and any tracked vendored/generated directories, then writes .preflight.md from what it found — not a generic template. If the naming convention is genuinely mixed, it says so rather than forcing a rule. Won't overwrite an existing .preflight.md unless you pass --force.

Auto-fixing the safe findings

/pr-preflight:fix

Runs the same review as /pr-preflight:preflight, but applies the findings that are mechanical and safe to change without a design decision (leftover debug output, unused imports) directly, and reports the rest as usual. Secrets, missing tests, TODOs, and naming/error-handling drift are never auto-applied — those need a human decision, and a secret needs rotation, not just deletion.

Reminder before you push or open a PR

The plugin also installs a hook: right before Claude runs git push or gh pr create, it surfaces a one-line reminder to run /pr-preflight:preflight first, if you haven't this session. It doesn't block the command — it's a permission-confirm prompt with the reminder as the reason, so you can approve and continue immediately. Any other command is untouched; the hook only reacts to those two.

Usage stats (maintainer)

There's no install-count API for Claude Code plugins. scripts/stats.sh pulls the closest real signal from GitHub instead — stars/forks (public, unlimited history) and clone/view counts (needs push access, 14-day window):

./scripts/stats.sh

Why this and not just asking Claude to "review my diff"

A generic review prompt checks against a style guide it's guessing at. This one reads the actual file and its siblings first, so what it flags is a real deviation from how this code already looks — not a generic lint pass.