alex-mvp 发布 Alex：像资深工程师一样审查每个 PR

Alex reviews every pull request like a senior engineer:
a verdict you can read in five seconds, and the exact fix in thirty.

What a review looks like

Every review posts three things:


Summary card	Impact, merge verdict, confidence score from 0 to 5, numbered findings, file breakdown, a diagram of the change, and a review counter. Updated in place on every re-review.
Inline comments	On the exact line, with 🔴 P0, 🟠 P1 or 🟡 P2 badges and a one-click Commit suggestion fix. Never posted twice.
Status check	`Alex review · Confidence 4/5 · 1 P1` on the commit. Can block merging while a P0 is open.

The confidence score is the model's judgment inside hard limits, so it can never contradict the findings: any P0 caps it at 2, two P1s at 3, one P1 at 4. Details in the design system.

How it works

flowchart LR
  A["PR opened or /alex"] --> B["Gate: trusted author, not a draft"]
  B --> C["Engine reads the diff and the repo"]
  C --> D["Card: summary, score, diagram"]
  D --> E["Inline comments, card, status check"]

Alex checks out the base branch and reads PR code as git objects. It never runs code from the PR, so it is safe on untrusted changes.

Add Alex to a repo

Copy examples/workflows/alex-review.yml to .github/workflows/alex-review.yml.
Add the repo secret ALEX_LLM_API_KEY: any OpenAI-compatible key. The default endpoint is Fireworks.
Open a pull request. Alex reacts 👀, reviews, then 👍.

Command	What it does
`/alex`	Review the PR again

More commands (/alex review this at plus, questions, explain, tests) arrive in phase 2.

Optional: a bot with its own name and avatar

Create a GitHub App (permissions: Contents read, Pull requests write, Issues write, Commit statuses write), install it on your repos, then set the variable ALEX_APP_ID and the secret ALEX_APP_PRIVATE_KEY. Use docs/assets/alex-logo.svg as the avatar.

Action inputs

Input	Default
`llm_api_key`	required	API key, from a secret
`llm_url`	Fireworks	OpenAI-compatible base URL
`model`	DeepSeek V4.1 Flash	Model that finds issues
`card_model`	same as `model`	Model that writes the card
`token_budget`	`1500000`	Spending cap per review
`fail_on_p0`	`true`	Status check fails while a P0 is open
`engine_version`	`1.12.10`	Pinned engine version
`app_id`, `app_private_key`	empty	Optional bot identity

Roadmap

Alex is built phase by phase, and each phase is compared side by side with Greptile on a lab repo of real historical bugs.

Phase	Scope
1	Triggers, summary card, inline comments, safety gate, status check	✅ done (comparison)
2	`@alexreview-ai` bot identity, review tiers (base, plus, apex), `.alex/` config, install anywhere, questions	planned
3	Security scan, auto-approve, related-repo context	planned
4	Learning from reactions and replies, analytics, CLI	planned
5	Run tests in a sandbox, codebase index, hosted app	optional

Full plan: docs/blueprint.md.

Repository layout

alex/                 review logic: card, scoring, inline comments, GitHub client
tests/                unit tests
action.yml            the GitHub Action
examples/workflows/   workflow file to copy into a repo
docs/                 blueprint, design system, lab cases, comparisons, assets
evals/aacr/           AACR-Bench harness (industry benchmark)
evals/lab/            lab tooling: run the engine on a case, redact and scan snapshots
scripts/              developer helpers

Develop

python -m unittest discover -s tests -v
python -m alex.main preview --ocr-json engine.json --diff pr.diff --title "PR title" --out out --no-llm

Standard library only, nothing to install. Conventions: docs/engineering.md.

License

MIT