kimi-code-bot

Repo-aware AI code review bot, powered by kimi-cli.

Instead of dumping a diff into a model, kimi-code-bot lets the review agent browse the repository like a human reviewer (read-only tools provided by kimi-cli), then suppresses false positives with multi-pass parallel review + cluster voting + an independent verifier. All LLM/agent-loop concerns are delegated to kimi-cli — this repo is a thin orchestration layer (a Python port of hoverstare's architecture, minus rig).

💡 Like this project? Also check out hoverstare — the original Rust implementation of the same architecture (powered by rig instead of kimi-cli), with a self-hosted serve mode and an agent develop mode.

Features

• 🔍 Targeted verification: the agent reads files / greps / checks git show base:file before claiming a bug
• 🗳 Multi-pass voting: 3 lenses (correctness / concurrency / security) run in parallel; findings need ≥2 votes, single-vote findings go to a verifier ("rejection needs evidence, doubt favors keeping")
• 💬 Inline review comments with severity (🔴🟠🟡🔵), suggestion blocks, and drift-immune fingerprints
• ⏩ Incremental review on synchronize: only the delta since the last review is re-analyzed
• ✅ Auto-resolve fixed threads; ✅ confirmed fixed fallback without a PAT
• 🏷 Status checks: kimi-code-bot / kimi-code-bot-findings
• 🗣 @kimi-code-bot review|explain|help commands (collaborators only)
• 🛡 Fail-open: analysis failures never redden your CI (exit 0); config errors and publish failures exit 1

Usage (GitHub Action)

name: kimi-code-bot
on:
  pull_request: { types: [opened, reopened, synchronize] }
  issue_comment: { types: [created] }
  pull_request_review_comment: { types: [created] }

permissions:
  pull-requests: write
  issues: write
  contents: read
  statuses: write

env:
  KIMI_API_KEY: ${{ secrets.KIMI_API_KEY }}

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: liuchong/kimi-code-bot@v0.0.1
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}

LLM credentials belong to kimi-cli: set KIMI_API_KEY (+ optional KIMI_BASE_URL / KIMI_MODEL_NAME) as job env, or any provider kimi-cli supports.

Configuration

.github/kimi-code-bot.toml in the reviewed repo (all fields optional):

model = "kimi-for-coding"              # kimi-cli model alias
reformat_model = "kimi-for-coding"     # cheap model for the JSON reformat pass
passes = 3                             # review lenses (1-3)
verify = true                          # verifier for single-vote findings
severity_threshold = "medium"          # below this -> Nitpicks
ignore = ["*.lock", "dist/**"]
max_diff_kb = 400
max_tool_calls = 20
timeout_secs = 900
review_drafts = false
fail_closed = false                    # true: analysis failures exit 1
status_checks = true
language = "en"                        # en | zh
mention = "@kimi-code-bot"             # @command trigger (avoid collisions with real users)
instructions = ["AGENTS.md"]           # loaded from the BASE branch

Every field can be overridden with a KIMIBOT_* env var (env > toml > defaults).

Local CLI

uv tool install kimi-code-bot
export GITHUB_TOKEN=... KIMI_API_KEY=...
kimi-code-bot review --repo owner/name --pr 123 [--full] [--dry-run]

Development

uv venv && uv pip install -e '.[dev]'
uv run pytest

Layout: agent.py is the only module allowed to import kimi_cli (the backend switch point); everything else is framework-agnostic orchestration.

License

Apache-2.0