docc-code-review

docc-code-review

Independent code review via a second AI model (OpenRouter) to catch what self-review in the same session misses.

Tools required

• Bash (git commands)
• openrouterai MCP (chat_completion)

Configuration via env vars

Var	Default	Purpose
DOCC_REVIEW_MODEL	deepseek/deepseek-v4-pro	Primary review model
DOCC_REVIEW_FALLBACK	qwen/qwen3-coder-next	Fallback model
DOCC_REVIEW_STACK	(empty)	Extra stack context injected into system prompt

Set DOCC_REVIEW_STACK to describe your project's tech stack for more relevant feedback:

export DOCC_REVIEW_STACK="Next.js 15 App Router, TypeScript strict, tRPC v11, Drizzle ORM"

---

Workflow

Step 1: Get the diff

git branch --show-current

# On a feature/story branch:
git diff origin/main...HEAD
# or against dev:
git diff origin/dev...HEAD

# Already merged — diff the merge commit:
git log --merges --oneline -5
git diff <merge-commit>^1 <merge-commit>

Also look for a story/spec file if the project uses BMAD:

ls _bmad-output/implementation-artifacts/ 2>/dev/null | grep <story-id>

If found, read it for acceptance criteria context.

Step 2: Build system prompt

Base prompt:

You are a senior software engineer doing a thorough code review.
Focus on correctness, security, and maintainability.

Rules:
- No type-unsafe code without justification
- Auth/permission checks on all mutations
- No hardcoded secrets or credentials
- DB migrations must be additive only — never modify existing migrations
- Error handling must be explicit

Output — skip style nitpicks, be direct:
## 🔴 Bugs / Logic Errors
## 🟠 Security / Auth Issues
## 🟡 Code Quality Issues
## 🟢 Looks Good
## ⚪ Out of Scope

For each issue: file + line + description + suggested fix.
If nothing critical — say so explicitly.

If DOCC_REVIEW_STACK is set, append: Stack: <value> If story ACs were found, append them to the user message.

Step 3: Call OpenRouter

Call openrouterai:chat_completion with:

• model: DOCC_REVIEW_MODEL (default deepseek/deepseek-v4-pro)
• messages: system prompt + user message with full diff (+ ACs if available)

If primary model errors, retry once with DOCC_REVIEW_FALLBACK.

Step 4: Report

Present model output as-is. Prepend one-line summary:

Review (<model>): X critical, Y warnings. [Fix before merge / Clear to merge]

Categorize findings:

• Fix before merge — bugs, security holes, broken migrations, business rule violations
• Backlog — non-blocking improvements worth tracking

Do not auto-fix. Wait for human decision.