prev

README

PREV

AI-powered code review CLI tool for diffs, commits, branches, and merge/pull requests.

Supports multiple AI providers: OpenAI, Anthropic (Claude), Azure OpenAI, Gemini, Ollama, Groq, Together, LM Studio, and any OpenAI-compatible API.

Supported AI Providers

Provider key	Service	Type	Typical auth/config
openai	OpenAI	Native	OPENAI_API_KEY
anthropic	Anthropic / Claude	Native	ANTHROPIC_API_KEY
azure	Azure OpenAI	Native	AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, deployment/model
gemini	Google Gemini	OpenAI-compatible	GEMINI_API_KEY
ollama	Ollama	OpenAI-compatible	local base_url
groq	Groq	OpenAI-compatible	PREV_GROQ_API_KEY or config
together	Together	OpenAI-compatible	PREV_TOGETHER_API_KEY or config
lmstudio	LM Studio	OpenAI-compatible	local base_url
openai-compat	Generic OpenAI-compatible endpoint	OpenAI-compatible	config file recommended

Use prev ai list to see the providers available in your runtime and their default models.

How prev Compares to Other Open-Source AI Reviewers

Feature	prev	PR-Agent (Qodo)	CodeRabbit	ai-pr-reviewer	Others
Two-pass review pipeline (walkthrough + detailed)	Yes	No	Yes	No	No
AI providers (OpenAI, Claude, Gemini, Ollama, Groq...)	9	3-4	Proprietary	1	1-2
Offline/local LLM support (Ollama, LM Studio)	Yes	Limited	No	No	Rare
Persistent cross-MR review memory	Yes	No	No	No	No
Markdown-backed memory portability (repo-storable, human-readable)	Yes	No	No	No	Rare
Symbol impact mapping (function reference tracking)	Yes	No	No	No	No
Concurrency/race-risk detection (Go)	Yes	No	No	No	No
Native deterministic precheck before AI (impact + risk hints)	Yes	No	No	No	No
Serena MCP integration (symbol-level context)	Yes	No	No	No	No
Thread continuity (reuse threads across pushes)	Yes	Partial	Yes	No	No
Duplicate-thread suppression on subsequent pushes	Yes	Partial	Partial	No	Rare
Resolved-thread awareness (ignore resolved discussions)	Yes	Partial	Partial	No	Rare
Fixed thread commands via MR comments (prev ...)	Yes	No	No	No	No
Auto-discover repo guidelines (CLAUDE.md, AGENTS.md, copilot-instructions)	Yes	No	Partial	No	No
Strictness levels (strict/normal/lenient)	Yes	Partial	Yes	No	Rare
AI fix prompt blocks in inline comments	Yes	Yes	Yes	No	No
Hunk consolidation (merge findings per hunk)	Yes	No	Partial	No	No
Token-aware batching	Yes	Yes	Yes	No	Rare
Incremental MR reviews using baseline markers	Yes	No	Partial	No	Rare
Unified GitLab + GitHub support in one CLI workflow	Yes	Partial	Yes	No	Rare
Config hierarchy (CLI > env > file > defaults)	Yes	Partial	SaaS	Minimal	Varies
CLI-native (no SaaS dependency)	Yes	Yes	No	GitHub Action	Varies

Installation

# Homebrew
brew install sanix-darker/tap/prev

# From source
go install github.com/sanix-darker/prev@latest

# Or clone and build
git clone https://github.com/sanix-darker/prev.git
cd prev
go build -o prev .

# Docker
docker build -t prev .
docker run --rm -e OPENAI_API_KEY=sk-xxx prev version

For maintainers: Homebrew tap publishing in release automation is guarded by repository variable ENABLE_HOMEBREW=true and secret TAP_GITHUB_TOKEN. If not enabled, release artifacts still publish using .goreleaser.nohomebrew.yaml.

Quick Start

# 1. Set up your AI provider credentials
export OPENAI_API_KEY=sk-xxx          # For OpenAI (default)
# OR
export ANTHROPIC_API_KEY=sk-ant-xxx   # For Claude
# OR run locally with Ollama (no key needed)

# 2. Review a diff between two files
prev diff fixtures/test_diff1.py,fixtures/test_diff2.py

# 3. Review a git commit
prev commit abc123 --repo /path/to/repo

# 4. Review a branch (two-pass pipeline with walkthrough + detailed review)
prev branch feature-branch --repo /path/to/repo

# 5. Optimize a code file
prev optim myfile.py

# 6. Review a GitLab/GitHub merge/pull request
export GITLAB_TOKEN=glpat-xxxx
prev mr review my-group/my-project 42

# 7. Review with strict mode
prev branch feature-branch --strictness strict

Commands

Review Commands

Command	Description
prev diff <file1,file2>	Review diff between two files
prev commit <hash>	Review a git commit
prev branch <name>	Review a branch diff (two-pass pipeline)
prev optim <file|clipboard>	Optimize code

Merge/Pull Request Commands

Command	Description
prev mr review <project> <mr_id>	Review a merge/pull request using AI
prev mr diff <project> <mr_id>	Show MR diff locally (no AI)
prev mr list <project>	List open merge requests

Provider & Config Commands

Command	Description
prev ai list	List available AI providers
prev ai show	Show current provider and model
prev config show	Show current configuration
prev config init	Create default config file
prev config effective	Show merged effective configuration (with env/flags applied)
prev config validate	Validate configuration keys and provider requirements
prev memory show	Show persistent review memory (markdown or JSON)
prev memory prune	Prune old/low-value memory entries
prev memory export <path>	Export memory as markdown/json
prev memory reset --yes	Reset persistent review memory
prev version	Print version info

Branch Review Pipeline

The prev branch command uses a two-pass review pipeline inspired by CodeRabbit:

1. Pass 1 -- Walkthrough: The AI receives abbreviated diffs and diff stats for all changed files, then produces a high-level summary and a changes table.
2. Pass 2 -- Detailed Review: Files are batched by token budget and sent for detailed file-by-file review. The walkthrough summary is included as context so the AI understands the broader purpose of each change. Related walkthrough and detailed-review calls reuse the same logical conversation history to reduce context loss across passes.

Between the two passes, files are enriched with surrounding code context (configurable via --context) so the AI can see the code around each hunk, not just the raw diff.

Branch Flags

Flag	Default	Description
--context	10	Number of surrounding context lines for review
--max-tokens	80000	Maximum tokens per AI batch
--per-commit	false	Review each commit individually
--legacy	false	Use legacy single-prompt review mode
--serena	auto	Serena MCP mode: auto, on, off

Example Output

# Branch Review: feature-branch -> main

## Walkthrough
This branch adds user authentication with JWT tokens...

## Changes
| File | Type | Summary |
|------|------|---------|
| auth.go | New | JWT token generation and validation |
| middleware.go | Modified | Added auth middleware |

## Detailed Review

### auth.go
**auth.go:42** [HIGH]: JWT secret should not be hardcoded...

### middleware.go
No significant issues found.

## Statistics
- Files reviewed: 2
- Issues: 1 (1 HIGH)
- Changes: +145/-3

Strictness Levels

Control review thoroughness with --strictness:

Level	Severity Filter	Description
strict	All (CRITICAL, HIGH, MEDIUM, LOW)	Report all issues including style nits. Be thorough.
normal (default)	MEDIUM and above	Focus on bugs, security, and significant code quality issues.
lenient	HIGH and above	Only report CRITICAL and HIGH severity issues.

# Strict: catch everything
prev branch feature --strictness strict

# Lenient: only critical issues
prev mr review my-project 42 --strictness lenient

Serena Integration

Serena is an MCP (Model Context Protocol) server that provides symbol-level code intelligence. When enabled, prev replaces raw context lines with enclosing function/class bodies, giving the AI better understanding of surrounding code.

Modes

Mode	Behavior
auto (default)	Use Serena if available, fall back to line-based context if not
on	Require Serena; fail with error if not installed
off	Disable Serena entirely

Prerequisites

# Install uv (includes uvx)
python -m pip install --user uv
# or: curl -LsSf https://astral.sh/uv/install.sh | sh

# Serena is fetched automatically via uvx on first use
uvx --from git+https://github.com/oraios/serena serena --help
prev branch feature --serena=on

CI runners must install uv/uvx before running commands with --serena=on.

GitHub Actions example:

- uses: astral-sh/setup-uv@v6
- run: uvx --version

MR/PR Reviews

Review GitLab merge requests or GitHub pull requests directly from your terminal.

VCS provider is auto-detected: if GITLAB_TOKEN is set, GitLab is used; if GITHUB_TOKEN is set, GitHub is used. Override with --vcs. If both tokens are set, gitlab is selected unless you pass --vcs github.

# GitLab
export GITLAB_TOKEN=glpat-xxxx
export GITLAB_URL=https://gitlab.com   # optional, defaults to gitlab.com

# GitHub
export GITHUB_TOKEN=ghp_xxxx

# Review an MR (prints review to terminal)
prev mr review my-group/my-project 42 --dry-run

# Review and post comments to the MR/PR
prev mr review my-group/my-project 42

# Post only a summary comment (no inline comments)
prev mr review my-group/my-project 42 --summary-only

# Use a specific AI provider
prev mr review my-group/my-project 42 --provider anthropic

# Control inline comment filtering
prev mr review my-group/my-project 42 --strictness strict

# Inspect the fully resolved provider/model/config if behavior surprises you
prev config effective

Review Continuity

MR review now keeps a logical conversation history across related AI calls:

• walkthrough -> detailed review batches
• re-review passes in review.passes
• inline finding recovery
• thread replies and top-level reply generation

This continuity is provider-agnostic. When a provider exposes a native response ID it is retained inside the logical conversation, but prev does not require provider-specific chat IDs to preserve context.

MR Flags

Flag	Description
--dry-run	Print review to terminal without posting to VCS
--summary-only	Post only a summary comment, no inline comments
--vcs	VCS provider: gitlab, github (auto-detected from env)
--gitlab-token	GitLab token (or use GITLAB_TOKEN env)
--gitlab-url	GitLab instance URL (or use GITLAB_URL env)
--strictness	Review strictness: strict, normal, lenient
--nitpick	Sensitivity 0..10 for lower-severity suggestions (combined with strictness)
--max-comments	Max inline comments to post (0 = unlimited; prioritizes highest severity)
--review-passes	Number of AI review passes (0 = config/default 1)
--inline-only	Post inline comments only (skip summary notes and thread replies)
--incremental	Review only file-level deltas since the last baseline marker
--filter-mode	Inline filtering mode: added, diff_context, file, nofilter
--serena	Serena MCP mode for MR context: auto, on, off
--context	Number of surrounding context lines used for MR enrichment
--max-tokens	Max token budget used by MR context enrichment
--memory	Enable/disable persistent cross-MR reviewer memory
--memory-file	Markdown memory file path (default: .prev/review-memory.md)
--memory-max	Max memory items injected into each review prompt
--native-impact	Enable deterministic native impact/risk precheck
--native-impact-max-symbols	Max changed symbols used for impact map
--fix-prompt	Include AI fix prompt block in inline comments: off, auto, always
--structured-output	Request/parse JSON findings with markdown fallback
--mr-diff-source	MR diff source strategy: auto, git, raw, api

Persistent Review Memory

prev can keep persistent reviewer memory across MRs in a markdown file:

• Default file: .prev/review-memory.md
• Format: human-readable markdown + a machine-readable JSON fenced block
• Tracks recurring open findings and fixed findings
• Injects relevant history into new MR review prompts

This helps reduce repeated comments on already-fixed issues while keeping pressure on recurring unresolved ones.

# default behavior (memory on)
prev mr review my-group/my-project 42

# custom markdown memory path
prev mr review my-group/my-project 42 --memory-file .prev/team-memory.md

# disable memory for one run
prev mr review my-group/my-project 42 --memory=false

AI Fix Prompt Block

prev can include a structured "AI agent fix prompt" inside inline comments when direct patch suggestions are absent.

• --fix-prompt off (default): disabled
• --fix-prompt auto: only for high-impact findings without direct patch suggestion
• --fix-prompt always: always include

prev mr review my-group/my-project 42 --fix-prompt auto

Native Impact/Risk Precheck

prev now injects a deterministic precheck block into MR review prompts:

• changed-symbol impact map (repo-wide reference counts and changed-file hits)
• concurrency/race-risk signals from changed Go hunks (goroutines, lock symmetry, channel flow, shared map mutations)

Flags:

prev mr review my-group/my-project 42 --native-impact --native-impact-max-symbols 16

Memory Commands

# Show markdown memory
prev memory show

# Show machine JSON
prev memory show --json

# Prune fixed findings older than 45 days and keep max 400 entries
prev memory prune --fixed-older-than-days 45 --max-entries 400

# Export as JSON
prev memory export ./artifacts/review-memory.json --format json

# Reset memory (explicit confirmation required)
prev memory reset --yes

MR Thread Commands

MR comments can control bot behavior using the configured mention handle. By default this is prev; override it with review.mention_handle in config or PREV_MENTION_HANDLE.

Use the plain keyword in sentence text. Do not rely on @prev, because prev may be an actual user handle on the hosting platform.

• prev pause: pause reviews for the MR/thread
• prev resume: resume paused MR/thread reviews
• prev review: force review processing for that thread and clear any prior thread-level ignore state
• prev summary: post one top-level summary note (idempotent)
• prev reply: bot posts a thread reply
• prev ignore: ignore the current finding/thread on future reruns until a reviewer re-requests it with prev review

Inline continuity behavior:

• Finds multiple issues in one changed hunk and posts them as key points in a single inline comment.
• Reuses matching unresolved discussions on later pushes (reply in-thread) instead of opening duplicate new threads.
• Keeps default inline comments short, direct, and low-fluff.
• Keeps default prev reply responses concise; expands only when the reviewer explicitly asks for more detail.
• Strips emojis from bot review comments and replies.

Platform trigger behavior:

• GitHub: the bundled prev-review workflow listens to pull request updates, top-level PR comments, and inline review-comment replies, so prev reply and prev summary can trigger a new review run directly.
• GitLab: the CLI understands the same plain prev commands, but comment-event reruns depend on your GitLab automation. Standard MR pipelines usually run on merge request events, so discussion commands are processed on the next pipeline run unless you add webhook-triggered or note-triggered automation.

Repository Guidelines Mapping

prev can automatically load repository review guidelines and inject them into review prompts so findings align with project conventions.

Auto-discovered files (when present):

• AGENTS.md
• CLAUDE.md
• .claude/*.md (including .claude/CLAUDE.md)
• .github/copilot-instructions.md
• .github/instructions/*.md
• .copilot-instructions.md

This applies to diff, commit, branch, and mr review. For MR review, CI_PROJECT_DIR is used when available (for CI runners).

You can combine this with config-based guidelines (review.guidelines in ~/.config/prev/config.yml).

Text-Only Review Scope

prev reviews text/code changes only. Binary assets are skipped from AI review and inline findings, including formats such as:

• PDFs (.pdf)
• Images (.png, .jpg, .gif, ...)
• Archives/binaries (.zip, .tar, .exe, .so, ...)

AI Providers

Use --provider and --model flags to override the default provider for any command.

Provider resolution order:

1. --provider CLI flag
2. PREV_PROVIDER environment variable
3. provider key in config file (~/.config/prev/config.yml)
4. Fallback to openai

OpenAI (default)

export OPENAI_API_KEY=sk-xxx
export OPENAI_API_MODEL=gpt-4o     # optional, defaults to gpt-4o

prev diff file1.py,file2.py

Anthropic (Claude)

export ANTHROPIC_API_KEY=sk-ant-xxx
export ANTHROPIC_MODEL=claude-sonnet-4-20250514  # optional

prev diff file1.py,file2.py --provider anthropic

Azure OpenAI

export AZURE_OPENAI_API_KEY=xxx
export AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com
export AZURE_OPENAI_DEPLOYMENT=gpt-4o

prev diff file1.py,file2.py --provider azure

Ollama (local, free)

# Start ollama
ollama serve &
ollama pull llama3

prev diff file1.py,file2.py --provider ollama --model llama3

Other OpenAI-compatible APIs

# Groq
export GROQ_API_KEY=gsk_xxx
prev diff file1.py,file2.py --provider groq --model llama-3.3-70b-versatile

# Together
export TOGETHER_API_KEY=xxx
prev diff file1.py,file2.py --provider together --model meta-llama/Llama-3-70b-chat-hf

# LM Studio
prev diff file1.py,file2.py --provider lmstudio --model local-model

# Gemini (OpenAI-compatible endpoint)
export GEMINI_API_KEY=xxx
prev diff file1.py,file2.py --provider gemini --model gemini-2.0-flash

# Any OpenAI-compatible endpoint
export OPENAI_COMPAT_API_KEY=xxx
export OPENAI_COMPAT_BASE_URL=https://your-api.example.com/v1
prev diff file1.py,file2.py --provider openai-compat --model your-model

Configuration

Create a config file at ~/.config/prev/config.yml:

prev config init

Additional repository resources:

• context_prev.md: deep technical onboarding context for code agents/maintainers
• examples/configs/*.yml: ready-to-use config profiles (CI, strict MR, Gemini, local ollama, minimal)
• docs/prev/binary-size-audit.md: current stripped binary baseline, dependency audit, and size-reduction options

Full config example:

# prev configuration
# Active provider (openai | anthropic | azure | gemini | ollama | custom).
provider: openai

# Provider-specific settings. Each block corresponds to a registered provider.
providers:
  openai:
    # api_key can also be set via OPENAI_API_KEY env var.
    api_key: ""
    model: "gpt-4o"
    # base_url: "https://api.openai.com/v1"  # override for proxies
    max_tokens: 1024
    timeout: 30s

  anthropic:
    # api_key can also be set via ANTHROPIC_API_KEY env var.
    api_key: ""
    model: "claude-sonnet-4-20250514"
    max_tokens: 1024
    timeout: 30s

  azure:
    # api_key can also be set via AZURE_OPENAI_API_KEY env var.
    api_key: ""
    base_url: ""  # e.g. https://<resource>.openai.azure.com
    model: ""     # deployment name
    api_version: "2024-02-01"
    max_tokens: 1024
    timeout: 30s

  gemini:
    # api_key can also be set via GEMINI_API_KEY env var.
    api_key: ""
    base_url: "https://generativelanguage.googleapis.com/v1beta/openai"
    model: "gemini-2.0-flash"
    max_tokens: 1024
    timeout: 30s

  # Example: self-hosted Ollama or any OpenAI-compatible endpoint.
  ollama:
    base_url: "http://localhost:11434/v1"
    model: "llama3"
    max_tokens: 1024
    timeout: 60s

# Retry configuration (applies to all providers).
retry:
  max_retries: 3
  initial_interval: 1s
  max_interval: 30s
  multiplier: 2.0

# Review policy and conventions.
review:
  # 1: critical only, 10: include nits and minor suggestions.
  nitpick: 5
  # Optional strictness default for MR review when CLI flag is not provided.
  # Allowed: strict | normal | lenient
  # strictness: "normal"
  # Number of AI review passes (re-review loop) for MR review.
  passes: 1
  # Maximum inline comments for MR review (0 = unlimited).
  max_comments: 0
  # Inline filtering mode: added | diff_context | file | nofilter
  filter_mode: "diff_context"
  # MR diff source strategy: auto | git | raw | api
  mr_diff_source: "auto"
  # Enable structured JSON findings output parsing (with markdown fallback).
  structured_output: false
  # Enable incremental review scope using baseline markers.
  incremental: false
  # Post inline comments only (skip summary notes and thread replies).
  inline_only: false
  # Persistent review memory and deterministic prechecks.
  memory: true
  memory_file: ".prev/review-memory.md"
  memory_max: 12
  native_impact: true
  native_impact_max_symbols: 12
  # Include AI fix prompt blocks in inline comments: off | auto | always.
  fix_prompt: "off"
  # Plain keyword used in MR thread commands. Config accepts an optional leading @, but comments should use the bare keyword.
  mention_handle: "prev"
  # Optional Serena/context defaults for MR review.
  # serena_mode: "auto"
  # context_lines: 10
  # max_tokens: 80000
  conventions:
    labels: ["issue", "suggestion", "remark"]
  # Optional custom instructions injected into review prompts.
  guidelines: |
    Prioritize correctness, security, and maintainability.
    Keep findings concrete and actionable.

# Display options.
debug: false
max_key_points: 3
max_characters_per_key_point: 100
explain: false

Global Flags

Flag	Description
--provider, -P	AI provider to use (openai, anthropic, azure, gemini, ollama, etc.)
--model, -m	Model to use for the AI provider
--stream, -s	Enable streaming output (default: true)
--strictness	Review strictness: strict, normal, lenient (default: normal)
--debug	Enable debug output
--help, -h	Help for any command

Per-command flags:

Flag	Commands	Description
--repo, -r	commit, branch	Path to git repository
--path, -p	commit, branch	Filter diff to specific file paths

Binary Size Notes

A stripped local build using go build -trimpath -ldflags='-s -w' -o /tmp/prev-audit ./ currently produces a binary around 7.4M on Linux amd64. There are no obvious dead production dependencies to remove without changing behavior. The next meaningful reductions come from feature tradeoffs such as optional clipboard support, minimal config/output profiles, or replacing Cobra/pflag with a smaller custom CLI layer.

See docs/prev/binary-size-audit.md for the current audit and recommended directions.

Development

# Run unit tests
make test-unit

# Run E2E tests
make test-e2e

# Build
go build -o prev .

# Docker build
docker build -t prev .

License

See LICENSE file.

Documentation

Overview

Licensed under Apache 2.0 License. */