nitpicker

Multi-reviewer code review using LLMs. Spawns parallel agents with different models/prompts, aggregates their feedback into a final verdict. Supports two modes — parallel aggregation and actor-critic debate — across two task types: code review and free-form questions.

Free Web version is available for open source projects.

Each reviewer is an agentic loop that can call tools (read files, grep, glob, git commands) to explore the repo before writing its review. Discovery roles build a quick initial map and default to one early, disjoint subagent wave for multi-surface targets; validation roles delegate only bounded verification of submitted claims. Tool outputs include lightweight headers and clearer truncation/no-match messages so agents can reason about partial evidence more reliably. A separate aggregator model deduplicates and synthesizes the individual reviews into a final verdict.

Diff and PR reviews capture a frozen orientation snapshot before reviewers fan out: full HEAD, resolved base and merge-base revisions, the exact committed comparison, working-tree status, and committed/uncommitted file maps. Every lane and later debate round receives the same snapshot; agents still inspect the actual hunks and code with tools.

Requirements

• Rust toolchain
• A git repository to review
• At least one configured LLM (API key or Gemini OAuth)

Installation

cargo install nitpicker

Quick start

export ANTHROPIC_API_KEY="your-api-key-here"

Review

nitpicker
nitpicker --repo /path/to/repo
nitpicker --repo /path/to/repo --prompt "focus on src/api/"
nitpicker --fallback  # try the next configured reviewer if a model fails
nitpicker --analyze src/components/
nitpicker --analyze  # entire repo

Parallel Mode

nitpicker --no-debate
nitpicker --no-debate --analyze src/
nitpicker --no-debate --max-turns 40

PR review

nitpicker pr
nitpicker pr https://github.com/owner/repo/pull/42
nitpicker pr --no-comment
nitpicker pr https://github.com/owner/repo/pull/42 --no-comment
# force a fresh temp clone even when the URL points to your current repo
nitpicker pr https://github.com/owner/repo/pull/42 --clone
# machine-readable output for embedding (one JSON object on stdout)
nitpicker pr https://github.com/owner/repo/pull/42 --no-comment --json

Ask

nitpicker ask "should we use eyre or thiserror for error handling?"
nitpicker ask --no-debate "is this authentication flow secure?"
nitpicker ask --rounds 3 "should we split this module?"
nitpicker ask --max-turns 40 "should we split this module?"

Configuration

Configuration is loaded from (first match wins):

1. --config <path> (explicit flag)
2. nitpicker.toml in repo root
3. ~/.nitpicker/config.toml (global config)

# create a config in current directory
nitpicker init

# prefer OpenRouter experimental free models when OPENROUTER_API_KEY is set
nitpicker init --free

# create a global config at ~/.nitpicker/config.toml
nitpicker init --global

Example nitpicker.toml:

[defaults]
debate = true          # optional, default: true
fallback = true        # optional, default: false; use reviewer order as a failover ring
max_turns = 100        # optional, default: 100
log_trajectories = false # optional, default: false
# presets = ["correctness", "security"]  # optional; default also includes performance, simplicity

[aggregator]
model = "claude-sonnet-5"
provider = "anthropic"
max_tokens = 16384       # optional, default: 16384

[[reviewer]]
name = "claude"          # used in output headers and logs
model = "claude-sonnet-5"
provider = "anthropic"
# max_tokens = 32768     # optional, default: unset (the provider's own per-model limit)

[[reviewer]]
name = "gpt"
model = "gpt-5.6-sol"
provider = "openai_compatible"
base_url = "https://api.openai.com/v1"
api_key_env = "OPENAI_API_KEY"

# optional: define a custom review angle (or override a built-in by using its name)
[presets.api-security]
prompt = """
Review trust boundaries, authentication, authorization, input handling, and secret exposure.
Require a concrete attacker-controlled path and plausible impact for every finding.
"""

Tip: Use providers that were not used for the initial building of your codebase to enforce diversity of thought.

Review presets

A preset is one named review angle — a rubric that tells a reviewer what to investigate; the execution mode (parallel, debate, alloy) decides how. Every review run resolves an ordered preset list: --preset on the command line beats [defaults].presets, which beats the four-angle built-in default (correctness, security, performance, simplicity). Domain-specific built-ins ai-systems, ml-rigor, and tone are opt-in. general is a standalone broad review for unusual targets or user-defined concerns and cannot be combined with another preset. A [presets.<name>] table with a built-in's name replaces it.

nitpicker --preset security                      # one focused angle
nitpicker --preset security,ml-rigor             # commas split
nitpicker --preset ai-systems                    # agent/prompt/tool/context audit
nitpicker --preset general --prompt "review the plugin contract"
nitpicker pr --preset api-security               # project-defined preset

Fan-out: parallel mode runs every configured reviewer against every selected preset (reviewers × presets jobs); debate mode runs one independent Reviewer/Validator debate per preset, lanes concurrent, with a single meta-review across all lanes. Spend and wall-clock scale with the selection: the untouched default runs four lanes (or 4× the parallel jobs) where 0.8.x ran one combined review. Names are case-sensitive; unknown or empty names, mixing general with another preset, or selecting more than 16 presets fails before any model call. Every final finding includes a Lens field naming the angle that produced it (or all contributing angles when synthesis merges duplicates). ask, init, and reflect take no presets — the flag is rejected there.

Built-in rubrics and review/debate protocols live as auditable Markdown under prompts/ and are compiled into the binary. Generic loop contracts such as compaction and final-turn handling live under crates/nitpicker-agent/prompts/ and are compiled into the library that interprets them. Rust owns selection and interpolation, not the prompt prose.

Unknown config keys are rejected. For example, use max_tokens for output length; token_limit is not a supported field.

max_tokens caps a single response, and on a reasoning model it is a budget for reasoning plus the answer — set too low, the model spends it all thinking and returns empty content, which is indistinguishable from a model that said nothing. Reviewers therefore default to no cap (the provider applies its own per-model limit); set one only to bound spend. The aggregator writes one bounded synthesis and defaults to 16384. Two exceptions: Anthropic's API requires the field, so an unset reviewer cap becomes 8192 there — raise it explicitly if your model reasons past that; and auth = "codex" ignores the setting entirely, since that endpoint rejects max_output_tokens.

Debate mode is enabled by default for nitpicker, nitpicker ask, and nitpicker pr. Pass --no-debate to use parallel aggregation for a single run. Use [defaults].max_turns or --max-turns to control the per-agent tool-use loop limit.

Fallback mode is opt-in with [defaults].fallback = true or --fallback and requires at least two reviewers. Each logical reviewer keeps its normal primary, then tries subsequent [[reviewer]] entries in declaration order, wrapping at the end. Failover retries only the failed completion with the existing conversation history; it does not restart the agent. The successful route remains active for that agent, and a quota-limited route is skipped by the other jobs for the rest of the run. The aggregator tries its configured model first, then the reviewer list. A successful fallback is logged but does not make the verdict degraded. With Alloy, each completion still chooses its first healthy reviewer randomly; a failed choice then follows declaration order.

Set [defaults].log_trajectories = true to save per-agent JSONL traces and a final aggregation.json under ~/.nitpicker/sessions/session-<timestamp>-<pid>/.

Provider types

provider	Auth	Notes
anthropic	ANTHROPIC_API_KEY env var (or api_key_env), or auth = "azure-ad"	base_url optional
gemini	GEMINI_API_KEY/GOOGLE_AI_API_KEY env var (or api_key_env), or auth = "agy-keyring"	base_url optional (e.g. a local Gemini-compatible server); agy-keyring reuses the Antigravity CLI OAuth token from the system keyring — research only, see warning
openai	OPENAI_API_KEY env var (or api_key_env), auth = "azure-ad", or auth = "codex"	codex reuses your ChatGPT subscription via the Codex CLI token — research only, see warning
openrouter	OPENROUTER_API_KEY env var (or api_key_env)	explicit model names are recommended; model = "free" is experimental

anthropic_compatible and openai_compatible are accepted as aliases for backward compatibility.

auth = "azure-ad" authenticates with a refreshing Azure AD (Entra ID) token instead of a static key — for OpenAI and Anthropic models hosted on Azure AI Foundry. Requires a build with the azure feature, see below.

auth = "codex" authenticates with your ChatGPT Plus/Pro (Codex) subscription instead of a paid API key, reusing the token the Codex CLI stores on disk, see below.

OpenRouter models

openrouter supports both explicit pinned models and an experimental free auto-selection mode.

Pinned models are the supported default and the recommended setup:

# recommended: explicit model
[[reviewer]]
name = "qwen"
model = "qwen/qwen3-30b-a3b"
provider = "openrouter"

Experimental best-effort free auto-selection is also available:

# experimental: auto-select a currently available free model
# omit `model` or set model = "free"
[[reviewer]]
provider = "openrouter"

# explicit experimental form
[[reviewer]]
model = "free"
provider = "openrouter"

When model is omitted or set to "free", nitpicker tries to pick a currently working free model at startup.

This mode is convenient, but it is not production-stable and may fail due to upstream availability, routing differences, or timeouts.

If you want predictable behavior, pin explicit model names instead of relying on free auto-selection.

export OPENROUTER_API_KEY="your-key"

A free OpenRouter account is sufficient for the experimental free mode — no credit card required, just rate limits.

Antigravity Keyring (research only)

AG2 is Google's current agentic IDE, succeeding both the older Gemini CLI OAuth path and the earlier AG1 preview. The gemini-3.x family ships only through AG2's CloudCode backend, so auth = "agy-keyring" exists purely as a research path to compare those models against the rest of the reviewer pool, with full awareness of the ToS posture above.

The proxy reads agy's OAuth token from the system keyring (service=gemini, account=antigravity) via the keyring crate (Secret Service on Linux, Keychain on macOS, Credential Manager on Windows), relies on agy to refresh it, and routes chat through CloudCode's v1internal:streamGenerateContent SSE endpoint. Run agy and complete its login first. NITPICKER_ANTIGRAVITY_PLATFORM can override the auto-detected platform enum if needed.

This path requires a build with the antigravity feature (off by default, since it pulls in the local proxy stack — axum — and the keyring crate with its native backends):

cargo build --release --features antigravity
# or: cargo install --features antigravity ...

Without the feature, auth = "agy-keyring" is rejected at config validation with a build hint, and nitpicker init won't offer the keyring reviewer.

Tested AG2 models (current author config): gemini-3.1-pro-low, gemini-3.5-flash-low. Other IDs returned by fetchAvailableModels (e.g. gemini-3-flash-agent) likely work but have not been exercised.

[aggregator]
model = "gemini-3.5-flash-low"
provider = "gemini"
auth = "agy-keyring"

[[reviewer]]
name = "gemini"
model = "gemini-3.1-pro-low"
provider = "gemini"
auth = "agy-keyring"

ChatGPT/Codex subscription (research only)

auth = "codex" (on an openai reviewer/aggregator) reuses the OAuth token the Codex CLI stores in ~/.codex/auth.json. Log in once with codex login (choosing your ChatGPT account, not an API key); nitpicker reads the token read-only and refreshes the short-lived access token in-memory via the refresh token — it never writes back to auth.json. Set CODEX_HOME to override the token directory.

Under the hood this talks to the Codex subscription endpoint (chatgpt.com/backend-api/codex/responses), which speaks the OpenAI Responses API with subscription-specific quirks (a required top-level system prompt, mandatory streaming, store: false, no max_output_tokens, and encrypted reasoning items round-tripped across turns since nothing is server-side persisted); nitpicker handles all of that transparently. No API-key env var is needed.

Models are your subscription's Codex models (e.g. gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna):

[aggregator]
model = "gpt-5.4"
provider = "openai"
auth = "codex"

[[reviewer]]
name = "codex"
model = "gpt-5.4"
provider = "openai"
auth = "codex"

Azure AD (Azure AI Foundry)

Call OpenAI and Anthropic models hosted on Azure AI Foundry using a short-lived Azure AD (Entra ID) token instead of a static key. nitpicker acquires the token via the Azure SDK and transparently refreshes it (rebuilding the client before the token expires), so long reviews and debates don't die mid-run — the equivalent of the Python SDK's azure_ad_token_provider.

This path requires a build with the azure feature (off by default, since it pulls in the Azure SDK and needs Rust 1.88+):

cargo build --release --features azure
# or: cargo install --features azure ...

Set auth = "azure-ad" on an openai or anthropic reviewer/aggregator and point base_url at your Foundry endpoint:

[[reviewer]]
name = "gpt"
provider = "openai"                                                  # OpenAI models → /openai/v1
base_url = "https://<resource>.services.ai.azure.com/openai/v1"
model = "gpt-4o"                                                     # your Foundry deployment / model
auth = "azure-ad"

[[reviewer]]
name = "claude"
provider = "anthropic"                                               # Anthropic models → /anthropic
base_url = "https://<resource>.services.ai.azure.com/anthropic"
model = "claude-sonnet-4-5"
auth = "azure-ad"
azure_credentials = "dev"                                            # optional, see below

Optional per-reviewer/aggregator fields:

• azure_scope — AAD token scope. Defaults to https://cognitiveservices.azure.com/.default.

• azure_credentials — selects the credential chain, mirroring the Azure SDK's AZURE_TOKEN_CREDENTIALS:

  • "dev" — developer tools only (az login, Azure Developer CLI), excluding managed identity. Use on a VM where you want az login instead of a system-assigned managed identity.
  • "prod" — env service principal (AZURE_TENANT_ID/AZURE_CLIENT_ID/AZURE_CLIENT_SECRET), then managed identity.
  • unset / "auto" — env service principal → managed identity → developer tools, in that order.

  If unset, the AZURE_TOKEN_CREDENTIALS env var is honored as a fallback.

CLI reference

nitpicker [OPTIONS]
nitpicker ask [--no-debate] [--rounds N] [--max-turns N] [OPTIONS] <topic>
nitpicker pr [URL] [--no-comment] [--no-debate] [--rounds N] [--max-turns N] [OPTIONS]
nitpicker reflect [--sessions-dir DIR] [--n N] [OPTIONS]
nitpicker init [--global] [--free] [--repo <DIR>]

Review (default)

--repo <PATH>          git repository to review [default: .]
--config <PATH>        config file [default: <repo>/nitpicker.toml, then ~/.nitpicker/config.toml]
--prompt <TEXT>        review instructions (optional, has a sensible default)
--preset <NAME>        review angle(s) to run; repeatable, comma-separated, replaces the configured default list
--context-file <PATH>  inject a file's contents into the prompt; repeatable
--analyze [PATH]       analyze existing code instead of reviewing changes
--no-debate            use parallel aggregation instead of actor-critic debate
--fallback             try subsequent configured reviewers when a model fails
--rounds <N>           maximum debate rounds [default: 5]
--max-turns <N>        maximum tool-use turns per agent or debate turn [default: 100 via config]
-v, --verbose          show info-level logs (hidden by default)

--context-file

The agents' own tools are sandboxed to the repository, so they cannot open a design doc that lives outside it. --context-file reads such a file directly into the prompt:

nitpicker --context-file ~/notes/migration-plan.md --context-file /tmp/rfc.md

Available on the default review mode, ask, and pr; the flag may be given before or after the subcommand. Files are injected verbatim, in the order given, after the task and any --prompt text. Total injected size — contents plus each block's wrapper — is capped at 256 KiB; a missing, non-regular (FIFO, device node), binary, or non-UTF-8 file is an error, raised before any model is called. Because the contents are placed in the task prompt rather than the system prompt, spawned subagents do not inherit them.

PR subcommand

nitpicker pr [URL] [--no-comment] [--no-debate] [--rounds N] [--max-turns N] [--prompt TEXT] [--preset NAME] [--context-file PATH] [--repo .] [--config PATH] [--json] [-v]

Reviews a GitHub PR using its title, description, and diff. Requires the gh CLI (gh auth login to authenticate).

• Without URL: reviews the current branch's open PR (must be run inside the repo)
• With URL (https://github.com/owner/repo/pull/N): clones the repo into a temp dir, checks out the PR branch, reviews it, then cleans up
• By default, posts the review as a PR comment. Pass --no-comment to skip posting.
• --no-debate, --rounds, and --max-turns work the same as in the default review mode
• --json emits a single machine-readable JSON object on stdout (status, PR metadata, models, resolved presets, report_markdown, usage, …) instead of the human report, with all logs/progress on stderr — handy for calling nitpicker as a subprocess. Exits non-zero on failure, with a status: "error" object on stdout; a degraded run (some job or debate turn failed — degraded: true) exits 3 after emitting the envelope. The coverage key reports each preset's attempted/succeeded job or lane counts; entries with succeeded > 0 identify the angles that produced evidence, while the counts distinguish partial coverage such as 1 of N jobs. The usage block reports aggregate input_tokens/output_tokens/total_tokens, cached_input_tokens/cache_creation_input_tokens, and subagents_spawned for the run (best-effort: successful completions only). The cache fields are a breakdown of input_tokens, not an extra charge — a healthy multi-turn run shows most of its input served from cache. Cost formulas must discount them: input_tokens now counts cache reads on every provider (before 0.8.3 the Anthropic path omitted them), so a naive input_tokens * input_price over-states a cache-heavy run.

Ask subcommand

nitpicker ask [--no-debate] [--rounds N] [--max-turns N] [--context-file PATH] [--repo .] [--config PATH] [-v] <topic>

Runs agents on a free-form question instead of a code diff. By default, two agents take turns as Actor/Critic before a meta-reviewer concludes. Pass --no-debate to switch to the parallel reviewer plus aggregator flow.

Reflect subcommand

nitpicker reflect [--sessions-dir DIR] [--n N] [--repo .] [--config PATH]

Analyzes the most recent recorded sessions (20 by default) and reports recurring execution friction. Reflection first computes deterministic trace metrics, then analyzes each session with the first reviewer and uses the second reviewer (the critic) for calibrated cross-session synthesis, falling back to the first reviewer when only one is configured. The ordinary aggregator is intentionally not used because reflection reduction is a tool-using investigation, not report formatting. Pattern claims include exact session support, representative trace evidence, confidence, and a small measurable experiment. New session records also preserve staged reviewer/debate verdicts so reflection can compare execution with the outcome. Set [defaults].log_trajectories = true beforehand. Sessions recorded before staged verdicts became part of the aggregation schema are rejected, and reflect does not write extra trace copies.

Debate mode (default)

Two LLM agents take turns exploring the codebase with file/git tools and submitting verdicts. The Critic can signal agreement (agree=true) to end early. A meta-reviewer synthesizes the dialogue.

• reviewer[0] in config → Actor (review: Reviewer)
• reviewer[1] in config → Critic (review: Validator)
• aggregator → Meta-reviewer

Interactive text runs show a compact cast/progress view while debating, then print the final synthesized result. While a review or debate is active, the terminal tab/window title also shows nitpicker's fixed-outline target activity glyph plus the repository name (for example, ⊙ reviewer); it is TTY-only and cleared when the run ends, so redirected and JSON output stay unchanged. In a terminal, --verbose also shows intermediate debate output and the saved transcript path; redirected stdout stays final-report-only.

With --verbose, the transcript is saved to {tempdir}/debate-{timestamp}.md (ask) or review-debate-{timestamp}-{preset-slugs}.md (review; one section per preset lane). Non-verbose runs skip the write.

Exit codes (default review, ask, and pr)

code	meaning
0	clean verdict
1	hard failure — no verdict (bad config, missing key, every reviewer/turn failed)
2	CLI usage error (clap's exit code for bad arguments)
3	degraded verdict — report printed, but a reviewer or debate turn failed

Non-interactive, non-verbose stdout carries exactly the final report, so the binary can be driven as a subprocess: read stdout for the verdict, branch on the exit code. pr follows the same codes; in --json mode the envelope (status: ok|error, with degraded: true on an exit-3 run) is emitted and flushed before the exit.

Using the agent as a library

The agentic core — the loop, the file/git tools, and the provider-agnostic LLM clients — lives in a separate crate, nitpicker-agent, with none of the CLI/review/PR machinery. Use it to drive your own file-reading agent with subagents:

use nitpicker_agent::prelude::*;
use std::path::Path;

let client = client_from_env(LLMProvider::Anthropic { base_url: None, api_key_env: None })?;
let result = AgentBuilder::new("explorer", "claude-sonnet-5", "You explore codebases.", client)
    .subagent_system_prompt("You are a focused file-reading worker. Report findings concisely.")
    .run("Map the module layout of this repo.", &file_agent_tools(), Path::new("."))
    .await?;
println!("{}", result.text);

file_agent_tools() is the read-only file/git toolset plus spawn_subagent. You control the top-level prompt, the subagent prompt, the toolset, and the client; config-file-driven client construction is available via the config/provider modules. See crates/nitpicker-agent/examples/file_agent.rs.

Changelog

See CHANGELOG.md for release history.