Flaught

Your PR's designated skeptic.

The reviewer should not be the author. A model reviewing code it wrote itself tends to agree with its own choices. Same blind spots, same rationalizations. Flaught decouples the two: point it at any LLM, independent of whatever wrote the code, and get a genuinely adversarial second opinion instead of an echo.

bash

# install
npm install -g @flaught/core
# scaffold config + prompts, then review vs main
flaught init
flaught review --base main

Runs in GitHub Actions, GitLab CI, or any shell. No API key needed for the deterministic pass (flaught review --no-llm); set GROQ_API_KEY for the full LLM pass. Provider config →

01Five stages, one skeptical pass

Five stages produce findings tagged by source (deterministic tool vs. LLM), which pool together and pass through two gates (a noise budget that caps volume, then a severity check) to decide exit 0 (clean) or 1 (gated, blocks the PR). A fault anywhere in the pipeline (bad config, missing API key, LLM outage) skips both gates and exits 2 instead: a tool problem, not a verdict on the code.

Context assembly

Diff, changed files, and a one-hop dependency neighborhood so findings carry blast radius, not just a line number.
Deterministic tools deterministic

Semgrep, a linter, and a vuln scanner. Every finding is tagged source_type: "deterministic" so you know what came from a tool.
LLM adversarial pass llm

Structured skeptical review of security, architecture, scope-creep, and test quality, tagged source_type: "llm", distinct from the tool findings.
Test inversion

Runs the test suite on pre-change code and flags tests that pass on both sides: cheap mutation testing that catches vacuous tests.
Scope-creep detection

Heuristic plus LLM; flags hunks unrelated to the PR's stated intent.

Output: a Markdown PR comment plus a versioned JSON artifact you can keep (one snapshot per run, for trend tracking).

02See it work

Animated terminal recording: flaught review runs against main on a deliberately flawed pull request and reports findings, then a peek at the JSON artifact. — A deliberately flawed PR reviewed against `main`: a string-interpolated SQL query (critical, LLM-asserted), a scope-creep hunk in an unrelated file, and a vacuous test that passes on both sides of the change, caught by test inversion.

03Review with a different model than the one that wrote the code

Self-review is the weak spot in AI-assisted development: the model that wrote your PR is primed to defend it. Point .advreview.yml at a different provider and Flaught breaks that correlation, zero code change.

Groq default

groq/compound-mini

OpenAI

gpt-4o

Gemini

gemini-1.5-pro

Ollama

local + cloud

A sane pairing: code with Claude, review with Groq or GPT-4o. Code with anything, review with a different anything. Any OpenAI-compatible endpoint works via base_url; Anthropic has its own native adapter too, best paired as the coder here, not the reviewer.

04Extend it without forking it

The defaults are a starting point. Three surfaces let you reshape the review without touching Flaught's source: the prompt, the config, and the library API.

Prompt

.flaught-prompt/

Override any section (categories, severities, posture) or append team-specific rules on top of the defaults.

Config

.advreview.yml

Swap the linter, vuln scanner, or Semgrep ruleset; tune the noise budget and severity gate per project.

Library

runReview()

Call Flaught from Node for a custom dashboard or a CI system that isn't GitHub Actions.

Full guides: prompt templates, configuration, programmatic API.

05How it compares

Against the SaaS market leader and the closest open-source structural analog. The difference is what the review does: a tool that says “here are issues” is one thing; one that also asks “are these issues real?” (and whether your tests even test the change, and whether a hunk belongs in this PR) is another.

	Flaught	CodeRabbit	Open Code ReviewAlibaba · OCR
Open source, self-hostable	✓	✗	✓
Bring-your-own LLM (incl. local)	✓	✗	✓
Deterministic tools + LLM, provenance-tagged	✓	~¹	~²
Adversarial refute / skeptic pass (cross-model)	✓	✗	~³
Test inversion (vacuous-test detection)	✓	✗	✗
Scope-creep detection	✓	✗	✗

✓ yes ~ partial / varies ✗ not a named feature

¹ CodeRabbit runs 40+ linters + SAST + LLM, but doesn't tag each finding's provenance.
² OCR is a deterministic-engineering + agent hybrid; its deterministic layer is file selection, line positioning, and rule routing, not external tools (Semgrep/linter/vuln) with per-finding source tags.
³ OCR has a reflection module (same-pipeline self-check), similar spirit, but not a different model refuting findings with anti-correlation.

06The honest caveat

⚠ The JSON artifact is evidence that scrutiny occurred, not evidence that findings are correct. LLM-asserted findings may include hallucinations. Deterministic-tool findings have their own false-positive rates. Treat this as a prompt for human review, not audit-truth.

Named for Monsignor Flaught, the devil's advocate in A Canticle for Leibowitz.