Argus Review

Argus is an open-source code review tool built by Redesign Health.

It reviews pull requests by decomposing each change into the flows it participates in, dispatching repository-aware specialist agents in parallel against those flows, validating findings before surfacing them, and persisting every round so later rounds can check what was resolved, regressed, or still open instead of re-reviewing from scratch. The architecture borrows from the same ideas now showing up in the broader agentic-harness conversation: stochastic and deterministic hooks composed deliberately, just-in-time context retrieval, structured outputs, and explicit coverage accounting.

For the design principles and production results behind Argus, see the agentic code review harness write-up on our site.

The prompts Argus ships with are opinionated — they encode one team's production coding standards, not a neutral default. Expect to adapt them to your own codebase and conventions (see Adapting Argus to your team) rather than running them unmodified.

Cost Responsibility

Argus makes API calls to third-party services (Anthropic, OpenAI, and optionally GitHub) using your API keys. You are solely responsible for all token consumption and associated costs incurred by running Argus. Costs vary significantly depending on PR size, repository complexity, and configuration — in production use, a single review may cost anywhere from approximately $2 to $42 or more in API fees. Redesign Health does not monitor, limit, or reimburse these costs. By using Argus, you acknowledge that you bear full financial responsibility for any and all API charges incurred through your credentials.

No Warranty; Limitation of Liability

Argus is provided "AS IS" and "AS AVAILABLE," without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, accuracy, or non-infringement. Redesign Health Inc. makes no representation or warranty that Argus will be error-free, uninterrupted, secure, or free of bugs or defects.

To the fullest extent permitted by applicable law, in no event shall Redesign Health Inc., its affiliates, officers, directors, employees, or agents be liable for any indirect, incidental, special, consequential, or punitive damages, or any loss of profits, data, use, or goodwill, arising out of or in connection with your use of or inability to use Argus, regardless of the theory of liability (contract, tort, strict liability, or otherwise), even if advised of the possibility of such damages.

Your use of Argus is at your sole risk. This tool performs automated code review and may produce inaccurate, incomplete, or misleading findings. It is not a substitute for human code review, security auditing, or professional judgment. You are solely responsible for any decisions made or actions taken based on Argus's output.

For the full license terms, see the Apache License 2.0 included in the repository.

• Cost Responsibility
• No Warranty; Limitation of Liability
• How it works
• Getting started
• Using Argus
• Managing secrets
• Connecting a database
• Adapting Argus to your team
• Measured results
• Docs

How it works

A pull request goes in. Argus's planner groups the changed files into "system groups" (roughly: one per feature area touched), assigns specialist co-reviewers where the file patterns warrant it (security, SQL, infra, ...), and flags anything that needs a cross-file pass. Each group and specialist then runs as its own Claude Agent SDK session — with Read/Glob/Grep tools against a SHA-pinned local git worktree, not a shared context window — in parallel. A coverage check verifies every changed file got reviewed by someone (gaps get a targeted follow-up pass), and a final writer step consolidates every finding into one review with a verdict and severity tags.

fetch_diff → plan → [system reviewers ‖ specialists ‖ cross-cutting] (parallel)
           → check_coverage → (fill_gaps if needed) → write_review → persist

Full design rationale is in docs/ARCHITECTURE.md.

Getting started

Prerequisites:

• git
• Python 3.12+
• The claude CLI (Node.js) — the Claude Agent SDK spawns it as a subprocess for every reviewer session
• Three API keys — see Managing secrets

Run:

export ANTHROPIC_API_KEY=...
export GITHUB_TOKEN_RO=...
export OPENAI_API_KEY=...

uvx --from argus-code-review argus review owner/repo --pr 123

Prefer a .env file over exporting in your shell? Copy .env.example to .env, fill in the three keys above, and Argus picks it up automatically (see Managing secrets).

That's it for a first run. With no other configuration, Argus stores round history and pipeline checkpoints in a local SQLite database at ~/.local/share/argus/history.db — nothing else to set up.

uvx --from argus-code-review argus ... downloads the argus-code-review package into a throwaway environment and runs its argus console script — the --from is required because the installable package name (argus-code-review) and the console script it provides (argus) differ, and uvx only infers the script name from the package name when they match. If you'll be running Argus repeatedly, install it once instead of re-resolving it every invocation:

uv tool install argus-code-review
argus review owner/repo --pr 123

Using Argus

# Review an open PR
argus review owner/repo --pr 123

# Review a specific commit against a base ref (no open PR required)
argus review owner/repo --sha abc123 --base-ref main

# Write markdown + JSON output
argus review owner/repo --pr 123 -o review.md

# Publish the review as a PR comment (upserts on re-run) and set a commit status
argus review owner/repo --pr 123 --post --commit-status

# Dismiss a finding before the next round
argus review owner/repo --pr 123 --dismiss "B2 -- pre-existing, not from this PR"

# Override the reviewer models (or set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL instead).
# --frontier-model controls both the planner/coverage tier AND the cross-cutting
# reviewer -- claude-fable-5 here is already the planner/coverage default, but
# it also moves cross-cutting OFF its cheaper claude-opus-5 default onto fable-5.
# Cost note: --specialist-model here also moves the highest-volume path (system
# reviewer, specialists, writer, lite-review) onto the pricier Opus tier -- both
# flags in this example trade cost for reasoning headroom, don't use them together
# as a low-cost default. Any non-empty --specialist-model value also withholds
# the 1M-context beta from that same highest-volume path, since the beta is
# only verified against the unoverridden default (autocompact may thrash on
# long reviews under this override -- see argus/runners.py for the tradeoff).
argus review owner/repo --pr 123 --specialist-model claude-opus-5 --frontier-model claude-fable-5

# Clear an already-set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL for just this
# run by passing an empty string -- useful when the env var is set globally
# (e.g. in your shell profile or CI) but you want one invocation to use the
# hardcoded default instead.
argus review owner/repo --pr 123 --specialist-model ""

# Inspect and export the packaged prompts
argus prompts list
argus prompts export ./my-prompts

What you get back:

• A formatted review comment (Markdown) with a verdict (APPROVE / BLOCKING), a risk level, and findings tagged B1, B2, ... (blocking) or S1, S2, ... (suggestion).
• -o review.md writes that Markdown to a file, plus a sibling review.json with the full structured response.
• Re-running against the same PR (same owner/repo + --pr) picks up the prior round from storage automatically: round 2+ verifies whether previously reported findings were resolved, regressed, or are still open, scoped to just the commits since the last review — not a full re-review.

Managing secrets

Argus reads configuration from environment variables, or from a .env file in the working directory (add it to .gitignore — never commit it). .env.example lists every variable Argus reads, required and optional, with comments — copy it to .env and fill in what you need.

Three keys are required for a first run:

GITHUB_TOKEN_RO needs read-only access: Contents: read + Pull requests: read on a fine-grained PAT, or repo read on a classic one.

How you get these values into the environment is up to your setup — export them in your shell, load them from a secrets manager in CI, or use a .env file locally. Argus itself has no opinion and no cloud dependency for secrets; it just reads os.environ.

The full configuration surface (storage URLs, prompt overrides, tracing, etc.) is documented where each feature is introduced below, and summarized in docs/STORAGE.md for storage-specific variables.

Tracing reviewer sessions (optional)

Every reviewer subagent session is wrapped in a LangSmith @traceable span. Tracing is off by default and turns on automatically the moment LANGSMITH_API_KEY is set — no flag, no code change. LANGCHAIN_TRACING_V2=true also activates it (the module docstring in argus/runners.py documents the same behavior) — useful if you're already in a LangChain-heavy environment where that variable is set ambiently:

export LANGSMITH_API_KEY=...
export LANGSMITH_PROJECT=argus   # optional; defaults to LangSmith's own default project

Leaving LANGSMITH_API_KEY unset is the normal, fully-supported case — Argus runs identically either way, just without traces. There's nothing else to configure: no separate on/off switch, and nothing in Argus itself fails or degrades based on whether tracing is enabled.

If you're proxying ANTHROPIC_API_KEY/OPENAI_API_KEY through your own credential-vending layer (see Managing secrets), LangSmith isn't part of that Anthropic/OpenAI proxying by default — set LANGSMITH_API_KEY (and, if you're also proxying the LangSmith endpoint itself, LANGSMITH_ENDPOINT) the same way you'd set any other env var for this process.

Connecting a database

Argus needs somewhere to keep round history (so re-reviewing a PR knows what was already found) and pipeline checkpoints (so a crash mid-run can resume). Three modes are supported, resolved automatically from which environment variables are set — there's no --backend flag.

Recipe: local, zero-config (default)

Set nothing. Both round history and checkpoints land in a local SQLite file at ~/.local/share/argus/history.db (created automatically on first run). Good for solo use and for trying Argus out — this is what the quickstart above uses.

Recipe: shared Postgres, e.g. Supabase

Use this to share round history across a team or CI. Any Postgres works; Supabase is a common zero-ops choice.

1. Create a Supabase project (or use an existing one), then grab its pooler connection string (Project Settings → Database → Connection string → "Transaction" pooling mode, port 6543) — the pooler is what you want for a short-lived CLI process like Argus, not the direct connection on port 5432.
2. Apply the schema, in order:

   psql "$SUPABASE_DB_URL" -f schema/008_add_code_reviews.sql
   psql "$SUPABASE_DB_URL" -f schema/009_add_reviewer_version.sql
   psql "$SUPABASE_DB_URL" -f schema/010_add_review_patterns.sql
   psql "$SUPABASE_DB_URL" -f schema/011_add_review_progress_columns.sql
   psql "$SUPABASE_DB_URL" -f schema/015_create_agent_runs.sql
   

   The gaps in that numbering (no 001-007 or 012-014) aren't missing files — these five are self-contained extracts from a longer internal migration series, and 008 creates the schema from scratch, so nothing earlier is required. Apply exactly the five files above, in order, and you have the complete schema.
3. Set the connection string and run Argus as usual — that one variable switches both round history and checkpointing to Postgres:

   export SUPABASE_DB_URL="postgresql://postgres.xxxx:pass@aws-0-region.pooler.supabase.com:6543/postgres"
   argus review owner/repo --pr 123
   

   (SUPABASE_DB_URL is accepted as an alias for ARGUS_DB_URL — if you're pointing at a non-Supabase Postgres instance, use ARGUS_DB_URL instead; they're equivalent, and if both are set ARGUS_DB_URL wins.)

Recipe: HTTP storage backend

For environments that can reach HTTPS but not a raw Postgres port (e.g. a sandboxed agent runtime with egress restricted to port 443), or for teams that don't want to expose their database directly to wherever Argus runs: route storage through your own HTTP API with ARGUS_STORAGE_READ_URL / ARGUS_STORAGE_WRITE_URL (optionally ARGUS_STORAGE_AUTH for a bearer token your service checks). The full two-endpoint contract — everything you need to implement a compatible backend service — and mode-resolution details are in docs/STORAGE.md.

Adapting Argus to your team

Argus ships with a full set of production review prompts — 20 files, one per reviewer role — rather than a thin default. Adapting them to your team is the main lever for making Argus yours.

Editing what Argus looks for

argus prompts export ~/.config/argus/prompts   # copy the packaged prompts out
$EDITOR ~/.config/argus/prompts/pr-review-subagent.md
argus review owner/repo --pr 123               # no env var needed — it's already found

That's the standing-customization path: an explicit ARGUS_PROMPTS_DIR (if you need one), then a repo-local ./.argus/prompts/, then ~/.config/argus/prompts/, are all checked automatically, highest priority first — see docs/CUSTOMIZING_PROMPTS.md for the full resolution chain and a --no-prompt-overrides escape hatch.

Any file you don't touch keeps loading from the package, so adapt incrementally. Each prompt mixes two kinds of content — pull them apart as you edit:

• Transferable methodology — silent-fallback detection, stub completeness, loud-failure-over-silent-failure philosophy, coverage mechanics, the severity rubric. This is why the reviewer is effective; keep it unless you have a specific reason not to.
• House conventions from wherever the prompts were tuned — references to specific shared libraries, a particular lint-config or .cursorrules-like compliance check, a specific secrets-manager's path conventions, a specific orchestration/LLM framework stack. This is the part that should become your team's equivalent — your shared libraries, your lint rules, your infra stack, the failure modes your codebase actually has.

Start with pr-review-subagent.md (the core reviewer prompt) since it has the most coding-standards content per line. Leave the pipeline-mechanics prompts (planner, coverage-check, writer) alone until you have evidence they need to change — they shape how the pipeline runs, not what it looks for. To evaluate any change, re-run the same PR before and after and diff the findings. Full prompt-by-prompt anatomy: docs/CUSTOMIZING_PROMPTS.md.

Adding a new specialist reviewer

Nine specialists ship today (security, SQL, infra, orchestration, frontend, Slackbot integrations, deployment, LLM patterns, observability), assigned by the planner based on which files a group touches. Editing an existing specialist's prompt is just editing a file, as above.

Adding an entirely new specialist currently requires a small code change, not just a new prompt file: the specialist name has to be added to the SpecialistName type and the prompt-name mapping in argus/runners.py (a couple of lines each), so the planner's structured output can validate against it. This is a known rough edge — making the specialist roster fully prompt-file-driven (drop a file in a specialists/ folder, get a new specialist, no code change) is tracked as a planned improvement. Check the repo's issue tracker for its status before assuming you need to fork.

Measured results

Argus's pipeline has gone through three iterations (v1/v2/v3 in its internal history; v3 is what ships here). Measured across 22 real PRs:

* This early tracked-cost figure covered only the planner + writer LLM calls, not the Agent SDK subagent sessions that do most of the work — true cost per review in that same 22-PR set is closer to $2-4, and larger production samples put it at $6-42 depending on PR size. Full methodology: see the write-up.

Docs

• docs/CUSTOMIZING_PROMPTS.md — full prompt-by-prompt anatomy for adapting Argus to your team.
• docs/STORAGE.md — storage modes, schema, and the HTTP contract for self-hosters.
• docs/BUILDING_A_REVIEW_LOOP.md — how to wrap Argus in a review → fix → re-review automation loop gated on CI.
• docs/PRECHECKS.md — the deterministic, non-LLM gate that runs before the LLM pipeline: CI-status routing signal and custom semgrep rules mined from Argus's own recurring findings.
• docs/ARCHITECTURE.md — the pipeline design and why it looks the way it does.
• docs/RELEASING.md — how releases are cut, for contributors.
• Agentic code review harness write-up — design principles, production results, and a worked example of a full review round.

Contributing

See CONTRIBUTING.md for dev setup and PR expectations, SECURITY.md to report a vulnerability, CODE_OF_CONDUCT.md for community standards, and CHANGELOG.md for a history of notable changes.

License

Apache License 2.0.

Built by Redesign Health. Named by Nathan Marinoff, after Argus, the many-eyed giant of Greek mythology.