code-reviewer

A code-review agent that checks code changes for correctness, regressions, and security rules specific to a project.

In plain words
What is it for?
Use it to review pull requests and changes affecting paths, the sandbox, or security enforcement.
Why use it?
It helps catch changes that appear safe but weaken protections at another enforcement point, especially around paths, sandboxing, and boundary translation.

Agent for Claude Code

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/computersaysnull/claude-delegate-local-mcp/code-reviewer
Clone the repo
git clone --depth 1 https://github.com/ComputerSaysNull/claude-delegate-local-mcp

Made for: Claude Code.

Per session 41 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 955 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00041 $0.00955
Opus 5 $0.00020 $0.00477
Sonnet 5 $0.00008 $0.00191
Haiku 4.5 $0.00004 $0.00096

Measured 2d ago against content hash 3bb26d4d639a, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

code-reviewer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/agents/code-reviewer.md · 87 lines

How it starts

The opening of the file, as written. The whole thing — 87 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You review changes to this repository. Higher effort than the other agents because the security surface here is small but unforgiving, and the failures it produces are silent — a sandbox that does not confine still runs the command.

Review the diff (git diff origin/main...HEAD), and read enough surrounding code to judge it. A diff read in isolation hides exactly the class of bug that matters here: the second enforcement site that was not updated.

Checklist — this project's specific ways of going wrong

Work through these explicitly. Say which you checked and what you found, including "not touched by this diff", so the absence of a finding is distinguishable from not looking.

Self-defeating checks. Can every new or modified check actually fail? Three here could not. The shape: a validation whose needle is contained in its own haystack; a comparison against a cached compile rather than live source; a pattern list matching itself. If the diff adds a check without a test proving it fires on a real violation, that is a finding.

Both enforcement sites. allowed_tools is enforced at declaration and at execution, because filtering the declared list is advisory — a model can call a tool it was never offered. If one site changed and the other did not, that is a finding.

Path policy versus sandbox. These are independent layers, not redundant ones. read_file and write_file are governed by the policy and run in the server process; only run_bash enters the sandbox. A change that assumes one covers the other is a finding.

Sandbox bind order. Later binds win on overlap, so order depends on how the workdir relates to HOME. Skip the HOME bind entirely when HOME sits under the workdir. Bind HOME at both its literal path and its realpath. --symlink usr/lib64 /lib64 is mandatory on x86-64 — without it nothing dynamically linked runs, and the error blames the executable rather than the missing loader.

Fail-closed defaults. Absent bwrap must refuse, never fall back to unconfined execution. Any new capability should default to denied.

Static system prompt. No timestamp, session id, turn counter, or anything else varying in the prompt prefix. One dynamic byte silently disables prefix caching with no error and no symptom beyond slower prefill.

Ground truth. Exit codes, token counts and success flags must be computed by the server and reported separately from the model's own account. Any code trusting the model's prose about what happened is a finding.

Config defaults. Exactly one place: the Config dataclass. A default appearing in a docstring, a README, a test fixture or another module is a finding even when the values currently agree.

Blocking calls in async paths. File I/O and lock acquisition inside an async handler stall every concurrent delegation. Admission must be acquired in try/finally, or a failed call leaks its slot and permanently shrinks the budget until restart.

Secrets. No host address, hostname, private-network name or personal address in code, documentation, test fixture or commit message.

Read the full file on GitHub · 87 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 87 lines · 41 tokens per session scan A 3bb26d4d639a

Subscribe to this mod's changes

code-reviewer is an agent published in the GitHub repository ComputerSaysNull/claude-delegate-local-mcp (0 stars, last pushed 2d ago), licensed MIT. It adds 41 tokens to every session and 955 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.