dsh-review

Independent code review for Codex, powered by DeepSeek Harness (dsh) + DeepSeek V4 Flash. A second opinion from a different model family and a different harness — at near-zero cost.

Agent = Model + Harness. Most review tools re-run the same model and the same harness — which means the reviewer shares the author's blind spots. dsh-review sends your diff to a genuinely independent pair: DeepSeek's V4 Flash running inside DeepSeek's own harness, headlessly, and brings back a structured review you can act on.

Why you want a second opinion

• A model that writes code tends to miss its own blind spots (self-review bias).
• Cross-review with an independent third party is the 2026 community consensus: don't let the author be the judge.
• dsh-review turns that workflow into one command: gather diff → DeepSeek Harness reviews it → structured report back.

Features

• 🧠 Independent by design — different model family (DeepSeek V4 Flash) and different harness (DSH) than Codex
• ⚡ Fast — one review in seconds to a few minutes
• 💰 Cheap — V4 Flash cache-hit input pricing is roughly $0.0028/M; batch reviews without watching the meter
• 📜 Traceable — DSH's headless run keeps an append-only session log
• 🔌 Flexible scope — git diff, specific files, or a whole repository

Install

Requirements: Node.js, a DeepSeek API key (DEEPSEEK_API_KEY), network access.

One command:

git clone https://github.com/Tinzlu/dsh-review.git ~/.codex/skills/dsh-review

Or with the official Codex skill installer:

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py --repo Tinzlu/dsh-review --path .

Restart Codex, then just ask:

用 dsh-review 审查当前 diff

Run dsh-review on this PR / these files

Usage

The skill works from inside Codex (just describe the scope), or directly from a terminal:

# review a file
~/.codex/skills/dsh-review/scripts/dsh_review.sh "严格审查 /path/to/file.py，按严重程度列出问题"

# review from stdin
echo "审查下面这段 diff：$(git diff)" | ~/.codex/skills/dsh-review/scripts/dsh_review.sh

The script wraps npx -y @deepseek-ai/dsh --profile headless "<prompt>" — the first run downloads the package, later runs are fast.

Proof it works

• Sample report — a real headless DSH review, produced in a live session
• Cross-review comparison — the same file reviewed by Codex and by DSH + V4 Flash, side by side

How it works

Codex collects scope (git diff / files)
        ↓
dsh_review.sh → npx @deepseek-ai/dsh --profile headless "<review prompt>"
        ↓
DeepSeek V4 Flash reviews with DSH's toolset
        ↓
structured findings → back to Codex → delivered to you

FAQ

How much does it cost? A typical review costs cents or less. V4 Flash is cheap, and DSH's cache-friendly session design makes long, repeated reviews economical.

Does it replace my normal code review? No — it complements it. Use it as the independent second opinion, especially after Codex or Claude wrote the code.

Does my code leave the machine? Yes, the review request is sent to the DeepSeek API, like any DeepSeek usage. Never include secrets, tokens, or credentials in the reviewed scope.

DSH is a developer preview — what if it breaks? The skill documents a fallback: if dsh is unavailable (missing key, network, install failure), Codex performs a normal review and says so.

License

MIT