convez.ai
PluginDesenvolvimento WebIntermediário
zred0627

DS Rescue CC

zred0627/ds-rescue-cc

Segunda opinião DeepSeek para planos, esquemas e commits do Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

CLI que integra DeepSeek como revisor independente de código e planos para Claude Code e outras ferramentas LLM. Evita o efeito câmara de eco ao trazer uma segunda opinião de uma família de modelos diferente, detectando 1.7–2.3× mais bugs que auto-revisão.

Instalação

$npx convez-ai zred0627/ds-rescue-cc

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 6 meses
Versão
v0.1.3
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisar planos de implementação antes de execução
  • Analisar trade-offs em decisões arquiteturais
  • Validar diffs de commits com perspectiva cross-family

Veredito

Ferramenta pragmática que resolve um problema real: o viés sistemático de modelos da mesma família. A integração com DeepSeek é custo-efetiva (~$0.01 por revisão) e o CLI é direto. Ideal para equipes que usam Claude Code e querem rigor adicional em decisões críticas.

README

A tradução para o português deste README ainda está na fila.

ds-rescue: DeepSeek-Powered Cross-Family Code Review for LLM Coding Tools

ds-rescue is a single-binary CLI that uses DeepSeek API to give Claude Code (and other LLM CLIs) an independent second opinion on plans, schemes, and commits — zero echo chamber.

License: MIT Go 1.22+ DeepSeek v4-pro PRs welcome

English | 中文


TL;DR

  • What: 1-binary CLI for cross-family LLM review (plan / scheme / execution modes)
  • Who: Claude Code / opencode / Continue / Aider users wanting a non-Anthropic second opinion
  • Why: Same-family LLM self-review = echo chamber. Cross-family catches 1.7–2.3× more bugs.
  • How: go install github.com/zred0627/ds-rescue-cc/cmd/ds-rescue@latest && ds-rescue --check
  • Cost: DeepSeek v4-pro ~$0.01/review, v4-flash even cheaper

Three review modes

Mode: plan (default)

Reviews implementation plan documents: task DAGs, architecture decisions, risk registers. Surfaces missing assumptions, ordering bugs, rollback gaps, and unhandled failure branches.

ds-rescue docs/plans/my-feature.md           # auto-detected as plan mode
ds-rescue --mode plan docs/plans/my-feature.md
ds-rescue --plan docs/plans/my-feature.md    # alias

Mode: scheme

Reviews scheme comparison tables: A vs B vs C analysis, option trade-offs, technology selection matrices. Applies adversarial pressure to each option and challenges unstated assumptions.

ds-rescue --mode scheme architecture-choices.md
ds-rescue --scheme architecture-choices.md

Mode: execution

Reviews the git diff of the last commit or staged changes. The binary runs git diff HEAD~1 HEAD internally — no manual diff piping required.

ds-rescue --mode execution
ds-rescue --execution

Why cross-family review?

You finish a plan. Claude reviews it — clean architecture, no obvious gaps. You ship it.

Three days later: a race condition in the retry logic that Claude never flagged. Why? Because Claude wrote the plan. Asking the same model family to review its own output is structurally an echo chamber — models trained on similar corpora share systematic blind spots.

Real dogfooding example:

Claude self-review of a 3-service deployment plan: 5/5 PASS — no issues found.
DeepSeek review of the same plan:
  - Missing idempotency guard on the DB migration step (data corruption risk on retry)
  - Rollback order reverses dependency graph (service B torn down before service A)
  - Health-check timeout too short for cold-start on constrained VMs
  - No mention of what happens if Step 4 succeeds but Step 5 times out

Four findings. Zero from Claude. Same plan.

Research on code review effectiveness consistently shows that reviewers with different backgrounds find 1.7–2.3× more defects than same-background reviewers. The same applies to LLMs. The goal is not to replace Claude — it is to add an independent reviewer from a different family.


Why DeepSeek?

codex-plugin-cc pioneered bringing a non-Claude reviewer into Claude Code. ds-rescue-cc builds on that foundation.

ds-rescue uses DeepSeek because cross-family value requires a model from a materially different training distribution. DeepSeek's pricing makes habitual use practical — the marginal cost per review is low enough that you stop counting and just run it.

  • Cross-family complementarity: different training distribution from Anthropic models, non-overlapping blind spots
  • Cost-efficient: low enough to review every plan and commit without thinking about the bill
  • Chinese-domain reasoning: stronger for manufacturing / logistics / policy problems
  • Strong coding benchmark performance: good enough to catch real bugs

Quickstart

Install

# Recommended for all OSes (bypasses Windows Smart App Control)
go install github.com/zred0627/ds-rescue-cc/cmd/ds-rescue@latest

# Linux / macOS pre-built binary
curl -fsSL https://raw.githubusercontent.com/zred0627/ds-rescue-cc/main/scripts/install.sh | sh

# Windows PowerShell (with SAC detection and go install fallback)
iwr https://raw.githubusercontent.com/zred0627/ds-rescue-cc/main/scripts/install.ps1 | iex

Set your API key

export DEEPSEEK_API_KEY=sk-your-key-here
# or persist: mkdir -p ~/.ds-rescue && echo "sk-your-key-here" > ~/.ds-rescue/key

Get a DeepSeek API key at platform.deepseek.com.

First review

ds-rescue --check                             # self-check: key, binary, SKILL.md, model
ds-rescue docs/plans/my-feature.md           # plan mode
ds-rescue --scheme architecture-choices.md   # scheme mode
ds-rescue --execution                        # review last commit

Claude Code plugin

/plugin marketplace add zred0627/ds-rescue-cc
/plugin install ds-rescue

Then use /ds-rescue <file> from any Claude Code session:

/ds-rescue --plan path/to/plan.md
/ds-rescue --scheme path/to/scheme-comparison.md
/ds-rescue --execution

The plugin is a thin integration layer over the same CLI binary — surfaces findings inline in the conversation and checks for the binary on session start.


FAQ

How is ds-rescue different from codex-plugin-cc?

Dimensioncodex-plugin-ccds-rescue
Language / runtimeNode.jsGo (single binary, no runtime)
Model familyOpenAIDeepSeek (v4-pro / v4-flash)
Chinese reasoning qualityStandardStronger for CN-domain problems
Distribution (Windows)npm installgo install (SAC-safe) or pre-built binary
Pioneer creditYes — established the patternBuilds on that foundation

Why does Smart App Control block the Windows download?

Windows 11 SAC blocks downloaded unsigned binaries. Our binary is unsigned — EV code signing costs $300–500/year. Three workarounds in order of preference:

# 1. Compile locally — no MOTW flag, SAC-safe (RECOMMENDED)
go install github.com/zred0627/ds-rescue-cc/cmd/ds-rescue@latest

# 2. Unblock after download
Unblock-File "$env:USERPROFILE\bin\ds-rescue.exe"

# 3. Add an exclusion (requires Administrator)
Add-MpPreference -ExclusionPath "$env:USERPROFILE\bin\ds-rescue.exe"

Which DeepSeek models are supported?

AliasModelCharacteristics
pro (default)deepseek-chat (v4-pro)Best quality, ~$0.01/review
flashdeepseek-chat-fastFaster, ~$0.002/review

Can I customize the review persona?

Yes — fork or edit plugins/ds-rescue/skills/ds-plan-challenger/SKILL.md. The file defines review behavior for all three modes. Point ds-rescue at your version:

export DS_RESCUE_SKILL_PATH=/path/to/your/SKILL.md
# or per-invocation: ds-rescue --skill /path/to/your/SKILL.md --plan my-plan.md

Is my code sent to DeepSeek's servers?

Yes, as with any LLM API call. If your organization policy prohibits sending code to external LLM APIs, that constraint applies here as it does to all LLM-powered tools.


Configuration

API key resolution priority

  1. $DEEPSEEK_API_KEY environment variable
  2. $XDG_CONFIG_HOME/ds-rescue/key (Linux/macOS)
  3. %APPDATA%\ds-rescue\key (Windows)
  4. ~/.ds-rescue/key (fallback)

SKILL.md path resolution priority

  1. --skill <path> CLI flag
  2. $DS_RESCUE_SKILL_PATH environment variable
  3. $CLAUDE_PLUGIN_ROOT/skills/ds-plan-challenger/SKILL.md (Claude Code plugin)
  4. ~/.ds-rescue/skills/ds-plan-challenger/SKILL.md

CLI flags reference

FlagDefaultDescription
--mode {plan|scheme|execution}auto-detectReview mode
--plan / --scheme / --execution—Mode aliases
--skill <path>(see above)Path to SKILL.md
--model {pro|flash}proDeepSeek model alias
--api-timeout <seconds>90API call timeout
--exec-timeout <seconds>30Per-tool-call timeout
--no-toolsfalseDisable agentic tool-use loop
--verbosefalseLog every tool call to stderr
--version—Print version and git commit
--check—Self-check: key, binary, SKILL.md, model

Contributing

See CONTRIBUTING.md. PRs on SKILL.md prompt definitions are especially welcome — no Go knowledge required.


License

MIT — see LICENSE

Copyright (c) 2026 Daniel


Related work


Maintained by Daniel (zred0627) · v0.1.0 (2026-05)

Ver o original no GitHub

Bom para

Relacionados