deepseek-offload
Offload task to DeepSeek or other neural lanes (no Agent spawning). Primary DeepSeek entry point — use for flash extraction, pro reasoning, research, and NVIDIA lanes.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- deepseek-offload
- triggers
- [ "/offload", "/ds-flash", "/ds-pro", "/research", "/deepseek", "/nvidia", "/nvidia-deepseek" ]
- description
- Offload task to DeepSeek or other neural lanes (no Agent spawning). Primary DeepSeek entry point — use for flash extraction, pro reasoning, research, and NVIDIA lanes.
- routing_note
- Primary lane router. For structured forge/plan/execute orchestration use /deepseek-workhorse. NVIDIA lanes: /nvidia-deepseek (llama-70b), /nvidia-nemotron (requires elevated NIM access).
03 SKILL.md 正文
DeepSeek Offload Skill
Direct routing to DeepSeek V4 (Flash/Pro) or other neural lanes via Scripts/offload.sh. No Anthropic Agent spawning — cost-conscious execution.
Usage
Main command: /offload <lane> <prompt>
/offload deepseek-v4-flash "Extract key findings from this report"
/offload deepseek-v4-pro "Reason through this problem step-by-step"
/offload deepseek-cloud "Complex code review with thinking enabled"
Convenience aliases
/ds-flash "Quick extraction/summarization task"
/ds-pro "Deep reasoning, code review, complex analysis"
/research "Research task routed to best available lane"
Available Lanes
| Lane | Model | Thinking | Tokens | Timeout | Best For |
|---|---|---|---|---|---|
deepseek-v4-flash |
Flash | No | 4K | 60s | Fast extraction, summaries, simple tasks |
deepseek-v4-pro |
Pro | Yes | 8K | 120s | Code review, reasoning, complex analysis |
deepseek-cloud |
Pro | Yes | 8K | 120s | Same as Pro (alias) |
deepseek-v4-pro-lite-budget |
Pro | No | 1K | 45s | Budget-constrained work |
Implementation
Routes via Scripts/offload.sh → offload-runner.mjs. Environment:
DEEPSEEK_API_KEYmust be set in~/.zshrc(done)- Falls back to empty key → lane skips with SKIPPED_MISSING_KEY
Cost Notes
- Flash: ~$0.14 / 1M input, $0.28 / 1M output tokens
- Pro: ~$0.55 / 1M input, $2.19 / 1M output tokens (thinking adds cost)
- Pro-lite-budget: Pro pricing, 1K token cap (cost-effective for triage)
Use Flash for extraction; Pro only for reasoning that needs it.
Session Notes
2026-05-29
- session: 210m | peak ctx: 54% | compacts: 0
- tools: Bash×131, Read×93, Edit×92, TodoWrite×10, Write×7, StructuredOutput×5, ToolSearch×1, Workflow×1
- corrections: none
- errors: none
2026-05-17
- session: 160m | peak ctx: 0% | compacts: 0
- tools: Bash×119, Read×52, Edit×34, Write×5, ToolSearch×2, mcp×1, ExitPlanMode×1, Skill×1
- corrections: none
- errors: none
2026-05-17
- session: 157m | peak ctx: 0% | compacts: 0
- tools: Bash×116, Read×52, Edit×34, Write×5, ToolSearch×2, mcp×1, ExitPlanMode×1, Skill×1
- corrections: none
- errors: none
2026-05-05
- Skill created, all lanes verified wired
- DeepSeek API key in ~/.zshrc confirmed present
- Live test:
deepseek-v4-flash "What is 2+2?"→ "The answer is 4." ✓ - No Agent() calls used; direct offload.sh routing only
04 同仓库更多 Skills
zoom-out
Tell the agent to zoom out and give broader context or a higher-level perspective. Use when you're unfamiliar with a section of code or need to understand how it fits into the bigger picture.
yuri-shura
6-perspective adversarial review for high-stakes turns (architecture decisions, refactor planning, deployment review) — fans out architect, adversary (7-vector attack), maintainer, ops, product, and security perspectives in parallel, then consolidates. Use when user says 'review this decision', '/shura', 'adversarial review', 'what could go wrong', 'architecture review', or 'deployment review'.
yuri-sales-intelligence
Unified sales and outreach skill generating cold outreach openers, social selling sequences, objection rebuttals, and ad creative for any paid platform — multi-persona aware, adapting tone and offer frame to buyer type. Use when user says 'write a cold email', 'outreach sequence', 'objection handling', 'ad copy', or 'sales pitch'.
yuri-code-intelligence
Unified code quality intelligence layer that diagnoses code smells (Long Method, Feature Envy, design pattern issues), prescribes targeted Fowler refactoring techniques, and reviews code for security, performance, and clarity. Use when user says 'review this code', 'refactor this', 'what's wrong with this method', 'code smells', or 'clean this up'.
yuri-adw
Invoke when user asks to run the ADW loop / agentic engineering loop / software factory workflow.
writing-skills
Use when creating new skills, editing existing skills, or verifying skills work before deployment