Agent Skills for Synalinks

---

This repository contains skills for coding agents that read the open Agent Skills format (SKILL.md): Claude Code, Codex, OpenCode, pi, and others.

Quick install

Install it as a plugin in one command:

# Claude Code (2.1.275 or later)
/plugin install synalinks --marketplace SynaLinks/synalinks-skills

# Codex
codex plugin marketplace add SynaLinks/synalinks-skills
codex plugin add synalinks@synalinks-skills

# OpenCode
opencode plugin github:SynaLinks/synalinks-skills

Any other agent (Cursor, Gemini CLI, Copilot, pi, ...):

npx skills add SynaLinks/synalinks-skills --skill synalinks

Details, global installs and Claude.ai uploads are in Install.

What are Agent Skills?

Agent Skills are reusable instructions, in a plain SKILL.md file, that teach a coding agent how to perform a task the way you want it done. The format started at Anthropic and is now an open standard read by agents across the ecosystem, so one skill works in every agent that supports it, unmodified.

What is the goal of Synalinks skills?

The goal of Synalinks skills is to teach coding agents to use the Synalinks framework correctly. Synalinks is Keras-inspired, so without guidance LMs tend to mix Keras / LangChain / DSPy syntax, producing plausible-looking but broken code. These skills constrain the agent to idiomatic Synalinks usage.

The synalinks skill

This repo ships a single skill, synalinks, that covers the whole framework. The agent auto-activates it based on its SKILL.md description. The SKILL.md is a scannable overview, organized into one section per slice of the framework, each pointing to a runnable script (and its captured run log) and a deep-dive reference doc:

Section	What it covers
Core	DataModel, Field, Input, JSON operators (+ & | ^ ~), synalinks.ops, configuration, LanguageModel/EmbeddingModel basics, DecisionModel
Programs	Program class, four building APIs (Functional / Sequential / Subclassing / Mixed), multi-input/output graphs, build/call lifecycle, save/load, summary, get_module, plot_program, custom serialization
Modules	Generator, ChainOfThought, SelfCritique, Identity, PythonSynthesis, SequentialPlanSynthesis, custom modules via subclassing
Control flow	Decision, MultiDecision, Branch, parallel branches, self-consistency, XOR input/output guards, And/Or modules, branch merging
Decision models	DecisionModel (TypeSafe jev), fields as questions, the decision_model argument, min_confidence/threshold abstention, tuning thresholds with KerasTuner
Agents	FunctionCallingAgent, RLM, DeepAgent, Tool definitions, MCP integration (MultiServerMCPClient), subagents, trajectories
Knowledge	KnowledgeBase (DuckDB), EmbedKnowledge, UpdateKnowledge, RetrieveKnowledge, RAG/KAG, hybrid search, Entity/Relation graphs
Training	compile() / fit() / evaluate() / predict(), callbacks, ProgramCheckpoint, training workflow
Rewards	ExactMatch, CosineSimilarity, LMAsJudge, ProgramAsJudge, RubricsAsJudge and its presets (Faithfulness, Toxicity, TaskCompletion, ...), AgentAsJudge, DeepAgentAsJudge, RLMAsJudge, ComposableReward, BatchReward, F1Score, custom rewards/metrics, masking
Optimizers	RandomFewShot, OMEGA, Dominated Novelty Search, mutation/crossover, quality-diversity tuning
Providers	Provider prefixes (openai, anthropic, groq, openrouter, cohere, deepseek, together_ai, bedrock, doubleword, hosted_vllm, ...), local OpenAI-compatible servers (LMStudio/vLLM/ollama), OpenAI-compatible setups (e.g. OpenRouter embeddings via api_base)
Datasets

The skill folder contains:

• SKILL.md: frontmatter + scannable usage guide for the whole framework
• scripts/: runnable example scripts (each docstring states its Usage: and Run log:)
• references/: deep-dive reference docs plus the captured run logs (*.log)

Install

Pick the route that matches your agent:

Agent	Install with
Claude Code, Codex, OpenCode	the plugin (recommended)
Any other agent (Cursor, Gemini CLI, Copilot, pi, ...)	npx skills
Claude.ai (web) and Claude Desktop	a .skill upload

The skill content is identical in every case: one SKILL.md, no per-agent variants to maintain.

As a plugin

This repo is a plugin marketplace for Claude Code and Codex, and an OpenCode plugin. The agent fetches the skill itself: no Node.js, no copied files.

Claude Code (2.1.275 or later):

/plugin install synalinks --marketplace SynaLinks/synalinks-skills

Then run /skills: synalinks should appear.

Codex:

codex plugin marketplace add SynaLinks/synalinks-skills
codex plugin add synalinks@synalinks-skills

Start a new Codex session to load it.

OpenCode:

opencode plugin github:SynaLinks/synalinks-skills           # this project
opencode plugin github:SynaLinks/synalinks-skills --global  # every project

This adds the plugin to .opencode/opencode.json (or your global config). You can also list it yourself:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["github:SynaLinks/synalinks-skills"]
}

The plugin registers the repo's skills/ directory as a skill source, so OpenCode discovers the skill without copying files. Restart OpenCode to load it, then check with opencode debug skill.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗