Claude Code Plugins

Community-maintained marketplace

Feedback

Convex Agents Context

@Sstobo/convex-skills
4
0

Customizes what information the LLM receives for each generation. Use this to control message history, implement RAG context injection, search across threads, and provide custom context.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name Convex Agents Context
description Customizes what information the LLM receives for each generation. Use this to control message history, implement RAG context injection, search across threads, and provide custom context.

Purpose

By default, the Agent includes recent messages as context. This skill covers customizing that behavior for advanced patterns like cross-thread search, memory injection, summarization, and filtering.

When to Use This Skill

  • Limiting context window to prevent token overflow
  • Searching across multiple threads for relevant context
  • Injecting memories or user profiles into every prompt
  • Summarizing long conversations before continuing
  • Filtering out sensitive or irrelevant messages
  • Adding few-shot examples to guide LLM

Configure Default Context Options

const myAgent = new Agent(components.agent, {
  name: "My Agent",
  languageModel: openai.chat("gpt-4o-mini"),
  contextOptions: {
    recentMessages: 50,
    excludeToolMessages: true,
    searchOptions: {
      limit: 10,
      textSearch: true,
      vectorSearch: false,
    },
  },
});

Override Context Per Call

export const generateWithCustomContext = action({
  args: { threadId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, prompt }) => {
    const result = await myAgent.generateText(
      ctx,
      { threadId },
      { prompt },
      {
        contextOptions: {
          recentMessages: 20,
          searchOptions: {
            limit: 5,
            textSearch: true,
            vectorSearch: true,
          },
        },
      }
    );

    return result.text;
  },
});

Search Across Threads

export const generateWithCrossThreadContext = action({
  args: { threadId: v.string(), userId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, userId, prompt }) => {
    const result = await myAgent.generateText(
      ctx,
      { threadId, userId },
      { prompt },
      {
        contextOptions: {
          searchOtherThreads: true,
          searchOptions: {
            limit: 15,
            textSearch: true,
            vectorSearch: true,
          },
        },
      }
    );

    return result.text;
  },
});

Custom Context Handler

Completely customize context:

const myAgent = new Agent(components.agent, {
  name: "My Agent",
  languageModel: openai.chat("gpt-4o-mini"),
  contextHandler: async (ctx, args) => {
    const userMemories = await getUserMemories(ctx, args.userId);
    const examples = getExamples();

    return [
      ...userMemories,
      ...examples,
      ...args.search,
      ...args.recent,
      ...args.inputMessages,
    ];
  },
});

Fetch Context Manually

Get context without calling LLM:

import { fetchContextWithPrompt } from "@convex-dev/agent";

export const getContextForPrompt = action({
  args: { threadId: v.string(), prompt: v.string() },
  handler: async (ctx, { threadId, prompt }) => {
    const { messages } = await fetchContextWithPrompt(ctx, components.agent, {
      threadId,
      prompt,
      contextOptions: {
        recentMessages: 20,
        searchOptions: { limit: 10, textSearch: true },
      },
    });

    return messages;
  },
});

Key Principles

  • Default context is sufficient: Most use cases work with defaults
  • Search improves relevance: Enable for long conversations
  • userId required for cross-thread: Provide when searching multiple threads
  • Context handlers are powerful: Use for memories, examples, special formatting
  • Recent messages take precedence: Used after search in context order

Next Steps

  • See rag for knowledge base context injection
  • See fundamentals for agent setup
  • See rate-limiting for token management