Context

What it is

Context is everything the language model can "see" when generating a response. It is the working memory of an AI session — the system prompt, your conversation, tool outputs, file contents, and project-specific rules all packed into a single token budget.

What it is in real OpenCode terms

In OpenCode, context is the combined set of information fed to the LLM on each turn. It includes:

• System prompt - OpenCode's base instructions for how to behave
• Conversation history - all previous user messages and assistant responses in the current session
• Tool outputs - file contents from read, output from bash, search results, etc.
• Rules files - project-specific instructions from AGENTS.md or files in .opencode/rules/
• Instructions config - additional context files pointed to by the "instructions" field in opencode.json
• MCP server descriptions - tool schemas and descriptions from connected MCP servers
• Skill content - on-demand instruction content loaded by skills

Context is session-scoped. It is not shared between different sessions. Each session starts with its own clean context that grows as you interact.

Why it exists

• Grounding. Without context, the LLM would hallucinate project details. Context gives it real facts about your codebase.
• Customization. Rules files and instructions let you teach the agent your project's conventions, preferred patterns, and constraints.
• Tool integration. Tool outputs (file reads, search results, bash output) are part of context, so the agent can reason about real data.

Anatomy / main elements

Context has a finite size (measured in tokens). Here is roughly how it fills up:

┌─────────────────────────────────┐
│  System Prompt                  │  ← OpenCode's base instructions
├─────────────────────────────────┤
│  Rules / Instructions           │  ← AGENTS.md, .opencode/rules/
├─────────────────────────────────┤
│  Conversation History          │  ← Messages from this session
├─────────────────────────────────┤
│  Tool Outputs                  │  ← File contents, bash output, etc.
├─────────────────────────────────┤
│  MCP Tool Descriptions         │  ← Schemas from connected servers
├─────────────────────────────────┤
│  Skill Content                 │  ← Loaded on-demand by skills
└─────────────────────────────────┘

When this fills up, compaction kicks in to summarize and reduce the token count while preserving important state.

Where it lives

Context is assembled at runtime for each session. The pieces come from:

• opencode.json - the "instructions" field points to extra context files
• AGENTS.md - project root, contains agent instructions
• .opencode/rules/ - directory of rule files loaded into context

How to create one

You do not "create" context manually. OpenCode assembles it. But you can influence what goes into it:

# Add project instructions
cat > AGENTS.md <<'EOF'
# Project Rules
- Use TypeScript strict mode
- Prefer named exports over default exports
- Run `npm run lint` before committing
EOF

# Add more rule files
mkdir -p .opencode/rules
cat > .opencode/rules/testing.md <<'EOF'
# Testing Rules
- Write tests for all new functions
- Use vitest, not jest
- Test file goes next to source file with .test.ts suffix
EOF

# Point to additional context in opencode.json
cat > opencode.json <<'EOF'
{
  "instructions": ["./docs/ARCHITECTURE.md", "./docs/API.md"]
}
EOF

Minimal example

A project with well-structured context:

# Project structure
my-project/
  AGENTS.md                    # General agent instructions
  .opencode/
    rules/
      architecture.md          # Architecture decisions
      coding-style.md          # Code conventions
    plugins/
      lint-on-edit.ts          # Plugin with hooks
  opencode.json
    # { "instructions": ["./docs/CONTRIBUTING.md"] }

# When you start a session, context includes:
# - System prompt (built-in)
# - AGENTS.md content
# - All .opencode/rules/*.md files
# - docs/CONTRIBUTING.md content
# - Conversation history (grows as you chat)
# - Tool outputs (file reads, searches, etc.)

Related concepts

Common confusion