Source Nodes

Source nodes are documents whose body contains instructions for fetching live data from external services. They are first-class nodes in the context graph — authored, versioned, and checksummed like any other document.

Key distinction

When an agent resolves a type: document node, it reads the body as knowledge to absorb.

When an agent resolves a type: source node, it reads the body as instructions to follow — what calls to make, in what order, with what parameters, and how to interpret the results.

The markdown body is the instruction set. The agent is the execution engine.

Design principle

Frontmatter carries what machines index. The body carries what agents execute.

The frontmatter tells the resolver “this is a source node that uses the Jira MCP server and depends on another source.” The body tells the agent “call this tool with these parameters, then use the result like this.”

Example

---
title: "Current Sprint Tickets"
type: source
tags: ["#engineering", "#sprint"]
status: published
version: 2
source:
  transport: mcp
  server: jira
  tools: [jira_get_sprint_issues]
  depends_on:
    - contextnest://sources/project-config
  cache_ttl: 300
---
 
# Current Sprint Tickets
 
This source provides live sprint data from Jira.
 
## Fetch the sprint
 
Call `jira_get_sprint_issues` with:
- **project**: Read from the project-config source
- **sprint**: "current"
- **fields**: summary, status, assignee, priority
 
## Interpret the results
 
Group tickets by status (To Do / In Progress / Done).
Flag any ticket marked "blocked" or with no assignee.

Result lifecycle

Source results are session-scoped — they exist in the agent’s working memory, not in the vault. The vault stores the authored instructions, not transient API responses.