LearnNewsExamplesServices
Frontmatter
id9892
titledocs: Expand CodebaseOverview with Agent OS internals and DevIndex architecture
stateClosed
labels
documentationai
assigneestobiu
createdAtApr 11, 2026, 9:23 PM
updatedAtApr 12, 2026, 10:29 AM
githubUrlhttps://github.com/neomjs/neo/issues/9892
authortobiu
commentsCount1
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
blockedBy[]
blocking[]
closedAtApr 12, 2026, 10:29 AM

docs: Expand CodebaseOverview with Agent OS internals and DevIndex architecture

tobiu
tobiu commented on Apr 11, 2026, 9:23 PM

Context

CodebaseOverview.md is the primary onboarding document for AI agents. It's comprehensive for the frontend engine and component system but has significant gaps in the Agent OS internals discovered during Claude Opus 4.6's first deep-read sessions.

Missing Content

1. DevIndex as a Flagship Application

DevIndex is the only application showcasing Neo.core.Base as a backend runtime. Its service architecture (Spider.mjs, Updater.mjs, Storage.mjs) runs as Node.js singletons, not browser components. This is architecturally significant — it proves Neo's class system works identically in both environments. Currently absent from the Flagship Applications section.

2. Bridge Topology

The overview says "Neural Link (WebSocket JSON-RPC)" but doesn't explain the three-party architecture:

Bridge Process (Standalone WebSocket Hub)
    ↑                    ↑
    │                    │
Agent MCP Server    App Worker (Browser)
(Node.js client)    (Browser client)

This distinction matters because it enables multi-agent concurrent access to the same running application — a core capability for swarm operations.

3. Daemon Lifecycle Scripts

runSandman.mjs and runGoldenPath.mjs are the entry points for offline cognitive maintenance. New agents need to know these are node buildScripts/ai/runSandman.mjs invocations, not MCP protocol operations. The overview mentions DreamService abstractly but not these concrete runners.

4. Shared MCP Infrastructure

ai/mcp/ToolService.mjs is the universal dispatch layer for all 5 MCP servers. Understanding that callTool() is the single choke-point across ALL servers is architecturally critical for agents planning interceptors or middleware (e.g., the NL Action Recorder #9889).

5. Agent Profiles

ai/agent/profile/Librarian.mjs demonstrates the agent subclass pattern with configurable modelProvider. The overview doesn't mention that agents have profiles, or that local inference via MLX Server / ML Studio is a first-class deployment target (April 2026).

6. ConnectionService Session Awareness

Brief mention that the NL server tracks live App Worker sessions with appName, connectedAt, console logs, and window topology via ConnectionService.sessionData. This is what makes the Neural Link "session-aware."

A2A Context (Fat Ticket)

Files to Modify

File Change
learn/guides/fundamentals/CodebaseOverview.md Add DevIndex section under Flagship Applications; expand Agent OS section with Bridge topology, daemon scripts, agent profiles, shared ToolService, ConnectionService

Scope Boundary

This is a documentation-only ticket. No code changes. The prior #9875 (already merged) addressed AI infrastructure scale and boundaries. This ticket addresses the operational internals that agents need to understand to work effectively with the Agent OS.

Verification Plan

  1. Visual review of the updated document
  2. Confirm no duplication with existing content from #9875
tobiu added the documentation label on Apr 11, 2026, 9:23 PM
tobiu added the ai label on Apr 11, 2026, 9:23 PM
tobiu assigned to @tobiu on Apr 11, 2026, 9:23 PM
tobiu
tobiu Apr 11, 2026, 9:29 PM

Input from Claude Opus 4.6 (Antigravity):

❋ ## Scope Extension: ask_knowledge_base in CodebaseOverview

Connecting dots with #9893 — the CodebaseOverview's "Key Concepts: Query Entry Points" section (line 480) and "When Stuck" section (line 579) both reference query_documents exclusively. Neither mentions ask_knowledge_base.

For a ~479k-line platform, this distinction matters operationally:

  • query_documents → "Here are 10 relevant files to read" (agent must still read them)
  • ask_knowledge_base → "Here is the synthesized answer with citations" (zero file reads needed)

Proposed addition to CodebaseOverview: A brief callout in the "Why This Matters for You" section (line 461) or the "When Stuck" section (line 579):

Ask, Don't Search: For conceptual questions ("How does reactivity work?", "What is the config system?"), use ask_knowledge_base instead of query_documents. It synthesizes answers from multiple source files in a single call — essential for getting the top-level perspective in a codebase of this scale.

This should be included in the #9892 implementation scope alongside the DevIndex and Agent OS internals additions.

tobiu cross-referenced by PR #9894 on Apr 11, 2026, 9:56 PM
tobiu referenced in commit 43a11df - "docs: expand CodebaseOverview with Agent OS internals and DevIndex architecture (#9892)" on Apr 12, 2026, 10:09 AM
tobiu cross-referenced by PR #9897 on Apr 12, 2026, 10:10 AM
tobiu referenced in commit 84c52aa - "docs: expand CodebaseOverview with Agent OS internals and DevIndex architecture (#9892)" on Apr 12, 2026, 10:24 AM
tobiu referenced in commit 975f13a - "docs: expand CodebaseOverview with Agent OS internals and DevIndex architecture (#9892) (#9897)" on Apr 12, 2026, 10:29 AM
tobiu closed this issue on Apr 12, 2026, 10:29 AM
tobiu cross-referenced by #9898 on Apr 12, 2026, 10:36 AM