Enhancement: Code-review-graph integration for token-efficient agent context

[Dhwanil25]

Overview

Integrate code-review-graph as an optional enhancement to the multi-agent engine. Instead of feeding agents broad task descriptions with no structural knowledge of the codebase they're working on, this integration queries a local AST-based knowledge graph to compute the blast radius of a task — only the affected functions, callers, and dependency edges — and injects that as targeted context per agent.

Measured impact: 6.8x average token reduction (up to 49x on large monorepos). A review that consumed 12,507 tokens drops to 458 with equal or better quality.

---

Problem

When Agentis agents handle coding tasks today:

• Each agent receives the full task description and reasons about it blind
• No structural knowledge of the actual codebase is available
• Coder/Reviewer agents consume large token budgets re-deriving context that a graph already knows

---

Proposed Architecture

Data flow

User types task
  → (if graph enabled) query blast radius
  → orchestrator plans with graph context
  → each worker receives only its relevant code slice
  → agents work on targeted context (~450 tokens vs ~12,000)

Files to change

vite-plugin-agentis.ts
Add two middleware endpoints that proxy to the code-review-graph MCP sidecar:

• GET /agentis/graph/status — checks if an instance is running
• POST /agentis/graph/query — takes a task string, returns blast radius (files + functions in scope)

src/lib/multiAgentEngine.ts

• buildOrchestratorSystem() — pass a graphContext string (blast radius summary) so the orchestrator plans around actual affected code
• executeWorkers() — inject relevant graph nodes into each worker's prompt before streamLLM. Coder gets function signatures + callers. Reviewer gets dependency edges. Researcher gets the import graph.

src/components/pages/UniversePage.tsx

• "Attach codebase" toggle below the task textarea
• Pre-launch badge: "14 functions · 3 files in scope" so the user sees what was scoped before agents spawn

src/components/pages/SettingsPage.tsx

• New "Integrations" section with a code-review-graph card
• Toggle to enable/disable
• MCP server URL field (default localhost:8765)
• "Build graph" button to trigger initial indexing

---

Activation logic

Only activates for coding-type tasks. The orchestrator detects whether the task involves code, files, or functions — general tasks (research, writing, analysis) bypass the graph entirely so there's no overhead for non-coding workflows.

---

Implementation options

Option	Effort	Notes
Sidecar proxy (recommended)	Low	Run code-review-graph MCP server separately; Agentis engine proxies to it
Port to TypeScript	High	Full rewrite of AST parsing + SQLite graph in TS — Tree-sitter has TS bindings

Option 1 (sidecar) is the right starting point. The MCP server runs alongside the Agentis engine and the plugin middleware handles the bridge.

---

References

• code-review-graph — 6.8x token reduction, Tree-sitter + SQLite, 14 languages, MCP protocol
• Related: vite-plugin-agentis.ts middleware pattern already used for OpenClaw migration endpoints