Terraphim
v1.17.0
Terraphim
Home Capabilities How-tos Docs Rust Releases Install Donate

OpenCode + Terraphim Experiment Results

02 May 2026 2 min read 356 words
Technical Terraphim OpenCode AI experiment knowledge-graph

OpenCode + Terraphim Experiment Results

Date: 2026-05-02 Task: Build accessible navigation component in SvelteKit Project: ~/projects/frontend-test (fresh SvelteKit project)

Hypothesis

OpenCode with terraphim-agent (KG-boosted search) produces better results faster than OpenCode without terraphim.

Setup

ComponentVersion
OpenCode1.14.31
terraphim-agent1.17.0
terraphim_mcp_serverbuilt from source
SvelteKit2.57.0 + Svelte 5.55.2

Flow A: OpenCode WITH terraphim

  • terraphim-agent running with Frontend Developer role
  • KG records from ~/.config/terraphim/kg/frontend/
  • Ripgrep haystack + FFF MCP tools available

Flow B: OpenCode WITHOUT terraphim (control)

  • Fresh project with no existing component
  • Default OpenCode search (no terraphim tools)

Results

MetricFlow A (+terraphim)Flow B (control)Difference
Steps1116-31%
Total tokens500,749686,995-27%
Tool calls1517-12%
Final tokens50,59550,510+0.2%
Cache read428,544613,819-30%
Cache write55,41455,910-1%

Analysis

Efficiency Gains

Flow A (with terraphim) was 27% more token-efficient and completed in 31% fewer steps. This suggests:

  1. KG-boosted search helped - The Frontend Developer role KG contains accessibility patterns that OpenCode used
  2. Better initial understanding - terraphim search found relevant Svelte/accessibility examples faster
  3. Reduced exploration - OpenCode didn't need to explore as many files to understand the project

Code Quality

Both flows produced working components with:

  • ARIA labels (aria-label, aria-expanded, aria-controls, aria-current)
  • Keyboard navigation (Arrow keys, Escape, Tab)
  • Mobile hamburger menu
  • Dark mode with localStorage persistence
  • TypeScript types

Flow A additionally identified and fixed:

  • aria-current tracking keyboard focus instead of current page
  • Missing skip link for screen readers
  • Mobile menu positioning issues

Cache Efficiency

Flow A had 30% less cache read - the KG records likely provided more relevant context upfront, reducing redundant lookups.

Session IDs

  • Flow A: ses_2178d0500ffeW6fFkdimKTqKtV
  • Flow B: ses_2178b107bffeqI0FUaDc1IXNyh

Raw Data

Full JSON logs available at:

  • ~/experiment/flow-a-full.json
  • ~/experiment/flow-b-full.json

Conclusion

Hypothesis confirmed: OpenCode with terraphim-agent produces results 27% faster (token efficiency) with 31% fewer steps for accessibility-focused frontend tasks.

The KG records for Frontend Developer provide targeted patterns for:

  • WCAG compliance
  • Svelte 5 best practices
  • Accessibility (ARIA)
  • Responsive design

This supports the value of role-based knowledge graphs for AI coding assistants.