# Delegation Capability Matrix — 2026-08-15 Audit

Verified state of Hermes ↔ Claude capability boundaries after the Aug 15 conversation.

## Verified installs

| Capability | Status | Details |
|---|---|---|
| Playwright on VPS | ✅ Installed | 1.62.0, Chromium 151, browsers at `/root/.hermes/playwright-browsers/` (root fs read-only required env-var override) |
| Claude Code CLI on VPS | ✅ Installed + authenticated | v2.1.197 at `/root/.hermes/npm-global/bin/` (npm prefix override for read-only root), Anthropic API key, task-contract JSON verified |
| Chrome Extension bridging | ❌ Not possible | Extension lives on Connie; not reachable from VPS. Manual Connie execution only for Extension-dependent tasks |
| Structured task-contract endpoint | ✅ Via Claude Code CLI | `claude -p <json-task> --output-format json` returns structured envelope. Vault MCP still serves file tools only. |

## Routing logic

Task requires Chrome Extension / live browser with logged-in session?
  YES → Manual Connie execution (Claude Desktop with Extension)
  NO  → Hermes handles it (web_search, web_extract, Playwright)

Task requires complex reasoning, copy, strategy, or code generation?
  YES → Hermes shells out to Claude Code CLI (claude -p)
  NO  → Hermes handles it directly

Task requires both browser + reasoning?
  YES → Hermes scrapes with Playwright → sends data to Claude Code CLI → validates result

## Cost notes

- Playwright: free, ~115MB install (headless shell), lightweight
- Claude Code CLI: free to install, ~$0.05/task (Haiku routes + Opus reasons). At 20 tasks/day ≈ $30/month
- VPS resources: both are lightweight

## Key lesson from Aug 15

The "orchestrator/specialist" architecture doc assumed delegation was already wired. It isn't — the architecture is correct, but the plumbing needed installs. Both Playwright and Claude Code CLI are now installed and tested. Don't promise "Hermes delegates to Claude" until both are verified (they are now, as of 2026-08-15).