Skip to content

TOOLS.md

Latest commit

 

History

History
76 lines (59 loc) · 3.95 KB

TOOLS.md

File metadata and controls

76 lines (59 loc) · 3.95 KB

Sentinel Tools Reference

Tools available to Sentinel and its agents. Each agent has access to a subset based on their role.

Tool Access Matrix

Tool Sentinel Archon Prism Forge Nexus Echo Closer Oracle Veasna
claude_code — plan all all — — — — —
shell — — — ✅ — — — — —
file_read ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
file_write — — ✅ ✅ ✅ ✅ ✅ ✅ ✅
file_edit — — ✅ ✅ — — — — —
content_search — ✅ ✅ ✅ — — — ✅ —
glob_search — ✅ ✅ ✅ — — — ✅ —
git_operations — ✅ — ✅ — — — — —
web_search ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
web_fetch ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
browser — — ✅ ✅ — — — — —
screenshot — — ✅ ✅ — — — — —
memory_store ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
memory_recall ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
http_request — — — ✅ ✅ — ✅ ✅ —
pdf_read — — — ✅ — — — ✅ —
proposal_gen — — — — ✅ ✅ ✅ — —
delegate ✅ — — — — — — — —

Sentinel-Only Tools

Sentinel has exclusive access to orchestration tools:

  • delegate: Spawn a subagent with a task
  • memory_store / memory_recall: Persist and retrieve cross-session knowledge
  • web_search / web_fetch: Quick lookups when classifying tasks
  • file_read: Read workspace state for context building

Sentinel does NOT have: shell, file_write, file_edit, claude_code, browser. It delegates execution, never executes directly.

Technology Stack (Enforced by Forge & Prism)

When building, agents must use these defaults:

Category Technology
Bots Python (asyncio, httpx, pydantic)
Frontend React + Vite.js or Next.js, TypeScript, Tailwind CSS, shadcn/ui
Backend (speed) Rust (axum, tokio, serde)
Backend (rapid) Next.js API routes or Python FastAPI
Payment Stripe + Baray.io (AES-256-CBC encrypted)
Authentication KOOMPI ID OAuth 2.0 (@koompi/oauth SDK)
Database PostgreSQL (primary), SQLite (embedded), Redis (cache)

Cost-Aware Model Routing

Each persona picks the cheapest model that can do the job:

Persona Claude Code Model Rationale
Archon claude-haiku-4-5 Plan/review only — no code generation needed
Prism Default (Sonnet) Full implementation of frontend components
Forge claude-sonnet-4-6 Heavy implementation; escalate to Opus for complex tasks
Oracle claude-haiku-4-5 Read/search only — no code generation
Others No Claude Code access Business/research personas don't code

Cost savings: By routing plan/review to Haiku (~10x cheaper than Opus), a typical multi-agent workflow costs 60-80% less than running everything on Opus.

Runtime override: The model parameter on claude_code tool allows Sentinel to override at runtime when a task is harder than expected.

Tool Safety Rules

  1. Shell: Only Forge has shell access. All shell commands are logged.
  2. File writes: Agents cannot write outside the user's workspace directory.
  3. HTTP requests: Rate-limited. No requests to internal/private IPs without explicit config.
  4. Browser: Sandboxed. No credential autofill. No form submission without user approval.
  5. Memory: No secrets stored in memory. PII detection is active.
  6. Claude Code: Runs in --print mode. --dangerously-skip-permissions only in sandboxed environments.