FPF Agent Stack

Bounded Context: Offline AI Agent Runtime

1. Context

Problem. Cloud LLM APIs leak private data, incur per-token costs, and require network connectivity.

Solution. An embedded, offline agent runtime with auditable execution and tri-state safety guards.

Scope. Local development, private automation, air-gapped environments.

2. Intent (L-PROJ-*)

L-PROJ-01: Core Capabilities

Capability	Description
Embedded Model	Qwen 2.5 0.5B Instruct via Transformers.js — no API keys, no network
MCP Server	Exposes skills as tools via Model Context Protocol (Stdio)
AgentFS	Copy-on-write sandbox for safe, auditable file operations
Guard	Tri-state validation: pass / degrade / abstain
Skills	Dynamic loading from `core/skills/` with L/A/D/E contracts

L-PROJ-02: Project Structure (FPF E.4 Compliant)

fpf-agent-stack/ ├── core/ # CANON: Normative contracts │ └── skills/ # L/A/D/E skill definitions │ ├── file-ops/ │ └── repo-search/ ├── src/ # TOOLING: Implementation │ ├── cli.ts │ ├── mcp-server.ts │ ├── model-gateway.ts │ ├── guard.ts │ └── agent-fs.ts ├── test/ # TOOLING: Verification │ └── bdd/ │ └── features/ └── docs/ # PEDAGOGY: Rationale & tutorials └── *.md

3. Archetypal Grounding

U.System Example: AgentFS Session

# Start isolated session node dist/src/cli.js start-session ./my-project # All writes go to session layer, host untouched node dist/src/cli.js write <session> file.txt "content" # Audit trail available node dist/src/cli.js audit <session>

U.Episteme Example: Skill Contract

# core/skills/file-ops/SKILL.md ## L-TOOL-01: write_file Write text content to a file path within workspace. ## A-TOOL-01: Admissibility - Predicate: Active AgentFS session required - Evidence: session_id ## D-TOOL-01: Duties - Writes isolated to session layer - Text content only (no binary) ## E-TOOL-01: Evidence - Logged in AgentFS audit log

4. Getting Started

Prerequisites

Node.js 20+ (LTS)
8GB RAM minimum

Setup

npm install npm run build

Run

# Start MCP server npm start # Run tests npm test

5. Conformance Checklist

ID	Requirement	Verification
CC-PROJ-01	Runs without network	`npm start` succeeds offline
CC-PROJ-02	Model loads <30s	Manual timing on 8GB machine
CC-PROJ-03	File writes sandboxed	`npm test` — `@agentfs` green
CC-PROJ-04	Unknown tools abstain	`npm test` — `@guard` green