AgentOS

Open-source, human-supervised workflow-integrity layer for the Eijex FactorForge and ValidationHub pipeline.

The repository contains software, schemas, policy checks, and synthetic fixtures only. It does not contain partner records, proprietary sequences, credentials, or permissioned wet-lab evidence.

---

Role in 4-Repo Architecture

factorforge-cds  (open source) ← CDS design engine (API, CLI, profiles)
      │
      ▼
ValidationHub    (open source) ← Evidence-registry code and public schemas
      │
      ▼
AgentOS          (open source) ← THIS REPO: workflow-integrity layer
      │                        intake → boundary check → claim review → public projection
      ▼
eijex-mcp        (open source) ← MCP integration surface

eijex-agentOps is the operations layer, not a data store. It:

• Processes incoming sequence optimization requests
• Enforces public/private data boundaries
• Routes wet-lab results through human-in-the-loop review
• Generates public summaries only after approval

Permissioned data records live outside Git. ValidationHub defines their registry boundary, while public tools live in eijex-mcp.

---

Repo Structure

eijex-agentOps/
├── intake/              ← Incoming request processing
├── boundary/            ← Public/private data boundary enforcement
├── claim_review/        ← Wet-lab claim review & public projection
├── server/              ← FastAPI layer (api.py, sequence_registry, sanitizer)
├── tools/               ← ValidationHub writer, FactorForge runner adapter, GenBank exporter, mock agent harness
├── evals/               ← Deterministic guardrail eval fixtures + promptfoo configs
├── skills/              ← AgentOps skill pack (claim-boundary, no-raw-sequence, release-gate, trust-pack)
├── replays/             ← Synthetic agent run replays
├── schemas/             ← Shared refs (points to validationHub as source of truth)
└── scripts/             ← Operational utilities

---

Security Boundaries

Data type	Allowed here	Must stay in
Raw protein sequences	❌ Never	permissioned external storage
Private CDS outputs	❌ Never	permissioned external storage
ValidationHub intake records	read-only ref	validationHub
Public summaries (post-approval)	✅ Output only	factorforge-cds VALIDATION.md
Agent workflow logic	✅ Core	this repo

Human approval required before any external send or public summary generation.

---

Status

Component	Status
Repo scaffold	✅ v0.1
intake/ workflow	✅ v0.1 (parse_intake.py, schema-validated)
boundary/ checker	✅ (check_boundary.py, prohibited_fields.yaml, redaction.py)
claim_review/ pipeline	✅ (review_claims.py + approval_gate.py, human approval enforced)
server/ FastAPI layer	✅ v0.8 (8 endpoints, sequence_registry, sanitizer)
ValidationHub physical write	✅ v0.6 (tools/validationhub_writer.py)
FactorForge runner adapter	✅ (tools/factorforge_runner.py)
GenBank exporter	✅ v3.2.1 (tools/genbank_exporter.py)
Agent guardrail evals	✅ (evals/, deterministic fixtures)
AgentOps skill pack	✅ (skills/, policy evaluation corpus)
eijex-mcp integration	✅ Requires a separately running AgentOS service

---

Quick start

python -m pip install -r requirements.txt
python -m pip install pytest
python -m pytest -q
uvicorn server.api:app --host 127.0.0.1 --port 8787

Copy .env.example values into your local environment and replace every placeholder. Never commit real tokens, database URLs, sequences, or evidence records.

Open-source and data boundary

The code is licensed under GNU AGPL v3.0. Opening the software does not grant permission to publish third-party research records. Public projection remains allowlisted, consent-aware, and human-approved.