veille-supervisor 0.3.0

Control plane for production AI-agent work: plan, govern, and verify agent runs.

AI Runtime Supervisor / Veille Console

Control plane for production AI-agent work. The Supervisor plans, contextualizes, routes, governs, and verifies each agent run—reducing wasted spend and unreliable outcomes without requiring teams to rebuild their applications.

Current release: 0.2.0 — Phases 1–5 + Local Integration Console. Event schema 0.2.0. All capabilities are opt-in and off by default.

What exists today

• Versioned data contracts (task, events, plan, policy, validation)
• Python SDK (Supervisor + RunCollector) for zero-touch event emission
• LangGraph adapter with automatic callback-based instrumentation
• OpenAI Agents SDK adapter and OpenAI Responses API adapter (skeleton)
• Model provider port with 8 provider drivers (LiteLLM, OpenAI, Anthropic, Gemini, OpenRouter, Ollama, LM Studio, OpenAI-compatible)
• Synthetic cited market-research LangGraph demo workflow with mock tools
• Representative trace fixtures for success, expensive, and failed-validation runs
• Observe-only policy engine (duplicate detection, retry budget, cost overrun, validation)
• Opt-in enforcement (Phase 2): block / retry / pause / stop with full audit trail
• Budget tracking (BudgetTracker, in-memory default, Redis backend port)
• Run-explorer CLI and OpenTelemetry export (Console + OTLP)
• Advisory planning (Phase 3): tier selection, per-step context manifests, capability+tier model routing
• Adaptive optimization (Phase 4): semantic near-duplicate detection + idempotent result caching (SUPERVISOR_OPTIMIZE, dry-run default)
• Memory governance (Phase 5): memory store + scoring + governor (memory.retrieved/memory.expired manifest) with audited expiry, no automatic deletion (SUPERVISOR_MEMORY)
• Local Integration Console (veille CLI + FastAPI + React web UI) — register workflows, connect providers, run live, inspect execution
• Local development environment (Docker Compose + pytest)

Quickstart

Prerequisites

• Python 3.12+
• Docker Desktop (optional, for Postgres/Redis/MinIO scaffold)

Setup (Windows)

.\scripts\dev.ps1
.\.venv\Scripts\Activate.ps1

Run tests

pytest -v

Use the console

# Doctor — environment + safe-config report
veille doctor

# List provider connections
veille connections
veille connections validate openai

# List registered workflows
veille workflows

# Run the mock demo
veille demo mock

# Run the real-world demo
veille demo real-world

# List saved runs
veille runs

# Run an arbitrary registered workflow
veille run cited_market_research --input '{"scenario":"success"}'

# Start the web UI (then open http://127.0.0.1:8000)
veille serve

Run the classic demo workflow

# Successful run
python -m examples.cited_market_research.agent --scenario success

# All scenarios + write trace fixtures
python -m examples.cited_market_research.agent --scenario all --write-fixtures

# Phase 1 run explorer (inspect a captured run)
python -m supervisor.cli explore --run fixtures/traces/expensive_run.json --policy

# Live run with policy observations and OTel export
python -m supervisor.cli explore --live --scenario expensive --policy --otel

No API keys required. Mock models and tools are used by default.

Repository layout

src/supervisor/          Core contracts, SDK, adapters, analytics, policy, telemetry, CLI, console
examples/                Runnable demo workflows
fixtures/traces/         Synthetic trace JSON for tests and baselines
docs/                    Architecture, contracts, roadmap, ADRs
docs/development/        Source master prompt, blueprint, phase plans
ui/                      React+TypeScript+Vite web UI (veille console frontend)
templates/               Baseline measurement templates

Documentation

Document	Description
Architecture	System components and boundaries
Runtime chain	Stage-by-stage runtime flow
Data contracts	Schema reference
Roadmap	Phase status and deferrals
Integrations	Adapter contracts + provider drivers
Operations	Local dev, commands, runbooks
Policy engine	Policy modes and Phase 1 observe policies
ADR-013	Local Integration Console design

Setup guides

Guide	Description
Safe local setup	Running with mock providers (default)
Mock demo walkthrough	End-to-end mock demo
Real provider setup	Setting up real model providers
OpenRouter integration	Using OpenRouter as a gateway
OpenAI Agents SDK	Running an OpenAI Agents SDK workflow
LiteLLM integration	Using LiteLLM for multi-provider access

License

MIT