helm-ai 0.2.2

AI agent orchestration and token-optimisation system

Helm

AI agent orchestration and token-optimisation system.

A routing layer dispatches tasks to the cheapest capable worker — local Ollama, MCP expert, Playwright, or Claude Orchestrator — based on confidence scoring. A SENTINEL knowledge file accumulates lessons per-pod so routing improves over time. At the org level, a CEO orchestrator aggregates across pods into exec summaries and a risk register.

---

Stack

• Python 3.11+ — core orchestration package (helm/)
• Pydantic v2 — all schema validation
• TypeScript — MCP bridge (Sprint 3B)
• HTML — report output (Sprint 3B)

---

Project Structure

helm/                   # Python package
tests/
  fixtures/             # Canonical test fixtures — never modified during test runs
  mocks/
    claude_responses/   # Mock Claude API responses for unit tests
    ollama_responses/   # Mock Ollama responses for unit tests
docs/
  tasks/
    worker.md           # helm-builder task queue
    mcp.md              # mcp-builder task queue
  messages/             # Async inter-agent comms
  tracking/
    sprint.md           # Sprint board (source of truth)
  decisions/            # Architecture decision records
  lessons-learned.md    # Known failure patterns — read before every task
  guidelines/
    workflow.md         # Task sizing and completion protocol
CLAUDE.md               # Project rules — read every session via /go
AGENT_SYSTEM.md         # Operating manual for multi-agent coordination

---

Development Setup

Requires Python 3.11+. The codebase uses datetime.UTC and other 3.11 stdlib additions throughout. Running on an older Python will fail at import time.

Mac (first time)

brew install python@3.11          # install if not present
python3.11 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pip install fastapi "uvicorn[standard]" httpx
pytest --co -q                    # verify test discovery

Windows / home PC

# Install Python 3.11+ from python.org or winget install Python.Python.3.11
python -m venv .venv
.venv\Scripts\activate
pip install -e ".[dev]"
pip install fastapi "uvicorn[standard]" httpx
pytest --co -q

Cross-machine notes

• .python-version (repo root) is set to 3.11 — pyenv and most tooling will respect it.
• pyproject.toml enforces requires-python = ">=3.11" — pip install will fail loudly on older Pythons rather than silently producing broken installs.
• The .[dev] extras install pytest, pytest-timeout, and pytest-cov. FastAPI, uvicorn, and httpx are runtime/test dependencies for the web layer but not yet in [dev] extras (pending S25 dep review).

Run the test suite

pytest                             # all tests
pytest -m sprint25_gate -v        # current sprint gate only
pytest --co -q                    # discovery check (no execution)

---

Sprint Plan

Sprint	Focus	Gate
1	Foundation: schemas, briefing, routing, SENTINEL	37 tests green
2	Integration: Ollama worker, circuit breaker, pool queue, quota monitor	GP-001/002/003 pass
3A	Optimisation: real confidence scoring, shadow testing	Shadow tester live
3B	Report generator + MCP bridge	HTML output from real data
4	Hardening: security audit, 100-task baseline	≥74% token savings
5	Org layer: CEO orchestrator, Org-SENTINEL	GP-004 passes
6	Org hardening + optional live viewer	200-task baseline ≥74%

See docs/tracking/sprint.md for current status.

---

Contributing

Read CLAUDE.md and AGENT_SYSTEM.md before contributing. Branch from develop. PR to develop. Managing Director reviews all merges.