harnessweaver

The core of HarnessWeaver, the composable AI harness framework: the spec loader, the 13-point hook bus, envelope and budgets, the error taxonomy, the runners, the built-in modules and the test kit. It depends only on pydantic, opentelemetry-api, pyyaml, anyio and loguru.

uv add harnessweaver                 # the core
uv add "harnessweaver[openai]"       # plus a first-party plugin
uv add "harnessweaver[all]"          # every first-party plugin, the CLI and the sidecar

import harnessweaver as hw

harness = hw.load("harness.yaml")        # spec + org baseline + env overlay, validated at startup

# Wrap: govern a client the app already has
client = harness.instrument(AsyncOpenAI(base_url=GATEWAY))

# Session: your code owns the flow, the harness governs each step
async with harness.session("support", principal=user, tenant="acme") as s:
    route = await s.decide("intent", text=question)
    articles = await s.context("help").fetch(question)
    answer = await s.generate(prompt="answer@1", question=question, intent=route.choice)

# Run: the harness owns the lifecycle
result = await harness.run("extract_invoice", doc, principal=user)

The test kit is harnessweaver.testing, and installing the package adds the plugin conformance suite to pytest (pytest --harnessweaver-conformance=<namespace>).

Documentation: the docs, from Concepts to the spec reference.