First session (no API keys)

The fastest taste of apomesh needs no API keys, no network, and not even the running stack. You replay a cassette — a recorded session’s full event stream — through the SDK’s in-process daemon stub, and watch the SDK reconstruct the session end to end. It’s fully deterministic, so it’s the same every time.

Prerequisites

• Python 3.11+ and uv.

You do not need the Docker stack for this step, and you do not need any provider credentials.

Install the SDK

The SDK isn’t published to a package index yet, so install it from the repo. From the repo root:

cd sdks/python
uv sync --all-extras

uv run commands operate inside the SDK’s virtual environment automatically. This is the same install you’ll use again for the SDK dispatch step.

Replay the recorded sessions

The SDK ships recorded deep-research sessions as cassettes and a replay test suite that drives each one through an in-process stub daemon:

uv run pytest tests/integration/

To replay just the driving demo — a research goal in, a structured report out:

uv run pytest tests/integration/test_demo_driving.py

What just happened

Each cassette holds a real deep-research session’s complete envelope stream — the lifecycle events a live run would emit. The replay drives that stream through a stub daemon in the same process, so the SDK sees exactly what it would see from a real orchestrator, without any LLM calls.

The suite exercises the full deep-research shape from the driver’s side:

• The driving replay reconstructs the session’s lifecycle (planner, workers, validators, synthesizer) and decodes the terminal event into a typed DeepResearchReport — title, sections, claims with citations, and a summary.
• The verified-citation replay observes the recovery path when a worker fabricates a citation and the session retries with a revision.
• The coverage-replan replay observes the session re-spawning weak workers when a coverage check flags a gap.

A passing run means the SDK reconstructed each recorded session and its structured report correctly — a completed session, watched with no keys and no network.

Note

This replays recorded events; no live model runs. To dispatch a live session against a real LLM provider, you need the running stack and a credential — that’s the next step.

Next

Run a real-provider session.