AI Manager Core

Provider-neutral and project-neutral control-plane primitives for AI-assisted operations.

The package owns strict contracts, canonical identities, authority intersection, workflow transitions, receipts and provider/effect ports. Domain repositories continue to own their risk, compliance, credentials and execution policy.

Safety boundary

AI Manager Core can describe and validate trading, credential and system-update operations. It does not contain an exchange client, secret store, updater, publisher or production executor.

Operator commands follow this sequence:

structured command -> effect plan -> hash-bound approval -> broker
                   -> execution receipt -> independent readback

Free-form model output is never an approval.

Condor reference

The architecture adapts useful patterns from Hummingbot Condor:

• separate reasoning, operator handlers, routines and deterministic execution;
• typed one-shot and continuous routines;
• provider selection behind a stable agent boundary;
• operator role checks and confirmation challenges;
• direct Telegram trading, credential and update workflows.

AI Manager Core adds exact scope intersection, immutable receipts, independent readback and adapter-local authority. Condor is a reference, not a runtime dependency.

Development

python3.12 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/pytest
.venv/bin/ruff check .
.venv/bin/mypy src

Python 3.12 or newer is required.

Cross-repository conformance

conformance/ai-manager-core-v1.json is the portable golden-vector bundle for FlowMate and NOVA. Consumers must reproduce the exact canonical JSON bytes and SHA-256 digests before an adapter can claim compatibility. schema_bundle() and build_schema_manifest() expose deterministic contract descriptors; a manifest mismatch is fail-closed.