zero-network-after-setup.md

Zero Network After Setup

BaseMyAI's network rule is intentionally narrow:

• setup may fetch model files only with explicit consent;
• the embedder receives a local model path and never downloads;
• remember, recall, graph traversal, GC, and forgetting are local database operations;
• MCP sampling or user-configured LLM backends may use a network because the operator explicitly chose that integration.

What Is Already Tested

The provisioning tests cover the most important invariant: when the model is absent and consent is false, setup fails instead of silently downloading.

cargo test -p basemyai --features test-util provision_without_consent_fails_when_model_absent

Manual Offline Proof

After a successful setup:

basemyai setup --fetch

run remember and recall with the machine disconnected or with an invalid proxy:

export HTTPS_PROXY=http://127.0.0.1:9
export HTTP_PROXY=http://127.0.0.1:9
export BASEMYAI_DB_KEY='dev-proof-key'

basemyai init ./offline-proof.bmai
basemyai remember ./offline-proof.bmai --agent-id offline-agent --text "offline memory works"
basemyai recall ./offline-proof.bmai --agent-id offline-agent --query "offline memory"

Expected result: the memory roundtrip succeeds. If the model path is missing, the command must fail with setup guidance rather than fetching anything silently.

CI Hardening Still To Add

A future CI job should pre-provision a tiny local test model or use the test-util deterministic embedder, deny outbound sockets at the process level, and execute a full remember/recall roundtrip.

That job should be named zero-network-after-setup and linked from the README once it exists.