#112 Disable local llama-server probing while a cloud text provider is active

[Anonymous]

Originally created by: yablokolabs

Summary

A cloud-backed Atomic session can start successfully while still probing and warning about http://127.0.0.1:8080. These requests look like an active backend failure even when the selected OpenAI-compatible provider is healthy.

This is finding §9 from Yabloko Labs' evaluation. The behavior spans bootstrap, sidecar session start, per-turn profile refresh, and the TUI health poller.

Current behavior

Runtime bootstrap checks local llama health and resolves a local profile before wiring the active cloud provider:

• bootstrap.ts (github.com)

The sidecar performs another unconditional health check after buildRuntime:

• sidecar/main.ts (github.com)

The agent loop refreshes the local profile manager at turn start and between steps, causing /props traffic even when inference uses a cloud provider:

• agent-loop.ts (github.com)
• agent-loop.ts (github.com)

The TUI constructs and always starts a poller against the local llama URL:

• chat-orchestrator.ts (github.com)

Proposed behavior

Resolve the active text provider before initializing local text-model health/profile state.

When the active text provider is cloud/OpenAI-compatible:

• skip local text /health and /props requests in bootstrap and sidecar startup;
• do not refresh a local ModelProfileManager during cloud turns;
• use an explicit cloud/non-local prompt profile;
• stop the TUI local poller, or label it clearly as inactive local-backend status without probing.

Local embeddings are independent and should still be probed when explicitly enabled. Switching back to local-llama should lazily initialize and refresh its profile, context, and slot state before the first local turn.

Acceptance criteria

• A cloud-only text configuration makes zero requests to the local text /health or /props endpoints during runtime bootstrap, sidecar session start, and cloud turns.
• No local-text-backend failure is emitted for a healthy cloud-backed run.
• The TUI does not present an inactive local backend as the health of the active cloud provider; it either does not poll or labels local status as inactive.
• Cloud turns use a non-local/plain profile and never refresh the local profile manager.
• Selecting local-llama preserves or lazily restores health, profile, context-window, and slot probes before local inference.
• Explicitly enabled local embeddings continue to be probed independently.
• Tests assert local request counts for CLI bootstrap, sidecar start, TUI startup, cloud turns, local turns, cloud-text-plus-local-embeddings, and cloud-to-local hot switching.