Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

==== ### ====
* Simplicity and clarity > cleverness.
* Robustness over attempting to salvage broken states (aligned with nuke & repave).
* Small, explicit state machines; avoid hidden implicit behavior.
* Strong invariants (e.g., “slot ownership implies exactly one active request”).

===== - Unit tests for: - state transitions - slot accounting under concurrency - loop detector - tool-call parsing + resume - BIOS assembly ordering =====
* Integration tests using a fake/stub HTTP server to simulate: - slow headers, slow streams, stalls, disconnects - malformed tool calls - restart conditions and crash loops

===== - Use bounded buffers everywhere (partial output, logs, tool traces). =====
* Liveness probing interval should be configurable and not too frequent (avoid tight polling).
* Avoid heavy dependencies; prefer stdlib + one HTTP client.
* Avoid per-token processing that scales badly: - loop detector should operate on completed lines and minimal state - avoid storing whole outputs unless explicitly requested.

If you want next, we can turn this into a concrete module outline (files/classes/interfaces + the exact public method signatures + enums + config dataclasses), then code against that with tests driving the implementation.