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

=== Appendix L: Test plan and stub fixtures ===

==== L.1 Unit test matrix ====
* Prompting ordering and BIOS regeneration triggers.
* Repeated-line detector triggers on ~39-char lines under the new default thresholds.
* Slot accounting: no leaks under completion/failure/cancel/restart.
* Tool loop: - multiple tool calls in one iteration consume 1 budget, - exit tools recorded and non-terminating, - “no results” tool outputs are success, - fallback parsing success/failure paths.
* Result semantics: - get_result() returns partial output on failure/cancel/restart, - get_result() releases records; subsequent lookup is NOT_FOUND.

==== L.2 Integration tests with stub “fake llama-server” ====

Stub server supports:
* /v1/models ready/not-ready/flaky.
* /v1/chat/completions streaming: - text deltas, - structured tool calls (batch), - exit tool calls, - stalls, disconnects, errors.

Scenarios:
* slow headers → headers timeout paths,
* long silence with liveness updates (where applicable),
* mid-stream stall → idle stream timeout,
* disconnect mid-stream → request fails with partial output,
* tool iteration with multiple calls,
* exit tool emission recorded,
* crash-loop lockout behavior.

==== L.3 No-orphans test ====

Spawn a dummy process tree; assert stop/restart kills the entire process group.