Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
# A TEST-ONLY overlay for the scripted Strata worker, layered over the product profile
# (`cordis.patch.yml`) by `tool_flow_e2e.py --patch` order. It is not part of the shipped profile and
# nothing but that test applies it: the product profile must keep answering title requests.
#
# It exists because a *scripted* model is not a model. The worker replays a fixed list of scripts, one
# per request, so which reply a request receives depends on how many requests came before it -- and a
# sequence written by hand is only meaningful if the Harness sends exactly the requests the script
# author planned for.
#
# The session-title provider does not: on the first prompt it fires its own chat-completions request,
# before the turn's model call. The worker answers that request with the first script, so a
# two-script sequence meaning "call the bash tool, then report" is consumed as "title, report": the
# tool call is never emitted, the turn completes normally, and the run looks exactly like a model that
# chose not to use a tool. That is a silent, plausible-looking failure -- the worst kind -- and it is
# why this overlay disables the provider rather than adding a script.
#
# The deterministic first-prompt fallback title (`session-title`) stays enabled, so sessions are still
# named from their first message; only the extra model request is gone.
- id: session-title-llm
disabled: true