~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
These offline tests measure actual raw chat-context construction while varying history length, long/short/cited responses and interleaved progress steps. The complete expected context is generated independently. Assertions gate latest findings, order, user text, progress filtering, fresh content and a linear stripping-work budget. Window limits retain their existing zero/negative behavior.
pdm run pytest tests/performance/chat_context -q \
--chatctx-profile local --chatctx-report /tmp/chat-context.json
CI uses 8/16 findings, two iterations after one warmup; local uses 250/500,
seven after two. --chatctx-{profile,size,iterations,warmups,report,max-p95-ms}
overrides LDR_PERF_CHATCTX_*, with shared LDR_PERF_PROFILE fallback.
Constructor/input preparation, assertions, cleanup and log-sink I/O are untimed.
The operation includes a counter around real citation stripping. There are no
LLM or network calls. Per-unit p95 is a scaled batch statistic, not item latency.
The automatic Performance Assertions workflow runs serial CI profiles, checks JUnit for actual passing tests in every selected suite and rejects missing, failed or skipped evidence. Reports retain raw timings and implementation hashes. Wall-clock gates are opt-in; correctness and work limits always gate CI.