Switch the default chat model to DeepSeek V4 Flash via OpenRouter

Rescue account-state provider errors and page ops on provider incidents

Replace the client-level fallback with a turn-level rescue

Refresh MODEL_PRICING against current provider price sheets

Upgrade backend dependencies; adapt to langchain 1.3.15 and ruff 0.16

Archive private KB bundles for fast cold starts

Tag LangSmith traces by deployment environment

Remove Claude Haiku 4.5 from the selectable model list

Improve KB command ergonomics, tool-output streaming, and wiki linting

Improve retrieval streaming and tracing

Add model-aware prod v2 memory preset

fix: preserve DeepSeek thinking events

feat(experiments): DeepSeek stage-1 compaction arms + prefix-preserving summarization

fix(chat): keep one thread on one provider, drop the unselectable Gemini

fix(retrieval): store the BM25 index as gzipped JSON instead of pickle (#13)

Use direct DeepSeek as default model

Use OpenRouter DeepSeek with Gemini fallback

fix(activity): repeated Gemini search queries start a fresh cycle

fix(graphrag): query-relevant community reports, rerank floor parity

fix(retrieval): budget skip, standard RRF, fail-fast collection load

fix(config): restore default dotenv precedence (shell wins over .env)

fix(config): honor AI_TUTOR_KB_DIR for every KB path

fix(citations): keep labels of links with balanced brackets

perf(citations): cache the KB manifest indexes

fix(api): close request-size gaps in body and history caps

fix(api): emit stream error frame after cleanup parts

fix(citations): handle malformed links; count only consulted docs in activity (#12)

Public docs-only bundle: allowlist filtering, leak audit, kb.tar.gz cold start (#11)

wip: eval harness + public docs bundle; relocate eval docs into evals/ (#10)

fix(citations): drop code-block URLs from chips; default URL reading off

feat(prompts): add course content reproduction policy and related tests

Merge experiment #3: SLM knowledge compaction (Ollama small models, 32k window)

Merge experiment #5: context-compaction study (keep-all vs compaction vs retrieve, Gemini 2.5 Flash)

Merge experiment #2: GraphRAG vs classical RAG (opt-in retriever + scoped comparison)

feat(evals): single-turn latency finding (F27) + de-confound latency tooling

evals(slm): Axis-A memory-compaction arm on local SLMs

memory: add hierarchical and delta summarization presets (Axis A)

compaction study: harness for compaction-vs-keep-everything on a long lesson

graphrag: publish the index to HF for reproducibility (prod cold-start skips it)

style: apply ruff formatting to telemetry and report

Refactor telemetry pricing and evaluation reports

graphrag: validate retriever vs real index, fix bugs, scope by source

Add OpenRouter provider for open-model evals

graphrag: scaffold true GraphRAG retriever arm (opt-in, default unchanged)

Refine v2 eval plan and memory harness

Add Part C eval variants, LLM judge, and harness fixes

Add evaluation harness for memory/context-management experiments

Enhance knowledge base and API functionality

Enhance API validation and error handling

Enhance model validation and tool call management