UI: send-on-Enter + live reasoning trace; docs: answer-cache decision (#106)

cleanup: remove the translation module — the guard is English-only now (#105)

retrieval: remove the cross-encoder reranker (ablation showed no earned keep) (#100)

perf: hydrate retrieved sections concurrently — cut the dominant answer latency (#96)

Follow client-side redirects + read_page URL guard: real API content (#92)

Make the deploy usable: strip/redact API keys + gate non-English + faster QuOTE (#91)

Concise answers + immediate thinking indicator (Ultralytics-style UX) (#90)

Route questions: grounded first, loop for multi-source shapes (#89)

Answer-quality eval (LLM-judge) + hy3-led OpenRouter default (#81)

Harden the OpenRouter path: no-choices 200s retry, dead deepseek slug removed

Let the planner choose the content space: search_docs gains a kind filter

Guard: calibrated threshold (0.35) + fail-open on untranslated questions

Guard: replace the injection classifier with embedding-space membership

Guard: multilingual injection classifier first, open model as fallback

App backpressure and cleanups

Share a circuit breaker across LLM calls; cap Retry-After sleeps

Harden the input guard for production

Merge input guard branch (injection + topicality checks)

Serve questions concurrently instead of one-at-a-time (#49)

Add input guard: prompt-injection + topicality check on the user's question

Codebase quality pass: kill silent-truncation hacks, fix latent bugs, cover them (#46)

OpenRouter free-tier resilience + GEMINI alias + live HF Space smoke test (#41)

Self-heal LLM: fall back across providers when one is unreachable (#40)

Fix LLM transport drift, wire live static checks, clean config (#39)

Log base_url + full exception for openai-compat failures

Add diagnostics: log retrieve/hydrate counts + live-fetch failures

Loop: always retrieve once before the planner (robust to flaky free models)

Harden provider resolution for deploys: auto-select + default OpenRouter base_url

M3.4: LangGraph rewrite of the agent + loop-vs-langgraph comparison

M3: three tools + manual agent loop (planner-driven, budgeted)

openai-compat: fall back across a chain of free models on rate-limit

Translation must never crash the search: retry 429s, fall back on any error

Add non-English query translation (via OpenRouter) before retrieval

M2 wiring: grounded answers with validated citations + grounded eval mode

Eval workflow: model as input (free slugs rotate); fail fast on 404 model-gone

openai-compat: fall back to prompt-only JSON when host rejects response_format

Default openai-compat model: DeepSeek-V4-Flash ($0.09/$0.18 per M on DeepInfra)

Add openai-compat provider path: cheap open models on US/EU hosts

Wait out the rate-limit window on Anthropic 429s, matching the Gemini path

Make Gemini the default generation provider (free tier)

M1: Answer schema, LLM wrapper with retry+repair, static checks, 15-question eval

Scaffold repo structure per M0 (pyproject, ruff, pytest, pre-commit)