feat(tools): implement Pattern A - Dynamic Enum Injection for LLM tool constraints

refactor(redis): split cache key constants and align chat/model caching docs

feat: align think-mode semantics, harden background processors, and stabilize streaming chat UX

fix(qwen): resolve Responses API schema errors and tool parameter type coercion

perf(chat): optimize history context retrieval with Redis cache

feat(llm): add finishReason support to MistralAiAdapter for Auto-PASS

refactor(agent): implement strict dual-model orchestration pipeline

refactor: split agent-loop service and enhance chat/LLM rate limiting

feat(api): add DeepSeek adapter and password login; fix CI deploy workflow

refactor(api) & feat(web): architecture review fixes and UI enhancements

feat(storage): add file_assets tracking with PENDING lifecycle and chat message binding

refactor(web): restructure chat runtime and harden scroll pagination/viewport stability

fix(chat-scroll): eliminate 300ms push-up delay, fix blank space on new session, and fix keyboard content jump

refactor(web): reduce eslint disables and split speech/input helpers

feat(web): add voice conversation runtime and onnx voiceprint verification flow