feat: constrain agent questions to user-friendly abstraction level (#13)

feat: add model selector dropdown to UI

feat: add severity field to GeneratedQuestionCard

fix: pass max_tokens to LLM, add debug log for JSON parse fallback

feat: replace keyword gap analyzer with LLM-based analyze_gaps

refactor: update CrewOrchestrator to return ChatTurnResponse

refactor: update BaseOrchestrator and MockChatBackend to return ChatTurnResponse

refactor: type ContextVar design state as DesignState

refactor: type analyze_gaps and extract_decisions with AgentResponse

feat: add PreviewData and ChatTurnResponse pydantic models

refactor: convert AgentDef from dataclass to Pydantic BaseModel

refactor: add ToolConfig pydantic model, replace tool_config dicts

feat: add memory recall/remember, planning, and collaboration to Flow

feat: remove raw history, add Memory to CrewOrchestrator

refactor: migrate tools to BaseTool subclasses with args_schema

refactor: extract derive_part_name, remove dead code from prompts.py

fix: address code review findings from Flow refactor

refactor: wire AgentDispatchFlow, remove RoutingEngine

feat: add AgentDispatchFlow with router-gated agent dispatch

feat: add agent_flow models, routing, and readiness helpers

fix: ask agent confirm now correctly updates plan card fields

feat: add notes field to DesignPlan and field_agents config

fix: include geometry stats in CrewOrchestrator preview to prevent toFixed crash

fix: prevent question card loop when agent mentions keywords as context

feat: integrate gap analyzer into orchestrator, update CNC/CAM NOT READY prompts

feat: add gap analyzer with analyze_gaps() and generate_question_cards()

fix: path traversal in model endpoints, XSS in plan card, remove dead code

feat: include threemf_url in preview response

feat: add planning phase branching to CrewOrchestrator

feat: add DesignPlan model, compute_score(), phase/plan fields to DesignState

refactor: eliminate all Serializer classes, use Pydantic model_dump() directly

refactor: typed Pydantic config + Pydantic result models, eliminate dict access

fix: remove unused AGENTS import from orchestrator

refactor: remove SingleCallOrchestrator, unify all backends through CrewAI

fix: remove unused json import from crew_orchestrator

feat(crewai): rewrite orchestrator with native LLM, tools, structured output

feat(tools): add generate_gcode_tool, store last shape on execution

feat(cam): integrate CAM generation into orchestrator pipeline

fix: increase max_tokens to 8192, improve retry for truncated code

feat: add confirm/revise buttons to agent recommendations

fix: surface LLM backend errors to user instead of silent fallback

feat: agents build on each other's responses in sequential chain

feat: agents clarify before acting, CAD agent gates on sufficient specs

feat: add CrewAI CadQuery tools — execute_cad_tool, validate_cad_tool

feat: complete LLM adapter — enable function calling, add vision support

refactor: CrewOrchestrator inherits BaseOrchestrator, uses RoutingEngine + BackendFactory + settings

refactor: orchestrators inherit BaseOrchestrator, use RoutingEngine

refactor: move extract_decisions into DesignState.update_from_messages, config-drive domain data

refactor: load agent definitions from config.yaml

feat: add BaseOrchestrator ABC