[TRACKING] Feature roadmap

[GeigerJ2]

Status: ✅ in use · 🔄 in progress · ⬜ to adopt

Feature we want	Library	Status
Agent loop, tool calling, structured output, streaming	pydantic-ai	✅
Provider/model switching (cloud + local Ollama)	pydantic-ai models	✅
Tool server (MCP)	fastmcp	✅
Config from env / .env	pydantic-settings	✅ (#10)
Recover from failing tool calls	pydantic-ai (ModelRetry)	✅ (#11)
CLI commands, flags, subcommands (entry point)	click (reuse aiida.cmdline options/param types)	⬜ currently a raw input() loop
HITL / approval gate before writes	pydantic-ai built-in tool approval (deferred tools)	🔄 PR #8 is custom; move onto the primitive
Multi-line input + vim keybindings + history	prompt_toolkit (PromptSession(multiline=True, vi_mode=True))	⬜ replaces raw input()
Spinner, live output, Markdown rendering	rich (console.status, Live)	⬜ replaces the local custom spinner
Eval harness (golden queries, scoring, reports)	pydantic-evals (Dataset/Case/LLMJudge)	⬜ ADR-06; don't hand-roll
Tracing / per-run tokens + latency	Pydantic Logfire (OTel)	⬜ optional, high-leverage for cloud-vs-local study
Full TUI (only if we outgrow a REPL)	Textual	⬜ later, maybe never