Mega Context

Nanochat CLI TUI Status

2 min read

Nanochat CLI TUI Status

Tracks

  • Repo wiring: package scaffold, entrypoint, add Textual/PyYAML deps.
  • Config: prefab load/save/override, diff skeleton.
  • Checkpoints: scan + basic compatibility heuristics.
  • Train/Eval orchestration: subprocess launch + log streaming.
  • Setup flow: dependency detection + guided install inside TUI (no reliance on mc_setup/run*.sh). (basic uv/torch/pyarrow + path inputs wired; needs richer UX/validation)
  • Dataset tab: dataset download/prep orchestration. (now respects base_data paths, defers deps to Setup; needs progress + modal redirects)
  • Visualization panels: rich screens (loss, GPU, timers, MC LOD) with refresh cadences.
  • Command palette + keyboard UX for core actions.
  • Live screen editing + persistence.
  • Plugin/extension API: config schema/panels/telemetry hooks for MC and other projects (initial plugin registry landed).
  • Telemetry ingestion: WANDB streaming, OTEL spans, log tail parsing.
  • Mid-run config patching + checkpoint-now signaling.
  • Robust compatibility checks (checkpoint metadata, manual override).
  • UI polish: alerts/toasts, status bar details, color/contrast audit.
  • Resilience: retries/backoff, crash recovery of UI state.
  • Packaging/bootstrap: minimal install path on fresh boxes.
  • Feedback round 1 (current):
    • Config UX: dropdown selector, categorized sub-tabs/scroll, inline editors (no raw JSON/YAML), dirty tracking, save/save-as with reload guard. (dropdown + sub-tabs + binding hints + reload guard added; polish/layout still pending)
    • Rename “prefabs” → “configs”; display values (not raw text); auto-apply edits, save writes files.
    • Checkpoints: refresh via action; moved logic into checkpoints component.
    • Train: contextual start; no config display leakage.
    • Eval: renamed to Eval; continuation only.
  • Dataset tab before Train; detect/download/prep dataset; train blocks when missing with modal redirect. (gating present; missing deps now push to Setup; modal still TODO)
    • Setup first tab; no mc_setup dependency; self-managed setup; “Run” action; re-check via action not button. (env inputs + pyarrow install + path propagation; needs better UX/modal)
    • Tab order: Setup → Config → Dataset → Checkpoints → Train → Eval.

Notes

  • Current build is a functional skeleton; prioritizing setup/dataset flows and panel/telemetry next.

Graph View