micasa.

Your --help was the ugliest screen in the app

Kong parsed my arguments for four months. It worked fine. It also didn’t generate shell completions, so micasa had none. And --help was plain monochrome text next to a TUI with a color palette – the one screen that looked like it belonged to a different application.

Kong out, Cobra in#

#785 replaces Kong with Cobra. The main reason was completions – Kong doesn’t generate them, so micasa never had tab completion. Cobra builds them from the command tree at runtime. micasa completion bash writes a script to stdout, same for zsh and fish. Add a subcommand next week, the completions already know about it.

While I was in there, I overrode Cobra’s help function to use the Wong palette. Subcommands in one color, flags in another, descriptions in the adaptive foreground. --help looks like the rest of the app now.

Side benefit: Kong’s CLI tests compiled the full binary and exec’d it in a subprocess, which took about ten seconds per run. Cobra commands are functions – construct the root, set args, call Execute() against an io.Writer. CLI tests went from ~10s to ~100ms.

The LLM shows its work#

The extraction pipeline proposes database operations – create a vendor, update a title, link a quote – but you couldn’t see the details. You got a summary and an accept/reject choice.

Documents now have an Ops column (#776). It opens an interactive JSON tree: every proposed INSERT, UPDATE, and DELETE with field values inline. j/k to navigate, l to expand, h to collapse. Collapsed nodes show inline previews. Clickable too.

You can point at the exact field the LLM got wrong before you accept.

Faster the second time#

Re-extracting a document used to redo the full pipeline – OCR, text extraction, everything. Now it skips straight to the LLM with the cached text from the first run (#763).

Same PR: r in edit mode triggers extraction without opening a form. A Model column shows which LLM produced the extraction. The model name and operations JSON are persisted in the database – if you need to know what model read your invoice six months from now, it’s there.

Other things since last week#

  • Charm v2 – bubbletea, lipgloss, huh, bubblezone, and glamour all migrated to their v2 releases. Go 1.26 required. The bubbletea-overlay dependency got inlined. Nothing should look different, but if something does, open an issue (#788).
  • micasa demo – --demo is now a subcommand. micasa demo --years 10 instead of micasa --demo --years 10 (#787).
  • Keybinding hints – two-tier keycap rendering: pill keycaps for inline hints, bold accent for reference panels like the help overlay (#783).
  • Document restore – accepting an extraction on a soft-deleted document now restores it instead of silently writing to a hidden row (#777).
  • Hide-deleted – soft-deleting a row now respects your explicit hide-deleted toggle instead of overriding it (#774).
  • Sort – toggling sort didn’t visually activate until you pressed a navigation key; the cached viewport wasn’t being invalidated (#773).
  • Service log sync – closing the service log overlay auto-syncs and highlights the Last column so you see the update immediately (#772).
  • Error rendering – failed extraction step errors render as plain text instead of raw JSON (#778).

Try it#

go run github.com/cpcloud/micasa/cmd/micasa@latest demo

Tab completions:

source <(micasa completion bash)
source <(micasa completion zsh)
micasa completion fish | source

Binaries on the releases page.