Fella v0.1.3

A patch on top of v0.1.2: the Windows /update command now actually
installs the update it downloads.

Fixed

• Windows: /update never applied the download. Since /update landed
  in v0.1.1 it would find the new version, download it, verify it against
  SHA256SUMS, close Fella — and then nothing installed and the app never
  came back. Two things were wrong. The detached hand-off that runs the
  installer was created with no console at all, so powershell.exe
  exited during startup before running a single line — no install, no
  relaunch, no trace. It was also built as a cmd string whose quoting
  arrived mangled, and it waited a fixed two seconds that often wasn't long
  enough for Fella to release the lock on its own binary, so a silent
  installer couldn't overwrite it and simply gave up. The hand-off is now a
  powershell -File script with its own hidden console that waits for the
  lock to actually release, runs the installer to completion, relaunches,
  and writes each step to %TEMP%\fella-update\update.log. Verified end to
  end on a real Windows build (0.1.2-style install → /update → installs
  and relaunches). The download-and-checksum half was always fine, and
  nothing is touched on disk until the checksum passes, so a failed apply
  still can't produce a broken install — the verified installer is left in
  %TEMP%\fella-update\ to run by hand.

Install

macOS / Linux

curl -fsSL https://lilfella.app/install.sh | sh

Windows (PowerShell)

irm https://lilfella.app/install.ps1 | iex

Already on an earlier build? Open Fella and type /update — and on Windows,
this is the release where that finally works.

Or grab the build for your OS directly from this release: .dmg (macOS),
-setup.exe / .msi (Windows), .AppImage / .deb (Linux).

Verify your download

Builds are unsigned. Check against SHA256SUMS, attached to this release:

sha256sum -c SHA256SUMS          # macOS / Linux

Get-FileHash <file> -Algorithm SHA256   # Windows — compare by hand

Known limitations

• /update on macOS and Linux is still best-effort, not yet verified
  against a real install on those platforms — Windows is now tested end to
  end. Nothing is touched on disk until the checksum passes, so a failure
  there means "didn't update", never a broken install — fall back to the
  install command above.
• run_python is not a hostile-code sandbox — see v0.1.0's notes,
  unchanged in this release.
• Unsigned, no signed auto-updater. /update covers the common case;
  code signing and notarisation are still unplanned for now.

Fella v0.1.2

A patch on top of v0.1.1: reopen past conversations, correct totals on
messy number/text columns, a real app icon, a tighter default window, and a
Windows folder-picker fix.

Added

• /history <n> reopens a past conversation. /history now lists your
  saved conversations — a preview of the first question, the message count,
  the date, and which folder it was about — instead of only a count and a
  file path. /history <n> loads that conversation back into a new tab. If it
  was about a different folder than the one currently open, a note explains
  that a new question will answer from the current folder.
• parse_num() for messy numeric columns. A helper (usable from /sql
  and by the model) that reads a number a person wrote — a currency sign,
  thousands separators, a trailing %, accounting-style negatives (1,234) —
  and returns nothing for what it can't parse, so SUM/AVG over a
  partly-messy text column skip the unreadable rows instead of a bare CAST
  silently returning a wrong total. When a column is kept as text because it's
  only mostly numeric, its schema note now points at parse_num and a way to
  count what didn't parse.

Changed

• Custom app icon — the default Tauri template icon is replaced with a
  real Fella mark, a soft dome character in the brand teal, across every
  platform's icon set.
• Default window is 760×720 (was 1080×760), sized to the reading column so
  the welcome screen and conversations don't sit in a narrow strip with wide
  empty margins on first launch. Still resizable; the OS remembers a size you
  set.
• README brought back in line with the shipped app: five commands that
  weren't listed (/tab, /focus, /history, /retry, /help), the tab
  keyboard shortcuts, per-tab model, the full set of verification checks, all
  three pack kinds, and follow-up/session memory.

Fixed

• Windows: the mouse cursor went invisible while navigating the /open
  folder picker when /open was run by typing it and pressing Enter.
  Windows hides the pointer while you type and normally restores it on the
  next mouse move, but the modal picker takes the message loop before that
  happens. Fella now restores the pointer before opening the picker, matching
  the welcome-screen "Choose a folder" button.

Install

macOS / Linux

curl -fsSL https://lilfella.app/install.sh | sh

Windows (PowerShell)

irm https://lilfella.app/install.ps1 | iex

Already on an earlier build? Open Fella and type /update.

Or grab the build for your OS directly from this release: .dmg (macOS),
-setup.exe / .msi (Windows), .AppImage / .deb (Linux).

Verify your download

Builds are unsigned. Check against SHA256SUMS, attached to this release:

sha256sum -c SHA256SUMS          # macOS / Linux

Get-FileHash <file> -Algorithm SHA256   # Windows — compare by hand

Known limitations

• run_python is not a hostile-code sandbox — see v0.1.0's notes,
  unchanged in this release.
• Unsigned, no signed auto-updater. /update covers the common case;
  code signing and notarisation are still unplanned for now.
• /update's per-OS apply step is best-effort. Nothing is touched on disk
  until after the checksum passes, so a failure means "didn't update", never a
  broken install — fall back to the install command above.

Fella v0.1.1

A small update on top of v0.1.0: one new command, two real bugs fixed.

Added

• /update — check for a newer release and install it, without a full
  manual reinstall. Downloads the right installer for your OS, verifies it
  against SHA256SUMS (the same check the install scripts already do), then
  closes Fella and hands off to the installer. Manual only — it never checks
  automatically or in the background.

Fixed

• /reindex broke every already-loaded table. A workspace that opened
  fine could report the exact same, unchanged file as unreadable the moment
  /reindex ran later in that session. Affected any tabular file (CSV, TSV,
  JSON, NDJSON, XLSX), not a single file type.
• install.ps1's checksum check always failed on Windows PowerShell 5.1,
  regardless of what the release's SHA256SUMS actually contained. This was
  already fixed live before this release — the install scripts aren't
  versioned release artifacts — noted here for completeness.

Install

macOS / Linux

curl -fsSL https://lilfella.app/install.sh | sh

Windows (PowerShell)

irm https://lilfella.app/install.ps1 | iex

Already on v0.1.0? Open Fella and type /update.

Or grab the build for your OS directly from this release: .dmg (macOS),
-setup.exe / .msi (Windows), .AppImage / .deb (Linux).

Verify your download

Builds are still unsigned. Check against SHA256SUMS, attached to this
release:

sha256sum -c SHA256SUMS          # macOS / Linux

Get-FileHash <file> -Algorithm SHA256   # Windows — compare by hand

Known limitations

• /update's per-OS apply step is best-effort, not live-verified on
  every platform yet. Nothing is touched on disk until after the checksum
  passes, so a failure here means "didn't update", never a broken install —
  fall back to the install command above if it doesn't work for you.
• run_python is not a hostile-code sandbox — see v0.1.0's notes,
  unchanged in this release.
• Unsigned, no signed auto-updater. /update covers the common case;
  code signing and notarisation are still unplanned for now.

Fella v0.1.0

Fella is a small local-first desktop app for personal analytics. Point it at a
folder of your own files — bank statements, health exports, workout logs,
notes, receipts — and ask questions in plain language. Every answer is
produced by deterministic computation (SQL, or Python when SQL can't express
it) never by the model guessing, and every answer shows its working.

This is the first public build.

Core

• Plain-language questions over a folder of your own files — no schema, no
  setup. Point Fella at a folder and start asking.
• Deterministic answers. Every figure in a reply comes from a tool result —
  run_sql against an in-memory SQLite table, or run_python (a restricted
  subprocess) when SQL can't express the computation — never from the model
  directly.
• Verification pass. After answering, Fella re-runs the queries the answer
  cites and confirms every number it stated actually appears in a real
  result; a mismatch is surfaced, not hidden.
• The working. A fold under every answer (▸ working · N steps · …) shows
  every tool call, the exact SQL/Python, sample rows, timings, and the
  self-check results.
• Read-only. Fella reads the folder you point it at; it never writes,
  moves, or deletes anything there.

Supported files

• Tabular — .csv .tsv .json .ndjson. Column types are inferred,
  including currency written as text ($1,200.00, thousands separators,
  accounting-style negatives), and loaded as SQL tables.
• Spreadsheets — .xlsx, one table per sheet.
• Documents — .pdf .txt .md .log, searched with grep_files and
  read whole with read_file. No index; works on every model provider.
• .parquet needs a source build with --features duckdb (not in the
  default binary, kept small).

Models

• Local, private, default — a local Ollama install;
  nothing leaves the machine.
• Hosted, free — /login → Ollama Cloud, a free tier with starter
  models, no local install needed.
• Hosted, other providers — Vercel AI Gateway, OpenAI, xAI, OpenRouter, or
  any OpenAI-compatible endpoint via /login, at your discretion (typically
  paid per token).
• API keys live in a 0600 auth.json, never in the settings database,
  localStorage, or the transcript.

Extensibility (opt-in, none bundled)

• Packs — theme (colour tokens), skill (vocabulary/rules fed to the
  model), and mcp (connect a remote data source over the Model Context
  Protocol). Install from a local folder (/packs add <path>) or by id from
  a small seed catalog (/packs install <id>, hash-checked downloads).
• fella.md — drop one in your folder to tell Fella how your files are
  organised and what your terms mean. No pack required.

Interface

• One window: pick a folder, ask questions. Slash commands (/open,
  /files, /schema, /sql, /login, /model, /reindex, /packs,
  /connect, /clear) are a power-user shortcut — you never need them.
• Conversation tabs, Ctrl+K command palette, mid-run stop (Esc, or the
  pulsing dot next to the composer), Markdown-rendered answers.
• Configurable via environment variables: FELLA_MAX_STEPS (agent step
  budget), FELLA_QUERY_TIMEOUT_SECS (SQL timeout).

Install

macOS / Linux

curl -fsSL https://lilfella.app/install.sh | sh

Windows (PowerShell)

irm https://lilfella.app/install.ps1 | iex

Or grab the build for your OS directly from this release: .dmg (macOS —
drag Fella to Applications, then right-click it → Open the first time),
-setup.exe / .msi (Windows — More info → Run anyway if
SmartScreen warns), or .AppImage / .deb (Linux).

Verify your download

Builds are unsigned — there's no code signing or notarisation yet. Check
your download against SHA256SUMS, attached to this release:

sha256sum -c SHA256SUMS          # macOS / Linux

Get-FileHash <file> -Algorithm SHA256   # Windows — compare by hand

Known limitations

• Unsigned, no auto-updater. Re-run the install command (or re-download)
  to update.
• run_python is not a hostile-code sandbox. It runs in a restricted
  subprocess (python3 -I, a fresh temp working directory, a stripped
  environment, CPU/memory/file-size limits, a wall-clock timeout) — best-effort
  isolation for analysing your own data, not a security boundary. It does not
  confine filesystem reads to the workspace or block network access. Treat a
  snippet as code you chose to run on your own machine.
• The hosted pack marketplace isn't live yet. /packs install <id> pulls
  from a small seed catalog; /packs add <path> works fully offline.
• .parquet requires building from source with --features duckdb.