v0.4.3

Bug fixes.

v0.4.2

Bug fixes.

v0.4.1

• A Generic OpenAI provider - any OpenAI-compatible server should work with it.
• The web UI gets a dark theme switch.

v0.4.0

• Added web UI: otaku web opens the same session in a browser — the second frontend, over the
  same stories, the same lore and the same commands.
• Added native Windows support (10 and 11, 64-bit only).
• The context builder is reworked from the ground up, after docs/context_design.md; the new
  max_context setting keeps the prompt inside the window models still handle well for roleplay.
• The lore browser in terminal (/lore, /cast) is reworked to include the system message and
  the story messages.
• Prompt caching for OpenRouter.
• Sound notifications.
• A second, longer sample story.

v0.3.0

• New feature: character card import with the /card command.
• Commands /me and /you suggest and autocorrect cast names.
• New commands available mid-prompt only: /ooc and /cue, both hinting the LLM — /ooc
  introduces a standing note, /cue a one-time steer that the LLM won't see later.
• The browsers (/stories, /lore, /model) follow the terminal theme.
• Commands get highlighted.

otaku 0.2.2

• New providers: OpenRouter and NanoGPT (cloud), llama.cpp and LM Studio (local); the prompt for
  cloud providers is $ instead of >.
• Numerous UI improvements, including:
  • dialogue coloring;
  • undo erases the taken-back exchange from the screen;
  • regenerate erases the old reply and streams the new one in its place;
  • a rule drawn across the screen wherever the played story breaks;
  • providers configurable directly in the model picker (/model or Ctrl+O);
  • /system accepting a file in addition to text input.
• New commands: /clear, /last, /balance; command /rename renamed to /title.
• New CLI command: otaku update.

otaku 0.2.1

Packaging only — no functional changes.

Changed

• The required Python version is now 3.11, down from 3.14.
• Dependency bounds updated.

otaku 0.2.0

0.2 is a ground-up rewrite as a roleplay terminal client: chats are stories that can be
branched from any message, a background pass extracts lore from played messages — scenes,
characters, journals — and the context sent to the model keeps the opening and the recent tail
verbatim with scene summaries in between.

What carries over from 0.1: local models, one terminal, encrypted storage. The 0.1 chat and fleet
features (one-shot/pipe mode, otaku list/otaku stop, cross-provider RAM management,
/remember defaults) are gone. The database schema is new.

Added

• Stories: branch from any message, fork with its memory, resume where you left off; a
  full-screen story browser with message-level resume and in-place editing.
• The lore engine: idle-debounced background extraction closes scenes over played messages, keeps
  per-character journals with rolled-up histories, and feeds the story back to the model as a
  recap; /lore browses and edits the memory in place.
• Roleplay commands: /me, /you, /ooc — framing joined at wire time, bodies stored verbatim.
• Import and export: a lossless Markdown story document, SillyTavern .jsonl chats, and
  plain-text dismantling — the format detected from the file.
• A sample story seeded into a fresh database, so a first launch lands mid-story — with or
  without a reachable model.
• Day-rotated request, system, and error logs (otaku logs).

otaku 0.1.2

Hardening release: complete state-dir isolation, never-overwrite key handling, and REPL quality-of-life.

Fixed

• OTAKU_CONFIG_DIR is now fully isolated. Previously a custom state dir
  wrote a config pointing at the default ~/.otaku/history.db, and — much
  worse — its first run replaced the shared OS-keychain key, making the
  primary setup's history permanently undecryptable. Each dir now gets its own
  default database and its own keychain item; the default ~/.otaku keeps the
  legacy item name, so existing installs keep unlocking.
• Unlock failures no longer stutter ("Could not unlock encryption key: could
  not unlock…") and a key mismatch is spelled out instead of surfacing as an
  empty error.
• /remember no longer risks dropping other models' saved defaults when
  model_defaults.json is corrupt — it refuses instead, and writes atomically.

Changed

• Key setup never overwrites anything. Provisioning reuses an existing
  keychain or retrieve_command key instead of replacing it, keychain writes
  are add-only, and keys.json is created exclusively — so deleting
  keys.json no longer burns the key, and a restored backup keeps working.
  A stray pre-envelope history.key is left in place rather than deleted.
• Bare otaku unlocks encryption before the model picker, so interactive
  key ceremonies (passphrase, command providers) happen up front.
• /history rows fall back title → summary → first prompt.
• Config paths in notices are shown ~-shortened; shorter -nr banner.

Added

• /model remembers the switched-to model as the last used one (the picker
  preselects it next time).
• [defaults].verbose — set the stats line globally (a session-wide UI
  preference; deliberately not per-model, so /remember doesn't pin it).

otaku 0.1.1

Initial public release. (0.1.0 was a premature PyPI upload from a pre-release
tree — it was never tagged and is superseded by 0.1.1.)

Added

• Multi-backend client for Ollama, LM Studio, MLX (omlx), and any
  OpenAI-compatible server — from one terminal command.
• Zero-config first run: the initial ~/.otaku/config.toml auto-detects each
  built-in engine's port (and omlx's API key) from your environment
  (OLLAMA_HOST) or the engine's own settings file, falling back to the
  standard default. Runs once, at that first write; edit the sections freely
  afterwards.
• Cross-provider model management: otaku list (with --running to show only
  loaded models), load and unload from the picker, otaku stop --all, with a
  live RAM gauge. Provider queries run concurrently with a short (0.5s) probe
  timeout, so one configured-but-down provider no longer slows every command;
  when nothing is reachable, otaku names each provider, whether it answered,
  and points at ~/.otaku/config.toml to fix.
• Chat REPL: streaming responses, thinking-effort control, tok/s stats
  (/set verbose, off by default), triple-quoted multiline input, in-chat
  model switching (/model), /new (fresh conversation) vs /clear (reset
  context in place), and slash commands.
• Streaming markdown rendering: headers, lists, blockquotes, rules, and fenced
  code blocks (syntax-highlighted via Pygments) on top of inline emphasis/code.
• omlx output smoothing ([providers.omlx].smooth, default on): de-jitters
  omlx's bursty token delivery into steady typing, without affecting tok/s.
• Persistent session defaults: a [defaults] config section (system, think,
  parameters, no_record) plus per-model overrides keyed by bare model name;
  /remember saves the current settings as the model's defaults.
• One-shot / pipe mode: otaku <model> "prompt" and … | otaku <model>
  print a plain reply and exit (prompt + stdin combined instruction-first),
  so otaku works as a Unix filter.
• Encrypted conversation history (AES-256-GCM), searchable across all
  conversations by full message content, with background LLM-generated
  summaries (idle-debounced so they never block exit or reload a cold model;
  [defaults].create_summaries / summary_idle_seconds), user-set titles
  (/title, shown in the /history picker), and resume-from-any-turn.
• Get answers out: /copy (last reply or whole chat → clipboard, via the native
  tool or an OSC 52 fallback) and /save <file> (conversation → Markdown).
• Install via uv tool install or Homebrew (brew install enclavum/tap/otaku).
• Runs on macOS, Linux, and Windows. On Windows the streaming-time Ctrl+R
  (cancel + regenerate) shortcut is disabled — it needs a POSIX terminal — but
  everything else works; WSL gives full parity.