Skip to content

v0.14.0

Choose a tag to compare

@xuiltul xuiltul released this 24 Sep 05:40
· 921 commits to main since this release

AnimaWorks v0.14.0

This release adds a terminal chat UI and per-thread model selection, slims the agent harness and prompts, and fixes runaway chat context, rate-guard backoff and task resume.

Highlights

  • Terminal chat UI: animaworks chat NAME without a message opens a Textual TUI with a live Anima sidebar, activity feed, Markdown transcript, tool cards, a skill-aware slash palette, session resume and stream reattach. Install the tui extra (included in all-tools).
  • Model picker: choose the chat model per thread in the web chat and workspace. The model catalog is discovered from the installed CLIs and OpenAI-compatible endpoints.
  • Harness diet: the system prompt is split into resident, conditional and reference layers, behavior rules are consolidated, token estimation is CJK-aware, and priming uses one token budget with scored pointers. Mode S passes a strict MCP config and tool allowlist (about 101K → 43K prompt tokens).
  • Memory: action rules are attached to tool results instead of blocking calls. Nightly LLM pipelines are re-enabled, and forgetting and fact conflicts are handed to the weekly consolidation model.
  • Reliability: chat sessions no longer grow without bound. The rate guard honors provider-stated quota resets. Chat SSE streams have an idle timeout. Task TTL survives resume, and interrupted tasks can be resumed under their original ID.
  • Battle view for live tasks, an optional Atlas Cloud image backend (thanks @binyangzhu000-sudo), a Slack channel opt-out, Codex activity logging, and Docker onboarding fixes.

Compatibility and upgrade notes

  • Mode B is removed. ollama/* models and the legacy assisted mode map to Mode A. Replace any explicit Mode B configuration.
  • On first start, the v0140_harness_diet_resync migration overwrites runtime prompts/ and common_knowledge/ from the templates, removes four retired prompts, and maps "mode": "B" entries in models.json to "A". Local edits to shipped common_knowledge/ files are replaced; keep custom knowledge in separately named files.
  • The in-process LifecycleManager, AttentionResolver, the goal tool and several deprecated config keys are removed. Existing config.json files still load.

See CHANGELOG.md for the full list.

日本語

ターミナル用チャット UI とスレッドごとのモデル選択を追加し、ハーネスとプロンプトを軽量化したリリースです。

  • animaworks chat NAME をメッセージなしで実行すると、ターミナル UI(TUI)が開きます。サイドバー、アクティビティ、スキル一覧つきのスラッシュパレット、セッション再開に対応しています。利用には tui extra が必要です。
  • Web チャットでスレッドごとにモデルを選べます。モデル一覧はインストール済みの CLI から自動で取得します。
  • システムプロンプトを常駐・条件付き・参照の 3 層に分け、行動ルールを統合しました。priming を単一のトークン予算に整理し、Mode S ではツールを絞りました(約 101K → 43K トークン)。
  • action gate はツール呼び出しを止めず、関連ルールを結果に添える方式になりました。夜間 LLM パイプラインを再び有効にしました。
  • チャットのコンテキストが際限なく膨らむ問題、rate guard のバックオフ、SSE の停止、再開したタスクの TTL を修正しました。
  • Mode B を削除しました。 初回起動時に prompts/ と common_knowledge/ がテンプレートで上書きされます(同梱ファイルへの手動の変更は失われます)。