Changelog

All notable changes to AnimaWorks will be documented in this file.

Format based on Keep a Changelog, adhering to Semantic Versioning.

Unreleased

[0.7.0] - 2026-04-21

Added

Discord Integration

Discord連携を追加 — Gateway Bot + Webhook + チャネル同期 + UI (#156)
per-channelリードAnima — channel_members の先頭メンバーがチャネル担当
全ロールテンプレートのデフォルト外部ツールにDiscord・Notionを追加

Slack Enhancement

Slack受信メッセージにチャネル名・宛先注釈・observe intentを追加
Board→Slack双方向同期を有効化
Slackアバター自動同期 + 自動返信にAnimaアイコン表示
XSERVERへのSlackアバター自動アップロードとicon_url自動解決
AVATAR_URL__ で .env 経由のSlackアバターURL対応
default_anima のみチャンネルメッセージに自動応答
限定チャンネル対応 + Shared bot fallbackガード

Execution & Models

MICRO tier + compact_comm ツールプロファイル追加（8Kコンテキストモデル対応）
Gemma 4 モデルサポート追加（Mode A、コンテキストウィンドウ、ベンチマーク）
nanoGPT を LLM プロバイダとして追加
Codex サブスクリプションモデルをモデルピッカーに追加
ロールベースのローカルLLMプリセットを追加
Anthropic サブスクリプション認証、Claude Code CLI検出、Diffusersローカルバックエンド追加

TaskExec & Agent

TaskExec にフルコンテキストシステムプロンプトを付与（Minimal → Full）
TaskExec からの Agent/Task サブタスク生成をブロック

Notification

call_human にインタラクティブモード追加 — チャネル横断の承認ボタン

Memory & Prompt

S-mode アイドルコンパクションをactivity_log + セッション破棄ベースにリライト
Inbox (DM) の Priming とプロンプト構築をChat同等に強化

Usage & Dashboard

Usage ダッシュボード、Usage Governor、マルチプロバイダ対応追加
nanoGPT usage をダッシュボードに追加（3カラム: Claude / OpenAI / nanoGPT）
Runway・着地予測・Refresh ボタンをダッシュボードに追加
Usage Governor に時間比例のウィークリー使用制約を追加

Image Generation

IP-Adapter Face Reference サポートの改善（全画像生成バックエンド対応）
Face Reference 画像の直接インポートオプションをアセット管理に追加
remake-preview にStepsスライダー追加（品質/速度のユーザー制御）
Diffusers IP-Adapter で OpenCV による顔検出・クロップ
PIL高速パス（低VRAMでvibe reference利用時にSDXLスキップ）

CLI & Audit

animaworks anima audit に --date オプション追加（単日クエリ）
日本語/エイリアス名でのAnima解決をサポート

UI

Anima 詳細エディタ、パラメータ化ルート追加
権限UIにパス別read-onlyトグル追加
チームビルダーに8つの新ロール追加・OpenAIモデルリスト拡張
Web検索（DuckDuckGo）、アセットfullbody-onlyモード追加

CI/CD

PyPI publishing ワークフロー追加（Trusted Publishing）

Changed

Agent/Task ツールをハードブロック方式に変更（intercepting to pending → hard-block）
animaworks anima audit をレポートモードに統一（サマリーモードを廃止）
Anima name regex が日本語エイリアス・助詞（は、に、を等）に対応

Fixed

Supervisor & Background

Consolidation を busy_hang 検出から除外、cron command timeout を追加
全cronジョブ登録に misfire_grace_time=600 を追加
Consolidation IPC タイムアウトを600秒→1800秒に延長
Phase B でモデルをオーバーライドする際のexecutorスワップ修正
DEFAULT_CONSOLIDATION_MODEL から anthropic/ プレフィックスを削除
supervisor.stop() の NoneType poll エラーガード
不審な・POSIXのみのcronコマンドをWindowsでブロック

Slack

default_anima 欠損時に利用可能botへのフォールバック
Slackチャネル同期で missing_scope をdebugレベルに引き下げ
Slackアイコン URL解決の強化（不正テンプレート・非httpスキーム対応）+ 3層アイコンURL設定追加

Discord

inbox処理中の discord_channel_post ブロック（重複防止）
Discord auto_reply 指示の強化（重複投稿防止）
Guild の #dm-* チャネルをDMとして扱い intent=question をルーティング
discord_channel_sync の None チェック修正

Image Generation

IP-Adapter 使用後の UNet encoder_hid_dim_type 復元
SCRFD出力パースをInsightFace ONNX形式に対応（2-D shape、RGB→BGRチャネル順序修正）
IP-Adapter の毎回リロードでFace Reference安定性向上
remake-preview 中のアバター上書き防止（リトライフィードバックループ解消）
low-VRAM ステップキャップを10→25に引き上げ（顔品質改善）
LLM プロンプト合成に45秒タイムアウトを追加
remake-preview セッション中のアセット再調整をスキップ

Prompt & Priming

TIER_MICRO で Priming をスキップ
MICRO tier 境界を8192コンテキストウィンドウに調整

Governor

Governor notify で Messenger API 使用を修正、i18n subject追加
Heartbeat レイヤー違反の削除、Ollama プリセットガード追加
PR #149 機能のセーフデフォルト適用 (#153 #154 #155)

Other

Brave Search バックエンド復元（DuckDuckGo フォールバック付き）
pyproject.toml セクション順序とPEP 639準拠修正
サイレント except Exception: pass をlogger.debugに置換

[0.6.3] - 2026-04-02

Added

UI

Dashboard・Workspaceチャットに KaTeX 数式レンダリングを追加（KaTeX 0.16.44 + marked-katex-extension 5.1.7）

Templates & Documentation

anatomy/essentials.md を追加 — AnimaWorks全体像を1ページで俯瞰するガイド（実行パス、Heartbeat/Cron、マシン、チーム設計、記憶、コスト最適化）。[IMPORTANT] マーカーでRAG優先ブースト（ja/en/ko）
00_index.md に「Start Here」セクション追加、common_knowledge_hint.md にessentials.md参照を追加
team-designテンプレートを大幅拡充 — 法務（Legal）・財務（Finance）・トレーディング（Trading）・営業＆マーケティング（Sales & Marketing）・COO・経営企画（Corporate Planning）・カスタマーサクセス（Customer Success）・インフラ（Infrastructure）・秘書（Secretary）の9チーム追加（ja/en/ko、各ロールの injection/machine/checklist 付き）
org-chart-template.md 追加 — 組織図テンプレート
team-design/guide.md を新チームパターンで更新

Fixed

TaskExec誤完了を防止 — error/cancel/expired状態の検出、cancelled/expiredセンチネル値の正しいキューステータスへのマッピング、ストリーミングエラーチャンクの追跡、serial batchの failed_dependency パスでの _sync_task_queue 呼び出し漏れ、fuzzyマッチの最小文字数チェック（8文字）を追加 (#145)
非Anthropicクレデンシャル使用Animaでconsolidation時にモデルだけ上書きしAPIキー/ベースURLが元のまま残る問題を修正 — _resolve_consolidation_credential() でプロバイダ別クレデンシャルを正しく解決

0.6.2 - 2026-03-30

Added

Memory & Search

search_memory に activity_log スコープを追加 — 直近3日間の行動ログをBM25キーワード検索可能に。scope="all" 使用時はベクトル検索結果とRRF（Reciprocal Rank Fusion）で統合
search_memory に skills スコープを追加 — スキル・共通スキルをベクトル検索可能に
人間フィードバック・好み情報の write_memory_file 自動保存を強制化

Completion & Quality

completion_gate ツールを全実行モード（S/A/B/C/D/G）に追加 — 最終回答前の完了チェックリスト検証
Mode S: Agent SDK Stop hook による直接チェックリスト注入
Mode A: マーカーファイル方式で未呼び出し時に1回リトライ強制

Consolidation

2-phase multipass consolidation — tool_result全文を活用した高精度エピソード抽出 + エラートレース分析
Phase A: activity_logを時間チャンク分割 → LLM one-shotでエピソード抽出 → マージ・重複除去
Phase B: エラートレース分析（error + failed tool_result 収集・要約）→ 知識抽出・procedure自動生成

Organization & Delegation

委譲タスクのPriming表示 + 部下からの自動同期（sync_delegated）
スーパーバイザーの state/plans/ 読み取りアクセスを許可
cross-Anima 書き込み境界ガイダンスをプロンプト・common_knowledgeに追加

External Integrations

Discord連携ツール追加（REST API v10）— メッセージ送受信・チャネル操作
Slackユーザーメンション解決をインジェスト時に実行 + システムアノテーション付与 + intent="question" 即時トリガー最適化
call_human スレッド返信時の intent='question' 自動設定

Assets & UI

Animaアセットにアイコン追加・個別再生成・fullbodyアップロード対応 (#141)
セッション境界での current_state.md 自動アーカイブ (#143)

Templates & Documentation

team-designテンプレート追加・machine関連ドキュメントを operations/machine/ に統合
machine-tool-usageをロールベースワークフローガイドに再構成
skill-creatorを Use-when ガイド + lint_skill.py で刷新

Changed

skill ツールを廃止し read_memory_file に統合 — スキルカタログはシステムプロンプトに直接表示、read_memory_file(path="common_skills/.../SKILL.md") で全文取得
Priming Channel D（スキルマッチ）を削除 — スキル情報はシステムプロンプト内カタログに移行
common_skills/ 全スキル説明を Use-when パターンに移行（ja/en/ko）
read_memory_file に common_skills/ パス解決を追加
delegate_task スキーマに自己完結型の指示ルールをインライン化
Claude組込みの CronCreate/Delete/List ツールを拒否リストに追加
アセットルート強化・パイプラインドキュメント整合・コーディング規約修正

Fixed

POSIX cron day_of_week 番号（0=日曜）とAPScheduler（0=月曜）の互換性問題を修正
cron.md パーサーのフォーマット許容性向上（よくある書式ミスに耐性）
ローカル file:// 画像パスおよび絶対パスの解決を修正
ChromaDB InternalError を query() でキャッチしHTTP 500を防止
RAGメタデータの非整数 version/count フィールドに対するガード追加
テーブル要素のフォスターペアレンティングによるチャットレイアウト崩れを修正
data-* 属性のエスケープを属性安全方式に変更しDOM破損を防止
reply_routing のサイレント例外をデバッグログに置換
古い current_task.md が current_state.md 存在時に自動削除されない問題を修正
韓国語プロンプトのフォーマット修正

Security

litellm>=1.82.6 をピン留め — サプライチェーン攻撃バージョンの排除

Migration

v062_skill_removal_and_activity_log: テンプレート全同期（common_knowledge, prompts, reference, common_skills）+ DB tool_descriptions/guides再同期 + 旧 skill ツール記述削除

0.6.1 - 2026-03-21

Fixed

fix ruff lint violations: bare f-string (F541), missing raise from (B904), and formatting
add missing _is_port_listening and _get_daemon_log_path mocks to restart CLI tests
load GlobalPermissionsCache in E2E blocklist test after blocklist refactoring

0.6.0 - 2026-03-21

Added

New Execution Engines

Mode D (Cursor Agent) execution engine — Cursor Agent CLI subprocess with MCP integration agent loop
Mode D session continuity via cursor-agent --resume for cross-turn context preservation
Mode D system prompt optimization with A+C hybrid turn rotation
Mode G (Gemini CLI) execution engine — Gemini CLI subprocess with stream-JSON parse
6 execution modes (S/C/D/G/A/B) supported across docs, setup wizard, and settings UI (#139)

Meeting & Communication

Meeting Room mode for multi-Anima conferences with facilitator-driven discussion
block communication tools during meetings and add context summarization (#130)
message quality protocol for Anima-to-Anima communication — structured format, noise reduction (#137)
stream activity report generation via SSE

Security & Permissions

permissions.global.json — unified global command security config with startup verification
Mode S bypassPermissions — full built-in tool access without explicit allowed_tools list

Planning & Monitoring

todo_write session-scoped planning tool for Mode A agents
cron health check with cron.md parse validation and periodic monitoring
heartbeat quality improvement — observe evidence requirements, plan-outcome tracking, OK gate

Platform & i18n

Korean locale (ko) — full i18n support including prompt templates, common_knowledge, common_skills, reference docs, and web UI strings (#124)
Windows-native supervisor support (#136)
Codex login support in setup wizard and settings

Other

migration step for task_delegation_rules → common_knowledge move
token usage pricing correction and cache token tracking across all execution paths

Fixed

prevent restart helper from being killed by process scanner during shutdown
prevent activity log bloat from unbounded tool_result content
prevent delegate_task dual-trigger duplicate execution (#129)
clarify submit_tasks vs delegate_task tool descriptions to prevent misuse
enable real-time streaming for meeting mode chat bubbles
eliminate false positives in global permissions deny patterns
copy auth credentials and settings to Gemini CLI per-Anima workspace
map message:* trigger to chat session type for Mode D resume
add None guard to SDK stream cache token accumulation

Changed

move task_delegation_rules to common_knowledge and unify access
loopback host validation and i18n error messages

[0.5.5] - 2026-03-18

Added

unified animaworks migrate command — 22 migration steps across 5 categories (structural, per-anima, template sync, DB sync, version tracking) with --dry-run, --list, --force, --resync-db options
auto-migration on server startup via ensure_runtime_dir() — no manual action needed for most upgrades
migration_state.json for idempotent version tracking of applied migrations

[0.5.4] - 2026-03-18

Added

overhaul MessageDeduplicator with overflow_inbox individual file approach
fully separate ChromaDB from runner processes — server-only ownership
enforce MUST rules for memory verification and file reading before acting

Fixed

apply ruff format to dedup.py and handler_skills.py
replace silent except with logger.warning in http_store.close()
harden RAG process separation — bool returns, ABC bug, CLI safety
eliminate remaining ChromaDB ABC bypass and direct instantiation
use dedicated COMPACT_TIMEOUT_SEC (300s) for SDK idle compaction
save shortterm in idle compaction (Mode A/B/S fallback) and preserve in Mode A blocking threshold path
add extra_mcp_servers mock in test_loads_from_config_json

[0.5.3] - 2026-03-17

Added

improve priming query quality for chat/inbox/heartbeat
compress role specialty prompts and add report-formats common knowledge
improve priming quality — Channel B/C/E noise removal, knowledge injection, task priority
cap frequency_boost and implement per-anima access counting for shared collections
unify resolve_context_window with models.json as SSoT (#115)
vector-primary search_memory with rich results and episode chunking fix
align read_file budget with Claude Code Read tool constraints
Mode A context compaction — align tool output limits with Mode S + LLM one-shot compaction
replace permissions.md templates with JSON, add CLI permissions command
replace MD parsers with JSON loader, unify execution mode security
add PermissionsConfig model, loader, and MD→JSON migration
enable progressive streaming for Codex mode (C) over IPC
add --background CLI flag to machine tool & improve SKILL docs
restrict delegate_task to direct subordinates only
unify descendant permissions with direct subordinate (#108)
add replay time range selector UI (1h/3h/6h/12h/24h)
reinforce memory consolidation merge pipeline
add agent-browser skill and browser automation guide
prevent injection.md bloat — [IMPORTANT] always-prime, size governance, workspace separation
read-before-write guard and knowledge dedup hints for write_memory_file
add background parameter to Bash tool for async command execution with streaming output
centralize embedding inference into server process
SWE-bench multi-agent evaluation infrastructure
add replay mode to Dashboard (Business 2D) org-dashboard
workspace registry and alias resolution for Anima working directories
add real-time context window usage ring indicator to chat UI
graceful interrupt for Mode S session preservation
inject machine tool MUST directive into TaskExec prompt
enhance Dashboard (Business 2D) real-time visualization
add fuzzy CJK-Latin spacing tolerance to Edit tool
Phase 4 — CLI subcommands for supervisor/vault/internal tools + test fixes
unify tool schema — 18-tool Claude Code-compatible architecture
add machine-tool common skill and fix hint reference
show machine tool hint in system prompt for non-heartbeat triggers
machine tool engine priority with list discovery
unified CLI fallback routing — both entry points resolve all commands
dynamically hide machine tool engines based on CLI availability
add machine tool — external agent CLI as stateless power tools for Animas
Support prefers-color-scheme for default theme (fixes #56) (#87)
add housekeeping rotation for task_results/ and pending/failed/
add --since HH:MM filter to audit_subordinate tool and CLI
add conceptual integration of [IMPORTANT] memories (amygdala→semantic consolidation)
consolidation retry + PostToolUse knowledge frontmatter hook (#73)
task queue 2-layer sync — plan_tasks Layer 2 registration + PendingExec completion/failure sync
add task-architecture.md to common_knowledge anatomy
add AST-based hardcoded Japanese string detection test
add 2-stage heartbeat timeout (soft reminder + hard cutoff)
prevent current_task.md bloat with HB cleanup instruction and auto-pruning
task_queue compact with archival + list_tasks output optimization
deprecate send_message(intent="delegation"), enforce delegate_task usage
delegate_task writes to subordinate state/pending for immediate execution
enrich call_human thread replies with notification context
add advanced agent benchmark (Sonnet 4.6 vs Qwen3.5-35B)
add importance boost to RAG retriever (amygdala model)
add 4-model agent benchmark results and Qwen3.5-35B recommendation
add AnimaWorks agent benchmark runner for hina evaluation
add Team Presets API with industry x purpose templates
add AI Brainstorm feature with multi-character perspective generation
add Team Builder UI, External Tasks widget, fd_limits utility and watchdog graceful import
add common_knowledge access paths reference, update index files and tests
add reference/ shared directory infrastructure
move 8 detailed reference docs from common_knowledge to reference/
add streaming repetition detection as safety net
add penalty parameter support in models.json and _build_llm_kwargs
add Bedrock Kimi K2.5 thinking support via reasoning_config
pass enable_thinking via extra_body for openai/* models (vLLM)
unify thinking display across all execution modes
support thinking mode for Qwen models on AWS Bedrock
collapsible background sessions in chat main tab

Fixed

tighten session summary task extraction criteria to reduce false positives
use dict access in compressed summary keyword search test
use dict keys instead of tuple indices in conversation summary tests
use dict key instead of tuple index in procedures search test
align e2e activity log test with chat noise filtering
prevent race condition in restart helper when PID file is missing
update test assertions for compressed communication_rules template
address review feedback (iteration 1)
prevent thread context from truncating user messages in inbox
remove hiring_context from init.py, cli.py, compare_prompt_db.py
replace silent except-pass with logged debug message in rag_search
e2e test_hybrid_search_common_knowledge failing on CI
e2e test_search_memory_text_scope failing due to missing ChromaDB collection
remove dimension param from create_collection to prevent GPU model loading in runners
resolve ruff lint and format issues
update test mock paths after slack/chatwork submodule refactoring (9033eab8)
update test mock path for _call_compression_llm after refactoring
restore direct retriever injection in PrimingEngine._get_or_create_retriever
resolve ruff lint errors in core/ — I001 import sort + F401 unused import
replace parse_permitted_tools() with load_permissions() + get_permitted_tools()
inter-anima boundary check and test updates for permissions.json
eliminate SSE streaming race condition in _sse_tail event delivery
update E2E test to check task_queue.jsonl instead of current_state.md
ruff format core/memory/task_queue.py
address review feedback — remove remaining pending.md refs, batch task lookups
update budget constant test for Channel E (300 → 500)
Issue #114 — update templates and handler for current_state/pending separation
address review feedback (iteration 1)
replay now fetches all events for selected time range
set reasoning_content=None in mock to prevent MagicMock thinking injection
replace silent except Exception: pass with except OSError in pipe cleanup
update test assertions for openai/ thinking default and i18n baseline
initialize _read_paths in _make_handler for test_path_traversal
initialize _read_paths in _FakeWriteHandler and pre-populate in overwrite test
add archive_memory_file to MCP tools and strengthen consolidation prompts
block delegation tools during memory consolidation
exclude .archive/ and _archived/ from merge candidates and RAG indexing
ruff lint and format for core/ cli/ server/
replay time range now covers full requested hours
use raw vector similarity for merge candidates, improve archive exclusion
improve replay feature — event mapping, 24h range, 200x speed
address review feedback (iteration 2)
address review feedback (iteration 1)
default enable_thinking=True for openai/ models and detect untagged thinking
use polling for heartbeat intervals that don't divide evenly into 60
broaden _get_locale() exception handling to include ConfigError
remove unused imports (F401) in tests and scripts
remove extraneous f-prefix from f-strings without placeholders (F541)
remove unused imports and variables (F401/F841)
make sync tests async to match module-level pytestmark
implement strip_untagged_thinking for vLLM tag-free thinking detection
explicitly disable thinking for openai/ models when thinking=None
address review feedback — correct env var and cleanup
isolate SWE runtime from production ~/.animaworks/
mock load_auth and slack_socket_manager in server unit tests
correct invalid loop variable in schemas.py (F821)
export const チェックを test_websocket_imports_match_org_exports に追加
use visible tool in dashboard viz test
combine if-elif branches in activity.py (SIM114)
expand onDone search window from 1000 to 1200 chars
add sync comments for VISIBLE_TOOL_NAMES across FE/BE
address review feedback — buffer all WS events during replay
filter tool_use from Dashboard card streams
multi-layer defense against tag leakage into chat responses
repair org-dashboard crash + feat: add grid snap for card placement
isolate streaming updates per thread to prevent cross-thread bleed
restore context ring on Anima tab switch
wrap entire interrupt+receive in timeout guard
normalize readable locations in permission templates
add shared/ write permission to all role templates
address review feedback (iteration 1)
use horizontal whitespace only in fuzzy CJK-Latin pattern
align E2E tests with unified 18-tool architecture
map WebSearch 'limit' param to 'count' in dispatch layer
address review findings — _WRITE_TOOLS, path traversal, budget mapping
replace silent except-pass with logged exceptions in machine.py
add injection.md/status.json to Mode S subordinate file access
inject credentials from AnimaWorks config into machine subprocess
apply ruff format to machine.py and update i18n hardcode baseline
address review findings for machine tool
prevent thinking preview scroll-jump by patching DOM in-place
prevent thinking preview scroll-jump by patching DOM in-place
recalculate resolved_mode in _resolve_background_config
fall back to CPU when CUDA OOM during embedding model load
extend t() locale allowlist to include zh and ko (#78)
preserve on early-exit path for safety-net
remove misleading offset/limit hint from read_memory_file truncation message (#80)
map legacy 'ts' field to timestamp in Message validator
inboxの不正ファイルがメッセージ処理全体を停止させる問題を修正
replace IntervalTrigger with polling-based heartbeat for interval > 60min
correct claude_agent_sdk import in compact_session
grace IPC errors when process is alive to prevent false SIGTERM
IntervalTrigger の end_date によるハートビート停止バグを修正
add 2000-line truncation to read_memory_file to prevent prompt too large
resolve ruff lint/format errors (import sort, formatting)
resolve ruff lint errors in _parse_since (F821, UP037)
update remaining plan_tasks references in task_queue.py docstrings
skip parallel tasks with failed dependencies (mirror serial check)
make animaworks restart survive caller death via detached helper
ensure [IMPORTANT] tag is preserved and discoverable across memory lifecycle
depends_onを持つタスクの初期statusをpendingに修正
add trailing slash to knowledge_dir_str startswith check to prevent false matches
strip orphan tags from Qwen3.5 streaming output
address review feedback (iteration 1)
rename activity filter "タスク" → "タスク管理" to distinguish from "タスク実行"
add anatomy/task-architecture.md to expected file list
add soft < hard timeout validation to HeartbeatConfig
remove unused UTC import after datetime cleanup in handler_skills
remove redundant local datetime import shadowing now_iso in _handle_plan_tasks
update test_task_metrics mock for split list_tasks() calls
update audit.py for new list_tasks() default behavior
strip residual tags when multiple think blocks emitted
isolate TestSchedulerManagerE2E from system config
update skill-creator test assertion and restore tags field in ja template
use MagicMock for synchronous get_pid in restart_race tests
address review feedback (iteration 2)
update remaining en templates to deprecate send_message delegation intent
update remaining ja templates to deprecate send_message delegation intent
replace silent except-pass with logger.debug in call_human
save notification mapping in CLI call_human (Mode S reply routing)
update existing tests for delegation intent deprecation and SDK hook changes
use detail length for tool_detail events in debug log
multiple small fixes — enable_thinking=False on Bedrock + chunk counter off-by-one
use actual event name in debug log (was hardcoded as text_delta)
repair test_stream_exception_handling mock as proper async generator
set data-theme attribute on body for dark themes (closes #53)
add HEALTHCHECK to Dockerfile and docker-compose files (closes #52)
add .dockerignore for smaller Docker builds (closes #51)
downgrade debug-labeled stream logs from INFO to DEBUG
repair malformed tool-call JSON in non-streaming execute() path
update RepetitionDetector tests for n=10/threshold=10 defaults
relax RepetitionDetector thresholds to reduce false positives
repair malformed tool-call JSON from GLM-4.7 thinking mode
simplify StreamingThinkFilter and add Bedrock GLM thinking support
support vLLM reasoning parser that strips opening tag
resolve CI failures — ruff format + CSS hover/active parity
text-format tool call IDをイテレーションごとにユニーク化
Bedrock tool calling — keep toolConfig when history has toolUse/toolResult
Llama 4 Maverickのテキスト形式ツールコールをパースして実行する
update stale references to files moved to reference/
fall back to non-streaming for models without streaming tool use
update test expectations for moved reference files
set litellm.modify_params=True in all executors for Bedrock compatibility
immortalize SDK sessions — remove TTL, preserve on compaction failure
sanitize tool_use_id for Bedrock Converse API compatibility
address review feedback (iteration 2)
rename unused loop variables to underscore prefix (B007, 10件)
restrict synthetic thinking_blocks injection to Anthropic models only
use PID-unique temp file in save_config to prevent concurrent rename race
replace blind Exception with ValidationError in pytest.raises (B017/F841)
rename unused loop variables to _ prefix (B007/I001)
bind loop variable in lambda to prevent B023 closure bug
ruff format _litellm_streaming.py
add enable_thinking to _thinking_enabled check in streaming
ruff format core/execution/base.py
resolve UP032/UP015/SIM118/UP012/F401 lint warnings (5 files)
apply same Qwen Bedrock routing to assisted.py (Mode B)
resolve SIM110/SIM114/SIM103/B011/F841 lint warnings
add # noqa: F401 to try/except availability-check imports
replace invalid noqa directives and remove trailing whitespace
remove unnecessary f-prefix and unused imports (ruff F401/F541)
mock Agent SDK fallback in test_llm_failure_returns_none
untrack private-only files and harden .gitignore
prevent auto-scroll to bottom during thinking zone updates
patch Agent SDK transport to allow graceful CLI shutdown
pass thread_id to ConversationMemory and ShortTermMemory
pass source param to process_message_stream in streaming handler
update Message.source docstring to include googlechat
add source parameter to process_message for external platform awareness
add text_delta to mock_stream so archive_paths is called
add text_delta to mock_stream so archive_paths is called
move overflow:hidden to base .chat-bubble to cover all bubble types
wrap long lines in code blocks to prevent bubble overflow
infinite scroll not working for tool-heavy animas (hinata)
prevent chat bubble content overflow causing page-level scroll
task exec markdown rendering, compact header, italic muted body
session splitting on trigger change, subtle bg-session styling
don't archive inbox messages when LLM returns empty response
address review feedback (iteration 1)
bootstrapループ防止（3つの構造的バグ修正）
prevent bootstrap infinite loop (3 structural bugs)
update session_tool_uses tests for deferred-chaining design
remove stale inject_shortterm patches from litellm_loop test
SDK session complete isolation — chat-only resume, fresh for background

Changed

compress communication/messaging prompt templates
remove hiring_rules/hiring_context system prompt injection
remove per-line truncation and align Mode S Read default with CC
extract magic numbers to named constants in compaction logic
split _image_clients.py into image/ package
split handler_org.py into focused Mixin submodules
decompose PrimingEngine, config/models, and ConversationMemory god classes
split slack.py and chatwork.py into focused submodules
split lifecycle.py, builder.py, chat.py into submodules (WT-3)
split i18n.py and schemas.py into domain-based packages (Phase A)
extract agent_sdk.py into _sdk_interrupt, _sdk_options, extended _sdk_stream/_sdk_session
fix ambiguous test assertion for jira args
rename current_task.md to current_state.md and AnimaStatus.current_task to active_label
add missing trust levels, fix lint and spec typo
restrict emotion_instruction to chat trigger only
streamline EN/KO/ZH READMEs to match JA structure
fix indentation, remove redundant imports, update exception catches and tests
replace generic RuntimeError with domain-specific exceptions
update tests, docs, and scripts for task tool rename
trim conceptual integration prompt — remove PII examples, reduce verbosity
rename task tools — plan_tasks→submit_tasks, add_task→backlog_task
extract _handle_hard_timeout helper to stay within 85-line budget
remove all intent="delegation" references from templates
SDK hook delegates only when subordinate explicitly named
replace str+Enum with StrEnum in evaluation framework (UP042)
restructure common_knowledge 00_index.md for clarity
improve bootstrap loop fix

Performance

open output file once per stream thread instead of per line
optimize Board page — lazy load, caching, incremental polling

Other

フィードバックに対応しました
symple化
Anima個別のMCP設定を追加
GoogleTasks対応
Revert "fix: explicitly disable thinking for openai/ models when thinking=None"
Revert "fix: prevent bootstrap infinite loop (3 structural bugs)"
Revert "style: apply ruff format to bootstrap loop fix files"

[0.5.2] - 2026-03-09

Added

add md_to_chatwork() Markdown sanitizer for Chatwork messages
add encapsulation boundary classification to anima-anatomy docs
add Heartbeat and Cron tabs to chat sidebar
slack_channel_post / slack_channel_update as gated external tool actions
gated external tool actions — default-deny safety valve for dangerous sub-actions
add TextAnimator for smooth FE streaming text display
internationalize timezone handling — auto-detect system TZ + configurable override
add "Enter to send" toggle in Settings page
dual-query RAG strategy + language-agnostic keyword extraction
add slack_react tool for emoji reactions (#22)
add daily RAG indexing to ProcessSupervisor and fix per-anima vectordb
add per-anima RAG index builder script
add GitHub Release workflow with LLM-generated release notes
recursive directory indexing for all memory types (#20)
change ChromaDB distance metric to cosine similarity (#19)
normalize system prompt heading hierarchy with XML tags
add .ragignore support and retriever min_score threshold (#18)
DK summary injection — replace full-text with title+description lists
restrict recent_tool_results injection to Mode B only
budget-aware timeline thinning for LLM activity report input
i18n audit timeline strings and send plain text to Activity Report LLM
unified timeline audit with cross-anima merged view
update LLM prompt for key_activities and add CLI audit --mode report
add qualitative fields (key_activities, top_tools) to AnimaAuditEntry
redesign audit report mode with priority-based category display
rewrite audit_subordinate to match Issue specification
add Activity Report page (#17)
add one_shot_completion() with LiteLLM → Agent SDK fallback
add time-based activity schedule (night mode)
add global Activity Level slider for heartbeat cost control
add generic Notion external tool
add profile subcommand for multi-instance management
unified outbound budget with role defaults + status.json override
add demo interactive onboarding with 3-layer experience
migrate ~480 hardcoded Japanese strings to i18n t() system

Fixed

remove stale server.pid before starting in Docker container
address review feedback (iteration 2)
protect identity.md from self-modification in Mode S
update environment.md rule 7 to document subordinate management file permissions
replace get_event_loop() with get_running_loop() in async contexts
skip malformed activity log entries instead of failing entire file
remap 'event' key to 'type' in activity log loader
use resp is not None instead of isinstance(resp, dict) for SlackResponse
correct indentation error in test_rag_e2e.py
assign haiku model to general-role animas in demo
add missing demo examples and adjust_dates.sh to Dockerfile
correct indentation error in test_heartbeat_decomposition_e2e.py
replace date.today() with today_local() in all e2e and housekeeping tests
force-reset _app_tz in conftest to prevent timezone state leakage
replace date.today() with today_local() across all test files
use today_local() in conversation transcript tests to match JST implementation
use today_local() instead of date.today() in activity log rotation tests
patch get_credential in x_search tests to isolate from shared/credentials.json
patch get_credential in web_search tests to isolate from shared/credentials.json
replace silent except-pass with debug log in asset reconciliation config load
replace silent except-pass with debug log in gated action permission check
update template file count test for usecases/ directory
use safe .get() for threadId in send_message response
propagate context and reply_to fields in plan_tasks batch handler
improve Channel D skill matching — configurable threshold & personal-first sort
add explicit None checks for remaining get_vector_store() callers
plan_tasks _wake() callback receives unwanted self argument
graceful degradation when RAG/ChromaDB fails to initialize
update scheduler E2E test for daily indexing cron job
robust indexing_time/enabled extraction from consolidation config
TextAnimator accumulator-based timing and rate calculation
widen onDone handler scan window in streaming indicator test
update org context E2E test assertions for refactored comm rules
align Mode B skill injection test with priming-based matching
fix TextAnimator timing bug — reset _lastStepTime on idle ticks
address review feedback (iteration 1)
update record_access unit tests for DB-read access_count
address review feedback (iteration 2)
read access_count from DB in record_access to prevent stale increment
remove CHUNK-DEBUG temporary logging from chat streaming
add missing mock attrs for indexing cron + resolve lint violations
replace silent except-Exception-pass with debug logging
resolve ruff lint and format violations
reset ToolPromptStore singleton between tests to prevent cache leak
unify anima_factory locale fallback to locale→en→ja
remove unnecessary noqa comment from error handler
suppress Slack Bolt 404 for unhandled events
align tests with communication-rules refactor and remove date-sensitive timestamps
increase asyncio StreamReader buffer for Codex subprocess pipes (#21)
address review feedback (iteration 2)
align skills hash pattern with SKILL.md-only indexing
address review feedback (iteration 1)
address review feedback (iteration 1)
update activity report tests for generate_org_timeline mock
exclude self-addressed messages from outbound limit count
preserve LLM utils fixes and initial truncation safety net
remove in-place meta mutation and dead meta_copy code in audit
add backward compatibility for legacy 'days' parameter
address review findings (C1/H1-H3/M1/M5)
add localStorage fallback for activity settings and fix demo splash link
raise on LLM failure in conversation _call_llm to preserve compression behavior
add error handling to settings page API calls (night mode save/load)
replace broad 'except Exception: pass' with KeyError in scheduler_manager
address review feedback (iteration 1)
resolve ruff lint (I001 import sort) and format violations
replace undefined CSS variables in demo splash and suggest cards
downgrade demo models from opus/sonnet to sonnet/haiku for cost savings
reuse httpx.Client, add CLI tests (coverage 65% → 92%)
address review feedback (iteration 1)
address review feedback (iteration 1)
centralize consolidation model for all internal LLM calls
prevent zombie process accumulation via explicit wait() and periodic reaper
reset _last_progress_at on lock acquisition to prevent busy-hang false positives
update _run_priming mock return value to match new tuple signature
correct expected Slack user ID in test_slack_prefix_case_insensitive
skip Group 6 header for task trigger in system prompt builder
improve chat renderer and chat styling
resolve flaky tests caused by hardcoded dates and dedup cache pollution
restore docs/images/ needed by README

Changed

compress communication rules & messaging prompts (~1,000 tokens saved)
compress environment.md prompt — remove Claude-redundant instructions
simplify memory_guide — remove skill list, use counts
move public design docs from docs/implemented/ to docs/specs/
i18n hardcoded Japanese strings in skill_creator and related modules

Performance

truncate tool fields in conversation API and optimize history poll diff

Other

bug fix indexer.py
modify debug systemprompt

[0.5.1] - 2026-03-06

[0.5.0] - 2026-03-06

Added

update demo defaults to en-business preset with real avatar assets
add Anima identity (username/icon_url) to slack_send tool
add Slack notification icons for all Animas
workspace dashboard live status, KPI polling, and activity streams
replace tier-based prompt scaling with linear budget allocation
implement live activity cards for workspace org-dashboard
add message lines & avatar variants to org dashboard
replace org-dashboard with canvas node graph layout
progress-aware busy hang detection replacing counter-based kill
unified hot-reload system for config and connections
add copy/download action buttons to assistant chat bubbles
eliminate Any types with Protocol/TypedDict replacements
add animaworks anima rename CLI command
fix silent failures in task lifecycle and memory I/O
unify tool visibility across all execution modes (S/A/B)
unified housekeeping engine for disk rotation
auto-inject reply instruction metadata for external platform inbox messages
enforce MUST task creation in heartbeat/inbox prompts
config thread-safety and cleanup fixes
notification channel vault/shared credential support and robustness fixes
use os.replace for atomic writes and atomic truncation in cron_logger
handle Slack app_mention events with ts-based dedup and thread reply context
auto-assign intent on Slack mention/DM for immediate inbox processing
align skill-creator with Agent Skills spec and fix path resolution
document external message reception in messaging-guide (ja/en)
add call_human guide to common_knowledge (ja/en)
resync prompt DB sections migration + behavior_rules memory-backed advice rule
add --force flag to stop/restart commands
validate/repair broken knowledge frontmatter on write and startup
improve _fetch_node_content fallback and update_graph_incremental type inference
add LLM API retry for 429/5xx/network errors
write recovery_note on heartbeat process crash and fix unread_count
pending task failure safety — file move lifecycle and failure notifications
per-Anima Slack bot token resolution
add background_model override for heartbeat/inbox/cron cost reduction
expose credential vault tools (vault_get/vault_store/vault_list) to Animas
shared common_knowledge per-anima index with hash-based change detection
search_memory OR-split and priming keyword fallback
add AI-speed task deadline guidelines to environment.md
fix Agent tool intercept and add heartbeat task results visibility
spreading activation repair + episodes support
expose all supervisor tools in Mode S MCP + add CLI audit subcommand
consolidation quality improvements — frontmatter repair, REFLECTION extraction, smart activity filtering
incremental sync of common_skills/common_knowledge on startup
add gmail inbox, sent, and search subcommands

Fixed

regenerate female business avatars with correct gender prompt
add gender field to demo character sheets and fix prompt generation
add missing bubble-actions and voice-controls-slot to workspace CSS/HTML
add missing chat avatar styles to workspace CSS
always show amber spinner for any running stream entry
detect active groups on init so workspace spinners reflect ongoing tasks
replace silent except-pass with debug logging in reconciliation
org-dashboard spinner only spins during active tasks, not idle Running
address review feedback (iteration 1)
resolve ruff lint errors and time-dependent test failures
repair 3 broken unit tests
pass execution mode explicitly to _estimate_tool_overhead
add :active parity for .org-card:hover selector
address review feedback (iteration 1)
prevent error handler double-fault, fix exception hierarchy and restore fallbacks
address review feedback (iteration 1)
7 bugfix batch — error handling, activity log, health check, client lifecycle
update regression tests for new exception propagation
resolve per-anima Slack token in CLI path
MCP supervisor safe fallback + executor context window overrides
address review feedback (iteration 2)
remove stale lastChunkTime reference in chat-stream.js
address review feedback (iteration 1)
replace broad except Exception with specific custom exceptions
address review feedback (iteration 1)
address review feedback (iteration 2)
replace inbox file creation with activity log for external sends
address review feedback (iteration 1)
use _thread.LockType for Python 3.12 isinstance compatibility
resolve all 8 failing tests on main
rename E2E test to clarify regression guard intent
address review feedback (iteration 1)
correct test mocking targets for CLI staleness tests
CLI staleness — log path, --local deprecation, exit code, env var unification
use getattr for record.result_summary robustness
cron logger data integrity — KeyError on command entries and timezone mismatch
use record.result_summary for tool_end completed_tools summary
prevent broadcast() race condition with list snapshot
use Date comparison for liveIsNewer to handle UTC/JST timezone mismatch
route call_human replies back to originating Anima
add num_retries to Mode B LLM calls for transient error resilience
align repair metadata with ensure_knowledge_frontmatter
address review feedback (iteration 1)
address review feedback (iteration 2)
add defensive type check for reply_to in failure notification path
update base tool count 27→30 for vault tools addition
update E2E outbound test for per-Anima slack _send_via_slack signature
add vault tools to dispatch dict test expectations
resolve CI test failures — vault MCP schemas and vibe reference realistic path
disable spreading activation in RAGFilter tests
resolve review regressions — test alignment, silent exception, code quality
address review findings — result.id bug, C-method, config params, threading lock
remove Chatwork-specific Bash blocks and update S-mode tool docs
address review findings — remove dead imports, fix parse-fail path, include E2E test updates

Changed

apply ruff format to server/routes/system.py
apply ruff format to 3 remaining files
introduce ruff linting/formatting and improve CI pipeline
remove deprecated modules and simplify architecture
unify external platform source constants and deduplicate _detect_slack_intent
unify claude-opus-4-20250514 refs to claude-opus-4-6
simplify gmail tool — extract _fetch_emails, Email.to_dict, constants

Performance

reduce markdown re-render interval to 30ms / 10 chars
fix chat streaming stutter — incremental markdown, ASGI middleware, log reduction
fix streaming chat display jank — RAF batching, log reduction, SSE flush

[0.4.10] - 2026-03-04

Added

DK removal Phase 1+2 — full Channel C search + budget expansion
add Priming Channel F (episodes) and search_memory episodes support
add check_background_task / list_background_tasks MCP tools
add audit_subordinate supervisor tool for monitoring subordinate activity
expose plan_tasks via MCP and update heartbeat prompts to recommend it
add model info CLI commands, anima info, and comprehensive docs
add think tag strip filter for Qwen3.5 content-embedded reasoning
improve task awareness — origin_chain human bonus + heartbeat add_task guidance
implement live tool activity streaming — tool_detail SSE + subordinate activity broadcast
improve heartbeat effectiveness — raise tool limit, filter activity noise, enforce STALE tracking
implement live tool activity streaming with real-time UI updates
implement Board channel ACL (access control for shared channels)
add Qwen 3.5 model support (Mode A + 64K context window)
implement credential vault encryption with PyNaCl SealedBox
per-Anima Chatwork write token + fix streaming-controller container ref
idle conversation pre-compression, Claude auto-update, server PID detection fix
update system prompts for S-mode Task tool auto-routing
resolve avatar URLs for chat history from_person and workspace
add anima avatar display to chat bubbles
persist per-pane anima/thread selection across reloads
demo README & Quick Start — English/Japanese guides with Docker demo link
demo fictional runtime data — 3-day activity logs and state files
demo asset infrastructure — directory structure, generation and optimization scripts
add chatwork_delete tool for self-message deletion
NovaCraft world-building — 4 presets × 3 characters with full personality
demo Docker infrastructure with preset selection
Task tool delegation + SDK subagent for S-mode
add frontend image resize & cache module for avatar thumbnails
add token usage tracking and cost estimation

Fixed

auto-convert anime prompts to realistic in asset generation
adjust episode budget to 500 tokens per Issue spec
skip memory_eval e2e tests when experiments/ unavailable in CI
address review findings — ElevenLabs TTSSynthesisError, exception separation, tests
replace silent except-pass with debug logging in audit_subordinate
patch path for MeshyClient credential mock in asset optimization tests
per-thread interrupt event via ContextVar for parallel streams
await missing on neurogenesis_reorganize + catch-up missed consolidation jobs
transcribe tool broken via submit — composite name + subcommand mismatch
address review findings for audit_subordinate
update dispatch dict test to include audit_subordinate tool
token usage input/output_tokens always 0 — use dict.get() instead of getattr()
add missing tool category flags to Mode B AssistedExecutor
RC-1 無音嚥下防止 + RC-6 interrupt時response_done保証
TTS P0 — stop swallowing synthesis errors (RC-1) and guarantee response_done on interrupt (RC-6)
filter subordinate tool activity by org hierarchy on frontend
avoid buffering non-think content in StreamingThinkFilter
filter subordinate activity to prevent global tool_use event leakage
priming Channel C keyword extraction and search accuracy
auto-inject frontmatter for knowledge/procedures, repair Priming pipeline
prevent subprocess leak by re-raising GeneratorExit in async generators
unify started_at to milliseconds in stream_registry.py
prevent keepOnlyStreaming from clearing completed chat messages
address Critical/Important review findings for Board ACL
replace fragile split("---", 2) frontmatter parsing with line-based parser
voice TTS playback failure on reconnect — AudioContext state management and server robustness
update credential resolver test to expect vault.json in error message
guard app.js init() against double execution on Settings navigation
prevent mic button presence from shifting send button layout
prevent pane auto-focus on stream end, add flash notification
モバイル情報パネルが右にズレて崩れる問題を修正
address review findings for token usage tracking
voice chat UI improvements — layout, VAD loading, TTS sanitization, duplicate response prevention

Changed

simplify streaming-controller and session-manager
remove obsolete UI test scripts and re-enable realistic animations
separate streaming indicator animation for active vs inactive tabs

Performance

paginate activity API — cap per-Anima loading instead of O(N) full scan
zone-based partial DOM updates for streaming chat bubbles

Other

Revert "refactor: simplify streaming-controller and session-manager"

[0.4.9] - 2026-03-02

Added

Chat マルチペイン分割 — VS Code 風に複数チャットインスタンスを横並び表示（分割/閉じるボタン付き）
gmail_draft に添付ファイル対応とスレッド返信自動解決を追加
アイコンのみ表示時にアニマ名ツールチップをポップアップ表示（デスクトップ hover / モバイル tap）

Fixed

アイコンのみ表示時（サイドバー折りたたみ / モバイル）にアニマタブの閉じるボタンが非表示で操作不能だった問題を修正
モバイルでドロップダウンがペインの overflow:hidden にクリップされる問題を修正
Priming レイヤー5件のバグ修正（CUDA OOM, Channel B スコアリング, 表示肥大, Channel D フォールバック, outbound 切り詰め）
SDK セッション再開時に即座にコンパクションが発生する問題を10分タイムアウトで防止
gmail_draft attachments パラメータの文字列→リスト変換を追加
37件のテスト関数を現行実装に合わせて更新

[0.4.8] - 2026-03-02

Added

アセットRemakeで画像スタイル選択UI追加、configデフォルト参照に変更
style-aware prompt separation for realistic image generation
asset remake UX improvements — scratch generation, preview history, expression grid
unified anime/realistic avatar display and remake support
split display mode from color theme, add Settings page
smart scroll with floating scroll-to-bottom button
image generation pipeline enhancements and asset reconciler updates
mobile chat UX improvements, dashboard, and setup wizard enhancements
theme system with 10 color presets and dropdown selector
DAG scheduler for parallel task execution
restore transcript writing and add shared conversation log
thread tab styling improvements — subtle active state, streaming pulse, completion indicator

Fixed

add locale-based ethnicity to realistic prompt conversion
replace hardcoded anima names with generic placeholders in templates
sidebar active menu text invisible on dark themes
thinking inline preview text invisible on dark themes
dark theme input text color - use token instead of hardcoded black
streaming stop button and per-thread concurrency
update tests for asset reconciler default, streaming indicator path, and deferred trigger
update streaming indicator test for refactored chat JS
prevent Codex SDK LimitOverrunError on thread resume with large prompts
use JST date in tool_result_log and channels tests
handle additional Codex SDK event types (text.delta, response.completed)
use JST date in heartbeat history test for CI timezone compat
ensure chat queue auto-drains after streaming completes
use JST date for activity_log paths in more test files
preserve full message content in heartbeat dedup consolidation
use JST date in activity spec test to match now_iso() timezone
resolve CI test failures and add activity group_type filter

Changed

align private repo tracking with public, merge PR #1
rename dashboard anima list label to Org Chart

[0.4.7] - 2026-03-01

Added

publish.sh --release で自動バージョンインクリメント
cron中のinbox抑制 + per-anima flockによる多重起動防止
add origin metadata to RAG chunks and trust-separated priming output (provenance phase 4)
propagate origin_chain in Anima-to-Anima messaging (provenance phase 3)
add origin tracking at external data entry points (provenance phase 2)

Fixed

session chaining時にSDK session IDをクリアしてfreshセッションを開始
compaction空白地帯の解消 — context window是正・閾値スケール修正・Mode Sチェイニング有効化
Mode S streaming compaction failure — 2 root causes
prevent repeated content in session-chained responses
make status.json the SSoT for supervisor/speciality fields
detect stale cron/heartbeat schedules via file mtime reconciliation
detect stale cron/heartbeat schedules via file mtime reconciliation
load mode_s_auth from status.json into resolved config

[0.4.6] - 2026-02-28

Fixed

Mode C (Codex SDK) session-chaining: add CodexResultMessage adapter for num_turns/session_id interface
Mode C prompt selection unified with Mode S via _is_mcp_mode() helper (communication_rules, messaging, hiring_rules, tool guides)

[0.4.5] - 2026-02-28

Added

Mode S (Agent SDK) multimodal image input support
text artifact popup viewer for code blocks (file: cards)
ConversationMemory provider-specific credential injection (Bedrock/Azure/Vertex)
toggle for chat right-side status pane
Mode C (Codex SDK) interrupt_event support

Fixed

create-anima name parsing bug: Japanese headings no longer captured as anima name
create-anima documentation updated from deprecated create-anima to anima create
chat message duplication and cross-anima display contamination
concurrent per-anima streaming in chat UI
chat stream error recovery and reconnection handling
recovered chat content preservation on reload

Changed

major backend refactoring: split 6 God-class modules into focused Mixin files
- handler.py (3298 → 8 files)
- agent_sdk.py (1716 → 5 files)
- image_gen.py (2434 → 6 files)
- activity.py (1471 → 6 files)
- manager.py (1419 → 4 files)
- litellm_loop.py (1392 → 4 files)
frontend refactoring: split large JS modules
- chat.js (2490 → 12 controllers)
- character.js (1396 → 5 modules)
- office3d.js (1208 → 5 modules)
- chat-controller.js (1009 → 5 modules)
- app.js (716 → 4 modules)
- timeline.js (576 → 4 modules)
extract shared chat logic to shared/chat/ for Dashboard/Workspace reuse

[0.4.4] - 2026-02-27

Added

expose permitted external tools as native MCP tools in Mode S

Fixed

harden media proxy with extracted secure module
harden parsing paths and align e2e expectations
switch chat controls to SVG and add desktop sidebar toggle
improve workspace chat tab stream indicators
auto-refresh chat view every 5 seconds
enrich cron job parsing with next and last run data
streamline memory tabs and support scheduler job fallbacks
load scheduler jobs from split API fields
tighten chat input box height on chat and workspace
reduce workspace chat action button sizes
harden log viewer path handling and polish UI behaviors
resolve markdown image paths and add attachment fallback
scale down images in chat bubbles rendered via markdown
voice TTS not switching when changing anima tab
resume active stream on reload regardless of process status
streaming stop button now scoped to current anima+thread

0.4.3 - 2026-02-27

Added

surface assistant image artifacts in chat and history
add persistent dashboard chat tabs with unread stars
multi-thread chat backend + frontend
add frontend i18n with shared i18n.js module and locale JSON files (L6)
add core/i18n.py and externalize hardcoded Japanese strings in Python (L5)
add English translations for knowledge, skills, and roles templates (L4)
message queue with management UI and queue-anytime support
pending message queue, icon-only buttons, dynamic button state
add pending message queue with interrupt-and-send for chat UI
implement call_human reply routing from Slack (Issue #1)
intercept SDK Task tool → pending LLM task for background execution
add voice output sanitization and voice-mode suffix for TTS
migrate skill format from flat files to directory structure
Mode B補助輪強化 — ツール仕様テキスト改善 + インテント検出リプロンプト
add process control buttons and LLM session interrupt to WebUI
add web_fetch internal tool for Anima URL content retrieval

Fixed

improve add-conversation anima menu UX
refine chat tab UI behavior and responsive layout
preserve streaming partial responses during heartbeat relay
polish chat input UX and refresh static asset versions
make VAD auto mode fully hands-free
close remaining assistant-image review gaps
isolate thread histories in chat view and tighten voice reply length
improve workspace voice input UX and metadata handling
restore VAD auto mode and suppress silent STT hallucination
prevent duplicated voice stream updates on UI reinit
improve chat UI behaviors and thinking stream visibility
show avatar icons in anima chat tabs
restore old chat threads and improve tab UX
improve workspace streaming UX and sync docs navigation
improve workspace stream resume and thinking preview behavior
avoid task-intercept blocked misclassification
resume dashboard chat stream after page return
clear chat display immediately when switching threads
clear chat input immediately after send submission
make channel ID regex case-insensitive for consistency
thread_id validation and conversation view filtering
close prior activity group on next trigger
update review docs and chat sidebar responsiveness
persist chat drafts until successful send
resolve review revision findings for i18n L4-L6
update tests for i18n string changes and adapt evaluation docs
remove bottom whitespace in chat sidebar memory section
avoid no-response when stream events are missing
make send/queue buttons perfect circles with explicit width+height
prevent mobile auto-zoom on chat input focus
catch StopAsyncIteration on Agent SDK session resume
add voice chat balloon callbacks to standalone chat page
prevent duplicate process spawn via _starting/_restarting guards
add per-anima mode_s_auth to prevent shared API key rate limiting
address review findings — missing interrupt_event paths, CSS :active parity

Changed

address round-2 review findings for reply routing
address review findings for call_human reply routing
Cursor-style input — embed buttons inside textarea container
use English persona-aware message for Task intercept deny reason

Performance

parallelize anima startup and make web server start first

[0.4.2] - 2026-02-26

Added

voice chat bubble integration, thinking delta, mobile touch improvements
add Codex SDK execution mode C for OpenAI Codex CLI integration
thinking UI integration — all frontends + persistence + voice indicator
add thinking streaming events and collapsible UI component
integrate adaptive thinking into S/A execution engines
add thinking_effort schema, resolve_max_tokens, adaptive thinking helpers
add prompt i18n (ja/en) — locale-aware template system
add status.json hot-reload via IPC without process restart
enable extended thinking for Bedrock Claude models
add business theme UI with CSS design tokens and workspace org dashboard
add voice chat system (STT + TTS + WebSocket)
add AWS Bedrock provider support for LiteLLM execution
expand supervisor file access permissions for subordinate management
enable WebFetch/WebSearch native tools in S mode
add bustup image overlay on avatar click in chat page

Fixed

HTTPS reverse proxy auth loop and asset reconciliation infinite retry
preserve thinking_blocks in LiteLLM tool-call iterations
resolve frontend regressions and improve workspace org dashboard
sanitize thinking text rendering to prevent XSS
resolve review revision findings for prompt i18n
use output_config instead of reasoning_effort for Anthropic SDK
resolve all 98 test regressions from i18n restructuring + main rebase
update assisted.py adaptive thinking + fix config_reader test mocks
resolve review findings — cli.py path regression + DRY _get_locale
91件の失敗テストを現行実装に追従させる
address review findings — disposeOffice on view switch, Lucide createIcons timing
raise compaction threshold ceiling from 0.95 to 0.98
web_search dispatch bug and context threshold auto-scaling
add builder.py to silent-pass allowlist for org tree status.json fallback
board mobile scroll and channel switching
add missing tool result for ToolExecutionError in serial path
improve error handling with custom exception hierarchy
resolve type safety issues across codebase

Changed

move outbound section to PrimingEngine

[0.4.1] - 2026-02-25

Added

Cron sessions now use heartbeat-equivalent context (full identity + memory + org); removed separate heartbeat trigger from cron
Chatwork outbound messaging re-enabled with chatwork send support
Subagent CLI execution skill (common_skills/subagent-cli.md) for delegating shell tasks to Claude Code subprocess

Fixed

Chatwork send rejects with clear error when WRITE token is not configured (previously silent failure)
Skill/procedure invocation instruction in memory_guide updated to reference skill tool (progressive disclosure)

0.4.0 - 2026-02-25

Added

Execution & Architecture

3-path execution separation — Heartbeat/Inbox/TaskExec with independent locks and trigger-based prompt filtering
Tiered System Prompt — 4-tier progressive reduction (T1 Full → T4 Minimal) based on context window size
Prompt injection defense with boundary labeling (trusted/medium/untrusted trust levels on tool results and Priming data)
debug_superuser flag for unrestricted file/command access bypass (debug Anima support)

Supervisor & Organization

Supervisor tools expansion — 6 new tools for manager Animas (org_dashboard, ping_subordinate, read_subordinate_state, delegate_task, task_tracker, restart_subordinate)
Per-anima denied command enforcement from permissions.md

Memory & Knowledge

Tool result consolidation — persist tool results to long-term memory via daily consolidation
Procedures/knowledge separated injection into system prompt (distinct budget allocation)
write_memory_file enables common_knowledge/ writes with improved hints
read_file hardening — dynamic line limits, line numbers, code block formatting, safety notes, partial reads

Communication & UI

Messaging data model unification — DM/Message event names consolidated, dm_logs deprecated (activity_log is primary)
Board reverse pagination — newest messages first with infinite scroll
Board DM list UI improvements — mini avatars, sorting, layout optimization
Activity timeline — per-anima selection fix and trigger-based grouping
Chatwork files/download subcommands

Configuration

status.json as Single Source of Truth for Anima model configuration (2-layer resolution: status.json → anima_defaults)
DM log rotation registered as daily system cron in LifecycleManager
Orphan Anima archival before auto-deletion

Fixed

Context window exceeded: automatic tier downgrade with hard truncation fallback
max_tokens default raised from 4096 → 8192
LiteLLM streaming empty response diagnostics and ContextVar reset safety
MCP integer type validation auto-relaxation and rate limit error messages
Setup page password configuration failure
Activity timeline grouping — per-anima tracking eliminates cross-anima orphans
Command meta-character blanket rejection relaxed to blocklist approach
Final iteration tool exclusion to force final text answer
DM display issues — dedup, garbage pair filter, arrow notation
Board offset calculation for 3+ page channels
Board infinite scroll completion
Skill descriptions included in system prompt memory_guide section
Heartbeat tool instruction and org context directory scan
check_permissions external tool enumeration bug and task_tracker private method usage
permissions.md section header inconsistency and DM question intent
3-path execution review fixes — state_file_lock, inbox status, wake signal
24 failing tests updated to match current source code

Changed

injection.md model info abolished — status.json is now the sole model config source
identity.md/injection.md placed immediately after Group 1 to guarantee personality resolution before any context
Remove system prompt duplicates and unify legacy terminology

0.3.1 - 2026-02-25

Changed

Update all default models to current generation: Opus 4 → Opus 4.6 ($5/$25, 67% cheaper), Sonnet 4 → Sonnet 4.6 (same price, 1M context), Haiku 4.5 → GPT-4.1-mini / Gemini 2.5 Flash
Role templates updated: engineer/manager use claude-opus-4-6, writer/general/researcher use claude-sonnet-4-6
Context window map: add 1M entries for Opus 4.6, Sonnet 4.6, GPT-4.1 family
Setup wizard: update provider model lists to current generation
Centralize default model name into DEFAULT_ANIMA_MODEL constant — future model updates only need 1 line change + role templates

Fixed

Orphan directory prevention: 3-layer defense (pre-creation validation, automatic cleanup, config consistency check)
read_memory_file returns directory listing on File not found instead of bare error
ToolCallRecord dataclass JSON serialization error
ARG_MAX exceeded: oversized system prompts passed via temp file

Added

Azure OpenAI api_version / Vertex AI credential passthrough to LiteLLM
CHANGELOG.md with auto-generation script (scripts/generate_changelog.py)
SSE/IPC layer separation — producer task decoupled architecture
Unified outbound rate limiting for send_message / post_channel

Performance

Reduce system prompt bloat: remove s_builtin section, add knowledge budget cap

[0.3.0] - 2026-02-25

First official release. AnimaWorks is a framework that treats AI agents not as tools but as autonomous individuals ("Anima"), each with their own identity, memory, and decision-making criteria.

Added

Core Framework

Anima lifecycle management — create, delete, disable, enable with CLI and API
Process Supervisor — each Anima runs as an isolated child process with Unix Domain Socket IPC
Hierarchical organization — supervisor field defines reporting structure, messaging-based communication
Role templates — 6 preset roles (engineer, manager, writer, researcher, ops, general) with model/parameter defaults
3-layer config resolution — per-anima override > role template > global defaults
Unified credential management with config.json cascade

Execution Engine

Mode S (SDK) — Claude Agent SDK with Claude Code subprocess, streaming, PreCompact hooks
Mode A (Autonomous) — LiteLLM + tool_use loop for GPT-4o, Gemini Pro, Ollama models with tool support
Mode B (Basic) — framework-mediated I/O for lightweight/tool-less models
Automatic mode resolution via wildcard pattern matching on model names
Session chaining — automatic context overflow detection and new session creation
Streaming support across all execution modes with SSE relay
ARG_MAX protection — oversized system prompts passed via temp file
Context tracking with message_start event parsing (S mode)

Memory System

RAG engine — ChromaDB + intfloat/multilingual-e5-small (384-dim) with incremental indexing
Knowledge graph — NetworkX-based spreading activation with Personalized PageRank
Priming layer — 5-channel parallel automatic recall injected into system prompt
- A: Sender profile, B: Recent activity, C: Related knowledge, D: Skill match, E: Pending tasks
Dynamic budget allocation by message type (greeting/question/request/heartbeat)
Consolidation — daily episode→knowledge synthesis (NREM sleep analog) + weekly merge/compression
Active forgetting — 3-stage synaptic homeostasis (downscaling → reorganization → complete forgetting)
Unified activity log — all interactions recorded as JSONL timeline per Anima
Streaming journal — crash-resistant Write-Ahead Log for streaming output recovery
Conversation memory with automatic compression (display 20 / trigger 50 / retain 20)
Shared user memory — cross-Anima user profiles in shared/users/
Atomic file writes with fsync and two-stage recovery

Communication

Internal messaging via Messenger (send_message) with async delivery
Board — shared channels (append-only JSONL) with mentions and DM history
External messaging — Slack (Socket Mode + Webhook) and Chatwork integration
Unified outbound routing — auto-detect internal Anima vs external platform
Human notification — call_human with multi-channel support (Slack, Chatwork, LINE, Telegram, ntfy)
Outbound rate limiting — 3-layer cascade prevention for message storms
DM gratitude loop and board pollution suppression

Autonomy

Heartbeat — periodic self-check with customizable checklist
Cron — scheduled tasks with YAML definition, async parallel execution
Background task submission and execution
Task queue with persistence, deadline enforcement, and delegation prompt injection
Heartbeat/conversation parallelization with lock separation

System Prompt

6-group structured prompt builder (environment → identity → situation → memory → organization → meta)
Distilled knowledge injection (knowledge/ + procedures/, 10% of context budget)
Dynamic tool guide generation per execution mode
Behavior rules with MUST constraints

Web UI

FastAPI server with WebSocket real-time updates
SPA dashboard with activity timeline, status panels, and memory viewer
3D office workspace with pathfinding, idle behaviors, and desk layout
Visual novel-style conversation screen with expression variants
Board UI — channel/DM browsing and posting
Setup wizard (GUI) with language selection (17 languages)
Multi-user authentication (password + localhost trust)
Mobile responsive design (iPad Safari viewport fix, touch support)
SSE streaming for chat and heartbeat responses
Infinite scroll pagination for conversation history
Multimodal image input in chat

Tools

External tool framework — auto-discovery, creation, hot-reload, unified dispatch
Web search and X (Twitter) search
Slack and Chatwork messaging
Gmail integration
GitHub and AWS integration
Image generation — NovelAI API, fal.ai (Flux), Meshy (3D models)
Transcription (Whisper) and local LLM (Ollama)
Supervisor model control — parent Anima can change child's model and restart

Asset System

Character image generation with bust-up expression variants
Vibe Transfer for style consistency across team
3D model generation via Meshy API with FBX→glTF conversion
Asset reconciler — periodic batch generation for missing assets
3-layer cache (API / delivery / compression) for 3D models

CLI

animaworks init — workspace initialization with safe merge and full reset modes
animaworks server — start/stop/restart with PID file management
animaworks chat / animaworks send — interactive and one-shot messaging
animaworks create-anima / animaworks delete-anima — Anima management from character sheets
animaworks index — RAG index management
animaworks config — configuration management with export-sections
animaworks board — Board channel management
One-liner setup script for quick installation

DevOps

Publish script — private→public repo sync with rsync, PII scan (Claude Code), and changelog
Apache-2.0 license with SPDX headers
Memory evaluation framework — 3 ablation experiments with synthetic datasets
Comprehensive test suite (unit + E2E)

Changed

Migrated from distributed architecture (Gateway + Worker + Redis) to monolithic FastAPI server
Renamed execution modes from A1/A2/B to S/A/B
Switched license from AGPL-3.0 to Apache-2.0
Moved model mode patterns from config.json to models.json
Tool permissions changed from whitelist to default-allow (blacklist) model

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

Changelog

Unreleased

[0.7.0] - 2026-04-21

Added

Discord Integration

Slack Enhancement

Execution & Models

TaskExec & Agent

Notification

Memory & Prompt

Usage & Dashboard

Image Generation

CLI & Audit

UI

CI/CD

Changed

Fixed

Supervisor & Background

Slack

Discord

Image Generation

Prompt & Priming

Governor

Other

[0.6.3] - 2026-04-02

Added

UI

Templates & Documentation

Fixed

0.6.2 - 2026-03-30

Added

Memory & Search

Completion & Quality

Consolidation

Organization & Delegation

External Integrations

Assets & UI

Templates & Documentation

Changed

Fixed

Security

Migration

0.6.1 - 2026-03-21

Fixed

0.6.0 - 2026-03-21

Added

New Execution Engines

Meeting & Communication

Security & Permissions

Planning & Monitoring

Platform & i18n

Other

Fixed

Changed

[0.5.5] - 2026-03-18

Added

[0.5.4] - 2026-03-18

Added

Fixed

[0.5.3] - 2026-03-17

Added

Fixed

Changed

Performance

Other

[0.5.2] - 2026-03-09

Added

Fixed

Changed

Performance

Other

[0.5.1] - 2026-03-06

[0.5.0] - 2026-03-06

Added