Clouds Coder 2026.03.30-Stable #10
FonaTech
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
CHANGELOG 2026-03-30
Based on 0325 Version. Bugs Fixes and Some Strategy Adjustments.
English
Headline: Universal Skills Ecosystem + Dual RAG Architecture + Core Reliability Fixes
1. Universal Skills Ecosystem Compatibility (Critical)
5-ecosystem compatibility: Clouds Coder now loads and executes skills from any of the five major skill ecosystems without any per-provider adapters:
awesome-claude-skillsMinimax-skillsskills-mainkimi-agent-internalsacademic-pptx-skill-mainRoot cause of previous failures: The Execution Guide injection (lines 11094–11131) forced
read_filecalls on virtual skill paths that don't exist in the filesystem, causing the model to loop indefinitely trying to read non-existent files instead of following the skill's SKILL.md instructions.Fixes and simplifications:
_skill_chain_completion_blocker,_record_skill_chain_entry) — eliminated over-engineered interception that silently blocked skill execution_broadcast_loaded_skillblackboard writes from 16 fields → 6 fields (name, path, description, loaded_at, trigger_context, source)_loaded_skills_prompt_hintfrom ~350 tokens → ~120 tokens: compact in-context hint that tells the model which skills are loaded without cluttering its context budgetmax_tokens=120) so the model makes its own judgment about what skill fits the task type — no keyword-based forced triggersTodoWritefor coordinating plan steps with skill executionNew:
_preload_skills_from_plan_steps— scans plan step text for skill name mentions and proactively preloads them before execution begins, reducing skill-load latency mid-execution.Shell path auto-quoting (
_rewrite_shell_virtual_paths): paths containing spaces are now automatically wrapped in double quotes before shell dispatch, fixing execution failures on macOS paths with spaces.Plan expansion:
2. Dual RAG Architecture — Code RAG + Data RAG (High)
Architecture: Two independent ingestion and retrieval engines, both built on TF_G_IDF_RAG:
RAGIngestionService(Data RAG): handles documents, PDFs, structured data files — general knowledge baseCodeIngestionService(Code RAG): handles source code files with code-aware tokenization — code-specific knowledge baseUnified retrieval:
query_knowledge_library(query, top_k)searches both libraries in parallel and returns a merged ranked result, so the model always queries one interface regardless of content type.RAG guide injection: Both
research-orchestrator-proandscientific-reasoning-labnow include a full retrieval guide documenting thequery_knowledge_libraryinterface, parameter meanings, response format, and best-practice query patterns. The model can leverage the knowledge base directly from within a loaded skill.3. Built-in Skills Overhaul: research-orchestrator-pro & scientific-reasoning-lab (High)
research-orchestrator-prorewritten as cooperative decision hub:scientific-reasoning-labrewritten as 5-step self-iterating reasoning engine:4. Multi-Factor Priority Context Compression (High)
Problem: Previous
_auto_compactdiscarded messages chronologically (oldest first), which could drop task-critical information (current plan step, recent errors, active skills) while retaining low-value content from early in a session.New
_classify_message_priority— 10-factor scoring (0–10):TodoWrite,plan_step,finish_task): +2Error:, exception traces): +2<loaded-skill>,skill loaded): +1compact-resumenote: forced to 10 (always preserved)New
_priority_compress_messages— priority-based three-tier compression:_build_state_handoffenhanced with four new structured fields:PLAN_PROGRESS: completed/total plan stepsCURRENT_STEP: text of the active plan stepACTIVE_SKILLS: list of currently loaded skillsRECENT_TOOLS: summary of last 5 tool calls_auto_compactintegration: priority compression runs first; original chronologicalpop(0)is preserved as a safety fallback if priority compression doesn't reduce tokens far enough.5. Anti-stall Mechanism Optimization (Medium)
Problem:
_manager_apply_anti_stalltriggered "CHANGE YOUR APPROACH" after only 2 consecutive delegations to the same target, interrupting agents that were legitimately making incremental progress across multiple turns.Changes:
6. Critical Bug Fixes (High)
Fix 1 —
CodeIngestionService._flush_lock(AttributeError)AttributeError: 'CodeIngestionService' object has no attribute '_flush_lock'when uploading code files to the Code LibraryCodeIngestionService.__init__completely overrides parentRAGIngestionService.__init__without callingsuper().__init__(), so_flush_lock = threading.Lock()(initialized in the parent) was never createdself._flush_lock = threading.Lock()at the end ofCodeIngestionService.__init__Fix 2 — Frontend
setTaskLevel()complexity selector resetssetTaskLevel()calledupdateLevelBtn(lvl)to update the UI button but never calledscheduleSnapshot(), so the next SSE snapshot refresh overwrote the button state with stale server datascheduleSnapshot({forceFull:false, delayMs:80, allowWhenFrozen:true})afterupdateLevelBtn(lvl), matching the pattern already used byapplyModel()Fix 3 —
_sync_todos_from_blackboarddrops worker TodoWrite itemsTodoWritefrom developer/explorer/reviewer agents only persist for one round, then disappear on the next blackboard syncowner ∈ {developer, explorer, reviewer}were being filtered out fromnon_system_rows(as non-system items) but were not included insystem_rowseither — so they were silently lost on every sync cycleworker_rowslist and merged with priority (placed between system rows and non-system rows), protected from sync overwritesFix 4 — Anti-stall threshold and instruction (see item 5 above)
Fix 5 — Multi-factor context compression (see item 4 above)
2026-03-25 Summary
query_knowledge_libraryretrieval interface and injected retrieval guides in built-in skillsresearch-orchestrator-proredesigned as a non-interfering analysis hub;scientific-reasoning-labrebuilt as a 5-phase self-iterating reasoning engine_flush_lockAttributeError, task-level selector UI revert, worker todo preservation, and the two above中文
标题:Skills 生态系统兼容 & 双库 RAG 架构 & 内核可靠性修复
1. Skills 生态系统全面兼容(重要)
5 大生态系统兼容:Clouds Coder 现在可原生加载并执行来自以下五大生态系统的 skills,无需任何 per-provider 适配器:
awesome-claude-skillsMinimax-skillsskills-mainkimi-agent-internalsacademic-pptx-skill-main此前失败的根因:Execution Guide 注入(行 11094–11131)强制对虚拟 skill 路径发起
read_file调用,但这些路径在文件系统中根本不存在,导致模型陷入循环,无法按 SKILL.md 指令正常执行。修复与简化:
_skill_chain_completion_blocker、_record_skill_chain_entry等)— 消除了过度工程化的链路拦截,曾静默阻断 skill 执行_broadcast_loaded_skill黑板写入:16 字段 → 6 字段(name/path/description/loaded_at/trigger_context/source)_loaded_skills_prompt_hint:约 350 tokens → 约 120 tokens,在不浪费上下文预算的前提下向模型告知已加载的 skillsmax_tokens=120),模型根据任务类型自主判断调用哪个 skill,而非关键词强触发TodoWrite协调 plan steps 与 skill 执行新增
_preload_skills_from_plan_steps:扫描 plan steps 文本中的 skill 名称,在执行开始前提前预加载,减少执行中途 skill 加载的延迟。Shell 路径自动引号(
_rewrite_shell_virtual_paths):含空格的路径在 shell 分发前自动加双引号,修复 macOS 含空格路径的执行失败问题。Plan 扩展:
2. 双库 RAG 架构 — Code RAG + Data RAG(高)
架构:两个独立的摄取与检索引擎,均基于 TF_G_IDF_RAG:
RAGIngestionService(Data RAG):处理文档、PDF、结构化数据文件 — 通用知识库CodeIngestionService(Code RAG):处理源代码文件,使用代码感知分词 — 代码专用知识库统一检索:
query_knowledge_library(query, top_k)并行检索两个库并返回合并排序的结果,模型使用单一接口检索任意内容类型。RAG 指南注入:
research-orchestrator-pro和scientific-reasoning-lab均内置完整的 RAG 检索指南,记录query_knowledge_library接口说明、参数含义、响应格式和最佳实践查询模式。模型可在 skill 内直接利用知识库。3. 内置 Skills 重写:research-orchestrator-pro & scientific-reasoning-lab(高)
research-orchestrator-pro重写为协作型决策中枢:scientific-reasoning-lab重写为 5 步自迭代推理引擎:4. 多因素优先级上下文压缩(高)
问题:旧版
_auto_compact按时间顺序(最旧的先丢弃)裁剪消息,可能丢弃任务关键信息(当前 plan step、近期错误、活跃 skills),同时保留会话初期的低价值内容。新增
_classify_message_priority— 10 因素评分(0-10):TodoWrite、plan_step、finish_task):+2Error:、异常 traceback):+2<loaded-skill>、skill loaded):+1compact-resume注释:强制设为 10(始终保留)新增
_priority_compress_messages— 优先级三级压缩:_build_state_handoff增强,新增四个结构化字段:PLAN_PROGRESS:已完成/总计 plan stepsCURRENT_STEP:当前活跃 plan step 的文本ACTIVE_SKILLS:当前已加载的 skills 列表RECENT_TOOLS:最近 5 次工具调用摘要_auto_compact整合:优先级压缩优先执行;原有的时序pop(0)作为保底 fallback 保留。5. Anti-stall 机制优化(中)
问题:
_manager_apply_anti_stall在连续 2 次委派给相同 target 后就触发 "CHANGE YOUR APPROACH",打断了正在合理分步推进的 agent。变更:
6. 关键 Bug 修复(高)
修复 1 —
CodeIngestionService._flush_lock(AttributeError)AttributeError: 'CodeIngestionService' object has no attribute '_flush_lock'CodeIngestionService.__init__完全重写了父类RAGIngestionService.__init__但未调用super().__init__(),导致父类初始化的_flush_lock = threading.Lock()从未创建CodeIngestionService.__init__末尾添加self._flush_lock = threading.Lock()修复 2 — 前端
setTaskLevel()复杂度选择器回弹setTaskLevel()调用updateLevelBtn(lvl)更新了 UI 按钮,但未调用scheduleSnapshot(),导致下一次 SSE 快照刷新用服务端旧数据覆盖按钮状态updateLevelBtn(lvl)后添加scheduleSnapshot({forceFull:false, delayMs:80, allowWhenFrozen:true}),与applyModel()已有的模式保持一致修复 3 —
_sync_todos_from_blackboard丢失 Worker TodoWrite 条目TodoWrite写入的条目仅在当前轮次有效,下一次黑板同步后消失owner ∈ {developer, explorer, reviewer}的条目在同步时被从non_system_rows中过滤出去(作为非系统条目),但又没有被加入system_rows,导致每个同步周期都静默丢弃这些条目worker_rows列表,合并时优先保留(位于 system rows 与 non-system rows 之间),不再被同步覆盖修复 4 — Anti-stall 阈值与指令(见第 5 项)
修复 5 — 多因素上下文压缩(见第 4 项)
2026-03-25 修复总结
query_knowledge_library,RAG 检索指南注入到内置 skillsresearch-orchestrator-pro重设计为非干扰性分析中枢;scientific-reasoning-lab重构为 5 阶段自迭代推理引擎_flush_lockAttributeError、任务级别选择器 UI 回弹、worker todo 保护,以及上述两项日本語
タイトル:Skills エコシステム互換 & デュアル RAG アーキテクチャ & コア信頼性修正
1. Skills エコシステム全面対応(重大)
5 エコシステム対応:Clouds Coder は以下 5 つの主要 skills エコシステムから skills を読み込み、per-provider アダプターなしで実行可能:
awesome-claude-skillsMinimax-skillsskills-mainkimi-agent-internalsacademic-pptx-skill-main以前の失敗の根本原因:Execution Guide インジェクション(行 11094–11131)がファイルシステムに存在しない仮想 skill パスに
read_fileを強制し、モデルが SKILL.md 指示に従う代わりに無限ループに陥っていた。修正と簡素化:
_skill_chain_completion_blocker、_record_skill_chain_entry等) — skill 実行をサイレントにブロックしていた過剰エンジニアリングの排除_broadcast_loaded_skillブラックボード書き込み簡素化:16 フィールド → 6 フィールド_loaded_skills_prompt_hint簡素化:約 350 tokens → 約 120 tokensmax_tokens=120)、モデルがタスクタイプに応じて自律的に skill を判断TodoWriteを使用可能新機能
_preload_skills_from_plan_steps:plan step テキストの skill 名前を事前スキャンして先行プリロード。Shell パス自動クォート(
_rewrite_shell_virtual_paths):スペースを含むパスを自動でダブルクォートで囲み、macOS のパス実行失敗を修正。Plan 拡張:
2. デュアル RAG アーキテクチャ — Code RAG + Data RAG(高)
アーキテクチャ:TF_G_IDF_RAG 上に構築された 2 つの独立した取り込み・検索エンジン:
RAGIngestionService(Data RAG):ドキュメント、PDF、構造化データファイル — 汎用知識ベースCodeIngestionService(Code RAG):コードファイル専用、コード認識トークナイザー統一検索:
query_knowledge_library(query, top_k)が両ライブラリを並列検索し、マージされたランク付き結果を返す。RAG ガイド注入:
research-orchestrator-proとscientific-reasoning-lab両方に、query_knowledge_libraryインターフェース、パラメータ、レスポンス形式、ベストプラクティスクエリパターンを説明する完全な RAG 検索ガイドを内蔵。3. 内蔵 Skills 全面リライト:research-orchestrator-pro & scientific-reasoning-lab(高)
research-orchestrator-proを協調型決定ハブとしてリライト:scientific-reasoning-labを 5 ステップ自己反復推論エンジンとしてリライト:4. 多因子優先度コンテキスト圧縮(高)
問題:以前の
_auto_compactは時系列順(古いものから)でメッセージを削除し、重要なタスク情報(現在の plan step、最近のエラー、アクティブ skills)を失う可能性があった。新
_classify_message_priority— 10 因子スコアリング(0–10):TodoWrite、plan_step、finish_task):+2Error:、例外トレース):+2compact-resumeノート:強制的に 10(常に保持)新
_priority_compress_messages— 優先度ベース 3 段階圧縮:_build_state_handoff強化:PLAN_PROGRESS、CURRENT_STEP、ACTIVE_SKILLS、RECENT_TOOLS フィールドを追加。_auto_compact統合:優先度圧縮を先行実行;元の時系列pop(0)はセーフティフォールバックとして保持。5. Anti-stall メカニズム最適化(中)
変更:
6. 重大バグ修正(高)
修正 1 —
CodeIngestionService._flush_lock(AttributeError)CodeIngestionService.__init__が親クラスRAGIngestionService.__init__を完全に上書きしsuper().__init__()を呼ばないため、親で初期化される_flush_lockが存在しないCodeIngestionService.__init__の末尾にself._flush_lock = threading.Lock()を追加修正 2 — フロントエンド
setTaskLevel()複雑度セレクターリセットsetTaskLevel()がupdateLevelBtn(lvl)を呼ぶがscheduleSnapshot()を呼ばないため、次の SSE スナップショット更新で UI が古いサーバーデータで上書きされるupdateLevelBtn(lvl)の後にscheduleSnapshot({forceFull:false, delayMs:80, allowWhenFrozen:true})を追加修正 3 —
_sync_todos_from_blackboardが Worker TodoWrite アイテムを消失owner ∈ {developer, explorer, reviewer}のアイテムがnon_system_rowsからフィルタリングされるがsystem_rowsにも追加されず、毎回のブラックボード同期でサイレントに消失worker_rowsリストに収集して優先的にマージ修正 4 — Anti-stall 閾値と指示(第 5 項参照)
修正 5 — 多因子コンテキスト圧縮(第 4 項参照)
2026-03-25 修正サマリー
query_knowledge_library、内蔵 skills への RAG 検索ガイド注入research-orchestrator-proを非干渉の分析ハブとして再設計;scientific-reasoning-labを 5 フェーズ自己反復推論エンジンとしてリビルド_flush_lockAttributeError、タスクレベルセレクター UI リバート、worker todo 保護、および上記 2 項This discussion was created from the release Clouds Coder 2026.03.30-Stable.
Beta Was this translation helpful? Give feedback.
All reactions