Everything agent.
All related.

From the release notes

Release v4.7.9

From the release notes

Initial release release(typesafe): bump to 0.0.1a2 (#40576) feat(typesafe): experimental AutoModeMiddleware (#40545) feat(typesafe): experimental ModelRouterMiddleware (#40543) fix(typesafe): trace usage metadata (#40570) feat(typesafe): TypeSafeClassifier (#40542)

From the release notes

What changed Fixes Finding B of #3327: memory_search({ smart: true }) was reporting SmartRetrieval's internal RRF fusion score in the similarity field instead of the underlying retrieval relevance score — e.g. a raw retrieval relevance of 0.8 was showing up as 0.03252247488101534. similarity now carries the raw retrieval relevance (matches smart: false behavior) the fused SmartRetrieval ranking score is now exposed separately as rankingScore threshold behavior is unchanged (still applied pre-RRF/MMR/recency) 14 new regression tests across the SmartRetrieval package and MCP handler Fixed in #3340 (thanks @takagibit18). Finding A of #3327 (reasoningBank/vectorBackend reporting enabled:true but routing through bridge-fallback) remains open and unclaimed. Note on version history 3.42.1–3.42.3 were published to npm without matching version-bump commits on main. Verified their tags (v3.42.0, v3.42.1, v3.42.3) are all ancestors of this release, so 3.42.4 is a strict superset of everything previously shipped. Co-Authored-By: RuFlo ruv@ruv.net

From the release notes

Initial release feat(typesafe): TypeSafeClassifier (#40542)

From the release notes

cut the revison v2.1.0-rc0

From the release notes

Release nightly created from 1ff3d96 at 2026-09-17 21:18:32+08:00

From the release notes

fix(api): align Go proxy contract parity (#19723)

From the release notes

Changes since langchain==1.4.0 release(langchain): 1.4.1 (#40498) fix(langchain): preserve open MCP object arguments (#40414) fix(langchain): correct InterruptOnConfig documentation (#40140)

From the release notes

Fixes hooks: quote-aware split for RUFLO_HOOK_CLI_OVERRIDE (#3344) — the test-only CLI-override parser in plugins/ruflo-core/scripts/ruflo-hook.cjs did a naive .split(' '), which broke process.execPath into C:\Program plus stray tokens on a standard Windows Node install (C:\Program Files\nodejs\node.exe). Found while verifying #3332's Windows argv fix on a real Windows machine rather than windows-latest CI (whose hosted Node path has no space). No production impact — RUFLO_HOOK_CLI_OVERRIDE is never set outside the test harness. release: spawn npm.cmd with shell:true on Windows (#3346, #3348) — scripts/stage-internal-runtime-bundles.mjs and scripts/prepare-root-publish.mjs both spawned npm.cmd directly without shell:true, which throws EINVAL on any current Node/Windows combo (CreateProcess can't launch a .cmd, and Node has refused to shell out to one implicitly since CVE-2024-27980). This blocked every Windows publish of @claude-flow/cli and the claude-flow umbrella package. Found live while publishing this release from a Windows machine. Verification plugins/ruflo-core/scripts/test-hooks.mjs: 32/32 (was 29/32 before #3344) v3/@claude-flow/cli/__tests__/ruflo-hook-windows-argv.test.ts, escape-cmd-arg.test.cjs, mcp-launch.test.cjs, scripts/smoke-ruflo-hook-cjs.mjs, cross-platform/env-var audits: all clean Full Windows publish pipeline (@claude-flow/cli → claude-flow → ruflo) exercised end-to-end on a real Windows machine after the fixes landed latest / alpha / v3alpha dist-tags verified aligned at 3.42.3 across all three packages 🤖 Generated with RuFlo

From the release notes

Feature: Elixir alias/import/require/use targets now resolve onto the module's defmodule node across files, so the internal module dependency graph is no longer dropped as dangling. Only top-level modules are indexed (a nested defmodule, labeled with its bare inner name, cannot capture an unrelated use <Name> from another file), and a same-file reference is left unresolved so it cannot clobber the structural contains edge (#3603, thanks @ayushcodes10). Feature: a Rust self.method() call now resolves to a method defined on the same type in another file (the common split-impl-block layout), pooling methods across every impl of one type and refusing to link when two unrelated types share a bare name (#3602, thanks @ayushcodes10). Feature: a Ruby member call obj.foo on a known-type receiver now resolves to a method foo inherited from a superclass, including across files, using the same conservative promotion as the implicit-self resolver — a single owning class, matching method kind, and one unambiguous ancestry chain, or it stays dangling (#3585, thanks @oleksii-tumanov). Fix: every edge endpoint written to graph.json is now a declared node. An imports/imports_from/re_exports edge to an external module (stdlib, a third-party dependency) mints a typed external stub node instead of leaving a dangling endpoint that loaders materialise as an attribute-less phantom, and a cross-repo merge unifies the same external module into one global node instead of fragmenting it per repo; sourceless external call targets stay suppressed (#2873, #2878, thanks @AromalBiju1). Fix: a Markdown code-span mention edge now survives an incremental rebuild instead of being pruned as an unauthored link, and a dotted span (Foo.bar) resolves using its qualifiers as evidence, rejecting misleading matches such as time.sleep or pyproject.toml (#3587, thanks @AstroMined).

From the release notes

The PageIndex SDK, local or cloud — vectorless, reasoning-based RAG, end to end. Much faster indexing — the PageIndex Flash engine gets the tree from layout stats: no LLM involved for the structure generation itself, LLMs only write the node summaries, and tree expansion proposes a wave of nodes concurrently. client = PageIndexClient() client.submit_document("report.pdf") client.chat("What does the report conclude?") Index, to chat, to agent integration, one client. Local mode needs no server, no vector DB, no PageIndex API key. Highlights Flash engine: the local default (mode="standard" keeps the classic LLM pipeline). Embedded bookmarks are consumed when trustworthy, and tree optimization is on by default — optimize="merge" for the deterministic LLM-free pass, "full" (default) adds LLM expand, which runs a wave of nodes concurrently instead of one round-trip at a time. One complete surface, local and cloud: PageIndexLocalClient(storage_path=...) is the same client as cloud — submit, tree, page content, chat, and agent tools all present in both modes, so code moves between them unchanged. Cloud documents, your own model: api_key decides where your documents live; a configured chat model decides who answers — and the two combine. PageIndexClient(api_key="pi-...", chat_model="openai/gpt-5.2") runs the same in-process document-QA engine over the live cloud tool set. Page content flows through your process to your provider on your credentials; doc_id targets at the prompt level; enable_citations stays with the managed chat. Agent integration: the cloud MCP tool contract, in-process — client.agent_tools() (plain functions), as_openai_tools(), as_anthropic_tools(), as_claude_mcp(), plus one-call openai_agent_config() / anthropic_runner_config() / claude_agent_config() bundles and agent_instructions() for the system prompt. Cloud clients get the live server tool set over the MCP bridge (read-only endpoint by default); local clients get the in-process subset with the same schemas and envelopes — agent prompts port unchanged. Chat surfaces: chat() — question in, answer out, on any backend; chat(stream=True) shows the run as it happens, thinking and tool calls woven into the text, or as typed events via .events; chat(protocol="responses" | "messages") drives the OpenAI Responses or Anthropic Messages API natively with that protocol's own shapes, and chat_completions() keeps the OpenAI-compatible envelope — all with doc_id targeting, streaming, honest usage accounting, and prompt-cache continuity across turns. Transcripts append verbatim: a protocol lane's output goes back into the next request unchanged. Model & connection knobs: index_model / chat_model, index_backend / chat_backend (and per-call backend) passed verbatim to each lane — LiteLLM-routed providers, keyless OpenAI-compatible servers, Azure/Bedrock/Vertex included. index= / chat= slots: the grouped spelling of the flat arguments — a string shorthand or a mapping (index={"model": ..

From the release notes

🚀 Nexent：开源智能体平台 / Nexent: Open Source Intelligent Agent Platform 我们很高兴地宣布Nexent v2.6.0 正式发布！🎉 Nexent 是一个开源智能体平台，能够将流程的自然语言转化为完整的多模态智能体 —— 无需编排，无需复杂拖拉拽。基于 MCP 工具生态，Nexent 提供强大的模型集成、数据处理、知识库管理、零代码智能体开发能力。我们的目标很简单：将数据、模型和工具整合到一个智能中心中，使日常工作流程更智能、更互联。 We are excited to announce that Nexent v2.6.0 is released! 🎉 Nexent is an open-source agent platform that turns process-level natural language into complete multimodal agents — no diagrams, no wiring. Built on the MCP tool ecosystem, Nexent provides model integration, data processing, knowledge-base management, and zero-code agent development. Our goal is simple: to bring data, models, and tools together in one smart hub, making daily workflows smarter and more connected. 新功能 / New Features 重构模型管理页面：统一模型编辑对话框，新增模型容量自动推荐（内置 LiteLLM 模型目录、离线可用）、自定义参数校验与推理参数配置。 / Revamped the model management page: unified model edit dialog, automatic model capacity suggestions (bundled LiteLLM catalog, works offline), custom parameter validation, and inference parameter configuration. 新增完整会话线程（Thread）管理能力，覆盖创建、切换与并发控制，并配套遥测指标。 / Added full conversation thread management, covering creation, switching, and concurrency control, with telemetry support. 上线外部记忆体系：智能体可在最终回答后自动提取记忆，支持外部记忆提供方检索（top_k、超时可配置），并提供 AIDP Memory 插件与插件热加载。 / Introduced the external memory system: agents automatically extract memories after final answers, with configurable external provider search (top_k, timeout), an AIDP Memory plugin, and hot-reloadable plugins. 优化 Agent 创建体验：NL2Agent 生成流程与落地卡片打磨，新增 NL2Agent 优化建议，提示词同步升级。 / Improved the agent creation experience: polished NL2Agent generation flow and landing cards, with new NL2Agent optimization suggestions and upgraded prompts. 检索引用交互升级：回答中的引用支持统一悬浮卡片与点击定位高亮。 / Upgraded retrieval citations: references in answers now support unified hover cards with click-to-focus highlighting. 重构统一标签（Label）系统，支持标签库权限、资源打标与筛选，并为多版本 Agent 提供协作版本标签。 / Refactored the unified label system with tag library permissions, resource tagging, and filtering, plus collaborative version labels for multi-version agents. 支持模型优先级排序，可按优先级为智能体选择模型。 / Added model priority sorting so agents can select models by priority. 沙箱适配 Anthropic Skills，可直接运行 Anthropic 技能格式。 / Adapted the sandbox to run Anthropic skills natively. S3 上传/下载工具支持用户自定义配置。 / The upload_to_s3 and download_from_s3 tools now support user configuration. 日志支持落盘为文件，便于排查与归档；服务启动时自动恢复中断的后台任务。 / Logs are now written to disk as files for easier troubleshooting, and interrupted background tasks are automatically recovered on service startup. Bug 修复 / Bug Fixes 完成上下文预算契约迁移，增加有界压缩与真实溢出恢复，长对话不再丢失上下文。 / Completed the context budget contract migration with bounded compaction and real overflow recovery, preventing context loss in long conversations. 修复会话链路多处问题：返回后恢复会话、线程切换保留智能体、会话标题使用会话模型生成、正确渲染执行代码与对比模式下的历史保留。 / Fixed several conversation-flow issues: conversation restore after returning, agent retention across thread switches, session-model-based title generation, correc

From the release notes

v3.42.1 — SONA env-default, review batch, and reliability fixes A patch release: one small feature (opt-in, fully backward-compatible) plus a batch of independently-reviewed bug fixes, a performance fix, and a dependency bump — all reviewed, tested, and merged in this pass. What's new for you SONA learning mode can now default from the environment RUFLO_INTELLIGENCE_MODE sets a fleet-wide default learning mode (real-time, balanced, research, edge, batch) without threading mode/sonaMode through every call site — useful for a managed deployment that wants every session to learn in a specific profile. An explicit per-call mode always wins; an unset or unrecognised value falls through to balanced, so a typo can never silently pick a profile. (Originally landed with a bug where one of the two places reading this variable only checked it once per process instead of per session — fixed before release, with new regression tests in both places.) Diagnostics and memory fixes ruflo doctor no longer guesses your memory driver from a raw table count (a heuristic that never actually correlated with which driver you're using) — it now reports only what it verifies directly. ruflo doctor now resolves the installed MetaHarness version correctly instead of reporting stale/incorrect data. memory_stats now answers from the live store instead of a snapshot that couldn't see in-flight writes — a working memory store no longer reports itself as uninitialized. Also fixes a namespace named __proto__ silently vanishing from the breakdown, and a hardcoded row cap that was truncating large stores without saying so. A database handle used during memory initialization now closes on every code path, including the failure path that used to leak it. Opening the graph database now verifies the handle is actually persistent before using it — a bad argument shape used to silently produce an in-memory-only graph that vanished on exit, with no error. Fixed a memory-key bug on Windows where a drive letter's colon wasn't sanitized consistently with how Claude Code derives the same key. Federation and security The federation gateway's untrusted-relay data is now read through the same validated envelope shape the gateway actually sends, closing a gap where some tools could miscount or silently see zero messages. WITNESS_ED25519_ROOT now actually resolves @noble/ed25519 for witness-manifest signing (this only affects the maintainer signing path — the separate node:crypto-based signature verification path was never affected). Performance AgentBBS now dedupes redundant peer-hello and watch-replay traffic, and computing the next sequence number no longer requires scanning an entire (potentially large) room log. Housekeeping swarm status/swarm stop now report a real swarm id and no longer get confused by a stale "stopped" state file shadowing a live swarm. Bumped agent-browser to 3.37.1 for recording/WebMCP improvements and a Windows headless-Chrome cleanup fix. Upgrading

From the release notes

fix(bedrock): report the expected key schema instead of a JSONDecodeE…

From the release notes

Feature: Terraform module calls with a literal local source (./… or ../…) now resolve to a directory-scoped module node, exposing the caller→implementation topology (e.g. environment → application → base); remote and registry sources and source expressions are left unresolved and never fabricate a target. After upgrading an existing graph, run graphify update . once to regenerate Terraform ids and topology (#3571, thanks @vstepko). Feature: a bare implicit-self call in a Ruby method now resolves up the inheritance chain to a method defined on a superclass, including across files. The promotion is conservative — it only upgrades an existing inferred edge when a single owning class, a matching method kind, and one unambiguous ancestry chain all agree, and it bails on mixins, reopened classes, and dynamic superclasses rather than guess (#3569, thanks @oleksii-tumanov). Feature: an inline code span in a Markdown file (e.g. `Widget.render`) now emits a references edge to the symbol it names, but only when exactly one corpus callable matches — generic words and ambiguous names link to nothing, so docs join the graph without a false-edge firehose (#3562, thanks @AstroMined). Fix: a prose file whose stem is the bare plural "tokens" (e.g. TOKENS.md) is no longer excluded as a credential dump; singular token.md and other keywords' bare plurals still exclude, and content-based secret detection is unchanged (#3527, thanks @HARSHAVARDHAN-RAJU5). Fix: a call whose only resolution target is a sourceless external stub no longer emits a spurious calls edge, so an unresolved external name stops accruing a phantom god-node (#3156, thanks @DevChiniwala). Fix: calls made inside a module-level anonymous closure (an IIFE or a top-level callback) are now attributed to the file node instead of being dropped (#3124, thanks @ayushcodes10). Fix: cross-repo merge now respects C# namespace/using scoping — a parked member call binds to a same-named type in another repo only when the caller's namespace or an in-scope using resolves it, so a third-party receiver no longer binds to an unrelated repo-local type (#3360, thanks @DevChiniwala). Fix: a bare Lua require("mod") statement (no assignment) now emits an imports edge, including the chained require("lazy").setup({}) LazyVim idiom, so Neovim configs graph correctly instead of coming out as disconnected files (#3320, thanks @A-Levin). Fix: the Claude/CodeBuddy PreToolUse guard hooks now carry a timeout, so a wedged filesystem stat can no longer stall a tool call indefinitely (#3314, thanks @ayushcodes10). Fix: id-prefix reconstruction is now Unicode-aware — a non-Latin path segment is slugified with the canonical normalizer instead of being dropped, so files under non-ASCII directories keep stable ids (#3559, thanks @ayushcodes10). Fix: cohesion_score now excludes self-loops (e.g. a recursive call) from the edge count, so a community's score stays within 0..1 instead of being inflated above 1 (#3558, thanks @jordanale

From the release notes

v3.42.0 — Swarm, Memory & Security Reliability Batch This release lands 14 fixes produced by Ruflo's own nightly "Dream Cycle" research pipeline — each one found a real bug or gap in the codebase, built a fix, proved it with before/after tests, and (in several cases) survived an independent adversarial review before merging. Nothing here is speculative: every change below has a passing regression test that failed before the fix and passes after it. What's new for you 🐝 Swarm consensus actually uses trust weights now When multiple agents vote on a decision, Ruflo has long computed a "trust weight" for each voter — but that number was silently discarded before it ever reached the vote count. Every strategy (Raft, Byzantine, gossip) was voting as if every agent were equally trustworthy. Now the weight is actually used, so a vote from an agent with a strong track record counts for more than a vote from an untested one. 🔍 Vector search is faster and more accurate HNSW product-quantization (a memory-compression technique for embeddings) was built and tested but never actually wired into the search path — searches were running on meaningless data. It's now dispatched correctly, more than doubling search accuracy (recall) in our benchmark. MMR re-ranking (used to diversify search results so you don't get 10 near-duplicate hits) was re-tokenizing every candidate on every pass, even when it didn't need to. It's now ~6x faster with byte-identical results. Memory deduplication used to only catch exact byte-for-byte duplicates. It now also catches near-duplicates (paraphrases, reformatted text) using the embeddings Ruflo already computes — no extra cost, same-quality matching. Fixed a bug where a memory pattern's retrieval score (how well it matched your query) was overwriting its learned reliability (how trustworthy that pattern has proven to be over time) — which meant the "only trust high-confidence patterns" safety check was silently checking the wrong number. 🔐 Security hardening MCP tool governance is now enforceable. Ruflo has shipped a governance policy file (audit logging, per-session call limits) for a while, but nothing in the running server ever actually read it. It's now wired in as an opt-in flag (RUFLO_MCP_ENFORCE_POLICY=1) — off by default, so nothing changes unless you turn it on. The call-limit for that policy now resets properly. It used to be a lifetime cap that, once hit, locked a session out permanently until restart. It's now a rolling time window, matching how every other rate limiter (Cloudflare's, etc.) does it. Hive-mind voting can no longer be Sybil-attacked. Previously, any caller could cast a vote under a made-up agent name and it would count toward quorum — meaning one caller could out-vote everyone else with fabricated identities. Votes are now checked against the actual roster of agents that joined the hive. Joining or leaving a hive-mind now requires proof you're allowed to. This closes the gap the vote-fix abov

From the release notes

Fix agentdb_pattern-store/agentdb_pattern-search now surface degraded: true whenever the response comes from anything other than a healthy reasoningBank controller (the bridge-fallback and memory-store-fallback paths alike), instead of only in a note string a caller has no structural reason to check. Fixes #3288. Includes a follow-up correction (b9ca64f80) after review caught that the initial fix only handled the case where the controller registry is entirely absent, missing the more common case where the registry exists but ReasoningBank itself is unusable (controller: 'bridge-fallback') — the degraded flag is now applied at the result.controller === 'reasoningBank' boundary in both handlers rather than per fallback-label, so it can't be reopened by a future fallback label. See #3323 for the full discussion. This does not address why vectorBackend/reasoningBank fail to activate in a given deployment — that's tracked separately in #1228 (ADR-053) and #2296.

From the release notes

Release v4.7.8

From the release notes

Security fix Federation event identity: fetchRecent/fetchManyOn/fetchChannel now apply the signature-verified id/pubkey/created_at after the publisher's own message content, instead of before — a publisher could otherwise put those keys in their own content and overwrite their verified identity downstream, which the swarm claims ledger (reduceClaims) trusts for release/handoff authorization. Same fix applied to x_federation_channel_read in @claude-flow/cli, which additionally now verifies event signatures at all (reqEvents() previously trusted unverified events). The live ruflo-x-gateway production service was redeployed with this fix ahead of this release. Windows hook argument escaping: ruflo-hook.cjs's Windows shell:true npm-shim invocation path now escapes every argument (cmd.exe metacharacter neutralization), so a hook-derived value can no longer be reinterpreted as a separate command or redirection. See #3322 for the full technical writeup and test coverage.

From the release notes

🌐 English | 中文 This is a refinement release for the 2.1.8 multi-Agent version, building on it with a series of improvements and fixes across multi-Agent collaboration, model capabilities, and channel integration. 🤝 Multi-Agent Experience Further polish around using multi-Agent teams: Multi-Agent session selection: the in-session multi-Agent selector adds a lead-Agent marker that clearly separates the lead Agent from members, while simplifying the view in single-Agent setups. Collaboration fix: fixed failures when delegating tasks from non-lead Agents in multi-Agent group chats, now correctly identifying member Agents. Member Agent image rendering: fixed images generated by non-default Agents failing to render on the Web console and desktop client. Group-chat entry: the desktop Team page adds a group-chat entry, so you can start a group session directly from an Agent team. 📡 Channel Integration Fixes Fixed key issues along the console's channel start/stop path for more stable channel integration: Desktop channel integration fix: fixed channels failing to start in the desktop client. Channel stop fix: fixed a lost app_module lookup during the channel stop flow that caused disconnect errors (a regression introduced by #3141). QQ channel reconnection: fixed the QQ channel not reconnecting after a silent WebSocket disconnect; it now reconnects automatically. 🤖 Model Capabilities Model list configuration: model providers can now configure a model list, setting the model name, model type, context window, max output, and more. Thanks @alanyz106 (#3146) Multiple fallback models: when the primary model fails, CowAgent now tries an ordered fallback chain in turn, replacing the previous single fallback model. Thanks @alanyz106 (#3125) Fallback credential fix: fixed credential resolution during fallback, which now resolves credentials by the provider actually routed to rather than the global configuration. Thanks @alanyz106 (#3142) New image models: added gpt-image-2.5-flare and gpt-image-2.5-sunburst. Thanks @cowagent (#3140) 🛠 Improvements & Fixes Config file tolerance: tolerate a UTF-8 BOM in config.json, and stop the pre-login 401 polling to avoid invalid requests before sign-in. Image generation fixes: fixed Baidu Ernie IMAGE_CREATE crashing when unsupported, now returning an error message instead (Thanks @c020627 #3144). fixed Midjourney reply retries not recursing into the send logic correctly (Thanks @c020627 #3151). Voice fix: fixed silk voice transcoding by first decoding silk into a standalone wav and then re-encoding to mp3, avoiding overwriting the source file (Thanks @c020627 #3149). Media URL classification: determine media type from the resolved URL path, fixing misclassified media links (Thanks @c020627 #3150). Baidu Translate: fixed Baidu Translate not reporting the API error after all retries failed (Thanks @c020627 #3147). Background command detection: restrict run_in_background to long-running processes, preventi

From the release notes

v2026.09.14 — 三語正式版 / Trilingual release 繁體中文 reader-ux 三語 README 移除動畫 banner 下方重複的靜態圖與文件站捷徑，讓首頁入口更乾淨。動畫停止時仍保留完整構圖；文件站的減少動態、無 JavaScript 靜態 PNG 與三語 PDF fallback 都繼續保留。 來源 1 · 來源 2 maintenance 修正 Stage 6 五組範例的三語收合區塊，以及 CLI 工具表與 Stage 3 摘要中的行內程式碼：粗體、清單、行內程式碼與連結現在會由 MkDocs 正確解析，不再把 **Context** 或反引號這類 Markdown 記號直接顯示給讀者。新增 post-build gate，會檢查實際 HTML，而不是只看 .md 原始碼；程式碼範例中的記號仍被允許。 來源 1 learning-content Stage 7 改為「Agent 上線工程：可測、可看、可停、可恢復」。同一個查資料案例依序解釋 Harness、Loop、Graph、Eval、Observability、Approval、Checkpoint、Resume 與 Recovery；完整 Eval Case 明確包含 input、初始狀態、成功條件、禁止行為、選用參考答案、評分方法與案例資訊，避免把 Golden Set 誤解成只有輸入。 來源 1 learning-content Stage 7.5 改成「有證據才增加複雜度」的進階選讀，不再重教 Stage 7 基礎。可見核心詞表先用白話說明 Evaluator–Optimizer／Agent-as-Judge、Failure Injection／Chaos Eval、Autonomy Gradients／Trust Layers 與 Model–Harness Fit，再用兩組三語亮色圖協助讀者依失敗證據選擇最小做法。 來源 1 简体中文 reader-ux 三语 README 移除动画 banner 下方重复的静态图与文档站快捷入口，让首页入口更干净。动画停止时仍保留完整构图；文档站的减少动态、无 JavaScript 静态 PNG 与三语 PDF fallback 都继续保留。 來源 1 · 來源 2 maintenance 修正 Stage 6 五组示例的三语折叠区块，以及 CLI 工具表和 Stage 3 摘要中的行内代码：粗体、列表、行内代码与链接现在会由 MkDocs 正确解析，不再把 **Context** 或反引号这类 Markdown 标记直接显示给读者。新增 post-build gate，会检查实际 HTML，而不是只看 .md 源码；代码示例中的标记仍被允许。 來源 1 learning-content Stage 7 改为“Agent 上线工程：可测、可看、可停、可恢复”。同一个资料查询案例依次解释 Harness、Loop、Graph、Eval、Observability、Approval、Checkpoint、Resume 与 Recovery；完整 Eval Case 明确包含 input、初始状态、成功条件、禁止行为、可选参考答案、评分方法与案例信息，避免把 Golden Set 误解为只有输入。 來源 1 learning-content Stage 7.5 改成“有证据才增加复杂度”的进阶选读，不再重复讲解 Stage 7 基础。可见核心词表先用白话说明 Evaluator–Optimizer／Agent-as-Judge、Failure Injection／Chaos Eval、Autonomy Gradients／Trust Layers 与 Model–Harness Fit，再用两组三语亮色图帮助读者根据失败证据选择最小做法。 來源 1 English reader-ux The three READMEs remove the duplicate static-image and documentation-site shortcuts below the animated banner for a cleaner entry point. The complete composition remains visible when animation stops, while reduced-motion, no-JavaScript PNG, and trilingual PDF fallbacks remain available. source 1 · source 2 maintenance Fixed the trilingual collapsible sections in five Stage 6 examples, plus inline code in the CLI tool table and Stage 3 summaries. MkDocs now renders emphasis, lists, inline code, and links instead of showing Markdown markers such as **Context** or backticks to readers. A new post-build gate checks the actual HTML rather than source .md alone, while still allowing markers inside code examples. source 1 learning-content Stage 7 is now “Production Engineering for Agents: Testable, Observable, Stoppable, and Recoverable.” One research-assistant example progressively explains Harness, Loop, Graph, Eval, Observability, Approval, Checkpoint, Resume, and Recovery. A complete Eval Case explicitly includes input, initial state, success criteria, forbidden actions, an optional reference answer, grading method, and case metadata, so a Golden Set is not mistaken for inputs alone. source 1 learning-content Stage 7.5 is now an advanced reading map that adds complexity only when evidence justifies it, without reteaching Sta

From the release notes

Summary Removed unnecessary beartype hooks from three package imports. This prevents unwanted import-time instrumentation and dependency coupling. based on #15258

From the release notes

v2026.09.13 — 三語正式版 / Trilingual release 繁體中文 maintenance 發布前完整連結掃描確認 YIELD INTELLIGENCE 的 hosted MCP、替代路徑與 pricing 入口都回傳 404。三語精選目錄已移除這筆目前無法使用的服務與 repository snapshot，不以 README 仍寫著「live」取代實際端點證據。 來源 1 reader-ux 首頁下方的角色分流圖加入圖示動畫試版：只讓五個代表性角色圖示依序輕動，並支援停止／播放、減少動態與靜態 PNG。這組圖沿用三張舊原圖，因此圖中文字尚未完全對齊；英文圖的 Analysts 也比正文的 Knowledge workers 範圍窄。正式校正前，請以圖片下方保持展開的角色表為準。 來源 1 governance 貢獻規則改成繁中先定稿、英文與簡中在同一個 PR 同步。三語的概念、URL、數字、評分與安全限制必須一致；新增資源不再以會變動的 GitHub stars 當門檻，改用來源、維護狀態與教學價值判斷。 來源 1 learning-content Stage 7 新增可見的 Eval 基礎：Golden Set、Reference Set、Trial、Grader、Baseline、Regression 與 Holdout Set 都先用白話解釋，再保留正式名稱。三語同版圖把「定義成功、建立案例、重複測試、比較基準、人工複查」串成一個證據迴圈。 來源 1 examples Stage 7 的 Eval 範例改用 8 題版本化 JSON，分成 5 題開發案例與 3 題凍結 holdout。Ollama 與 Anthropic 共用同一個 runner，可選多次 trial、保存報告並和 baseline 比較；資料格式不相容或空輸出會直接失敗，離線測試不讀 API key。 來源 1 简体中文 maintenance 发布前完整链接扫描确认 YIELD INTELLIGENCE 的 hosted MCP、替代路径与 pricing 入口都返回 404。三语精选目录已移除这项目前无法使用的服务与 repository snapshot，不以 README 仍写着“live”取代实际端点证据。 來源 1 reader-ux 首页下方的角色分流图加入图标动画试版：只让五个代表性角色图标依次轻动，并支持停止／播放、减少动态与静态 PNG。这组图沿用三张旧原图，因此图中文字尚未完全对齐；英文图的 Analysts 也比正文的 Knowledge workers 范围窄。正式校正前，请以图片下方保持展开的角色表为准。 來源 1 governance 贡献规则改为繁中先定稿、英文与简中在同一个 PR 同步。三语的概念、URL、数字、评分与安全限制必须一致；新增资源不再以会变动的 GitHub stars 作为门槛，而是根据来源、维护状态与教学价值判断。 來源 1 learning-content Stage 7 新增可见的 Eval 基础：Golden Set、Reference Set、Trial、Grader、Baseline、Regression 与 Holdout Set 都先用白话解释，再保留正式名称。三语同版图把“定义成功、建立案例、重复测试、比较基准、人工复查”串成一个证据循环。 來源 1 examples Stage 7 的 Eval 范例改用 8 题版本化 JSON，分为 5 题开发案例与 3 题冻结 holdout。Ollama 与 Anthropic 共用同一个 runner，可选择多次 trial、保存报告并与 baseline 比较；数据格式不兼容或空输出会直接失败，离线测试不读取 API key。 來源 1 English maintenance The full pre-release link scan confirmed that YIELD INTELLIGENCE's hosted MCP endpoint, alternate route, and pricing endpoint all return 404. The unavailable service and its repository snapshot were removed from all three curated catalogs; a README that still says “live” does not outweigh direct endpoint evidence. source 1 reader-ux The role map below the homepage now has a pilot icon animation. Only the five representative role icons move in sequence, with stop/play controls, reduced-motion support, and static PNGs. Because it reuses three older source images, wording inside the images is not yet fully aligned; the English image's “Analysts” label is also narrower than “Knowledge workers” in the text. Until the artwork is corrected, use the always-visible role table below it as the canonical guide. source 1 governance Contribution rules now finalize Traditional Chinese first and synchronize English and Simplified Chinese in the same PR. Concepts, URLs, numbers, ratings, and safety limits must match across locales; new resources are judged by sources, maintenance, and teaching value instead of changing GitHub star counts. source 1 learning-content Stage 7 now teaches visible Eval foundations. Golden Set, Reference Set, Trial, Grader, Baseline, Regression, and Holdout Set are explained in plain language before retaining their formal names

From the release notes

The PageIndex SDK, local or cloud — vectorless, reasoning-based RAG, end to end. Much faster indexing — the PageIndex Flash engine gets the tree from layout stats: no LLM involved for the structure generation itself, LLMs only write the node summaries, and tree expansion proposes a wave of nodes concurrently. client = PageIndexClient() client.submit_document("report.pdf") client.chat("What does the report conclude?") Index, to chat, to agent integration, one client. Local mode needs no server, no vector DB, no PageIndex API key. Highlights Flash engine: the local default (mode="standard" keeps the classic LLM pipeline). Embedded bookmarks are consumed when trustworthy, and tree optimization is on by default — optimize="merge" for the deterministic LLM-free pass, "full" (default) adds LLM expand, which runs a wave of nodes concurrently instead of one round-trip at a time. One complete surface, local and cloud: PageIndexLocalClient(storage_path=...) is the same client as cloud — submit, tree, page content, chat, and agent tools all present in both modes, so code moves between them unchanged. Cloud documents, your own model: api_key decides where your documents live; a configured chat model decides who answers — and the two combine. PageIndexClient(api_key="pi-...", chat_model="openai/gpt-5.2") runs the same in-process document-QA engine over the live cloud tool set. Page content flows through your process to your provider on your credentials; doc_id targets at the prompt level; enable_citations stays with the managed chat. Agent integration: the cloud MCP tool contract, in-process — client.agent_tools() (plain functions), as_openai_tools(), as_anthropic_tools(), as_claude_mcp(), plus one-call openai_agent_config() / anthropic_runner_config() / claude_agent_config() bundles and agent_instructions() for the system prompt. Cloud clients get the live server tool set over the MCP bridge (read-only endpoint by default); local clients get the in-process subset with the same schemas and envelopes — agent prompts port unchanged. Chat surfaces: chat() — question in, answer out, on any backend; chat(stream=True) shows the run as it happens, thinking and tool calls woven into the text, or as typed events via .events; chat(protocol="responses" | "messages") drives the OpenAI Responses or Anthropic Messages API natively with that protocol's own shapes, and chat_completions() keeps the OpenAI-compatible envelope — all with doc_id targeting, streaming, honest usage accounting, and prompt-cache continuity across turns. Transcripts append verbatim: a protocol lane's output goes back into the next request unchanged. Model & connection knobs: index_model / chat_model, index_backend / chat_backend (and per-call backend) passed verbatim to each lane — LiteLLM-routed providers, keyless OpenAI-compatible servers, Azure/Bedrock/Vertex included. index= / chat= slots: the grouped spelling of the flat arguments — a string shorthand or a mapping (index={"model": ..

From the release notes

What's Changed Custom target connector sink API changes Container sinks now fulfill child target providers through per-action child slots (TargetActionSink.from_fn_with_children / from_async_fn_with_children); returning a ChildTargetDef list is deprecated, and in the Rust SDK that closure shape moved to from_async_fn_with_child_defs by @georgeh0 in #2407 Sinks whose callbacks compare equal but differ in type no longer share one sink identity, and callbacks that can't be weak-referenced (e.g. NamedTuple instances) now raise TypeError by @georgeh0 in #2404 fix(engine): confine a merged target-action batch failure to the failing component, instead of failing every component merged into the same sink call by @georgeh0 in #2406 Fixes for spurious re-execution of memoized functions Garbage-collecting an identical redefinition (re-run notebook cell, importlib.reload, function factory) no longer invalidates the live function's memo by @georgeh0 in #2405 copy.copy / copy.deepcopy of a coco function return the function itself, so dropping a copy no longer invalidates the original's memo by @georgeh0 in #2409 fix: component exception handlers receive the original typed Python exception instead of a wrapping RuntimeError by @hardness1020 in #2383 fix: preserve tuple and bytes component-path segments through state-store decoding, so target-state ownership for those components is no longer leaked or misattributed by @hardness1020 in #2403 Full Changelog: v1.0.22...v1.0.23

From the release notes

Fix: graphify.serve now imports cleanly on Python 3.12 and 3.13. The chinese extra pins jieba-py from 3.12 onward (0.9.60 mistakenly kept the old jieba until 3.14, and its invalid regex escapes are a hard error on 3.12+), and the jieba import now suppresses the tokenizer's SyntaxWarning regardless of message or line so it never escalates under -W error. Fix: the git hook's rebuild-root guard now rejects a symlink-loop or dangling .graphify_root on Python 3.13, whose Path.resolve() no longer raises on a loop — the saved root must resolve to a real directory inside the repo before it is adopted.

From the release notes

Fix: atomic writes now fall back correctly on Windows WinError 17 (cannot move to a different drive), not just PermissionError — a shared os_replace_with_fallback copies through a temp in the target directory and restores the original if the swap fails, keeping install/export/cache writes crash-safe (#3508, thanks @ayushcodes10). Fix: files that could not be classified into any language or type are now surfaced (a count and top extensions in the console and GRAPH_REPORT) instead of vanishing from a "successful" run; noise and ignored paths are unaffected (#3511, thanks @ayushcodes10). Fix: a C or C++ header with no trailing newline no longer trips a tree-sitter syntax error — a newline is appended before parsing when the source lacks one (#3513, thanks @imanolpg). Fix: Office and Workspace document sidecars converted into graphify-out/converted/ are no longer dropped when graphify-out/ is gitignored — the tool stops ignoring its own output (#3504, thanks @ayushcodes10). Fix: a call inside an exported function (export function f(){ g() }, export const f = () => g()) now resolves, including calls to aliased imports (import { x as y }; y()) (#3346, thanks @abhay-codes07). Fix: a package.json exports map is now resolved by importer platform with the runtime condition preferred over types, so cross-package edges no longer drop to a non-existent .d.ts (#3487, thanks @dajiaohuang). Feature: Python 3.14 is now supported — optional dependencies without 3.14 wheels are gated to drop-in replacements (jieba to jieba-py, graspologic to graspologic-native) by interpreter version, and vulnerable dependency floors (setuptools, pypdf, yt-dlp) were raised (#3490, thanks @taazbro).

From the release notes

Fix: an incremental rebuild no longer drops cross-file concept nodes from files it didn't touch — global dedup during a merge now protects existing nodes from untouched files instead of collapsing same-labeled ones across them (#3477, thanks @hopstreax). Fix: extraction now falls back to sequential in-process work when the process pool cannot start (e.g. POSIX semaphore exhaustion), instead of aborting the whole run (#3497, thanks @curtismu7). Performance: Python symbol resolution is roughly 47% faster — path resolution is memoized, each file parses once across both resolution passes, and the tree walk is iterative rather than recursive; extraction output is unchanged (#3500 / #3501 / #3502, thanks @abhay-codes07). Fix: graphify explain now accepts a path::Symbol form to disambiguate a symbol that shares its name with its file, and the ambiguity hint now shows a form the resolver actually accepts (#3485, thanks @ayushcodes10). Fix: the git hook now keeps its rebuild root inside the repository — a committed .graphify_root pointing outside the worktree is ignored and falls back to the repo top, so a checked-in marker can't steer the hook to scan or write outside the tree (#3265, thanks @ayushcodes10).

From the release notes

0.10.18 (2026-09-12) fix(providers): don't throw on reasoning-only /v1/responses output (e116f4ed5cbcecb35f0ddbb350ff2fa0ea42eec3)

From the release notes

0.10.17 (2026-09-11) fix(providers): route /v1/responses when system content is cache-marked blocks (82f028e55e23c6b5f03413be2bdff819fef76ca6)

From the release notes

Changes since langchain-core==1.6.2 release(core): 1.6.3 (#40407) feat(core): Allow model name and provider tracing metadata override based on gateway response (#40406) test(core): cover the deprecated .text() access path (#40243) docs(core): remove stale Args/Raises entries from FileCallbackHandler._write and ChatGeneration.set_text (#40211)

From the release notes

🌐 English | 中文 🤝 Multi-Agent Teams CowAgent is now a multi-Agent framework. You can create multiple Agents, each with its own role, to form a team and collaborate on complex tasks within a single session. Supported on the Web console, the desktop client, and channel integrations: Team management: create and maintain Agents, configure name, responsibilities, default model, skills, and knowledge, and copy the setup from an existing Agent. Resource isolation: each Agent has its own workspace, memory, and sessions; the knowledge base and skills can be shared or kept independent. Group collaboration: add multiple Agents to one session and use @ to direct a message to a specific member, with each Agent using its own configured model. Task delegation: a group chat has a lead Agent that can hand a task off to the member best suited for it based on their responsibilities, with support for multi-level delegation and configurable scope, depth, and timeout. Multiple channel instances: an IM channel of the same type can now run multiple instances, each bound to a single Agent or an Agent team. Supported on WeChat, WeCom smart bots, Feishu, DingTalk, QQ, Telegram, Slack, Discord, and more. Thanks @AaronZ345 (#2975, #2973), @zkjqd (#3096), @cowagent (#3118) Docs: Agent Teams ⏰ Scheduled Task Upgrades Scheduled tasks can now be created and managed from the console, with channel recipients and execution records added: Manual task creation: create, edit, and run tasks directly on the tasks page, and select the channel instance and recipient. Channel recipient management: maintain trusted recipients per channel instance, so a task can be delivered to a designated recipient rather than only the channel where it was created. Thanks @dajiaohuang (#3043) Execution records: a new execution-records page lets you review the history list and the details of each run. Thanks @mengluo04 (#3107) Concurrent-update fix: fixed tasks being lost on concurrent updates (Thanks @Whxuan0701 #3062). 📊 Context-Usage Visualization Context consumption is now visible and controllable, making token cost easier to manage: Context-usage view: a ring view replaces the previous clear-context button and shows a full breakdown of what makes up the context. Context actions: below the ring view you can trigger smart compaction, clear the context, or set the maximum context size. Context calculation: the usage breakdown is derived automatically from the model's reported usage, with agent_max_context_tokens available as a user-configurable cost cap. Thanks @chimyves (#3087), @sufan721 (#3079) 🤖 Model Capabilities New models: added deepseek-flash (DeepSeek V4.1 Flash), gpt-6-astra, claude-fable-5-1, qwen3.8-flash, glm-5.3-flash, gemini-3.8-flash, and the vision model deepseek-v4-flash-vision-exp (Thanks @a1094174619 #3094). Model fallback: configure a fallback model under Model Config > Primary Model, which takes over automatically when the primary model fails, preventing interr

From the release notes

Fix: a call to a Python function defined nested inside another function now resolves to that inner definition per lexical scope, instead of leaking to a same-named function elsewhere; direct recursion is preserved as a self-loop (#3410, thanks @hopstreax). Fix: submodule imports inside a PEP 420 namespace package (a directory with no __init__.py) now resolve to the target module instead of being dropped (#3429, thanks @flaukowski). Fix: a bare-name import of a module sitting next to the importing file (a flat script dir with no package) now resolves to that sibling — matching CPython's sys.path[0] behavior — without over-resolving a genuine third-party name (#3430, thanks @hopstreax). Fix: a PHP use Foo\Bar as Baz; import keys its local binding on the alias, so later Baz references resolve to the real class, and distinct aliased same-named classes stay separate (#3421, thanks @ayushcodes10). Fix: a directory literally named out is skipped as build output only when there is build-output evidence, so a real source out/ is no longer silently dropped (#3347, thanks @abhay-codes07). Fix: a shell script invoked in exec position through a variable path ("$SCRIPT_DIR/foo.sh") now resolves to the target when the variable holds a constant, matching the existing source handling (#3416, thanks @edwardselby). Fix: added the missing Iterable import in build.py so its type annotations resolve (they were an undefined name) (#3462, thanks @xiongjianxu). Fix: graphify global add no longer infers an empty repo tag for a path like /tmp/graph.json; it degrades to a sensible non-empty tag via the same helper merge-graphs uses (#3464, thanks @xiongjianxu). Feature: SQL extraction now emits index nodes for CREATE [UNIQUE] INDEX, linked to the table they index (#3467, thanks @L4XB). Fix: a TypeScript export * re-export no longer resolves a name to a same-named interface method — only module-level exports are candidates (#3436, thanks @L4XB). Feature: Rust module-level static and const declarations are now extracted as nodes (#3471, thanks @L4XB). Fix: an incremental rebuild no longer drops a node's rationale/summary when a semantic node's absolute source_file collides with its AST twin's relative path during dedup — source paths are normalized first (#3472, thanks @hopstreax). Fix: graphify install no longer aborts when the always-on registration target is unwritable (read-only or symlinked config); it skips that step with an actionable warning and still installs the skill (#3474, thanks @dajiaohuang). Fix: community labelling falls back to an installed claude CLI resolved at run time instead of pinning a path that expires (e.g. under snap/nvm), so labelling keeps working across updates (#3475, thanks @ktsang622). Fix: the all extra now includes psycopg[binary], so pip install 'graphifyy[all]' provides the postgres driver (#3482, thanks @L4XB).

From the release notes

v2026.09.10 — 三語正式版 / Trilingual release 繁體中文 maintenance 發布前掃描發現 notebooklm-skill 已封存，三語資源表與 Cookbook 改列歷史範例，不再推薦新流程使用。Graphthulhu 原入口回傳 404，移除失效連結並保留歷史名稱與評分；其授權目前無法重新核對。原編輯星級不變，沒有新增替代專案。 來源 1 reader-ux 首頁頂部 Banner 保留原版構圖、配色、文字與圖示，加入 18 秒循環動畫。先走共用基礎，再依序展示 A、B 路線；每語言 13 個圖示區域會做對應的小動作，包括 CLI 游標輸入、工具輕轉與 Hub 箭頭旋轉。文字與卡片固定，最後 2 秒保持靜止。 來源 1 accessibility 三語文件站提供停止／播放控制；系統偏好減少動態或停用 JavaScript 時使用靜態圖。README 保留靜態圖入口，三語 PDF 也使用完整 PNG。SVG 不載入外部圖片、字型或 JavaScript，圖片容量上限未提高。GitHub 嵌入圖片是否遵守真實系統的減少動態偏好仍未確認，需要靜態閱讀時請使用靜態圖入口或文件站控制。 來源 1 known-limitations 首頁下方的舊學習地圖仍待修正，不屬於本次頂部 Banner 更新。Track A 的正確順序是 A1 → A2 → Stage 5 → A3 → Stage 8；Stage 7.5 是進階選讀。修圖前請以正文中保持展開的路線表為準；這張舊圖尚未加入動畫。 來源 1 简体中文 maintenance 发布前扫描发现 notebooklm-skill 已归档，三语资源表与 Cookbook 改列历史范例，不再推荐新流程使用。Graphthulhu 原入口返回 404，移除失效链接并保留历史名称与评分；其授权目前无法重新核对。原编辑星级不变，没有新增替代项目。 來源 1 reader-ux 首页顶部 Banner 保留原版构图、配色、文字与图标，加入 18 秒循环动画。先走共用基础，再依次展示 A、B 路线；每语言 13 个图标区域会做对应的小动作，包括 CLI 光标输入、工具轻转与 Hub 箭头旋转。文字与卡片固定，最后 2 秒保持静止。 來源 1 accessibility 三语文档站提供停止／播放控制；系统偏好减少动态或禁用 JavaScript 时使用静态图。README 保留静态图入口，三语 PDF 也使用完整 PNG。SVG 不加载外部图片、字体或 JavaScript，图片容量上限未提高。GitHub 嵌入图片是否遵守真实系统的减少动态偏好仍未确认，需要静态阅读时请使用静态图入口或文档站控制。 來源 1 known-limitations 首页下方的旧学习地图仍待修正，不属于本次顶部 Banner 更新。Track A 的正确顺序是 A1 → A2 → Stage 5 → A3 → Stage 8；Stage 7.5 是进阶选读。修图前请以正文中保持展开的路线表为准；这张旧图尚未加入动画。 來源 1 English maintenance The pre-release scan found notebooklm-skill archived. All three catalogs and Cookbooks now keep it as a historical example, not a recommendation for new workflows. Graphthulhu's original address returns 404; its dead link was removed while retaining its historical name and rating. Its license cannot currently be reverified. Editorial ratings are unchanged, and no replacement project was added. source 1 reader-ux The homepage's top banner keeps the original composition, colors, text, and icons, with an 18-second animation loop. Shared foundations come first, followed by tracks A and B; 13 icon regions per language make small, meaningful movements, including typing CLI cursors, gently turning tools, and rotating Hub arrows. Text and cards stay fixed, with 2 still seconds at the end. source 1 accessibility All three documentation locales provide stop/play controls and use a static image when reduced motion is preferred or JavaScript is disabled. Each README links to its static image, and all three PDFs use the complete PNG. SVGs load no external images, fonts, or JavaScript; image budgets have not increased. Whether GitHub's embedded image follows the actual system reduced-motion preference remains unverified; use the static-image link or documentation controls for still viewing. source 1 known-limitations The older learning map lower on the homepage still needs correction and is not part of this top-banner update. The correct Track A order is A1 → A2 → Stage 5 → A3 → Stage 8; Stage 7.5 is an advanced option. Until that image is corrected, follow the always-visible route tables in the text. The old map has not been animated. source

From the release notes

0.10.16 (2026-09-10) fix(providers): correct GPT-5.6 cache retention and add the shipped 5.6 siblings (7f55072d02ee1d73383539312ebba2f5c13fbeb6)

From the release notes

0.10.15 (2026-09-10) feat(providers): support GPT-6 (gpt-6-astra) across the OpenAI param guards (f33595810842c5a0e8a411d104b655c07ae938ed)

From the release notes

Patch release. Three fixes from the issue backlog, each confirmed in source before any change and each with a regression test proven to fail without the fix. A curate can no longer destroy a user's memory index (#3224) AutoMemoryBridge.curateIndex() rebuilt MEMORY.md from its own topic files and overwrote whatever was there. The #1556 guard only skipped the write when nothing matched — but the bridge writes the first topic file itself, so the guard protected exactly one sync and then stopped applying. The reporter measured a hand-curated index of 75 lines and 49 links reduced to a 6-line stub. A curate may now grow or reorder the index and never shrink it. When the generated view is smaller, the existing file is left untouched, the generated view is written beside it as MEMORY.generated.md, and the bridge emits index:preserved. The old failure was silent; a silent refusal would only be a quieter version of the same problem. autoStart: false is honoured (#3278) Two mismatches were stacked: what init writes what the check read file .claude/settings.json claude-flow.config.json key claudeFlow.daemon.autoStart daemon.autostart A different file and a different capital S, so the setting a fresh project ships was never consulted — under a comment reading "Opt-in only — prevents unintended token consumption (#1427, #1330)". Both files and both spellings are now read. Reads disclose the store they did not search (#3228, in part) 3.41.1 shipped the sibling-store disclosure in memory list alone. retrieve, search and an empty list kept returning a clean negative while the rows sat in the file they do not read — on the reported Windows install, 31,673 rows answering found:false. Still not fixed: the Windows 0xC0000409 abort on a ~4.4 GB allocation with the native bridge enabled. That is #2948 and this release does not touch it. On Windows the split is now visible, not healed. Install npx ruflo@3.41.2 Merge commit df87b0db338a.

From the release notes

Patch release. Two fixes, both reported by people hitting them rather than found in-house. Memory path identity (#3196) An explicit database path now selects that database. The bridge cached one registry globally, so the first caller to touch it decided the file for the whole process and every later caller's --path or CLAUDE_FLOW_DB_PATH was accepted and then silently ignored. A CLI write and an MCP write landed in two files, both reporting success, and neither interface could see the other's rows. memory list also stopped printing a confident total for one store while rows sat unread in its sibling — it now names the store it did not read. Reported downstream by pacphi/agentic-kit#213. Not done deliberately: merging the two database files. #2786 gave AgentDB its own plaintext file so native better-sqlite3 never opens an encrypted memory.db. Unifying them would re-break encryption at rest. Seraphina truncation (#3267) Seraphina no longer passes a truncated reasoning dump off as coordination guidance. The routed model spent its whole token budget reasoning and never emitted its JSON, and the resulting empty proposal list read as the swarm needs nothing — the one conclusion a coordinator must never reach by accident. Install npx ruflo@3.41.1 Merge commit cda8b04f6ae1.

From the release notes

Release v4.7.7

From the release notes

Changes since langchain-anthropic==1.7.1 release(anthropic): 1.7.2 (#40387) fix(anthropic): preserve invalid tool use blocks (#40372)

From the release notes

Retain the shared handoff runtime source for installed plugins; not a…

From the release notes

The PageIndex SDK, local or cloud — vectorless, reasoning-based RAG, end to end. Much faster indexing — the PageIndex Flash engine gets the tree from layout stats: no LLM involved for the structure generation itself, LLMs only write the node summaries, and tree expansion proposes a wave of nodes concurrently. client = PageIndexClient() client.submit_document("report.pdf") client.chat("What does the report conclude?") Index, to chat, to agent integration, one client. Local mode needs no server, no vector DB, no PageIndex API key. Highlights Flash engine: the local default (mode="standard" keeps the classic LLM pipeline). Embedded bookmarks are consumed when trustworthy, and tree optimization is on by default — optimize="merge" for the deterministic LLM-free pass, "full" (default) adds LLM expand, which runs a wave of nodes concurrently instead of one round-trip at a time. One complete surface, local and cloud: PageIndexLocalClient(storage_path=...) is the same client as cloud — submit, tree, page content, chat, and agent tools all present in both modes, so code moves between them unchanged. Cloud documents, your own model: api_key decides where your documents live; a configured chat model decides who answers — and the two combine. PageIndexClient(api_key="pi-...", chat_model="openai/gpt-5.2") runs the same in-process document-QA engine over the live cloud tool set. Page content flows through your process to your provider on your credentials; doc_id targets at the prompt level; enable_citations stays with the managed chat. Agent integration: the cloud MCP tool contract, in-process — client.agent_tools() (plain functions), as_openai_tools(), as_anthropic_tools(), as_claude_mcp(), plus one-call openai_agent_config() / anthropic_runner_config() / claude_agent_config() bundles and agent_instructions() for the system prompt. Cloud clients get the live server tool set over the MCP bridge (read-only endpoint by default); local clients get the in-process subset with the same schemas and envelopes — agent prompts port unchanged. Chat surfaces: chat() — question in, answer out, on any backend; chat(stream=True) shows the run as it happens, thinking and tool calls woven into the text, or as typed events via .events; chat(protocol="responses" | "messages") drives the OpenAI Responses or Anthropic Messages API natively with that protocol's own shapes, and chat_completions() keeps the OpenAI-compatible envelope — all with doc_id targeting, streaming, honest usage accounting, and prompt-cache continuity across turns. Transcripts append verbatim: a protocol lane's output goes back into the next request unchanged. Model & connection knobs: index_model / chat_model, index_backend / chat_backend (and per-call backend) passed verbatim to each lane — LiteLLM-routed providers, keyless OpenAI-compatible servers, Azure/Bedrock/Vertex included. index= / chat= slots: the grouped spelling of the flat arguments — a string shorthand or a mapping (index={"model": ..

From the release notes

RuFlo 3.41.0 — Join the open swarm in four commands Any machine running RuFlo can now join a shared, signed, invite-gated swarm at x.ruv.io with a key only it holds. No Tailscale, no VPN, no account. Step-by-step guide with troubleshooting: https://gist.github.com/ruvnet/6d3202cc7030e32169e8d5ac73b4e870 Illustrated explainer: The Swarm Wall (Cognitum Media plates) Try it npx ruflo@latest federation join --code v2.… # 1 · get in with your own key (code comes to you privately) npx ruflo@latest federation roster # 2 · you are listed npx ruflo@latest federation claims # 3 · who owns what npx ruflo@latest federation sync # 4 · read what the swarm posted Then in Claude Code: claude mcp add --transport http --scope user x-ruv-io https://x.ruv.io/mcp What's new ruflo federation command (#3256): join, sync, roster, claims, registry, plus admin-gated invite, admit, publish. Self-service join generates ~/.ruflo/nostr.key (0600), redeems the invite with a NIP-98 signature, and proves membership with NIP-42. Public and private swarm channels (#3261, ADR-386): ruflo federation channel --action create|grant|accept|publish|read|list. Private channels are NIP-44 encrypted with keys the gateway never sees. x.ruv.io gateway plugin (#3255): MCP at https://x.ruv.io/mcp, WebSocket proxy, ruv://federation/registry, ruv://swarm/roster, ruv://claims/board, ruv://swarm/channels. Admin writes are token-gated and fail closed. Seraphina (#3256): swarm coordinator via the cognitum meta-llm. seraphina_guidance { goal } returns guidance, proposals and risks. Advisory only. Canonical relay wss://relay.ruv.io. Sign the NIP-42 relay tag with this exact address; the old *.run.app host still works as legacyRelay. 15 federation MCP tools ship in the RuFlo MCP server. CI: @metaharness/darwin pinned to 0.10.2 so the pin-drift and clean-install gates are green again (#3262); skills and gateway docs cover channels (#3263). Validated before this release 67 CLI tests and 12 gateway tests green; all 23 workflows green on main. Fresh npm install ruflo@latest on a clean cache: federation join with a new key, registry, roster, sync, and a public-channel publish/read round trip against the live relay. Relay E2E on three paths: direct wss://relay.ruv.io, the legacy host, and proxied through wss://x.ruv.io. Private-channel E2E from a second, independent key: plaintext for members, ciphertext only for non-members. Known ruflo verify reports the same one stale witness marker (#1697 rvf-wasm overrides) as 3.40.0; manifest refresh is tracked separately. Packages @claude-flow/cli, claude-flow, ruflo → 3.41.0 (latest, alpha, v3alpha).

From the release notes

New features Agentic RAG: Refactored the Agentic RAG retrieval framework, significantly improving reasoning speed and benchmark performance. (#19046, #19112, #19172, #19209, #19424, #19459) Knowledge compilation runtime settings (#19254) Enforce a rate limit on the knowledge compilation pipeline (#19156) Knowledge graph shows total and current node counts (#19142,#19095, #19114) Graph search highlights the node when pressing Enter on an entity name (#19320) New Sitemap data source for sitemap.xml-based web ingestion (#19344) WebDAV connector supports custom CA certificates (#16901) MonkeyOCRv2 parsing support (#18887) Self-hosted PaddleOCR-VL deployment support (#19011) New Sofya search tool for agent (#19323) EPUB document preview (#19242) Excel citation sources located in document preview (#18853) Model Support New Hubris model provider (#19341) New llmman model provider (#19030) Improvements New RERANK_TOKEN_LIMIT_MODE configuration (#19404) Highlight in retrieval testing and search (#18981) Similarity threshold displayed as a percentage (#18974) UI/UX Completed Turkish locale (#19009) Search match highlighting in Team settings tables (#15985) Bug fixes Fixed images uploaded in chat being ignored by vision models (#18884) Fixed non-UTF-8 encodings in CSV/TXT preview and document decoding (#19222, #19253) What's Changed fix(agentic_rag): include configured system prompt in RAGTools.sys_prompt() by @chakshu-dhannawat in #18927 fix(web): make sidebar i18n reactive and refine template card/icon UI by @yooxdb in #16646 fix(dataset): align parser methods order with python by @6ba3i in #18877 feat(webdav): support custom CA certificates by @taek105 in #16901 Fix Azure embedding endpoint handling by @zcxGGmu in #16993 fix go: migrate missing table chatchannel by @myf-bee in #18944 Python && Go: fix parse status by @JinHai-CN in #18841 fix(ollama): preserve native API base URL in OllamaCV by @iwasaki1112 in #17013 fix(file): stop is_parent_folder_exist raising on a missing parent by @marmar9615-cloud in #18930 fix(agent): enforce MCP task deadlines in worker by @Lem0nTea2002 in #18869 fix: upgrade starlette to 1.3.1 (CVE-2026-54283) by @anupamme in #18947 refactor: delete the unused discord_svr.py bot script by @noron12234 in #17682 Refactor: make all url inside tooltip as link by @wangq8 in #18943 Py: add update_time in select list when list data sync logs by @JinHai-CN in #18867 fix: include filename in file download Content-Disposition header by @huang-aoqin in #17105 fix(embed): drop_params: True was rejected by Together AI and other strict providers by @Harsh23Kashyap in #18627 fix(rag): key Infinity field_map by original column names, not pinyin by @Taranum01 in #18318 feat(go-syncer): add Zendesk connector by @mkaaad in #18961 Fix: Remove the empty response field from the retrieval operator. by @cike8899 in #18963 fix(agent): reject fractional integer DSL parameters by @6ba3i in #18964 Fix: update t

From the release notes

Warning Self-hosted deployments using the bundled Weaviate must complete a manual, staged upgrade before starting 1.17.1. The bundled Weaviate server moves from 1.27.0 to 1.39.2 — 12 minor versions — and skipping minors is unsupported. Pulling and restarting can silently and permanently break vector search. See Upgrade Guide below. Fresh deployments, external Weaviate, and other vector stores are unaffected. New Features Dataset-Scoped Knowledge Base API Keys Knowledge base service-API keys were scoped to the whole workspace: one key could read and write every knowledge base in the tenant, so giving an integration access to a single knowledge base meant giving it access to all of them. Keys can now be bound to specific knowledge bases from the knowledge base's API access panel, with a workspace / this-knowledge-base scope selector at creation time. A bound key returns 403 for any other dataset — both documents and retrieval — and for endpoints that carry no dataset ID, such as list-all. Existing keys are unbound and keep their workspace-wide behavior, so nothing changes when you upgrade. (#37569) Keyboard Movement for Workflow Nodes and Comments The workflow canvas required a mouse to reposition anything. Nodes, multi-node selections, comment markers, and draft comments can now be moved with the arrow keys — hold Shift for larger steps — including Note, Iteration, and Loop nodes. Continuous movement is committed on key release or blur, so it produces one undo entry rather than one per keypress, and read-only mode, container boundaries, and collaboration updates are all still respected. (#41967) Marketplace Creator Profiles and Homepage Redesign Plugin authors now have public profile pages at /marketplace/creator/<handle>, listing their creations with a publisher sidebar, reachable from a new Creator Center entry in the sidebar help menu. The marketplace homepage was rebuilt around a hero, a banner carousel, sticky search, catalog tabs, and tag/language filters, with a stacked-and-swipeable mobile layout. Bundles are no longer mixed into plugin categories. (#41538) Configurable Markdown Form Field Name Length The Markdown form field-name limit was hard-coded at 128 characters, with no way to accommodate longer generated field names. It is now configurable through MARKDOWN_FORM_FIELD_NAME_MAX_LENGTH (Docker) or NEXT_PUBLIC_MARKDOWN_FORM_FIELD_NAME_MAX_LENGTH (direct Web deployments), defaulting to the previous 128. Character validation and prototype-pollution protections are unchanged. (#41697) Bug Fixes Knowledge Base This release fixes a cluster of extraction bugs that changed document content without reporting an error. Re-import affected documents after upgrading; indexed text is not retroactively repaired. CSV cells were passed through pandas type inference before indexing. Leading-zero codes became floats (00123 → 123.0), empty cells became the literal string nan, and cells containing NA or NULL became missing values. The extract

From the release notes

0.8.0 (2026-09-10) Features core: support ordered cube queries (#2677) (8d23753)

From the release notes

What's Changed ci: cap build-test at 60 min and stream prek hook progress by @georgeh0 in #2378 ci(deps): bump taiki-e/install-action from 2.87.0 to 2.87.5 by @dependabot[bot] in #2384 fix(litellm): align embedding responses to inputs by index by @hardness1020 in #2381 fix(engine): skip child-existence reconciliation on demote-only deletes by @georgeh0 in #2379 fix(core): track component activity explicitly instead of by reference count by @georgeh0 in #2385 ci(deps): bump svgo from 4.0.2 to 4.1.0 in /docs by @dependabot[bot] in #2386 ci(deps): bump sharp from 0.35.3 to 0.35.4 in /docs by @dependabot[bot] in #2388 ci(deps): bump js-yaml from 4.3.1 to 4.3.2 in /docs by @dependabot[bot] in #2387 feat(litellm): bound concurrent embedding requests to the backend by @georgeh0 in #2392 fix(litellm): make embedding timeouts terminal, add timeout param by @georgeh0 in #2394 feat(litellm): retry transcription requests under one timeout by @georgeh0 in #2395 docs(litellm): one shared-behavior section for embedder and transcriber by @georgeh0 in #2396 Full Changelog: v1.0.21...v1.0.22

From the release notes

0.4.0 (2026-09-09) Miscellaneous Chores wren-core-base: Synchronize wren-rust-core versions

From the release notes

0.4.0 (2026-09-09) Features core: support ordered cube queries (#2677) (8d23753) Bug Fixes core: reject collapsed relationship joins (#2697) (4000bea)

From the release notes

0.4.0 (2026-09-09) Miscellaneous Chores wren-manifest-macro: Synchronize wren-rust-core versions

From the release notes

🕸️ Cross-Host Federation + Claims Coordinate agents across machines with signed, verifiable messages and work claims. Two coordinators — use either alone or together: Local mesh (agentbbs): HTTP pull, Ed25519-signed, pinned keys. Works over Tailscale, LAN, VPN, or loopback — no Tailscale required. Slack: coordinate in a channel thread with structured messages. Zero infrastructure. Install npx ruflo@latest --version # 3.40.0 Quick start — join the mesh (run on each host) 1. federation_bbs_identity {} → your nodeId + publicKey 2. federation_bbs_peer_add { nodeId, publicKey, url } → pin every other node 3. federation_bbs_serve { bindHost:"<your routable IP>", port:7777 } 4. federation_bbs_register { roomLabel:"#coordination" } 5. federation_bbs_sync { roomId:"<from step 4>" } → on a 15–30s timer Unsigned, misattributed, oversize, or over-hop messages are dropped and counted — a hostile or looping peer can't corrupt the log. Claims — coordinate who owns what Runtime tools: claims_claim, claims_release, claims_handoff, claims_status. Cross-host, publish claim events: { "type":"ClaimIssued", "from":"nodeA", "resourceId":"deploy-api", "ttlSeconds":3600 } One owner per resource. Claim → wait for the coordinator's ClaimAck → then do the work. Security Every mesh message is Ed25519-verified against a pinned key. Registry-anchored pinning refuses any endpoint presenting a different key for a known node (identity-hijack / MITM protection). Private keys never leave the host. Don't put secrets in message payloads. What's also in 3.40.0 Restored memory search recall (a real regression: the default threshold made keyword recall impossible) — #3252 Full CLI test suite green (69 → 0 failures) 📖 Full step-by-step guide, protocol reference, GCP-secret bootstrap, and troubleshooting: https://gist.github.com/ruvnet/2fbad4d0e3bc02cc7d47b4f6fdd2c205 🤖 Generated with RuFlo

From the release notes

Changes since langchain-openai==1.6.1 release(openai): 1.6.2 (#40339) fix(openai): add GPT-6 Astra reasoning efforts (#40330) chore(deps): bump httpx2 from 2.10.0 to 2.12.0 in /libs/partners/openai (#40309)

From the release notes

Patch release shipping #3250. Fixed federation_bbs_* tools were unconditionally degraded. All four tools (register / publish / watch / human_join) gated on await import('agentbbs'). The published agentbbs package is a CLI-only launcher — its tarball contains only bin/agentbbs.js, package.json and README.md, with no main, no exports, and no root index.js for Node to fall back to. That import could never resolve, so the tools reported degraded: true no matter whether agentbbs was installed and working. Replaced with a subprocess presence probe. Probe shell gated to win32. The follow-up hardening: shell: true was unconditional while the probed binary comes from AGENTBBS_BIN, and under a shell that argument is command-interpreted — a crafted value ran a trailing command (verified locally). POSIX now spawns without a shell; Windows keeps cmd.exe so the agentbbs.cmd shim still resolves via PATHEXT. Matches the existing convention in browser-tools.ts, commands/init.ts and init/helpers-generator.ts. @noble/ed25519 no longer externalized in tests. It is a hard dependency of @claude-flow/cli, not an optional one, and externalizing it broke dynamic-import resolution under Vite's SSR transform. Scope This makes the Phase 1 local room/envelope surface functional when agentbbs is genuinely installed. It does not implement cross-host federation — that remains Phase 2+. Install npx ruflo@latest All three packages (@claude-flow/cli, claude-flow, ruflo) are at 3.39.3 on latest, alpha, and v3alpha. 🤖 Generated with RuFlo https://claude.ai/code/session_013u4pmL9ZUAXb6usVQgNo67