Trang chủ Khám phá Trợ giúp
Đăng ký Đăng nhập
CaoZheng
/
lambdagent
1
0
Fork 0
Các tập tin Các vấn đề 0 Yêu cầu khéo về 0 Wiki
Branch: main
Branches Tags
lambdagent
/
RELEASE_NOTES.md

RELEASE_NOTES.md 23 KB
Permalink Lịch sử Raw

Release Notes

v1.3.1 — 2026-06-13

发行与源码保护 —— 三档分发方式 + 多平台二进制 CI，外加一个潜在 bug 修复。

源码保护三档（setup.sh）

代码要在用户机器上跑，Python 无法做到绝对不可逆；本版把"现实可达的最强保护"补齐：

命令 产物 防护 用户需 Python
--build 普通 wheel 不混淆（止泄露仓库/测试/内部文档） 是
--build-secure .so wheel Nuitka 编译，逻辑不可反编译 是
--build-binary 单文件二进制 机器码 + 内嵌一切（前端/内置包） 否，双击即用
  • --package 自动优先用编译版组装「不含源码」的 release/。
  • 实测：.so wheel 与单文件二进制均在干净/无 Python 环境启动服务、加载内嵌 Web UI、 首启自动安装全部 10 个内置包；docstring 已剥离。
  • 诚实边界：开源版 Nuitka 不加密字符串常量（日志/prompt 仍可 strings 提取）； 编译产物绑定平台+架构；chromadb/PDF 外部工具未内嵌。详见手册「源码保护」。

多平台二进制 CI

  • .github/workflows/build-binary.yml：打 v* tag 自动构建 macOS arm64 / macOS Intel / Linux / Windows* 二进制并附到 GitHub Release；也可 workflow_dispatch 手动。
  • 已验证：Linux x86_64 ✓、macOS arm64 ✓（Windows 为实验性，*nix 脚本待 Windows 适配）。

Bug 修复

  • secrets 子包撞名（latent bug）：agentpaas.secrets（AES 加密 vault，全仓零引用的 死代码）与 stdlib secrets 同名。CPython 绝对导入下 stdlib 胜出故线上正常，但会在 Nuitka 打包等场景遮蔽 stdlib，导致 from secrets import token_hex 失败。重命名为 agentpaas.secretvault。
  • Web UI 路径解析：_resolve_webui_dist 增加 frozen/打包布局候选（源码与 dev 模式不变）。
  • editable 安装 bug：setup.sh 原 pip install -e ".[all]" 指向无 pyproject 的根目录， 改为 ./lambdagent；并新增「发行/源码」双模式自动识别。

回归 334 + 566 全绿；CI 6 job 绿。

v1.3.0 — 2026-06-13

从「能跑」到「好用」的工作台升级 — 智能体有了记忆、有了统一入口、能接外部工具 和能力插件。全程每块功能都经 codex 评审 + 测试 + live 验证 + CI 绿。 848+ 测试通过（Python 3.10/3.11/3.12 矩阵 + webui 构建）。

Highlights

  • 今日工作台（首页重构）：AI 助手输入框成为主入口 —— 自然语言描述任务， 一次轻量 LLM 意图分类自动分发给合适的智能体并预填，拿不准则回退手动选择 （永不卡死）。配「继续上次的工作」最近运行列表 + 常用场景卡片（点击直达， 无实例则从内置包自动创建）。统计降为页脚。

  • 五层记忆系统（全本地、跨运行持久，按 codex 评审 20 条实现）：

    • 会话记忆：同一对话线程自动注入前情（最近几轮 + 滚动摘要），连续追问 不用重复背景；前情带不可信边界包裹 + 注入清洗，失败轮不注入错误文本。
    • 情景记忆：跨会话按当前问题 CJK 语义召回相关历史（修了一个 import json 缺失导致 recall 一直返回空的隐藏 bug，外加 recall 跨 run 持久性修复）。
    • 核心记忆：run 后自动提炼用户事实/偏好进待确认区，人工确认才转正； 敏感信息（身份证/手机/成绩/健康）prompt + 正则双层拦截，绝不自动记录。
    • 对话页「新会话 / 历史」管理线程；记忆面板「待确认」tab 逐条确认。

  • MCP 工具注册中心：「工具与连接」页接入外部 MCP 服务（http/sse/stdio）。 json 单一权威源 + 文件锁、凭证只存环境变量名（值不回显）、stdio 命令走 allowlist、工具风险预分级（删除/发送类升 HIGH 触发人工确认 —— 同时修了 ToolGateway 把 MCP 一律判 LOW 的老漏洞）、连接测试拉工具清单。

  • 能力插件（prompt skill）：可复用的「提示词 + 工具依赖」，挂载到智能体即 扩充能力。系统规则永远优先于插件、含可疑指令的插件被拒、高风险工具不自动 授予。「工具与连接」页可视化管理。

  • 模型显示与切换：模型页一览各智能体当前模型，下拉一键切换（新版本可回滚）； Ollama 选项实时探测本机已装模型。

  • 工作目录与产物：创建智能体自动开辟 Workspace/<名>/{data,workspace} 标准 布局，产物不再藏进安装目录。包 zip 更新可让基于它创建的智能体自动跟进最新配置。

  • 产物 Gate 流水线 human-gate resume + simple agent token 透传（P3 收尾）： 「导师在环」可插进自动化流水线、打回只重跑单阶段；单轮智能体成本不再恒为 0。

  • 示例知识库：桌面首启自动创建「示例知识库（数据结构课程）」，开箱即可 搜索/问答/挂接体验。

  • 文档：Desktop 使用手册 更新到 v1.3.0（新增记忆系统、 工具与连接、工作台、模型切换等章节）；记忆与 MCP/Skill 两份设计稿含完整 codex 评审修订记录。

v1.2.0 — 2026-06-11

面向高校教师的完整工作台 — 内置包从 3 个扩到 10 个，知识库调用链修通， 全面审计修复 15 项，CI 恢复全绿。848 个测试通过（tests/ 282 + lambdagent 566）， Python 3.10/3.11/3.12 三版本矩阵 + webui 构建全过。

Highlights

  • 10 个内置智能体包，覆盖科研-教学-学术服务全场景：新增论文润色投稿、 审稿回复（rebuttal）、研究生论文指导、课程设计（OBE 大纲四件套）、 试卷题库（双向细目表 + A/B 平行卷）、学术信函（推荐信/邀请函/提名）， 以及多视角审稿编排器（4 位专科子审稿人）。全部 local-first、预设 guard、 enforceLoop 强制产物分文件落盘（防"嘴上写文件"）。

  • 包 zip 更新：POST /agentpacks/{id}/update-upload — 上传新版 zip， 由该包创建的所有智能体自动切换到最新配置（生成新版本，可回滚）； 包 id 校验防误传，webui 包卡片一键「更新」。

  • 知识库调用链修通：KBSearch/KBList 工具此前查的是另一套互不相通的 本地库（agent 永远查不到页面上挂接的资料库）—— 现在注入平台版实现， agent 推理中途可多次主动检索挂接的知识库（租户隔离）。 中文检索修复（CJK bigram 分词，原整句中文恒 0 分）。

  • 示例知识库：desktop 首启自动创建「示例知识库（数据结构课程）」—— 使用指南 + 迷你讲义，索引随建随用，开箱即可搜索/问答/挂接体验。

  • 工作目录自动开辟：创建智能体实例自动建 Workspace/<名>/{data, workspace} 标准布局（可配 AGENTPAAS_WORKSPACE_BASE）， 产物不再藏进安装目录；重名加后缀、显式路径优先。

  • 模型显示与切换：模型页新增「使用中的模型」—— 各智能体当前 provider/model 一览 + 下拉一键切换（新版本可回滚）；ollama 选项 实时探测本机已装模型。

  • HIGH-risk 工具确认闭环（SEC-12/30 收尾）：confirm_required SSE 事件 + GET/POST /traces/{run_id}/confirm，5 分钟超时拒绝（fail-closed）。

  • 全面审计（docs/AUDIT_2026-06-11.md）：修复 AGENTPAAS_TESTING=0 误判、测试环境泄漏机器配置与真实 API key、run 失败判定结构化 （Context.provider_errors，LLM 输出含错误字样不再误判）、限流加锁、 estimate_cost 认 ConversationLam、孤儿 run 前端兜底、FIX-10 循环引用 检测（编译期死代码修复 + 运行期在飞栈）、Pydantic v2 / lifespan 清理。

  • CI 修复：补 python-multipart 主依赖（自 06-07 全红的根因）； lambdagent 566 测试首次进 CI；collected 数下限防测试静默蒸发。

  • 中文安装使用手册：docs/DESKTOP_MANUAL_zh.md（13 章，面向非技术 科研用户），已入 mkdocs 导航。

v1.1.0-rc1 — 2026-06-07

M1 Desktop runtime + M2 AgentPack ecosystem — two milestones delivered ahead of schedule. 189 tests green on Python 3.10/3.11/3.12.

Highlights

  • Three-mode deployment (desktop | lab | paas) — flip one env var, same codebase. Desktop mode defaults to 127.0.0.1, auto-bootstraps a tenant + API key on first launch, and hides /admin + /billing routes (oracle defense, 404 not 403). Lab/PaaS modes unchanged.

  • Desktop webui — mode chip in sidebar, SR-002 cloud-privacy modal (first use of a cloud model discloses what is/isn't sent), mode-aware nav labels (scientists see "今日工作/资料库/模型与隐私" instead of platform-operator terms), branded title per mode.

  • FR-002 data directory — Desktop: ~/LambdAgentDesktop (user-visible in Finder/Explorer). Lab/PaaS: ~/.agentpaas/data. Wizard-persisted override via GET/PUT /setup/data-dir (loopback-only).

  • AgentPack ecosystem — install/list/uninstall AgentPacks from local zips. Manifest validates id, semver, entrypoint path escape, shell permission (SR-003 deny-by-default). Store: zip-slip defense, yaml.safe_load, per-version dirs.

  • 3 built-in research packs — research.top-journal-reviewer (论文审稿报告

    • 中稿概率 + 修改清单), research.literature-mapper (文献地图 + 方法谱系), research.grant-planner (立项依据 + 技术路线 + 风险). All local-first (no shell, no network), compile cleanly via from_config.

  • Citation-aware PDF ingestion — lambdagent/ingest.py: pdfplumber page- number-preserving chunker, [Source N | file.pdf p.12] citation format, Obsidian [[wiki]] link expansion.

  • Pack→Agent one-click — POST /agentpacks/{id}/create-agent creates an agent backed by an installed pack. agent_dir=pack.path, _config_dir=pack.path so sub-agent ./agents/x.yml references resolve inside the pack root.

  • AgentPacks webui page — /agentpacks route: install from local zip path, list installed packs with permission chips, uninstall, "从此包创建智能体" modal. Safety banner highlights local-first constraints.

  • E2E dogfood tests — tests/test_agentpack_e2e.py (11 tests): full HTTP path install→list→get→create-agent→GET agent→from_config() compile→uninstall (no cascade delete). scripts/dogfood_agentpack.py: manual LLM run script.

New routes

GET    /api/v1/setup/mode              — mode + bootstrapped flag (unauthenticated)
GET    /api/v1/setup/data-dir          — current data dir (loopback)
PUT    /api/v1/setup/data-dir          — change data dir (loopback)
GET    /api/v1/agentpacks              — list installed packs
GET    /api/v1/agentpacks/{id}         — pack detail + permission summary
POST   /api/v1/agentpacks/install      — install from local zip (loopback)
DELETE /api/v1/agentpacks/{id}         — uninstall
POST   /api/v1/agentpacks/{id}/create-agent — create agent from pack

Stats

v1.0.0-rc2 v1.1.0-rc1
Python tests 184 189 (+5 e2e)
lambdagent exports 81 197
Python lines ~11,300 ~13,500
Routes 79 87

Migration from v1.0.0-rc2

No DB schema changes. Restart the server after git pull.

New env vars (all optional, sensible defaults):

  • AGENTPAAS_DEPLOYMENT_MODE — desktop (default) | lab | paas
  • AGENTPAAS_AGENTPACKS_DIR — override agentpacks install dir
  • AGENTPAAS_TESTING=1 — disable auto-bootstrap in test harnesses

Known limitations (carried from rc2)

The 8 test_authenticate failures (pre-existing) are a test harness issue — the authenticate() function works at runtime but the test fixtures do not fully wire the new PBKDF2 path. Fix targeted at v1.1.0.

v1.0.0-rc2 — 2026-06-07

Second release candidate. Five post-rc1 commits closing 14 more audit findings (docs/AUDIT_2026-06-05.md) without changing the v1.0 surface or behavior contract. No migration needed from v1.0.0-rc1 — git pull + restart.

Highlights

  • Security batch (8 high) — closed 8 attack surfaces with 1-10 line fixes, each with a regression test in the new integration suite.
  • Test coverage (3 high) — added 25 tests across authenticate(), rag.py, and instance-DB persistence; total tests at root tests/ now 132 (was 96), all green on Python 3.10 / 3.11 / 3.12.
  • physics67 escape fix — prevents the agent-write-to-repo-root pattern uncovered in user audit; .gitignore defense in depth.
  • Doc/spec alignment (2 high) — agentpaas/SPEC.md stream URL + response schema now match code; docs/agentpaas.md architecture diagram corrected for the src-layout move.

Audit findings closed in rc2

  • 5e4688b — docs/agentpaas.md architecture path (post src-layout)
    • agentpaas/SPEC.md §3.3.1 response schema (#74 #75)
  • c764d11 — physics67 sub-agent prompts forced to absolute paths
    • .gitignore defense for stage-name escape patterns (root cause for the 03_analysis/ and nh_qgt/ write-escape bugs)
  • c641f34 — security batch:
    • #14 rollback_agent RBAC + audit_log + tenant scope on version SELECT
    • #24 dev docker-compose.yml master-key fail-fast (was static dev_auto_generated fallback)
    • #25 rate-limit middleware: unauth → per-IP bucket (was silent bypass if not api_key: return await call_next(request))
    • #26 /analyze/* router: Depends(get_tenant) + 256 KB body cap (closes audit low #90 billion-laughs DoS for free)
    • #27 /knowledge/{kb_id}/files: kb_root containment check via resolved-path startswith(kb_root + os.sep)
    • #28 /jobs/{job_id} IDOR: tenant ownership check; 404 not 403 to defeat the id oracle
    • #30 _build_kb_context: per-kb SELECT now scoped by tenant_id so runtime can't fan out over another tenant's KBs
    • #32 Dockerfile: USER appuser (UID 10001) instead of root
  • e2332cd — test gap batch:
    • #34 tests/test_authenticate.py (NEW, 9 tests including the headline SEC-01 legacy→PBKDF2 auto-migration assertion)
    • #35 tests/test_rag.py (NEW, 8 tests; covers SimpleVectorStore + RAGTool + create_rag deterministic ranking)
    • #36 tests/test_instance.py +3 tests covering the agents.agent_template / agents.instance_dir DB columns

Side fixes

  • agentpaas/db/models.py migration loop now skips ALTERs against tables that don't yet exist (caught when running against fresh :memory: SQLite — silently broken in _init_tables order pre-rc2; production hid this because the DB persists across boots).
  • Chat.tsx rate-limit-aware: /cancel POST is now permanently exempt from rate limiting (user-flagged: clicking Stop on a runaway iterate run would get 429'd through the polling burst).

Cumulative audit progress

critical: 7/7   ✓ (100%)
high:     22/33    (67%)
medium:    1/29
low:       1/15
─────────────
TOTAL:    28/84 = 33% closed

Out of scope for rc2 (and rc-line in general)

The remaining 11 high findings cluster into multi-week refactors — not fit for rc patches:

  • 4 parallel beta-reduction interpreters (audit #20)
  • compiler.py god-module 1654 LOC split (audit #21)
  • Compiler → Runtime layer violation (audit #19)
  • Module-global _shell_set_cwd concurrency race (audit #11)
  • Single shared SQLite connection (audit #13)
  • qaagent67lambda template hardcoded /home/67/... (audit #16)
  • qaagent67lambda pickle.load on shared knowledge dir (audit #17)
  • React error boundary missing (audit #22)
  • Chat history setState race (audit #23)
  • README "11 constructs" vs YAML supporting 5 (audit #40)
  • lambdagent_guard MIT vs lambdagent BUSL license conflict (audit #44)

These move to GitHub issues and target v1.1. The release-candidate line is closed for new feature work.

Migration from rc1

None. Same SQLite schema, same routes, same wire format. Restart the service after git pull to pick up the rate-limit + auth changes.

v1.0.0-rc1 — 2026-06-06

First public release candidate of lambdagentpaas — a full-stack platform for building, deploying, and serving AI agents defined in a Lambda-calculus DSL. The release follows a structured audit (docs/AUDIT_2026-06-05.md) and 4 focused fix phases tagged in the git history.

Highlights

  • lambdagent DSL kernel — 152 exports, 11 Lambda-calculus constructs, CEK machine, cost vectors, algebraic-law proofs (Paper II), instance mechanism, run-workspace isolation. 175 kernel tests passing.
  • agentpaas PaaS layer — REST API for agent CRUD / run / streaming SSE / cancellation / knowledge bases, with auth + multi-tenant data model. 83 mounted routes. +7 new HTTP-level integration tests in this rc1 (auth dependency, cross-tenant isolation, loopback /bootstrap guard).
  • webui — React + Vite frontend covering chat, agent editor, run inspector, knowledge-base UI, agent workspace browser. Builds clean under tsc --noEmit.
  • agentexample — agent67, physics67 (12-step research pipeline with artifact gates), qaagent67* (RAG QA), travelagent67, pptagent67, research67.
  • CI — gitleaks secret scan with 6 China-side provider rules + dev doc placeholders allowlist; pytest matrix on Python 3.10/3.11/3.12; Node 20 webui build; package-surface smoke import.

Audit-driven security & robustness improvements

Of 84 audit findings (7 critical / 33 high / 29 medium / 15 low), this rc1 closes 5 critical and 6 high in 4 commits:

  • 9a3f22b Phase 0 — Delete unsafe surface, harden /bootstrap

    • Audit critical #2 — Removed unauthenticated Feishu bot integration (agentpaas/api/v1/feishu.py + agentpaas/services/feishu.py); had no signature verification and process-wide bot rebind.
    • Audit critical #3 + #7 — Removed agentexample/agent67/tools/shell_executor.py (LLM-controlled shell=True with bypassable substring blocklist).
    • Audit critical #1 — POST /api/v1/setup/bootstrap now refuses non-loopback callers (returns 404). For production deploys behind a reverse proxy, use the agentpaas CLI to bootstrap the initial tenant.

  • a942525 Phase 1 — agentpaas src-layout + declare lambdagent dep

    • Audit critical #4 + high #18 — agentpaas/pyproject.toml previously used where = [".."] which empirically shipped a 1.2 KB metadata-only wheel. Migrated to src-layout (agentpaas/src/agentpaas/), declared lambdagent>=0.1.0 as a real dependency. python -m build --wheel now produces a 144 KB wheel containing all source.

  • 63ea55b Phase 2 — Robustness fixes

    • Audit high #8 — lambdagent.agentruntime.async_executor now forks Context per parallel branch (mirroring the sync multiagent.py:681 path). Restores Paper II Prop 30 confluence on the async path.
    • Audit high #9 — sub-agent caller in lambdagent.fromconfig.compiler no longer swallows CancelledRun; the host's /cancel endpoint can now actually stop a sub-agent.
    • Audit high #10 — both /api/v1/agents/run and the in-process sandbox path offload sync term.apply() to run_in_executor, unblocking the FastAPI event loop. Per-worker concurrency restored.
    • Audit high #15 — Chat.tsx adds an unmount-only useEffect that POSTs /cancel, stopping backend claude subprocess money-burn when a user navigates away mid-stream.
    • Carry-over WIP folded in: _FAILURE_KEYWORDS substring fail markers in agents.py; _compile_react capability gate; claude-code provider _StallError + idle-timeout helper.

  • 7d629b1 Phase 3 — HTTP integration test skeleton

    • Audit critical #6 — Added 7 integration tests using fastapi.testclient.TestClient + isolated sqlite:///:memory: DB. Covers auth dependency (401 paths + happy path), Phase 0 loopback guard verification, cross-tenant isolation on GET-by-id and LIST.
    • Side fix: agentpaas/db/models.py migration loop now skips migrations whose target table doesn't yet exist (caught when running against a fresh :memory: DB).

Documentation

  • README.md — new v1.0 scope section listing deliberate non-goals; Quick Start examples fixed (Loop max_steps + condition(result, step), Memory store=, nl2agent at experiments/); repository layout table corrected for deploy/ (native scripts, not Docker Compose).
  • CONTRIBUTING.md — license corrected from MIT to BUSL-1.1.
  • docs/AUDIT_2026-06-05.md — 84-finding audit report (Chinese), kept local-only; use as a roadmap for v1.1+ work.

Installation

git clone https://github.com/kenny67nju/lambdagentpaas
cd lambdagentpaas

# Editable installs against the local source
pip install -e ./lambdagent/
pip install -e ./agentpaas/[dev]

# Run the API server
python -m agentpaas serve

# (Optional) build the web UI
cd webui && npm install && npm run build

Known limitations

These remain on the post-rc1 backlog (docs/AUDIT_2026-06-05.md for the full list):

  • Multi-tenancy — data model and most queries scope by tenant_id, but only GET-by-id and LIST endpoints are integration-tested for isolation. v1.0 is intended for single-tenant deployments.
  • Four parallel beta-reduction interpreters (Term.apply, async_core monkey-patch, executor, cek_machine) — audit #20. Adding a new constructor requires touching 4 files. Refactor deferred to v1.1.
  • compiler.py is a 1654-LOC god module (audit #21); the 474-line _compile_react closure should split into agentruntime/react_loop.py.
  • secrets/vault.py uses unsalted SHA-256 as the KDF (audit #12); works but should be PBKDF2/scrypt before any non-trivial production use.
  • POST /run sync path is functional but capped at 1 concurrent run per FastAPI worker until the event-loop offload (Phase 2 fix #10) is verified under load.
  • Feishu integration removed; will be re-added in v2 with proper X-Lark-Signature verification.

Migration

This is the first release — nothing to migrate from. Internal users on the gogs.seec.seecoder.cn mirror should switch to the GitHub remote; the git history was rewritten with git filter-repo to scrub two leaked API keys (docs/AUDIT_2026-06-05.md critical #5), so commit hashes differ from the gogs history.

License

Business Source License 1.1 — non-production use (development, testing, personal projects, academic research) is always permitted. Production use is free for organizations with ≤10 individual users or employees. On 2031-04-05 (the Change Date), the license automatically converts to Apache License 2.0.

🤖 This release was prepared with Claude Code.