v0.5.0

ToolSeed v0.5.0

让 Agent 将值得重复利用的任务方法、错误教训与工具用法转化为能力，在合适时机召回、应用并更新。
本版本仍为 Alpha。

新增与改进

• 经验学习：保留方法、适用条件、失败教训、工具配方和复用检查，脚本无需作为能力主体。
• 主动召回：DSH / Codex 根据任务、结构化错误与工具线索取得少量候选，由宿主模型判断是否采用。
• 使用反馈：观察绑定原能力、版本、哈希和实际交付，重复使用不会仅因任务不同而再建一份能力。
• 修订门禁：完整修订校验读取基线；DSH 可通过宿主工具执行新场景与回归检查后启用候选版本。
• 宿主适配：共享学习状态规则，保留各自权限与事件接入；DSH 支持已验证的 rc.2 / rc.3。
• 完整升级：固定 GitHub commit，独立安装 Core，同步 Skill、MCP 和适配器，备份并试迁移已有库后自检。
• 改善中英文场景与别名召回，保留版本绑定的结构化错误经验，拆分 Core 领域服务。

安装与升级

PyPI 发布成功后：

python -m pip install --upgrade toolseed==0.5.0
toolseed update codex --ref v0.5.0
# DSH 用户先完全退出 Desktop，再运行：
toolseed update dsh --ref v0.5.0

也可以安装 GitHub Release 附带的 wheel。升级继续使用实际宿主中的能力库目录，不自动合并两个库。

兼容与范围

• Python 3.11+，依赖官方 MCP SDK 2.3+（小于 3）与 PyYAML 6。
• 数据库 schema 从 v6 升到 v7；完整升级会备份并在副本上验证迁移。
• 五个公共 MCP 工具不变。完整定向修订需提供读取所得 base_version_id 与 base_content_hash。
• Core 交付指引，操作由宿主执行；方法交付、模型自述与工具活动均不能替代独立任务结果。
• 两轮真实 DSH 模型任务观察到经验沉淀与跨会话复用；Codex 原生链路验证含脚本驱动场景。
  这些证据不构成普遍成功率、稳定自主学习或提效百分比的证明。
• 本地基线检查为 Python 442 项（1 项条件跳过）、Node 157 项；发布结果以 Actions 实际检查为准。

项目说明 ·
演示素材入口

v0.4.1

ToolSeed v0.4.1

Patch release over v0.4.0. Git install is the supported distribution path for this release;
PyPI is not published (see "Distribution" below).

What changed

• Fix — evidence and maintenance lifecycle hardening (0f1648c): evidence validity,
  maintenance decisions, capability observations, version promotion and rollback paths are
  tightened so a failed write cannot be mistaken for an applied decision, and old decisions
  without a record-order marker are rejected instead of being guessed.
• Fix — atomic decision resolution (c0cac12): applying a maintenance decision now
  resolves and records it atomically; a commit failure rolls back instead of leaving a
  half-applied decision behind.
• Docs — Alpha usage path and delivery boundary (277f38f): a new Alpha quickstart in
  the README (install, MCP stdio configuration, confirming the four public tools, setting and
  verifying the persistence directory, search/save/deliver, and reading Invocation/Evidence/
  unknown), plus a new docs/ALPHA-STATUS.md that states what is verified, what reuse was
  observed in real agent work, what net benefit is not proven, and the unsupported surface
  and known limits. The stale "MCP not implemented" note is gone.
• Skill — save guidance (277f38f): the packaged Skill now states every enforced manifest
  limit, requires a pre-submit check, and defines a minimal correction from the actual error
  message instead of blind retries. Saving stays optional per task. No Core contract was
  relaxed.
• Test — contract guard (277f38f): a new test pins the documented limits to the Core
  constants and keeps the usage documents free of stale claims.

Compatibility

• No database schema migration: the schema is identical to v0.4.0.
• No change to the four public tools (toolbox_search, toolbox_save, toolbox_run,
  toolbox_stats) and no change to the candidate contract.
• Python >= 3.11; the only runtime dependency is mcp>=2.3,<3.

Install (Git)

python -m venv .venv-toolseed
.venv-toolseed/Scripts/pip install "git+https://github.com/SimulationDuTe/ToolSeed@v0.4.1"
# or: .venv-toolseed/bin/pip install "git+https://github.com/SimulationDuTe/ToolSeed@v0.4.1"

Then register the MCP server (Claude Code / Cursor style):

{
  "mcpServers": {
    "toolseed": {
      "command": "<venv>/Scripts/toolseed-mcp.exe",
      "args": ["--data-dir", "<abs-path>/.toolseed"]
    }
  }
}

Codex style (~/.codex/config.toml):

[mcp_servers.toolseed]
command = "<venv>/Scripts/toolseed-mcp.exe"
args = ["--data-dir", "<abs-path>/.toolseed"]

Verify the install: the server must list exactly the four tools above, toolbox_stats must
return structured data, and the data directory must contain toolseed.sqlite3.

Distribution

• GitHub Release: pushing the v0.4.1 tag runs the release workflow, which builds the
  sdist and wheel, runs twine check, writes SHA256SUMS and attaches the artifacts to a
  GitHub Release for this tag.
• PyPI: not published, and this release does not publish there. The Git install path above
  is the supported route; a PyPI upload would be a separate, separately authorised step.

Known limits

• Net benefit is not proven: reuse was observed in real agent work, but the measured cost
  of the method-library path was not lower. See docs/ALPHA-STATUS.md.
• Procedures are untrusted text; nothing is executed, and there is no sandbox.
• Single-user, local storage; provider token and billing figures are not observable here.
• Documentation lives in the repository, not inside the wheel or sdist.

v0.4.0

Full Changelog: v0.3.0...v0.4.0

v0.3.0

Full Changelog: https://github.com/SimulationDuTe/ToolSeed/commits/v0.3.0