collab-charter
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@collab-charterhelp me refactor this module with API compatibility, run the full pipeline"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
collab-charter · 协作宪章执行内核
把《用户协作宪章》从一份提示词文档,变成 Agent 可加载、可调用、可验证的执行层。
90 条条款 · 8 站管线 · 14 个 MCP 工具 · 368 项测试 · 条款↔实现双向追溯 · 330 条知识库
这是什么
一份写在 Markdown 里的协作规范,模型只能"尽力遵守"。本项目把它编译成能跑的东西:
宪章原文 | 在本项目中 |
0.3 路由管线(八站) |
|
0.4 五级优先级 / 0.5 五条冲突规则 |
|
第十章 安全闸门 |
|
0.6 记忆流(沉淀/再激活/衰减) |
|
0.8 回流前自检(六项) |
|
第十一章 知识库与语义场 |
|
第十二章 符号锚定 |
|
—— | 外加:把图像提示词规范 |
一句话:宪章不再是一段"请遵守"的文字,而是一个能被 curl、能被 MCP 调用、
能被测试断言、能在改坏时立刻报错的系统。
Related MCP server: org-as-code MCP Server
快速开始
1. 命令行(零安装)
git clone https://github.com/LucasDevKid9/collab-charter && cd collab-charter
# 会话开始:第一章要求的接收自检(≤3 行)
PYTHONPATH=src python3 -m collab_charter.cli receipt
# 这次该走全量还是快速通道?
PYTHONPATH=src python3 -m collab_charter.cli route "帮我重构这个模块"
# → Route(全量): 安全闸门 → 意图解析 → 认知镜像 → 记忆检索 → 角色协作 → 综合裁决 → 输出 → 反馈回流
# 跑完整管线,看逐站轨迹 + 输出 + 自检结论
PYTHONPATH=src python3 -m collab_charter.cli handle "重构这个模块" --constraint 保持API兼容
# 查任意条款原文(「查来源」用)
PYTHONPATH=src python3 -m collab_charter.cli clause 0.5-R4
# 宪章完整性自检
PYTHONPATH=src python3 -m collab_charter.cli verify
# 知识库摘要(第十一章:词条数 / 断链 / 跨文件同名 / 未验证)
PYTHONPATH=src python3 -m collab_charter.cli kb
# 语义场深挖:加权词条 + 分布权重 + 闸门结论 + 覆盖缺口
PYTHONPATH=src python3 -m collab_charter.cli mine "谷圈 痛包 晒谷" --state 收敛期
# 潜空间语义场:关联网络 / 中心性 / 社群 / 缺口
PYTHONPATH=src python3 -m collab_charter.cli field "谷圈 痛包 晒谷" --state 收敛期
# 符号锚定:多条件权衡交给 z3 硬算(需 pip install z3-solver)
PYTHONPATH=src python3 -m collab_charter.anchor \
--declare "x:可行性,y:预算充足,z:时间允许,w:风险可控" \
--constraints "x+y+z+w>=3; Implies(y==0,w==1)" \
--decisions "做:x+y+z+w>=4; 观望:x+y+z+w==3; 不做:x+y+z+w<3" \
--rounds "|y==0"
# 治理内核:回放审计账本(每次 handle 自动记录,决策可解释、可复盘)
PYTHONPATH=src python3 -m collab_charter.cli replay --sessions # 列出审计会话
PYTHONPATH=src python3 -m collab_charter.cli replay <session> --step # 逐事件回放
PYTHONPATH=src python3 -m collab_charter.cli replay --why <decision_id> # 解释某条决策(为什么拦/放)
PYTHONPATH=src python3 -m collab_charter.cli replay <session> --json # 机器可读导出
# 宪章编译为策略IR + 风险概览 + 宪章差分(治理内核·P2-10 轻量版)
PYTHONPATH=src python3 -m collab_charter.cli policy --summary # 风险档概览
PYTHONPATH=src python3 -m collab_charter.cli policy --risk critical # 只看高危条款
PYTHONPATH=src python3 -m collab_charter.cli policy --diff 另一份.yaml # 宪章差分(改了什么)
PYTHONPATH=src python3 -m collab_charter.cli policy --json # 机器可读导出2. Python
from collab_charter import Charter, Session
from collab_charter.flows import user_instruction, model_speculation
session = Session(Charter.load())
session.register_tool(ToolSpec("web_search", "联网检索", my_search_fn, {"q": {"type": "string"}}))
session.ask("帮我查一下 z3 的最新技术选型")
session.call_tool("web_search", q="z3 solver 2024")
out = session.handle("顺便评估下迁移成本", constraints=["不能停机"])
print(out.text) # 输出
print(out.trace) # 逐站轨迹(「讲思路」时展开)
print(out.explain_failures()) # 自检没通过时说清楚哪里没通过回流自检(0.8-C4)说明:全量路径下,handle() 默认不自动沉淀记忆,
因此 0.8-C4(反馈是否已回流)会显示未通过——这是有意的安全闸,防止
未经确认就把内容写入长期记忆。两种让它通过的方式:
# 方式 A:显式回流(沉淀整个会话的关键流)
session.reflow()
out = session.handle("再确认一次", constraints=[])
# 方式 B:白名单自动沉淀(命中消息文本或 memory_key 的流自动入库,本次即通过)
out = session.handle("记录一下:先过闸再断言", auto_deposit_keys=("先过闸",))
assert out.selfcheck.passed认知执行注入(Agent 平台接线点):stage_hooks 对全部 8 个站点统一生效,
mirror_fn / council_fn 让你用真实 LLM 回调替代认知镜像与三角角色协作的
确定性模板(无回调时回退模板,保持零依赖):
from collab_charter.router import PipelineStage
out = session.handle(
"帮我重构这个模块",
mirror_fn=lambda flows: "镜像:你的隐含诉求是低风险重构",
council_fn=lambda flows: real_llm_council(flows), # 真实三角视角
stage_hooks={PipelineStage.INTENT: tag_flows}, # 任意站点拦截
)3. MCP(推荐给 Agent 平台)
{
"mcpServers": {
"collab-charter": {
"command": "python3",
"args": ["/绝对路径/collab-charter/interfaces/mcp_server.py"]
}
}
}零依赖(只用标准库实现 JSON-RPC over stdio),拷到任何沙箱都能直接跑,
并已通过官方 mcp SDK 的 ClientSession 端到端冒烟(initialize →
tools/list → tools/call),可在 Claude Code / Cursor 等真实客户端下握手。
14 个工具:条款查询 / 通览 / 接收自检 / 路由决策 / 管线执行 / 冲突裁决 /
安全闸门 / 提示词诊断 / 记忆沉淀 / 记忆检索 / 符号锚定求解 / 知识库检索 /
语义场深挖 / 潜空间语义场。
(anchor_solve 需要可选依赖 z3-solver:pip install collab-charter[anchor];
未安装时返回明确安装指引,不影响其余 13 个工具。)
4. 特征簇提示词诊断
PYTHONPATH=tools python3 -m fcslint "古早立绘,半身像,赛璐璐平涂,16-bit indexed color,硬边缘,4:3"输出六大功能簇归类、权重梯度偏离、冲突清单、控制力量化评分(0–100)、
以及同范式的簇结构保留优化稿(重排 / 降权 / 删无效项 / 补锚点)。
词库已升级为可扩容知识库(knowledge/,330 条),诊断与深挖共用同一份数据。
项目结构
charter/ ① 宪章层 CHARTER.md(人读) · charter.yaml(★单一事实源) · compact.md(精简版)
knowledge/ ② 知识层 lexicon.yaml · styles.yaml · subcultures.yaml(330 条,11.1)
src/collab_charter/ ③ 内核层 flows · gates · router · arbiter · memory · selfcheck · session · cli
· lexicon/(知识库四件套) · anchor/(符号求解器)
interfaces/ ④ 接口层 mcp_server.py(零依赖 MCP,14 工具)
tools/ ⑤ 工具层 fcslint/(特征簇诊断) · anchor/solve.py(兼容 shim)
skills/ ⑥ 技能层 using-charter · info-flow-protocol · feature-cluster-syntax
· deep-research · authoring-skills
harnesses/ ⑦ 适配层 claude/codex/cursor 清单 · session-start 钩子 · tool-mapping.md
tests/ ⑧ 验证层 368 项测试,含条款↔实现双向追溯
docs/ ANALYSIS · ARCHITECTURE · SEMANTIC-FIELD · TRACEABILITY四个值得一看的设计
① 单一事实源 + 双向追溯。 条款只写在 charter/charter.yaml,每条有全局唯一 id。
代码申报"我执行哪些条款",测试机械验证"条款真实存在 → 模块真实存在 → 有指名测试",
并做反向检查——不允许声明实现一个宪章里不存在的条款。
见 docs/TRACEABILITY.md。
② 把不变量写成断言,而不是注释。 宪章 0.3 规定了八个站点。如果代码里少一个站, 普通实现会静默跑偏;本项目启动即校验,不一致直接抛错。同样手法用于闸门矩阵覆盖度、 自检项一致性、知识库断链、MCP 工具数。
③ 技能正文不指名工具。 技能只写动作(「查一条条款」),平台差异收敛到 harnesses/tool-mapping.md 一张映射表。 新增平台不修改任何技能与源码——这是本项目最重要的不变式。
④ 确定性优先于聪明。 向量空间不用 hashing trick(330 条规模下碰撞噪声可观测), 社群划分不用随机初始化,空间按任务临时建构用后即弃。一切为了 11.2-V1: 同输入 → 同输出 → 可复现、可审计。见 docs/SEMANTIC-FIELD.md。
平台接入
平台 | 接入方式 |
Claude Code |
|
Codex |
|
Cursor |
|
Gemini CLI |
|
其他 / 自研 Agent |
|
各平台能力缺口与降级方案(无记忆 / 无文件访问 / 无联网 / 无子代理)见 tool-mapping.md。
测试
python3 -m pytest -q # 368 项
python3 -m pytest tests/test_traceability.py -v # 条款追溯
PYTHONPATH=src python3 -m collab_charter.cli verify依赖:Python ≥ 3.9,pyyaml(仅此一项;MCP server 与 fcslint 核心零依赖)。
fastapi/uvicorn 为可选(HTTP 接口);z3-solver 为可选(符号锚定,
pip install collab-charter[anchor])。
文档
docs/ANALYSIS.md —— 对 superpowers / cognitive-core / 特征簇总纲 / 协作宪章 四份材料的实证式深度分析与缺陷清单(按 P0–P3 分级),以及本项目的自我约束
docs/ARCHITECTURE.md —— 七个关键设计决策、备选方案取舍、诚实边界
docs/SEMANTIC-FIELD.md —— 知识库 / 向量征量空间 / 分布权重 / 潜空间语义场的深度说明与诚实边界
docs/TRACEABILITY.md —— 90 条条款逐条归属(代码执行 / 技能承载)
许可
MIT,见 LICENSE。
English
collab-charter turns a Chinese-language user-side collaboration charter from a prompt document into an executable layer for AI agents.
90 clauses with globally unique, addressable IDs in a single source of truth (
charter/charter.yaml); human-readable text and code both reference it.8-stage pipeline (safety gate → intent → mirror → recall → council → adjudicate → emit → reflow) with a fast path for trivial requests.
Deterministic conflict arbitration with 5-level priority, every verdict citing its clause.
Safety gate as a data table (sensitivity × permission), executed, not hard-coded.
330-entry knowledge base (dictionary / styles / internet subcultures) with a deterministic sparse measure space, distribution weights, and latent-field network analysis — see
docs/SEMANTIC-FIELD.md.Symbolic anchoring (cognitive-anchor v2.0 merged in): z3 full enumeration, AST whitelist sandbox, multi-round drift detection.
14 MCP tools, zero third-party dependencies (stdlib JSON-RPC over stdio).
368 tests, including bidirectional clause↔implementation traceability.
Skills describe actions, never tool names; per-platform differences live in one mapping table, so adding a platform touches no skill or source file.
Use it to give any agent a consistent, auditable, verifiable collaboration protocol — and to lint image-generation prompts against a formal feature-cluster grammar.
This server cannot be deployed
Maintenance
Related MCP Connectors
Shared control plane for AI coding agents — tasks, memory, decisions, file locks. 12 tools.
AI Reasoning Cache & Consensus Layer with 11 MCP tools via Streamable HTTP.
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
Context and control for any AI. Agent identity, scoped access, and proof across tools.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceAutonomous spec-to-product coding-agent CLI. Its MCP server exposes 34 tools over stdio: project state and task-queue ops, memory retrieve/store, code search, quality and verification reports, repo hotspots/co-changes, and structured findings/learnings.3,973 npm1,096Business Source 1.1
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to interact with a Git-native organizational governance system, supporting proposal-validation workflows, decision-making, and audit trail management through 21 MCP tools over stdio transport.1MIT

Nexxupp MCP Serverofficial
AlicenseBqualityCmaintenanceEnables AI agents to store and search private memory, explore and edit workspace stations, databases, tasks, and charts, and schedule calendar events through natural language. Supports local stdio and hosted HTTP modes with per-user OAuth scoping of tools.22MIT- AlicenseNot gradedqualityBmaintenanceExposes a living systems memory vault to autonomous agents, letting them search notes with BM25 ranking, read sections within token budgets, extract API/port/schema contracts, compute dependency blast radius, and run machine-verifiable invariant assertions before mutating code. It also supports surgical frontmatter edits, schema linting, wikilink integrity checks, and worklog recording, over stdio MCP.MIT