repoctx

io.github.nugehs由 GitHub 账号 nugehs 发布v2.6.0更新于 Oct 25, 2026

Local-first code context, impact analysis, and merge-readiness verdicts for AI agents.

已验证STDIO仅桌面知识与记忆开发者工具

概览

AI 生成的概览

一个本地优先的 MCP 服务器,为 AI 助手提供确定性的仓库上下文、变更影响排序和合并就绪判定,代码不出本机。

可以做什么
检查并绘制仓库结构,按任务意图索引和搜索本地项目,不使用嵌入模型也不联网。生成任务感知的上下文包,列出请求所需的文件、测试和命令。排序变更的影响范围,并评估差异与所述意图的偏离程度,生成绑定到确切代码树的凭据。记录请求预期触及的文件并对暂存树做门禁,指出未声明的文件,标记密钥路径和高风险变更。输出评审上下文、判定结果、跨仓库工作区报告、代理体验评分、模型层级建议和代理脚手架文件。
适用场景
适用于代理实际修改代码仓库、且希望以本地可重复的方式核查变更触及范围与是否可合并的场景。适合在 Claude Code、Codex、Cursor 等宿主旁需要确定性、不依赖模型的证据,且不希望使用账号或托管服务的团队。
运行要求
以本地 stdio 进程在用户机器上运行,从 npm 包安装或用 npx 启动,因此需要 Node.js。它读取本地 Git 仓库及其暂存树;未声明任何账号、API 密钥、环境变量或请求头。README 提到可选的在线模式和可选托管组件,但核心命令无需网络。
安装前请注意
门禁会在仓库内写入和读取状态,包括意图文件和凭据,仓库地图查询使用外部按用户缓存。本地门禁通过明确不等于合并批准,托管 CI、代码评审和人工发布决定仍各自独立。启用可选的在线模式和托管组件前,请先确认它们发送哪些数据,因为默认情况下代码不会离开本机。

安装

在 SourceWeft 中

  1. 打开 控制台中的 repoctx,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Solumbe

Models generate the change. Solumbe proves whether it is safe to merge.

Formerly Òtítọ́ (@bashbop/otito), renamed in 4.0.0; the CHANGELOG maps every old name to its new one.

[CI] [npm] [license: MIT] [node] [Listed on mcpservers.org] [Solumbe MCP server – quality and maintenance score on Glama]

[solumbe demo]

[Solumbe MCP server – quality and maintenance score on Glama]

Solumbe is a local-first, deterministic, model-agnostic trust layer for AI-assisted development. It builds task-aware repository context before an agent edits, scores how much a change actually touches, and gates merge readiness against the exact staged tree, with no server, no account, and no code leaving the machine.

It does not replace Claude Code, Codex, Cursor, Gemini, or any native agent harness. It runs beside them and keeps working as the models change underneath.

A passing local gate is never an automatic merge approval: hosted CI, GitHub review, CODEOWNERS, and the human release decision remain separate authorities.

Install

bash
npm install -g @bashbop/solumbesolumbe doctorsolumbe install --host all

The last command connects Solumbe's MCP server to every agent host it finds: Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code, Gemini CLI and Kimi Code CLI. Add --dry-run to see the plan first, or name hosts with --host claude-code,cursor. Restart each host afterwards.

Or without installing: npx -y @bashbop/solumbe doctor.

What it does

Rank what a change actually touches, from the request alone, with no model, no embeddings, and no network:

console
$ solumbe impact . "add refund handling to checkout" --top 3
  concepts: money flow
   1.  src/payment/checkout.service.ts    score 155       |- role: required       |- path matches: checkout       |- symbol matches: checkout, refund       |- concept match: money flow (×1.4 + 6.0)       |- risk: money flow
   2.  src/payment/stripe.webhook.ts    score 6       |- role: advisory       |- concept match: money flow (×1.4 + 6.0)

Gate the exact staged tree, and say precisely why:

console
$ solumbe gate . --staged --base origin/main --request "add refund handling to checkout" --min-convergence 80
  [OK]    Changed files          1 changed file found.  [OK]    Staged snapshot        Changed-file scope and convergence evidence are captured from the exact staged Git tree.     |- Tree: 7d0b513dd256877d6bd309a6406468a9b1eaeedc     |- Base: b776f04fcfbd02c8ce8d63e1a0c2450be58de039  [OK]    Secret safety          No secret file paths and no credential values found in the changed content.  [WARN]  Risk review            Risk-sensitive files changed; maintainer review should be explicit.     |- src/payment/checkout.service.ts  [OK]    Convergence            Task and diff satisfy the convergence requirement.     |- Score: 100/100 (aligned)     |- Receipt handle: rcpt_a8d31efae06e     |- Subject tree: 7d0b513dd256877d6bd309a6406468a9b1eaeedc
  VERDICT  WARN

The convergence score is the part a model cannot grade for itself: it compares the stated intent against the files the diff actually changed, and produces a receipt bound to the exact base, parent, and staged-tree identity.

Hold the change to what was declared. Record the files a request is expected to touch before the edit, and the gate names any file that was never declared:

console
$ solumbe declare . "add refund handling to checkout"$ # ... the agent edits ...$ solumbe gate . --staged --base HEAD --intent .solumbe/intent.json
  [FAIL]  Scope contract         Touched `src/auth/session.ts`, which was never declared. Amend the intent with the reason it belongs (`solumbe amend <file> --reason "<why>"`), or take it out of the change.     |- Request: "add refund handling to checkout"     |- Contract: fe655c2af2b31b7e3edec6aedd248b379235811d1228c829f87ff73796ace487     |- Changed files: 3 (1 declared, 0 amended, 1 implied, 1 undeclared)     |- Implied: tests/checkout.service.test.ts (owner-test of src/payment/checkout.service.ts)     |- Undeclared on a risk-sensitive path: src/auth/session.ts [auth/security]

An amendment adds a file with its reason, hashed onto the contract, so a reviewer reads why the scope grew. Over MCP it is change_impact with declare: true, and review_gate with intent.

The core

Four commands are the product. They call no model, open no socket, and read nothing outside the repository.

StepCommandWhat it proves
Before the editsolumbe context "add refund handling" --path .Which files, tests and commands the request needs
After the editsolumbe converge . "add refund handling" --base origin/main --stagedHow far the diff is from the request, with a receipt
Before the mergesolumbe gate . --staged --base origin/main --request "add refund handling"PASS / WARN / FAIL on the exact staged tree
After the mergesolumbe attest . --verdict verdict.json --merge <sha>A hash-chained record of what shipped, under which verdict

solumbe help shows these and how to set up; solumbe help --all lists every command, including the ones that support them (impact, declare, pr, review, map, calibrate, route, ax, dashboard and the rest). Nothing was removed.

Nothing hosted is required, and nothing hosted can change a verdict. Local Core, Optional Hosted lists exactly what the optional pieces send, and where.

Supporting commands

GoalCommand
Rank change blast radiussolumbe impact . "add refund handling" --top 12
Score intent vs. executionsolumbe converge "add refunds" --path . --base HEAD --staged
Score a committed change exactlysolumbe converge "add refunds" --path . --base HEAD~1 --head HEAD
Gate a product change across repossolumbe workspace-gate ../web ../api --request "ship change"
Grade risk flags against historysolumbe calibrate . --window 30
Grade route tiers against historysolumbe regret . --window 30 --offline
Regrade a saved run, same answerssolumbe regret --rescore run.json
Score Agent Experiencesolumbe ax . "add a new MCP tool"
Recommend a model tiersolumbe route . "add a new MCP tool"
Sharpen context with a model readsolumbe context "add a new MCP tool" --path . --online
Inspect one reposolumbe repo . --json
Build a code mapsolumbe map . --json
Index and search local projectssolumbe index ~/projects --discover then solumbe search "events controller"
Generate an agent harnesssolumbe harness . --out .solumbe/harness.md
Run the MCP serversolumbe mcp

Every command takes --json, and solumbe help lists the full set with flags.

MCP

Solumbe ships a stdio MCP server exposing 14 tools: repo_inspect, repo_map, repo_index, repo_search, context_pack, change_impact, agent_experience, model_route, convergence_score, review_context, review_gate, review_verdict, workspace_report, and repo_harness.

json
{  "mcpServers": {    "solumbe": {      "command": "npx",      "args": ["-y", "@bashbop/solumbe", "mcp"]    }  }}

solumbe install --host all writes this entry for you, using absolute paths so hosts that do not inherit your shell PATH can still start it. Published in the MCP Registry as io.github.BASHBOP/solumbe. Repo-map lookups use an external per-user cache and never write into the inspected repository. Host-specific setup for Claude Code, Claude Desktop, Codex, Cursor, VS Code, Gemini CLI, and Kimi Code is in MCP and Agent Workflows.

How it compares

ApproachStrengthsWhere solumbe differs
Sourcegraph / Cody contextPowerful hosted code search and embedding-based context across an orgLocal-first and deterministic: no server, no account, no code leaves the machine, and the same query always yields the same packet
Hand-written CLAUDE.md / rules filesCurated, intent-rich guidanceHand-written context goes stale; solumbe regenerates context from the actual code (symbols, imports, routes, tests) on every run and complements a short CLAUDE.md
grep / ripgrepFast, universal text matchingsolumbe ranks whole files by task intent across paths, symbols, exports, and tests, then adds patterns and validation commands, producing a context packet rather than a list of matching lines

Documentation

Full command reference, agent workflows, release process, evaluation method, and the design theses behind the trust layer:

bashbop.github.io/solumbe

Contributing

bash
git clone https://github.com/BASHBOP/solumbe.gitcd solumbe && npm ci && npm run ci

npm run ci is the full gate: format, lint, typecheck, version check, tests, coverage floors (70% lines / 60% branches / 75% functions), three evaluation corpora, dependency audit, and a packaged-tarball smoke test. Run it before requesting review.

Start with CONTRIBUTING.md and the Code of Conduct. All changes need maintainer review; main requires passing gates and resolved conversations. Solumbe follows Semantic Versioning, so say whether a PR is no-impact, patch, minor, or major.


Part of the toolchain

solumbe is one of four tools that form a deterministic trust layer for AI-assisted development. Each uses static analysis to answer a question people keep handing to an LLM.

  • solumbe (this tool), for context: what does this change actually touch?
  • tieline, for contracts: did the front end and back end quietly stop agreeing?
  • bouncer, for compliance: could you defend this to Ofcom?
  • aiglare, for governance: where can the model do something you can't undo?

More at segunolumbe.com. static analysis, never the model.

来源:README.md,提交 02eaad0

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v2.6.0最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分