SkillFoxx

Spine: an architecture control loop for GigaCode and Claude Code

spine-bank

The arch-be binary with an MCP server, hooks and 79 skills checks agent code against architecture rules and does not let the evidence be faked

  • 2 ★ on the repository
  • Last commit
  • MIT
CLI

Extended access

Why this level

  • arch-be connect edits .claude/settings.json and installs hooks that block agent actions when the gate fails
  • The binary is downloaded from a GitHub release and runs on the developer machine, so its integrity must be checked against SHA256SUMS
All reasons and checks
Russian stack

romannekrasovaillm/spine-bankAuthor:romannekrasovaillm

Install

Works in
Services

Manual install

curl -L -o arch-be https://github.com/romannekrasovaillm/spine-bank/releases/latest/download/arch-be-core-linux-x86_64 && chmod +x arch-be && mv arch-be ~/.local/bin/

On macOS arm64 the file is arch-be-core-macos-arm64 and on Windows arch-be-core-windows-x86_64.exe; integrity is checked against SHA256SUMS from the release.

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • arch-be connect edits .claude/settings.json and installs hooks that block agent actions when the gate fails
  • The binary is downloaded from a GitHub release and runs on the developer machine, so its integrity must be checked against SHA256SUMS

This is third-party code. Review the repository files before installing.

What it does

Spine attaches to a coding agent (Claude Code, Qwen Code, GigaCode) as the spine MCP server with 45 read-only tools, a set of hooks and 79 skills, 12 of them spine-* playbooks. The main core edition runs without its own language model: the agent thinks and the verdict is the exit code of arch-be gate. The arch-be arch-diff --base main command shows in one screen what a PR changed in the architecture: new links, external systems, contracts and touched invariants, each with file and line and no LLM. arch-be connect writes the hooks into .claude/settings.json. GigaCode has its own guide, docs/GIGACODE.md, and the sources hold the src/llm/gigachat.rs adapter.

Who it is for. Team leads and architects whose teams write code with GigaCode or Claude Code and want agent decisions checked by deterministic rules.

Good fit when

  • A reviewer wants to know what this PR changed in the architecture, with file and line
  • A GigaCode agent edits code and a gate has to reject violations of the CONSTRAINTS.yaml rules
  • The agent must not weaken rules just to go green: rule tampering is caught by the gate
  • You want lessons from past failures to reach the agent context at session start

Not a fit when

  • A regular linter is enough: Spine pays off where there is a system model and architecture rules
  • You want a tool that ships its own language model: the core edition has none
  • You cannot put a prebuilt binary from a GitHub release into the developer environment

Example request

Show the architecture diff of my branch against main and run arch-be gate on the project

Limitations

The core edition ships without its own language model, while the GigaChat adapter sits in the sources at src/llm/gigachat.rs and belongs to other editions. According to the author, the GigaCode connection was verified on qwen-code 0.0.5 and 0.24.0, and the GigaCode CLI inherits its settings from it. The binary installs from a GitHub release for Linux, macOS and Windows, and the sources are Rust. A verdict of INCOMPLETE signals an incomplete check and must not be read as green. There is one author, two stars and the documentation is mostly in Russian.

How to disable. Delete the arch-be binary from ~/.local/bin, remove the hooks block and the spine MCP server from .claude/settings.json (or the GigaCode settings) and the .arch-handoff directory in the project.

FAQ

Does it need its own language model?

Not in the core edition: your agent makes the decisions and Spine checks them with rules and returns an exit code.

Which agents are supported?

The README says six harnesses were verified; Claude Code, Qwen Code and GigaCode are named, and GigaCode has its own guide.

What does arch-diff give without setup?

A one-screen view of what a PR changed in the architecture, with file and line per fact; the model is optional and the graph is built from build manifests.

How do I verify the downloaded binary?

Compare SHA256SUMS from the same release: sha256sum --check SHA256SUMS on Linux or shasum -a 256 --check SHA256SUMS on macOS.

Can the agent approve a decision itself?

No: signing a binding and accepting rubrics stay with a person, and attempts to weaken rules are flagged by the gate as rule_weakened or self_approved.

Discussion

No questions or reviews yet.

All GigaChat tools

GigaCode Bridge MCP

Gigacode-bridge-mcp

Russian stack

MCP server that delegates isolated code edits to GigaCode CLI in temporary Git worktrees and returns a patch for independent review

MCP serverExtended access
Russian stack

A small Python MCP server for Sber's GigaChat API: authorizes with a Basic key and calls the models directly from an agent, tested on OpenCode under Windows

MCP serverExtended access

README in short

The README explains why the loop exists (checking agent code and blocking forged evidence), eight reasons, the flow from agent through hooks to the gate, the delivery editions, step 0 with the binary download and step 1 with an architecture diff in five minutes.

README badge

Report a problem