convez.ai
MCP Server
astrojones

raisl

astrojones/raisl

Ver no GitHub
InicialPluginScore · de 0 a 100

The repo agent harness as a Claude Code plugin — a bundled, auto-connecting MCP server (repo_* tools + proxied serena_* code navigation), safety hooks, and generic coding-workflow skills/subagents. Born-harnessed repos for any MCP client.

Instalação

$npx convez-ai astrojones/raisl

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
1
Acompanhando
0
Saúde
50%
Último commit
há 2 semanas
Versão
v4.2.0
Issues abertas
16
Contribuidores
7

README

A tradução para o português deste README ainda está na fila.

astrojones

The repo agent harness as a Claude Code plugin. Install it once and every git repo you open gets safe, deterministic, repo-aware tooling and symbol-level code navigation — instead of a coding agent improvising with raw shell, and with no harness files written into the repo.

astrojones is client-agnostic in spirit: the same harness server runs for Claude Code (bundled here, auto-connecting) and — pinned into a repo's .mcp.json — for CI and other MCP clients. It is generic: nothing in it is tied to any one org. (The astrojones-specific deploy layer lives in the separate, private deploy plugin, which builds on astrojones.)

Install

/plugin marketplace add astrojones/claude-plugins
/plugin install astrojones@astrojones

That's it. The harness MCP server is bundled and auto-connects at session start — no per-repo setup.

What you get

  • A bundled, auto-connecting MCP server — deterministic repo_* tools (overview, status, search, precise range reads, impact, diff, change-verification, health, bootstrap) instead of raw shell.
  • Serena, proxied through the same server as serena_* tools for semantic symbol navigation and editing — find definitions, references, call sites; edit by symbol, not by line guesswork.
  • Zero-footprint by default — opening a repo writes no harness files into it (serena keeps its symbol index under .serena/, which it gitignores). The bundled server (tools, an always-read instructions guide, and the workflow prompts) is everything Claude Code needs. Materialize the on-disk harness (agent/ policies + tools, an AGENTS.md guide) only on demand with repo_bootstrap / /harness-init — for per-repo customization or non-MCP clients (opencode/CI).
  • Safety hooks — a safe-shell + secret-read guard (PreToolUse) and a post-edit verification nudge (PostToolUse).
  • Generic coding-workflow skills — bugfix, feature, refactor, test, implement, commit (served as MCP prompts; see the drift note below), plus the Claude-Code-only plan skill: harness-native plan mode that drives the read-only explorer and architect subagents in place of the built-in Explore/Plan agents.
  • Workflow subagents — implementer, reviewer, test-runner (TDD streams, diff review, narrow verification), explorer (read-only symbol navigation — the harness-native replacement for the built-in Explore agent; prefer it for any code exploration), and architect (read-only design/planning — the harness-native replacement for the built-in Plan agent; hands the build to implementer).

Materializing the on-disk harness (opt-in)

By default the harness writes nothing into your repos — the bundled server delivers the tools, the instructions guide, and the workflow prompts directly. Materialize the on-disk harness only when you want editable per-repo config or need to support a non-MCP client:

PathWhenMechanism
In-session, explicitOn demand — to edit agent/ policies/health, or to harness a freshly-created repoCall the repo_bootstrap MCP tool (optionally path=... to target another repo, pin=... for a project .mcp.json).
Fallback / CINon-Claude-Code clients (opencode), CI, or the MCP server is unreachable/harness-init runs the bundled CLI: uv run --project "${CLAUDE_PLUGIN_ROOT}/servers/harness-mcp" repo-agent-harness bootstrap --target both.

The workflow skills (bugfix, feature, …) and /harness-init live in the harness server as prompts (the single source of truth), exposed via @mcp.prompt() (Claude Code) and a repo_prompt_get(name) tool (clients that only surface tools, e.g. opencode). The skills/<name>/SKILL.md files are offline copies; repo_drift_check reports divergence as a warning, never an error, and repo_drift_sync refreshes them.

The discipline the harness expects

Orient with repo_context_overview; locate with Serena; read precise ranges (never dump whole files); check blast radius with repo_impact_file before a cross-file edit; edit small; then run repo_verify_changed (narrow lint/typecheck/tests for what changed, not the whole suite). Raw shell is policy-bounded — destructive commands and secret reads are denied with an actionable reason.

The server

The MCP server is a vendored Python package at servers/harness-mcp/ (repo-agent-harness), launched by run-mcp.sh via uv run --project servers/harness-mcp. It is tested in this repo:

uv run --project servers/harness-mcp pytest

Local development

claude --plugin-dir /path/to/astrojones    # test before publishing

The plugin index lives in astrojones/claude-plugins; this repo is just the plugin.

Ver o original no GitHub

Bom para

Relacionados