Claude Code

Claude Code decides which tool to call from a skill's description and its supporting files. Without one naming Rift's tools, an agent falls back to grepping the repository even when a rift mcp server already indexes it. rift install claude writes that skill.

Install

Run this from a codebase Rift serves:

rift install claude

The command writes SKILL.md and a references/tools.md sidecar under .claude/skills/rift/ in the current workspace. Claude Code reads SKILL.md's frontmatter to decide when to load the skill, and loads the sidecar on demand for each tool's parameters.

Pass --user to write under the operator's home directory instead of the workspace, so every project the operator opens with Claude Code picks up the skill:

rift install claude --user

Pass --remove to delete the generated skill from either scope:

rift install claude --remove
rift install claude --user --remove

Install as a plugin

The repository doubles as a Claude Code plugin marketplace, so the skill installs without the rift CLI:

/plugin marketplace add volarized/rift
/plugin install rift@rift

The plugin ships the same generated skill plus an .mcp.json naming the rift mcp server, so one install covers every project the operator opens. The plugin's server entry is additive and scoped to the plugin: Claude Code leaves the operator's own .mcp.json and settings untouched, and uninstalling the plugin removes the entry with it.

The rift executable is still what serves the tools. When it is missing, Claude Code skips the plugin's server entry and the skill's own install section names the installer commands; the quickstart carries the same commands.

The steering hook stays with rift install claude: the plugin carries the skill and the server entry alone. A workspace already carrying a rift entry written by rift install claude runs a second proxy beside the plugin's; both reach the same workspace server. Keep one: remove the project entry or disable the plugin for that project.

What the skill teaches

The generated SKILL.md carries a decision table pointing the agent at the tool that answers its situation:

• search answers when the target is unknown.
• get_symbol answers when the declaration name is known.
• nodes answers when the syntax structure at one position is needed.
• search with a traversal block answers a symbol's neighbors, its impact (who breaks when it changes), or a path between two symbols.
• The edit tools - patch, replace_node, insert_node, insert_symbol, replace_symbol, rename_symbol, remove_node, remove_symbol, and move_file - answer when the code needs to change, over raw file writes: the server recomputes each address's witness before writing and runs the workspace's configured hooks.

The skill also points the agent at rift://map as the first read in an unfamiliar repository, and at rift://logs when a refusal alone doesn't say why.

Steering

A skill changes what an agent reaches for; habit still wins the first time it needs to search. Claude Code agents call Grep and Glob from training, even in a workspace rift mcp already indexes. rift steer catches that first call and redirects it.

rift install claude writes a PreToolUse hook alongside the skill:

{
  "hooks": {
    "PreToolUse": [
      {
        "hooks": [
          {
            "command": "rift steer",
            "type": "command"
          }
        ],
        "matcher": "Grep|Glob"
      }
    ]
  }
}

Claude Code runs rift steer before the first Grep or Glob call in a session and pipes it the tool call as JSON on stdin. In a workspace Rift has indexed and version control tracks, that first call is denied: the reason names the search call that answers the same question, get_symbol for a known name, and rift://map for orientation.

The same Grep or Glob call passes if the agent retries it, and every later call in that session passes without a prompt. rift steer marks the session with a file under .rift/steer/, so it denies at most once per session.

Every other case passes without a prompt too: a workspace with no index yet, one with no version control, and a session that already saw its one redirect.

Set RIFT_STEER=0 in the environment Claude Code runs in to disable the hook entirely; every call then passes.

rift install claude --remove strips the hook entry from .claude/settings.json, leaving every other entry in the file untouched.

Regeneration

rift install claude reads the same tool router rift mcp serves, so the skill's tool names can never drift from what this build actually serves. Rerun the command after every upgrade; it overwrites the previous skill with the current one.