Skip to content

tool-ids.md

Latest commit

 

History

History
34 lines (26 loc) · 1.76 KB

tool-ids.md

File metadata and controls

34 lines (26 loc) · 1.76 KB
Raw

Tool ids

Every tool has one id. The id is the MCP tool name and the agent's tool name, and its kebab-case form is the CLI command.

The scheme

<area>_<object>_<action>, all lower case, words joined by _:

  • area: core, native, managed, java, android, apple, harmony, js, web, app, firmware, formats, radio, packets, db, assets, game, project, rebuild, build, runtime, adb, frida, engines, settings, plugins, setup (the list is registry::AREAS).
  • object: what it works on, plural for collections (functions, page, feed, inbox).
  • action: a plain verb (list, show, open, capture, compare, start, answer).

Examples: native_functions_list, native_function_pseudocode, web_page_capture, project_feed_show, rebuild_start, engines_ghidra_install.

Ids can't contain dots or spaces: MCP clients and model APIs accept only letters, digits, _ and - in tool names. The CLI spelling swaps _ for - (craft-anything-cli web-page-capture).

Rules

  • New tools use the scheme from the start; a test checks every id follows it.
  • The Rust function behind a tool has the tool's id as its name.
  • Text that agents and people read (prompts, summaries, errors) names tools by their ids.
  • Code never decides on a tool's name as text: it compares Registry::canonical(name) (which covers CLI names and mcp__<server>__<name>), reads a declared fact from the tool (Effects::CARRIES_TARGET_CONTENT), or uses the id constant kept next to the tool (PSEUDOCODE_TOOL, APPROVE_TOOL, SESSION_TOOL).
  • MCP prompts follow the same scheme (core_feature_explain, native_crash_explain, core_unknowns_review, runtime_run_plan, app_graphs_compare, app_rebuild_claim_check).