Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

feat(mcp): add the debug tool group and move compare_models into it

Jack Rudenko0a28db6862cf
diff --git a/packages/cli/src/cli.ts b/packages/cli/src/cli.ts
index f24e672b..dd14f8bb 100644
--- a/packages/cli/src/cli.ts
+++ b/packages/cli/src/cli.ts
@@ -2480,7 +2480,7 @@ ${h("ENVIRONMENT VARIABLES")}
${blue("CLAUDISH_DIAG_MODE")} Diagnostic output: auto / logfile / off
${blue("CLAUDISH_DEBUG")} Always enable debug logging: 1 / true ${dim("(same as -d)")}
${blue("CLAUDISH_ANTHROPIC_API_BILLING")} Bill native Claude to your API key ${dim("(see --anthropic-api-billing)")}
- ${blue("CLAUDISH_MCP_TOOLS")} MCP tool gating: all / low-level / agentic / session
+ ${blue("CLAUDISH_MCP_TOOLS")} MCP tool gating: all / low-level / agentic / session / debug
${blue("CLAUDISH_MODEL_OPUS")} Override model for Opus role
${blue("CLAUDISH_MODEL_SONNET")} Override model for Sonnet role
${blue("CLAUDISH_MODEL_HAIKU")} Override model for Haiku role
diff --git a/packages/cli/src/mcp-server.ts b/packages/cli/src/mcp-server.ts
index a8d0b843..41358827 100644
--- a/packages/cli/src/mcp-server.ts
+++ b/packages/cli/src/mcp-server.ts
@@ -157,7 +157,7 @@ The session_id in the channel tag's meta attributes is the key for all tool call
// ─── Types ───────────────────────────────────────────────────────────────────
-type ToolGroup = "low-level" | "agentic" | "session";
+type ToolGroup = "low-level" | "agentic" | "session" | "debug";
/**
* Per-call context handed to every tool handler. Built fresh in the CallTool
@@ -1181,21 +1181,26 @@ function defineTools(
tools.push({
name: "compare_models",
- description: "Run the same prompt through multiple models and compare responses",
+ description:
+ "DEBUG. Send the same single prompt to several models, one after another, and show " +
+ "their answers side by side — to check that routes answer or to eyeball raw answers. " +
+ "Each model gets exactly what run_prompt sends: one request, no tools, no Claude Code " +
+ "session. Not for real work: use team for a blind-judged multi-model run, or " +
+ "create_session for one model doing a task.",
inputSchema: {
type: "object",
properties: {
models: {
type: "array",
items: { type: "string" },
- description: "List of model IDs to compare",
+ description: "List of model names or model specs to compare, in the order to run them",
},
prompt: { type: "string", description: "The prompt to send to all models" },
system_prompt: { type: "string", description: "Optional system prompt" },
},
required: ["models", "prompt"],
},
- group: "low-level",
+ group: "debug",
// N models run SEQUENTIALLY below, so the total blocking time is the sum of
// every model's latency — the easiest of the three tools to push past the
// client's idle window.
@@ -2345,10 +2350,12 @@ function resolveToolGroups(mode: string): Set<ToolGroup> {
return new Set(["agentic"]);
case "session":
return new Set(["session"]);
+ case "debug":
+ return new Set(["debug"]);
default:
// Every other value — `all`, unset, a typo, or the removed `channel` — is
// everything. `channel` gets no alias: a rename removes the old name.
- return new Set(["low-level", "agentic", "session"]);
+ return new Set(["low-level", "agentic", "session", "debug"]);
}
}