ConduitMcp.Client.ToolCall (ConduitMCP v0.11.0)

One intended tools/call: which server, which tool, with which arguments.

This is the value hooks, policy and approvers see. args_digest is the lower-case hex SHA-256 of the arguments' canonical JSON (object keys sorted recursively, atom and string keys treated alike), so an approval can be bound to the exact arguments that were shown and a later change to them is detectable.

meta is user _meta for the request; reserved keys are added (and win) when the request is encoded, not here.

call_id is the LLM provider's id for the call (Anthropic tool_use.id, OpenAI call_id, Gemini functionCall.id), set by the ConduitMcp.Client.Adapters.* modules so a result can be routed back to the model. It is never sent to the server; it is kept out of meta because meta is forwarded as the request's _meta.

Summary

Types

t()

Functions

new(server_id, name, args, opts \\ [])

Builds a call and computes its args_digest.

put_args(call, args)

Replaces the arguments and recomputes the digest (used when a pre-hook rewrites them).

Types

t()

@type t() :: %ConduitMcp.Client.ToolCall{
  args: map(),
  args_digest: String.t(),
  call_id: String.t() | nil,
  meta: map(),
  name: String.t(),
  server_id: String.t()
}

Functions

new(server_id, name, args, opts \\ [])

@spec new(String.t(), String.t(), map(), keyword()) :: t()

Builds a call and computes its args_digest.

Raises when args is not JSON-encodable (including when it contains a ConduitMcp.Client.Secret).

Options

:meta - user _meta for the request (default %{}).
:call_id - the LLM provider's call id (default nil).

Example

iex> a = ConduitMcp.Client.ToolCall.new("files", "read", %{"path" => "/a", "n" => 1})
iex> b = ConduitMcp.Client.ToolCall.new("files", "read", %{n: 1, path: "/a"})
iex> a.args_digest == b.args_digest
true

put_args(call, args)

@spec put_args(t(), map()) :: t()

Replaces the arguments and recomputes the digest (used when a pre-hook rewrites them).