mirror of
https://github.com/we-promise/sure.git
synced 2026-09-09 08:34:26 +00:00
Brandon
ce92b36351
* feat(bills): assistant and MCP tools for bills Last of three chunks carved out of #3083, stacked on the UI bundle. Exposes bills to the builtin assistant and to MCP clients. Everything here is gated behind preview features, so the tools are absent from tools/list until a user opts in. Seven tools: - get_bills, get_bill_details and get_paycheck_plan for reads - get_bill_audit, a deterministic review that surfaces likely duplicates, price changes, trials about to convert, upcoming renewals and long-overdue bills - create_bill, update_bill and record_bill_payment for writes Shared argument parsing, permission checks and error shapes live in BillsSupport, so every tool answers with the same {error, hint} contract the existing tools use, and a bad argument never aborts the turn. The write tools mutate financial records on a model's say-so, so they refuse rather than guess: a payment cannot exceed what its cycle still owes, a repeated settle will not quietly close next month, an unrecognized frequency is an error instead of a silent monthly default, and non-finite or negative amounts are rejected before they reach the database. The read tools say what they filtered. An empty result names the statuses that do hold matches, the paycheck plan discloses the unconfirmed series it excluded from spending headroom, and history and price-change windows report their real totals rather than letting a caller sum a truncated list. A not-found no longer returns the scoped relation's SQL, which handed any MCP client the access-control schema for the cost of a guessed id. The in-page AI helpers are not here. Smart fill and smart configuration are buttons on the bills pages, so they ship with the UI bundle along with the provider-side suggester they call. Suite 7,854 runs, 0 failures. Rubocop clean, eager loading verified. * Address the ready-review round * Reject an out-of-range audit lookback out loud * Speak the cycle remainder guard through the allocator locale
97 lines
3.2 KiB
Ruby
97 lines
3.2 KiB
Ruby
require "test_helper"
|
|
|
|
class Assistant::ResponderTest < ActiveSupport::TestCase
|
|
include ProviderTestHelper
|
|
|
|
class EchoFunction < Assistant::Function
|
|
def self.name = "echo"
|
|
def self.description = "Echoes the received arguments"
|
|
def call(params = {}) = params
|
|
end
|
|
|
|
setup do
|
|
@chat = chats(:two)
|
|
@message = @chat.messages.create!(
|
|
type: "UserMessage",
|
|
content: "What is my net worth?",
|
|
ai_model: "gpt-4.1"
|
|
)
|
|
@llm = mock
|
|
@llm.stubs(:supports_responses_endpoint?).returns(true)
|
|
@responder = Assistant::Responder.new(
|
|
message: @message,
|
|
instructions: "instructions",
|
|
function_tool_caller: Assistant::FunctionToolCaller.new([ EchoFunction.new(nil) ]),
|
|
llm: @llm
|
|
)
|
|
end
|
|
|
|
test "default iteration cap is eight rounds" do
|
|
assert_equal 8, Assistant::Responder::DEFAULT_MAX_TOOL_CALL_ITERATIONS
|
|
end
|
|
|
|
test "final round requests forbid tool calls via tool_choice while keeping tool definitions" do
|
|
function_request = Provider::LlmConcept::ChatFunctionRequest.new(
|
|
id: "1", call_id: "1", function_name: "echo", function_args: "{}"
|
|
)
|
|
tool_response = Provider::LlmConcept::ChatResponse.new(
|
|
id: "1", model: "gpt-4.1", messages: [], function_requests: [ function_request ]
|
|
)
|
|
text_response = Provider::LlmConcept::ChatResponse.new(
|
|
id: "2", model: "gpt-4.1",
|
|
messages: [ Provider::LlmConcept::ChatMessage.new(id: "2", output_text: "Here is what I found") ],
|
|
function_requests: []
|
|
)
|
|
|
|
functions_seen = []
|
|
tool_choices_seen = []
|
|
|
|
@llm.stubs(:chat_response).with do |_prompt, **kwargs|
|
|
functions_seen << kwargs[:functions]
|
|
tool_choices_seen << kwargs[:tool_choice]
|
|
true
|
|
end.returns(
|
|
provider_success_response(tool_response),
|
|
provider_success_response(tool_response),
|
|
provider_success_response(text_response)
|
|
)
|
|
|
|
with_iteration_cap(2) do
|
|
@responder.respond
|
|
end
|
|
|
|
# Every request carries the real tool definitions — Anthropic rejects
|
|
# messages containing tool blocks when no tools are defined — and only the
|
|
# request after the final permitted round forbids further calls.
|
|
assert_operator functions_seen.size, :>=, 2
|
|
assert functions_seen.last.present?
|
|
assert_equal functions_seen.first, functions_seen.last
|
|
assert_equal :none, tool_choices_seen.last
|
|
assert tool_choices_seen[0..-2].all?(&:nil?)
|
|
end
|
|
|
|
test "a model that never stops calling tools hits the cap and raises" do
|
|
function_request = Provider::LlmConcept::ChatFunctionRequest.new(
|
|
id: "1", call_id: "1", function_name: "echo", function_args: "{}"
|
|
)
|
|
tool_response = Provider::LlmConcept::ChatResponse.new(
|
|
id: "1", model: "gpt-4.1", messages: [], function_requests: [ function_request ]
|
|
)
|
|
|
|
@llm.stubs(:chat_response).returns(provider_success_response(tool_response))
|
|
|
|
with_iteration_cap(2) do
|
|
assert_raises(Assistant::Responder::ToolCallLimitError) { @responder.respond }
|
|
end
|
|
end
|
|
|
|
private
|
|
def with_iteration_cap(value)
|
|
previous = ENV["ASSISTANT_MAX_TOOL_CALL_ITERATIONS"]
|
|
ENV["ASSISTANT_MAX_TOOL_CALL_ITERATIONS"] = value.to_s
|
|
yield
|
|
ensure
|
|
ENV["ASSISTANT_MAX_TOOL_CALL_ITERATIONS"] = previous
|
|
end
|
|
end
|