backend_fiesta/services/agents_test.go

package services

import (
	"os"
	"path/filepath"
	"strings"
	"testing"

	"nearle/services/tools"
	"nearle/utils"
)

// realRegistry builds every tool the server registers, exactly as the facade
// does — with nil services, because nothing here calls a handler.
//
// Replaces a hand-written list of tool names that went stale twice: once when
// the help corpus arrived and once when the first write did. Both times the
// shipped agents were correct and the TEST was wrong, which is the worst way
// round — a check that has to be remembered is a check that will not be.
func realRegistry(t *testing.T) *tools.Registry {
	t.Helper()

	corpus, err := tools.LoadHelp()
	if err != nil {
		t.Fatalf("loading the help corpus: %v", err)
	}

	r := tools.New(nil)
	for _, tool := range []tools.Tool{
		tools.StuckOrders(nil, nil),
		tools.DeliveryProgress(nil),
		tools.BranchPerformance(nil),
		tools.PendingApprovals(nil, nil),
		tools.LowStock(nil),
		tools.TillsNotSyncing(nil),
		tools.SalesByChannel(nil, nil, nil),
		tools.Help(corpus),
		tools.ApproveStockRequest(nil, nil),
	} {
		if err := r.Register(tool); err != nil {
			t.Fatalf("registering %s: %v", tool.Name, err)
		}
	}
	return r
}

// knownTool asks the real registry, so an agent naming a tool that does not
// exist fails here and not on somebody's deployment.
func knownTool(name string) bool { return liveRegistry.Has(name) }

// liveRegistry is built once, in TestMain, because knownTool has no *testing.T.
var liveRegistry *tools.Registry

func TestMain(m *testing.M) {
	fake := &testing.T{}
	liveRegistry = realRegistry(fake)
	os.Exit(m.Run())
}

// writeAgents drops YAML into a temporary directory and loads it.
func writeAgents(t *testing.T, files map[string]string) (map[string]Agent, error) {
	t.Helper()
	dir := t.TempDir()
	for name, body := range files {
		if err := os.WriteFile(filepath.Join(dir, name), []byte(body), 0o600); err != nil {
			t.Fatalf("writing %s: %v", name, err)
		}
	}
	return LoadAgents(dir, knownTool)
}

/* ── The agents that ship ──────────────────────────────────────────────── */

func TestTheEmbeddedAgentsLoad(t *testing.T) {
	// The defaults are compiled in so a binary always has working agents and a
	// deployment cannot be broken by a missing directory.
	agents, err := LoadAgents("", knownTool)
	if err != nil {
		t.Fatalf("the shipped agents do not load: %v", err)
	}
	for _, want := range []string{"orders", "inventory", "shopfloor"} {
		if _, ok := agents[want]; !ok {
			t.Fatalf("%q is missing: %v", want, AgentNames(agents))
		}
	}
}

func TestEveryShippedAgentNamesOnlyRealTools(t *testing.T) {
	// The check that makes a typo a startup failure rather than an agent that
	// silently cannot do one of the things it claims.
	if _, err := LoadAgents("", knownTool); err != nil {
		t.Fatalf("a shipped agent names a tool that does not exist: %v", err)
	}
}

func TestAddingAnAgentTakesNoCodeChange(t *testing.T) {
	// Phase 3's whole point, stated as a test.
	agents, err := writeAgents(t, map[string]string{
		"sixth.yaml": "name: sixth\ntier: fast\ntools:\n  - low_stock\n",
	})
	if err != nil {
		t.Fatalf("loading: %v", err)
	}
	if _, ok := agents["sixth"]; !ok {
		t.Fatal("a new file did not become an agent")
	}
}

/* ── The shared prompt ─────────────────────────────────────────────────── */

func TestEveryAgentInheritsTheRulesThatMustNotDrift(t *testing.T) {
	// Repeating "say when a result was truncated" in three files is three
	// chances for one of them to lose it.
	agents, err := LoadAgents("", knownTool)
	if err != nil {
		t.Fatalf("loading: %v", err)
	}
	for name, agent := range agents {
		if !strings.Contains(agent.System, "truncated") {
			t.Fatalf("agent %q was not told about truncated results", name)
		}
		if !strings.Contains(agent.System, "never estimate") {
			t.Fatalf("agent %q was not told to answer only from tools", name)
		}
	}
}

func TestAnAgentsOwnWordsAreAddedNotSubstituted(t *testing.T) {
	agents, err := writeAgents(t, map[string]string{
		"one.yaml": "name: one\ntools:\n  - low_stock\nsystem: |\n  Mind the gap.\n",
	})
	if err != nil {
		t.Fatalf("loading: %v", err)
	}
	system := agents["one"].System
	if !strings.Contains(system, "Mind the gap.") {
		t.Fatal("the agent's own words were dropped")
	}
	if !strings.Contains(system, "never estimate") {
		t.Fatal("the agent's own words replaced the shared rules")
	}
}

/* ── What must not load quietly ────────────────────────────────────────── */

func TestATypoInAToolNameStopsTheServer(t *testing.T) {
	// Invisible at runtime otherwise: the agent never calls it, the model says
	// it cannot look something up, and everything reports healthy.
	_, err := writeAgents(t, map[string]string{
		"one.yaml": "name: one\ntools:\n  - stcuk_orders\n",
	})
	if err == nil {
		t.Fatal("an agent naming a tool that does not exist was accepted")
	}
	if !strings.Contains(err.Error(), "stcuk_orders") {
		t.Fatalf("the error does not name the typo: %v", err)
	}
}

func TestAnAgentWithNoToolsIsRefused(t *testing.T) {
	// It could only answer from its prompt, which is the one thing this
	// assistant is built not to do.
	if _, err := writeAgents(t, map[string]string{"one.yaml": "name: one\ntier: fast\n"}); err == nil {
		t.Fatal("an agent with no tools was accepted")
	}
}

func TestAnUnknownTierIsRefusedRatherThanFallingBack(t *testing.T) {
	// The gateway falls back to the balanced model for anything it does not
	// recognise, so a "deep" agent with a typo would run on the cheap model for
	// months with nothing to show for it.
	if _, err := writeAgents(t, map[string]string{
		"one.yaml": "name: one\ntier: thorough\ntools:\n  - low_stock\n",
	}); err == nil {
		t.Fatal("an unknown tier was accepted")
	}
}

func TestAFileThatDoesNotNameItsAgentIsRefused(t *testing.T) {
	// Named by contents, not filename: a file renamed on deploy would otherwise
	// silently become a different agent.
	if _, err := writeAgents(t, map[string]string{
		"one.yaml": "tier: fast\ntools:\n  - low_stock\n",
	}); err == nil {
		t.Fatal("a file with no agent name was accepted")
	}
}

func TestTwoAgentsCannotShareAName(t *testing.T) {
	// One of them is being ignored, and which depends on directory order.
	_, err := writeAgents(t, map[string]string{
		"a.yaml": "name: same\ntools:\n  - low_stock\n",
		"b.yaml": "name: same\ntools:\n  - till_status\n",
	})
	if err == nil {
		t.Fatal("two agents took the same name")
	}
}

func TestBrokenYAMLIsRefusedWithTheFileNamed(t *testing.T) {
	_, err := writeAgents(t, map[string]string{
		"one.yaml": "name: one\n  tools: [\n",
	})
	if err == nil {
		t.Fatal("unparseable YAML was accepted")
	}
	if !strings.Contains(err.Error(), "one.yaml") {
		t.Fatalf("the error does not say which file: %v", err)
	}
}

func TestAnEmptyDirectoryIsRefusedNotTreatedAsNoAgents(t *testing.T) {
	// Zero agents means every question answers "no assistant called that",
	// which reads as a bug in the console rather than a missing config.
	if _, err := LoadAgents(t.TempDir(), knownTool); err == nil {
		t.Fatal("an empty agents directory was accepted")
	}
}

func TestAMissingDirectoryIsAnError(t *testing.T) {
	if _, err := LoadAgents(filepath.Join(t.TempDir(), "nope"), knownTool); err == nil {
		t.Fatal("a missing agents directory was accepted")
	}
}

/* ── Defaults ──────────────────────────────────────────────────────────── */

func TestAnAgentThatNamesNoLimitsGetsSensibleOnes(t *testing.T) {
	// Zero would mean a loop that runs no steps at all and answers nothing.
	agents, err := writeAgents(t, map[string]string{
		"one.yaml": "name: one\ntools:\n  - low_stock\n",
	})
	if err != nil {
		t.Fatalf("loading: %v", err)
	}
	agent := agents["one"]
	if agent.MaxSteps <= 0 || agent.MaxToolCalls <= 0 {
		t.Fatalf("limits are unusable: %+v", agent)
	}
	if agent.Tier != utils.TierBalanced {
		t.Fatalf("an agent with no tier got %q", agent.Tier)
	}
}

func TestTheDiskDirectoryReplacesTheEmbeddedAgentsRatherThanMerging(t *testing.T) {
	// Merging would leave a deployment guessing which half of an agent it is
	// running.
	agents, err := writeAgents(t, map[string]string{
		"only.yaml": "name: only\ntools:\n  - low_stock\n",
	})
	if err != nil {
		t.Fatalf("loading: %v", err)
	}
	if len(agents) != 1 {
		t.Fatalf("the embedded agents leaked in: %v", AgentNames(agents))
	}
}

func TestAgentNamesAreSortedSoTwoDeploymentsLogTheSameLine(t *testing.T) {
	names := AgentNames(map[string]Agent{"zulu": {}, "alpha": {}, "mike": {}})
	if names[0] != "alpha" || names[2] != "zulu" {
		t.Fatalf("not sorted: %v", names)
	}
}

// The platform agent may only hold tools a caller with no tenant can run.
//
// Nearle's own staff are `issuperadmin`: they read across every merchant and
// belong to none, so `tools.Caller.Tenantid` is zero and `satisfies` refuses
// anything declaring RequiresTenant or RequiresBranch. An agent holding those
// would accept every question and answer "pick a shop first" to all of them,
// which reads as a broken assistant rather than a scoped one.
//
// Asserted rather than trusted, because adding a tool to a YAML file is one
// line and the failure it produces is invisible until somebody in the platform
// workspace asks a question.
func TestThePlatformAgentOnlyHoldsToolsThatNeedNoTenant(t *testing.T) {
	registry := realRegistry(t)

	agents, err := LoadAgents("", registry.Has)
	if err != nil {
		t.Fatalf("agents: %v", err)
	}

	platform, ok := agents["platform"]
	if !ok {
		t.Fatal("no platform agent — Nearle Admin has no assistant at all")
	}
	if len(platform.Tools) == 0 {
		t.Fatal("the platform agent holds no tools")
	}

	for _, name := range platform.Tools {
		tool, ok := registry.Tool(name)
		if !ok {
			t.Fatalf("platform names %q, which is not in the registry", name)
		}
		if tool.Needs != tools.RequiresNothing {
			t.Fatalf(
				"platform holds %q, which needs a shop — a platform account has none, so every "+
					"question using it is refused", name)
		}
	}
}