mirror of
https://github.com/netbirdio/netbird.git
synced 2026-09-01 12:31:42 +02:00
Merge branch 'main' into fix/remove-math-rand
This commit is contained in:
@@ -18,21 +18,26 @@ import (
|
||||
// contract between the proxy and management; management flattens them into
|
||||
// queryable columns. Keep in sync with the proxy side.
|
||||
const (
|
||||
metaKeyProvider = "llm.provider"
|
||||
metaKeyModel = "llm.model"
|
||||
metaKeyResolvedProviderID = "llm.resolved_provider_id"
|
||||
metaKeySelectedPolicyID = "llm.selected_policy_id"
|
||||
metaKeyPolicyDecision = "llm_policy.decision"
|
||||
metaKeyPolicyReason = "llm_policy.reason"
|
||||
metaKeyInputTokens = "llm.input_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyOutputTokens = "llm.output_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyTotalTokens = "llm.total_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyCostUSDTotal = "cost.usd_total"
|
||||
metaKeyStream = "llm.stream"
|
||||
metaKeySessionID = "llm.session_id"
|
||||
metaKeyAuthorisingGroups = "llm.authorising_groups"
|
||||
metaKeyRequestPrompt = "llm.request_prompt"
|
||||
metaKeyResponseCompletion = "llm.response_completion"
|
||||
metaKeyProvider = "llm.provider"
|
||||
metaKeyModel = "llm.model"
|
||||
metaKeyResolvedProviderID = "llm.resolved_provider_id"
|
||||
metaKeySelectedPolicyID = "llm.selected_policy_id"
|
||||
metaKeyPolicyDecision = "llm_policy.decision"
|
||||
metaKeyPolicyReason = "llm_policy.reason"
|
||||
metaKeyInputTokens = "llm.input_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyOutputTokens = "llm.output_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyTotalTokens = "llm.total_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyCachedInputTokens = "llm.cached_input_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyCacheCreationTokens = "llm.cache_creation_tokens" //nolint:gosec // metadata key name, not a credential
|
||||
metaKeyCostUSDInput = "cost.usd_input"
|
||||
metaKeyCostUSDCachedInput = "cost.usd_cached_input"
|
||||
metaKeyCostUSDCacheCreate = "cost.usd_cache_creation"
|
||||
metaKeyCostUSDOutput = "cost.usd_output"
|
||||
metaKeyStream = "llm.stream"
|
||||
metaKeySessionID = "llm.session_id"
|
||||
metaKeyAuthorisingGroups = "llm.authorising_groups"
|
||||
metaKeyRequestPrompt = "llm.request_prompt"
|
||||
metaKeyResponseCompletion = "llm.response_completion"
|
||||
)
|
||||
|
||||
// IngestAccessLog flattens the metadata-bearing reverse-proxy access-log entry
|
||||
@@ -108,20 +113,25 @@ func flattenAccessLog(e *accesslogs.AccessLogEntry) (*types.AgentNetworkAccessLo
|
||||
BytesUpload: e.BytesUpload,
|
||||
BytesDownload: e.BytesDownload,
|
||||
|
||||
Provider: meta[metaKeyProvider],
|
||||
Model: meta[metaKeyModel],
|
||||
SessionID: meta[metaKeySessionID],
|
||||
ResolvedProviderID: meta[metaKeyResolvedProviderID],
|
||||
SelectedPolicyID: meta[metaKeySelectedPolicyID],
|
||||
Decision: meta[metaKeyPolicyDecision],
|
||||
DenyReason: meta[metaKeyPolicyReason],
|
||||
InputTokens: parseMetaInt(meta, metaKeyInputTokens),
|
||||
OutputTokens: parseMetaInt(meta, metaKeyOutputTokens),
|
||||
TotalTokens: parseMetaInt(meta, metaKeyTotalTokens),
|
||||
CostUSD: parseMetaFloat(meta, metaKeyCostUSDTotal),
|
||||
Stream: parseMetaBool(meta, metaKeyStream),
|
||||
RequestPrompt: meta[metaKeyRequestPrompt],
|
||||
ResponseCompletion: meta[metaKeyResponseCompletion],
|
||||
Provider: meta[metaKeyProvider],
|
||||
Model: meta[metaKeyModel],
|
||||
SessionID: meta[metaKeySessionID],
|
||||
ResolvedProviderID: meta[metaKeyResolvedProviderID],
|
||||
SelectedPolicyID: meta[metaKeySelectedPolicyID],
|
||||
Decision: meta[metaKeyPolicyDecision],
|
||||
DenyReason: meta[metaKeyPolicyReason],
|
||||
InputTokens: parseMetaInt(meta, metaKeyInputTokens),
|
||||
OutputTokens: parseMetaInt(meta, metaKeyOutputTokens),
|
||||
TotalTokens: parseMetaInt(meta, metaKeyTotalTokens),
|
||||
CachedInputTokens: parseMetaInt(meta, metaKeyCachedInputTokens),
|
||||
CacheCreationTokens: parseMetaInt(meta, metaKeyCacheCreationTokens),
|
||||
InputCostUSD: parseMetaFloat(meta, metaKeyCostUSDInput),
|
||||
CachedInputCostUSD: parseMetaFloat(meta, metaKeyCostUSDCachedInput),
|
||||
CacheCreationCostUSD: parseMetaFloat(meta, metaKeyCostUSDCacheCreate),
|
||||
OutputCostUSD: parseMetaFloat(meta, metaKeyCostUSDOutput),
|
||||
Stream: parseMetaBool(meta, metaKeyStream),
|
||||
RequestPrompt: meta[metaKeyRequestPrompt],
|
||||
ResponseCompletion: meta[metaKeyResponseCompletion],
|
||||
}
|
||||
|
||||
var groups []types.AgentNetworkAccessLogGroup
|
||||
@@ -140,18 +150,23 @@ func flattenAccessLog(e *accesslogs.AccessLogEntry) (*types.AgentNetworkAccessLo
|
||||
// log's ID so the two correlate.
|
||||
func usageFromFlattenedLog(e *types.AgentNetworkAccessLog, groups []types.AgentNetworkAccessLogGroup) (*types.AgentNetworkUsage, []types.AgentNetworkUsageGroup) {
|
||||
usage := &types.AgentNetworkUsage{
|
||||
ID: e.ID,
|
||||
AccountID: e.AccountID,
|
||||
Timestamp: e.Timestamp,
|
||||
UserID: e.UserID,
|
||||
ResolvedProviderID: e.ResolvedProviderID,
|
||||
Provider: e.Provider,
|
||||
Model: e.Model,
|
||||
SessionID: e.SessionID,
|
||||
InputTokens: e.InputTokens,
|
||||
OutputTokens: e.OutputTokens,
|
||||
TotalTokens: e.TotalTokens,
|
||||
CostUSD: e.CostUSD,
|
||||
ID: e.ID,
|
||||
AccountID: e.AccountID,
|
||||
Timestamp: e.Timestamp,
|
||||
UserID: e.UserID,
|
||||
ResolvedProviderID: e.ResolvedProviderID,
|
||||
Provider: e.Provider,
|
||||
Model: e.Model,
|
||||
SessionID: e.SessionID,
|
||||
InputTokens: e.InputTokens,
|
||||
OutputTokens: e.OutputTokens,
|
||||
TotalTokens: e.TotalTokens,
|
||||
CachedInputTokens: e.CachedInputTokens,
|
||||
CacheCreationTokens: e.CacheCreationTokens,
|
||||
InputCostUSD: e.InputCostUSD,
|
||||
CachedInputCostUSD: e.CachedInputCostUSD,
|
||||
CacheCreationCostUSD: e.CacheCreationCostUSD,
|
||||
OutputCostUSD: e.OutputCostUSD,
|
||||
}
|
||||
|
||||
usageGroups := make([]types.AgentNetworkUsageGroup, 0, len(groups))
|
||||
|
||||
@@ -28,17 +28,22 @@ func newIngestTestEntry() *accesslogs.AccessLogEntry {
|
||||
UserId: "user-1",
|
||||
AgentNetwork: true,
|
||||
Metadata: map[string]string{
|
||||
metaKeyProvider: "openai",
|
||||
metaKeyModel: "gpt-5.4",
|
||||
metaKeyResolvedProviderID: "prov-1",
|
||||
metaKeySessionID: "sess-1",
|
||||
metaKeyInputTokens: "100",
|
||||
metaKeyOutputTokens: "50",
|
||||
metaKeyTotalTokens: "150",
|
||||
metaKeyCostUSDTotal: "0.0123",
|
||||
metaKeyStream: "true",
|
||||
metaKeyRequestPrompt: "hello",
|
||||
metaKeyResponseCompletion: "world",
|
||||
metaKeyProvider: "openai",
|
||||
metaKeyModel: "gpt-5.4",
|
||||
metaKeyResolvedProviderID: "prov-1",
|
||||
metaKeySessionID: "sess-1",
|
||||
metaKeyInputTokens: "100",
|
||||
metaKeyOutputTokens: "50",
|
||||
metaKeyTotalTokens: "1174",
|
||||
metaKeyCachedInputTokens: "256",
|
||||
metaKeyCacheCreationTokens: "768",
|
||||
metaKeyCostUSDInput: "0.0071",
|
||||
metaKeyCostUSDCachedInput: "0.0009",
|
||||
metaKeyCostUSDCacheCreate: "0.0020",
|
||||
metaKeyCostUSDOutput: "0.0023",
|
||||
metaKeyStream: "true",
|
||||
metaKeyRequestPrompt: "hello",
|
||||
metaKeyResponseCompletion: "world",
|
||||
// repeated id must be de-duplicated before the group rows insert.
|
||||
metaKeyAuthorisingGroups: "grp-eng,grp-eng,grp-ops",
|
||||
},
|
||||
@@ -65,7 +70,19 @@ func TestIngestAccessLog_RealStore_LogCollectionOff(t *testing.T) {
|
||||
require.Len(t, usage, 1, "usage row must be written even with log collection off")
|
||||
assert.Equal(t, int64(100), usage[0].InputTokens, "input tokens must round-trip from metadata")
|
||||
assert.Equal(t, int64(50), usage[0].OutputTokens, "output tokens must round-trip from metadata")
|
||||
assert.InDelta(t, 0.0123, usage[0].CostUSD, 1e-9, "cost must round-trip from metadata")
|
||||
assert.Equal(t, int64(256), usage[0].CachedInputTokens, "cache-read tokens must round-trip from metadata")
|
||||
assert.Equal(t, int64(768), usage[0].CacheCreationTokens, "cache-write tokens must round-trip from metadata")
|
||||
// The per-bucket breakdown is the only cost state stored, and must survive
|
||||
// the write/read cycle as real columns — usage rows are the only cost
|
||||
// record for accounts with log collection off, so a dropped column here
|
||||
// loses the split permanently.
|
||||
assert.InDelta(t, 0.0071, usage[0].InputCostUSD, 1e-9, "input cost must round-trip from metadata")
|
||||
assert.InDelta(t, 0.0009, usage[0].CachedInputCostUSD, 1e-9, "cache-read cost must round-trip from metadata")
|
||||
assert.InDelta(t, 0.0020, usage[0].CacheCreationCostUSD, 1e-9, "cache-write cost must round-trip from metadata")
|
||||
assert.InDelta(t, 0.0023, usage[0].OutputCostUSD, 1e-9, "output cost must round-trip from metadata")
|
||||
// Aggregates are derived from the stored columns, never stored themselves.
|
||||
assert.InDelta(t, 0.0123, usage[0].TotalCostUSD(), 1e-9, "total is derived from the stored buckets")
|
||||
assert.InDelta(t, 0.0029, usage[0].CacheCostUSD(), 1e-9, "cache cost is derived from the two cache buckets")
|
||||
|
||||
logs, _, err := s.GetAgentNetworkAccessLogs(ctx, store.LockingStrengthNone, testAccountID, types.AgentNetworkAccessLogFilter{})
|
||||
require.NoError(t, err)
|
||||
@@ -96,6 +113,14 @@ func TestIngestAccessLog_RealStore_LogCollectionOn(t *testing.T) {
|
||||
require.Equal(t, int64(1), total, "exactly one access-log row expected")
|
||||
require.Len(t, logs, 1, "full access-log row must be written when log collection is on")
|
||||
assert.Equal(t, "gpt-5.4", logs[0].Model, "model must flatten from metadata")
|
||||
assert.Equal(t, int64(256), logs[0].CachedInputTokens, "cache-read tokens must flatten from metadata")
|
||||
assert.Equal(t, int64(768), logs[0].CacheCreationTokens, "cache-write tokens must flatten from metadata")
|
||||
assert.InDelta(t, 0.0029, logs[0].CacheCostUSD(), 1e-9, "cache cost is derived from the two cache buckets")
|
||||
assert.InDelta(t, 0.0123, logs[0].TotalCostUSD(), 1e-9, "total is derived from the stored buckets")
|
||||
assert.InDelta(t, 0.0071, logs[0].InputCostUSD, 1e-9, "input cost must flatten from metadata")
|
||||
assert.InDelta(t, 0.0009, logs[0].CachedInputCostUSD, 1e-9, "cache-read cost must flatten from metadata")
|
||||
assert.InDelta(t, 0.0020, logs[0].CacheCreationCostUSD, 1e-9, "cache-write cost must flatten from metadata")
|
||||
assert.InDelta(t, 0.0023, logs[0].OutputCostUSD, 1e-9, "output cost must flatten from metadata")
|
||||
assert.Equal(t, "hello", logs[0].RequestPrompt, "prompt must be retained when log collection is on")
|
||||
assert.Equal(t, "world", logs[0].ResponseCompletion, "completion must be retained when log collection is on")
|
||||
assert.True(t, logs[0].Stream, "stream flag must flatten from metadata")
|
||||
|
||||
@@ -38,7 +38,7 @@ func accessLogRow(id, sessionID string, ts time.Time, opts ...func(*types.AgentN
|
||||
InputTokens: 100,
|
||||
OutputTokens: 50,
|
||||
TotalTokens: 150,
|
||||
CostUSD: 0.01,
|
||||
InputCostUSD: 0.01,
|
||||
}
|
||||
for _, o := range opts {
|
||||
o(e)
|
||||
@@ -74,7 +74,7 @@ func withTokens(in, out, total int64, cost float64) func(*types.AgentNetworkAcce
|
||||
e.InputTokens = in
|
||||
e.OutputTokens = out
|
||||
e.TotalTokens = total
|
||||
e.CostUSD = cost
|
||||
e.InputCostUSD = cost
|
||||
}
|
||||
}
|
||||
|
||||
@@ -155,7 +155,7 @@ func TestAccessLogSessions_FoldAndAggregate(t *testing.T) {
|
||||
assert.Equal(t, int64(310), a.InputTokens, "input tokens summed")
|
||||
assert.Equal(t, int64(135), a.OutputTokens, "output tokens summed")
|
||||
assert.Equal(t, int64(445), a.TotalTokens, "total tokens summed")
|
||||
assert.InDelta(t, 0.031, a.CostUSD, 1e-9, "cost summed")
|
||||
assert.InDelta(t, 0.031, a.TotalCostUSD(), 1e-9, "cost summed")
|
||||
assert.Equal(t, "deny", a.Decision, "any deny makes the session a deny")
|
||||
assert.ElementsMatch(t, []string{"openai", "anthropic"}, a.Providers, "distinct providers")
|
||||
assert.ElementsMatch(t, []string{"gpt-5.4", "claude-haiku-4-5"}, a.Models, "distinct models")
|
||||
|
||||
@@ -420,6 +420,47 @@ var providers = []Provider{
|
||||
{ID: "mistral-embed", Label: "Mistral Embed", InputPer1k: 0.0001, OutputPer1k: 0, ContextWindow: 8192},
|
||||
},
|
||||
},
|
||||
{
|
||||
ID: "kimi_api",
|
||||
Kind: KindProvider,
|
||||
Name: "Kimi (Moonshot AI) API",
|
||||
Description: "Kimi K3 / K2 models via the Moonshot AI platform",
|
||||
DefaultHost: "api.moonshot.ai",
|
||||
AuthHeaderName: "Authorization",
|
||||
AuthHeaderTemplate: "Bearer ${API_KEY}",
|
||||
DefaultContentType: "application/json",
|
||||
BrandColor: "#1A1A2E",
|
||||
// ParserID empty on purpose: Moonshot serves two body shapes on
|
||||
// the same host and key, and the proxy's URL sniffer dispatches
|
||||
// both (same pattern as Bifrost). /v1/chat/completions matches
|
||||
// OpenAIParser; the Anthropic-compatible endpoint the official
|
||||
// Claude Code guide uses (/anthropic/v1/messages) contains
|
||||
// "/v1/messages" and matches AnthropicParser. Pinning "openai"
|
||||
// here would misparse the Claude Code path — the primary way
|
||||
// teams consume Kimi for coding today. Both endpoints accept the
|
||||
// same Moonshot key via Authorization: Bearer (Claude Code's
|
||||
// ANTHROPIC_AUTH_TOKEN rides that header too).
|
||||
//
|
||||
// api.moonshot.ai is the international platform; mainland-China
|
||||
// accounts live on api.moonshot.cn with separate billing —
|
||||
// operators there override the host on the provider record. The
|
||||
// kimi.com subscription coding endpoint (api.kimi.com/coding,
|
||||
// model id "k3") is account-bound seat licensing rather than a
|
||||
// meterable platform key, so it's deliberately not the default.
|
||||
ParserID: "",
|
||||
// Pricing per Moonshot's platform rates at K3 launch (July 2026):
|
||||
// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
|
||||
// window. kimi-k3 is the ONLY model the platform serves newer
|
||||
// accounts — K2-era ids (kimi-k2-thinking) and even the kimi-latest
|
||||
// alias return resource_not_found_error, verified live 2026-07-21 —
|
||||
// so it's the only catalog entry. Grandfathered accounts with K2
|
||||
// access can still type those ids on the provider's model rows.
|
||||
// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
|
||||
// doesn't appear here.
|
||||
Models: []Model{
|
||||
{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
|
||||
},
|
||||
},
|
||||
{
|
||||
ID: "litellm_proxy",
|
||||
Kind: KindGateway,
|
||||
|
||||
@@ -85,12 +85,17 @@ type Manager interface {
|
||||
|
||||
// PolicySelectionInput is the per-request selection envelope. The
|
||||
// proxy populates it from CapturedData (account, user, groups) plus
|
||||
// the provider llm_router resolved.
|
||||
// the provider llm_router resolved and the model it extracted.
|
||||
type PolicySelectionInput struct {
|
||||
AccountID string
|
||||
UserID string
|
||||
GroupIDs []string
|
||||
ProviderID string
|
||||
// Model is the already-normalised upstream model id the proxy extracted
|
||||
// (parser strips Bedrock region/version, Vertex @version), so a
|
||||
// case-insensitive compare suffices. Empty = undetermined → not permitted
|
||||
// (fail closed).
|
||||
Model string
|
||||
}
|
||||
|
||||
// PolicySelectionResult names the policy that "pays" for this request
|
||||
|
||||
@@ -5,6 +5,7 @@ import (
|
||||
"fmt"
|
||||
"math"
|
||||
"sort"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
|
||||
@@ -35,6 +36,10 @@ const (
|
||||
denyCodeAccountTokenCapExceeded = "llm_account.token_cap_exceeded"
|
||||
//nolint:gosec // account deny code label, not a credential
|
||||
denyCodeAccountBudgetCapExceeded = "llm_account.budget_cap_exceeded"
|
||||
// denyCodeModelBlocked is returned when policies govern the request's
|
||||
// (provider, caller-groups) but none permits the model. Matches the proxy
|
||||
// guardrail's code so both layers surface the same label.
|
||||
denyCodeModelBlocked = "llm_policy.model_blocked"
|
||||
)
|
||||
|
||||
// consumptionCache holds the consumption counters prefetched for one
|
||||
@@ -159,6 +164,25 @@ func (m *managerImpl) SelectPolicyForRequest(ctx context.Context, in PolicySelec
|
||||
}
|
||||
candidates := filterApplicablePolicies(policies, in)
|
||||
|
||||
// Model-allowlist gate scoped to the matched policies: keep candidates whose
|
||||
// guardrails permit the model (none enabled = unrestricted), deny when
|
||||
// policies apply but none permits it. Skip the load when none has a guardrail.
|
||||
if len(candidates) > 0 && anyPolicyHasGuardrails(candidates) {
|
||||
guardrailsByID, gErr := m.loadGuardrailsByID(ctx, in.AccountID)
|
||||
if gErr != nil {
|
||||
return nil, gErr
|
||||
}
|
||||
permitted := filterModelPermittedPolicies(candidates, guardrailsByID, in.Model)
|
||||
if len(permitted) == 0 {
|
||||
return &PolicySelectionResult{
|
||||
Allow: false,
|
||||
DenyCode: denyCodeModelBlocked,
|
||||
DenyReason: modelBlockedReason(in.Model),
|
||||
}, nil
|
||||
}
|
||||
candidates = permitted
|
||||
}
|
||||
|
||||
// Prefetch every consumption counter the ceiling + candidate policies will
|
||||
// read, in a single store round-trip, then score against the cache.
|
||||
cache, err := m.prefetchConsumption(ctx, in, rules, candidates, now)
|
||||
@@ -250,6 +274,90 @@ func filterApplicablePolicies(policies []*types.Policy, in PolicySelectionInput)
|
||||
return out
|
||||
}
|
||||
|
||||
// anyPolicyHasGuardrails reports whether any policy references at least one
|
||||
// guardrail, so the selector can skip loading guardrails when none do.
|
||||
func anyPolicyHasGuardrails(policies []*types.Policy) bool {
|
||||
for _, p := range policies {
|
||||
if p != nil && len(p.GuardrailIDs) > 0 {
|
||||
return true
|
||||
}
|
||||
}
|
||||
return false
|
||||
}
|
||||
|
||||
// loadGuardrailsByID loads the account's guardrails indexed by ID. Used by the
|
||||
// model-allowlist gate to resolve each candidate policy's attached guardrails.
|
||||
func (m *managerImpl) loadGuardrailsByID(ctx context.Context, accountID string) (map[string]*types.Guardrail, error) {
|
||||
guardrails, err := m.store.GetAccountAgentNetworkGuardrails(ctx, store.LockingStrengthNone, accountID)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("list account guardrails: %w", err)
|
||||
}
|
||||
byID := make(map[string]*types.Guardrail, len(guardrails))
|
||||
for _, g := range guardrails {
|
||||
if g != nil {
|
||||
byID[g.ID] = g
|
||||
}
|
||||
}
|
||||
return byID, nil
|
||||
}
|
||||
|
||||
// filterModelPermittedPolicies returns the subset of policies whose guardrails
|
||||
// permit the model. Order is preserved so downstream scoring is unaffected.
|
||||
func filterModelPermittedPolicies(policies []*types.Policy, byID map[string]*types.Guardrail, model string) []*types.Policy {
|
||||
out := make([]*types.Policy, 0, len(policies))
|
||||
for _, p := range policies {
|
||||
if policyPermitsModel(p, byID, model) {
|
||||
out = append(out, p)
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// policyPermitsModel reports whether a policy permits the model. No
|
||||
// allowlist-enabled guardrail = unrestricted (permits any, incl. empty);
|
||||
// otherwise the model must be in the union of its allowlists, so an
|
||||
// empty/undetermined model fails closed.
|
||||
func policyPermitsModel(p *types.Policy, byID map[string]*types.Guardrail, model string) bool {
|
||||
if p == nil {
|
||||
return false
|
||||
}
|
||||
wanted := normaliseModelID(model)
|
||||
restricted := false
|
||||
for _, gID := range p.GuardrailIDs {
|
||||
g, ok := byID[gID]
|
||||
if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
|
||||
continue
|
||||
}
|
||||
restricted = true
|
||||
if wanted == "" {
|
||||
continue
|
||||
}
|
||||
for _, allowed := range g.Checks.ModelAllowlist.Models {
|
||||
if normaliseModelID(allowed) == wanted {
|
||||
return true
|
||||
}
|
||||
}
|
||||
}
|
||||
return !restricted
|
||||
}
|
||||
|
||||
// normaliseModelID lowercases and trims a model identifier so the allowlist
|
||||
// compare is case-insensitive and trim-tolerant. Mirrors the proxy guardrail's
|
||||
// normaliseModel so both layers agree on what "same model" means.
|
||||
func normaliseModelID(model string) string {
|
||||
return strings.ToLower(strings.TrimSpace(model))
|
||||
}
|
||||
|
||||
// modelBlockedReason builds the human-readable deny reason for a model-allowlist
|
||||
// rejection. The model is quoted when known; an undetermined model is reported
|
||||
// as such so the access log distinguishes "wrong model" from "no model".
|
||||
func modelBlockedReason(model string) string {
|
||||
if normaliseModelID(model) == "" {
|
||||
return "request model could not be determined for the policy allowlist"
|
||||
}
|
||||
return fmt.Sprintf("model %q is not permitted by any applicable policy allowlist", model)
|
||||
}
|
||||
|
||||
// candidate is the per-policy intermediate the selector ranks. A
|
||||
// policy that's been exhausted on any enabled cap never makes it
|
||||
// into this slice; the selector's deny envelope carries the latest
|
||||
|
||||
@@ -0,0 +1,329 @@
|
||||
package agentnetwork
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
"github.com/golang/mock/gomock"
|
||||
"github.com/stretchr/testify/assert"
|
||||
"github.com/stretchr/testify/require"
|
||||
|
||||
"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
|
||||
"github.com/netbirdio/netbird/management/server/store"
|
||||
)
|
||||
|
||||
// guardedPolicy builds an enabled, uncapped policy that authorises sourceGroups
|
||||
// to reach providerID under the given guardrails. Uncapped keeps the selector's
|
||||
// headroom scoring trivial so these tests isolate the model-allowlist gate.
|
||||
func guardedPolicy(id, account string, sourceGroups []string, providerID string, guardrailIDs ...string) *types.Policy {
|
||||
return &types.Policy{
|
||||
ID: id,
|
||||
AccountID: account,
|
||||
Enabled: true,
|
||||
SourceGroups: sourceGroups,
|
||||
DestinationProviderIDs: []string{providerID},
|
||||
GuardrailIDs: guardrailIDs,
|
||||
CreatedAt: time.Now().UTC(),
|
||||
}
|
||||
}
|
||||
|
||||
// allowlistGuardrail builds a guardrail whose model allowlist is enabled and
|
||||
// carries the given models.
|
||||
func allowlistGuardrail(id, account string, models ...string) *types.Guardrail {
|
||||
return &types.Guardrail{
|
||||
ID: id,
|
||||
AccountID: account,
|
||||
Checks: types.GuardrailChecks{
|
||||
ModelAllowlist: types.GuardrailModelAllowlist{Enabled: true, Models: models},
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
func expectPolicies(mockStore *store.MockStore, account string, policies ...*types.Policy) {
|
||||
mockStore.EXPECT().
|
||||
GetAccountAgentNetworkPolicies(gomock.Any(), gomock.Any(), account).
|
||||
Return(policies, nil)
|
||||
}
|
||||
|
||||
func expectGuardrails(mockStore *store.MockStore, account string, guardrails ...*types.Guardrail) {
|
||||
mockStore.EXPECT().
|
||||
GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), account).
|
||||
Return(guardrails, nil)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_ModelBlockedByAllowlist proves the authoritative allowlist
|
||||
// decision: a policy authorises the (provider, group) but restricts the model,
|
||||
// and the requested model isn't on the list, so the request is denied.
|
||||
func TestSelectPolicy_ModelBlockedByAllowlist(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
UserID: "user-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.False(t, res.Allow, "a model outside the only applicable policy's allowlist must be denied")
|
||||
assert.Equal(t, denyCodeModelBlocked, res.DenyCode, "deny code must be model_blocked")
|
||||
assert.NotEmpty(t, res.DenyReason, "deny reason must be populated")
|
||||
}
|
||||
|
||||
// TestSelectPolicy_ModelAllowedByAllowlist is the allow counterpart: the model
|
||||
// is on the applicable policy's allowlist, so selection proceeds normally.
|
||||
func TestSelectPolicy_ModelAllowedByAllowlist(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o", "claude-opus-4"))
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
UserID: "user-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "a model on the applicable policy's allowlist must be allowed")
|
||||
assert.Equal(t, "pol-A", res.SelectedPolicyID)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_CaseInsensitiveModelMatch proves the compare tolerates case
|
||||
// and surrounding whitespace, matching the proxy guardrail's normalisation.
|
||||
func TestSelectPolicy_CaseInsensitiveModelMatch(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", " GPT-4o "))
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "gpt-4o",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "case/whitespace variants must match the allowlist entry")
|
||||
}
|
||||
|
||||
// TestSelectPolicy_UnguardedPolicyIsUnrestricted is the false-deny fix: when two
|
||||
// policies authorise the same (provider, group) and one has no guardrail, that
|
||||
// policy makes the request unrestricted — not caught by the other's allowlist.
|
||||
func TestSelectPolicy_UnguardedPolicyIsUnrestricted(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
restricted := guardedPolicy("pol-restricted", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
open := guardedPolicy("pol-open", "acc-1", []string{"grp-eng"}, "prov-1") // no guardrail
|
||||
expectPolicies(mockStore, "acc-1", restricted, open)
|
||||
expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "an un-guardrailed policy for the same (provider, group) must leave the request unrestricted")
|
||||
assert.Equal(t, "pol-open", res.SelectedPolicyID, "the unrestricted policy must be the one that pays")
|
||||
}
|
||||
|
||||
// TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups is the false-allow fix: a
|
||||
// model allowlisted only for grp-b must not be usable by a grp-a caller. The
|
||||
// selector considers only policies applicable to the caller's groups.
|
||||
func TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
polA := guardedPolicy("pol-a", "acc-1", []string{"grp-a"}, "prov-1", "g-a")
|
||||
polB := guardedPolicy("pol-b", "acc-1", []string{"grp-b"}, "prov-1", "g-b")
|
||||
expectPolicies(mockStore, "acc-1", polA, polB)
|
||||
expectGuardrails(mockStore, "acc-1",
|
||||
allowlistGuardrail("g-a", "acc-1", "gpt-4o"),
|
||||
allowlistGuardrail("g-b", "acc-1", "claude-opus-4"),
|
||||
)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-a"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4", // only allowed for grp-b
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.False(t, res.Allow, "grp-b's allowlisted model must not leak to a grp-a caller")
|
||||
assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_UndeterminedModelFailsClosed proves the fail-closed contract
|
||||
// mirrors the proxy: with a restricted applicable policy and an empty model
|
||||
// (e.g. a path-routed shape the parser couldn't map), the request is denied.
|
||||
func TestSelectPolicy_UndeterminedModelFailsClosed(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "", // undetermined
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.False(t, res.Allow, "an undetermined model must fail closed against a restricted policy")
|
||||
assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_DisabledAllowlistDoesNotRestrict proves a guardrail whose
|
||||
// model allowlist is disabled imposes no model restriction, even though the
|
||||
// policy references it.
|
||||
func TestSelectPolicy_DisabledAllowlistDoesNotRestrict(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
disabled := &types.Guardrail{
|
||||
ID: "g-1",
|
||||
AccountID: "acc-1",
|
||||
Checks: types.GuardrailChecks{
|
||||
ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}},
|
||||
},
|
||||
}
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1", disabled)
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "anything-goes",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "a disabled allowlist must not restrict the model")
|
||||
assert.Equal(t, "pol-A", res.SelectedPolicyID)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_UnionAcrossPolicyGuardrails proves a policy with multiple
|
||||
// allowlist guardrails permits the union of their models (not just the first).
|
||||
func TestSelectPolicy_UnionAcrossPolicyGuardrails(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1", "g-2")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1",
|
||||
allowlistGuardrail("g-1", "acc-1", "gpt-4o"),
|
||||
allowlistGuardrail("g-2", "acc-1", "claude-opus-4"),
|
||||
)
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4", // only in the second guardrail's list
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "a model in any of the policy's allowlist guardrails must be permitted")
|
||||
assert.Equal(t, "pol-A", res.SelectedPolicyID)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_GuardrailLookupErrorPropagates proves a store failure while
|
||||
// resolving the candidate policies' guardrails surfaces as an error, not a
|
||||
// silent allow/deny.
|
||||
func TestSelectPolicy_GuardrailLookupErrorPropagates(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
mockStore.EXPECT().
|
||||
GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), "acc-1").
|
||||
Return(nil, errors.New("store unavailable"))
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "gpt-4o",
|
||||
})
|
||||
require.Error(t, err, "a guardrail-lookup failure must surface as an error")
|
||||
assert.Nil(t, res)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted proves a
|
||||
// policy referencing a guardrail ID absent from the account's set (a stale
|
||||
// reference) imposes no model restriction — same as no guardrail.
|
||||
func TestSelectPolicy_MissingGuardrailReferenceTreatedAsUnrestricted(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-missing")
|
||||
expectPolicies(mockStore, "acc-1", policy)
|
||||
expectGuardrails(mockStore, "acc-1")
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "anything-goes",
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow, "an orphaned guardrail reference must not restrict the model")
|
||||
assert.Equal(t, "pol-A", res.SelectedPolicyID)
|
||||
}
|
||||
|
||||
// TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter proves the model
|
||||
// gate narrows candidates before cap scoring: the permitting policy is selected
|
||||
// even though the blocked one has a larger, more attractive cap.
|
||||
func TestSelectPolicy_PartialCandidatesPermittedAfterModelFilter(t *testing.T) {
|
||||
ctrl := gomock.NewController(t)
|
||||
mgr, mockStore := newSelectorMgr(t, ctrl)
|
||||
|
||||
polBig := guardedPolicy("pol-big", "acc-1", []string{"grp-eng"}, "prov-1", "g-restrict")
|
||||
polBig.Limits = types.PolicyLimits{
|
||||
TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 1_000_000, WindowSeconds: 3600},
|
||||
}
|
||||
polSmall := guardedPolicy("pol-small", "acc-1", []string{"grp-eng"}, "prov-1", "g-permit")
|
||||
polSmall.Limits = types.PolicyLimits{
|
||||
TokenLimit: types.PolicyTokenLimit{Enabled: true, GroupCap: 100, WindowSeconds: 3600},
|
||||
}
|
||||
expectPolicies(mockStore, "acc-1", polBig, polSmall)
|
||||
expectGuardrails(mockStore, "acc-1",
|
||||
allowlistGuardrail("g-restrict", "acc-1", "gpt-4o"),
|
||||
allowlistGuardrail("g-permit", "acc-1", "claude-opus-4"),
|
||||
)
|
||||
expectConsumptionBatch(mockStore, nil)
|
||||
|
||||
res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
|
||||
AccountID: "acc-1",
|
||||
GroupIDs: []string{"grp-eng"},
|
||||
ProviderID: "prov-1",
|
||||
Model: "claude-opus-4", // only pol-small's guardrail permits this
|
||||
})
|
||||
require.NoError(t, err)
|
||||
assert.True(t, res.Allow)
|
||||
assert.Equal(t, "pol-small", res.SelectedPolicyID,
|
||||
"the model filter must exclude pol-big before cap scoring")
|
||||
}
|
||||
@@ -235,7 +235,12 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
|
||||
|
||||
mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
|
||||
applyAccountCollectionControls(&mergedGuardrails, settings)
|
||||
guardrailJSON, err := marshalGuardrailConfig(mergedGuardrails)
|
||||
// The proxy guardrail is a per-provider fail-closed backstop; the
|
||||
// authoritative per-policy/group decision is management's
|
||||
// SelectPolicyForRequest. A provider lands in this map only when every
|
||||
// authorising policy restricts models.
|
||||
providerAllowlists := buildProviderAllowlists(enabledPolicies, guardrailsByID)
|
||||
guardrailJSON, err := marshalGuardrailConfig(providerAllowlists, mergedGuardrails.PromptCapture)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
@@ -780,10 +785,12 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byt
|
||||
|
||||
// guardrailConfig is the JSON shape the proxy-side llm_guardrail
|
||||
// middleware expects. Mirrors the proxy registration documented in
|
||||
// the management→proxy contract.
|
||||
// the management→proxy contract. provider_allowlists is keyed by the
|
||||
// resolved provider id llm_router stamps; a provider absent from the map is
|
||||
// unrestricted at the proxy layer.
|
||||
type guardrailConfig struct {
|
||||
ModelAllowlist []string `json:"model_allowlist,omitempty"`
|
||||
PromptCapture guardrailPromptCapture `json:"prompt_capture"`
|
||||
ProviderAllowlists map[string][]string `json:"provider_allowlists,omitempty"`
|
||||
PromptCapture guardrailPromptCapture `json:"prompt_capture"`
|
||||
}
|
||||
|
||||
type guardrailPromptCapture struct {
|
||||
@@ -828,13 +835,10 @@ func applyAccountCollectionControls(merged *MergedGuardrails, settings *types.Se
|
||||
merged.PromptCapture.RedactPii = settings.RedactPii || merged.PromptCapture.RedactPii
|
||||
}
|
||||
|
||||
func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
|
||||
func marshalGuardrailConfig(providerAllowlists map[string][]string, capture MergedPromptCapture) ([]byte, error) {
|
||||
cfg := guardrailConfig{
|
||||
ModelAllowlist: merged.ModelAllowlist,
|
||||
PromptCapture: guardrailPromptCapture{
|
||||
Enabled: merged.PromptCapture.Enabled,
|
||||
RedactPii: merged.PromptCapture.RedactPii,
|
||||
},
|
||||
ProviderAllowlists: providerAllowlists,
|
||||
PromptCapture: guardrailPromptCapture(capture),
|
||||
}
|
||||
out, err := json.Marshal(cfg)
|
||||
if err != nil {
|
||||
@@ -843,6 +847,74 @@ func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
|
||||
return out, nil
|
||||
}
|
||||
|
||||
// buildProviderAllowlists returns the proxy's per-provider backstop: a provider
|
||||
// is included only when every authorising policy restricts models (their union);
|
||||
// if any leaves it unrestricted it is omitted, so management decides per group.
|
||||
func buildProviderAllowlists(policies []*types.Policy, byID map[string]*types.Guardrail) map[string][]string {
|
||||
type providerAcc struct {
|
||||
models map[string]struct{}
|
||||
anyUnrestricted bool
|
||||
}
|
||||
accs := make(map[string]*providerAcc)
|
||||
for _, p := range policies {
|
||||
if p == nil {
|
||||
continue
|
||||
}
|
||||
restricted, models := policyModelAllowlist(p, byID)
|
||||
for _, providerID := range p.DestinationProviderIDs {
|
||||
if providerID == "" {
|
||||
continue
|
||||
}
|
||||
acc, ok := accs[providerID]
|
||||
if !ok {
|
||||
acc = &providerAcc{models: make(map[string]struct{})}
|
||||
accs[providerID] = acc
|
||||
}
|
||||
if !restricted {
|
||||
acc.anyUnrestricted = true
|
||||
continue
|
||||
}
|
||||
for _, m := range models {
|
||||
acc.models[m] = struct{}{}
|
||||
}
|
||||
}
|
||||
}
|
||||
out := make(map[string][]string, len(accs))
|
||||
for providerID, acc := range accs {
|
||||
if acc.anyUnrestricted {
|
||||
continue
|
||||
}
|
||||
models := make([]string, 0, len(acc.models))
|
||||
for m := range acc.models {
|
||||
models = append(models, m)
|
||||
}
|
||||
sort.Strings(models)
|
||||
out[providerID] = models
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// policyModelAllowlist reports whether a policy restricts models (has an
|
||||
// allowlist-enabled guardrail) and the union of allowed models. Models are
|
||||
// verbatim; the proxy factory lowercases/trims them at decode time.
|
||||
func policyModelAllowlist(p *types.Policy, byID map[string]*types.Guardrail) (bool, []string) {
|
||||
restricted := false
|
||||
var models []string
|
||||
for _, gID := range p.GuardrailIDs {
|
||||
g, ok := byID[gID]
|
||||
if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
|
||||
continue
|
||||
}
|
||||
restricted = true
|
||||
for _, m := range g.Checks.ModelAllowlist.Models {
|
||||
if m != "" {
|
||||
models = append(models, m)
|
||||
}
|
||||
}
|
||||
}
|
||||
return restricted, models
|
||||
}
|
||||
|
||||
// buildAccountService composes the per-account gateway Service. The
|
||||
// target carries the noop placeholder URL — the router middleware
|
||||
// rewrites every request to the matched provider's upstream before the
|
||||
@@ -986,38 +1058,11 @@ func unionSourceGroups(policies []*types.Policy) []string {
|
||||
return out
|
||||
}
|
||||
|
||||
// MergedGuardrails is the JSON shape passed to the proxy via the
|
||||
// guardrail middleware's config_json. Mirrors the proxy-side
|
||||
// expectations and is intentionally distinct from
|
||||
// types.GuardrailChecks so we can evolve either side independently.
|
||||
// MergedGuardrails is the synthesiser's fold target. Only prompt capture is
|
||||
// merged here — the model allowlist is emitted per-provider, and
|
||||
// token/budget/retention moved onto Policy.Limits and account Settings.
|
||||
type MergedGuardrails struct {
|
||||
ModelAllowlist []string `json:"model_allowlist,omitempty"`
|
||||
TokenLimits MergedTokenLimits `json:"token_limits"`
|
||||
Budget MergedBudget `json:"budget"`
|
||||
PromptCapture MergedPromptCapture `json:"prompt_capture"`
|
||||
Retention MergedRetention `json:"retention"`
|
||||
}
|
||||
|
||||
type MergedTokenLimits struct {
|
||||
Hourly *MergedTokenWindow `json:"hourly,omitempty"`
|
||||
Daily *MergedTokenWindow `json:"daily,omitempty"`
|
||||
Monthly *MergedTokenWindow `json:"monthly,omitempty"`
|
||||
}
|
||||
|
||||
type MergedTokenWindow struct {
|
||||
MaxInputTokens int `json:"max_input_tokens,omitempty"`
|
||||
MaxOutputTokens int `json:"max_output_tokens,omitempty"`
|
||||
}
|
||||
|
||||
type MergedBudget struct {
|
||||
Hourly *MergedBudgetWindow `json:"hourly,omitempty"`
|
||||
Daily *MergedBudgetWindow `json:"daily,omitempty"`
|
||||
Monthly *MergedBudgetWindow `json:"monthly,omitempty"`
|
||||
}
|
||||
|
||||
type MergedBudgetWindow struct {
|
||||
SoftCapUSD float64 `json:"soft_cap_usd,omitempty"`
|
||||
HardCapUSD float64 `json:"hard_cap_usd,omitempty"`
|
||||
PromptCapture MergedPromptCapture
|
||||
}
|
||||
|
||||
type MergedPromptCapture struct {
|
||||
@@ -1025,64 +1070,31 @@ type MergedPromptCapture struct {
|
||||
RedactPii bool `json:"redact_pii"`
|
||||
}
|
||||
|
||||
type MergedRetention struct {
|
||||
Enabled bool `json:"enabled"`
|
||||
Days int `json:"days"`
|
||||
}
|
||||
|
||||
// mergeGuardrails computes the effective guardrail spec applied at the
|
||||
// proxy, given the referencing policies and the account's guardrail
|
||||
// catalogue. Policy enabled-ness is the caller's responsibility — only
|
||||
// enabled policies should be passed in.
|
||||
// mergeGuardrails folds the referencing policies' guardrails into the
|
||||
// prompt-capture decision only. The model allowlist is enforced per-policy/group
|
||||
// in management and shipped per-provider; token/budget/retention live off
|
||||
// guardrails now.
|
||||
//
|
||||
// Merge rules:
|
||||
// - Model allowlist: union of allowlists across policies that enable it.
|
||||
// - Token / Budget: most-restrictive (min of non-zero caps) per window.
|
||||
// - Prompt capture: enabled if any policy enables it; redact_pii sticks
|
||||
// if any enabling policy turns it on.
|
||||
// - Retention: enabled if any enables it; smallest non-zero days wins.
|
||||
// Merge rule — prompt capture: enabled if any policy enables it; redact_pii
|
||||
// sticks if any enabling policy turns it on.
|
||||
func mergeGuardrails(policies []*types.Policy, byID map[string]*types.Guardrail) MergedGuardrails {
|
||||
merged := MergedGuardrails{}
|
||||
allowlist := make(map[string]struct{})
|
||||
allowlistEnabled := false
|
||||
|
||||
for _, policy := range policies {
|
||||
for _, gID := range policy.GuardrailIDs {
|
||||
g, ok := byID[gID]
|
||||
if !ok || g == nil {
|
||||
continue
|
||||
}
|
||||
mergeGuardrail(g, &merged, allowlist, &allowlistEnabled)
|
||||
mergeGuardrail(g, &merged)
|
||||
}
|
||||
}
|
||||
|
||||
if allowlistEnabled {
|
||||
merged.ModelAllowlist = make([]string, 0, len(allowlist))
|
||||
for m := range allowlist {
|
||||
merged.ModelAllowlist = append(merged.ModelAllowlist, m)
|
||||
}
|
||||
sort.Strings(merged.ModelAllowlist)
|
||||
}
|
||||
return merged
|
||||
}
|
||||
|
||||
// mergeGuardrail folds a single guardrail's enabled checks into the
|
||||
// running merge: model-allowlist models join the shared set (and flip
|
||||
// allowlistEnabled), and prompt-capture / redact-pii stick once any
|
||||
// enabling guardrail turns them on.
|
||||
//
|
||||
// TokenLimits, Budget, and Retention have moved off guardrails — token
|
||||
// and budget caps now live on the Policy itself (Policy.Limits) and
|
||||
// retention moves to account-level Settings — so they are not merged here.
|
||||
func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails, allowlist map[string]struct{}, allowlistEnabled *bool) {
|
||||
if g.Checks.ModelAllowlist.Enabled {
|
||||
*allowlistEnabled = true
|
||||
for _, m := range g.Checks.ModelAllowlist.Models {
|
||||
if m != "" {
|
||||
allowlist[m] = struct{}{}
|
||||
}
|
||||
}
|
||||
}
|
||||
// mergeGuardrail folds a single guardrail's prompt-capture settings into the
|
||||
// running merge: enabled / redact-pii stick once any enabling guardrail turns
|
||||
// them on.
|
||||
func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails) {
|
||||
if g.Checks.PromptCapture.Enabled {
|
||||
merged.PromptCapture.Enabled = true
|
||||
if g.Checks.PromptCapture.RedactPii {
|
||||
|
||||
@@ -81,8 +81,8 @@ func TestSynthesizeServices_RealStore_PromptCaptureAccountIsSoleControl(t *testi
|
||||
require.Len(t, services, 1)
|
||||
|
||||
cfg := decodeServiceGuardrailConfig(t, services[0])
|
||||
assert.Equal(t, []string{"gpt-5.4"}, cfg.ModelAllowlist,
|
||||
"model allowlist is a pure policy guardrail and must always reach the config")
|
||||
assert.Equal(t, map[string][]string{"prov-1": {"gpt-5.4"}}, cfg.ProviderAllowlists,
|
||||
"model allowlist is a pure policy guardrail and must reach the per-provider config")
|
||||
assert.False(t, cfg.PromptCapture.Enabled,
|
||||
"prompt capture must be off when the account toggle is off, even with a capture-enabled guardrail")
|
||||
}
|
||||
@@ -172,7 +172,7 @@ func TestSynthesizeServices_RealStore_NoGuardrail_CaptureOff(t *testing.T) {
|
||||
require.Len(t, services, 1, "exactly one synth service expected")
|
||||
|
||||
cfg := decodeServiceGuardrailConfig(t, services[0])
|
||||
assert.Empty(t, cfg.ModelAllowlist, "no guardrail → no allowlist")
|
||||
assert.Empty(t, cfg.ProviderAllowlists, "no guardrail → provider unrestricted (absent from map)")
|
||||
assert.False(t, cfg.PromptCapture.Enabled, "no guardrail → prompt capture off by default")
|
||||
assert.False(t, cfg.PromptCapture.RedactPii, "no guardrail → redact off by default")
|
||||
}
|
||||
|
||||
@@ -0,0 +1,95 @@
|
||||
package agentnetwork
|
||||
|
||||
import (
|
||||
"testing"
|
||||
|
||||
"github.com/stretchr/testify/assert"
|
||||
|
||||
"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
|
||||
)
|
||||
|
||||
// policyForProviders builds an enabled policy authorising the given providers
|
||||
// under the given guardrails (both optional). Groups are irrelevant to
|
||||
// buildProviderAllowlists, which keys purely on destination provider.
|
||||
func policyForProviders(id string, guardrailIDs []string, providerIDs ...string) *types.Policy {
|
||||
return &types.Policy{
|
||||
ID: id,
|
||||
Enabled: true,
|
||||
DestinationProviderIDs: providerIDs,
|
||||
GuardrailIDs: guardrailIDs,
|
||||
}
|
||||
}
|
||||
|
||||
func TestBuildProviderAllowlists(t *testing.T) {
|
||||
byID := map[string]*types.Guardrail{
|
||||
"g-4o": allowlistGuardrail("g-4o", "acc-1", "gpt-4o"),
|
||||
"g-opus": allowlistGuardrail("g-opus", "acc-1", "claude-opus-4"),
|
||||
"g-disabled": {ID: "g-disabled", Checks: types.GuardrailChecks{ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}}}},
|
||||
}
|
||||
|
||||
t.Run("all authorising policies restrict yields per-provider union", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-4o"}, "prov-x"),
|
||||
policyForProviders("p2", []string{"g-opus"}, "prov-x"),
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.Equal(t, map[string][]string{"prov-x": {"claude-opus-4", "gpt-4o"}}, got,
|
||||
"a provider every policy restricts carries the sorted union of their models")
|
||||
})
|
||||
|
||||
t.Run("any un-guardrailed policy leaves the provider unrestricted (omitted)", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-4o"}, "prov-x"),
|
||||
policyForProviders("p2", nil, "prov-x"), // no guardrail
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.NotContains(t, got, "prov-x",
|
||||
"a provider reachable by an un-guardrailed policy must be omitted so the proxy treats it as unrestricted")
|
||||
})
|
||||
|
||||
t.Run("a disabled allowlist counts as unrestricted", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-disabled"}, "prov-x"),
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.NotContains(t, got, "prov-x",
|
||||
"a policy whose only guardrail has a disabled allowlist is unrestricted")
|
||||
})
|
||||
|
||||
t.Run("providers are isolated from one another", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-4o"}, "prov-x"),
|
||||
policyForProviders("p2", []string{"g-opus"}, "prov-y"),
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.Equal(t, []string{"gpt-4o"}, got["prov-x"], "prov-x keeps only its own model")
|
||||
assert.Equal(t, []string{"claude-opus-4"}, got["prov-y"], "prov-y keeps only its own model")
|
||||
})
|
||||
|
||||
t.Run("one policy authorising two providers restricts both", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-4o"}, "prov-x", "prov-y"),
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.Equal(t, []string{"gpt-4o"}, got["prov-x"])
|
||||
assert.Equal(t, []string{"gpt-4o"}, got["prov-y"])
|
||||
})
|
||||
|
||||
t.Run("union across a single policy's guardrails", func(t *testing.T) {
|
||||
policies := []*types.Policy{
|
||||
policyForProviders("p1", []string{"g-4o", "g-opus"}, "prov-x"),
|
||||
}
|
||||
got := buildProviderAllowlists(policies, byID)
|
||||
assert.ElementsMatch(t, []string{"claude-opus-4", "gpt-4o"}, got["prov-x"],
|
||||
"a policy's own multiple allowlist guardrails union together")
|
||||
})
|
||||
|
||||
t.Run("an enabled allowlist with no models denies everything", func(t *testing.T) {
|
||||
empty := map[string]*types.Guardrail{"g-empty": allowlistGuardrail("g-empty", "acc-1")}
|
||||
got := buildProviderAllowlists([]*types.Policy{
|
||||
policyForProviders("p1", []string{"g-empty"}, "prov-x"),
|
||||
}, empty)
|
||||
assert.Equal(t, map[string][]string{"prov-x": {}}, got,
|
||||
"an enabled-but-empty allowlist is restricted with an empty set, not unrestricted")
|
||||
})
|
||||
}
|
||||
@@ -1031,8 +1031,12 @@ func TestSynthesizeServices_GuardrailMerge_AllowlistUnion_LimitsRestrictive(t *t
|
||||
|
||||
var cfg guardrailConfig
|
||||
require.NoError(t, json.Unmarshal(guardrailJSON, &cfg), "guardrail config must unmarshal cleanly")
|
||||
assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ModelAllowlist,
|
||||
"model allowlist union must keep both models")
|
||||
// Both policies restrict the same provider, so the per-provider backstop
|
||||
// carries the union of their models — a coarse gate that management's
|
||||
// per-policy/group check narrows; it only blocks models outside the union
|
||||
// when management is down.
|
||||
assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ProviderAllowlists["prov-1"],
|
||||
"per-provider allowlist union must keep both models")
|
||||
}
|
||||
|
||||
func TestSynthesizeServices_BackfillsMissingSessionKeys(t *testing.T) {
|
||||
|
||||
@@ -41,8 +41,24 @@ type AgentNetworkAccessLog struct {
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CostUSD float64
|
||||
Stream bool
|
||||
// Prompt-cache buckets: read + write token counts.
|
||||
CachedInputTokens int64
|
||||
CacheCreationTokens int64
|
||||
// Per-bucket cost breakdown — one column per token bucket the provider
|
||||
// bills separately. These four are the only cost state stored: the total
|
||||
// and the cache portion are derived on read (TotalCostUSD / CacheCostUSD)
|
||||
// rather than stored alongside, so a stored aggregate can never drift out
|
||||
// of step with the components it summarises.
|
||||
//
|
||||
// default:0 matters on upgrade: these columns are ALTER TABLE ADD COLUMN
|
||||
// on an existing table, and without it every historical row holds NULL —
|
||||
// which a raw SUM()/scan into float64 can't read. The default backfills
|
||||
// them as 0, so pre-upgrade rows report an unknown split, not an error.
|
||||
InputCostUSD float64 `gorm:"not null;default:0"`
|
||||
CachedInputCostUSD float64 `gorm:"not null;default:0"`
|
||||
CacheCreationCostUSD float64 `gorm:"not null;default:0"`
|
||||
OutputCostUSD float64 `gorm:"not null;default:0"`
|
||||
Stream bool
|
||||
|
||||
// Prompt capture. Only populated when prompt collection is enabled
|
||||
// (account master switch AND policy guardrail). Heavy free text.
|
||||
@@ -60,19 +76,44 @@ type AgentNetworkAccessLog struct {
|
||||
// the reverse-proxy AccessLogEntry table.
|
||||
func (AgentNetworkAccessLog) TableName() string { return "agent_network_access_log" }
|
||||
|
||||
// CostUSDSQLExpr is the SQL sum of the per-bucket cost columns — the total cost
|
||||
// of a row. Used wherever a query has to sort or aggregate on total cost now
|
||||
// that no cost_usd column is stored. Plain arithmetic over NOT NULL columns, so
|
||||
// it stays portable across SQLite and Postgres.
|
||||
const CostUSDSQLExpr = "(input_cost_usd + cached_input_cost_usd + cache_creation_cost_usd + output_cost_usd)"
|
||||
|
||||
// TotalCostUSD is the request's total cost: the sum of the four per-bucket
|
||||
// costs. Derived rather than stored so it cannot disagree with the breakdown.
|
||||
func (a *AgentNetworkAccessLog) TotalCostUSD() float64 {
|
||||
return a.InputCostUSD + a.CachedInputCostUSD + a.CacheCreationCostUSD + a.OutputCostUSD
|
||||
}
|
||||
|
||||
// CacheCostUSD is the portion of the total billed for prompt-cache buckets:
|
||||
// cache reads plus cache writes.
|
||||
func (a *AgentNetworkAccessLog) CacheCostUSD() float64 {
|
||||
return a.CachedInputCostUSD + a.CacheCreationCostUSD
|
||||
}
|
||||
|
||||
// ToAPIResponse renders the flattened entry as the API representation.
|
||||
func (a *AgentNetworkAccessLog) ToAPIResponse() api.AgentNetworkAccessLog {
|
||||
out := api.AgentNetworkAccessLog{
|
||||
Id: a.ID,
|
||||
ServiceId: a.ServiceID,
|
||||
Timestamp: a.Timestamp,
|
||||
StatusCode: a.StatusCode,
|
||||
DurationMs: int(a.Duration.Milliseconds()),
|
||||
InputTokens: a.InputTokens,
|
||||
OutputTokens: a.OutputTokens,
|
||||
TotalTokens: a.TotalTokens,
|
||||
CostUsd: a.CostUSD,
|
||||
Stream: &a.Stream,
|
||||
Id: a.ID,
|
||||
ServiceId: a.ServiceID,
|
||||
Timestamp: a.Timestamp,
|
||||
StatusCode: a.StatusCode,
|
||||
DurationMs: int(a.Duration.Milliseconds()),
|
||||
InputTokens: a.InputTokens,
|
||||
OutputTokens: a.OutputTokens,
|
||||
TotalTokens: a.TotalTokens,
|
||||
CachedInputTokens: a.CachedInputTokens,
|
||||
CacheCreationTokens: a.CacheCreationTokens,
|
||||
InputCostUsd: a.InputCostUSD,
|
||||
CachedInputCostUsd: a.CachedInputCostUSD,
|
||||
CacheCreationCostUsd: a.CacheCreationCostUSD,
|
||||
OutputCostUsd: a.OutputCostUSD,
|
||||
CostUsd: a.TotalCostUSD(),
|
||||
CacheCostUsd: a.CacheCostUSD(),
|
||||
Stream: &a.Stream,
|
||||
}
|
||||
|
||||
out.UserId = strPtr(a.UserID)
|
||||
@@ -112,20 +153,36 @@ func strPtr(s string) *string {
|
||||
// summary plus its ordered entries. Assembled in Go from a page of entries — it
|
||||
// is not a stored table.
|
||||
type AgentNetworkAccessLogSession struct {
|
||||
SessionID string // empty for a session-less (singleton) request
|
||||
UserID string
|
||||
GroupIDs []string // union of the entries' authorising groups
|
||||
StartedAt time.Time
|
||||
EndedAt time.Time
|
||||
RequestCount int
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CostUSD float64
|
||||
Providers []string // distinct vendors seen in the session
|
||||
Models []string // distinct models seen in the session
|
||||
Decision string // "deny" if any entry was denied, else "allow"
|
||||
Entries []*AgentNetworkAccessLog
|
||||
SessionID string // empty for a session-less (singleton) request
|
||||
UserID string
|
||||
GroupIDs []string // union of the entries' authorising groups
|
||||
StartedAt time.Time
|
||||
EndedAt time.Time
|
||||
RequestCount int
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CachedInputTokens int64
|
||||
CacheCreationTokens int64
|
||||
InputCostUSD float64
|
||||
CachedInputCostUSD float64
|
||||
CacheCreationCostUSD float64
|
||||
OutputCostUSD float64
|
||||
Providers []string // distinct vendors seen in the session
|
||||
Models []string // distinct models seen in the session
|
||||
Decision string // "deny" if any entry was denied, else "allow"
|
||||
Entries []*AgentNetworkAccessLog
|
||||
}
|
||||
|
||||
// TotalCostUSD is the session's total cost: the sum of the four per-bucket
|
||||
// costs accumulated across its entries.
|
||||
func (sess *AgentNetworkAccessLogSession) TotalCostUSD() float64 {
|
||||
return sess.InputCostUSD + sess.CachedInputCostUSD + sess.CacheCreationCostUSD + sess.OutputCostUSD
|
||||
}
|
||||
|
||||
// CacheCostUSD is the session's prompt-cache spend: cache reads plus writes.
|
||||
func (sess *AgentNetworkAccessLogSession) CacheCostUSD() float64 {
|
||||
return sess.CachedInputCostUSD + sess.CacheCreationCostUSD
|
||||
}
|
||||
|
||||
// sessionKey is the grouping key for an entry: its session id, or — when the
|
||||
@@ -205,7 +262,12 @@ func (sess *AgentNetworkAccessLogSession) foldEntry(sk *sessionSeen, e *AgentNet
|
||||
sess.InputTokens += e.InputTokens
|
||||
sess.OutputTokens += e.OutputTokens
|
||||
sess.TotalTokens += e.TotalTokens
|
||||
sess.CostUSD += e.CostUSD
|
||||
sess.CachedInputTokens += e.CachedInputTokens
|
||||
sess.CacheCreationTokens += e.CacheCreationTokens
|
||||
sess.InputCostUSD += e.InputCostUSD
|
||||
sess.CachedInputCostUSD += e.CachedInputCostUSD
|
||||
sess.CacheCreationCostUSD += e.CacheCreationCostUSD
|
||||
sess.OutputCostUSD += e.OutputCostUSD
|
||||
if e.Timestamp.Before(sess.StartedAt) {
|
||||
sess.StartedAt = e.Timestamp
|
||||
}
|
||||
@@ -248,15 +310,22 @@ func (sess *AgentNetworkAccessLogSession) ToAPIResponse() api.AgentNetworkAccess
|
||||
}
|
||||
|
||||
out := api.AgentNetworkAccessLogSession{
|
||||
StartedAt: sess.StartedAt,
|
||||
EndedAt: sess.EndedAt,
|
||||
RequestCount: sess.RequestCount,
|
||||
InputTokens: sess.InputTokens,
|
||||
OutputTokens: sess.OutputTokens,
|
||||
TotalTokens: sess.TotalTokens,
|
||||
CostUsd: sess.CostUSD,
|
||||
Decision: sess.Decision,
|
||||
Entries: entries,
|
||||
StartedAt: sess.StartedAt,
|
||||
EndedAt: sess.EndedAt,
|
||||
RequestCount: sess.RequestCount,
|
||||
InputTokens: sess.InputTokens,
|
||||
OutputTokens: sess.OutputTokens,
|
||||
TotalTokens: sess.TotalTokens,
|
||||
CachedInputTokens: sess.CachedInputTokens,
|
||||
CacheCreationTokens: sess.CacheCreationTokens,
|
||||
InputCostUsd: sess.InputCostUSD,
|
||||
CachedInputCostUsd: sess.CachedInputCostUSD,
|
||||
CacheCreationCostUsd: sess.CacheCreationCostUSD,
|
||||
OutputCostUsd: sess.OutputCostUSD,
|
||||
CostUsd: sess.TotalCostUSD(),
|
||||
CacheCostUsd: sess.CacheCostUSD(),
|
||||
Decision: sess.Decision,
|
||||
Entries: entries,
|
||||
}
|
||||
out.SessionId = strPtr(sess.SessionID)
|
||||
out.UserId = strPtr(sess.UserID)
|
||||
|
||||
@@ -54,7 +54,7 @@ var accessLogSortFields = map[string]string{
|
||||
"provider": "provider",
|
||||
"status_code": "status_code",
|
||||
"duration": "duration",
|
||||
"cost_usd": "cost_usd",
|
||||
"cost_usd": CostUSDSQLExpr,
|
||||
"total_tokens": "total_tokens",
|
||||
"user_id": "user_id",
|
||||
"decision": "decision",
|
||||
@@ -70,7 +70,7 @@ var accessLogSortFields = map[string]string{
|
||||
var sessionSortExprs = map[string]string{ //nolint:gosec // G101 false positive: "total_tokens" sort key, not a credential
|
||||
"timestamp": "MAX(timestamp)",
|
||||
"started_at": "MIN(timestamp)",
|
||||
"cost_usd": "SUM(cost_usd)",
|
||||
"cost_usd": "SUM" + CostUSDSQLExpr,
|
||||
"total_tokens": "SUM(total_tokens)",
|
||||
"duration": "SUM(duration)",
|
||||
"request_count": "COUNT(*)",
|
||||
|
||||
124
management/internals/modules/agentnetwork/types/cost_test.go
Normal file
124
management/internals/modules/agentnetwork/types/cost_test.go
Normal file
@@ -0,0 +1,124 @@
|
||||
package types
|
||||
|
||||
import (
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
"github.com/stretchr/testify/assert"
|
||||
"github.com/stretchr/testify/require"
|
||||
)
|
||||
|
||||
// costRow builds an access-log entry carrying only a cost breakdown — the rest
|
||||
// of the row is irrelevant to the summation identities under test.
|
||||
func costRow(id, session string, ts time.Time, in, cachedIn, cacheCreate, out float64) *AgentNetworkAccessLog {
|
||||
return &AgentNetworkAccessLog{
|
||||
ID: id,
|
||||
SessionID: session,
|
||||
Timestamp: ts,
|
||||
InputCostUSD: in,
|
||||
CachedInputCostUSD: cachedIn,
|
||||
CacheCreationCostUSD: cacheCreate,
|
||||
OutputCostUSD: out,
|
||||
}
|
||||
}
|
||||
|
||||
// TestAPIResponse_CostComponentsSumToAggregates is the contract a client adding
|
||||
// up an API response depends on: within a single rendered object, the four
|
||||
// per-bucket fields sum to cost_usd, and the two cache fields sum to
|
||||
// cache_cost_usd. Uses rates that are not exactly representable in binary
|
||||
// floating point, so the identity is checked against real arithmetic rather
|
||||
// than round numbers.
|
||||
func TestAPIResponse_CostComponentsSumToAggregates(t *testing.T) {
|
||||
row := costRow("r1", "s1", time.Now(), 0.000768, 0.0002304, 0.00192, 0.003)
|
||||
|
||||
api := row.ToAPIResponse()
|
||||
assert.InDelta(t, api.InputCostUsd+api.CachedInputCostUsd+api.CacheCreationCostUsd+api.OutputCostUsd,
|
||||
api.CostUsd, 1e-12, "rendered buckets must sum to the rendered cost_usd")
|
||||
assert.InDelta(t, api.CachedInputCostUsd+api.CacheCreationCostUsd, api.CacheCostUsd, 1e-12,
|
||||
"rendered cache buckets must sum to the rendered cache_cost_usd")
|
||||
assert.InDelta(t, 0.0059184, api.CostUsd, 1e-12, "total is the exact sum, not a separately rounded figure")
|
||||
assert.InDelta(t, 0.0021504, api.CacheCostUsd, 1e-12, "cache cost is the exact sum of the two cache buckets")
|
||||
}
|
||||
|
||||
// TestSessionSummary_SumsMatchSummedEntries proves a session summary equals the
|
||||
// sum of the entries it renders: a client that adds up the entries itself must
|
||||
// land on the same number the summary reports, per bucket and in total.
|
||||
func TestSessionSummary_SumsMatchSummedEntries(t *testing.T) {
|
||||
base := time.Date(2026, 5, 5, 10, 0, 0, 0, time.UTC)
|
||||
entries := []*AgentNetworkAccessLog{
|
||||
costRow("r1", "s1", base, 0.000768, 0.0002304, 0.00192, 0.003),
|
||||
costRow("r2", "s1", base.Add(time.Minute), 0.000625, 0.0009375, 0, 0.005),
|
||||
costRow("r3", "s1", base.Add(2*time.Minute), 0.0000016, 0, 0, 0.0000032),
|
||||
}
|
||||
|
||||
sessions := FoldAccessLogSessions([]string{"s1"}, entries)
|
||||
require.Len(t, sessions, 1)
|
||||
sess := sessions[0].ToAPIResponse()
|
||||
|
||||
var wantInput, wantCachedInput, wantCacheCreation, wantOutput float64
|
||||
for _, e := range entries {
|
||||
wantInput += e.InputCostUSD
|
||||
wantCachedInput += e.CachedInputCostUSD
|
||||
wantCacheCreation += e.CacheCreationCostUSD
|
||||
wantOutput += e.OutputCostUSD
|
||||
}
|
||||
|
||||
assert.InDelta(t, wantInput, sess.InputCostUsd, 1e-12, "session input cost is the sum of its entries")
|
||||
assert.InDelta(t, wantCachedInput, sess.CachedInputCostUsd, 1e-12, "session cache-read cost is the sum of its entries")
|
||||
assert.InDelta(t, wantCacheCreation, sess.CacheCreationCostUsd, 1e-12, "session cache-write cost is the sum of its entries")
|
||||
assert.InDelta(t, wantOutput, sess.OutputCostUsd, 1e-12, "session output cost is the sum of its entries")
|
||||
assert.InDelta(t, wantInput+wantCachedInput+wantCacheCreation+wantOutput, sess.CostUsd, 1e-12,
|
||||
"session total equals the summed entry buckets")
|
||||
|
||||
// Summing the rendered entries must give the same answer as reading the
|
||||
// summary — the property a UI relies on when it totals a table itself.
|
||||
var fromEntries float64
|
||||
for _, e := range sess.Entries {
|
||||
fromEntries += e.CostUsd
|
||||
}
|
||||
assert.InDelta(t, sess.CostUsd, fromEntries, 1e-12, "summary total must match the summed rendered entries")
|
||||
|
||||
// The sub-microdollar row must still contribute; it would vanish under
|
||||
// 6-decimal quantisation.
|
||||
assert.Greater(t, sess.InputCostUsd, 0.001393, "small-cost rows must not be quantised away")
|
||||
}
|
||||
|
||||
// TestUsageBuckets_SumsMatchSummedRows proves the same identity one level up:
|
||||
// a usage bucket equals the sum of the ledger rows folded into it, and the
|
||||
// buckets together equal the whole range.
|
||||
func TestUsageBuckets_SumsMatchSummedRows(t *testing.T) {
|
||||
day1 := time.Date(2026, 5, 5, 9, 0, 0, 0, time.UTC)
|
||||
day2 := time.Date(2026, 5, 6, 9, 0, 0, 0, time.UTC)
|
||||
rows := []*AgentNetworkUsage{
|
||||
{ID: "u1", Timestamp: day1, InputCostUSD: 0.000768, CachedInputCostUSD: 0.0002304, CacheCreationCostUSD: 0.00192, OutputCostUSD: 0.003},
|
||||
{ID: "u2", Timestamp: day1.Add(time.Hour), InputCostUSD: 0.000625, CachedInputCostUSD: 0.0009375, OutputCostUSD: 0.005},
|
||||
{ID: "u3", Timestamp: day2, InputCostUSD: 0.0000016, OutputCostUSD: 0.0000032},
|
||||
}
|
||||
|
||||
buckets := AggregateUsageByGranularity(rows, UsageGranularityDay)
|
||||
require.Len(t, buckets, 2, "two distinct days expected")
|
||||
|
||||
var total, cache float64
|
||||
for _, b := range buckets {
|
||||
api := b.ToAPIResponse()
|
||||
assert.InDelta(t, api.InputCostUsd+api.CachedInputCostUsd+api.CacheCreationCostUsd+api.OutputCostUsd,
|
||||
api.CostUsd, 1e-12, "each bucket's components must sum to its cost_usd")
|
||||
total += api.CostUsd
|
||||
cache += api.CacheCostUsd
|
||||
}
|
||||
|
||||
var wantTotal, wantCache float64
|
||||
for _, r := range rows {
|
||||
wantTotal += r.TotalCostUSD()
|
||||
wantCache += r.CacheCostUSD()
|
||||
}
|
||||
assert.InDelta(t, wantTotal, total, 1e-12, "buckets must sum to the total across all ledger rows")
|
||||
assert.InDelta(t, wantCache, cache, 1e-12, "buckets must sum to the cache spend across all ledger rows")
|
||||
|
||||
// A month bucket over the same rows must total identically — regrouping
|
||||
// changes the partition, never the sum.
|
||||
monthly := AggregateUsageByGranularity(rows, UsageGranularityMonth)
|
||||
require.Len(t, monthly, 1)
|
||||
assert.InDelta(t, wantTotal, monthly[0].ToAPIResponse().CostUsd, 1e-12,
|
||||
"re-bucketing at a different granularity must preserve the total")
|
||||
}
|
||||
@@ -25,8 +25,19 @@ type AgentNetworkUsage struct {
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CostUSD float64
|
||||
CreatedAt time.Time
|
||||
// Prompt-cache buckets: read + write token counts.
|
||||
CachedInputTokens int64
|
||||
CacheCreationTokens int64
|
||||
// Per-bucket cost breakdown, mirroring AgentNetworkAccessLog — the only
|
||||
// cost state stored; total and cache portion are derived on read. Kept on
|
||||
// the usage ledger too so spend can be attributed per bucket even for
|
||||
// accounts with log collection turned off. See AgentNetworkAccessLog for
|
||||
// why the columns carry a zero default.
|
||||
InputCostUSD float64 `gorm:"not null;default:0"`
|
||||
CachedInputCostUSD float64 `gorm:"not null;default:0"`
|
||||
CacheCreationCostUSD float64 `gorm:"not null;default:0"`
|
||||
OutputCostUSD float64 `gorm:"not null;default:0"`
|
||||
CreatedAt time.Time
|
||||
}
|
||||
|
||||
// TableName keeps usage records in their own stripped table. Named
|
||||
@@ -34,6 +45,17 @@ type AgentNetworkUsage struct {
|
||||
// agent_network_usage table in a shared database.
|
||||
func (AgentNetworkUsage) TableName() string { return "agent_network_request_usage" }
|
||||
|
||||
// TotalCostUSD is the request's total cost: the sum of the four per-bucket
|
||||
// costs. Derived rather than stored so it cannot disagree with the breakdown.
|
||||
func (u *AgentNetworkUsage) TotalCostUSD() float64 {
|
||||
return u.InputCostUSD + u.CachedInputCostUSD + u.CacheCreationCostUSD + u.OutputCostUSD
|
||||
}
|
||||
|
||||
// CacheCostUSD is the portion of the total billed for prompt-cache buckets.
|
||||
func (u *AgentNetworkUsage) CacheCostUSD() float64 {
|
||||
return u.CachedInputCostUSD + u.CacheCreationCostUSD
|
||||
}
|
||||
|
||||
// AgentNetworkUsageGroup is the normalised many-to-many row linking a usage
|
||||
// record to one authorising group, mirroring AgentNetworkAccessLogGroup so the
|
||||
// usage overview can filter by group with a `group_id IN (...)` join.
|
||||
|
||||
@@ -33,21 +33,45 @@ func ParseUsageGranularity(s string) UsageGranularity {
|
||||
// AgentNetworkUsageBucket is one aggregated usage time bucket. PeriodStart is
|
||||
// the UTC start of the bucket as YYYY-MM-DD.
|
||||
type AgentNetworkUsageBucket struct {
|
||||
PeriodStart string
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CostUSD float64
|
||||
PeriodStart string
|
||||
InputTokens int64
|
||||
OutputTokens int64
|
||||
TotalTokens int64
|
||||
CachedInputTokens int64
|
||||
CacheCreationTokens int64
|
||||
InputCostUSD float64
|
||||
CachedInputCostUSD float64
|
||||
CacheCreationCostUSD float64
|
||||
OutputCostUSD float64
|
||||
}
|
||||
|
||||
// TotalCostUSD is the bucket's total spend: the sum of the four per-bucket
|
||||
// costs. Derived rather than accumulated separately so it cannot disagree with
|
||||
// the components.
|
||||
func (b *AgentNetworkUsageBucket) TotalCostUSD() float64 {
|
||||
return b.InputCostUSD + b.CachedInputCostUSD + b.CacheCreationCostUSD + b.OutputCostUSD
|
||||
}
|
||||
|
||||
// CacheCostUSD is the bucket's prompt-cache spend: cache reads plus writes.
|
||||
func (b *AgentNetworkUsageBucket) CacheCostUSD() float64 {
|
||||
return b.CachedInputCostUSD + b.CacheCreationCostUSD
|
||||
}
|
||||
|
||||
// ToAPIResponse renders the bucket as the API representation.
|
||||
func (b *AgentNetworkUsageBucket) ToAPIResponse() api.AgentNetworkUsageBucket {
|
||||
return api.AgentNetworkUsageBucket{
|
||||
PeriodStart: b.PeriodStart,
|
||||
InputTokens: b.InputTokens,
|
||||
OutputTokens: b.OutputTokens,
|
||||
TotalTokens: b.TotalTokens,
|
||||
CostUsd: b.CostUSD,
|
||||
PeriodStart: b.PeriodStart,
|
||||
InputTokens: b.InputTokens,
|
||||
OutputTokens: b.OutputTokens,
|
||||
TotalTokens: b.TotalTokens,
|
||||
CachedInputTokens: b.CachedInputTokens,
|
||||
CacheCreationTokens: b.CacheCreationTokens,
|
||||
InputCostUsd: b.InputCostUSD,
|
||||
CachedInputCostUsd: b.CachedInputCostUSD,
|
||||
CacheCreationCostUsd: b.CacheCreationCostUSD,
|
||||
OutputCostUsd: b.OutputCostUSD,
|
||||
CostUsd: b.TotalCostUSD(),
|
||||
CacheCostUsd: b.CacheCostUSD(),
|
||||
}
|
||||
}
|
||||
|
||||
@@ -84,7 +108,12 @@ func AggregateUsageByGranularity(rows []*AgentNetworkUsage, g UsageGranularity)
|
||||
b.InputTokens += r.InputTokens
|
||||
b.OutputTokens += r.OutputTokens
|
||||
b.TotalTokens += r.TotalTokens
|
||||
b.CostUSD += r.CostUSD
|
||||
b.CachedInputTokens += r.CachedInputTokens
|
||||
b.CacheCreationTokens += r.CacheCreationTokens
|
||||
b.InputCostUSD += r.InputCostUSD
|
||||
b.CachedInputCostUSD += r.CachedInputCostUSD
|
||||
b.CacheCreationCostUSD += r.CacheCreationCostUSD
|
||||
b.OutputCostUSD += r.OutputCostUSD
|
||||
}
|
||||
|
||||
out := make([]*AgentNetworkUsageBucket, 0, len(byPeriod))
|
||||
|
||||
Reference in New Issue
Block a user