Add Linux ARM64 UI build

## Describe your changes

Export agent version info for iOS

## Issue ticket number and link

## Stack

<!-- branch-stack -->

### Checklist
- [ ] Is it a bug fix
- [ ] Is a typo/documentation fix
- [ ] Is a feature enhancement
- [x] It is a refactor
- [ ] Created tests that fail without the change (if possible)
- [ ] This change does **not** modify the public API, gRPC protocols,
functionality behavior, CLI / service flags, or introduce a new feature
— **OR** I have discussed it with the NetBird team beforehand (link the
issue / Slack thread in the description). See
[CONTRIBUTING.md](https://github.com/netbirdio/netbird/blob/main/CONTRIBUTING.md#discuss-changes-with-the-netbird-team-first).

> By submitting this pull request, you confirm that you have read and
agree to the terms of the [Contributor License
Agreement](https://github.com/netbirdio/netbird/blob/main/CONTRIBUTOR_LICENSE_AGREEMENT.md).

## Documentation
Select exactly one:

- [ ] I added/updated documentation for this change
- [x] Documentation is **not needed** for this change (explain why)

### Docs PR URL (required if "docs added" is checked)
Paste the PR link from https://github.com/netbirdio/docs here:

https://github.com/netbirdio/docs/pull/__

<!-- codesmith:footer -->
---
<a
href="https://app.blacksmith.sh/netbirdio/codesmith/netbird/pr/6918"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-dark-v2.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-light-v2.svg"><img
alt="View with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/view-with-codesmith-dark-v2.svg"></picture></a>
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1787750541&installation_model_id=427504&pr_number=6918&repository=netbirdio%2Fnetbird&return_to=https%3A%2F%2Fgithub.com%2Fnetbirdio%2Fnetbird%2Fpull%2F6918&signature=c2de74d89c36b01aac5866422b0be0b7525f7c6e26e46174efc280339eb864b6"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"></picture></a>
<sup>Need help on this PR? Tag <code>@codesmith-bot</code> with what you
need. Autofix is disabled.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->

.github/workflows/agent-network-e2e.yml

@@ -5,6 +5,13 @@ on:
   schedule:
     - cron: "0 3 * * *"
   workflow_dispatch:
+    inputs:
+      bedrock_model:
+        description: >-
+          Bedrock inference-profile id to drive the matrix with, exactly as
+          AWS issues it. Leave empty for the Sonnet 4.6 default.
+        required: false
+        default: ""
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -62,6 +69,8 @@ jobs:
           CLOUDFLARE_TOKEN: ${{ secrets.E2E_CLOUDFLARE_TOKEN }}
           AWS_BEARER_TOKEN_BEDROCK: ${{ secrets.E2E_AWS_BEARER_TOKEN_BEDROCK }}
           AWS_REGION: ${{ secrets.E2E_AWS_REGION }}
+          # Bedrock model override: dispatch input wins, then the repo variable, else the test default.
+          AWS_BEDROCK_MODEL: ${{ inputs.bedrock_model || vars.E2E_AWS_BEDROCK_MODEL }}
           # Vertex (Anthropic-on-Vertex): SA + project required; region defaults
           # to "global", model to a pinned claude snapshot.
           GOOGLE_VERTEX_SA_BASE64: ${{ secrets.E2E_GOOGLE_VERTEX_SA_BASE64 }}

.goreleaser.yaml

@@ -273,8 +273,8 @@ dockers_v2:
        - netbirdio/netbird
        - ghcr.io/netbirdio/netbird
      tags:
-       - "v{{ .Version }}-rootless"
-       - "{{ if eq .Env.SKIP_PUBLISH \"false\" }}latest{{ end }}"
+       - "{{ .Version }}-rootless"
+       - "{{ if eq .Env.SKIP_PUBLISH \"false\" }}rootless-latest{{ end }}"
      dockerfile: client/Dockerfile-rootless
      extra_files:
        - client/netbird-entrypoint.sh

.goreleaser_ui.yaml

@@ -21,6 +21,7 @@ builds:
       - linux
     goarch:
       - amd64
+      - arm64
     ldflags:
       - -s -w -X github.com/netbirdio/netbird/version.version={{.Version}} -X main.commit={{.Commit}} -X main.date={{.CommitDate}} -X main.builtBy=goreleaser
     mod_timestamp: "{{ .CommitTimestamp }}"

client/ios/NetBirdSDK/version.go

@@ -0,0 +1,12 @@
+//go:build ios
+
+package NetBirdSDK
+
+import "github.com/netbirdio/netbird/version"
+
+// GoClientVersion returns the NetBird Go client version that was baked into
+// the framework at compile time via
+// -ldflags "-X github.com/netbirdio/netbird/version.version=<version>".
+func GoClientVersion() string {
+	return version.NetbirdVersion()
+}

docs/agent-networks/01-end-to-end-flows.md

@@ -109,7 +109,7 @@ sequenceDiagram
         Chk->>Inj: continue
         Inj->>Inj: inject NetBird identity headers per provider config
         Inj->>Grd: continue
-        Grd->>Grd: enforce per-provider allowlist (fail-closed backstop)
+        Grd->>Grd: enforce model allowlist
         Grd->>Up: forward (over WireGuard)
         Up-->>Resp: response (JSON or SSE stream)
         Resp->>Resp: parse usage tokens, completion
@@ -135,21 +135,6 @@ sequenceDiagram
   (`redact_pii = settings.RedactPii`). Phones, emails, credit cards,
   PII names — see `redact.go` for the full set. See
   [`modules/31-proxy-middleware-builtin.md`](modules/31-proxy-middleware-builtin.md).
-- The model allowlist is enforced in TWO places. `CheckLLMPolicyLimits`
-  is authoritative: it resolves the policy that governs this
-  (provider, caller-groups) and denies (`deny_code = llm_policy.model_blocked`)
-  when no applicable policy permits the model — so an allowlist scoped to
-  one group/provider never leaks to another, and an un-guardrailed policy
-  is genuinely unrestricted. `llm_guardrail` is a per-provider fail-closed
-  backstop: it only carries an allowlist for a provider every authorising
-  policy restricts, and blocks unknown/undetermined models even when
-  management is unreachable. Because that backstop allowlist is the UNION
-  of every restricting policy's models, per-group narrowing lives only in
-  the authoritative check: during a `CheckLLMPolicyLimits` outage
-  `llm_limit_check` fails open, so a caller can reach any model in the
-  provider's union — a group scoped to model A could reach model B if
-  another group restricts the same provider to B. This is the documented
-  fail-open trade-off; a future flag may switch it to fail-closed.
 - SSE streaming requires special handling on the response side; the
   parser must handle partial chunks without buffering the whole
   stream. See [`modules/32-proxy-llm-parsers.md`](modules/32-proxy-llm-parsers.md).

docs/agent-networks/modules/21-management-agentnetwork.md

@@ -122,7 +122,7 @@ At request time the path is independent: the proxy calls `SelectPolicyForRequest
   | on_request | 1 | `llm_router` | `{"providers":[{id, models[], upstream_*, auth_header_*, allowed_group_ids[]}]}` | **true** |
   | on_request | 2 | `llm_limit_check` | `{}` | – |
   | on_request | 3 | `llm_identity_inject` | `{"providers":[{provider_id, header_pair?, json_metadata?, extra_headers?}]}` | **true** |
-  | on_request | 4 | `llm_guardrail` | `{"provider_allowlists"?: {providerID: []model}, "prompt_capture":{enabled,redact_pii}}` | – |
+  | on_request | 4 | `llm_guardrail` | `{"model_allowlist"?, "prompt_capture":{enabled,redact_pii}}` | – |
   | on_response | 5 | `llm_limit_record` | `{}` (runs LAST at runtime) | – |
   | on_response | 6 | `cost_meter` | `{}` | – |
   | on_response | 7 | `llm_response_parser` | `{"capture_completion": <bool>, "redact_pii"?: true}` | – |

docs/agent-networks/modules/31-proxy-middleware-builtin.md

@@ -244,7 +244,7 @@ no mocks. Tests: `TestChain_AllowPath_StampsAttributionAndRecordsCounter`
 | `llm_router` | `{providers: [{id, models, upstream_scheme, upstream_host, upstream_path?, auth_header_name, auth_header_value, allowed_group_ids}]}` |
 | `llm_limit_check` | `{}` — pulls `MgmtClient` from `FactoryContext` |
 | `llm_identity_inject` | `{providers: [{provider_id, header_pair?|json_metadata?, extra_headers?}]}` |
-| `llm_guardrail` | `{provider_allowlists: {providerID: []string}, prompt_capture: {enabled, redact_pii}}` — allowlist keyed by resolved provider id; a provider absent from the map is unrestricted (fail-closed backstop; authoritative per-policy/group check is management's `CheckLLMPolicyLimits`) |
+| `llm_guardrail` | `{model_allowlist: []string, prompt_capture: {enabled, redact_pii}}` |
 | `llm_response_parser` | `{redact_pii?, capture_completion?: *bool}` |
 | `cost_meter` | `{pricing_path?}` (basename inside data-dir; defaults `pricing.yaml`) |
 | `llm_limit_record` | `{}` — same pattern as `llm_limit_check` |

e2e/agentnetwork/chat_test.go

@@ -9,12 +9,220 @@ import (
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"gorm.io/driver/sqlite"
+	"gorm.io/gorm"
 
 	"github.com/netbirdio/netbird/e2e/harness"
 	"github.com/netbirdio/netbird/shared/management/http/api"
 )
 
+// per1k is a model's published USD rates per 1k tokens. read is the prompt-cache read rate
+// (OpenAI: the cached-input discount rate); write is the cache-creation rate where one exists.
+type per1k struct{ in, out, read, write float64 }
+
+// publishedPer1k hardcodes the vendors' PUBLISHED rates for the models the live matrix can drive,
+// keyed by the normalized model id the proxy stamps. Deliberately independent of the proxy's
+// pricing table so a wrong embedded rate or a broken normalization fails the run.
+var publishedPer1k = map[string]per1k{
+	"gpt-4o-mini":                 {0.00015, 0.0006, 0.000075, 0},
+	"gpt-4o":                      {0.0025, 0.01, 0.00125, 0},
+	"claude-haiku-4-5":            {0.001, 0.005, 0.0001, 0.00125},
+	"claude-sonnet-4-5":           {0.003, 0.015, 0.0003, 0.00375},
+	"claude-sonnet-4-6":           {0.003, 0.015, 0.0003, 0.00375},
+	"kimi-k3":                     {0.003, 0.015, 0.0003, 0.003}, // no published write rate: bills at the input rate
+	"anthropic.claude-haiku-4-5":  {0.001, 0.005, 0.0001, 0.00125},
+	"anthropic.claude-sonnet-4-5": {0.003, 0.015, 0.0003, 0.00375},
+	"anthropic.claude-sonnet-4-6": {0.003, 0.015, 0.0003, 0.00375},
+}
+
+// rawCostVerificationSQL is the operator-facing double-check, run straight against the management
+// sqlite store: recompute each usage row's expected total and cache cost from its own persisted
+// token buckets and hardcoded published rates. OpenAI counts cached tokens as a subset of input;
+// Anthropic-shape providers count cache buckets additively.
+const rawCostVerificationSQL = `
+WITH rates(model, in_rate, out_rate, read_rate, write_rate) AS (
+  VALUES
+    ('gpt-4o-mini',                 0.00015, 0.0006, 0.000075, 0.0),
+    ('gpt-4o',                      0.0025,  0.01,   0.00125,  0.0),
+    ('claude-haiku-4-5',            0.001,   0.005,  0.0001,   0.00125),
+    ('claude-sonnet-4-5',           0.003,   0.015,  0.0003,   0.00375),
+    ('claude-sonnet-4-6',           0.003,   0.015,  0.0003,   0.00375),
+    ('kimi-k3',                     0.003,   0.015,  0.0003,   0.003),
+    ('anthropic.claude-haiku-4-5',  0.001,   0.005,  0.0001,   0.00125),
+    ('anthropic.claude-sonnet-4-5', 0.003,   0.015,  0.0003,   0.00375),
+    ('anthropic.claude-sonnet-4-6', 0.003,   0.015,  0.0003,   0.00375)
+)
+SELECT
+  u.provider,
+  u.model,
+  u.input_tokens,
+  u.output_tokens,
+  u.cached_input_tokens,
+  u.cache_creation_tokens,
+  u.input_cost_usd,
+  u.cached_input_cost_usd,
+  u.cache_creation_cost_usd,
+  u.output_cost_usd,
+  -- No cost_usd / cache_cost_usd columns are stored: both are derived from the
+  -- four per-bucket columns above, exactly as the API renders them.
+  (u.input_cost_usd + u.cached_input_cost_usd + u.cache_creation_cost_usd + u.output_cost_usd) AS cost_usd,
+  (u.cached_input_cost_usd + u.cache_creation_cost_usd) AS cache_cost_usd,
+  CASE WHEN u.provider = 'openai' THEN
+    (u.input_tokens - MIN(u.cached_input_tokens, u.input_tokens))*r.in_rate/1000.0
+  ELSE
+    u.input_tokens*r.in_rate/1000.0
+  END AS expected_input,
+  CASE WHEN u.provider = 'openai' THEN
+    MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+  ELSE
+    u.cached_input_tokens*r.read_rate/1000.0
+  END AS expected_cached_input,
+  CASE WHEN u.provider = 'openai' THEN
+    0.0
+  ELSE
+    u.cache_creation_tokens*r.write_rate/1000.0
+  END AS expected_cache_creation,
+  u.output_tokens*r.out_rate/1000.0 AS expected_output,
+  CASE WHEN u.provider = 'openai' THEN
+    (u.input_tokens - MIN(u.cached_input_tokens, u.input_tokens))*r.in_rate/1000.0
+      + MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+      + u.output_tokens*r.out_rate/1000.0
+  ELSE
+    u.input_tokens*r.in_rate/1000.0 + u.cached_input_tokens*r.read_rate/1000.0
+      + u.cache_creation_tokens*r.write_rate/1000.0 + u.output_tokens*r.out_rate/1000.0
+  END AS expected_total,
+  CASE WHEN u.provider = 'openai' THEN
+    MIN(u.cached_input_tokens, u.input_tokens)*r.read_rate/1000.0
+  ELSE
+    u.cached_input_tokens*r.read_rate/1000.0 + u.cache_creation_tokens*r.write_rate/1000.0
+  END AS expected_cache
+FROM agent_network_request_usage u
+JOIN rates r ON r.model = u.model
+ORDER BY u.timestamp`
+
+// verifyUsageRowsSQL re-checks every persisted usage row directly in the management sqlite store,
+// bypassing the API path — the same audit an operator can run on a production store.db.
+func verifyUsageRowsSQL(t *testing.T, srv *harness.Combined) {
+	t.Helper()
+
+	dbPath, err := srv.SnapshotStoreDB(t.TempDir())
+	require.NoError(t, err, "snapshot management sqlite store")
+
+	db, err := gorm.Open(sqlite.Open(dbPath), &gorm.Config{})
+	require.NoError(t, err, "open store snapshot")
+	sqlDB, err := db.DB()
+	require.NoError(t, err)
+	defer func() { _ = sqlDB.Close() }()
+
+	rows, err := db.Raw(rawCostVerificationSQL).Rows()
+	require.NoError(t, err, "run raw cost verification query")
+	defer func() { _ = rows.Close() }()
+
+	verified := 0
+	for rows.Next() {
+		var provider, model string
+		var inTok, outTok, readTok, writeTok int64
+		var inCost, cachedInCost, cacheCreateCost, outCost, cost, cacheCost float64
+		var wantInput, wantCachedInput, wantCacheCreation, wantOutput, wantTotal, wantCache float64
+		require.NoError(t, rows.Scan(&provider, &model, &inTok, &outTok, &readTok, &writeTok,
+			&inCost, &cachedInCost, &cacheCreateCost, &outCost, &cost, &cacheCost,
+			&wantInput, &wantCachedInput, &wantCacheCreation, &wantOutput, &wantTotal, &wantCache), "scan usage row")
+		t.Logf("[sql] %s/%s: in=%d out=%d cache_read=%d cache_write=%d stored in/cached/create/out=$%.6f/$%.6f/$%.6f/$%.6f total=$%.6f cache=$%.6f expected total=$%.6f cache=$%.6f",
+			provider, model, inTok, outTok, readTok, writeTok,
+			inCost, cachedInCost, cacheCreateCost, outCost, cost, cacheCost, wantTotal, wantCache)
+		assert.InDeltaf(t, wantInput, inCost, 1e-6, "stored input_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCachedInput, cachedInCost, 1e-6, "stored cached_input_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCacheCreation, cacheCreateCost, 1e-6, "stored cache_creation_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantOutput, outCost, 1e-6, "stored output_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantTotal, cost, 1e-6, "derived cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, wantCache, cacheCost, 1e-6, "derived cache_cost_usd for %s/%s must match the published-rate recompute", provider, model)
+		assert.InDeltaf(t, inCost+cachedInCost+cacheCreateCost+outCost, cost, 1e-9,
+			"stored buckets must sum to the derived cost_usd for %s/%s", provider, model)
+		verified++
+	}
+	require.NoError(t, rows.Err(), "iterate usage rows")
+	require.Positive(t, verified, "raw SQL check must cover at least one usage row")
+	t.Logf("[sql] verified %d usage rows in store.db against published rates", verified)
+
+	gwRows, err := db.Raw(`SELECT model,
+	  (input_cost_usd + cached_input_cost_usd + cache_creation_cost_usd + output_cost_usd) AS cost_usd
+	  FROM agent_network_request_usage WHERE model LIKE '%/%'`).Rows()
+	require.NoError(t, err, "query gateway-prefixed usage rows")
+	defer func() { _ = gwRows.Close() }()
+	for gwRows.Next() {
+		var model string
+		var cost float64
+		require.NoError(t, gwRows.Scan(&model, &cost), "scan gateway usage row")
+		t.Logf("[sql] gateway %s: stored=$%.6f (must be 0 — deliberately unpriced)", model, cost)
+		assert.Zerof(t, cost, "gateway-prefixed model %q must store cost 0, never a guessed rate", model)
+	}
+	require.NoError(t, gwRows.Err(), "iterate gateway usage rows")
+}
+
+// validateAccessLogCost recomputes a live access-log row's expected total and cache cost from the
+// published per-1k rates and the row's persisted token buckets, and asserts both stored values.
+// Gateway-prefixed model ids the proxy deliberately does not price must store cost 0.
+func validateAccessLogCost(t *testing.T, pc providerCase, row api.AgentNetworkAccessLog) {
+	t.Helper()
+	model := catalogModel(pc)
+	provider := ""
+	if row.Provider != nil {
+		provider = *row.Provider
+	}
+	t.Logf("[cost] %s: provider=%s model=%s in=%d out=%d total=%d cache_read=%d cache_write=%d cost=$%.6f cache_cost=$%.6f",
+		pc.name, provider, model, row.InputTokens, row.OutputTokens, row.TotalTokens,
+		row.CachedInputTokens, row.CacheCreationTokens, row.CostUsd, row.CacheCostUsd)
+
+	rates, known := publishedPer1k[model]
+	if !known {
+		if strings.Contains(model, "/") {
+			assert.Zerof(t, row.CostUsd, "gateway-prefixed model %q is not priced so the cost meter must skip (cost 0)", model)
+			return
+		}
+		t.Logf("[cost] %s: no published rate on file for model %q (env-overridden?); skipping cost validation", pc.name, model)
+		return
+	}
+
+	// input_tokens may legitimately be 0: Moonshot/Kimi reports fully cached prompts under the cache
+	// buckets only. Output and total must always be present on a priced row.
+	require.Positive(t, row.OutputTokens, "priced row must carry output tokens")
+	require.Positive(t, row.TotalTokens, "priced row must carry total tokens")
+
+	var wantInput, wantCachedInput, wantCacheCreation float64
+	if provider == "openai" {
+		cached := min(row.CachedInputTokens, row.InputTokens) // cached is a subset of input
+		wantInput = float64(row.InputTokens-cached) / 1000 * rates.in
+		wantCachedInput = float64(cached) / 1000 * rates.read
+		// OpenAI has no cache-write bucket; wantCacheCreation stays 0.
+	} else {
+		// Anthropic / Bedrock shape: cache buckets are additive to input_tokens.
+		wantInput = float64(row.InputTokens) / 1000 * rates.in
+		wantCachedInput = float64(row.CachedInputTokens) / 1000 * rates.read
+		wantCacheCreation = float64(row.CacheCreationTokens) / 1000 * rates.write
+	}
+	wantOutput := float64(row.OutputTokens) / 1000 * rates.out
+	wantCache := wantCachedInput + wantCacheCreation
+	wantTotal := wantInput + wantCache + wantOutput
+
+	t.Logf("[cost] %s: expecting input=$%.6f cached_input=$%.6f cache_creation=$%.6f output=$%.6f total=$%.6f cache=$%.6f from published rates",
+		pc.name, wantInput, wantCachedInput, wantCacheCreation, wantOutput, wantTotal, wantCache)
+	assert.InDeltaf(t, wantInput, row.InputCostUsd, 1e-6, "stored input_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCachedInput, row.CachedInputCostUsd, 1e-6, "stored cached_input_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCacheCreation, row.CacheCreationCostUsd, 1e-6, "stored cache_creation_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantOutput, row.OutputCostUsd, 1e-6, "stored output_cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantTotal, row.CostUsd, 1e-6, "derived cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, wantCache, row.CacheCostUsd, 1e-6, "derived cache_cost_usd for %s (%s)", pc.name, model)
+
+	// The aggregates must be exactly the sum of the stored components, not an
+	// independently-computed figure that could drift from the breakdown.
+	assert.InDeltaf(t, row.InputCostUsd+row.CachedInputCostUsd+row.CacheCreationCostUsd+row.OutputCostUsd,
+		row.CostUsd, 1e-9, "stored buckets must sum to the derived cost_usd for %s (%s)", pc.name, model)
+	assert.InDeltaf(t, row.CachedInputCostUsd+row.CacheCreationCostUsd,
+		row.CacheCostUsd, 1e-9, "stored cache buckets must sum to the derived cache_cost_usd for %s (%s)", pc.name, model)
+}
+
 // providerCase is one entry in the live provider matrix. The same scenario runs
 // for every available provider; availability is keyed off env vars so the suite
 // covers whatever credentials are present (source ~/.llm-keys locally / set the
@@ -116,12 +324,12 @@ func availableProviders() []providerCase {
 		if region == "" {
 			region = "eu-central-1"
 		}
-		// A valid Bedrock inference-profile id (region prefix + date + version),
-		// overridable per account. `global.` profiles can be invoked from any
-		// region; set AWS_BEDROCK_MODEL to match the enabled profile for the token.
+		// A valid Bedrock inference-profile id, overridable per account (AWS_BEDROCK_MODEL, also the
+		// workflow's bedrock_model dispatch input). `global.` profiles work from any region. Defaults to
+		// Sonnet 4.6, whose id convention dropped the -YYYYMMDD-v1:0 suffix that Haiku 4.5 still carries.
 		model := os.Getenv("AWS_BEDROCK_MODEL")
 		if model == "" {
-			model = "global.anthropic.claude-haiku-4-5-20251001-v1:0"
+			model = "global.anthropic.claude-sonnet-4-6"
 		}
 		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: model, kind: harness.WireBedrock})
 	}
@@ -257,6 +465,10 @@ func TestProvidersMatrix(t *testing.T) {
 			// session id and confirm the marker propagated end-to-end.
 			sessionID := "e2e-session-" + pc.name
 
+			// A long-form prompt so completions carry realistic token counts for cost validation;
+			// max_tokens in the harness bodies (2048) lets the full answer through.
+			const matrixPrompt = "explain GitHub workflow in 1000 words"
+
 			// Retry briefly to absorb tunnel/DNS jitter on the first call.
 			var code int
 			var body string
@@ -267,11 +479,11 @@ func TestProvidersMatrix(t *testing.T) {
 				var cerr error
 				switch pc.kind {
 				case harness.WireVertex:
-					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, matrixPrompt, sessionID)
 				case harness.WireBedrock:
-					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, matrixPrompt, sessionID)
 				default:
-					c, b, cerr = cl.ChatPrefixed(ctx, settings.Endpoint, proxyIP, pc.pathPrefix, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
+					c, b, cerr = cl.ChatPrefixed(ctx, settings.Endpoint, proxyIP, pc.pathPrefix, pc.kind, pc.model, matrixPrompt, sessionID)
 				}
 				if cerr == nil {
 					code, body = c, b
@@ -290,6 +502,7 @@ func TestProvidersMatrix(t *testing.T) {
 
 			// The session id sent as x-session-id must round-trip into the
 			// access-log row for this provider.
+			var row api.AgentNetworkAccessLog
 			require.Eventually(t, func() bool {
 				logs, lerr := srv.ListAccessLogs(ctx)
 				if lerr != nil {
@@ -297,11 +510,15 @@ func TestProvidersMatrix(t *testing.T) {
 				}
 				for _, r := range logs.Data {
 					if r.SessionId != nil && *r.SessionId == sessionID {
+						row = r
 						return true
 					}
 				}
 				return false
 			}, 30*time.Second, 2*time.Second, "session id %q must be recorded in an access-log row for %s", sessionID, pc.name)
+
+			// Stored total and cache cost must match the published rates applied to the row's buckets.
+			validateAccessLogCost(t, pc, row)
 		})
 	}
 
@@ -322,4 +539,7 @@ func TestProvidersMatrix(t *testing.T) {
 		}
 		return false
 	}, 60*time.Second, 3*time.Second, "consumption must be recorded with positive token counts after live traffic")
+
+	// Final raw-SQL audit: bypass the API and re-verify every persisted usage row in the store.
+	verifyUsageRowsSQL(t, srv)
 }

e2e/agentnetwork/guardrail_block_test.go

@@ -1,209 +0,0 @@
-//go:build e2e
-
-package agentnetwork
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/e2e/harness"
-	"github.com/netbirdio/netbird/shared/management/http/api"
-)
-
-// pathRoutedGuardrailCase is one provider's self-contained scenario: its own
-// provider, its own guardrail whose allowlist holds ONLY that provider's
-// allowed model, and its own policy. Each case runs in isolation (its own
-// proxy + client), so the guardrail the proxy enforces contains exactly this
-// provider's model — never a mixed cross-provider list.
-type pathRoutedGuardrailCase struct {
-	name       string
-	catalogID  string // agent-network catalog provider id
-	wire       string // harness.WireVertex | harness.WireBedrock
-	allowEntry string // the single model id put on the guardrail allowlist
-	allowModel string // model id sent that MUST be served (200)
-	blockModel string // model id sent that MUST be denied (403 model_blocked)
-}
-
-// TestGuardrailBlocksUnselectedModel_PathRouted is the end-to-end regression
-// guard for the customer report that a model-allowlist guardrail attached to a
-// policy has no effect for PATH-ROUTED providers — where the model travels in
-// the URL, not the JSON body: Google Vertex (…/models/{model}:rawPredict) and
-// AWS Bedrock (/model/{id}/invoke).
-//
-// Each provider is tested in isolation with a guardrail allowlisting a single
-// model of its own: the allowed model (in the URL path) is served (200) and an
-// unselected model (in the URL path) is denied 403 by the guardrail
-// (llm_policy.model_blocked) before the upstream. The Vertex case mirrors the
-// customer verbatim — allow Sonnet, and the unselected model is the exact
-// claude-opus-4-6 they reported reaching the model unblocked. The Bedrock case
-// sends a region-prefixed, versioned inference-profile id so URL-path model
-// normalization is exercised too.
-//
-// The provider is catch-all (no models), so the router forwards any model and a
-// 403 can only come from the guardrail, never model_not_routable. Only the
-// upstream LLM is mocked (the vLLM nginx answers any path with 200); management
-// synth/reconcile, the proxy middleware chain (URL-path model extraction,
-// router, guardrail) and the tunnel are all real, and the guardrail denies
-// before the upstream is dialed so the mock cannot influence the block. A
-// static bearer api key is used so the router injects a static Authorization
-// header instead of minting a GCP token — the only reason path-routed providers
-// normally need live credentials — so the test runs with none and is always on.
-func TestGuardrailBlocksUnselectedModel_PathRouted(t *testing.T) {
-	cases := []pathRoutedGuardrailCase{
-		{
-			name:       "vertex",
-			catalogID:  "vertex_ai_api",
-			wire:       harness.WireVertex,
-			allowEntry: "claude-sonnet-4-5",
-			allowModel: "claude-sonnet-4-5",
-			blockModel: "claude-opus-4-6", // the customer-reported model
-		},
-		{
-			name:       "bedrock",
-			catalogID:  "bedrock_api",
-			wire:       harness.WireBedrock,
-			allowEntry: "anthropic.claude-sonnet-4-5", // normalized catalog id
-			allowModel: "us.anthropic.claude-sonnet-4-5-v1:0",
-			blockModel: "us.anthropic.claude-opus-4-8-v1:0",
-		},
-	}
-
-	for _, tc := range cases {
-		tc := tc
-		t.Run(tc.name, func(t *testing.T) {
-			runPathRoutedGuardrailCase(t, tc)
-		})
-	}
-}
-
-func runPathRoutedGuardrailCase(t *testing.T, tc pathRoutedGuardrailCase) {
-	t.Helper()
-
-	const (
-		vertexProject = "e2e-project"
-		vertexRegion  = "global"
-	)
-
-	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
-	defer cancel()
-
-	vllm, err := harness.StartVLLM(ctx, srv)
-	require.NoError(t, err, "start mock upstream")
-	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
-
-	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-" + tc.name})
-	require.NoError(t, err, "create group")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
-
-	ephemeral := false
-	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
-		Name:       "e2e-guardrail-" + tc.name + "-client",
-		Type:       "reusable",
-		ExpiresIn:  86400,
-		UsageLimit: 0,
-		AutoGroups: []string{grp.Id},
-		Ephemeral:  &ephemeral,
-	})
-	require.NoError(t, err, "mint setup key")
-	require.NotEmpty(t, sk.Key, "setup key plaintext")
-
-	// Catch-all provider (no models) so the router forwards any model; a static
-	// bearer key means the router injects a static auth header instead of minting
-	// a GCP token. Bootstraps the cluster if it isn't already.
-	staticKey := "static-e2e-token"
-	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:             tc.name,
-		ProviderId:       tc.catalogID,
-		UpstreamUrl:      vllm.URL,
-		ApiKey:           &staticKey,
-		Enabled:          ptr(true),
-		BootstrapCluster: ptr(harness.AgentNetworkCluster),
-	})
-	require.NoError(t, err, "create %s provider", tc.name)
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
-
-	// Guardrail allowlisting ONLY this provider's allowed model.
-	var gr api.AgentNetworkGuardrailRequest
-	gr.Name = "e2e-guardrail-" + tc.name
-	gr.Checks.ModelAllowlist.Enabled = true
-	gr.Checks.ModelAllowlist.Models = []string{tc.allowEntry}
-	guard, err := srv.CreateGuardrail(ctx, gr)
-	require.NoError(t, err, "create guardrail")
-	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
-
-	enabled := true
-	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-guardrail-" + tc.name,
-		Enabled:                &enabled,
-		SourceGroups:           []string{grp.Id},
-		DestinationProviderIds: []string{prov.Id},
-		GuardrailIds:           &[]string{guard.Id},
-	})
-	require.NoError(t, err, "create policy")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
-
-	settings, err := srv.GetSettings(ctx)
-	require.NoError(t, err, "read settings")
-	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
-
-	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-guardrail-"+tc.name+"-proxy")
-	require.NoError(t, err, "mint proxy token")
-	px, err := harness.StartProxy(ctx, srv, proxyToken)
-	require.NoError(t, err, "start proxy")
-	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
-
-	cl, err := harness.StartClient(ctx, srv, sk.Key)
-	require.NoError(t, err, "start client")
-	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
-
-	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
-	// Probe first: the GET resolves the endpoint and its first packet wakes the
-	// lazy proxy peer, so WaitProxyPeer then observes it connected.
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
-	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
-		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
-	}
-
-	send := func(model string) (int, string) {
-		var code int
-		var body string
-		var cerr error
-		switch tc.wire {
-		case harness.WireVertex:
-			code, body, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, vertexProject, vertexRegion, model, "Reply with exactly: pong", "")
-		case harness.WireBedrock:
-			code, body, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, model, "Reply with exactly: pong", "")
-		default:
-			t.Fatalf("unsupported wire %q", tc.wire)
-		}
-		require.NoError(t, cerr, "request must reach the proxy for %s", tc.name)
-		return code, body
-	}
-
-	// Allowed model (in the URL path) is served. Retry to absorb tunnel/DNS
-	// jitter on the first call over the freshly warmed tunnel.
-	var code int
-	var body string
-	deadline := time.Now().Add(90 * time.Second)
-	for time.Now().Before(deadline) {
-		code, body = send(tc.allowModel)
-		if code == 200 {
-			break
-		}
-		time.Sleep(5 * time.Second)
-	}
-	assert.Equal(t, 200, code,
-		"allowed %s model (URL path) must be served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
-
-	// Unselected model (in the URL path) must be blocked by the guardrail.
-	code, body = send(tc.blockModel)
-	assert.Equal(t, 403, code,
-		"unselected %s model (URL path) must be denied, not served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
-	assert.Contains(t, body, "llm_policy.model_blocked",
-		"%s denial must come from the guardrail allowlist, not routing; body: %s", tc.name, body)
-}

e2e/agentnetwork/guardrail_groupswitch_test.go

@@ -1,205 +0,0 @@
-//go:build e2e
-
-package agentnetwork
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/e2e/harness"
-	"github.com/netbirdio/netbird/shared/management/http/api"
-)
-
-// TestGuardrailGroupSwitchTakesEffectAfterTTL proves that moving a peer between
-// groups flips its model-allowlist decision once the proxy's tunnel-peer cache
-// expires. The peer's groups reach the guardrail via ValidateTunnelPeer, which
-// the proxy caches; the switch is invisible until that cache expires. The proxy
-// runs with a short NB_PROXY_TUNNEL_CACHE_TTL so the flip happens in seconds
-// instead of the 5-minute default.
-//
-// Setup: one catch-all provider declaring modelA + modelB; polA (grpA -> allow
-// modelA) and polB (grpB -> allow modelB). The client starts in grpA. modelA is
-// served and modelB denied; after switching the client grpA -> grpB, modelB is
-// served and modelA denied. The cross-group deny comes from management's
-// per-policy/group CheckLLMPolicyLimits (the proxy backstop carries the union).
-func TestGuardrailGroupSwitchTakesEffectAfterTTL(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
-	defer cancel()
-
-	const (
-		modelA = "e2e-model-a"
-		modelB = "e2e-model-b"
-		// Short tunnel-cache TTL so a group switch propagates in seconds.
-		// Exercises the NB_PROXY_TUNNEL_CACHE_TTL override.
-		cacheTTL = 3 * time.Second
-	)
-
-	vllm, err := harness.StartVLLM(ctx, srv)
-	require.NoError(t, err, "start mock upstream")
-	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
-
-	grpA, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-gswitch-a"})
-	require.NoError(t, err, "create group A")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpA.Id) })
-
-	grpB, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-gswitch-b"})
-	require.NoError(t, err, "create group B")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpB.Id) })
-
-	ephemeral := false
-	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
-		Name:       "e2e-gswitch-client",
-		Type:       "reusable",
-		ExpiresIn:  86400,
-		UsageLimit: 0,
-		AutoGroups: []string{grpA.Id}, // client starts in group A
-		Ephemeral:  &ephemeral,
-	})
-	require.NoError(t, err, "mint setup key")
-	require.NotEmpty(t, sk.Key, "setup key plaintext")
-
-	staticKey := "static-e2e-token"
-	prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:        "gswitch",
-		ProviderId:  "openai_api",
-		UpstreamUrl: vllm.URL,
-		ApiKey:      &staticKey,
-		Enabled:     ptr(true),
-		Models: &[]api.AgentNetworkProviderModel{
-			{Id: modelA, InputPer1k: 0.001, OutputPer1k: 0.001},
-			{Id: modelB, InputPer1k: 0.001, OutputPer1k: 0.001},
-		},
-		BootstrapCluster: ptr(harness.AgentNetworkCluster),
-	})
-	require.NoError(t, err, "create provider")
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
-
-	mkGuard := func(name, model string) api.AgentNetworkGuardrail {
-		var gr api.AgentNetworkGuardrailRequest
-		gr.Name = name
-		gr.Checks.ModelAllowlist.Enabled = true
-		gr.Checks.ModelAllowlist.Models = []string{model}
-		g, gerr := srv.CreateGuardrail(ctx, gr)
-		require.NoError(t, gerr, "create guardrail %s", name)
-		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
-		return g
-	}
-	gA := mkGuard("e2e-gswitch-a", modelA)
-	gB := mkGuard("e2e-gswitch-b", modelB)
-
-	enabled := true
-	polA, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-gswitch-a",
-		Enabled:                &enabled,
-		SourceGroups:           []string{grpA.Id},
-		DestinationProviderIds: []string{prov.Id},
-		GuardrailIds:           &[]string{gA.Id},
-	})
-	require.NoError(t, err, "create policy A")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polA.Id) })
-
-	polB, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-gswitch-b",
-		Enabled:                &enabled,
-		SourceGroups:           []string{grpB.Id},
-		DestinationProviderIds: []string{prov.Id},
-		GuardrailIds:           &[]string{gB.Id},
-	})
-	require.NoError(t, err, "create policy B")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polB.Id) })
-
-	settings, err := srv.GetSettings(ctx)
-	require.NoError(t, err, "read settings")
-	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
-
-	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-gswitch-proxy")
-	require.NoError(t, err, "mint proxy token")
-	px, err := harness.StartProxy(ctx, srv, proxyToken, map[string]string{
-		"NB_PROXY_TUNNEL_CACHE_TTL": cacheTTL.String(),
-	})
-	require.NoError(t, err, "start proxy")
-	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
-
-	cl, err := harness.StartClient(ctx, srv, sk.Key)
-	require.NoError(t, err, "start client")
-	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
-
-	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
-	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
-		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
-	}
-
-	send := func(model string) (int, string) {
-		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
-		require.NoError(t, cerr, "request must reach the proxy")
-		return code, body
-	}
-	sendUntil := func(model string, want int, timeout time.Duration) (int, string) {
-		var code int
-		var body string
-		deadline := time.Now().Add(timeout)
-		for time.Now().Before(deadline) {
-			code, body = send(model)
-			if code == want {
-				return code, body
-			}
-			time.Sleep(2 * time.Second)
-		}
-		return code, body
-	}
-
-	// Phase 1 — client is in group A: modelA served, modelB denied.
-	code, body := sendUntil(modelA, 200, 90*time.Second)
-	assert.Equal(t, 200, code,
-		"group-A model must be served while the client is in group A; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-	code, body = send(modelB)
-	assert.Equal(t, 403, code,
-		"group-B model must be denied while the client is in group A; body: %s", body)
-	assert.Contains(t, body, "llm_policy.model_blocked")
-
-	// Switch the client peer from group A to group B.
-	peerID := clientPeerInGroup(t, ctx, grpA.Id)
-	_, err = srv.API().Groups.Update(ctx, grpB.Id, api.PutApiGroupsGroupIdJSONRequestBody{
-		Name:  grpB.Name,
-		Peers: &[]string{peerID},
-	})
-	require.NoError(t, err, "add peer to group B")
-	_, err = srv.API().Groups.Update(ctx, grpA.Id, api.PutApiGroupsGroupIdJSONRequestBody{
-		Name:  grpA.Name,
-		Peers: &[]string{},
-	})
-	require.NoError(t, err, "remove peer from group A")
-
-	// Phase 2 — after the short TTL expires the proxy re-validates the peer,
-	// sees group B, and the decision flips. Poll to absorb TTL + re-validation.
-	code, body = sendUntil(modelB, 200, 60*time.Second)
-	assert.Equal(t, 200, code,
-		"after the group switch + TTL, the group-B model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-	code, body = send(modelA)
-	assert.Equal(t, 403, code,
-		"after the switch, the old group-A model must be denied; body: %s", body)
-	assert.Contains(t, body, "llm_policy.model_blocked")
-}
-
-// clientPeerInGroup returns the id of the single peer that is a member of the
-// given group — the test client. The proxy peer is never added to test groups.
-func clientPeerInGroup(t *testing.T, ctx context.Context, groupID string) string {
-	t.Helper()
-	peers, err := srv.API().Peers.List(ctx)
-	require.NoError(t, err, "list peers")
-	for _, p := range peers {
-		for _, g := range p.Groups {
-			if g.Id == groupID {
-				return p.Id
-			}
-		}
-	}
-	t.Fatalf("no peer found in group %s", groupID)
-	return ""
-}

e2e/agentnetwork/guardrail_multipolicy_test.go

@@ -1,201 +0,0 @@
-//go:build e2e
-
-package agentnetwork
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/e2e/harness"
-	"github.com/netbirdio/netbird/shared/management/http/api"
-)
-
-// TestGuardrailMultiPolicyModelAllowlist: modelSelected served (200), grpOther's
-// modelOther denied for the grpMain client (403 model_blocked, no cross-group
-// leak), and openModel on the un-guardrailed policy's provider served (200).
-func TestGuardrailMultiPolicyModelAllowlist(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
-	defer cancel()
-
-	const (
-		modelSelected = "e2e-selected"
-		modelOther    = "e2e-other"
-		openModel     = "e2e-open"
-	)
-
-	vllm, err := harness.StartVLLM(ctx, srv)
-	require.NoError(t, err, "start mock upstream")
-	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
-
-	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-main"})
-	require.NoError(t, err, "create main group")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpMain.Id) })
-
-	grpOther, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-other"})
-	require.NoError(t, err, "create other group")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
-
-	ephemeral := false
-	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
-		Name:       "e2e-guardrail-mp-client",
-		Type:       "reusable",
-		ExpiresIn:  86400,
-		UsageLimit: 0,
-		AutoGroups: []string{grpMain.Id}, // client joins grpMain only
-		Ephemeral:  &ephemeral,
-	})
-	require.NoError(t, err, "mint setup key")
-	require.NotEmpty(t, sk.Key, "setup key plaintext")
-
-	staticKey := "static-e2e-token"
-	models := func(ids ...string) *[]api.AgentNetworkProviderModel {
-		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
-		for _, id := range ids {
-			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
-		}
-		return &out
-	}
-
-	// pRestricted declares the two guardrailed models so routing is deterministic
-	// (model -> provider). Created first, so it carries the bootstrap cluster.
-	pRestricted, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:             "restricted",
-		ProviderId:       "openai_api",
-		UpstreamUrl:      vllm.URL,
-		ApiKey:           &staticKey,
-		Enabled:          ptr(true),
-		Models:           models(modelSelected, modelOther),
-		BootstrapCluster: ptr(harness.AgentNetworkCluster),
-	})
-	require.NoError(t, err, "create restricted provider")
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pRestricted.Id) })
-
-	pOpen, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:        "open",
-		ProviderId:  "openai_api",
-		UpstreamUrl: vllm.URL,
-		ApiKey:      &staticKey,
-		Enabled:     ptr(true),
-		Models:      models(openModel),
-	})
-	require.NoError(t, err, "create open provider")
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pOpen.Id) })
-
-	mkGuardrail := func(name, model string) api.AgentNetworkGuardrail {
-		var gr api.AgentNetworkGuardrailRequest
-		gr.Name = name
-		gr.Checks.ModelAllowlist.Enabled = true
-		gr.Checks.ModelAllowlist.Models = []string{model}
-		g, gerr := srv.CreateGuardrail(ctx, gr)
-		require.NoError(t, gerr, "create guardrail %s", name)
-		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
-		return g
-	}
-	gMain := mkGuardrail("e2e-guardrail-mp-main", modelSelected)
-	gOther := mkGuardrail("e2e-guardrail-mp-other", modelOther)
-
-	enabled := true
-	// polMain: grpMain restricted to modelSelected on pRestricted.
-	polMain, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-guardrail-mp-main",
-		Enabled:                &enabled,
-		SourceGroups:           []string{grpMain.Id},
-		DestinationProviderIds: []string{pRestricted.Id},
-		GuardrailIds:           &[]string{gMain.Id},
-	})
-	require.NoError(t, err, "create main policy")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
-
-	// polOther: grpOther restricted to modelOther on the SAME provider. The
-	// client is not in grpOther, so modelOther must never be usable by it.
-	polOther, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-guardrail-mp-other",
-		Enabled:                &enabled,
-		SourceGroups:           []string{grpOther.Id},
-		DestinationProviderIds: []string{pRestricted.Id},
-		GuardrailIds:           &[]string{gOther.Id},
-	})
-	require.NoError(t, err, "create other policy")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
-
-	// polOpen: grpMain on pOpen with NO guardrail — unrestricted.
-	polOpen, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-guardrail-mp-open",
-		Enabled:                &enabled,
-		SourceGroups:           []string{grpMain.Id},
-		DestinationProviderIds: []string{pOpen.Id},
-	})
-	require.NoError(t, err, "create open policy")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOpen.Id) })
-
-	settings, err := srv.GetSettings(ctx)
-	require.NoError(t, err, "read settings")
-	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
-
-	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-guardrail-mp-proxy")
-	require.NoError(t, err, "mint proxy token")
-	px, err := harness.StartProxy(ctx, srv, proxyToken)
-	require.NoError(t, err, "start proxy")
-	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
-
-	cl, err := harness.StartClient(ctx, srv, sk.Key)
-	require.NoError(t, err, "start client")
-	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
-
-	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
-	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
-		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
-	}
-
-	send := func(model string) (int, string) {
-		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
-		require.NoError(t, cerr, "request must reach the proxy")
-		return code, body
-	}
-	// sendUntil200 absorbs first-call tunnel/DNS jitter on the freshly warmed tunnel.
-	sendUntil200 := func(model string) (int, string) {
-		var code int
-		var body string
-		deadline := time.Now().Add(90 * time.Second)
-		for time.Now().Before(deadline) {
-			code, body = send(model)
-			if code == 200 {
-				break
-			}
-			time.Sleep(5 * time.Second)
-		}
-		return code, body
-	}
-
-	t.Run("selected model allowed for its group", func(t *testing.T) {
-		code, body := sendUntil200(modelSelected)
-		assert.Equal(t, 200, code,
-			"grpMain's allowlisted model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-	})
-
-	t.Run("other group's model does not leak", func(t *testing.T) {
-		// modelOther is allowlisted only for grpOther. The grpMain client must be
-		// denied by management's per-policy/group check — not waved through by an
-		// account-wide union. This is the security-critical wrong-ALLOW guard.
-		code, body := send(modelOther)
-		assert.Equal(t, 403, code,
-			"another group's allowlisted model must be denied for this caller; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-		assert.Contains(t, body, "llm_policy.model_blocked",
-			"denial must be a model-allowlist decision; body: %s", body)
-	})
-
-	t.Run("unguarded policy leaves its provider unrestricted", func(t *testing.T) {
-		// polOpen carries no guardrail, so pOpen is unrestricted for grpMain. The
-		// old account-wide union would have blocked openModel (it is on no
-		// allowlist); it must now be served — the false-DENY guard.
-		code, body := sendUntil200(openModel)
-		assert.Equal(t, 200, code,
-			"an un-guardrailed policy's provider must not be blocked by another policy's allowlist; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-	})
-}

e2e/agentnetwork/guardrail_pergroup_providers_test.go

@@ -1,422 +0,0 @@
-//go:build e2e
-
-package agentnetwork
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/e2e/harness"
-	"github.com/netbirdio/netbird/shared/management/http/api"
-)
-
-// pergroupCase describes one provider surface for the per-group allowlist matrix.
-// selectedReq/otherReq are the model identifiers as they travel in the request
-// (URL path for Bedrock/Vertex, body "model" for chat/messages). selectedAllow/
-// otherAllow are the (normalized) forms the guardrail allowlist holds — for
-// Bedrock these differ from the request form so path normalization is exercised.
-type pergroupCase struct {
-	name          string
-	catalogID     string
-	wire          string // "chat", "messages", "vertex", "bedrock"
-	models        *[]api.AgentNetworkProviderModel
-	selectedReq   string
-	selectedAllow string
-	otherReq      string
-	otherAllow    string
-
-	providerID string // filled during setup
-}
-
-// TestGuardrailPerGroupAllowlist_AllProviders proves the per-policy/group model
-// allowlist end to end across every always-on provider surface, including the
-// path-routed ones (Vertex, Bedrock) where the model travels in the URL.
-//
-// For each provider two policies target it: grpMain (the client) is allowed only
-// selectedReq; grpOther (which the client is NOT in) is allowed only otherReq.
-// The client must get selectedReq served (200) and otherReq denied (403,
-// llm_policy.model_blocked) — the cross-group no-leak property. The deny is the
-// authoritative per-policy/group decision from management (the proxy per-provider
-// backstop carries the union of both models), so this also confirms management
-// receives the correct normalized model for path-routed providers.
-func TestGuardrailPerGroupAllowlist_AllProviders(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
-	defer cancel()
-
-	const (
-		vertexProject = "e2e-project"
-		vertexRegion  = "global"
-	)
-
-	priced := func(ids ...string) *[]api.AgentNetworkProviderModel {
-		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
-		for _, id := range ids {
-			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
-		}
-		return &out
-	}
-
-	cases := []*pergroupCase{
-		{
-			name: "openai", catalogID: "openai_api", wire: harness.WireChat,
-			models:      priced("oai-model-a", "oai-model-b"),
-			selectedReq: "oai-model-a", selectedAllow: "oai-model-a",
-			otherReq: "oai-model-b", otherAllow: "oai-model-b",
-		},
-		{
-			name: "anthropic", catalogID: "anthropic_api", wire: harness.WireMessages,
-			models:      priced("ant-model-a", "ant-model-b"),
-			selectedReq: "ant-model-a", selectedAllow: "ant-model-a",
-			otherReq: "ant-model-b", otherAllow: "ant-model-b",
-		},
-		{
-			// Vertex catalog ids travel bare in the rawPredict path.
-			name: "vertex", catalogID: "vertex_ai_api", wire: "vertex",
-			selectedReq: "claude-sonnet-4-5", selectedAllow: "claude-sonnet-4-5",
-			otherReq: "claude-opus-4-6", otherAllow: "claude-opus-4-6",
-		},
-		{
-			// Bedrock request ids are region-prefixed/versioned; the parser
-			// normalizes them to the catalog key the allowlist holds.
-			name: "bedrock", catalogID: "bedrock_api", wire: "bedrock",
-			selectedReq: "us.anthropic.claude-sonnet-4-5-v1:0", selectedAllow: "anthropic.claude-sonnet-4-5",
-			otherReq: "us.anthropic.claude-opus-4-8-v1:0", otherAllow: "anthropic.claude-opus-4-8",
-		},
-	}
-
-	vllm, err := harness.StartVLLM(ctx, srv)
-	require.NoError(t, err, "start mock upstream")
-	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
-
-	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-pergroup-main"})
-	require.NoError(t, err, "create main group")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpMain.Id) })
-
-	grpOther, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-pergroup-other"})
-	require.NoError(t, err, "create other group")
-	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
-
-	ephemeral := false
-	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
-		Name:       "e2e-pergroup-client",
-		Type:       "reusable",
-		ExpiresIn:  86400,
-		UsageLimit: 0,
-		AutoGroups: []string{grpMain.Id},
-		Ephemeral:  &ephemeral,
-	})
-	require.NoError(t, err, "mint setup key")
-	require.NotEmpty(t, sk.Key, "setup key plaintext")
-
-	staticKey := "static-e2e-token"
-	enabled := true
-
-	for i, c := range cases {
-		req := api.AgentNetworkProviderRequest{
-			Name:        "e2e-pergroup-" + c.name,
-			ProviderId:  c.catalogID,
-			UpstreamUrl: vllm.URL,
-			ApiKey:      &staticKey,
-			Enabled:     ptr(true),
-			Models:      c.models,
-		}
-		if i == 0 {
-			req.BootstrapCluster = ptr(harness.AgentNetworkCluster)
-		}
-		prov, perr := srv.CreateProvider(ctx, req)
-		require.NoError(t, perr, "create provider %s", c.name)
-		c.providerID = prov.Id
-		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
-
-		gSel := mkAllowGuardrail(t, ctx, "e2e-pergroup-"+c.name+"-sel", c.selectedAllow)
-		gOth := mkAllowGuardrail(t, ctx, "e2e-pergroup-"+c.name+"-oth", c.otherAllow)
-
-		polMain, merr := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-			Name:                   "e2e-pergroup-" + c.name + "-main",
-			Enabled:                &enabled,
-			SourceGroups:           []string{grpMain.Id},
-			DestinationProviderIds: []string{prov.Id},
-			GuardrailIds:           &[]string{gSel.Id},
-		})
-		require.NoError(t, merr, "create main policy %s", c.name)
-		t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
-
-		polOther, oerr := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-			Name:                   "e2e-pergroup-" + c.name + "-other",
-			Enabled:                &enabled,
-			SourceGroups:           []string{grpOther.Id},
-			DestinationProviderIds: []string{prov.Id},
-			GuardrailIds:           &[]string{gOth.Id},
-		})
-		require.NoError(t, oerr, "create other policy %s", c.name)
-		t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
-	}
-
-	settings, err := srv.GetSettings(ctx)
-	require.NoError(t, err, "read settings")
-	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
-
-	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-pergroup-proxy")
-	require.NoError(t, err, "mint proxy token")
-	px, err := harness.StartProxy(ctx, srv, proxyToken)
-	require.NoError(t, err, "start proxy")
-	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
-
-	cl, err := harness.StartClient(ctx, srv, sk.Key)
-	require.NoError(t, err, "start client")
-	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
-
-	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
-	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
-		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
-	}
-
-	send := func(c *pergroupCase, model string) (int, string) {
-		var code int
-		var body string
-		var cerr error
-		switch c.wire {
-		case "vertex":
-			code, body, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, vertexProject, vertexRegion, model, "Reply with exactly: pong", "")
-		case "bedrock":
-			code, body, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, model, "Reply with exactly: pong", "")
-		default:
-			code, body, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, c.wire, model, "Reply with exactly: pong", "")
-		}
-		require.NoError(t, cerr, "request must reach the proxy for %s", c.name)
-		return code, body
-	}
-
-	for _, c := range cases {
-		t.Run(c.name, func(t *testing.T) {
-			// grpMain's own model is served. Retry to absorb tunnel/DNS jitter on
-			// the first call over the freshly warmed tunnel.
-			var code int
-			var body string
-			deadline := time.Now().Add(90 * time.Second)
-			for time.Now().Before(deadline) {
-				code, body = send(c, c.selectedReq)
-				if code == 200 {
-					break
-				}
-				time.Sleep(5 * time.Second)
-			}
-			assert.Equal(t, 200, code,
-				"%s: grpMain's allowlisted model must be served; body: %s\n=== proxy logs ===\n%s", c.name, body, px.Logs(context.Background()))
-
-			// grpOther's model must NOT leak to the grpMain client.
-			code, body = send(c, c.otherReq)
-			assert.Equal(t, 403, code,
-				"%s: another group's allowlisted model must be denied for this caller; body: %s\n=== proxy logs ===\n%s", c.name, body, px.Logs(context.Background()))
-			assert.Contains(t, body, "llm_policy.model_blocked",
-				"%s: denial must be a model-allowlist decision, not routing; body: %s", c.name, body)
-		})
-	}
-}
-
-// TestGuardrailMultiGroupUser proves the per-policy/group decision for a caller
-// that belongs to MULTIPLE groups at once. Two scenarios, one shared stack:
-//
-//   - union across the user's groups: the client is in gUX and gUY, each with
-//     its own policy+guardrail on provider P1 (gUX->union-a, gUY->union-b). The
-//     client may use BOTH models (the union of its groups' allowlists) while a
-//     third, un-allowlisted model is denied.
-//   - an un-guardrailed group lifts the restriction: the client is in gMP and
-//     gMQ on provider P2, where gMP restricts to mix-a but gMQ's policy carries
-//     NO guardrail. Because one applicable policy is unrestricted, the client may
-//     use a model on no allowlist (mix-z) as well as mix-a.
-func TestGuardrailMultiGroupUser(t *testing.T) {
-	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
-	defer cancel()
-
-	const (
-		unionA = "mg-union-a"
-		unionB = "mg-union-b"
-		unionC = "mg-union-c" // allowlisted by neither group
-		mixA   = "mg-mix-a"
-		mixZ   = "mg-mix-z" // on no allowlist; reachable only via the un-guardrailed policy
-	)
-
-	priced := func(ids ...string) *[]api.AgentNetworkProviderModel {
-		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
-		for _, id := range ids {
-			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
-		}
-		return &out
-	}
-
-	vllm, err := harness.StartVLLM(ctx, srv)
-	require.NoError(t, err, "start mock upstream")
-	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
-
-	mkGroup := func(name string) *api.Group {
-		g, gerr := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: name})
-		require.NoError(t, gerr, "create group %s", name)
-		t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), g.Id) })
-		return g
-	}
-	gUX := mkGroup("e2e-mg-union-x")
-	gUY := mkGroup("e2e-mg-union-y")
-	gMP := mkGroup("e2e-mg-mix-p")
-	gMQ := mkGroup("e2e-mg-mix-q")
-
-	ephemeral := false
-	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
-		Name:       "e2e-mg-client",
-		Type:       "reusable",
-		ExpiresIn:  86400,
-		UsageLimit: 0,
-		AutoGroups: []string{gUX.Id, gUY.Id, gMP.Id, gMQ.Id}, // client in all four groups
-		Ephemeral:  &ephemeral,
-	})
-	require.NoError(t, err, "mint setup key")
-	require.NotEmpty(t, sk.Key, "setup key plaintext")
-
-	staticKey := "static-e2e-token"
-	enabled := true
-
-	// P1 — union scenario: two restricting policies, one per group.
-	p1, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:             "e2e-mg-union",
-		ProviderId:       "openai_api",
-		UpstreamUrl:      vllm.URL,
-		ApiKey:           &staticKey,
-		Enabled:          ptr(true),
-		Models:           priced(unionA, unionB, unionC),
-		BootstrapCluster: ptr(harness.AgentNetworkCluster),
-	})
-	require.NoError(t, err, "create union provider")
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), p1.Id) })
-
-	polUX, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-mg-union-x",
-		Enabled:                &enabled,
-		SourceGroups:           []string{gUX.Id},
-		DestinationProviderIds: []string{p1.Id},
-		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-union-x", unionA).Id},
-	})
-	require.NoError(t, err, "create union policy X")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polUX.Id) })
-
-	polUY, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-mg-union-y",
-		Enabled:                &enabled,
-		SourceGroups:           []string{gUY.Id},
-		DestinationProviderIds: []string{p1.Id},
-		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-union-y", unionB).Id},
-	})
-	require.NoError(t, err, "create union policy Y")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polUY.Id) })
-
-	// P2 — mixed scenario: one restricting policy + one un-guardrailed policy.
-	p2, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
-		Name:        "e2e-mg-mix",
-		ProviderId:  "openai_api",
-		UpstreamUrl: vllm.URL,
-		ApiKey:      &staticKey,
-		Enabled:     ptr(true),
-		Models:      priced(mixA, mixZ),
-	})
-	require.NoError(t, err, "create mix provider")
-	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), p2.Id) })
-
-	polMP, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-mg-mix-p",
-		Enabled:                &enabled,
-		SourceGroups:           []string{gMP.Id},
-		DestinationProviderIds: []string{p2.Id},
-		GuardrailIds:           &[]string{mkAllowGuardrail(t, ctx, "e2e-mg-mix-p", mixA).Id},
-	})
-	require.NoError(t, err, "create mix policy P")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMP.Id) })
-
-	polMQ, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
-		Name:                   "e2e-mg-mix-q",
-		Enabled:                &enabled,
-		SourceGroups:           []string{gMQ.Id},
-		DestinationProviderIds: []string{p2.Id}, // NO guardrail -> unrestricted
-	})
-	require.NoError(t, err, "create mix policy Q")
-	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMQ.Id) })
-
-	settings, err := srv.GetSettings(ctx)
-	require.NoError(t, err, "read settings")
-	require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
-
-	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-mg-proxy")
-	require.NoError(t, err, "mint proxy token")
-	px, err := harness.StartProxy(ctx, srv, proxyToken)
-	require.NoError(t, err, "start proxy")
-	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
-
-	cl, err := harness.StartClient(ctx, srv, sk.Key)
-	require.NoError(t, err, "start client")
-	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
-
-	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
-	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
-	require.NoError(t, err, "resolve endpoint to proxy IP")
-	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
-		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
-	}
-
-	send := func(model string) (int, string) {
-		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
-		require.NoError(t, cerr, "request must reach the proxy")
-		return code, body
-	}
-	sendUntil200 := func(model string) (int, string) {
-		var code int
-		var body string
-		deadline := time.Now().Add(90 * time.Second)
-		for time.Now().Before(deadline) {
-			code, body = send(model)
-			if code == 200 {
-				break
-			}
-			time.Sleep(5 * time.Second)
-		}
-		return code, body
-	}
-
-	t.Run("union across the user's groups", func(t *testing.T) {
-		code, body := sendUntil200(unionA)
-		assert.Equal(t, 200, code, "model allowed by group X must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-
-		code, body = sendUntil200(unionB)
-		assert.Equal(t, 200, code, "model allowed by group Y must also be served (union across the user's groups); body: %s", body)
-
-		code, body = send(unionC)
-		assert.Equal(t, 403, code, "a model on neither group's allowlist must be denied; body: %s", body)
-		assert.Contains(t, body, "llm_policy.model_blocked")
-	})
-
-	t.Run("an un-guardrailed group lifts the restriction", func(t *testing.T) {
-		code, body := sendUntil200(mixA)
-		assert.Equal(t, 200, code, "the restricted group's model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-
-		code, body = sendUntil200(mixZ)
-		assert.Equal(t, 200, code,
-			"a non-allowlisted model must be served because the user is also in a group whose policy has no guardrail; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
-	})
-}
-
-// mkAllowGuardrail creates a guardrail whose model allowlist is enabled and holds
-// exactly the given model, registering cleanup.
-func mkAllowGuardrail(t *testing.T, ctx context.Context, name, model string) api.AgentNetworkGuardrail {
-	t.Helper()
-	var gr api.AgentNetworkGuardrailRequest
-	gr.Name = name
-	gr.Checks.ModelAllowlist.Enabled = true
-	gr.Checks.ModelAllowlist.Models = []string{model}
-	g, err := srv.CreateGuardrail(ctx, gr)
-	require.NoError(t, err, "create guardrail %s", name)
-	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
-	return g
-}

e2e/harness/client.go

@@ -256,7 +256,7 @@ func (cl *Client) ChatPrefixed(ctx context.Context, endpoint, proxyIP, pathPrefi
 	case WireMessages:
 		path = "/v1/messages"
 		headers = []string{"anthropic-version: 2023-06-01"}
-		body = fmt.Sprintf(`{"model":%q,"max_tokens":64,"messages":[{"role":"user","content":%q}]}`, model, prompt)
+		body = fmt.Sprintf(`{"model":%q,"max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, model, prompt)
 	default:
 		path = "/v1/chat/completions"
 		body = fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":%q}]}`, model, prompt)
@@ -271,7 +271,7 @@ func (cl *Client) ChatPrefixed(ctx context.Context, endpoint, proxyIP, pathPrefi
 // is sent as the universal x-session-id header the proxy records.
 func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region, model, prompt, sessionID string) (int, string, error) {
 	path := fmt.Sprintf("/v1/projects/%s/locations/%s/publishers/anthropic/models/%s:rawPredict", project, region, model)
-	body := fmt.Sprintf(`{"anthropic_version":"vertex-2023-10-16","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	body := fmt.Sprintf(`{"anthropic_version":"vertex-2023-10-16","max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, prompt)
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
@@ -282,7 +282,7 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 // header the proxy records.
 func (cl *Client) Bedrock(ctx context.Context, endpoint, proxyIP, model, prompt, sessionID string) (int, string, error) {
 	path := "/model/" + model + "/invoke"
-	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":2048,"messages":[{"role":"user","content":%q}]}`, prompt)
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
@@ -341,7 +341,7 @@ func (cl *Client) Terminate(ctx context.Context) error {
 	return cl.container.Terminate(ctx)
 }
 
-// containerLogs reads up to 256 KiB of a container's logs for diagnostics.
+// containerLogs reads up to 4 MiB of a container's logs for diagnostics — enough for a whole provider-matrix run.
 func containerLogs(ctx context.Context, c testcontainers.Container) string {
 	if c == nil {
 		return ""
@@ -351,6 +351,6 @@ func containerLogs(ctx context.Context, c testcontainers.Container) string {
 		return fmt.Sprintf("<logs error: %v>", err)
 	}
 	defer r.Close()
-	b, _ := io.ReadAll(io.LimitReader(r, 256<<10))
+	b, _ := io.ReadAll(io.LimitReader(r, 4<<20))
 	return string(b)
 }

e2e/harness/combined.go

@@ -221,6 +221,29 @@ func (c *Combined) CreateProxyTokenCLI(ctx context.Context, name string) (string
 	return "", fmt.Errorf("token not found in CLI output: %s", string(out))
 }
 
+// SnapshotStoreDB copies the management sqlite store (with WAL/SHM sidecars) out of the bind-mounted
+// data dir into dstDir and returns the copy's path; reading a copy avoids locking against live writes.
+func (c *Combined) SnapshotStoreDB(dstDir string) (string, error) {
+	src := filepath.Join(c.workDir, "data", "store.db")
+	if _, err := os.Stat(src); err != nil {
+		return "", fmt.Errorf("management store not found at %s: %w", src, err)
+	}
+	dst := filepath.Join(dstDir, "store.db")
+	for _, suffix := range []string{"", "-wal", "-shm"} {
+		data, err := os.ReadFile(src + suffix)
+		if err != nil {
+			if os.IsNotExist(err) && suffix != "" {
+				continue // sidecar only exists in WAL mode
+			}
+			return "", fmt.Errorf("read %s: %w", src+suffix, err)
+		}
+		if err := os.WriteFile(dst+suffix, data, 0o600); err != nil {
+			return "", fmt.Errorf("write %s: %w", dst+suffix, err)
+		}
+	}
+	return dst, nil
+}
+
 // Logs returns the combined server container logs, for diagnostics.
 func (c *Combined) Logs(ctx context.Context) string {
 	return containerLogs(ctx, c.container)

e2e/harness/proxy.go

@@ -38,11 +38,7 @@ type Proxy struct {
 // network, registered via the given account proxy token and serving the
 // AgentNetworkCluster over a self-signed wildcard cert. It does not wait for
 // peer connectivity — callers poll management for the proxy peer.
-// StartProxy launches the reverse-proxy container. Optional envOverrides are
-// merged into the container environment after the defaults, so callers can set
-// or override any NB_PROXY_* var (e.g. NB_PROXY_TUNNEL_CACHE_TTL for tests that
-// need a short authorization-cache window).
-func StartProxy(ctx context.Context, c *Combined, proxyToken string, envOverrides ...map[string]string) (*Proxy, error) {
+func StartProxy(ctx context.Context, c *Combined, proxyToken string) (*Proxy, error) {
 	root, err := repoRoot()
 	if err != nil {
 		return nil, err
@@ -97,12 +93,6 @@ func StartProxy(ctx context.Context, c *Combined, proxyToken string, envOverride
 		WaitingFor: wait.ForLog("Initial mapping sync complete").WithStartupTimeout(90 * time.Second),
 	}
 
-	for _, ov := range envOverrides {
-		for k, v := range ov {
-			req.Env[k] = v
-		}
-	}
-
 	ctr, err := testcontainers.GenericContainer(ctx, testcontainers.GenericContainerRequest{
 		ContainerRequest: req,
 		Started:          true,

management/internals/modules/agentnetwork/accesslog_ingest.go

@@ -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))

management/internals/modules/agentnetwork/accesslog_ingest_realstore_test.go

@@ -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")

management/internals/modules/agentnetwork/accesslog_sessions_realstore_test.go

@@ -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")

management/internals/modules/agentnetwork/manager.go

@@ -86,17 +86,12 @@ 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 and the model it extracted.
+// the provider llm_router resolved.
 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

management/internals/modules/agentnetwork/policyselect.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"math"
 	"sort"
-	"strings"
 	"time"
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
@@ -36,10 +35,6 @@ 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
@@ -164,25 +159,6 @@ 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)
@@ -274,90 +250,6 @@ 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

management/internals/modules/agentnetwork/policyselect_model_test.go

@@ -1,329 +0,0 @@
-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")
-}

management/internals/modules/agentnetwork/synthesizer.go

@@ -235,12 +235,7 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
-	// 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)
+	guardrailJSON, err := marshalGuardrailConfig(mergedGuardrails)
 	if err != nil {
 		return nil, err
 	}
@@ -785,12 +780,10 @@ 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. provider_allowlists is keyed by the
-// resolved provider id llm_router stamps; a provider absent from the map is
-// unrestricted at the proxy layer.
+// the management→proxy contract.
 type guardrailConfig struct {
-	ProviderAllowlists map[string][]string    `json:"provider_allowlists,omitempty"`
-	PromptCapture      guardrailPromptCapture `json:"prompt_capture"`
+	ModelAllowlist []string               `json:"model_allowlist,omitempty"`
+	PromptCapture  guardrailPromptCapture `json:"prompt_capture"`
 }
 
 type guardrailPromptCapture struct {
@@ -835,10 +828,13 @@ func applyAccountCollectionControls(merged *MergedGuardrails, settings *types.Se
 	merged.PromptCapture.RedactPii = settings.RedactPii || merged.PromptCapture.RedactPii
 }
 
-func marshalGuardrailConfig(providerAllowlists map[string][]string, capture MergedPromptCapture) ([]byte, error) {
+func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
 	cfg := guardrailConfig{
-		ProviderAllowlists: providerAllowlists,
-		PromptCapture:      guardrailPromptCapture(capture),
+		ModelAllowlist: merged.ModelAllowlist,
+		PromptCapture: guardrailPromptCapture{
+			Enabled:   merged.PromptCapture.Enabled,
+			RedactPii: merged.PromptCapture.RedactPii,
+		},
 	}
 	out, err := json.Marshal(cfg)
 	if err != nil {
@@ -847,74 +843,6 @@ func marshalGuardrailConfig(providerAllowlists map[string][]string, capture Merg
 	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
@@ -1058,11 +986,38 @@ func unionSourceGroups(policies []*types.Policy) []string {
 	return out
 }
 
-// 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.
+// 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.
 type MergedGuardrails struct {
-	PromptCapture MergedPromptCapture
+	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"`
 }
 
 type MergedPromptCapture struct {
@@ -1070,31 +1025,64 @@ type MergedPromptCapture struct {
 	RedactPii bool `json:"redact_pii"`
 }
 
-// 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.
+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.
 //
-// Merge rule — prompt capture: enabled if any policy enables it; redact_pii
-// sticks if any enabling policy turns it on.
+// 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.
 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)
+			mergeGuardrail(g, &merged, allowlist, &allowlistEnabled)
 		}
 	}
+
+	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 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) {
+// 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{}{}
+			}
+		}
+	}
 	if g.Checks.PromptCapture.Enabled {
 		merged.PromptCapture.Enabled = true
 		if g.Checks.PromptCapture.RedactPii {

management/internals/modules/agentnetwork/synthesizer_guardrail_realstore_test.go

@@ -81,8 +81,8 @@ func TestSynthesizeServices_RealStore_PromptCaptureAccountIsSoleControl(t *testi
 	require.Len(t, services, 1)
 
 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	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.Equal(t, []string{"gpt-5.4"}, cfg.ModelAllowlist,
+		"model allowlist is a pure policy guardrail and must always reach the 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.ProviderAllowlists, "no guardrail → provider unrestricted (absent from map)")
+	assert.Empty(t, cfg.ModelAllowlist, "no guardrail → no allowlist")
 	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")
 }

management/internals/modules/agentnetwork/synthesizer_provider_allowlist_test.go

@@ -1,95 +0,0 @@
-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")
-	})
-}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -1031,12 +1031,8 @@ func TestSynthesizeServices_GuardrailMerge_AllowlistUnion_LimitsRestrictive(t *t
 
 	var cfg guardrailConfig
 	require.NoError(t, json.Unmarshal(guardrailJSON, &cfg), "guardrail config must unmarshal cleanly")
-	// 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")
+	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ModelAllowlist,
+		"model allowlist union must keep both models")
 }
 
 func TestSynthesizeServices_BackfillsMissingSessionKeys(t *testing.T) {

management/internals/modules/agentnetwork/types/accesslog.go

@@ -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)

management/internals/modules/agentnetwork/types/accesslogfilter.go

@@ -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(*)",

management/internals/modules/agentnetwork/types/cost_test.go

@@ -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")
+}

management/internals/modules/agentnetwork/types/usage.go

@@ -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.

management/internals/modules/agentnetwork/types/usageoverview.go

@@ -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))

management/internals/shared/grpc/proxy.go

@@ -285,7 +285,6 @@ func (s *ProxyServiceServer) CheckLLMPolicyLimits(ctx context.Context, req *prot
 		UserID:     req.GetUserId(),
 		GroupIDs:   req.GetGroupIds(),
 		ProviderID: req.GetProviderId(),
-		Model:      req.GetModel(),
 	})
 	if err != nil {
 		log.WithContext(ctx).Errorf("select policy for request: %v", err)

management/internals/shared/grpc/proxy_llm_policy_limits_test.go

@@ -1,138 +0,0 @@
-package grpc
-
-import (
-	"context"
-	"errors"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-	"google.golang.org/grpc/codes"
-	grpcstatus "google.golang.org/grpc/status"
-
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
-	"github.com/netbirdio/netbird/shared/management/proto"
-)
-
-// fakeAgentNetworkLimits records the PolicySelectionInput it was invoked with
-// and returns a pre-programmed result, so tests can assert what the handler
-// forwards to the selector.
-type fakeAgentNetworkLimits struct {
-	gotInput agentnetwork.PolicySelectionInput
-	result   *agentnetwork.PolicySelectionResult
-	err      error
-}
-
-func (f *fakeAgentNetworkLimits) SelectPolicyForRequest(_ context.Context, in agentnetwork.PolicySelectionInput) (*agentnetwork.PolicySelectionResult, error) {
-	f.gotInput = in
-	if f.err != nil {
-		return nil, f.err
-	}
-	return f.result, nil
-}
-
-func (f *fakeAgentNetworkLimits) RecordUsage(_ context.Context, _ agentnetwork.RecordUsageInput) error {
-	return nil
-}
-
-// TestCheckLLMPolicyLimits_ForwardsModelToSelector proves the wiring added here:
-// the model the proxy extracted must reach the selector's Model unchanged,
-// alongside the account/user/group/provider fields.
-func TestCheckLLMPolicyLimits_ForwardsModelToSelector(t *testing.T) {
-	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true, SelectedPolicyID: "pol-1"}}
-	s := &ProxyServiceServer{}
-	s.SetAgentNetworkLimitsService(fake)
-
-	req := &proto.CheckLLMPolicyLimitsRequest{
-		AccountId:  "acc-1",
-		UserId:     "user-1",
-		GroupIds:   []string{"grp-a", "grp-b"},
-		ProviderId: "prov-1",
-		Model:      "claude-opus-4",
-	}
-
-	resp, err := s.CheckLLMPolicyLimits(context.Background(), req)
-	require.NoError(t, err)
-	require.NotNil(t, resp)
-
-	assert.Equal(t, "acc-1", fake.gotInput.AccountID)
-	assert.Equal(t, "user-1", fake.gotInput.UserID)
-	assert.Equal(t, []string{"grp-a", "grp-b"}, fake.gotInput.GroupIDs)
-	assert.Equal(t, "prov-1", fake.gotInput.ProviderID)
-	assert.Equal(t, "claude-opus-4", fake.gotInput.Model,
-		"the request's model must be forwarded to the selector")
-}
-
-// TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty proves an undetermined
-// model (empty string) is forwarded as-is; the selector decides how to treat it.
-func TestCheckLLMPolicyLimits_EmptyModelForwardedAsEmpty(t *testing.T) {
-	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{Allow: true}}
-	s := &ProxyServiceServer{}
-	s.SetAgentNetworkLimitsService(fake)
-
-	req := &proto.CheckLLMPolicyLimitsRequest{
-		AccountId:  "acc-1",
-		ProviderId: "prov-1",
-	}
-
-	_, err := s.CheckLLMPolicyLimits(context.Background(), req)
-	require.NoError(t, err)
-	assert.Equal(t, "", fake.gotInput.Model, "an absent model must be forwarded as empty, not fabricated")
-}
-
-// TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode proves the deny
-// envelope surfaces the model-allowlist deny code + reason through the response.
-func TestCheckLLMPolicyLimits_DenyResponseCarriesModelBlockedCode(t *testing.T) {
-	fake := &fakeAgentNetworkLimits{result: &agentnetwork.PolicySelectionResult{
-		Allow:      false,
-		DenyCode:   "llm_policy.model_blocked",
-		DenyReason: `model "claude-opus-4" is not permitted by any applicable policy allowlist`,
-	}}
-	s := &ProxyServiceServer{}
-	s.SetAgentNetworkLimitsService(fake)
-
-	resp, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
-		AccountId:  "acc-1",
-		ProviderId: "prov-1",
-		Model:      "claude-opus-4",
-	})
-	require.NoError(t, err)
-	require.NotNil(t, resp)
-	assert.Equal(t, "deny", resp.Decision)
-	assert.Equal(t, "llm_policy.model_blocked", resp.DenyCode)
-	assert.NotEmpty(t, resp.DenyReason)
-	assert.Empty(t, resp.SelectedPolicyId, "a denied request must carry no selected policy")
-}
-
-// TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal proves a selector
-// failure surfaces as an Internal gRPC error rather than a silent allow.
-func TestCheckLLMPolicyLimits_SelectorErrorSurfacesAsInternal(t *testing.T) {
-	fake := &fakeAgentNetworkLimits{err: errors.New("boom")}
-	s := &ProxyServiceServer{}
-	s.SetAgentNetworkLimitsService(fake)
-
-	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
-		AccountId:  "acc-1",
-		ProviderId: "prov-1",
-		Model:      "gpt-4o",
-	})
-	require.Error(t, err)
-	st, ok := grpcstatus.FromError(err)
-	require.True(t, ok)
-	assert.Equal(t, codes.Internal, st.Code(), "selector errors must never fail open on the hot path")
-}
-
-// TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented locks the
-// fallback: with no limits service wired the RPC returns Unimplemented.
-func TestCheckLLMPolicyLimits_UnconfiguredServiceReturnsUnimplemented(t *testing.T) {
-	s := &ProxyServiceServer{}
-
-	_, err := s.CheckLLMPolicyLimits(context.Background(), &proto.CheckLLMPolicyLimitsRequest{
-		AccountId:  "acc-1",
-		ProviderId: "prov-1",
-	})
-	require.Error(t, err)
-	st, ok := grpcstatus.FromError(err)
-	require.True(t, ok)
-	assert.Equal(t, codes.Unimplemented, st.Code())
-}

management/server/migration/migration.go

@@ -683,3 +683,81 @@ func BackfillPublicIDs[T any](ctx context.Context, db *gorm.DB) error {
 	log.WithContext(ctx).Infof("Backfill of empty public_id in table %s completed", tableName)
 	return nil
 }
+
+// FoldCostAggregatesIntoBuckets migrates a per-request cost table from the old
+// "stored aggregate" shape (cost_usd + cache_cost_usd columns) to the per-bucket
+// breakdown, where the total and cache portion are derived on read instead.
+//
+// The fold preserves both aggregates exactly for historical rows: the cache
+// total moves into cached_input_cost_usd and the remainder into
+// input_cost_usd, so a row's derived total and cache cost still match what it
+// reported before the upgrade. The finer split is genuinely unknown for those
+// rows — the old schema never recorded a read/write or input/output division —
+// so it is lumped rather than guessed; only rows written after the upgrade
+// carry a true four-way split.
+//
+// Dropping the columns before folding would zero every historical row's cost,
+// so the update runs first and the drop only happens once it succeeds. A table
+// with no cost_usd column has already been migrated (or was created fresh) and
+// is skipped.
+func FoldCostAggregatesIntoBuckets[T any](ctx context.Context, db *gorm.DB) error {
+	var model T
+
+	if !db.Migrator().HasTable(&model) {
+		log.WithContext(ctx).Debugf("table for %T does not exist, no cost-bucket migration needed", model)
+		return nil
+	}
+	if !db.Migrator().HasColumn(&model, "cost_usd") {
+		log.WithContext(ctx).Debugf("table for %T has no cost_usd column, cost buckets already migrated", model)
+		return nil
+	}
+
+	stmt := &gorm.Statement{DB: db}
+	if err := stmt.Parse(&model); err != nil {
+		return fmt.Errorf("parse model schema: %w", err)
+	}
+	tableName := stmt.Schema.Table
+
+	// COALESCE guards rows whose new columns were added as NULL by an earlier
+	// AutoMigrate run that predates the NOT NULL default.
+	hasCacheColumn := db.Migrator().HasColumn(&model, "cache_cost_usd")
+	cacheExpr := "0"
+	if hasCacheColumn {
+		cacheExpr = "COALESCE(cache_cost_usd, 0)"
+	}
+
+	if err := db.Transaction(func(tx *gorm.DB) error {
+		// Only touch rows that carry a legacy total and no breakdown yet, so
+		// the migration is idempotent and never overwrites a true split.
+		update := fmt.Sprintf(`UPDATE %s
+			SET input_cost_usd = COALESCE(cost_usd, 0) - %s,
+			    cached_input_cost_usd = %s,
+			    cache_creation_cost_usd = 0,
+			    output_cost_usd = 0
+			WHERE COALESCE(cost_usd, 0) <> 0
+			  AND COALESCE(input_cost_usd, 0) = 0
+			  AND COALESCE(cached_input_cost_usd, 0) = 0
+			  AND COALESCE(cache_creation_cost_usd, 0) = 0
+			  AND COALESCE(output_cost_usd, 0) = 0`, tableName, cacheExpr, cacheExpr)
+		res := tx.Exec(update)
+		if res.Error != nil {
+			return fmt.Errorf("fold legacy cost aggregates in %s: %w", tableName, res.Error)
+		}
+		log.WithContext(ctx).Infof("folded legacy cost aggregates into per-bucket columns for %d rows in table %s", res.RowsAffected, tableName)
+
+		if err := tx.Migrator().DropColumn(&model, "cost_usd"); err != nil {
+			return fmt.Errorf("drop cost_usd from %s: %w", tableName, err)
+		}
+		if hasCacheColumn {
+			if err := tx.Migrator().DropColumn(&model, "cache_cost_usd"); err != nil {
+				return fmt.Errorf("drop cache_cost_usd from %s: %w", tableName, err)
+			}
+		}
+		return nil
+	}); err != nil {
+		return err
+	}
+
+	log.WithContext(ctx).Infof("migration of stored cost aggregates to per-bucket columns in table %s completed", tableName)
+	return nil
+}

management/server/migration/migration_test.go

@@ -16,6 +16,7 @@ import (
 	"gorm.io/driver/sqlite"
 	"gorm.io/gorm"
 
+	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
 	"github.com/netbirdio/netbird/management/server/migration"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/testutil"
@@ -639,3 +640,99 @@ func TestCleanupOrphanedResources_SkipsWhenForeignKeyExists(t *testing.T) {
 	db.Model(&testChildWithFK{}).Count(&count)
 	assert.Equal(t, int64(2), count, "Both rows should survive — migration must skip when FK constraint exists")
 }
+
+// legacyCostRow is the pre-breakdown shape of the usage table: cost was stored
+// as a total plus a cache portion, with no per-bucket columns. Used to build a
+// realistic pre-upgrade table for the fold migration to run against.
+type legacyCostRow struct {
+	ID           string `gorm:"primaryKey"`
+	AccountID    string
+	Model        string
+	CostUSD      float64
+	CacheCostUSD float64
+}
+
+func (legacyCostRow) TableName() string { return "agent_network_request_usage" }
+
+// TestFoldCostAggregatesIntoBuckets_PreservesHistoricalCost covers the upgrade
+// path: a table written under the old schema must come out with its per-row
+// total and cache cost unchanged, because dropping cost_usd without folding it
+// forward would silently zero every historical row's spend.
+func TestFoldCostAggregatesIntoBuckets_PreservesHistoricalCost(t *testing.T) {
+	ctx := context.Background()
+	db := setupDatabase(t)
+	// setupDatabase hands back a process-shared database, so start from a clean
+	// table rather than inheriting rows from another test.
+	require.NoError(t, db.Migrator().DropTable(&agentNetworkTypes.AgentNetworkUsage{}))
+
+	require.NoError(t, db.AutoMigrate(&legacyCostRow{}), "legacy table must be created")
+	require.NoError(t, db.Create(&legacyCostRow{
+		ID: "u1", AccountID: "acct-1", Model: "claude-sonnet-4-6", CostUSD: 0.0123, CacheCostUSD: 0.0029,
+	}).Error)
+	require.NoError(t, db.Create(&legacyCostRow{
+		ID: "u2", AccountID: "acct-1", Model: "gpt-4o", CostUSD: 0.5, CacheCostUSD: 0,
+	}).Error)
+	// A zero-cost row (denied / unpriced request) must stay zero, not be touched.
+	require.NoError(t, db.Create(&legacyCostRow{ID: "u3", AccountID: "acct-1", Model: "gw/unpriced"}).Error)
+
+	// AutoMigrate adds the per-bucket columns alongside the legacy ones, exactly
+	// as a real upgrade does before the post-auto migrations run.
+	require.NoError(t, db.AutoMigrate(&agentNetworkTypes.AgentNetworkUsage{}), "new columns must be added")
+
+	require.NoError(t, migration.FoldCostAggregatesIntoBuckets[agentNetworkTypes.AgentNetworkUsage](ctx, db))
+
+	assert.False(t, db.Migrator().HasColumn(&agentNetworkTypes.AgentNetworkUsage{}, "cost_usd"),
+		"legacy cost_usd column must be dropped once folded")
+	assert.False(t, db.Migrator().HasColumn(&agentNetworkTypes.AgentNetworkUsage{}, "cache_cost_usd"),
+		"legacy cache_cost_usd column must be dropped once folded")
+
+	var rows []*agentNetworkTypes.AgentNetworkUsage
+	require.NoError(t, db.Order("id").Find(&rows).Error)
+	require.Len(t, rows, 3)
+
+	// u1: total and cache portion both preserved; the read/write and
+	// input/output splits are unknowable for a legacy row, so the cache total
+	// lands on cached_input and the remainder on input.
+	assert.InDelta(t, 0.0123, rows[0].TotalCostUSD(), 1e-9, "historical total must survive the fold")
+	assert.InDelta(t, 0.0029, rows[0].CacheCostUSD(), 1e-9, "historical cache cost must survive the fold")
+	assert.InDelta(t, 0.0094, rows[0].InputCostUSD, 1e-9, "non-cache remainder lands on input")
+	assert.InDelta(t, 0.0029, rows[0].CachedInputCostUSD, 1e-9, "legacy cache total lands on cached input")
+	assert.Zero(t, rows[0].CacheCreationCostUSD, "legacy rows carry no read/write split to recover")
+	assert.Zero(t, rows[0].OutputCostUSD, "legacy rows carry no input/output split to recover")
+
+	// u2: no cache spend — the whole total is the non-cache remainder.
+	assert.InDelta(t, 0.5, rows[1].TotalCostUSD(), 1e-9, "cache-free historical total must survive")
+	assert.Zero(t, rows[1].CacheCostUSD(), "a cache-free row must stay cache-free")
+
+	// u3: zero stays zero rather than being rewritten.
+	assert.Zero(t, rows[2].TotalCostUSD(), "an unpriced row must remain unpriced")
+}
+
+// TestFoldCostAggregatesIntoBuckets_SkipsAlreadyMigrated proves the migration is
+// safe to re-run: with no legacy column present it is a no-op that leaves a
+// true four-way split untouched.
+func TestFoldCostAggregatesIntoBuckets_SkipsAlreadyMigrated(t *testing.T) {
+	ctx := context.Background()
+	db := setupDatabase(t)
+	require.NoError(t, db.Migrator().DropTable(&agentNetworkTypes.AgentNetworkUsage{}))
+
+	require.NoError(t, db.AutoMigrate(&agentNetworkTypes.AgentNetworkUsage{}))
+	// Timestamp must be set explicitly: a zero time.Time serialises as
+	// '0000-00-00 00:00:00', which MySQL rejects under strict mode.
+	require.NoError(t, db.Create(&agentNetworkTypes.AgentNetworkUsage{
+		ID: "u1", AccountID: "acct-1", Model: "claude-sonnet-4-6",
+		Timestamp:    time.Date(2026, 5, 5, 9, 0, 0, 0, time.UTC),
+		InputCostUSD: 0.001, CachedInputCostUSD: 0.002, CacheCreationCostUSD: 0.003, OutputCostUSD: 0.004,
+	}).Error)
+
+	require.NoError(t, migration.FoldCostAggregatesIntoBuckets[agentNetworkTypes.AgentNetworkUsage](ctx, db),
+		"running against an already-migrated table must be a no-op, not an error")
+
+	var row agentNetworkTypes.AgentNetworkUsage
+	require.NoError(t, db.First(&row, "id = ?", "u1").Error)
+	assert.InDelta(t, 0.001, row.InputCostUSD, 1e-9, "a true split must not be rewritten")
+	assert.InDelta(t, 0.002, row.CachedInputCostUSD, 1e-9)
+	assert.InDelta(t, 0.003, row.CacheCreationCostUSD, 1e-9)
+	assert.InDelta(t, 0.004, row.OutputCostUSD, 1e-9)
+	assert.InDelta(t, 0.01, row.TotalCostUSD(), 1e-9, "derived total sums the four buckets")
+}

management/server/store/sql_store_agentnetwork.go

@@ -71,7 +71,7 @@ func (s *SqlStore) GetAgentNetworkMetrics(ctx context.Context) (AgentNetworkMetr
 	usageRow := db.Model(&agentNetworkTypes.AgentNetworkUsage{}).
 		Select("COALESCE(SUM(input_tokens), 0) AS input_tokens, " +
 			"COALESCE(SUM(output_tokens), 0) AS output_tokens, " +
-			"COALESCE(SUM(cost_usd), 0) AS cost_usd").Row()
+			"COALESCE(SUM" + agentNetworkTypes.CostUSDSQLExpr + ", 0) AS cost_usd").Row()
 	if err := usageRow.Scan(&m.InputTokens, &m.OutputTokens, &m.CostUSD); err != nil {
 		return AgentNetworkMetrics{}, fmt.Errorf("scan agent network usage metrics: %w", err)
 	}

management/server/store/sql_store_agentnetwork_accesslog_test.go

@@ -37,7 +37,7 @@ func TestAgentNetworkUsage_RealStore_RoundTrip(t *testing.T) {
 		InputTokens:        1200,
 		OutputTokens:       640,
 		TotalTokens:        1840,
-		CostUSD:            0.0231,
+		InputCostUSD:       0.0231,
 	}
 	usageGroups := []agentNetworkTypes.AgentNetworkUsageGroup{
 		{UsageID: usage.ID, GroupID: "grp-eng", AccountID: accountID},
@@ -71,7 +71,7 @@ func TestAgentNetworkUsage_RealStore_RoundTrip(t *testing.T) {
 		InputTokens:  1200,
 		OutputTokens: 640,
 		TotalTokens:  1840,
-		CostUSD:      0.0231,
+		InputCostUSD: 0.0231,
 	}
 	entryGroups := []agentNetworkTypes.AgentNetworkAccessLogGroup{
 		{LogID: entry.ID, GroupID: "grp-eng", AccountID: accountID},
@@ -127,7 +127,7 @@ func TestAgentNetworkUsageOverview_DailyAggregation(t *testing.T) {
 	mk := func(id string, ts time.Time, model string, in, out int64, cost float64) *agentNetworkTypes.AgentNetworkUsage {
 		return &agentNetworkTypes.AgentNetworkUsage{
 			ID: id, AccountID: accountID, Timestamp: ts, Model: model,
-			InputTokens: in, OutputTokens: out, TotalTokens: in + out, CostUSD: cost,
+			InputTokens: in, OutputTokens: out, TotalTokens: in + out, InputCostUSD: cost,
 		}
 	}
 	require.NoError(t, s.CreateAgentNetworkUsage(ctx, mk("u1", day1, "gpt-4o", 100, 50, 0.10), nil))
@@ -143,7 +143,7 @@ func TestAgentNetworkUsageOverview_DailyAggregation(t *testing.T) {
 	assert.Equal(t, "2026-05-05", buckets[0].PeriodStart, "oldest-first ordering")
 	assert.Equal(t, int64(300), buckets[0].InputTokens, "same-day input tokens summed")
 	assert.Equal(t, int64(130), buckets[0].OutputTokens)
-	assert.InDelta(t, 0.30, buckets[0].CostUSD, 1e-9, "same-day cost summed")
+	assert.InDelta(t, 0.30, buckets[0].TotalCostUSD(), 1e-9, "same-day cost summed")
 	assert.Equal(t, "2026-05-06", buckets[1].PeriodStart)
 	assert.Equal(t, int64(15), buckets[1].TotalTokens)
 
@@ -174,7 +174,7 @@ func TestAgentNetworkAccessLogSessions_RealStore(t *testing.T) {
 			ID: id, AccountID: accountID, ServiceID: "svc", Timestamp: ts,
 			UserID: user, StatusCode: 200, Provider: provider, Model: model,
 			SessionID: session, Decision: decision,
-			InputTokens: 100, OutputTokens: 50, TotalTokens: 150, CostUSD: cost,
+			InputTokens: 100, OutputTokens: 50, TotalTokens: 150, InputCostUSD: cost,
 		}
 	}
 
@@ -207,7 +207,7 @@ func TestAgentNetworkAccessLogSessions_RealStore(t *testing.T) {
 	s1 := sessions[2]
 	assert.Equal(t, 2, s1.RequestCount, "s1 has two requests")
 	assert.Equal(t, int64(300), s1.TotalTokens, "tokens summed across the session")
-	assert.InDelta(t, 0.30, s1.CostUSD, 1e-9, "cost summed across the session")
+	assert.InDelta(t, 0.30, s1.TotalCostUSD(), 1e-9, "cost summed across the session")
 	assert.Equal(t, "alice", s1.UserID)
 	assert.Equal(t, "allow", s1.Decision)
 	// SQLite hands times back in time.Local; normalise to UTC so the instant is

management/server/store/store.go

@@ -650,6 +650,14 @@ func getMigrationsPostAuto(ctx context.Context) []migrationFunc {
 		func(db *gorm.DB) error {
 			return migration.DropIndex[proxy.Proxy](ctx, db, "idx_proxy_account_id_unique")
 		},
+		// Post-auto so the per-bucket cost columns already exist when the legacy
+		// aggregates are folded into them and dropped.
+		func(db *gorm.DB) error {
+			return migration.FoldCostAggregatesIntoBuckets[agentNetworkTypes.AgentNetworkAccessLog](ctx, db)
+		},
+		func(db *gorm.DB) error {
+			return migration.FoldCostAggregatesIntoBuckets[agentNetworkTypes.AgentNetworkUsage](ctx, db)
+		},
 	}
 }
 

proxy/internal/accesslog/logger.go

@@ -221,14 +221,21 @@ func (l *Logger) allowDenyLog(serviceID types.ServiceID, reason string) bool {
 // proxy/internal/middleware/keys.go — only the dimensions management needs to
 // record a usage row (provider / model / tokens / cost / groups).
 var usageMetadataKeys = map[string]struct{}{
-	"llm.provider":             {},
-	"llm.model":                {},
-	"llm.resolved_provider_id": {},
-	"llm.input_tokens":         {},
-	"llm.output_tokens":        {},
-	"llm.total_tokens":         {},
-	"cost.usd_total":           {},
-	"llm.authorising_groups":   {},
+	"llm.provider":              {},
+	"llm.model":                 {},
+	"llm.resolved_provider_id":  {},
+	"llm.input_tokens":          {},
+	"llm.output_tokens":         {},
+	"llm.total_tokens":          {},
+	"llm.cached_input_tokens":   {},
+	"llm.cache_creation_tokens": {},
+	"cost.usd_input":            {},
+	"cost.usd_cached_input":     {},
+	"cost.usd_cache_creation":   {},
+	"cost.usd_output":           {},
+	"cost.usd_total":            {},
+	"cost.usd_cache":            {},
+	"llm.authorising_groups":    {},
 }
 
 // stripAgentNetworkEntryForUsage returns the entry reduced to what's needed to

proxy/internal/auth/tunnel_cache.go

@@ -3,30 +3,20 @@ package auth
 import (
 	"context"
 	"net/netip"
-	"os"
-	"strings"
 	"sync"
 	"time"
 
-	log "github.com/sirupsen/logrus"
 	"golang.org/x/sync/singleflight"
 
 	"github.com/netbirdio/netbird/proxy/internal/types"
 	"github.com/netbirdio/netbird/shared/management/proto"
 )
 
-// tunnelCacheTTL is the default cap on how long a positive ValidateTunnelPeer
-// result is reused before re-fetching from management. 5 minutes balances
-// freshness against management load on busy mesh networks. Override it with
-// envTunnelCacheTTL when an account needs authorization changes to take effect
-// sooner (at the cost of more ValidateTunnelPeer RPCs).
+// tunnelCacheTTL caps how long a positive ValidateTunnelPeer result is
+// reused before re-fetching from management. 5 minutes balances freshness
+// against management load on busy mesh networks.
 const tunnelCacheTTL = 300 * time.Second
 
-// envTunnelCacheTTL overrides tunnelCacheTTL. The value is a Go duration string
-// (e.g. "30s", "2m"); an unset, unparseable, or non-positive value keeps the
-// default.
-const envTunnelCacheTTL = "NB_PROXY_TUNNEL_CACHE_TTL"
-
 // tunnelCachePerAccount caps the number of cached identities per account.
 // Bounded eviction avoids memory growth in pathological cases (huge peer
 // roster, brief request bursts) while staying generous for normal use.
@@ -70,35 +60,16 @@ type accountBucket struct {
 	order []tunnelCacheKey
 }
 
-// newTunnelValidationCache constructs a cache with the configured TTL
-// (envTunnelCacheTTL override or default) and default bounds.
+// newTunnelValidationCache constructs a cache with default TTL and bounds.
 func newTunnelValidationCache() *tunnelValidationCache {
 	return &tunnelValidationCache{
 		entries: make(map[types.AccountID]*accountBucket),
-		ttl:     tunnelCacheTTLFromEnv(),
+		ttl:     tunnelCacheTTL,
 		maxSize: tunnelCachePerAccount,
 		now:     time.Now,
 	}
 }
 
-// tunnelCacheTTLFromEnv returns the tunnel-cache TTL, honoring the
-// envTunnelCacheTTL override. The override must be a positive Go duration
-// string (e.g. "30s", "2m"); anything unset, unparseable, or non-positive
-// falls back to tunnelCacheTTL.
-func tunnelCacheTTLFromEnv() time.Duration {
-	raw := strings.TrimSpace(os.Getenv(envTunnelCacheTTL))
-	if raw == "" {
-		return tunnelCacheTTL
-	}
-	d, err := time.ParseDuration(raw)
-	if err != nil || d <= 0 {
-		log.Warnf("ignoring invalid %s=%q (want a positive Go duration like 30s or 2m); using default %s",
-			envTunnelCacheTTL, raw, tunnelCacheTTL)
-		return tunnelCacheTTL
-	}
-	return d
-}
-
 // get returns a cached response for the key, or nil when missing or
 // expired. Expired entries are evicted lazily on read.
 func (c *tunnelValidationCache) get(key tunnelCacheKey) *proto.ValidateTunnelPeerResponse {

proxy/internal/auth/tunnel_cache_test.go

@@ -169,32 +169,3 @@ func TestTunnelCache_BoundedSizeEvictsOldest(t *testing.T) {
 	assert.NotNil(t, cache.get(keys[1]), "second-newest must remain cached")
 	assert.NotNil(t, cache.get(keys[2]), "newest must remain cached")
 }
-
-func TestTunnelCacheTTLFromEnv(t *testing.T) {
-	t.Run("unset uses default", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "")
-		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
-	})
-	t.Run("valid duration overrides", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "45s")
-		assert.Equal(t, 45*time.Second, tunnelCacheTTLFromEnv())
-	})
-	t.Run("whitespace trimmed", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "  2m ")
-		assert.Equal(t, 2*time.Minute, tunnelCacheTTLFromEnv())
-	})
-	t.Run("unparseable uses default", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "nonsense")
-		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
-	})
-	t.Run("non-positive uses default", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "0s")
-		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
-		t.Setenv(envTunnelCacheTTL, "-30s")
-		assert.Equal(t, tunnelCacheTTL, tunnelCacheTTLFromEnv())
-	})
-	t.Run("constructor honors override", func(t *testing.T) {
-		t.Setenv(envTunnelCacheTTL, "90s")
-		assert.Equal(t, 90*time.Second, newTunnelValidationCache().ttl)
-	})
-}

proxy/internal/llm/bedrock.go

@@ -56,10 +56,12 @@ type bedrockResponse struct {
 		OutputTokens             int64 `json:"output_tokens"`
 		CacheReadInputTokens     int64 `json:"cache_read_input_tokens"`
 		CacheCreationInputTokens int64 `json:"cache_creation_input_tokens"`
-		// Converse — camelCase.
-		InputTokensCamel  int64 `json:"inputTokens"`
-		OutputTokensCamel int64 `json:"outputTokens"`
-		TotalTokensCamel  int64 `json:"totalTokens"`
+		// Converse — camelCase; cache buckets are additive to inputTokens (AWS names the write bucket cacheWriteInputTokens).
+		InputTokensCamel      int64 `json:"inputTokens"`
+		OutputTokensCamel     int64 `json:"outputTokens"`
+		TotalTokensCamel      int64 `json:"totalTokens"`
+		CacheReadTokensCamel  int64 `json:"cacheReadInputTokens"`
+		CacheWriteTokensCamel int64 `json:"cacheWriteInputTokens"`
 	} `json:"usage"`
 }
 
@@ -83,16 +85,18 @@ func (BedrockParser) ParseResponse(status int, contentType string, body []byte)
 	}
 	inTok := firstNonZero(resp.Usage.InputTokens, resp.Usage.InputTokensCamel)
 	outTok := firstNonZero(resp.Usage.OutputTokens, resp.Usage.OutputTokensCamel)
+	cacheRead := firstNonZero(resp.Usage.CacheReadInputTokens, resp.Usage.CacheReadTokensCamel)
+	cacheWrite := firstNonZero(resp.Usage.CacheCreationInputTokens, resp.Usage.CacheWriteTokensCamel)
 	total := resp.Usage.TotalTokensCamel
 	if total == 0 {
-		total = inTok + outTok + resp.Usage.CacheReadInputTokens + resp.Usage.CacheCreationInputTokens
+		total = inTok + outTok + cacheRead + cacheWrite
 	}
 	return Usage{
 		InputTokens:         inTok,
 		OutputTokens:        outTok,
 		TotalTokens:         total,
-		CachedInputTokens:   resp.Usage.CacheReadInputTokens,
-		CacheCreationTokens: resp.Usage.CacheCreationInputTokens,
+		CachedInputTokens:   cacheRead,
+		CacheCreationTokens: cacheWrite,
 	}, nil
 }
 

proxy/internal/llm/bedrock_test.go

@@ -26,6 +26,18 @@ func TestBedrockParser_ParseResponse_Converse(t *testing.T) {
 	require.Equal(t, int64(14), u.TotalTokens, "converse uses provider total")
 }
 
+// Converse camelCase cache fields must land in the billed Usage buckets, same as the InvokeModel snake_case fields.
+func TestBedrockParser_ParseResponse_ConverseCacheBuckets(t *testing.T) {
+	body := []byte(`{"usage":{"inputTokens":11,"outputTokens":3,"cacheReadInputTokens":7,"cacheWriteInputTokens":9}}`)
+	u, err := BedrockParser{}.ParseResponse(200, "application/json", body)
+	require.NoError(t, err)
+	require.Equal(t, int64(11), u.InputTokens, "converse input tokens")
+	require.Equal(t, int64(3), u.OutputTokens, "converse output tokens")
+	require.Equal(t, int64(7), u.CachedInputTokens, "converse cache-read tokens")
+	require.Equal(t, int64(9), u.CacheCreationTokens, "converse cache-write tokens")
+	require.Equal(t, int64(11+3+7+9), u.TotalTokens, "total backfill is additive when the provider omits totalTokens")
+}
+
 func TestBedrockParser_ParseResponse_StreamingUnsupported(t *testing.T) {
 	_, err := BedrockParser{}.ParseResponse(200, "application/vnd.amazon.eventstream", []byte("binary"))
 	require.ErrorIs(t, err, ErrStreamingUnsupported, "event-stream must route to the streaming accumulator")

proxy/internal/llm/pricing/defaults_coverage_test.go

@@ -28,12 +28,12 @@ func TestDefaultTable_FirstPartyModelCoverage(t *testing.T) {
 			"ministral-8b-latest", "mistral-embed",
 		},
 		"anthropic": {
-			"claude-fable-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
+			"claude-fable-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6",
 			"claude-opus-4-1", "claude-sonnet-4-6", "claude-sonnet-4-5", "claude-haiku-4-5",
 		},
 		// bedrock keys are the normalized ids the request parser emits.
 		"bedrock": {
-			"anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
+			"anthropic.claude-opus-5", "anthropic.claude-opus-4-8", "anthropic.claude-opus-4-7", "anthropic.claude-opus-4-6",
 			"anthropic.claude-opus-4-1", "anthropic.claude-sonnet-4-6", "anthropic.claude-sonnet-4-5",
 			"anthropic.claude-haiku-4-5", "meta.llama3-3-70b-instruct",
 			"amazon.nova-pro", "amazon.nova-lite", "amazon.nova-micro", "amazon.nova-2-lite",

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -180,6 +180,11 @@ anthropic:
     output_per_1k: 0.050
     cache_read_per_1k: 0.001
     cache_creation_per_1k: 0.0125
+  claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
   claude-opus-4-8:
     input_per_1k: 0.005
     output_per_1k: 0.025
@@ -236,6 +241,11 @@ bedrock:
   # eu.anthropic.claude-sonnet-4-5-20250929-v1:0 -> anthropic.claude-sonnet-4-5.
   # Anthropic-on-Bedrock keeps the additive cache buckets (read ≈0.1x input,
   # write ≈1.25x input); Nova / Llama report no cache, so cost is input+output.
+  anthropic.claude-opus-5:
+    input_per_1k: 0.005
+    output_per_1k: 0.025
+    cache_read_per_1k: 0.0005
+    cache_creation_per_1k: 0.00625
   anthropic.claude-opus-4-8:
     input_per_1k: 0.005
     output_per_1k: 0.025

proxy/internal/llm/pricing/pricing.go

@@ -128,6 +128,46 @@ type Table struct {
 //   - Other providers: cached and cacheCreation are ignored; cost is
 //     inTokens*InputPer1K + outTokens*OutputPer1K.
 func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (float64, bool) {
+	c, ok := t.Costs(provider, model, inTokens, outTokens, cachedInput, cacheCreation)
+	return c.TotalUSD, ok
+}
+
+// Costs is a per-request cost split. The four per-bucket fields are the base
+// of the breakdown — one per token bucket the provider bills separately — and
+// the two aggregates are derived from them:
+//
+//	TotalUSD = InputUSD + CachedInputUSD + CacheCreationUSD + OutputUSD
+//	CacheUSD = CachedInputUSD + CacheCreationUSD
+//
+// InputUSD is always the cost of the *non-cached* input bucket, for both
+// provider shapes: on OpenAI the cached subset is carved out of inTokens and
+// billed as CachedInputUSD, so the two never double-count. Buckets a provider
+// doesn't bill are zero, which keeps the identities above true everywhere.
+type Costs struct {
+	InputUSD         float64
+	CachedInputUSD   float64
+	CacheCreationUSD float64
+	OutputUSD        float64
+	TotalUSD         float64
+	CacheUSD         float64
+}
+
+// newCosts assembles a split from its per-bucket parts, deriving the two
+// aggregates so TotalUSD and CacheUSD can never drift from the breakdown.
+func newCosts(input, cachedInput, cacheCreation, output float64) Costs {
+	return Costs{
+		InputUSD:         input,
+		CachedInputUSD:   cachedInput,
+		CacheCreationUSD: cacheCreation,
+		OutputUSD:        output,
+		TotalUSD:         input + cachedInput + cacheCreation + output,
+		CacheUSD:         cachedInput + cacheCreation,
+	}
+}
+
+// Costs returns the estimated USD cost split for the given token counts, with
+// the same semantics as Cost.
+func (t *Table) Costs(provider, model string, inTokens, outTokens, cachedInput, cacheCreation int64) (Costs, bool) {
 	// Clamp negatives to zero before any pricing math so a malformed
 	// upstream count can never produce a negative cost.
 	if inTokens < 0 {
@@ -143,15 +183,15 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		cacheCreation = 0
 	}
 	if t == nil {
-		return 0, false
+		return Costs{}, false
 	}
 	byModel, ok := t.entries[provider]
 	if !ok {
-		return 0, false
+		return Costs{}, false
 	}
 	entry, ok := byModel[model]
 	if !ok {
-		return 0, false
+		return Costs{}, false
 	}
 	output := (float64(outTokens) / 1000.0) * entry.OutputPer1K
 	switch provider {
@@ -168,7 +208,7 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		}
 		nonCached := float64(inTokens-clamped) / 1000.0 * entry.InputPer1K
 		cached := float64(clamped) / 1000.0 * cachedRate
-		return nonCached + cached + output, true
+		return newCosts(nonCached, cached, 0, output), true
 	case "anthropic", "bedrock":
 		// Bedrock-Anthropic returns the same additive cache buckets as
 		// first-party Anthropic; non-Anthropic Bedrock models simply report
@@ -184,10 +224,10 @@ func (t *Table) Cost(provider, model string, inTokens, outTokens, cachedInput, c
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
 		read := float64(cachedInput) / 1000.0 * readRate
 		create := float64(cacheCreation) / 1000.0 * createRate
-		return input + read + create + output, true
+		return newCosts(input, read, create, output), true
 	default:
 		input := float64(inTokens) / 1000.0 * entry.InputPer1K
-		return input + output, true
+		return newCosts(input, 0, 0, output), true
 	}
 }
 

proxy/internal/middleware/builtin/cost_calculation_matrix_test.go

@@ -0,0 +1,329 @@
+package builtin_test
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"strconv"
+	"testing"
+
+	"github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_request_parser"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
+)
+
+// Drives the real pipeline (llm_request_parser → llm_response_parser → cost_meter) on the embedded default pricing
+// table and asserts exact USD amounts hardcoded from the vendors' published prices, including the cache split.
+func TestCostCalculation_ProviderMatrix(t *testing.T) {
+	// Empty data dir → embedded defaults, like a proxy with no pricing override.
+	builtin.Configure(context.Background(), t.TempDir(), nil, nil, nil)
+
+	reqMW, err := llm_request_parser.Factory{}.New(nil)
+	require.NoError(t, err, "build llm_request_parser")
+	respMW, err := llm_response_parser.Factory{}.New(nil)
+	require.NoError(t, err, "build llm_response_parser")
+	costMW, err := cost_meter.Factory{}.New(nil)
+	require.NoError(t, err, "build cost_meter")
+	t.Cleanup(func() { _ = costMW.Close() })
+
+	const jsonCT = "application/json"
+	const sseCT = "text/event-stream"
+	const awsCT = "application/vnd.amazon.eventstream"
+
+	cases := []struct {
+		name     string
+		url      string
+		reqBody  []byte
+		respCT   string
+		respBody []byte
+
+		wantProvider  string
+		wantModel     string
+		wantCost      float64 // exact expected USD; ignored when wantSkip is set
+		wantCacheCost float64 // expected cost.usd_cache portion of wantCost
+		wantSkip      string  // expected cost.skipped reason, "" when priced
+	}{
+		{
+			// gpt-4o-mini $0.15/$0.60 per MTok: 1000×0.15/1M + 500×0.60/1M.
+			name:         "openai chat completions",
+			url:          "https://api.openai.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"gpt-4o-mini","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"choices":[{"message":{"content":"pong"}}],"usage":{"prompt_tokens":1000,"completion_tokens":500,"total_tokens":1500}}`),
+			wantProvider: "openai",
+			wantModel:    "gpt-4o-mini",
+			wantCost:     0.00045,
+		},
+		{
+			// OpenAI cached tokens are a SUBSET of prompt_tokens at a discount; gpt-4o $2.50/$10 per MTok, cached $1.25/M:
+			// 250×2.5/1M + 750×1.25/1M + 500×10/1M.
+			name:          "openai cached subset discount",
+			url:           "https://api.openai.com/v1/chat/completions",
+			reqBody:       []byte(`{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":750}}}`),
+			wantProvider:  "openai",
+			wantModel:     "gpt-4o",
+			wantCost:      0.0065625,
+			wantCacheCost: 0.0009375,
+		},
+		{
+			// OpenAI streaming: usage rides the final SSE frame.
+			name:         "openai chat SSE stream",
+			url:          "https://api.openai.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"gpt-4o-mini","stream":true,"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       sseCT,
+			respBody:     sseBody(`{"choices":[{"delta":{"content":"po"}}]}`, `{"choices":[{"delta":{"content":"ng"}}]}`, `{"choices":[],"usage":{"prompt_tokens":1000,"completion_tokens":500}}`, "[DONE]"),
+			wantProvider: "openai",
+			wantModel:    "gpt-4o-mini",
+			wantCost:     0.00045,
+		},
+		{
+			// Mistral speaks the OpenAI shape: mistral-large-latest $0.50/$1.50 per MTok.
+			name:         "mistral via openai shape",
+			url:          "https://api.mistral.ai/v1/chat/completions",
+			reqBody:      []byte(`{"model":"mistral-large-latest","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":1000}}`),
+			wantProvider: "openai",
+			wantModel:    "mistral-large-latest",
+			wantCost:     0.002,
+		},
+		{
+			// The field report, minus caching: Bedrock Sonnet 4.6 $3/$15 per MTok, 3×3/1M + 1514×15/1M = $0.022719.
+			// Also covers inference-profile normalization of the region-prefixed versioned id in the URL.
+			name:         "bedrock invoke — reported scenario, no cache",
+			url:          "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke",
+			reqBody:      []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"content":[{"type":"text","text":"pong"}],"usage":{"input_tokens":3,"output_tokens":1514}}`),
+			wantProvider: "bedrock",
+			wantModel:    "anthropic.claude-sonnet-4-6",
+			wantCost:     0.022719,
+		},
+		{
+			// The field report as observed: the FIRST call of a session also wrote a 30,528-token prompt cache at
+			// 1.25× input ($3.75/M): 0.022719 + 30528×3.75/1M = $0.137199 — the reported $0.1372.
+			name:          "bedrock invoke — reported scenario with cache write",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"usage":{"input_tokens":3,"output_tokens":1514,"cache_creation_input_tokens":30528,"cache_read_input_tokens":0}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-sonnet-4-6",
+			wantCost:      0.137199,
+			wantCacheCost: 0.11448,
+		},
+		{
+			// Same numbers over the InvokeModel event-stream: message_start carries input + cache, message_delta the output.
+			name:          "bedrock invoke stream with cache write",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-sonnet-4-6-20260115-v1:0/invoke-with-response-stream",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        awsCT,
+			respBody:      bedrockInvokeStream(t, `{"type":"message_start","message":{"usage":{"input_tokens":3,"output_tokens":1,"cache_creation_input_tokens":30528}}}`, `{"type":"content_block_delta","delta":{"type":"text_delta","text":"pong"}}`, `{"type":"message_delta","usage":{"output_tokens":1514}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-sonnet-4-6",
+			wantCost:      0.137199,
+			wantCacheCost: 0.11448,
+		},
+		{
+			// Converse camelCase usage incl. cache buckets. Haiku 4.5 $1/$5 per MTok, read $0.10/M, write $1.25/M:
+			// 50×1/1M + 100×5/1M + 2000×0.1/1M + 1000×1.25/1M = $0.002.
+			name:          "bedrock converse with cache buckets",
+			url:           "https://bedrock-runtime.eu-central-1.amazonaws.com/model/eu.anthropic.claude-haiku-4-5-20251001-v1:0/converse",
+			reqBody:       []byte(`{"messages":[{"role":"user","content":[{"text":"hi"}]}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"output":{"message":{"content":[{"text":"pong"}]}},"usage":{"inputTokens":50,"outputTokens":100,"totalTokens":3150,"cacheReadInputTokens":2000,"cacheWriteInputTokens":1000}}`),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-haiku-4-5",
+			wantCost:      0.002,
+			wantCacheCost: 0.00145,
+		},
+		{
+			// Same numbers over converse-stream: usage rides the trailing metadata frame.
+			name:    "bedrock converse stream with cache buckets",
+			url:     "https://bedrock-runtime.eu-central-1.amazonaws.com/model/eu.anthropic.claude-haiku-4-5-20251001-v1:0/converse-stream",
+			reqBody: []byte(`{"messages":[{"role":"user","content":[{"text":"hi"}]}]}`),
+			respCT:  awsCT,
+			respBody: bedrockConverseStream(t,
+				`{"delta":{"text":"pong"}}`,
+				`{"usage":{"inputTokens":50,"outputTokens":100,"totalTokens":3150,"cacheReadInputTokens":2000,"cacheWriteInputTokens":1000}}`,
+			),
+			wantProvider:  "bedrock",
+			wantModel:     "anthropic.claude-haiku-4-5",
+			wantCost:      0.002,
+			wantCacheCost: 0.00145,
+		},
+		{
+			// First-party Anthropic, additive cache buckets. Sonnet 4.6:
+			// 256×3/1M + 200×15/1M + 768×0.3/1M + 512×3.75/1M.
+			name:          "anthropic messages with cache buckets",
+			url:           "https://api.anthropic.com/v1/messages",
+			reqBody:       []byte(`{"model":"claude-sonnet-4-6","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:        jsonCT,
+			respBody:      []byte(`{"content":[{"type":"text","text":"pong"}],"usage":{"input_tokens":256,"output_tokens":200,"cache_read_input_tokens":768,"cache_creation_input_tokens":512}}`),
+			wantProvider:  "anthropic",
+			wantModel:     "claude-sonnet-4-6",
+			wantCost:      0.0059184,
+			wantCacheCost: 0.0021504,
+		},
+		{
+			// Anthropic SSE: input from message_start, output from message_delta. Haiku 4.5: 1000×1/1M + 2000×5/1M.
+			name:         "anthropic SSE stream",
+			url:          "https://api.anthropic.com/v1/messages",
+			reqBody:      []byte(`{"model":"claude-haiku-4-5","stream":true,"messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       sseCT,
+			respBody:     sseBody(`{"type":"message_start","message":{"usage":{"input_tokens":1000,"output_tokens":2}}}`, `{"type":"content_block_delta","delta":{"type":"text_delta","text":"pong"}}`, `{"type":"message_delta","usage":{"output_tokens":2000}}`, `{"type":"message_stop"}`),
+			wantProvider: "anthropic",
+			wantModel:    "claude-haiku-4-5",
+			wantCost:     0.011,
+		},
+		{
+			// Kimi's Anthropic-compatible endpoint: kimi-k3 $3/$15 per MTok under the anthropic table.
+			name:         "kimi anthropic shape",
+			url:          "https://api.moonshot.ai/anthropic/v1/messages",
+			reqBody:      []byte(`{"model":"kimi-k3","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"input_tokens":1000,"output_tokens":1000}}`),
+			wantProvider: "anthropic",
+			wantModel:    "kimi-k3",
+			wantCost:     0.018,
+		},
+		{
+			// Vertex path-routed model with "@version" stripped; Anthropic-on-Vertex priced under the anthropic table.
+			name:         "vertex anthropic path-routed",
+			url:          "https://aiplatform.googleapis.com/v1/projects/p/locations/global/publishers/anthropic/models/claude-sonnet-4-6@20260115:rawPredict",
+			reqBody:      []byte(`{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"input_tokens":200,"output_tokens":100}}`),
+			wantProvider: "anthropic",
+			wantModel:    "claude-sonnet-4-6",
+			wantCost:     0.0021,
+		},
+		{
+			// Gateway-prefixed model ids are not in the pricing table: the meter must SKIP, never guess a rate.
+			name:         "gateway-prefixed model skips pricing",
+			url:          "https://gateway.example.com/v1/chat/completions",
+			reqBody:      []byte(`{"model":"openai/gpt-4o-mini","messages":[{"role":"user","content":"hi"}]}`),
+			respCT:       jsonCT,
+			respBody:     []byte(`{"usage":{"prompt_tokens":1000,"completion_tokens":500}}`),
+			wantProvider: "openai",
+			wantModel:    "openai/gpt-4o-mini",
+			wantSkip:     "unknown_model",
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			in := &middleware.Input{
+				Method:  "POST",
+				URL:     tc.url,
+				Headers: []middleware.KV{{Key: "Content-Type", Value: "application/json"}},
+				Body:    tc.reqBody,
+			}
+
+			reqOut, err := reqMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "request parser")
+			in.Metadata = append(in.Metadata, reqOut.Metadata...)
+
+			require.Equal(t, tc.wantProvider, metaKV(in.Metadata, middleware.KeyLLMProvider), "detected provider")
+			require.Equal(t, tc.wantModel, metaKV(in.Metadata, middleware.KeyLLMModel), "detected (normalized) model")
+
+			in.Status = 200
+			in.RespHeaders = []middleware.KV{{Key: "Content-Type", Value: tc.respCT}}
+			in.RespBody = tc.respBody
+
+			respOut, err := respMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "response parser")
+			in.Metadata = append(in.Metadata, respOut.Metadata...)
+
+			costOut, err := costMW.Invoke(context.Background(), in)
+			require.NoError(t, err, "cost meter")
+
+			if tc.wantSkip != "" {
+				assert.Equal(t, tc.wantSkip, metaKV(costOut.Metadata, middleware.KeyCostSkipped), "expected cost skip reason")
+				assert.Empty(t, metaKV(costOut.Metadata, middleware.KeyCostUSDTotal), "no cost may be emitted on skip")
+				return
+			}
+
+			raw := metaKV(costOut.Metadata, middleware.KeyCostUSDTotal)
+			require.NotEmpty(t, raw, "cost.usd_total must be emitted; skip=%q", metaKV(costOut.Metadata, middleware.KeyCostSkipped))
+			got, err := strconv.ParseFloat(raw, 64)
+			require.NoError(t, err, "cost must be a float")
+			// cost.usd_total is rendered with %.6f: allow half of the last printed digit on top of float error.
+			assert.InDelta(t, tc.wantCost, got, 5.1e-7, "USD cost for %s", tc.name)
+
+			rawCache := metaKV(costOut.Metadata, middleware.KeyCostUSDCache)
+			require.NotEmpty(t, rawCache, "cost.usd_cache must be emitted next to cost.usd_total")
+			gotCache, err := strconv.ParseFloat(rawCache, 64)
+			require.NoError(t, err, "cache cost must be a float")
+			assert.InDelta(t, tc.wantCacheCost, gotCache, 5.1e-7, "cache USD cost for %s", tc.name)
+		})
+	}
+}
+
+// metaKV returns the value for key in kvs, or "" when absent.
+func metaKV(kvs []middleware.KV, key string) string {
+	for _, kv := range kvs {
+		if kv.Key == key {
+			return kv.Value
+		}
+	}
+	return ""
+}
+
+// sseBody renders data frames as a text/event-stream body.
+func sseBody(frames ...string) []byte {
+	var b bytes.Buffer
+	for _, f := range frames {
+		b.WriteString("data: ")
+		b.WriteString(f)
+		b.WriteString("\n\n")
+	}
+	return b.Bytes()
+}
+
+// awsFrame encodes one AWS event-stream frame with the given :event-type.
+func awsFrame(t *testing.T, eventType string, payload []byte) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	enc := eventstream.NewEncoder()
+	require.NoError(t, enc.Encode(&buf, eventstream.Message{
+		Headers: eventstream.Headers{{Name: ":event-type", Value: eventstream.StringValue(eventType)}},
+		Payload: payload,
+	}), "encode event-stream frame")
+	return buf.Bytes()
+}
+
+// bedrockInvokeStream builds an invoke-with-response-stream body: each "chunk" frame wraps a base64 Anthropic event.
+func bedrockInvokeStream(t *testing.T, events ...string) []byte {
+	t.Helper()
+	var body bytes.Buffer
+	for _, ev := range events {
+		wrap, err := json.Marshal(map[string]string{"bytes": base64.StdEncoding.EncodeToString([]byte(ev))})
+		require.NoError(t, err)
+		body.Write(awsFrame(t, "chunk", wrap))
+	}
+	return body.Bytes()
+}
+
+// bedrockConverseStream builds a converse-stream body: contentBlockDelta frames plus a trailing metadata usage frame.
+func bedrockConverseStream(t *testing.T, deltas ...string) []byte {
+	t.Helper()
+	var body bytes.Buffer
+	for i, ev := range deltas {
+		eventType := "contentBlockDelta"
+		if i == len(deltas)-1 {
+			eventType = "metadata"
+		}
+		body.Write(awsFrame(t, eventType, []byte(ev)))
+	}
+	return body.Bytes()
+}

proxy/internal/middleware/builtin/cost_meter/middleware.go

@@ -32,7 +32,12 @@ const (
 )
 
 var metadataKeys = []string{
+	middleware.KeyCostUSDInput,
+	middleware.KeyCostUSDCachedInput,
+	middleware.KeyCostUSDCacheCreation,
+	middleware.KeyCostUSDOutput,
 	middleware.KeyCostUSDTotal,
+	middleware.KeyCostUSDCache,
 	middleware.KeyCostSkipped,
 }
 
@@ -140,18 +145,38 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 	}
 
 	table := m.loader.Get()
-	cost, ok := table.Cost(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
+	costs, ok := table.Costs(provider, model, inTokens, outTokens, cachedTokens, cacheCreationTokens)
 	if !ok {
 		out.Metadata = skip(skipUnknownModel)
 		return out, nil
 	}
 
+	// Per-bucket costs first: they're the base of the breakdown, and the two
+	// aggregates that follow are derived from exactly these four values.
 	out.Metadata = []middleware.KV{
-		{Key: middleware.KeyCostUSDTotal, Value: fmt.Sprintf("%.6f", cost)},
+		{Key: middleware.KeyCostUSDInput, Value: usd(costs.InputUSD)},
+		{Key: middleware.KeyCostUSDCachedInput, Value: usd(costs.CachedInputUSD)},
+		{Key: middleware.KeyCostUSDCacheCreation, Value: usd(costs.CacheCreationUSD)},
+		{Key: middleware.KeyCostUSDOutput, Value: usd(costs.OutputUSD)},
+		{Key: middleware.KeyCostUSDTotal, Value: usd(costs.TotalUSD)},
+		{Key: middleware.KeyCostUSDCache, Value: usd(costs.CacheUSD)},
 	}
 	return out, nil
 }
 
+// usd renders a cost as the fixed-precision string every cost.usd_* key
+// carries, so the per-bucket values and the aggregates round identically.
+//
+// 9 decimals, not 6: these values are summed downstream — per request, per
+// session, and per usage bucket — so the rounding step is applied once per
+// bucket per row and then accumulated. At 6 decimals a single row loses up to
+// 2e-6 across its four buckets (enough to break a 1e-6 reconciliation against
+// published rates), and a bucket smaller than half a microdollar quantises to
+// zero outright: 16 cache-read tokens on a cheap model is 1.6e-9, so summing
+// 10k such rows reports 0.02 instead of 0.016. Nano-dollar precision keeps the
+// per-row error ~1000x below the smallest realistic bucket.
+func usd(v float64) string { return fmt.Sprintf("%.9f", v) }
+
 // skip returns a single-entry metadata slice carrying the given skip
 // reason under KeyCostSkipped.
 func skip(reason string) []middleware.KV {

proxy/internal/middleware/builtin/cost_meter/middleware_test.go

@@ -67,7 +67,15 @@ func TestMiddleware_StaticSurface(t *testing.T) {
 	assert.NoError(t, mw.Close(), "Close on stateless middleware is a no-op")
 
 	keys := mw.MetadataKeys()
-	expected := []string{middleware.KeyCostUSDTotal, middleware.KeyCostSkipped}
+	expected := []string{
+		middleware.KeyCostUSDInput,
+		middleware.KeyCostUSDCachedInput,
+		middleware.KeyCostUSDCacheCreation,
+		middleware.KeyCostUSDOutput,
+		middleware.KeyCostUSDTotal,
+		middleware.KeyCostUSDCache,
+		middleware.KeyCostSkipped,
+	}
 	assert.Equal(t, expected, keys, "metadata key allowlist must match the spec")
 }
 
@@ -105,7 +113,7 @@ func TestFactory_DefaultPricingPathLoadsFixture(t *testing.T) {
 
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted for known model")
-	assert.Equal(t, "0.000750", value, "0.00015 + 0.0006 per 1k tokens, 6-decimal format")
+	assert.Equal(t, "0.000750000", value, "0.00015 + 0.0006 per 1k tokens, 9-decimal format")
 }
 
 func TestFactory_PricingPathOverride(t *testing.T) {
@@ -129,7 +137,7 @@ func TestFactory_PricingPathOverride(t *testing.T) {
 
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted with custom pricing path")
-	assert.Equal(t, "0.015000", value, "2*0.0025 + 1*0.01 = 0.015 with 6-decimal format")
+	assert.Equal(t, "0.015000000", value, "2*0.0025 + 1*0.01 = 0.015 with 9-decimal format")
 }
 
 func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
@@ -148,7 +156,7 @@ func TestInvoke_ComputesCostForKnownModel(t *testing.T) {
 
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cost.usd_total must be emitted")
-	assert.Equal(t, "0.018000", value, "0.003 + 0.015 = 0.018 with 6-decimal format")
+	assert.Equal(t, "0.018000000", value, "0.003 + 0.015 = 0.018 with 9-decimal format")
 	_, skipped := metaValue(t, out.Metadata, middleware.KeyCostSkipped)
 	assert.False(t, skipped, "cost.skipped must not be set when cost is computed")
 }
@@ -357,8 +365,25 @@ func TestInvoke_OpenAICachedSubsetDiscount(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "cached subset path must produce a cost — never a skip")
 	// 250 non-cached at 0.0025/1k + 750 cached at 0.00125/1k + 500 output at 0.01/1k.
-	assert.Equal(t, "0.006563", value,
+	assert.Equal(t, "0.006562500", value,
 		"cached subset must be billed at the discount rate, non-cached at the full rate; never double-billed")
+
+	cache, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDCache)
+	require.True(t, ok, "cost.usd_cache must be emitted alongside cost.usd_total")
+	// 750 cached at 0.00125/1k = 0.0009375.
+	assert.Equal(t, "0.000937500", cache, "cache cost is the discounted cost of the cached subset")
+
+	// Per-bucket breakdown. On OpenAI the cached subset is carved out of the
+	// input bucket, so input covers only the 250 non-cached tokens — the two
+	// must never double-count the same 750 tokens.
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDInput, "0.000625000",
+		"input bucket bills only the non-cached remainder")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCachedInput, "0.000937500",
+		"cached-input bucket bills the discounted subset")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCacheCreation, "0.000000000",
+		"OpenAI has no cache-write bucket")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDOutput, "0.005000000",
+		"output bucket bills 500 tokens at 0.01/1k")
 }
 
 // TestInvoke_AnthropicCacheBucketsAdditive proves the Anthropic
@@ -384,9 +409,33 @@ func TestInvoke_AnthropicCacheBucketsAdditive(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 256 input * 0.003 + 768 cache_read * 0.0003 + 512 cache_creation * 0.00375 + 200 output * 0.015
-	// = 0.000768 + 0.0002304 + 0.00192 + 0.003 = 0.0059184 → "0.005918" with 6-decimal format.
-	assert.Equal(t, "0.005918", value,
+	// = 0.000768 + 0.0002304 + 0.00192 + 0.003 = 0.0059184.
+	assert.Equal(t, "0.005918400", value,
 		"each Anthropic input bucket must bill at its own rate — cache_read cheap, cache_creation expensive, regular input mid")
+
+	cache, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDCache)
+	require.True(t, ok, "cost.usd_cache must be emitted alongside cost.usd_total")
+	// 768 cache_read * 0.0003 + 512 cache_creation * 0.00375 = 0.0021504.
+	assert.Equal(t, "0.002150400", cache, "cache cost sums the read and creation buckets")
+
+	// Per-bucket breakdown: four separately-billed buckets, each at its own rate.
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDInput, "0.000768000",
+		"input bucket bills 256 tokens at 0.003/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCachedInput, "0.000230400",
+		"cache-read bucket bills 768 tokens at the cheap 0.0003/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDCacheCreation, "0.001920000",
+		"cache-write bucket bills 512 tokens at the expensive 0.00375/1k")
+	assertBucket(t, out.Metadata, middleware.KeyCostUSDOutput, "0.003000000",
+		"output bucket bills 200 tokens at 0.015/1k")
+}
+
+// assertBucket asserts one per-bucket cost key carries the expected
+// 6-decimal value.
+func assertBucket(t *testing.T, md []middleware.KV, key, want, msg string) {
+	t.Helper()
+	got, ok := metaValue(t, md, key)
+	require.Truef(t, ok, "%s must be emitted", key)
+	assert.Equal(t, want, got, msg)
 }
 
 // TestInvoke_CachedTokensAbsentFallsBackToBaseFormula covers the
@@ -411,7 +460,7 @@ func TestInvoke_CachedTokensAbsentFallsBackToBaseFormula(t *testing.T) {
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok)
 	// 1000 input * 0.0025 + 500 output * 0.01 = 0.0025 + 0.005 = 0.0075
-	assert.Equal(t, "0.007500", value, "no cached metadata = same cost as before the feature landed")
+	assert.Equal(t, "0.007500000", value, "no cached metadata = same cost as before the feature landed")
 }
 
 // TestInvoke_UnparseableCachedTokensSkippedSilently proves the
@@ -435,7 +484,7 @@ func TestInvoke_UnparseableCachedTokensSkippedSilently(t *testing.T) {
 	require.NoError(t, err)
 	value, ok := metaValue(t, out.Metadata, middleware.KeyCostUSDTotal)
 	require.True(t, ok, "garbage cache metadata must NOT switch the response from a cost to a skip — fall back to 0 cached")
-	assert.Equal(t, "0.007500", value, "same as the no-cached-metadata path")
+	assert.Equal(t, "0.007500000", value, "same as the no-cached-metadata path")
 }
 
 // TestMiddleware_CloseCancelsReloader proves Close stops the per-instance

proxy/internal/middleware/builtin/llm_guardrail/factory.go

@@ -10,15 +10,11 @@ import (
 )
 
 // Config is the JSON-decoded shape accepted by the factory. The
-// runtime path consumes the normalised allowlists; raw config is not
+// runtime path consumes the normalised allowlist; raw config is not
 // retained beyond construction.
 type Config struct {
-	// ProviderAllowlists maps a resolved provider id (KeyLLMResolvedProviderID) to
-	// its model allowlist. A provider present is restricted to those models; one
-	// absent is unrestricted. Kept per-provider so one provider's list can't leak
-	// onto another.
-	ProviderAllowlists map[string][]string `json:"provider_allowlists,omitempty"`
-	PromptCapture      PromptCapture       `json:"prompt_capture"`
+	ModelAllowlist []string      `json:"model_allowlist"`
+	PromptCapture  PromptCapture `json:"prompt_capture"`
 }
 
 // PromptCapture toggles the optional prompt capture + redaction step
@@ -58,28 +54,21 @@ func isEmptyJSON(raw []byte) bool {
 	return false
 }
 
-// normaliseConfig lowercases and trims allowlist entries for case-insensitive
-// matching; empty entries drop. A provider whose entries all drop keeps an empty
-// (non-nil) list — "deny every model" — distinct from an absent provider
-// (unrestricted).
+// normaliseConfig lowercases and trims allowlist entries so the runtime
+// match is case-insensitive. Empty entries are dropped.
 func normaliseConfig(cfg Config) Config {
-	if len(cfg.ProviderAllowlists) == 0 {
-		cfg.ProviderAllowlists = nil
+	if len(cfg.ModelAllowlist) == 0 {
 		return cfg
 	}
-	cleaned := make(map[string][]string, len(cfg.ProviderAllowlists))
-	for provider, models := range cfg.ProviderAllowlists {
-		list := make([]string, 0, len(models))
-		for _, entry := range models {
-			n := normaliseModel(entry)
-			if n == "" {
-				continue
-			}
-			list = append(list, n)
+	cleaned := make([]string, 0, len(cfg.ModelAllowlist))
+	for _, entry := range cfg.ModelAllowlist {
+		n := normaliseModel(entry)
+		if n == "" {
+			continue
 		}
-		cleaned[provider] = list
+		cleaned = append(cleaned, n)
 	}
-	cfg.ProviderAllowlists = cleaned
+	cfg.ModelAllowlist = cleaned
 	return cfg
 }
 

proxy/internal/middleware/builtin/llm_guardrail/middleware.go

@@ -83,9 +83,8 @@ func (m *Middleware) MutationsSupported() bool { return false }
 // prompt capture only affects the metadata emitted alongside an allow.
 func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
 	model, modelPresent := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
-	providerID, _ := lookupMetadata(in.Metadata, middleware.KeyLLMResolvedProviderID)
 
-	if denial := m.evaluateAllowlist(providerID, model, modelPresent); denial != nil {
+	if denial := m.evaluateAllowlist(model, modelPresent); denial != nil {
 		return denial, nil
 	}
 
@@ -111,32 +110,20 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 // is a no-op.
 func (m *Middleware) Close() error { return nil }
 
-// evaluateAllowlist denies when the resolved provider's allowlist rejects the
-// model; nil means proceed. Scoped to the provider llm_router resolved, so an
-// unrestricted provider (absent from config) is never caught by another's list.
-func (m *Middleware) evaluateAllowlist(providerID, model string, modelPresent bool) *middleware.Output {
-	if len(m.cfg.ProviderAllowlists) == 0 {
+// evaluateAllowlist returns a deny Output when the configured allowlist
+// rejects the model. A nil return means the request should proceed.
+func (m *Middleware) evaluateAllowlist(model string, modelPresent bool) *middleware.Output {
+	if len(m.cfg.ModelAllowlist) == 0 {
 		return nil
 	}
-	// Restrictions exist but the resolved provider is unknown, so we can't tell
-	// if this request targets a restricted provider — fail closed. llm_router
-	// normally stamps the provider first, so this is a defensive guard.
-	if providerID == "" {
-		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
-	}
-	allowlist, restricted := m.cfg.ProviderAllowlists[providerID]
-	if !restricted {
-		// This provider has no allowlist (some authorising policy left it
-		// unrestricted); management owns any per-policy/group decision.
-		return nil
-	}
-	// Fail closed: with an allowlist in effect for this provider, a request whose
-	// model the parser couldn't extract (absent/empty) is denied. This enforces
-	// the allowlist for path-routed providers (Bedrock, Vertex) with no body model.
+	// Fail closed: with an allowlist configured, a request whose model the
+	// upstream parser could not extract (absent or empty) must be denied rather
+	// than allowed. This is what enforces the allowlist for URL/path-routed
+	// providers (Bedrock, Vertex, ...) whose model lives outside the JSON body.
 	if !modelPresent || normaliseModel(model) == "" {
 		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
-	if modelInAllowlist(allowlist, model) {
+	if m.modelInAllowlist(model) {
 		return nil
 	}
 	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
@@ -164,15 +151,14 @@ func denyModel(model, code, message, reason string) *middleware.Output {
 	}
 }
 
-// modelInAllowlist reports whether the model matches any entry in the supplied
-// (already-normalised) allowlist under the case-insensitive, trim-tolerant
-// comparison rule.
-func modelInAllowlist(allowlist []string, model string) bool {
+// modelInAllowlist reports whether the model matches any allowlist
+// entry under the case-insensitive, trim-tolerant comparison rule.
+func (m *Middleware) modelInAllowlist(model string) bool {
 	normalised := normaliseModel(model)
 	if normalised == "" {
 		return false
 	}
-	for _, allowed := range allowlist {
+	for _, allowed := range m.cfg.ModelAllowlist {
 		if allowed == normalised {
 			return true
 		}

proxy/internal/middleware/builtin/llm_guardrail/middleware_test.go

@@ -26,25 +26,6 @@ func newInput(meta ...middleware.KV) *middleware.Input {
 	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: meta}
 }
 
-const (
-	testProvider  = "prov-1"
-	otherProvider = "prov-2"
-)
-
-// providerCfg builds a Config restricting testProvider to the given models.
-func providerCfg(models ...string) Config {
-	return Config{ProviderAllowlists: map[string][]string{testProvider: models}}
-}
-
-// newInputProvider builds an input that carries a resolved provider id (as
-// llm_router would stamp) plus any extra metadata.
-func newInputProvider(provider string, meta ...middleware.KV) *middleware.Input {
-	all := make([]middleware.KV, 0, len(meta)+1)
-	all = append(all, middleware.KV{Key: middleware.KeyLLMResolvedProviderID, Value: provider})
-	all = append(all, meta...)
-	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: all}
-}
-
 func TestMiddlewareIdentity(t *testing.T) {
 	mw := New(Config{})
 	assert.Equal(t, ID, mw.ID(), "middleware ID must be llm_guardrail")
@@ -66,12 +47,12 @@ func TestMiddlewareIdentity(t *testing.T) {
 
 func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 	mw := New(Config{})
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	out, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no provider allowlists must allow any model")
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "empty allowlist must allow any model")
 	v, ok := metaValue(t, out.Metadata, middleware.KeyLLMPolicyDecision)
 	require.True(t, ok, "decision metadata must be emitted")
 	assert.Equal(t, "allow", v, "decision must be allow")
@@ -81,8 +62,8 @@ func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 }
 
 func TestAllowlistMatchAllows(t *testing.T) {
-	mw := New(providerCfg("gpt-4o", "claude-opus-4"))
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o", "claude-opus-4"}})
+	out, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
@@ -90,8 +71,8 @@ func TestAllowlistMatchAllows(t *testing.T) {
 }
 
 func TestAllowlistMissDenies(t *testing.T) {
-	mw := New(providerCfg("gpt-4o"))
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
+	out, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
 	))
 	require.NoError(t, err)
@@ -110,10 +91,10 @@ func TestAllowlistMissDenies(t *testing.T) {
 }
 
 func TestAllowlistCaseInsensitive(t *testing.T) {
-	mw := New(providerCfg("  GPT-4o  ", "Claude-OPUS-4"))
+	mw := New(Config{ModelAllowlist: []string{"  GPT-4o  ", "Claude-OPUS-4"}})
 	cases := []string{"gpt-4o", "GPT-4O", "  claude-opus-4 "}
 	for _, model := range cases {
-		out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		out, err := mw.Invoke(context.Background(), newInput(
 			middleware.KV{Key: middleware.KeyLLMModel, Value: model},
 		))
 		require.NoError(t, err)
@@ -122,15 +103,14 @@ func TestAllowlistCaseInsensitive(t *testing.T) {
 }
 
 func TestAllowlistMissingModelKeyDenies(t *testing.T) {
-	// Fail closed: with an allowlist in effect for the resolved provider, a
-	// request whose model the parser could not extract (URL/path-routed
-	// providers such as Bedrock or Vertex whose shape wasn't recognised) must be
-	// denied, not allowed.
-	mw := New(providerCfg("gpt-4o"))
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider))
+	// Fail closed: with an allowlist configured, a request whose model the
+	// parser could not extract (URL/path-routed providers such as Bedrock or
+	// Vertex whose shape wasn't recognised) must be denied, not allowed.
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
+	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when the provider is restricted")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when an allowlist is set")
 	assert.Equal(t, 403, out.DenyStatus, "deny status must be 403")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
@@ -142,101 +122,26 @@ func TestAllowlistMissingModelKeyDenies(t *testing.T) {
 
 func TestAllowlistEmptyModelValueDenies(t *testing.T) {
 	// A present-but-empty model is as undeterminable as an absent one.
-	mw := New(providerCfg("gpt-4o"))
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
+	out, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "   "},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when the provider is restricted")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when an allowlist is set")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
 }
 
 func TestAllowlistEmptyListAllowsMissingModel(t *testing.T) {
-	// Without any provider allowlists there is nothing to enforce, so a missing
-	// model is still allowed — the fail-closed rule only applies when a
-	// restriction is in effect.
+	// Without an allowlist there is nothing to enforce, so a missing model is
+	// still allowed — the fail-closed rule only applies when a list is set.
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no allowlist must allow even without a model")
 }
 
-func TestUnrestrictedProviderAllowsAnyModel(t *testing.T) {
-	// The request resolved to otherProvider, which has no allowlist, so its
-	// traffic must not be caught by testProvider's restriction — the
-	// cross-provider-leak / false-deny guard.
-	mw := New(providerCfg("gpt-4o"))
-	out, err := mw.Invoke(context.Background(), newInputProvider(otherProvider,
-		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
-	))
-	require.NoError(t, err)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "an unrestricted provider must not inherit another provider's allowlist")
-}
-
-func TestPerProviderAllowlistsAreIsolated(t *testing.T) {
-	// gpt-4o is allowed only on testProvider; claude-opus-4 only on
-	// otherProvider. A model allowlisted for one provider must not be usable on
-	// the other — the fail-closed layer never unions allowlists across providers.
-	mw := New(Config{ProviderAllowlists: map[string][]string{
-		testProvider:  {"gpt-4o"},
-		otherProvider: {"claude-opus-4"},
-	}})
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
-		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
-	))
-	require.NoError(t, err)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "claude-opus-4 is allowed only on otherProvider, not testProvider")
-	require.NotNil(t, out.DenyReason)
-	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "cross-provider model must be blocked, not model_unknown")
-}
-
-func TestRestrictionsButNoResolvedProviderFailsClosed(t *testing.T) {
-	// Restrictions exist for the account but the resolved provider id is absent,
-	// so the request cannot be scoped to a provider. Fail closed rather than
-	// wave it through.
-	mw := New(providerCfg("gpt-4o"))
-	out, err := mw.Invoke(context.Background(), newInput(
-		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-	))
-	require.NoError(t, err)
-	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "missing resolved provider must fail closed when restrictions exist")
-	require.NotNil(t, out.DenyReason)
-	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
-}
-
-func TestEnabledButEmptyAllowlistDeniesEveryModel(t *testing.T) {
-	// An allowlist-enabled provider with zero models is distinct from an
-	// unrestricted (absent) provider: it must deny every model.
-	mw := New(providerCfg())
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
-		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-	))
-	require.NoError(t, err)
-	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "an enabled-but-empty allowlist must deny every model")
-	require.NotNil(t, out.DenyReason)
-	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "deny code must be model_blocked, not model_unknown")
-}
-
-func TestFactoryAllEmptyEntriesDenyEveryModel(t *testing.T) {
-	// All the provider's entries are blank; they collapse to a non-nil empty
-	// list (deny everything for that provider), not "no restriction".
-	raw := []byte(`{"provider_allowlists":{"prov-1":["","   "]}}`)
-	mw, err := Factory{}.New(raw)
-	require.NoError(t, err)
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
-		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
-	))
-	require.NoError(t, err)
-	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "all-blank allowlist entries must still restrict the provider")
-	require.NotNil(t, out.DenyReason)
-	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "deny code must be model_blocked")
-}
-
 func TestPromptCaptureDisabledEmitsNoPrompt(t *testing.T) {
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput(
@@ -312,8 +217,8 @@ func TestFactoryAcceptsZeroConfigs(t *testing.T) {
 
 func TestFactoryDecodesValidConfig(t *testing.T) {
 	cfg := Config{
-		ProviderAllowlists: map[string][]string{testProvider: {"gpt-4o"}},
-		PromptCapture:      PromptCapture{Enabled: true, RedactPii: true},
+		ModelAllowlist: []string{"gpt-4o"},
+		PromptCapture:  PromptCapture{Enabled: true, RedactPii: true},
 	}
 	raw, err := json.Marshal(cfg)
 	require.NoError(t, err, "marshalling test config must succeed")
@@ -329,15 +234,15 @@ func TestFactoryRejectsMalformedJSON(t *testing.T) {
 }
 
 func TestFactoryNormalisesAllowlist(t *testing.T) {
-	raw := []byte(`{"provider_allowlists":{"prov-1":["  GPT-4o  ","",""," Claude-3 "]}}`)
+	raw := []byte(`{"model_allowlist":["  GPT-4o  ","",""," Claude-3 "]}`)
 	mw, err := Factory{}.New(raw)
 	require.NoError(t, err)
-	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	out, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "factory must lowercase + trim allowlist entries")
-	out2, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+	out2, err := mw.Invoke(context.Background(), newInput(
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-3"},
 	))
 	require.NoError(t, err)

proxy/internal/middleware/builtin/llm_limit_check/middleware.go

@@ -175,7 +175,7 @@ func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Ou
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
 			Code:    code,
-			Message: denyMessageForCode(code),
+			Message: "LLM policy limit exceeded",
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},
@@ -184,21 +184,6 @@ func denyFromManagement(resp *proto.CheckLLMPolicyLimitsResponse) *middleware.Ou
 	}
 }
 
-// denyMessageForCode maps a management deny code to a public message.
-// Model-allowlist rejections get a model-specific message matching the
-// local guardrail; everything else keeps the generic quota wording. The
-// message stays generic so it never leaks internal quota detail.
-func denyMessageForCode(code string) string {
-	switch code {
-	case "llm_policy.model_blocked":
-		return "model is not in the policy allowlist"
-	case "llm_policy.model_unknown":
-		return "request model could not be determined for the policy allowlist"
-	default:
-		return "LLM policy limit exceeded"
-	}
-}
-
 // lookupKV returns the value associated with key, or the empty
 // string when absent.
 func lookupKV(kvs []middleware.KV, key string) string {

proxy/internal/middleware/builtin/llm_limit_check/middleware_test.go

@@ -115,46 +115,6 @@ func TestInvoke_DenyConvertsToProxyDeny(t *testing.T) {
 	assert.NotContains(t, out.DenyReason.Message, "1000", "internal cap numbers must not reach the caller")
 }
 
-// TestInvoke_ModelDenyMessages proves a model-allowlist rejection gets a
-// model-specific public message rather than the generic quota wording, so a
-// blocked or undetermined model reads consistently with the local guardrail.
-func TestInvoke_ModelDenyMessages(t *testing.T) {
-	cases := []struct {
-		name    string
-		code    string
-		message string
-	}{
-		{"blocked", "llm_policy.model_blocked", "model is not in the policy allowlist"},
-		{"unknown", "llm_policy.model_unknown", "request model could not be determined for the policy allowlist"},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			mgmt := &fakeMgmt{
-				checkResp: &proto.CheckLLMPolicyLimitsResponse{
-					Decision: "deny",
-					DenyCode: tc.code,
-				},
-			}
-			m := New(mgmt, nil)
-
-			out := runInvoke(t, m, &middleware.Input{
-				AccountID:  "acc-1",
-				UserGroups: []string{"grp-engineers"},
-				Metadata: []middleware.KV{
-					{Key: middleware.KeyLLMResolvedProviderID, Value: "prov-1"},
-					{Key: middleware.KeyLLMModel, Value: "some-model"},
-				},
-			})
-
-			assert.Equal(t, middleware.DecisionDeny, out.Decision)
-			require.NotNil(t, out.DenyReason, "deny envelope must carry a reason payload")
-			assert.Equal(t, tc.code, out.DenyReason.Code, "canonical deny code surfaces to the caller")
-			assert.Equal(t, tc.message, out.DenyReason.Message,
-				"model denials must use a model-specific message, matching the local guardrail")
-		})
-	}
-}
-
 // TestInvoke_NoMgmtClientPassesThrough proves the partial-wiring
 // safety: a middleware constructed without a management client
 // allows every request without attribution. This makes a half-set-up

proxy/internal/middleware/builtin/llm_request_parser/guardrail_allowlist_test.go

@@ -25,17 +25,10 @@ func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []strin
 	})
 	require.NoError(t, err, "parser must not error")
 
-	const providerID = "prov-under-test"
-	guard := llm_guardrail.New(llm_guardrail.Config{
-		ProviderAllowlists: map[string][]string{providerID: allowlist},
-	})
-	// The real chain has llm_router stamp the resolved provider id before the
-	// guardrail runs; the parser doesn't, so add it here so the guardrail can
-	// scope the allowlist to this provider.
-	meta := append([]middleware.KV{{Key: middleware.KeyLLMResolvedProviderID, Value: providerID}}, parsed.Metadata...)
+	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
 	out, err := guard.Invoke(context.Background(), &middleware.Input{
 		Slot:     middleware.SlotOnRequest,
-		Metadata: meta,
+		Metadata: parsed.Metadata,
 	})
 	require.NoError(t, err, "guardrail must not error")
 	require.NotNil(t, out, "guardrail must return an output")

proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock.go

@@ -69,15 +69,18 @@ func applyBedrockInvokeChunk(payload []byte, usage *llm.Usage, completion *strin
 }
 
 // converseStreamEvent captures the Converse stream frames carrying completion
-// text (contentBlockDelta) and the final token usage (metadata).
+// text (contentBlockDelta) and the final token usage (metadata). Cache buckets
+// are additive to inputTokens (AWS write bucket: cacheWriteInputTokens).
 type converseStreamEvent struct {
 	Delta *struct {
 		Text string `json:"text"`
 	} `json:"delta"`
 	Usage *struct {
-		InputTokens  int64 `json:"inputTokens"`
-		OutputTokens int64 `json:"outputTokens"`
-		TotalTokens  int64 `json:"totalTokens"`
+		InputTokens      int64 `json:"inputTokens"`
+		OutputTokens     int64 `json:"outputTokens"`
+		TotalTokens      int64 `json:"totalTokens"`
+		CacheReadTokens  int64 `json:"cacheReadInputTokens"`
+		CacheWriteTokens int64 `json:"cacheWriteInputTokens"`
 	} `json:"usage"`
 }
 
@@ -105,6 +108,12 @@ func applyConverseStreamEvent(eventType string, payload []byte, usage *llm.Usage
 			if ev.Usage.TotalTokens > 0 {
 				usage.TotalTokens = ev.Usage.TotalTokens
 			}
+			if ev.Usage.CacheReadTokens > 0 {
+				usage.CachedInputTokens = ev.Usage.CacheReadTokens
+			}
+			if ev.Usage.CacheWriteTokens > 0 {
+				usage.CacheCreationTokens = ev.Usage.CacheWriteTokens
+			}
 		}
 	}
 }

proxy/internal/middleware/builtin/llm_response_parser/streaming_bedrock_test.go

@@ -66,6 +66,24 @@ func TestAccumulateBedrockStream_Converse(t *testing.T) {
 	require.Equal(t, "pong", completion, "converse text deltas concatenated")
 }
 
+// The converse-stream metadata frame's camelCase cache fields must reach the billed cache buckets.
+func TestAccumulateBedrockStream_ConverseCacheBuckets(t *testing.T) {
+	var body bytes.Buffer
+	body.Write(bedrockFrame(t, "contentBlockDelta", mustJSON(t, map[string]any{"delta": map[string]any{"text": "pong"}})))
+	body.Write(bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{
+		"inputTokens": 11, "outputTokens": 3, "totalTokens": 30,
+		"cacheReadInputTokens": 7, "cacheWriteInputTokens": 9,
+	}})))
+
+	usage, completion := accumulateBedrockStream(body.Bytes())
+	require.Equal(t, int64(11), usage.InputTokens, "input tokens from metadata frame")
+	require.Equal(t, int64(3), usage.OutputTokens, "output tokens from metadata frame")
+	require.Equal(t, int64(7), usage.CachedInputTokens, "cache-read tokens from metadata frame")
+	require.Equal(t, int64(9), usage.CacheCreationTokens, "cache-write tokens from metadata frame")
+	require.Equal(t, int64(30), usage.TotalTokens, "provider-reported total wins")
+	require.Equal(t, "pong", completion)
+}
+
 func TestAccumulateBedrockStream_Truncated(t *testing.T) {
 	// A body cut mid-frame must not panic; partial usage is returned.
 	full := bedrockFrame(t, "metadata", mustJSON(t, map[string]any{"usage": map[string]any{"inputTokens": 11, "outputTokens": 3}}))

proxy/internal/middleware/keys.go

@@ -75,8 +75,19 @@ const (
 	KeyLLMAttributionGroupID = "llm.attribution_group_id"
 	KeyLLMAttributionWindowS = "llm.attribution_window_seconds"
 
-	// Cost metering (emitted by cost_meter).
-	KeyCostUSDTotal = "cost.usd_total"
+	// Cost metering (emitted by cost_meter). The four per-bucket keys are the
+	// base of the breakdown — one per token bucket the provider bills
+	// separately — and the two aggregates below are derived from them:
+	// usd_total is their sum, usd_cache is cached_input + cache_creation.
+	KeyCostUSDInput = "cost.usd_input"
+	// KeyCostUSDCachedInput is the cost of the cache-read bucket (Anthropic cache_read; OpenAI's discounted cached subset of input).
+	KeyCostUSDCachedInput = "cost.usd_cached_input"
+	// KeyCostUSDCacheCreation is the cost of the cache-write bucket. Zero for providers without one.
+	KeyCostUSDCacheCreation = "cost.usd_cache_creation"
+	KeyCostUSDOutput        = "cost.usd_output"
+	KeyCostUSDTotal         = "cost.usd_total"
+	// KeyCostUSDCache is the portion of cost.usd_total billed for prompt-cache buckets (cache read/creation, or OpenAI's cached input subset).
+	KeyCostUSDCache = "cost.usd_cache"
 	KeyCostSkipped  = "cost.skipped"
 
 	// Framework-emitted error markers. Use the mw.<id>.* prefix to

shared/management/http/api/openapi.yml

@@ -5822,13 +5822,48 @@ components:
         total_tokens:
           type: integer
           format: int64
-          description: Total tokens consumed.
+          description: Total tokens consumed, including prompt-cache tokens.
           example: 1840
+        cached_input_tokens:
+          type: integer
+          format: int64
+          description: Input tokens read from the provider's prompt cache. Additive to input_tokens for Anthropic-shape providers; a subset of input_tokens for OpenAI.
+          example: 0
+        cache_creation_tokens:
+          type: integer
+          format: int64
+          description: Input tokens written to the provider's prompt cache. Zero for providers without a cache-write bucket.
+          example: 30528
         cost_usd:
           type: number
           format: double
           description: Estimated USD cost of the request.
           example: 0.0231
+        input_cost_usd:
+          type: number
+          format: double
+          description: Cost of the non-cached input tokens. Base component of cost_usd.
+          example: 0.0048
+        cached_input_cost_usd:
+          type: number
+          format: double
+          description: Cost of the prompt-cache read tokens. Base component of cost_usd, and part of cache_cost_usd.
+          example: 0.0015
+        cache_creation_cost_usd:
+          type: number
+          format: double
+          description: Cost of the prompt-cache write tokens. Base component of cost_usd, and part of cache_cost_usd.
+          example: 0.1130
+        output_cost_usd:
+          type: number
+          format: double
+          description: Cost of the output tokens. Base component of cost_usd.
+          example: 0.0038
+        cache_cost_usd:
+          type: number
+          format: double
+          description: Portion of cost_usd billed for prompt-cache usage.
+          example: 0.1145
         stream:
           type: boolean
           description: Whether the request was a streaming completion.
@@ -5852,7 +5887,14 @@ components:
         - input_tokens
         - output_tokens
         - total_tokens
+        - cached_input_tokens
+        - cache_creation_tokens
+        - input_cost_usd
+        - cached_input_cost_usd
+        - cache_creation_cost_usd
+        - output_cost_usd
         - cost_usd
+        - cache_cost_usd
     AgentNetworkAccessLogsResponse:
       type: object
       properties:
@@ -5926,13 +5968,48 @@ components:
         total_tokens:
           type: integer
           format: int64
-          description: Total tokens across the session.
+          description: Total tokens across the session, including prompt-cache tokens.
           example: 12880
+        cached_input_tokens:
+          type: integer
+          format: int64
+          description: Total prompt-cache read tokens across the session.
+          example: 0
+        cache_creation_tokens:
+          type: integer
+          format: int64
+          description: Total prompt-cache write tokens across the session.
+          example: 30528
         cost_usd:
           type: number
           format: double
           description: Total estimated USD cost across the session.
           example: 0.1617
+        input_cost_usd:
+          type: number
+          format: double
+          description: Total cost of non-cached input tokens across the session.
+          example: 0.0210
+        cached_input_cost_usd:
+          type: number
+          format: double
+          description: Total cost of prompt-cache read tokens across the session.
+          example: 0.0015
+        cache_creation_cost_usd:
+          type: number
+          format: double
+          description: Total cost of prompt-cache write tokens across the session.
+          example: 0.1130
+        output_cost_usd:
+          type: number
+          format: double
+          description: Total cost of output tokens across the session.
+          example: 0.0262
+        cache_cost_usd:
+          type: number
+          format: double
+          description: Portion of cost_usd billed for prompt-cache usage across the session.
+          example: 0.1145
         providers:
           type: array
           items:
@@ -5959,7 +6036,14 @@ components:
         - input_tokens
         - output_tokens
         - total_tokens
+        - cached_input_tokens
+        - cache_creation_tokens
+        - input_cost_usd
+        - cached_input_cost_usd
+        - cache_creation_cost_usd
+        - output_cost_usd
         - cost_usd
+        - cache_cost_usd
         - decision
         - entries
     AgentNetworkAccessLogSessionsResponse:
@@ -6013,19 +6097,61 @@ components:
         total_tokens:
           type: integer
           format: int64
-          description: Total tokens in the bucket.
+          description: Total tokens in the bucket, including prompt-cache tokens.
           example: 184000
+        cached_input_tokens:
+          type: integer
+          format: int64
+          description: Total prompt-cache read tokens in the bucket.
+          example: 20000
+        cache_creation_tokens:
+          type: integer
+          format: int64
+          description: Total prompt-cache write tokens in the bucket.
+          example: 45000
+        input_cost_usd:
+          type: number
+          format: double
+          description: Total cost of non-cached input tokens in the bucket.
+          example: 1.12
+        cached_input_cost_usd:
+          type: number
+          format: double
+          description: Total cost of prompt-cache read tokens in the bucket.
+          example: 0.06
+        cache_creation_cost_usd:
+          type: number
+          format: double
+          description: Total cost of prompt-cache write tokens in the bucket.
+          example: 0.36
+        output_cost_usd:
+          type: number
+          format: double
+          description: Total cost of output tokens in the bucket.
+          example: 0.77
         cost_usd:
           type: number
           format: double
           description: Total estimated USD spend in the bucket.
           example: 2.31
+        cache_cost_usd:
+          type: number
+          format: double
+          description: Portion of cost_usd billed for prompt-cache usage in the bucket.
+          example: 0.42
       required:
         - period_start
         - input_tokens
         - output_tokens
         - total_tokens
+        - cached_input_tokens
+        - cache_creation_tokens
+        - input_cost_usd
+        - cached_input_cost_usd
+        - cache_creation_cost_usd
+        - output_cost_usd
         - cost_usd
+        - cache_cost_usd
     AgentNetworkSettings:
       type: object
       description: Per-account Agent Network gateway settings. One row per account; cluster and subdomain are auto-assigned on first provider create and immutable thereafter.

shared/management/http/api/types.gen.go

@@ -1732,6 +1732,21 @@ type AccountSettings struct {
 
 // AgentNetworkAccessLog One per-request agent-network (LLM) access log entry with flattened, queryable LLM dimensions.
 type AgentNetworkAccessLog struct {
+	// CacheCostUsd Portion of cost_usd billed for prompt-cache usage.
+	CacheCostUsd float64 `json:"cache_cost_usd"`
+
+	// CacheCreationCostUsd Cost of the prompt-cache write tokens. Base component of cost_usd, and part of cache_cost_usd.
+	CacheCreationCostUsd float64 `json:"cache_creation_cost_usd"`
+
+	// CacheCreationTokens Input tokens written to the provider's prompt cache. Zero for providers without a cache-write bucket.
+	CacheCreationTokens int64 `json:"cache_creation_tokens"`
+
+	// CachedInputCostUsd Cost of the prompt-cache read tokens. Base component of cost_usd, and part of cache_cost_usd.
+	CachedInputCostUsd float64 `json:"cached_input_cost_usd"`
+
+	// CachedInputTokens Input tokens read from the provider's prompt cache. Additive to input_tokens for Anthropic-shape providers; a subset of input_tokens for OpenAI.
+	CachedInputTokens int64 `json:"cached_input_tokens"`
+
 	// CostUsd Estimated USD cost of the request.
 	CostUsd float64 `json:"cost_usd"`
 
@@ -1753,6 +1768,9 @@ type AgentNetworkAccessLog struct {
 	// Id Unique identifier for the access log entry.
 	Id string `json:"id"`
 
+	// InputCostUsd Cost of the non-cached input tokens. Base component of cost_usd.
+	InputCostUsd float64 `json:"input_cost_usd"`
+
 	// InputTokens Input (prompt) tokens consumed.
 	InputTokens int64 `json:"input_tokens"`
 
@@ -1762,6 +1780,9 @@ type AgentNetworkAccessLog struct {
 	// Model Requested LLM model.
 	Model *string `json:"model,omitempty"`
 
+	// OutputCostUsd Cost of the output tokens. Base component of cost_usd.
+	OutputCostUsd float64 `json:"output_cost_usd"`
+
 	// OutputTokens Output (completion) tokens produced.
 	OutputTokens int64 `json:"output_tokens"`
 
@@ -1801,7 +1822,7 @@ type AgentNetworkAccessLog struct {
 	// Timestamp Timestamp when the request was made.
 	Timestamp time.Time `json:"timestamp"`
 
-	// TotalTokens Total tokens consumed.
+	// TotalTokens Total tokens consumed, including prompt-cache tokens.
 	TotalTokens int64 `json:"total_tokens"`
 
 	// UserId NetBird user id of the authenticated caller, if applicable.
@@ -1810,6 +1831,21 @@ type AgentNetworkAccessLog struct {
 
 // AgentNetworkAccessLogSession A session-grouped view of agent-network access logs — all requests sharing a session id (or a single session-less request) folded into one summary plus its ordered entries.
 type AgentNetworkAccessLogSession struct {
+	// CacheCostUsd Portion of cost_usd billed for prompt-cache usage across the session.
+	CacheCostUsd float64 `json:"cache_cost_usd"`
+
+	// CacheCreationCostUsd Total cost of prompt-cache write tokens across the session.
+	CacheCreationCostUsd float64 `json:"cache_creation_cost_usd"`
+
+	// CacheCreationTokens Total prompt-cache write tokens across the session.
+	CacheCreationTokens int64 `json:"cache_creation_tokens"`
+
+	// CachedInputCostUsd Total cost of prompt-cache read tokens across the session.
+	CachedInputCostUsd float64 `json:"cached_input_cost_usd"`
+
+	// CachedInputTokens Total prompt-cache read tokens across the session.
+	CachedInputTokens int64 `json:"cached_input_tokens"`
+
 	// CostUsd Total estimated USD cost across the session.
 	CostUsd float64 `json:"cost_usd"`
 
@@ -1825,12 +1861,18 @@ type AgentNetworkAccessLogSession struct {
 	// GroupIds Union of the authorising group ids across the session's entries.
 	GroupIds *[]string `json:"group_ids,omitempty"`
 
+	// InputCostUsd Total cost of non-cached input tokens across the session.
+	InputCostUsd float64 `json:"input_cost_usd"`
+
 	// InputTokens Total input (prompt) tokens across the session.
 	InputTokens int64 `json:"input_tokens"`
 
 	// Models Distinct models seen in the session.
 	Models *[]string `json:"models,omitempty"`
 
+	// OutputCostUsd Total cost of output tokens across the session.
+	OutputCostUsd float64 `json:"output_cost_usd"`
+
 	// OutputTokens Total output (completion) tokens across the session.
 	OutputTokens int64 `json:"output_tokens"`
 
@@ -1846,7 +1888,7 @@ type AgentNetworkAccessLogSession struct {
 	// StartedAt Timestamp of the session's earliest request.
 	StartedAt time.Time `json:"started_at"`
 
-	// TotalTokens Total tokens across the session.
+	// TotalTokens Total tokens across the session, including prompt-cache tokens.
 	TotalTokens int64 `json:"total_tokens"`
 
 	// UserId NetBird user id of the session's caller.
@@ -2347,19 +2389,40 @@ type AgentNetworkSettingsRequest struct {
 
 // AgentNetworkUsageBucket One aggregated agent-network usage time bucket (UTC). The bucket width is set by the request's granularity.
 type AgentNetworkUsageBucket struct {
+	// CacheCostUsd Portion of cost_usd billed for prompt-cache usage in the bucket.
+	CacheCostUsd float64 `json:"cache_cost_usd"`
+
+	// CacheCreationCostUsd Total cost of prompt-cache write tokens in the bucket.
+	CacheCreationCostUsd float64 `json:"cache_creation_cost_usd"`
+
+	// CacheCreationTokens Total prompt-cache write tokens in the bucket.
+	CacheCreationTokens int64 `json:"cache_creation_tokens"`
+
+	// CachedInputCostUsd Total cost of prompt-cache read tokens in the bucket.
+	CachedInputCostUsd float64 `json:"cached_input_cost_usd"`
+
+	// CachedInputTokens Total prompt-cache read tokens in the bucket.
+	CachedInputTokens int64 `json:"cached_input_tokens"`
+
 	// CostUsd Total estimated USD spend in the bucket.
 	CostUsd float64 `json:"cost_usd"`
 
+	// InputCostUsd Total cost of non-cached input tokens in the bucket.
+	InputCostUsd float64 `json:"input_cost_usd"`
+
 	// InputTokens Total input (prompt) tokens in the bucket.
 	InputTokens int64 `json:"input_tokens"`
 
+	// OutputCostUsd Total cost of output tokens in the bucket.
+	OutputCostUsd float64 `json:"output_cost_usd"`
+
 	// OutputTokens Total output (completion) tokens in the bucket.
 	OutputTokens int64 `json:"output_tokens"`
 
 	// PeriodStart Start of the bucket in YYYY-MM-DD (UTC) — the day, the week start (Monday), or the month start, depending on granularity.
 	PeriodStart string `json:"period_start"`
 
-	// TotalTokens Total tokens in the bucket.
+	// TotalTokens Total tokens in the bucket, including prompt-cache tokens.
 	TotalTokens int64 `json:"total_tokens"`
 }