merge: expose token-accounting provenance

admin.go

@@ -303,9 +303,20 @@ type adminProviderView struct {
 	// declared after the scalar block as costing extra GC pointer-scan
 	// bytes. See timeout.go's watchdogBody for the identical convention.
 	Latency map[string]adminLatencyView `json:"latency,omitempty"`
-	Type    string                      `json:"type"`
-	BaseURL string                      `json:"baseUrl"`
-	LastErr string                      `json:"lastErr,omitempty"`
+	// Provenance summarizes non-reported usage accounting for this
+	// provider (feat: expose token-accounting provenance) — keyed by
+	// provenance kind, "estimated" or "unbilled" only (see
+	// adminProvenanceView's own doc comment for why "reported" is never a
+	// key here). Sourced from the SAME in-process g.provenance
+	// accumulator /metrics reads (metrics.go). nil (omitted) for a
+	// provider with no estimated/unbilled outcome yet, or for a
+	// deployment with metrics collection gated off entirely — mirroring
+	// Latency's identical "nil means no observations yet" convention
+	// immediately above.
+	Provenance map[string]adminProvenanceView `json:"provenance,omitempty"`
+	Type       string                         `json:"type"`
+	BaseURL    string                         `json:"baseUrl"`
+	LastErr    string                         `json:"lastErr,omitempty"`
 	// HealthState is this provider's discovery circuit breaker state
 	// (feat/provider-health): "closed" (normal), "open" (backing off
 	// after repeated discovery failures — maybeRefresh skips it until
@@ -417,6 +428,48 @@ func buildAdminLatencyViews(snaps []latencySnapshot) map[string]map[string]admin
 	return out
 }
 
+// adminProvenanceView is one provider's one non-reported provenance
+// kind's compact usage-accounting summary (feat: expose token-accounting
+// provenance) — adminProviderView.Provenance's value type, keyed by
+// provenance kind ("estimated" or "unbilled" — provenanceEstimated/
+// provenanceUnbilled, metrics.go) in adminProviderView.Provenance's outer
+// map. "reported" is deliberately never a key here: it is the default,
+// healthy case AttemptsDay/FailuresDay above already partially describe,
+// and duplicating its own count would not answer any operator question
+// this compact surface (buildAdminLatencyViews' own "compact, no
+// per-model breakdown" ruling, GET /admin/api/overview's ~113KB/5s-poll
+// budget) is not already better placed to answer — the full reported/
+// estimated/unbilled breakdown, by request AND by token, lives on
+// /metrics (llmgateway_usage_provenance_requests_total/-_tokens_total)
+// for an operator who wants the precise fraction.
+type adminProvenanceView struct {
+	Requests int64 `json:"requests"`
+	Tokens   int64 `json:"tokens"`
+}
+
+// buildAdminProvenanceViews groups snaps (g.provenance.snapshot(),
+// metrics.go) by provider, discarding the "reported" provenance entirely
+// (adminProvenanceView's own doc comment) — mirroring
+// buildAdminLatencyViews' identical "compact, provider-keyed outer map"
+// shape immediately above. A provider with no estimated/unbilled
+// observations at all is simply absent from the outer map, so
+// adminOverviewResponse's own lookup (buildAdminOverview, below) yields
+// nil for it — adminProviderView.Provenance's own documented "nil means
+// no estimated/unbilled accounting yet".
+func buildAdminProvenanceViews(snaps []provenanceSnapshot) map[string]map[string]adminProvenanceView {
+	out := make(map[string]map[string]adminProvenanceView)
+	for _, s := range snaps {
+		if s.key.provenance == provenanceReported {
+			continue
+		}
+		if out[s.key.provider] == nil {
+			out[s.key.provider] = make(map[string]adminProvenanceView)
+		}
+		out[s.key.provider][s.key.provenance] = adminProvenanceView{Requests: s.requests, Tokens: s.tokens}
+	}
+	return out
+}
+
 // adminModelMetaView is one upstream model's resolved metadata (feature
 // v0.23: context window, per-token cost) within its provider —
 // adminProviderView.ModelMeta's value type. Every field is a pointer,
@@ -559,6 +612,10 @@ func (g *Gateway) buildAdminOverview() adminOverviewResponse {
 	// for the whole response, mirroring providerUsage's own one-batched-
 	// read-for-every-provider shape above, not a per-provider read.
 	latencyViews := buildAdminLatencyViews(g.latency.snapshot())
+	// feat: expose token-accounting provenance — one g.provenance.
+	// snapshot() call for the whole response, mirroring g.latency.
+	// snapshot()'s own one-batched-read-for-every-provider shape above.
+	provenanceViews := buildAdminProvenanceViews(g.provenance.snapshot())
 
 	providers := make([]adminProviderView, len(snaps))
 	mi := 0
@@ -593,6 +650,7 @@ func (g *Gateway) buildAdminOverview() adminOverviewResponse {
 			ModelRates:       modelRates,
 			ModelMeta:        modelMeta,
 			Latency:          latencyViews[s.name],
+			Provenance:       provenanceViews[s.name],
 		}
 	}
 

admin_test.go

@@ -631,6 +631,78 @@ func TestAdminOverview_LatencySummary_AvgReflectsSeededObservations(t *testing.T
 	}
 }
 
+// TestAdminOverview_ProvenanceSummary_ExcludesReportedIncludesEstimatedAndUnbilled
+// proves buildAdminProvenanceViews' own filtering (admin.go, feat:
+// expose token-accounting provenance): "reported" observations never
+// surface on this compact admin summary (the operator already sees
+// AttemptsDay/FailuresDay for the healthy default), while "estimated"
+// and "unbilled" both do, as DISTINCT keys — never merged.
+//
+// MUTATION VERIFIED: removing the `if s.key.provenance ==
+// provenanceReported { continue }` guard from buildAdminProvenanceViews
+// (admin.go) made this test fail — zeta.Provenance gained a leaked
+// "reported" entry (Requests=1, Tokens=15) that must never appear here.
+// Reverted before committing.
+func TestAdminOverview_ProvenanceSummary_ExcludesReportedIncludesEstimatedAndUnbilled(t *testing.T) {
+	t.Parallel()
+	cfg := newAdminTestConfig()
+	h, gw := newAdminGatewayHandle(t, cfg)
+
+	// zeta: one reported observation (must never surface below) and one
+	// estimated observation.
+	gw.provenance.record("zeta", provenanceReported, 15)
+	gw.provenance.record("zeta", provenanceEstimated, 7)
+
+	// alpha: two unbilled observations, seeded separately to prove
+	// requests accumulates across calls rather than overwriting.
+	gw.provenance.record("alpha", provenanceUnbilled, 0)
+	gw.provenance.record("alpha", provenanceUnbilled, 0)
+
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, adminRequest(http.MethodGet, adminOverviewPath, "sk-admin1"))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200, body=%s", rec.Code, rec.Body.String())
+	}
+	var got adminOverviewResponse
+	if err := json.Unmarshal(rec.Body.Bytes(), &got); err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	var zeta, alpha *adminProviderView
+	for i := range got.Providers {
+		switch got.Providers[i].Name {
+		case "zeta":
+			zeta = &got.Providers[i]
+		case "alpha":
+			alpha = &got.Providers[i]
+		}
+	}
+	if zeta == nil || alpha == nil {
+		t.Fatalf("providers = %+v, want entries named zeta and alpha", got.Providers)
+	}
+
+	if _, ok := zeta.Provenance[provenanceReported]; ok {
+		t.Errorf(`zeta.Provenance has a %q entry: %+v — "reported" must never surface on this compact admin summary`, provenanceReported, zeta.Provenance)
+	}
+	est, ok := zeta.Provenance[provenanceEstimated]
+	if !ok {
+		t.Fatalf("zeta.Provenance missing %q entry: %+v", provenanceEstimated, zeta.Provenance)
+	}
+	if est.Requests != 1 || est.Tokens != 7 {
+		t.Errorf("zeta estimated = %+v, want {Requests:1 Tokens:7}", est)
+	}
+
+	unb, ok := alpha.Provenance[provenanceUnbilled]
+	if !ok {
+		t.Fatalf("alpha.Provenance missing %q entry: %+v", provenanceUnbilled, alpha.Provenance)
+	}
+	if unb.Requests != 2 || unb.Tokens != 0 {
+		t.Errorf("alpha unbilled = %+v, want {Requests:2 Tokens:0}", unb)
+	}
+	if _, ok := alpha.Provenance[provenanceEstimated]; ok {
+		t.Errorf(`alpha.Provenance has a %q entry despite only unbilled observations being seeded: %+v — estimated and unbilled must never conflate`, provenanceEstimated, alpha.Provenance)
+	}
+}
+
 // --- usage: math against seeded counters ---
 
 func TestAdminUsage_MathAgainstSeededCounters(t *testing.T) {

llmgateway.go

@@ -597,7 +597,17 @@ type Gateway struct {
 	// identical "safe even off a bare &Gateway{} literal" convention
 	// above.
 	latency *latencyStore
-	name    string
+	// provenance is the in-process usage-accounting-provenance accumulator
+	// (feat: expose token-accounting provenance, metrics.go's
+	// provenanceStore) — recordUsageProvenance's only write target, and
+	// writeProvenanceMetrics/buildAdminProvenanceViews' (admin.go) own
+	// read source. Always constructed, here, by newGateway, even when
+	// Config.Metrics is nil or disabled, mirroring g.latency's own
+	// identical "always-present, usually-empty, nil-receiver-safe"
+	// convention immediately above — see that field's own doc comment for
+	// the full reasoning, which applies here unchanged.
+	provenance *provenanceStore
+	name       string
 	// failoverLogGate rate-limits runMeteredCall's generic "failing over"
 	// log line (routes_unified.go) to once per storeErrorLogEvery
 	// (adversarial-review fix, F10) — reuses auth.go's own logGate type,
@@ -744,6 +754,10 @@ func newGateway(ctx context.Context, next http.Handler, config *Config, name str
 	// comment above for why an always-present, usually-empty store is the
 	// right default rather than a conditional nil.
 	g.latency = newLatencyStore()
+	// feat: expose token-accounting provenance — always constructed, here,
+	// regardless of whether Config.Metrics is nil/disabled, mirroring
+	// g.latency's own identical "always-present" ruling immediately above.
+	g.provenance = newProvenanceStore()
 
 	// A malformed Config.Metrics.AllowedCIDRs entry is a constructor
 	// error (validate-at-construction), not a silently-ignored allowlist

metrics.go

@@ -431,10 +431,13 @@ func (m *metricWriter) histogram(name string, labels []metricLabel, buckets []in
 const aggregationNoteStoreBacked = " AGGREGATION ACROSS REPLICAS: with Redis configured, this value is the gateway-wide total read from shared limit state and is IDENTICAL on every replica's own /metrics response — aggregate with max(), never sum(), or a multi-replica deployment overcounts by the replica count. Without Redis (the in-memory fallback), this SAME family becomes genuinely per-process instead, and sum() becomes the correct aggregation. Know which one you are running."
 
 // aggregationNotePerProcess is appended to llmgateway_rate_limit_
-// rejections_total's HELP text: limiter.rejections (limits.go) is a
-// bare in-process map, never written to or read from Redis regardless
-// of Config.Redis — unlike every store-backed family above, this one
-// never flips, and sum() is always the right cross-replica aggregation.
+// rejections_total's HELP text, and to both usage-provenance counter
+// families' HELP text below (llmgateway_usage_provenance_requests_total/
+// -_tokens_total, feat: expose token-accounting provenance): each is a
+// bare in-process store — limiter.rejections and g.provenance
+// respectively — never written to or read from Redis regardless of
+// Config.Redis. Unlike every store-backed family above, none of these
+// ever flips, and sum() is always the right cross-replica aggregation.
 const aggregationNotePerProcess = " AGGREGATION ACROSS REPLICAS: this counter lives only in this process's memory, never in Redis, regardless of Config.Redis — aggregate with sum(), which is always correct for it, unlike the store-backed families above."
 
 // aggregationNoteProviderHealthy is appended to llmgateway_provider_
@@ -468,6 +471,7 @@ func (g *Gateway) renderMetrics() []byte {
 	g.writeUsageMetrics(&m)
 	g.writeProviderMetrics(&m)
 	g.writeLatencyMetrics(&m)
+	g.writeProvenanceMetrics(&m)
 	g.writeRejectionMetrics(&m)
 	g.writeStoreHealthMetrics(&m)
 	return m.buf.Bytes()
@@ -961,6 +965,164 @@ func (g *Gateway) writeLatencyMetrics(m *metricWriter) {
 	}
 }
 
+// --- usage provenance (feat: expose token-accounting provenance) ---
+//
+// Token accounting has three provenances (usage.estimated's own doc
+// comment, limits.go), and before this section nothing on any read
+// surface distinguished them: REPORTED (the provider returned real
+// usage — the normal, accurate case), ESTIMATED (a non-streaming
+// response carried none, so prompt is substituted with
+// ceil(len(body)/4) and completion is billed as zero — runUnified's own
+// recording call site, routes_unified.go), and UNBILLED (a streaming
+// response carried none, so the request is counted but zero tokens are
+// billed). This section adds the metric surface only; billing itself is
+// unchanged — see runUnified's own doc comment at its recording call
+// site for the guarantee that no value recorded here ever feeds back
+// into what gets billed.
+//
+// ESTIMATED and UNBILLED are deliberately never merged into one "not
+// reported" signal: estimated substitutes an approximation (a nonzero
+// prompt count is still billed), unbilled charges nothing at all. A
+// provider silently ignoring stream_options.include_usage serves
+// completions entirely free against every configured budget, and that
+// is the one fact this feature exists to make visible — collapsing the
+// two would hide it again.
+
+const (
+	provenanceReported  = "reported"
+	provenanceEstimated = "estimated"
+	provenanceUnbilled  = "unbilled"
+)
+
+// provenanceKey identifies one (provider, provenance) series pair
+// provenanceStore accumulates under. provider is always one of this
+// deployment's configured provider names (config-bounded, never user
+// input, exactly like latencyKey.provider above), and provenance is
+// always one of the three package-level constants above — the total key
+// space is bounded by (configured providers) x 3, so no rotation/cap
+// logic is needed here, unlike rejectionCounter's own user/group-name
+// keys (limits.go), which ARE operator/end-user controlled and can churn
+// without bound.
+type provenanceKey struct {
+	provider   string
+	provenance string
+}
+
+// provenanceCounts is one provenanceKey's accumulated totals: requests
+// (how many accounted candidate outcomes were recorded under this
+// provenance) and tokens (their summed result.total() — the SAME number
+// runUnified already passed to accounting, never a second count),
+// answering the task brief's "what fraction of billed tokens came from
+// real reported usage" question directly: sum(tokens where provenance ==
+// reported) / sum(tokens across every provenance).
+type provenanceCounts struct {
+	requests int64
+	tokens   int64
+}
+
+// provenanceStore is the Gateway's in-process, per-replica usage-
+// provenance accumulator — mirrors latencyStore's own shape and
+// nil-receiver-safe convention above for the identical reason: a Gateway
+// assembled directly as a bare &Gateway{} literal (bypassing newGateway,
+// as a few older tests do) degrades to "no provenance observed" rather
+// than a nil-pointer panic.
+type provenanceStore struct {
+	data map[provenanceKey]*provenanceCounts
+	mu   sync.Mutex
+}
+
+// newProvenanceStore returns an empty, ready-to-use provenanceStore.
+func newProvenanceStore() *provenanceStore {
+	return &provenanceStore{data: make(map[provenanceKey]*provenanceCounts)}
+}
+
+// record accumulates one candidate outcome under (provider, provenance):
+// one request, plus tokens — allocating a fresh provenanceCounts the
+// first time key is seen.
+func (s *provenanceStore) record(provider, provenance string, tokens int64) {
+	if s == nil {
+		return
+	}
+	key := provenanceKey{provider: provider, provenance: provenance}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	c := s.data[key]
+	if c == nil {
+		c = &provenanceCounts{}
+		s.data[key] = c
+	}
+	c.requests++
+	c.tokens += tokens
+}
+
+// provenanceSnapshot is one provenanceKey's fully-copied counts, safe to
+// read after provenanceStore.snapshot returns without holding its lock.
+type provenanceSnapshot struct {
+	key      provenanceKey
+	requests int64
+	tokens   int64
+}
+
+// snapshot returns a fully-copied view of every provenanceKey s
+// currently holds, safe to read without s's lock — writeProvenanceMetrics
+// and buildAdminProvenanceViews (admin.go) both read through this rather
+// than s.data directly.
+func (s *provenanceStore) snapshot() []provenanceSnapshot {
+	if s == nil {
+		return nil
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := make([]provenanceSnapshot, 0, len(s.data))
+	for k, c := range s.data {
+		out = append(out, provenanceSnapshot{key: k, requests: c.requests, tokens: c.tokens})
+	}
+	return out
+}
+
+// recordUsageProvenance accumulates one candidate outcome into
+// g.provenance — runUnified's (routes_unified.go) only write path into
+// it, reached only for a candidate whose own response actually completed
+// (callErr == nil), gated on metricsEnabled exactly like
+// withLatencyRecorder's own wiring (routes_unified.go): a deployment
+// with metrics off calls this exactly zero times.
+func (g *Gateway) recordUsageProvenance(provider, provenance string, tokens int64) {
+	g.provenance.record(provider, provenance, tokens)
+}
+
+// writeProvenanceMetrics emits llmgateway_usage_provenance_requests_total
+// and llmgateway_usage_provenance_tokens_total — this feature's own two
+// counter families. Absent entirely (no # HELP/TYPE lines emitted) when
+// g.provenance has recorded nothing yet, mirroring writeLatencyMetrics'
+// own "no series until the first observation" convention above.
+//
+// snaps is sorted (provider, then provenance) purely for a deterministic
+// scrape body across repeated scrapes of identical state, matching
+// writeLatencyMetrics' own reasoning.
+func (g *Gateway) writeProvenanceMetrics(m *metricWriter) {
+	snaps := g.provenance.snapshot()
+	if len(snaps) == 0 {
+		return
+	}
+	sort.Slice(snaps, func(i, j int) bool {
+		a, b := snaps[i].key, snaps[j].key
+		if a.provider != b.provider {
+			return a.provider < b.provider
+		}
+		return a.provenance < b.provenance
+	})
+
+	m.family("llmgateway_usage_provenance_requests_total", "counter",
+		"Total accounted candidate outcomes (a successful upstream response, callErr == nil, that reached usage accounting in runUnified, routes_unified.go), by provider and provenance. provenance is \"reported\" (the provider returned real usage), \"estimated\" (a non-streaming response carried none, so prompt was substituted with ceil(len(body)/4) and completion billed as zero — usage.estimated, limits.go), or \"unbilled\" (a streaming response carried none, so the request is counted but zero tokens are billed). This family only reports how an already-billed number was arrived at; it never changes what gets billed."+aggregationNotePerProcess)
+	m.family("llmgateway_usage_provenance_tokens_total", "counter",
+		"Total tokens billed under each provenance (the SAME result.total() runUnified already passed to accounting, routes_unified.go — never a second, independent count), by provider and provenance. sum(...{provenance=\"reported\"}) / sum(...) across every provenance is the fraction of billed tokens that came from real provider-reported usage."+aggregationNotePerProcess)
+	for _, s := range snaps {
+		labels := []metricLabel{{"provider", s.key.provider}, {"provenance", s.key.provenance}}
+		m.sampleInt("llmgateway_usage_provenance_requests_total", labels, s.requests)
+		m.sampleInt("llmgateway_usage_provenance_tokens_total", labels, s.tokens)
+	}
+}
+
 // writeRejectionMetrics emits llmgateway_rate_limit_rejections_total from
 // limiter.rejections (limits.go) — the one metric family in this file
 // with no existing counterStore-backed accounting to read: checkAndCount

metrics_test.go

@@ -493,6 +493,8 @@ func TestMetrics_OutputParsesAsValidExposition(t *testing.T) {
 		"llmgateway_limit_store_up",
 		"llmgateway_upstream_ttfb_seconds",
 		"llmgateway_upstream_duration_seconds",
+		"llmgateway_usage_provenance_requests_total",
+		"llmgateway_usage_provenance_tokens_total",
 	}
 	for _, name := range wantFamilies {
 		if _, ok := types[name]; !ok {
@@ -1334,6 +1336,123 @@ func TestMetrics_LatencyHistogram_StreamingAndNonStreaming_DistinctLabelSets(t *
 	}
 }
 
+// --- usage provenance (feat: expose token-accounting provenance) ---
+
+// TestMetrics_UsageProvenance_EstimatedAndUnbilled_DistinctSeries proves
+// this feature's central correctness requirement end to end, through the
+// real /metrics exposition: a non-streaming zero-usage response
+// (estimated) and a streaming zero-usage response (unbilled) against the
+// SAME provider render as two DISTINCT series, never merged into one
+// "not reported" signal. Distinct provider names would make this trivial
+// (two unrelated series); using the identical provider for both is the
+// stronger proof — the two provenances would collide into a single
+// series if the recording code ever conflated them, exactly the
+// regression this test exists to catch.
+//
+// MUTATION VERIFIED: changing routes_unified.go's recording switch case
+// `case streaming && result.total() == 0:` to `case result.estimated:`
+// (merging the unbilled classification into the estimated one) made this
+// test fail: the "unbilled" series vanished from /metrics entirely and
+// the "estimated" series' requests count doubled to 2, proving this test
+// actually catches the two provenances collapsing into one — the exact
+// regression this feature exists to prevent. Reverted before committing.
+func TestMetrics_UsageProvenance_EstimatedAndUnbilled_DistinctSeries(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]any
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		if streaming, _ := body["stream"].(bool); streaming {
+			w.Header().Set("Content-Type", "text/event-stream")
+			w.WriteHeader(http.StatusOK)
+			fl := w.(http.Flusher)
+			_, _ = fmt.Fprint(w, "data: {\"id\":\"c\",\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n")
+			fl.Flush()
+			_, _ = fmt.Fprint(w, "data: [DONE]\n\n") // no usage chunk: unbilled
+			fl.Flush()
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		// No "usage" field at all: estimated.
+		_, _ = w.Write([]byte(`{"id":"c1","object":"chat.completion","model":"gpt-test","choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}]}`))
+	}))
+	defer srv.Close()
+
+	cfg := newMetricsTestConfig()
+	cfg.Providers["openai"].BaseURL = srv.URL
+	h, _ := newMetricsGatewayHandle(t, cfg)
+
+	nonStreamBody := map[string]any{"model": "gpt-test", "messages": []any{map[string]any{"role": "user", "content": "hi"}}}
+	streamBody := map[string]any{"model": "gpt-test", "stream": true, "messages": []any{map[string]any{"role": "user", "content": "hi"}}}
+
+	rec1 := httptest.NewRecorder()
+	h.ServeHTTP(rec1, newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", nonStreamBody))
+	if rec1.Code != http.StatusOK {
+		t.Fatalf("non-streaming request status = %d, want 200, body=%s", rec1.Code, rec1.Body.String())
+	}
+	rec2 := httptest.NewRecorder()
+	h.ServeHTTP(rec2, newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", streamBody))
+	if rec2.Code != http.StatusOK {
+		t.Fatalf("streaming request status = %d, want 200, body=%s", rec2.Code, rec2.Body.String())
+	}
+
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, metricsRequest(http.MethodGet, metricsPathDefault, "sk-admin1"))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("GET /metrics status = %d, want 200, body=%s", rec.Code, rec.Body.String())
+	}
+	_, samples := parsePrometheusText(t, rec.Body.Bytes())
+
+	var estimatedReq, unbilledReq *promSample
+	for i := range samples {
+		s := &samples[i]
+		if s.name != "llmgateway_usage_provenance_requests_total" || s.labels["provider"] != "openai" {
+			continue
+		}
+		switch s.labels["provenance"] {
+		case provenanceEstimated:
+			estimatedReq = s
+		case provenanceUnbilled:
+			unbilledReq = s
+		}
+	}
+	if estimatedReq == nil {
+		t.Fatal(`no llmgateway_usage_provenance_requests_total{provider="openai",provenance="estimated"} sample — the non-streaming zero-usage request must land under "estimated"`)
+	}
+	if unbilledReq == nil {
+		t.Fatal(`no llmgateway_usage_provenance_requests_total{provider="openai",provenance="unbilled"} sample — the streaming zero-usage request must land under "unbilled", DISTINCT from "estimated"`)
+	}
+	if estimatedReq.value != "1" {
+		t.Errorf(`estimated requests = %s, want 1`, estimatedReq.value)
+	}
+	if unbilledReq.value != "1" {
+		t.Errorf(`unbilled requests = %s, want 1`, unbilledReq.value)
+	}
+
+	var estimatedTok, unbilledTok *promSample
+	for i := range samples {
+		s := &samples[i]
+		if s.name != "llmgateway_usage_provenance_tokens_total" || s.labels["provider"] != "openai" {
+			continue
+		}
+		switch s.labels["provenance"] {
+		case provenanceEstimated:
+			estimatedTok = s
+		case provenanceUnbilled:
+			unbilledTok = s
+		}
+	}
+	if estimatedTok == nil || unbilledTok == nil {
+		t.Fatalf("llmgateway_usage_provenance_tokens_total missing estimated/unbilled series: estimated=%v unbilled=%v", estimatedTok, unbilledTok)
+	}
+	if estimatedTok.value == "0" {
+		t.Error(`estimated tokens = 0, want > 0 — estimated substitutes a nonzero prompt count, unlike unbilled`)
+	}
+	if unbilledTok.value != "0" {
+		t.Errorf(`unbilled tokens = %s, want 0 — unbilled charges nothing at all, unlike estimated`, unbilledTok.value)
+	}
+}
+
 // TestRecordLatency_MetricsDisabled_NoObservationCollected proves the
 // "no collection cost, no series emitted" requirement at the
 // Gateway/collection level (not merely "the renderer skips an empty

routes_unified.go

@@ -555,6 +555,34 @@ func (g *Gateway) runMeteredCall(sw *statusTrackingWriter, r *http.Request, scop
 			g.logf("%s: usage for model %q logged as estimated (%d prompt tokens derived from request body size, not the provider's reported usage)", logPrefix, cand.canonical, result.prompt)
 		}
 
+		// Usage provenance (feat: expose token-accounting provenance),
+		// recorded ONLY for a candidate whose own response actually
+		// completed (callErr == nil): a candidate that failed over never
+		// bills more than a zero-total no-op (this function's own doc
+		// comment above), and counting THAT as "unbilled" would drown the
+		// one signal this feature exists to surface — a provider silently
+		// serving completions free against every budget — in ordinary
+		// failover noise. "unbilled" is re-derived from streaming &&
+		// result.total() == 0 rather than a separate flag carried out of
+		// the block above: result.prompt/result.estimated are only ever
+		// mutated by that block's non-streaming branch, so this condition
+		// is true here if and only if the streaming branch above ran.
+		// Gated on metricsEnabled exactly like withLatencyRecorder's own
+		// wiring above, so a deployment with metrics off pays nothing
+		// extra here. Observation only: nothing recorded here feeds back
+		// into result or the account call above, both of which already
+		// ran — see g.recordUsageProvenance's own doc comment (metrics.go).
+		if callErr == nil && metricsEnabled(g.cfg) {
+			provenance := provenanceReported
+			switch {
+			case result.estimated:
+				provenance = provenanceEstimated
+			case streaming && result.total() == 0:
+				provenance = provenanceUnbilled
+			}
+			g.recordUsageProvenance(providerName, provenance, result.total())
+		}
+
 		// A miss stores the response after everything above has already
 		// run — accounting must never be skipped or delayed waiting on a
 		// cache write. Only a genuine upstream 200 is stored (spec §2's

routes_unified_test.go

@@ -8,6 +8,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math"
 	"net"
 	"net/http"
 	"net/http/httptest"
@@ -228,6 +229,250 @@ func TestHandleChat_Streaming_FlushesIncrementally(t *testing.T) {
 	}
 }
 
+// --- usage provenance (feat: expose token-accounting provenance) ---
+//
+// Token accounting has three provenances (usage.estimated's own doc
+// comment, limits.go): reported (real provider usage), estimated (a
+// non-streaming response with no usage, prompt substituted from body
+// size), and unbilled (a streaming response with no usage, request
+// counted but zero tokens billed). These three tests drive each real
+// HTTP shape through Gateway.ServeHTTP and assert both the provenance
+// classification AND that billing itself is byte-identical to the
+// pre-feature contract — this feature must never change a billed number.
+
+// TestUsageProvenance_ReportedUsage_ClassifiedReported proves the normal,
+// accurate case (a provider that returns real usage) is classified
+// "reported", not merely "not estimated" by omission.
+//
+// MUTATION VERIFIED: changing routes_unified.go's provenance switch
+// default from `provenance := provenanceReported` to `provenance :=
+// provenanceEstimated` made this test fail — the seen provenance became
+// "estimated" instead of "reported". Reverted before committing.
+func TestUsageProvenance_ReportedUsage_ClassifiedReported(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"id":"c1","object":"chat.completion","model":"gpt-test","choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":5}}`))
+	}))
+	defer srv.Close()
+
+	cfg := CreateConfig()
+	cfg.Metrics = &MetricsConfig{Enabled: true}
+	cfg.Providers = map[string]*ProviderConfig{
+		"openai": {Type: "openai", BaseURL: srv.URL, APIKey: "sk-up", Models: []string{"gpt-test"}},
+	}
+	cfg.Groups = map[string]*GroupConfig{"default": {}}
+	cfg.Users = &UsersConfig{Inline: []*UserConfig{{Name: "alice", Group: "default", APIKey: "sk-alice"}}}
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {})
+	h, err := New(context.Background(), next, cfg, "llmgw")
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+	gw, ok := h.(*Gateway)
+	if !ok {
+		t.Fatal("handler is not *Gateway")
+	}
+
+	body := map[string]any{"model": "gpt-test", "messages": []any{map[string]any{"role": "user", "content": "hi"}}}
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", body))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200, body=%s", rec.Code, rec.Body.String())
+	}
+
+	snaps := gw.provenance.snapshot()
+	if len(snaps) != 1 {
+		t.Fatalf("g.provenance has %d entries, want 1: %+v", len(snaps), snaps)
+	}
+	s := snaps[0]
+	if s.key.provider != "openai" || s.key.provenance != provenanceReported {
+		t.Errorf("provenance key = %+v, want provider=openai provenance=reported", s.key)
+	}
+	if s.requests != 1 {
+		t.Errorf("requests = %d, want 1", s.requests)
+	}
+	if s.tokens != 15 {
+		t.Errorf("tokens = %d, want 15 (10 prompt + 5 completion, the real reported usage)", s.tokens)
+	}
+
+	// Billing itself must be exactly what TestHandleChat_HappyPath_
+	// NonStreaming_AccountsUsage already proves for this identical
+	// response shape — this feature must never change a billed number.
+	tokIn, _ := gw.limiter.getCounter("user", "alice", metricTokIn, windowDay, time.Now())
+	tokOut, _ := gw.limiter.getCounter("user", "alice", metricTokOut, windowDay, time.Now())
+	if tokIn != 10 || tokOut != 5 {
+		t.Errorf("billed tokin/tokout = %d/%d, want 10/5 — provenance recording must never change what gets billed", tokIn, tokOut)
+	}
+}
+
+// TestUsageProvenance_NonStreamingZeroUsage_ClassifiedEstimated proves a
+// non-streaming response that carries no usage at all is classified
+// "estimated", with the SAME prompt-estimate/zero-completion billing
+// routes_unified.go already applied before this feature existed —
+// ceil(request body size / 4) prompt, zero completion.
+//
+// MUTATION VERIFIED: changing the recording switch's `case
+// result.estimated:` guard (routes_unified.go) to `case false:` made
+// this test fail — the seen provenance stayed "reported" instead of
+// becoming "estimated". Reverted before committing.
+func TestUsageProvenance_NonStreamingZeroUsage_ClassifiedEstimated(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		// No "usage" field at all — decodes to chatUsagePayload's zero
+		// value (provider_openai.go's forwardJSON), the identical shape a
+		// provider that omits usage from a non-streaming response sends.
+		_, _ = w.Write([]byte(`{"id":"c1","object":"chat.completion","model":"gpt-test","choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}]}`))
+	}))
+	defer srv.Close()
+
+	cfg := CreateConfig()
+	cfg.Metrics = &MetricsConfig{Enabled: true}
+	cfg.Providers = map[string]*ProviderConfig{
+		"openai": {Type: "openai", BaseURL: srv.URL, APIKey: "sk-up", Models: []string{"gpt-test"}},
+	}
+	cfg.Groups = map[string]*GroupConfig{"default": {}}
+	cfg.Users = &UsersConfig{Inline: []*UserConfig{{Name: "alice", Group: "default", APIKey: "sk-alice"}}}
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {})
+	h, err := New(context.Background(), next, cfg, "llmgw")
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+	gw, ok := h.(*Gateway)
+	if !ok {
+		t.Fatal("handler is not *Gateway")
+	}
+
+	reqBody := map[string]any{"model": "gpt-test", "messages": []any{map[string]any{"role": "user", "content": "hi"}}}
+	// The exact bytes newUnifiedRequest below marshals reqBody into —
+	// routes_unified.go's own estimate is ceil(len(body) / 4) over these
+	// SAME bytes, so this reproduces the source's formula against a
+	// black-box-observable input (the request the test itself sends),
+	// not a re-implementation of internal logic.
+	marshaled, err := json.Marshal(reqBody)
+	if err != nil {
+		t.Fatalf("marshal request body: %v", err)
+	}
+	wantPrompt := int64(math.Ceil(float64(len(marshaled)) / 4))
+
+	rec := httptest.NewRecorder()
+	h.ServeHTTP(rec, newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", reqBody))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want 200, body=%s", rec.Code, rec.Body.String())
+	}
+
+	snaps := gw.provenance.snapshot()
+	if len(snaps) != 1 {
+		t.Fatalf("g.provenance has %d entries, want 1: %+v", len(snaps), snaps)
+	}
+	s := snaps[0]
+	if s.key.provider != "openai" || s.key.provenance != provenanceEstimated {
+		t.Errorf("provenance key = %+v, want provider=openai provenance=estimated", s.key)
+	}
+	if s.tokens != wantPrompt {
+		t.Errorf("provenance tokens = %d, want %d (ceil(request body size / 4), the estimate)", s.tokens, wantPrompt)
+	}
+
+	// Billing unchanged from the pre-feature contract: prompt substituted
+	// with the estimate, completion billed as zero.
+	tokIn, _ := gw.limiter.getCounter("user", "alice", metricTokIn, windowDay, time.Now())
+	tokOut, _ := gw.limiter.getCounter("user", "alice", metricTokOut, windowDay, time.Now())
+	if tokIn != wantPrompt {
+		t.Errorf("billed tokin = %d, want %d (ceil(request body size / 4))", tokIn, wantPrompt)
+	}
+	if tokOut != 0 {
+		t.Errorf("billed tokout = %d, want 0 — completion is billed as zero for an estimated response", tokOut)
+	}
+}
+
+// TestUsageProvenance_StreamingZeroUsage_ClassifiedUnbilled_DistinctFromEstimated
+// proves a streaming response that carries no usage at all is classified
+// "unbilled" — DISTINCT from "estimated" (this feature's own
+// correctness requirement: collapsing the two into one "not reported"
+// signal would hide that a provider serving completions entirely free
+// is a materially different failure than an approximated bill) — and
+// that billing stays exactly what routes_unified.go already applied
+// before this feature existed: the request counted, zero tokens billed.
+//
+// MUTATION VERIFIED: changing the recording switch's `case streaming &&
+// result.total() == 0:` guard (routes_unified.go) to `case
+// result.estimated:` (merging unbilled into the estimated branch) made
+// this test fail — the seen provenance became "estimated" instead of
+// "unbilled", and the distinctness assertion below failed too. Reverted
+// before committing.
+func TestUsageProvenance_StreamingZeroUsage_ClassifiedUnbilled_DistinctFromEstimated(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(http.StatusOK)
+		fl, ok := w.(http.Flusher)
+		if !ok {
+			t.Fatal("fake upstream ResponseWriter does not support Flush")
+		}
+		_, _ = w.Write([]byte("data: {\"id\":\"c1\",\"object\":\"chat.completion.chunk\",\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n"))
+		fl.Flush()
+		// No usage-only chunk — the provider never honored
+		// stream_options.include_usage, the exact shape this feature
+		// exists to surface.
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		fl.Flush()
+	}))
+	defer srv.Close()
+
+	cfg := CreateConfig()
+	cfg.Metrics = &MetricsConfig{Enabled: true}
+	cfg.Providers = map[string]*ProviderConfig{
+		"openai": {Type: "openai", BaseURL: srv.URL, APIKey: "sk-up", Models: []string{"gpt-test"}},
+	}
+	cfg.Groups = map[string]*GroupConfig{"default": {}}
+	cfg.Users = &UsersConfig{Inline: []*UserConfig{{Name: "alice", Group: "default", APIKey: "sk-alice"}}}
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {})
+	h, err := New(context.Background(), next, cfg, "llmgw")
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+	gw, ok := h.(*Gateway)
+	if !ok {
+		t.Fatal("handler is not *Gateway")
+	}
+
+	body := map[string]any{"model": "gpt-test", "stream": true, "messages": []any{map[string]any{"role": "user", "content": "hi"}}}
+	rw := newRecordingWriter()
+	h.ServeHTTP(rw, newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", body))
+	if rw.status != http.StatusOK {
+		t.Fatalf("status = %d, want 200", rw.status)
+	}
+
+	snaps := gw.provenance.snapshot()
+	if len(snaps) != 1 {
+		t.Fatalf("g.provenance has %d entries, want 1: %+v", len(snaps), snaps)
+	}
+	s := snaps[0]
+	if s.key.provider != "openai" || s.key.provenance != provenanceUnbilled {
+		t.Errorf("provenance key = %+v, want provider=openai provenance=unbilled", s.key)
+	}
+	if s.key.provenance == provenanceEstimated {
+		t.Errorf("provenance = estimated, want unbilled — a streaming zero-usage response must never be classified the same as a non-streaming one")
+	}
+	if s.tokens != 0 {
+		t.Errorf("provenance tokens = %d, want 0 — unbilled charges nothing at all", s.tokens)
+	}
+
+	// Billing unchanged from the pre-feature contract: the request is
+	// counted (req/min bumped) but zero tokens are billed.
+	tokIn, _ := gw.limiter.getCounter("user", "alice", metricTokIn, windowDay, time.Now())
+	tokOut, _ := gw.limiter.getCounter("user", "alice", metricTokOut, windowDay, time.Now())
+	if tokIn != 0 || tokOut != 0 {
+		t.Errorf("billed tokin/tokout = %d/%d, want 0/0 — a streaming zero-usage response bills nothing", tokIn, tokOut)
+	}
+	reqCount, ok2 := gw.limiter.getCounter("user", "alice", metricReq, windowMin, time.Now())
+	if !ok2 || reqCount != 1 {
+		t.Errorf("request/min counter = %d (ok=%v), want 1 — the request must still be COUNTED even though zero tokens are billed", reqCount, ok2)
+	}
+}
+
 // TestHandleChat_NoAPIKey_Returns401 asserts the unified routes require
 // authentication before any resolve/limit work happens.
 func TestHandleChat_NoAPIKey_Returns401(t *testing.T) {

tools/yaegi-check/main.go

@@ -164,6 +164,26 @@ const (
 	hungBodyUpstreamSleep          = 600 * time.Millisecond
 )
 
+// provenanceEstProviderName/provenanceEstProviderModel back
+// exerciseUsageProvenance (feat: expose token-accounting provenance): a
+// provider whose upstream answers 200 with NO "usage" field at all —
+// decodes to chatUsagePayload's zero value (provider_openai.go's
+// forwardJSON), the exact shape runUnified's own estimation fallback
+// (routes_unified.go) classifies as "estimated" — proving the
+// substitution math (ceil(len(body)/4)) and the new
+// llmgateway_usage_provenance_requests_total/-_tokens_total families
+// (metrics.go) all run correctly under the REAL interpreter, not merely
+// compiled. "reported" is already exercised, under the interpreter, by
+// exerciseAttemptAccounting's own real, usage-bearing "openai" traffic
+// above — no separate fixture needed for that provenance here.
+const (
+	provenanceEstProviderName = "provenance-est"
+	// provenanceEstProviderModel: same reasoning as timeoutProviderModel's
+	// own doc comment above — must not collide with testDataWantModel or
+	// any other probe's model id.
+	provenanceEstProviderModel = "gpt-provenance-est-test"
+)
+
 // excludedTopLevelDirs lists repo-root directories the GOPATH copy must
 // never include: build tooling, integration fixtures, planning docs and
 // VCS metadata have nothing to do with the plugin package Yaegi imports.
@@ -348,6 +368,18 @@ func run() error {
 	}))
 	defer hungBodyUpstream.Close()
 
+	// provenanceEstUpstream backs exerciseUsageProvenance (feat: expose
+	// token-accounting provenance): answers 200 with NO "usage" field at
+	// all — see provenanceEstProviderName's own const block doc comment
+	// for why this is the exact shape runUnified's estimation fallback
+	// classifies as "estimated".
+	provenanceEstUpstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"id":"c-prov","object":"chat.completion","model":"` + provenanceEstProviderModel + `","choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}]}`))
+	}))
+	defer provenanceEstUpstream.Close()
+
 	// mcpProbeUpstream answers federation's outbound tools/call with a
 	// body deliberately larger than mcpBackendCallResponseMaxBytes, so
 	// exerciseHandler's POST /mcp probe below drives doBackendJSONRPC's
@@ -492,6 +524,9 @@ func run() error {
 		// comment above for why this provider exists alongside
 		// timeoutProviderName.
 		`"` + hungBodyProviderName + `":{"type":"openai","baseUrl":"` + hungBodyUpstream.URL + `","apiKey":"sk-up","requestTimeout":"` + hungBodyProviderRequestTimeout + `","models":["` + hungBodyProviderModel + `"]},` +
+		// provenanceEstProviderName (feat: expose token-accounting
+		// provenance) — see its own const block doc comment above.
+		`"` + provenanceEstProviderName + `":{"type":"openai","baseUrl":"` + provenanceEstUpstream.URL + `","apiKey":"sk-up","models":["` + provenanceEstProviderModel + `"]},` +
 		// "anthropic" backs the /v1/messages passthrough probes
 		// (exerciseMessagesRoute, below): a real anthropic-type provider,
 		// interpreted end to end through
@@ -591,14 +626,21 @@ func run() error {
 		return err
 	}
 	// exerciseLatencyMetrics (feat: instrument upstream latency) runs
-	// LAST, for the same reason exerciseMetricsRoute itself used to be
-	// last: it reads the exact same /metrics endpoint and must not race
-	// or perturb anything exerciseBreaker/exerciseMetricsRoute still care
-	// about. It relies on exerciseHandler's own exerciseAttemptAccounting
-	// sub-probe (already run, above) having driven a real, non-streaming
-	// POST /v1/chat/completions against the "openai" provider — see its
-	// own doc comment.
-	return exerciseLatencyMetrics(handler)
+	// after exerciseMetricsRoute, for the same reason exerciseMetricsRoute
+	// itself used to be last: it reads the exact same /metrics endpoint
+	// and must not race or perturb anything exerciseBreaker/
+	// exerciseMetricsRoute still care about. It relies on exerciseHandler's
+	// own exerciseAttemptAccounting sub-probe (already run, above) having
+	// driven a real, non-streaming POST /v1/chat/completions against the
+	// "openai" provider — see its own doc comment.
+	if err := exerciseLatencyMetrics(handler); err != nil {
+		return err
+	}
+	// exerciseUsageProvenance (feat: expose token-accounting provenance)
+	// runs LAST, for the identical reason: it reads /metrics and GET
+	// /admin/api/overview one more time and must not perturb any counter
+	// an earlier probe already asserted on.
+	return exerciseUsageProvenance(handler)
 }
 
 // builtinLookupModelID is a real, stable entry in the generated
@@ -1035,6 +1077,154 @@ func exerciseLatencyMetrics(handler http.Handler) error {
 	return nil
 }
 
+// exerciseUsageProvenance proves the usage-provenance metric surface
+// (feat: expose token-accounting provenance, metrics.go's
+// provenanceStore/writeProvenanceMetrics, admin.go's
+// buildAdminProvenanceViews) renders correctly under the REAL
+// interpreter — not merely that a compiled `go test` accepts the same
+// bytes, this package's own doc comment's exact concern.
+//
+// Drives one real, non-streaming POST /v1/chat/completions against
+// provenanceEstProviderName, whose upstream (provenanceEstUpstream, run())
+// answers 200 with no "usage" field at all — the exact shape
+// runUnified's estimation fallback (routes_unified.go) classifies as
+// "estimated". "reported" is not driven fresh here: exerciseHandler's own
+// exerciseAttemptAccounting sub-probe (already run, above) already drove
+// a real, usage-bearing POST /v1/chat/completions against "openai",
+// which is reported provenance by construction — reusing that traffic
+// instead of a third fixture, mirroring exerciseLatencyMetrics' own
+// "no new traffic" reasoning for its own duration-histogram assertion.
+func exerciseUsageProvenance(handler http.Handler) error {
+	chatReq := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(
+		`{"model":"`+provenanceEstProviderName+`/`+provenanceEstProviderModel+`","messages":[{"role":"user","content":"hi"}]}`,
+	))
+	chatReq.Header.Set("Authorization", "Bearer "+testDataUserAPIKey)
+	chatReq.Header.Set("Content-Type", "application/json")
+	chatRec := httptest.NewRecorder()
+	handler.ServeHTTP(chatRec, chatReq)
+	if chatRec.Code != http.StatusOK {
+		return fmt.Errorf("POST /v1/chat/completions (usage-provenance harness): status = %d, want 200, body=%s", chatRec.Code, chatRec.Body.String())
+	}
+
+	req := httptest.NewRequest(http.MethodGet, "/metrics", nil)
+	req.RemoteAddr = metricsProbeOutsideAddr
+	req.Header.Set("Authorization", "Bearer "+attemptAccountingAdminAPIKey)
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		return fmt.Errorf("GET /metrics (usage-provenance probe): status = %d, want 200, body=%s", rec.Code, rec.Body.String())
+	}
+
+	parser := expfmt.NewTextParser(model.LegacyValidation)
+	families, err := parser.TextToMetricFamilies(strings.NewReader(rec.Body.String()))
+	if err != nil {
+		return fmt.Errorf("GET /metrics: a real Prometheus TextParser rejected the interpreted rendering: %w (body=%s)", err, rec.Body.String())
+	}
+
+	requestsFamily, ok := families["llmgateway_usage_provenance_requests_total"]
+	if !ok {
+		return fmt.Errorf("GET /metrics: no llmgateway_usage_provenance_requests_total family in the parsed output (body=%s)", rec.Body.String())
+	}
+	if requestsFamily.GetType() != dto.MetricType_COUNTER {
+		return fmt.Errorf("llmgateway_usage_provenance_requests_total type = %s, want COUNTER", requestsFamily.GetType())
+	}
+	tokensFamily, ok := families["llmgateway_usage_provenance_tokens_total"]
+	if !ok {
+		return fmt.Errorf("GET /metrics: no llmgateway_usage_provenance_tokens_total family in the parsed output (body=%s)", rec.Body.String())
+	}
+	if tokensFamily.GetType() != dto.MetricType_COUNTER {
+		return fmt.Errorf("llmgateway_usage_provenance_tokens_total type = %s, want COUNTER", tokensFamily.GetType())
+	}
+
+	findSample := func(fam *dto.MetricFamily, provider, provenance string) *dto.Metric {
+		for _, m := range fam.GetMetric() {
+			var p, k string
+			for _, lp := range m.GetLabel() {
+				switch lp.GetName() {
+				case "provider":
+					p = lp.GetValue()
+				case "provenance":
+					k = lp.GetValue()
+				}
+			}
+			if p == provider && k == provenance {
+				return m
+			}
+		}
+		return nil
+	}
+
+	estReq := findSample(requestsFamily, provenanceEstProviderName, "estimated")
+	if estReq == nil {
+		return fmt.Errorf(`GET /metrics: no llmgateway_usage_provenance_requests_total{provider=%q,provenance="estimated"} sample found (body=%s)`, provenanceEstProviderName, rec.Body.String())
+	}
+	if estReq.GetCounter().GetValue() != 1 {
+		return fmt.Errorf(`llmgateway_usage_provenance_requests_total{provider=%q,provenance="estimated"} = %v, want 1`, provenanceEstProviderName, estReq.GetCounter().GetValue())
+	}
+	estTok := findSample(tokensFamily, provenanceEstProviderName, "estimated")
+	if estTok == nil {
+		return fmt.Errorf(`GET /metrics: no llmgateway_usage_provenance_tokens_total{provider=%q,provenance="estimated"} sample found (body=%s)`, provenanceEstProviderName, rec.Body.String())
+	}
+	if estTok.GetCounter().GetValue() <= 0 {
+		return fmt.Errorf(`llmgateway_usage_provenance_tokens_total{provider=%q,provenance="estimated"} = %v, want > 0 (estimated substitutes a nonzero prompt count)`, provenanceEstProviderName, estTok.GetCounter().GetValue())
+	}
+
+	reportedReq := findSample(requestsFamily, "openai", "reported")
+	if reportedReq == nil {
+		return fmt.Errorf(`GET /metrics: no llmgateway_usage_provenance_requests_total{provider="openai",provenance="reported"} sample found (body=%s) — exerciseAttemptAccounting's own earlier real, usage-bearing traffic must have been classified "reported"`, rec.Body.String())
+	}
+	if reportedReq.GetCounter().GetValue() < 1 {
+		return fmt.Errorf(`llmgateway_usage_provenance_requests_total{provider="openai",provenance="reported"} = %v, want >= 1`, reportedReq.GetCounter().GetValue())
+	}
+
+	// GET /admin/api/overview: proves admin.go's buildAdminProvenanceViews
+	// wiring also runs correctly interpreted, decoded generically (this
+	// harness module is compiled, not interpreted; adminOverviewResponse/
+	// adminProviderView are unexported — see readProviderAttemptCounters'
+	// own doc comment for why a generic decode is simpler here than
+	// exporting test-only types across that boundary).
+	adminReq := httptest.NewRequest(http.MethodGet, "/admin/api/overview", nil)
+	adminReq.Header.Set("Authorization", "Bearer "+attemptAccountingAdminAPIKey)
+	adminRec := httptest.NewRecorder()
+	handler.ServeHTTP(adminRec, adminReq)
+	if adminRec.Code != http.StatusOK {
+		return fmt.Errorf("GET /admin/api/overview (usage-provenance probe): status = %d, want 200, body=%s", adminRec.Code, adminRec.Body.String())
+	}
+	var overview map[string]any
+	if err := json.Unmarshal(adminRec.Body.Bytes(), &overview); err != nil {
+		return fmt.Errorf("decode GET /admin/api/overview body: %w", err)
+	}
+	providers, _ := overview["providers"].([]any)
+	var provenanceEstProvider map[string]any
+	for _, p := range providers {
+		pm, ok := p.(map[string]any)
+		if !ok {
+			continue
+		}
+		if pm["name"] == provenanceEstProviderName {
+			provenanceEstProvider = pm
+			break
+		}
+	}
+	if provenanceEstProvider == nil {
+		return fmt.Errorf("GET /admin/api/overview: no provider entry named %q (body=%s)", provenanceEstProviderName, adminRec.Body.String())
+	}
+	provenanceField, _ := provenanceEstProvider["provenance"].(map[string]any)
+	estimatedEntry, _ := provenanceField["estimated"].(map[string]any)
+	if estimatedEntry == nil {
+		return fmt.Errorf(`GET /admin/api/overview: provider %q has no provenance.estimated entry: %+v`, provenanceEstProviderName, provenanceField)
+	}
+	if reqs, _ := estimatedEntry["requests"].(float64); reqs != 1 {
+		return fmt.Errorf("GET /admin/api/overview: provider %q provenance.estimated.requests = %v, want 1", provenanceEstProviderName, estimatedEntry["requests"])
+	}
+	if _, hasReported := provenanceField["reported"]; hasReported {
+		return fmt.Errorf(`GET /admin/api/overview: provider %q provenance has a "reported" key, want it excluded from this compact admin summary: %+v`, provenanceEstProviderName, provenanceField)
+	}
+
+	fmt.Println("yaegi-check: usage-provenance metrics (llmgateway_usage_provenance_requests_total/-_tokens_total) and admin overview parsed by a real Prometheus TextParser/JSON decoder; estimated and reported provenances both present and distinct")
+	return nil
+}
+
 // exerciseHandler runs two real requests against the interpreted
 // handler, drawn from the plugin's own .traefik.yml testData: an
 // authenticated GET /v1/models, which must return the configured

webui/src/components/ProvidersView.vue

@@ -25,7 +25,7 @@ import { formatAgo, formatContextWindow, formatLatencyMs, formatUntil, refreshDe
 import { isModelDegraded } from '@/lib/provider-rate'
 import { type ExpandState, clearExpandOverrides, computeExpandedItems, toggleItemExpand } from '@/lib/search-expand'
 import { useDashboardStore } from '@/stores/dashboard'
-import type { AdminAliasView, AdminLatencyView, AdminModelMetaView, AdminModelRateView, AdminProviderView } from '@/types/api'
+import type { AdminAliasView, AdminLatencyView, AdminModelMetaView, AdminModelRateView, AdminProvenanceView, AdminProviderView } from '@/types/api'
 
 // This component backs the "Providers" tab (App.vue) — a UI-label rename
 // only. The data it renders still comes from GET /admin/api/overview
@@ -251,6 +251,40 @@ function nonStreamingLatencyTitle(v: AdminLatencyView): string {
   return `Average total response time, including generation — not a load-sensitive signal like streaming ttfb, since this provider buffers the whole completion before sending anything. ${latencyObservationCount(v)}. ${PER_REPLICA_CAVEAT}`
 }
 
+// --- usage provenance (feat: expose token-accounting provenance) ---
+//
+// Token accounting has three provenances (see AdminProvenanceView's own
+// doc comment): reported (the healthy default — no badge, mirroring the
+// health/latency badges' own "nothing to show reads as no badge"
+// convention), estimated (a non-streaming response carried no usage, so
+// prompt was substituted from request body size), and unbilled (a
+// streaming response carried no usage, so the request was counted but
+// zero tokens were billed — a provider potentially serving completions
+// entirely free against every budget). estimated/unbilled are two
+// DISTINCT badges, never merged into one "not reported" indicator —
+// collapsing them would hide exactly the fact this feature exists to
+// surface (task brief's own correctness rule).
+
+/** estimatedProvenance/unbilledProvenance pull p.provenance's two known kind keys (admin.go's buildAdminProvenanceViews emits only these two strings, never "reported") — undefined when that kind has no observations yet. */
+function estimatedProvenance(p: AdminProviderView): AdminProvenanceView | undefined {
+  return p.provenance?.estimated
+}
+function unbilledProvenance(p: AdminProviderView): AdminProvenanceView | undefined {
+  return p.provenance?.unbilled
+}
+
+/** estimatedProvenanceTitle is the estimated-accounting badge's hover/title detail. */
+function estimatedProvenanceTitle(v: AdminProvenanceView): string {
+  const plural = v.requests === 1 ? '' : 's'
+  return `${v.requests} non-streaming response${plural} carried no usage from the provider — prompt tokens were estimated from request body size instead (${v.tokens} tokens billed from the estimate), completion billed as zero. ${PER_REPLICA_CAVEAT}`
+}
+
+/** unbilledProvenanceTitle is the unbilled-accounting badge's hover/title detail — explicit about the budget-exposure implication (task brief's own framing). */
+function unbilledProvenanceTitle(v: AdminProvenanceView): string {
+  const plural = v.requests === 1 ? '' : 's'
+  return `${v.requests} streaming response${plural} carried no usage from the provider — the request was counted but zero tokens were billed. If this keeps happening, this provider may be serving completions entirely free against every budget. ${PER_REPLICA_CAVEAT}`
+}
+
 // --- model aliases (sortable DataTable, operator directive) ---
 //
 // The alias id gets the ModelChip copy treatment (an alias name IS the
@@ -419,6 +453,33 @@ const aliasEmptyMessage = computed(() =>
                 >
                   non-streaming avg {{ formatLatencyMs(nonStreamingLatency(p)!.avgDurationMs as number) }}
                 </Badge>
+                <!--
+                  Usage provenance (feat: expose token-accounting
+                  provenance): no badge at all for a kind with no
+                  observations yet, mirroring the health/latency badges'
+                  own convention above. estimated and unbilled are two
+                  distinct badges — never merged (correctness rule, task
+                  brief) — since one substitutes an approximation and the
+                  other bills nothing at all.
+                -->
+                <Badge
+                  v-if="estimatedProvenance(p)"
+                  as="span"
+                  variant="secondary"
+                  class="font-normal tabular-nums"
+                  :title="estimatedProvenanceTitle(estimatedProvenance(p) as AdminProvenanceView)"
+                >
+                  estimated {{ estimatedProvenance(p)!.requests }}
+                </Badge>
+                <Badge
+                  v-if="unbilledProvenance(p)"
+                  as="span"
+                  variant="destructive"
+                  class="font-normal tabular-nums"
+                  :title="unbilledProvenanceTitle(unbilledProvenance(p) as AdminProvenanceView)"
+                >
+                  unbilled {{ unbilledProvenance(p)!.requests }}
+                </Badge>
                 <FontAwesomeIcon
                   v-if="p.lastErr"
                   :icon="faTriangleExclamation"

webui/src/types/api.ts

@@ -58,6 +58,26 @@ export interface AdminLatencyView {
   count?: number
 }
 
+/**
+ * One provider's one non-reported provenance kind's compact usage-
+ * accounting summary (admin.go: adminProvenanceView — feat: expose
+ * token-accounting provenance). "estimated" means a non-streaming
+ * response carried no usage from the provider, so prompt was substituted
+ * with an estimate (ceil(request body length / 4)) and completion billed
+ * as zero. "unbilled" means a streaming response carried no usage, so
+ * the request was counted but zero tokens were billed — a provider
+ * silently ignoring stream_options.include_usage serves that completion
+ * entirely free against every budget. Billing itself is unaffected by
+ * this data; it only reports how an already-billed number was arrived
+ * at. requests/tokens are never fabricated zeros: this view only ever
+ * exists (see AdminProviderView.provenance) for a kind that was actually
+ * observed at least once.
+ */
+export interface AdminProvenanceView {
+  requests: number
+  tokens: number
+}
+
 export interface AdminProviderView {
   name: string
   type: string
@@ -142,6 +162,22 @@ export interface AdminProviderView {
    * together or present either one as a bare, unlabeled "latency".
    */
   latency?: Partial<Record<'streaming' | 'non-streaming', AdminLatencyView>>
+  /**
+   * provenance summarizes non-reported usage accounting for this
+   * provider (admin.go: adminProviderView.Provenance — feat: expose
+   * token-accounting provenance), keyed by provenance kind: 'estimated'
+   * or 'unbilled' only — 'reported' (the default, healthy case) is never
+   * a key here, see AdminProvenanceView's own doc comment for why.
+   * Undefined entirely for a provider with no estimated/unbilled outcome
+   * yet, or a deployment with metrics collection gated off — mirroring
+   * latency's identical "undefined, never an empty object" convention
+   * above.
+   *
+   * PER-REPLICA, IN-PROCESS: sourced from the same in-process accumulator
+   * as latency above — see that field's own caveat, which applies here
+   * unchanged.
+   */
+  provenance?: Partial<Record<'estimated' | 'unbilled', AdminProvenanceView>>
 }
 
 export interface AdminRedisView {