From b2c2a782499bee81abb4c05f29c546207f5f79a1 Mon Sep 17 00:00:00 2001
From: Lukin <mylukin@gmail.com>
Date: Sun, 5 Jul 2026 00:12:53 +0800
Subject: [PATCH] Record serving subscription provider accounts

Co-Authored-By: Codex <noreply@openai.com>
---
 apps/admin/src/lib/api/requests.test.ts       | 21 ++++
 apps/admin/src/lib/api/requests.ts            | 73 +++++++++++++-
 .../src/lib/components/DecisionChain.svelte   | 16 ++-
 .../src/lib/components/DecisionChain.test.ts  |  9 ++
 .../src/lib/components/RequestsTable.svelte   | 98 +++++++++++--------
 .../src/lib/components/RequestsTable.test.ts  |  2 +
 apps/admin/src/locales/en.json                |  5 +
 apps/admin/src/locales/ja.json                |  5 +
 apps/admin/src/locales/ko.json                |  5 +
 apps/admin/src/locales/zh-hans.json           |  5 +
 apps/admin/src/locales/zh-hant.json           |  5 +
 .../routes/keys/[keyId]/key-detail.test.ts    |  2 +
 .../routes/requests/[traceId]/+page.svelte    | 41 +++++---
 .../src/routes/requests/requests.test.ts      | 23 +++--
 apps/gateway/src/routes/admin/admin.test.ts   |  1 +
 apps/gateway/src/routes/chat.ts               |  4 +-
 apps/gateway/src/routes/gemini.ts             |  4 +
 apps/gateway/src/routes/image-telemetry.ts    |  1 +
 apps/gateway/src/routes/messages-pipeline.ts  |  3 +
 apps/gateway/src/routes/messages.ts           |  5 +
 .../src/routes/payload-capture.test.ts        |  1 +
 apps/gateway/src/routes/responses.ts          |  4 +
 .../src/runtime/serving-account.test.ts       | 23 +++++
 apps/gateway/src/runtime/serving-account.ts   | 15 +++
 implementation-notes.md                       | 33 +++----
 packages/core/src/executor/attempt-record.ts  | 17 +++-
 packages/core/src/routing/route-request.ts    |  4 +
 packages/core/src/signals/aggregate.test.ts   |  1 +
 packages/core/src/signals/collector.test.ts   |  1 +
 packages/core/src/store/ports.test.ts         |  1 +
 .../core/src/store/sqlite/telemetry.test.ts   |  1 +
 .../core/src/store/store-contract.test.ts     |  1 +
 packages/core/src/telemetry/decision.test.ts  | 23 +++++
 packages/core/src/telemetry/decision.ts       | 49 +++++++---
 packages/shared/src/decision/schema.test.ts   | 13 +++
 packages/shared/src/decision/schema.ts        | 10 ++
 packages/shared/src/index.ts                  |  2 +
 37 files changed, 432 insertions(+), 95 deletions(-)

diff --git a/apps/admin/src/lib/api/requests.test.ts b/apps/admin/src/lib/api/requests.test.ts
index ed351d4d..89ebda5b 100644
--- a/apps/admin/src/lib/api/requests.test.ts
+++ b/apps/admin/src/lib/api/requests.test.ts
@@ -39,6 +39,8 @@ function rawRecord(overrides: Record<string, unknown> = {}): Record<string, unkn
       },
       {
         alias: 'premium',
+        provider_name: 'anthropic',
+        provider_model: 'claude-x',
         skipped: false,
         skip_reason: null,
         status: 'ok',
@@ -53,6 +55,7 @@ function rawRecord(overrides: Record<string, unknown> = {}): Record<string, unkn
       status: 'ok',
       error_reason: null,
     },
+    serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
     latency_total_ms: 460,
     fallback_count: 1,
     cost_breakdown: { eval_usd: 0.0002, completion_usd: 0.01, total_usd: 0.0102 },
@@ -76,6 +79,16 @@ describe('toListItem', () => {
     expect(row.decided_by).toBe('eval');
   });
 
+  it('surfaces the served provider and final subscription account on list rows', () => {
+    const row = toListItem(rawRecord());
+    expect(row.served_provider).toBe('anthropic');
+    expect(row.serving_account).toEqual({ provider_id: 'anthropic', account: 'claude-team-a' });
+
+    const nonSubscription = toListItem(rawRecord({ serving_account: null }));
+    expect(nonSubscription.served_provider).toBe('anthropic');
+    expect(nonSubscription.serving_account).toBeNull();
+  });
+
   it('falls back to "—" for key_prefix when the record carries none (never plaintext)', () => {
     const row = toListItem(rawRecord({ key_prefix: null }));
     expect(row.key_prefix).toBe('—');
@@ -136,6 +149,8 @@ describe('toDetail', () => {
     expect(d.key_name).toBe('Production backend');
     expect(d.requested_model).toBe('gpt-4o');
     expect(d.final_model).toBe('premium'); // served model alias
+    expect(d.served_provider).toBe('anthropic');
+    expect(d.serving_account).toEqual({ provider_id: 'anthropic', account: 'claude-team-a' });
     expect(d.lane).toBe('coding');
     expect(d.status).toBe('ok');
     expect(d.latency_ms).toBe(460);
@@ -254,6 +269,12 @@ describe('toDetail', () => {
     });
     // The ok attempt carries no detail.
     expect(d.provider_attempts[1]?.error_detail ?? null).toBeNull();
+    expect(d.provider_attempts[1]?.provider).toBe('anthropic');
+    expect(d.provider_attempts[1]?.provider_model).toBe('claude-x');
+    expect(d.provider_attempts[1]?.serving_account).toEqual({
+      provider_id: 'anthropic',
+      account: 'claude-team-a',
+    });
   });
 
   it('maps a missing error_detail to null (legacy records)', () => {
diff --git a/apps/admin/src/lib/api/requests.ts b/apps/admin/src/lib/api/requests.ts
index 30112897..1d9299d0 100644
--- a/apps/admin/src/lib/api/requests.ts
+++ b/apps/admin/src/lib/api/requests.ts
@@ -46,6 +46,8 @@ export interface RequestListItem {
   complexity: string;
   decided_by: 'rules' | 'eval' | 'default' | 'fallback'; // decision layer (classification stage)
   lane: string;
+  served_provider: string | null;
+  serving_account: ServingAccountView | null;
   final_model: string | null;
   fallback_count: number; // execution fallback count (provider attempts - 1)
   status: 'ok' | 'error';
@@ -80,6 +82,11 @@ export interface TokenUsageView {
   total: number | null; // input + output when present; null when neither is measured
 }
 
+export interface ServingAccountView {
+  provider_id: string;
+  account: string;
+}
+
 // Redacted per-attempt upstream failure detail (admin-debug-error-detail).
 // Present only for an attempt that failed at the upstream; the backend has
 // already key-scrubbed `provider_raw` (Principle 7), so this is safe to display.
@@ -91,7 +98,9 @@ export interface AttemptErrorDetail {
 
 export interface ProviderAttempt {
   model: string;
-  provider: string;
+  provider: string | null;
+  provider_model: string | null;
+  serving_account: ServingAccountView | null;
   outcome: 'success' | 'error' | 'timeout' | 'rate_limited' | 'circuit_open' | 'skipped';
   skip_reason?: string;
   latency_ms: number;
@@ -111,6 +120,8 @@ export interface RequestDetail {
   key_prefix: string | null;
   key_name: string | null;
   requested_model: string | null; // what the client asked for
+  served_provider: string | null; // concrete provider that served the request
+  serving_account: ServingAccountView | null; // final subscription account, if any
   final_model: string | null; // the served model alias (null = no provider served)
   lane: string; // selected lane ('' on a legacy record)
   status: 'ok' | 'error';
@@ -183,6 +194,7 @@ interface RawAttempt {
   error_class?: string | null;
   latency_ms?: number;
   cost_usd?: number | null;
+  provider_name?: string | null;
   provider_model?: string | null;
   // Redacted upstream failure detail (admin-debug-error-detail). Null/absent on
   // ok/skipped rows and on legacy records.
@@ -254,6 +266,10 @@ interface RawDecisionRecord {
     status?: string;
     error_reason?: string | null;
   };
+  serving_account?: {
+    provider_id?: string | null;
+    account?: string | null;
+  } | null;
 }
 
 const BASE = '/admin/api/requests';
@@ -329,6 +345,42 @@ function attemptOutcome(a: RawAttempt): ProviderAttempt['outcome'] {
   return 'error';
 }
 
+function nonEmptyString(v: unknown): string | null {
+  return typeof v === 'string' && v.length > 0 ? v : null;
+}
+
+function aliasPrefix(alias: string | null): string | null {
+  if (!alias) return null;
+  const slash = alias.indexOf('/');
+  return slash > 0 ? alias.slice(0, slash) : null;
+}
+
+function normalizeServingAccount(raw: RawDecisionRecord): ServingAccountView | null {
+  const providerId = nonEmptyString(raw.serving_account?.provider_id);
+  const account = nonEmptyString(raw.serving_account?.account);
+  return providerId && account ? { provider_id: providerId, account } : null;
+}
+
+function successfulAttempt(raw: RawDecisionRecord, attempts: RawAttempt[]): RawAttempt | null {
+  const finalAlias = nonEmptyString(raw.final?.model_alias);
+  if (finalAlias) {
+    const exact = attempts.find(
+      (a) => a.status === 'ok' && a.skipped !== true && a.alias === finalAlias,
+    );
+    if (exact) return exact;
+  }
+  return attempts.find((a) => a.status === 'ok' && a.skipped !== true) ?? null;
+}
+
+function servedProvider(raw: RawDecisionRecord, attempts: RawAttempt[]): string | null {
+  const attempt = successfulAttempt(raw, attempts);
+  return (
+    nonEmptyString(attempt?.provider_name) ??
+    aliasPrefix(nonEmptyString(raw.final?.model_alias)) ??
+    aliasPrefix(nonEmptyString(attempt?.alias))
+  );
+}
+
 // Project the recorded `usage` block -> the UI token view. Reads only the four
 // recorded counts (Principle 1 — never recomputes upstream figures); DERIVES
 // `nonCached` (input − cached, clamped ≥0) and `total` (input + output) for the
@@ -376,6 +428,7 @@ export function computeTtfbMs(raw: RawDecisionRecord): number | null {
 // record does not carry are derived or safely defaulted; NEVER fabricated.
 export function toListItem(raw: RawDecisionRecord): RequestListItem {
   const attempts = Array.isArray(raw.provider_attempts) ? raw.provider_attempts : [];
+  const account = normalizeServingAccount(raw);
   const status: RequestListItem['status'] = raw.final?.status === 'error' ? 'error' : 'ok';
   const errorClass =
     status === 'error'
@@ -404,6 +457,8 @@ export function toListItem(raw: RawDecisionRecord): RequestListItem {
     complexity: raw.classifier?.complexity ?? '',
     decided_by: normalizeDecidedBy(raw.classifier?.decided_by),
     lane: raw.lane?.selected_lane ?? '',
+    served_provider: servedProvider(raw, attempts),
+    serving_account: account,
     final_model: raw.final?.model_alias ?? null,
     // Prefer the recorded value; fall back to deriving from attempts for legacy
     // records (Principle 5: execution-stage count, distinct from decided_by).
@@ -473,6 +528,7 @@ export function toDetail(raw: RawDecisionRecord): RequestDetail {
   const completion = sumCost(attempts);
   const status = raw.final?.status === 'error' ? 'error' : 'ok';
   const evalCacheHit = raw.classifier?.eval_cache_hit ?? null;
+  const account = normalizeServingAccount(raw);
   return {
     trace_id: String(raw.trace_id ?? raw.request_id ?? ''),
     // Same source as the list "Time" column (created_at, flattened by the detail
@@ -486,6 +542,8 @@ export function toDetail(raw: RawDecisionRecord): RequestDetail {
       typeof raw.key_prefix === 'string' && raw.key_prefix.length > 0 ? raw.key_prefix : null,
     key_name: typeof raw.key_name === 'string' && raw.key_name.length > 0 ? raw.key_name : null,
     requested_model: raw.requested_model ?? null,
+    served_provider: servedProvider(raw, attempts),
+    serving_account: account,
     final_model: raw.final?.model_alias ?? null,
     lane: raw.lane?.selected_lane ?? '',
     status,
@@ -518,7 +576,18 @@ export function toDetail(raw: RawDecisionRecord): RequestDetail {
     lane_candidates: Array.isArray(raw.lane?.candidate_chain) ? raw.lane.candidate_chain : [],
     provider_attempts: attempts.map((a) => ({
       model: String(a.alias ?? ''),
-      provider: String(a.provider_model ?? a.alias ?? ''),
+      provider:
+        nonEmptyString(a.provider_name) ??
+        aliasPrefix(nonEmptyString(a.alias)) ??
+        nonEmptyString(a.provider_model),
+      provider_model: nonEmptyString(a.provider_model),
+      serving_account:
+        account &&
+        a.status === 'ok' &&
+        a.skipped !== true &&
+        nonEmptyString(a.alias) === nonEmptyString(raw.final?.model_alias)
+          ? account
+          : null,
       outcome: attemptOutcome(a),
       skip_reason: a.skip_reason ?? undefined,
       latency_ms: typeof a.latency_ms === 'number' ? a.latency_ms : 0,
diff --git a/apps/admin/src/lib/components/DecisionChain.svelte b/apps/admin/src/lib/components/DecisionChain.svelte
index 1718b624..16b20a16 100644
--- a/apps/admin/src/lib/components/DecisionChain.svelte
+++ b/apps/admin/src/lib/components/DecisionChain.svelte
@@ -74,6 +74,10 @@
       return String(value);
     }
   }
+
+  function accountTitle(account: RequestDetail['provider_attempts'][number]['serving_account']): string {
+    return account ? `${account.provider_id}/${account.account}` : '';
+  }
 </script>
 
 <div class="flex flex-col gap-4">
@@ -236,8 +240,18 @@
       {#each detail.provider_attempts as a, i (i)}
         <li data-testid="attempt-row" class="flex flex-col gap-1 text-sm">
           <div class="flex flex-wrap items-center gap-2">
-            <span class="font-mono text-ink-strong">{a.provider}</span>
+            <span class="font-mono text-ink-strong">{a.provider ?? '—'}</span>
+            {#if a.serving_account}
+              <span class="badge-neutral" title={accountTitle(a.serving_account)}
+                >{a.serving_account.account}</span
+              >
+            {/if}
             <span class="text-ink-muted">{a.model}</span>
+            {#if a.provider_model && a.provider_model !== a.model}
+              <span class="text-ink-muted" title={$t('Provider model')}
+                >{$t('wire:')} {a.provider_model}</span
+              >
+            {/if}
             <span class={outcomeBadge(a.outcome)} title={a.outcome}
               >{$t(attemptCodeLabel(a.outcome))}</span
             >
diff --git a/apps/admin/src/lib/components/DecisionChain.test.ts b/apps/admin/src/lib/components/DecisionChain.test.ts
index d982c5f7..3c7e3086 100644
--- a/apps/admin/src/lib/components/DecisionChain.test.ts
+++ b/apps/admin/src/lib/components/DecisionChain.test.ts
@@ -19,6 +19,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
     key_prefix: 'helm_live_ab12',
     key_name: null,
     requested_model: 'gpt-4o',
+    served_provider: 'anthropic',
+    serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
     final_model: 'claude-x',
     lane: 'premium',
     status: 'ok',
@@ -45,6 +47,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
       {
         model: 'gpt-x',
         provider: 'openai',
+        provider_model: 'gpt-x',
+        serving_account: null,
         outcome: 'error',
         latency_ms: 120,
         error_class: 'upstream_error',
@@ -57,6 +61,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
       {
         model: 'claude-x',
         provider: 'anthropic',
+        provider_model: 'claude-x',
+        serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
         outcome: 'success',
         latency_ms: 340,
         error_detail: null,
@@ -64,6 +70,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
       {
         model: 'small-x',
         provider: 'local',
+        provider_model: 'small-x',
+        serving_account: null,
         outcome: 'skipped',
         skip_reason: 'capability_unsatisfiable',
         latency_ms: 0,
@@ -237,6 +245,7 @@ describe('DecisionChain', () => {
     expect(rows[0]).toHaveTextContent(/error/i);
     expect(rows[0]).toHaveTextContent('120');
     expect(rows[1]).toHaveTextContent(/success/i);
+    expect(rows[1]).toHaveTextContent('claude-team-a');
     expect(rows[2]).toHaveTextContent(/skipped/i);
     // Codes render as human labels; the raw code is preserved in the title tooltip.
     expect(rows[2]).toHaveTextContent('No compatible model');
diff --git a/apps/admin/src/lib/components/RequestsTable.svelte b/apps/admin/src/lib/components/RequestsTable.svelte
index 40b5645d..f410fdee 100644
--- a/apps/admin/src/lib/components/RequestsTable.svelte
+++ b/apps/admin/src/lib/components/RequestsTable.svelte
@@ -58,6 +58,10 @@
         return 'badge-neutral';
     }
   }
+
+  function accountTitle(account: RequestListItem['serving_account']): string | undefined {
+    return account ? `${account.provider_id}/${account.account}` : undefined;
+  }
 </script>
 
 <!-- Key cell contents (name + prefix subtitle, or bare prefix). Shared between the
@@ -69,16 +73,17 @@
   {:else}
     <code class="font-mono text-ink-strong">{r.key_prefix}</code>
   {/if}
+  <span aria-hidden="true"> </span>
 {/snippet}
 
 <div class="table-wrap">
   <table class="table-base">
     <thead class="table-head">
       <tr>
-        <th class="px-3 py-2" title={$t('The unique trace ID recorded for this request.')}
-          >{$t('Request ID')}</th
-        >
         <th class="px-3 py-2" title={$t('When the gateway received the request.')}>{$t('Time')}</th>
+        <th class="px-3 py-2" title={$t('Whether the request succeeded or returned an error.')}
+          >{$t('Status')}</th
+        >
         {#if showKey}
           <th
             class="px-3 py-2"
@@ -86,36 +91,28 @@
             >{$t('Key')}</th
           >
         {/if}
-        <th class="px-3 py-2" title={$t('The model the client asked for in its request.')}
-          >{$t('Requested model')}</th
+        <th class="px-3 py-2" title={$t('The concrete provider that served this request.')}
+          >{$t('Provider')}</th
         >
         <th
           class="px-3 py-2"
-          title={$t('Task type the classifier detected (e.g. coding, json, vision).')}
-          >{$t('Task')}</th
+          title={$t('The subscription account that served this request, when applicable.')}
+          >{$t('Subscription account')}</th
         >
-        <th
-          class="px-3 py-2"
-          title={$t('Estimated request difficulty used to pick a quality tier.')}
-          >{$t('Complexity')}</th
+        <th class="px-3 py-2" title={$t('The model that actually handled the request.')}
+          >{$t('Served model')}</th
         >
-        <th class="px-3 py-2" title={$t('Which classification layer chose the lane.')}
-          >{$t('Decided by')}</th
+        <th class="px-3 py-2" title={$t('The model the client asked for in its request.')}
+          >{$t('Requested model')}</th
         >
         <th class="px-3 py-2" title={$t('The quality/cost tier the request was routed to.')}
           >{$t('Lane')}</th
         >
-        <th class="px-3 py-2" title={$t('The model that actually handled the request.')}
-          >{$t('Served model')}</th
-        >
         <th
           class="px-3 py-2"
           title={$t('How many fallback models were tried before one succeeded.')}
           >{$t('Fallbacks')}</th
         >
-        <th class="px-3 py-2" title={$t('Whether the request succeeded or returned an error.')}
-          >{$t('Status')}</th
-        >
         <th class="px-3 py-2" title={$t('End-to-end response time in milliseconds.')}
           >{$t('Latency')}</th
         >
@@ -131,9 +128,25 @@
         <th class="px-3 py-2" title={$t('Estimated cost of the request in US dollars.')}
           >{$t('Cost')}</th
         >
+        <th
+          class="px-3 py-2"
+          title={$t('Task type the classifier detected (e.g. coding, json, vision).')}
+          >{$t('Task')}</th
+        >
+        <th
+          class="px-3 py-2"
+          title={$t('Estimated request difficulty used to pick a quality tier.')}
+          >{$t('Complexity')}</th
+        >
+        <th class="px-3 py-2" title={$t('Which classification layer chose the lane.')}
+          >{$t('Decided by')}</th
+        >
         <th class="px-3 py-2" title={$t('The error class, if the request failed.')}
           >{$t('Error')}</th
         >
+        <th class="px-3 py-2" title={$t('The unique trace ID recorded for this request.')}
+          >{$t('Request ID')}</th
+        >
       </tr>
     </thead>
     <tbody>
@@ -146,14 +159,14 @@
           class="table-row cursor-pointer"
           onclick={(e) => onRowClick(e, r.trace_id)}
         >
+          <td class="px-3 py-2 text-ink-body">{formatTs(r.ts)}</td>
           <td class="px-3 py-2">
-            <a
-              class="link-inline block max-w-[7rem] truncate font-mono text-ink-strong lg:max-w-none"
-              href={detailHref(r.trace_id)}
-              title={r.trace_id}>{r.trace_id}</a
-            >
+            {#if r.status === 'error'}
+              <span class="badge-error">{$t('error')}</span>
+            {:else}
+              <span class="badge-ok">{$t('ok')}</span>
+            {/if}
           </td>
-          <td class="px-3 py-2 text-ink-body">{formatTs(r.ts)}</td>
           {#if showKey}
             <td class="px-3 py-2">
               {#if r.key_id && onKeyFilter}
@@ -184,34 +197,41 @@
               {/if}
             </td>
           {/if}
-          <td class="px-3 py-2 text-ink-body">{r.requested_model ?? '—'}</td>
-          <td class="px-3 py-2 text-ink-body">{r.task_type || '—'}</td>
-          <td class="px-3 py-2 text-ink-body">{r.complexity || '—'}</td>
-          <td class="px-3 py-2">
-            <span data-testid="decided-by" class={decidedByClass(r.decided_by)}>{r.decided_by}</span
-            >
+          <td class="px-3 py-2 font-mono text-ink-body">{r.served_provider ?? '—'}</td>
+          <td
+            class="px-3 py-2 font-mono text-ink-body"
+            title={accountTitle(r.serving_account)}
+          >
+            {r.serving_account?.account ?? '—'}
           </td>
-          <td class="px-3 py-2 text-ink-body">{r.lane || '—'}</td>
           <td class="px-3 py-2 text-ink-body">{r.final_model ?? '—'}</td>
+          <td class="px-3 py-2 text-ink-body">{r.requested_model ?? '—'}</td>
+          <td class="px-3 py-2 text-ink-body">{r.lane || '—'}</td>
           <td class="px-3 py-2 text-ink-body">{r.fallback_count}</td>
-          <td class="px-3 py-2">
-            {#if r.status === 'error'}
-              <span class="badge-error">{$t('error')}</span>
-            {:else}
-              <span class="badge-ok">{$t('ok')}</span>
-            {/if}
-          </td>
           <td class="px-3 py-2 font-mono text-ink-body">{r.latency_ms}ms</td>
           <td data-testid="cell-tps" class="px-3 py-2 font-mono text-ink-body"
             >{formatTps(r.tps)}</td
           >
           <td class="px-3 py-2"><TokensCell usage={r.usage} /></td>
           <td class="px-3 py-2 font-mono text-ink-body">{formatUsd(r.cost_usd)}</td>
+          <td class="px-3 py-2 text-ink-body">{r.task_type || '—'}</td>
+          <td class="px-3 py-2 text-ink-body">{r.complexity || '—'}</td>
+          <td class="px-3 py-2">
+            <span data-testid="decided-by" class={decidedByClass(r.decided_by)}>{r.decided_by}</span
+            >
+          </td>
           <td
             class="px-3 py-2 {r.error_class ? 'text-red-600' : 'text-ink-muted'}"
             title={r.error_class ?? undefined}
             >{r.error_class ? $t(attemptCodeLabel(r.error_class)) : '—'}</td
           >
+          <td class="px-3 py-2">
+            <a
+              class="link-inline block max-w-[7rem] truncate font-mono text-ink-strong lg:max-w-none"
+              href={detailHref(r.trace_id)}
+              title={r.trace_id}>{r.trace_id}</a
+            >
+          </td>
         </tr>
       {/each}
     </tbody>
diff --git a/apps/admin/src/lib/components/RequestsTable.test.ts b/apps/admin/src/lib/components/RequestsTable.test.ts
index 1fe4dd1c..81a755fc 100644
--- a/apps/admin/src/lib/components/RequestsTable.test.ts
+++ b/apps/admin/src/lib/components/RequestsTable.test.ts
@@ -19,6 +19,8 @@ function item(overrides: Partial<RequestListItem> = {}): RequestListItem {
     complexity: 'high',
     decided_by: 'rules',
     lane: 'premium',
+    served_provider: 'anthropic',
+    serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
     final_model: 'claude-x',
     fallback_count: 1,
     status: 'ok',
diff --git a/apps/admin/src/locales/en.json b/apps/admin/src/locales/en.json
index 59ea4271..c752c219 100644
--- a/apps/admin/src/locales/en.json
+++ b/apps/admin/src/locales/en.json
@@ -564,6 +564,7 @@
   "Provider attempts (execution fallback)": "Provider attempts (execution fallback)",
   "provider error": "provider error",
   "Provider error": "Provider error",
+  "Provider model": "Provider model",
   "Provider not configured": "Provider not configured",
   "Providers": "Providers",
   "Proxy": "Proxy",
@@ -725,6 +726,7 @@
   "Streaming required": "Streaming required",
   "Streaming response — assembled from the recorded SSE stream.": "Streaming response — assembled from the recorded SSE stream.",
   "Subject": "Subject",
+  "Subscription account": "Subscription account",
   "Subscription Providers": "Subscription Providers",
   "Success": "Success",
   "Success rate": "Success rate",
@@ -743,6 +745,7 @@
   "The": "The",
   "The API key that authenticated the client, shown by prefix only.": "The API key that authenticated the client, shown by prefix only.",
   "The classifier decides which lane each request is routed to.": "The classifier decides which lane each request is routed to.",
+  "The concrete provider that served this request.": "The concrete provider that served this request.",
   "The error class, if the request failed.": "The error class, if the request failed.",
   "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.": "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.",
   "The fallback limit for any key without its own per-key value. 0 means unlimited.": "The fallback limit for any key without its own per-key value. 0 means unlimited.",
@@ -757,6 +760,7 @@
   "The request body as received from the client — before memory injection and translation.": "The request body as received from the client — before memory injection and translation.",
   "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.": "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.",
   "The subject is fixed — it identifies which fact a newer one supersedes.": "The subject is fixed — it identifies which fact a newer one supersedes.",
+  "The subscription account that served this request, when applicable.": "The subscription account that served this request, when applicable.",
   "The trace may not exist or has expired.": "The trace may not exist or has expired.",
   "The unique trace ID recorded for this request.": "The unique trace ID recorded for this request.",
   "The verdict that routed this request — the badge shows which layer decided it.": "The verdict that routed this request — the badge shows which layer decided it.",
@@ -836,6 +840,7 @@
   "View full key": "View full key",
   "View image": "View image",
   "Visual context compression": "Visual context compression",
+  "wire:": "wire:",
   "vs day before yesterday": "vs day before yesterday",
   "vs yesterday": "vs yesterday",
   "Waiting for authorization…": "Waiting for authorization…",
diff --git a/apps/admin/src/locales/ja.json b/apps/admin/src/locales/ja.json
index f17bcdbf..1b8e3854 100644
--- a/apps/admin/src/locales/ja.json
+++ b/apps/admin/src/locales/ja.json
@@ -564,6 +564,7 @@
   "Provider attempts (execution fallback)": "プロバイダー試行（実行フォールバック）",
   "provider error": "プロバイダーエラー",
   "Provider error": "プロバイダーエラー",
+  "Provider model": "上流モデル",
   "Provider not configured": "プロバイダー未設定",
   "Providers": "プロバイダー",
   "Proxy": "プロキシ",
@@ -725,6 +726,7 @@
   "Streaming required": "ストリーミングが必須",
   "Streaming response — assembled from the recorded SSE stream.": "ストリーミング応答——記録された SSE ストリームから結合して復元したものです。",
   "Subject": "サブジェクト",
+  "Subscription account": "サブスクリプションアカウント",
   "Subscription Providers": "サブスクリプションプロバイダー",
   "Success": "成功",
   "Success rate": "成功率",
@@ -743,6 +745,7 @@
   "The": "",
   "The API key that authenticated the client, shown by prefix only.": "クライアントを認証したAPIキー。プレフィックスのみ表示されます。",
   "The classifier decides which lane each request is routed to.": "分類器は各リクエストをどのレーンにルーティングするかを決定します。",
+  "The concrete provider that served this request.": "このリクエストを実際に処理した具体的なプロバイダー。",
   "The error class, if the request failed.": "リクエストが失敗した場合のエラークラス。",
   "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.": "プロバイダーに送信された正確なリクエスト — メモリ注入とプロトコル変換後。これはモデルが実際に受信した内容です。",
   "The fallback limit for any key without its own per-key value. 0 means unlimited.": "キー独自の値が設定されていないキーに適用されるフォールバック制限。0 は無制限を意味します。",
@@ -757,6 +760,7 @@
   "The request body as received from the client — before memory injection and translation.": "クライアントから受信したリクエスト本文 — メモリ注入と変換の前。",
   "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.": "サーバーはベースとなるオリジン + /mcp で MCP Streamable HTTP トランスポートを使用し、API キーを Bearer トークンとして認証します。ゲートウェイで memory.mcp.enabled を有効にしてください。それまで /mcp は 404 を返します。",
   "The subject is fixed — it identifies which fact a newer one supersedes.": "サブジェクトは固定です — 新しいファクトがどのファクトを置き換えるかを識別します。",
+  "The subscription account that served this request, when applicable.": "該当する場合、このリクエストを処理したサブスクリプションアカウント。",
   "The trace may not exist or has expired.": "その trace は存在しないか、期限切れの可能性があります。",
   "The unique trace ID recorded for this request.": "このリクエストに記録された一意のトレース ID。",
   "The verdict that routed this request — the badge shows which layer decided it.": "このリクエストをルーティングした判定結果——バッジがどのレイヤーの判定かを示します。",
@@ -836,6 +840,7 @@
   "View full key": "完全なキーを表示",
   "View image": "画像を表示",
   "Visual context compression": "視覚コンテキスト圧縮",
+  "wire:": "上流:",
   "vs day before yesterday": "一昨日比",
   "vs yesterday": "前日比",
   "Waiting for authorization…": "認可を待機中…",
diff --git a/apps/admin/src/locales/ko.json b/apps/admin/src/locales/ko.json
index 19d6fe7c..b7bce3d9 100644
--- a/apps/admin/src/locales/ko.json
+++ b/apps/admin/src/locales/ko.json
@@ -564,6 +564,7 @@
   "Provider attempts (execution fallback)": "공급자 시도 (실행 폴백)",
   "provider error": "공급자 오류",
   "Provider error": "공급자 오류",
+  "Provider model": "상위 모델",
   "Provider not configured": "공급자 미설정",
   "Providers": "공급자",
   "Proxy": "프록시",
@@ -725,6 +726,7 @@
   "Streaming required": "스트리밍 필수",
   "Streaming response — assembled from the recorded SSE stream.": "스트리밍 응답 — 기록된 SSE 스트림에서 조합해 복원했습니다.",
   "Subject": "주제",
+  "Subscription account": "구독 계정",
   "Subscription Providers": "구독 제공업체",
   "Success": "성공",
   "Success rate": "성공률",
@@ -743,6 +745,7 @@
   "The": "",
   "The API key that authenticated the client, shown by prefix only.": "클라이언트를 인증한 API 키이며, 접두사로만 표시됩니다.",
   "The classifier decides which lane each request is routed to.": "분류기는 각 요청이 라우팅될 레인을 결정합니다.",
+  "The concrete provider that served this request.": "이 요청을 실제로 처리한 구체적인 제공업체입니다.",
   "The error class, if the request failed.": "요청이 실패한 경우 오류 클래스입니다.",
   "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.": "제공자에게 전송된 정확한 요청 — 메모리 주입 및 프로토콜 변환 이후. 이것이 모델이 실제로 받은 내용입니다.",
   "The fallback limit for any key without its own per-key value. 0 means unlimited.": "키별 값이 없는 키에 적용되는 대체 제한입니다. 0은 무제한을 의미합니다.",
@@ -757,6 +760,7 @@
   "The request body as received from the client — before memory injection and translation.": "클라이언트에서 수신된 요청 본문 — 메모리 주입 및 번역 이전.",
   "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.": "서버는 베이스 오리진 + /mcp에서 MCP Streamable HTTP 전송을 사용하며, API 키를 Bearer 토큰으로 인증합니다. 게이트웨이에서 memory.mcp.enabled로 활성화하세요. 그 전까지 /mcp는 404를 반환합니다.",
   "The subject is fixed — it identifies which fact a newer one supersedes.": "주제는 고정되어 있습니다 — 어떤 사실이 더 새로운 사실로 대체되는지 식별합니다.",
+  "The subscription account that served this request, when applicable.": "해당하는 경우 이 요청을 처리한 구독 계정입니다.",
   "The trace may not exist or has expired.": "해당 trace가 존재하지 않거나 만료되었을 수 있습니다.",
   "The unique trace ID recorded for this request.": "이 요청에 대해 기록된 고유 추적 ID입니다.",
   "The verdict that routed this request — the badge shows which layer decided it.": "이 요청을 라우팅한 판정 결과 — 배지가 어느 레이어의 판정인지 표시합니다.",
@@ -836,6 +840,7 @@
   "View full key": "전체 키 보기",
   "View image": "이미지 보기",
   "Visual context compression": "시각 컨텍스트 압축",
+  "wire:": "상위:",
   "vs day before yesterday": "그저께 대비",
   "vs yesterday": "전일 대비",
   "Waiting for authorization…": "승인 대기 중…",
diff --git a/apps/admin/src/locales/zh-hans.json b/apps/admin/src/locales/zh-hans.json
index 25df7957..6cf7450c 100644
--- a/apps/admin/src/locales/zh-hans.json
+++ b/apps/admin/src/locales/zh-hans.json
@@ -564,6 +564,7 @@
   "Provider attempts (execution fallback)": "供应商尝试（执行回退）",
   "provider error": "供应商错误",
   "Provider error": "供应商返回错误",
+  "Provider model": "上游模型",
   "Provider not configured": "供应商未配置",
   "Providers": "提供商",
   "Proxy": "代理",
@@ -725,6 +726,7 @@
   "Streaming required": "必须使用流式",
   "Streaming response — assembled from the recorded SSE stream.": "流式响应——已从记录的 SSE 流合并还原。",
   "Subject": "主题",
+  "Subscription account": "订阅账号",
   "Subscription Providers": "订阅提供商",
   "Success": "成功",
   "Success rate": "成功率",
@@ -743,6 +745,7 @@
   "The": "",
   "The API key that authenticated the client, shown by prefix only.": "用于客户端鉴权的 API 密钥，仅显示前缀。",
   "The classifier decides which lane each request is routed to.": "分类器决定每个请求被路由到哪个 lane。",
+  "The concrete provider that served this request.": "实际处理此请求的具体提供商。",
   "The error class, if the request failed.": "请求失败时的错误类别。",
   "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.": "发送给提供商的确切请求——在记忆注入和协议转换之后。这就是模型实际收到的内容。",
   "The fallback limit for any key without its own per-key value. 0 means unlimited.": "未单独设置速率的密钥的兜底限制。0 表示不限。",
@@ -757,6 +760,7 @@
   "The request body as received from the client — before memory injection and translation.": "从客户端收到的请求正文——在记忆注入和转换之前。",
   "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.": "服务器在裸源地址 + /mcp 上使用 MCP Streamable HTTP 传输，并以 Bearer 令牌形式用你的 API 密钥进行认证。需在网关上通过 memory.mcp.enabled 启用——在此之前 /mcp 返回 404。",
   "The subject is fixed — it identifies which fact a newer one supersedes.": "主题是固定的——它标识较新的事实取代了哪个事实。",
+  "The subscription account that served this request, when applicable.": "实际处理此请求的订阅账号（如适用）。",
   "The trace may not exist or has expired.": "该 trace 可能不存在或已过期。",
   "The unique trace ID recorded for this request.": "为此请求记录的唯一跟踪 ID。",
   "The verdict that routed this request — the badge shows which layer decided it.": "本请求最终采用的分类结论——徽章标明由哪一层判定。",
@@ -836,6 +840,7 @@
   "View full key": "查看完整密钥",
   "View image": "查看图片",
   "Visual context compression": "视觉上下文压缩",
+  "wire:": "上游：",
   "vs day before yesterday": "对比前天",
   "vs yesterday": "对比昨日",
   "Waiting for authorization…": "正在等待授权…",
diff --git a/apps/admin/src/locales/zh-hant.json b/apps/admin/src/locales/zh-hant.json
index 725bdd92..320ce67f 100644
--- a/apps/admin/src/locales/zh-hant.json
+++ b/apps/admin/src/locales/zh-hant.json
@@ -564,6 +564,7 @@
   "Provider attempts (execution fallback)": "供應商嘗試（執行回退）",
   "provider error": "供應商錯誤",
   "Provider error": "供應商回傳錯誤",
+  "Provider model": "上游模型",
   "Provider not configured": "供應商未設定",
   "Providers": "供應商",
   "Proxy": "代理",
@@ -725,6 +726,7 @@
   "Streaming required": "必須使用串流",
   "Streaming response — assembled from the recorded SSE stream.": "串流回應——已從記錄的 SSE 串流合併還原。",
   "Subject": "主題",
+  "Subscription account": "訂閱帳號",
   "Subscription Providers": "訂閱提供者",
   "Success": "成功",
   "Success rate": "成功率",
@@ -743,6 +745,7 @@
   "The": "",
   "The API key that authenticated the client, shown by prefix only.": "用於用戶端鑑權的 API 金鑰，僅顯示前綴。",
   "The classifier decides which lane each request is routed to.": "分類器決定每個請求被路由到哪個 lane。",
+  "The concrete provider that served this request.": "實際處理此請求的具體供應商。",
   "The error class, if the request failed.": "請求失敗時的錯誤類別。",
   "The exact request sent to the provider — after memory injection and protocol translation. This is what the model actually received.": "傳送給供應商的確切請求——在記憶注入和協定轉換之後。這就是模型實際收到的內容。",
   "The fallback limit for any key without its own per-key value. 0 means unlimited.": "未單獨設定速率的金鑰的兜底限制。0 表示不限。",
@@ -757,6 +760,7 @@
   "The request body as received from the client — before memory injection and translation.": "從用戶端收到的請求正文——在記憶注入和轉換之前。",
   "The server uses the MCP Streamable HTTP transport at the bare origin + /mcp and authenticates with your API key as a bearer token. Enable it on the gateway with memory.mcp.enabled — until then /mcp returns 404.": "伺服器在裸源位址 + /mcp 上使用 MCP Streamable HTTP 傳輸，並以 Bearer 權杖形式用你的 API 金鑰進行認證。需在閘道上透過 memory.mcp.enabled 啟用——在此之前 /mcp 回傳 404。",
   "The subject is fixed — it identifies which fact a newer one supersedes.": "主題是固定的——它標識較新的事實取代了哪個事實。",
+  "The subscription account that served this request, when applicable.": "實際處理此請求的訂閱帳號（如適用）。",
   "The trace may not exist or has expired.": "該 trace 可能不存在或已過期。",
   "The unique trace ID recorded for this request.": "為此請求記錄的唯一追蹤 ID。",
   "The verdict that routed this request — the badge shows which layer decided it.": "本請求最終採用的分類結論——徽章標明由哪一層判定。",
@@ -836,6 +840,7 @@
   "View full key": "查看完整金鑰",
   "View image": "檢視圖片",
   "Visual context compression": "視覺上下文壓縮",
+  "wire:": "上游：",
   "vs day before yesterday": "對比前天",
   "vs yesterday": "對比昨日",
   "Waiting for authorization…": "正在等待授權…",
diff --git a/apps/admin/src/routes/keys/[keyId]/key-detail.test.ts b/apps/admin/src/routes/keys/[keyId]/key-detail.test.ts
index 61d67fd6..a88f1627 100644
--- a/apps/admin/src/routes/keys/[keyId]/key-detail.test.ts
+++ b/apps/admin/src/routes/keys/[keyId]/key-detail.test.ts
@@ -153,6 +153,8 @@ function requestItem(traceId: string): RequestListItem {
     complexity: 'complex',
     decided_by: 'rules',
     lane: 'balanced',
+    served_provider: 'openai',
+    serving_account: null,
     final_model: 'gpt-4o',
     fallback_count: 0,
     status: 'ok',
diff --git a/apps/admin/src/routes/requests/[traceId]/+page.svelte b/apps/admin/src/routes/requests/[traceId]/+page.svelte
index fbeb1eb2..2f8db35a 100644
--- a/apps/admin/src/routes/requests/[traceId]/+page.svelte
+++ b/apps/admin/src/routes/requests/[traceId]/+page.svelte
@@ -184,18 +184,6 @@
             {/if}
           </dd>
         </div>
-        <div>
-          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Requested model')}</dt>
-          <dd class="mt-0.5 font-mono text-ink-body">{d.requested_model ?? '—'}</dd>
-        </div>
-        <div>
-          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Served model')}</dt>
-          <dd class="mt-0.5 font-mono text-ink-body">{d.final_model ?? '—'}</dd>
-        </div>
-        <div>
-          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Lane')}</dt>
-          <dd class="mt-0.5 text-ink-body">{d.lane || '—'}</dd>
-        </div>
         <div>
           <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Status')}</dt>
           <dd class="mt-0.5">
@@ -206,6 +194,35 @@
             {/if}
           </dd>
         </div>
+        <div>
+          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Provider')}</dt>
+          <dd class="mt-0.5 font-mono text-ink-body">{d.served_provider ?? '—'}</dd>
+        </div>
+        <div>
+          <dt class="text-xs uppercase tracking-wide text-slate-400">
+            {$t('Subscription account')}
+          </dt>
+          <dd
+            class="mt-0.5 font-mono text-ink-body"
+            title={d.serving_account
+              ? `${d.serving_account.provider_id}/${d.serving_account.account}`
+              : undefined}
+          >
+            {d.serving_account?.account ?? '—'}
+          </dd>
+        </div>
+        <div>
+          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Served model')}</dt>
+          <dd class="mt-0.5 font-mono text-ink-body">{d.final_model ?? '—'}</dd>
+        </div>
+        <div>
+          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Requested model')}</dt>
+          <dd class="mt-0.5 font-mono text-ink-body">{d.requested_model ?? '—'}</dd>
+        </div>
+        <div>
+          <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Lane')}</dt>
+          <dd class="mt-0.5 text-ink-body">{d.lane || '—'}</dd>
+        </div>
         <div>
           <dt class="text-xs uppercase tracking-wide text-slate-400">{$t('Latency')}</dt>
           <dd class="mt-0.5 font-mono text-ink-body">
diff --git a/apps/admin/src/routes/requests/requests.test.ts b/apps/admin/src/routes/requests/requests.test.ts
index 14743524..382b59dc 100644
--- a/apps/admin/src/routes/requests/requests.test.ts
+++ b/apps/admin/src/routes/requests/requests.test.ts
@@ -30,6 +30,8 @@ function item(traceId: string, overrides: Partial<RequestListItem> = {}): Reques
     complexity: 'high',
     decided_by: 'rules',
     lane: 'premium',
+    served_provider: 'anthropic',
+    serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
     final_model: 'claude-x',
     fallback_count: 1,
     status: 'ok',
@@ -70,6 +72,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
     key_prefix: 'helm_live_ab12',
     key_name: 'Production backend',
     requested_model: 'gpt-4o',
+    served_provider: 'anthropic',
+    serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
     final_model: 'claude-x',
     lane: 'premium',
     status: 'ok',
@@ -96,6 +100,8 @@ function detail(overrides: Partial<RequestDetail> = {}): RequestDetail {
       {
         model: 'claude-x',
         provider: 'anthropic',
+        provider_model: 'claude-x',
+        serving_account: { provider_id: 'anthropic', account: 'claude-team-a' },
         outcome: 'success',
         latency_ms: 340,
         error_detail: null,
@@ -145,6 +151,8 @@ describe('requests list page', () => {
     expect(first).toHaveTextContent('high'); // complexity
     expect(first).toHaveTextContent('rules'); // decided_by
     expect(first).toHaveTextContent('premium'); // lane
+    expect(first).toHaveTextContent('anthropic'); // provider
+    expect(first).toHaveTextContent('claude-team-a'); // subscription account
     expect(first).toHaveTextContent('claude-x'); // final_model
     expect(first).toHaveTextContent('460'); // latency_ms
     expect(within(first).getByTestId('cell-tps')).toHaveTextContent('200 tok/s'); // true TPS
@@ -206,15 +214,16 @@ describe('requests list page', () => {
     );
   });
 
-  it('shows the request ID as the first column and the recorded time, with no separate "view" action', () => {
+  it('puts high-signal columns before classifier details, with trace ID at the end', () => {
     render(ListPage, {
       data: listData([item('tr_first', { ts: '2026-05-31T10:00:00Z' })]),
     });
     const cells = screen.getByTestId('request-row').querySelectorAll('td');
-    // Request ID is the FIRST column.
-    expect(cells[0]).toHaveTextContent('tr_first');
-    // The time column (second) renders the recorded timestamp (year is locale-stable).
-    expect(cells[1]).toHaveTextContent('2026');
+    expect(cells[0]).toHaveTextContent('2026'); // time first
+    expect(cells[1]).toHaveTextContent('ok'); // status before diagnostics
+    expect(cells[3]).toHaveTextContent('anthropic'); // provider before model/classifier
+    expect(cells[4]).toHaveTextContent('claude-team-a'); // concrete subscription account
+    expect(cells[cells.length - 1]).toHaveTextContent('tr_first'); // trace id still available
     // The trailing "view" link is gone — the whole row is the link now.
     expect(screen.queryByText('view')).not.toBeInTheDocument();
   });
@@ -367,13 +376,15 @@ describe('requests detail page', () => {
     expect(screen.getByTestId('payload-summary')).toHaveTextContent(/not recorded/i);
   });
 
-  it('renders a Request summary card with key (name+prefix), requested+served model, lane, status, latency', () => {
+  it('renders a Request summary card with key, provider/account, requested+served model, lane, status, latency', () => {
     render(DetailPage, {
       data: { detail: detail(), payload: { captured: false }, traceId: 'tr_1' },
     });
     const summary = screen.getByTestId('request-summary');
     expect(summary).toHaveTextContent('Production backend'); // key name
     expect(summary).toHaveTextContent('helm_live_ab12'); // key prefix (traceability)
+    expect(summary).toHaveTextContent('anthropic'); // concrete provider
+    expect(summary).toHaveTextContent('claude-team-a'); // concrete subscription account
     expect(summary).toHaveTextContent('gpt-4o'); // requested model
     expect(summary).toHaveTextContent('claude-x'); // served model
     expect(summary).toHaveTextContent('premium'); // lane
diff --git a/apps/gateway/src/routes/admin/admin.test.ts b/apps/gateway/src/routes/admin/admin.test.ts
index 3a595eda..af32e98b 100644
--- a/apps/gateway/src/routes/admin/admin.test.ts
+++ b/apps/gateway/src/routes/admin/admin.test.ts
@@ -296,6 +296,7 @@ function decision(traceId: string, lane: string): DecisionRecord {
     memory: null,
     usage: null,
     generation_ms: null,
+    serving_account: null,
   };
 }
 
diff --git a/apps/gateway/src/routes/chat.ts b/apps/gateway/src/routes/chat.ts
index 41623c64..5ab93e0d 100644
--- a/apps/gateway/src/routes/chat.ts
+++ b/apps/gateway/src/routes/chat.ts
@@ -34,7 +34,7 @@ import { streamSSE } from "hono/streaming";
 import type { AppEnv } from "../app.js";
 import { INTERNAL_API_KEY_ID } from "../internal-key.js";
 import { HelmHttpError } from "../middleware/error-handler.js";
-import type { ServingAccount } from "../runtime/serving-account.js";
+import { type ServingAccount, stampServingAccount } from "../runtime/serving-account.js";
 import type { WriteQueue } from "../runtime/write-queue.js";
 import { downgradeClientFastModeIfDisallowed } from "./fast-mode.js";
 import { atEventBoundary, HEARTBEAT_COMMENT, withHeartbeat } from "./heartbeat.js";
@@ -867,6 +867,7 @@ export function registerChatRoutes(app: Hono<AppEnv>, deps: ChatRouteDeps): void
             c.get("logger").log("warn", "cost.stream_backfill_failed", { trace_id: traceId });
           }
           await capturePayload(captureOn ? rawSse : null, result.upstreamRequest ?? null);
+          stampServingAccount(result.decision, servingAccount);
           await persist(result.decision);
           // Usage-budget settle (streamed): runs HERE — after the usage tail
           // backfilled the streamed cost — so the spend dimension settles the real
@@ -918,6 +919,7 @@ export function registerChatRoutes(app: Hono<AppEnv>, deps: ChatRouteDeps): void
       result.body !== null ? JSON.stringify(result.body) : null,
       result.upstreamRequest ?? null,
     );
+    stampServingAccount(result.decision, servingAccount);
     await persist(result.decision);
     if (result.final.status === "error" || result.body === null) {
       const error =
diff --git a/apps/gateway/src/routes/gemini.ts b/apps/gateway/src/routes/gemini.ts
index eba38051..3947871d 100644
--- a/apps/gateway/src/routes/gemini.ts
+++ b/apps/gateway/src/routes/gemini.ts
@@ -9,6 +9,7 @@ import { streamSSE } from "hono/streaming";
 import type { AppEnv } from "../app.js";
 import { type ConcurrencyGatePort, concurrencyReleaseGuard } from "../middleware/concurrency.js";
 import { estimateRequestTokens } from "../middleware/estimate-tokens.js";
+import { stampServingAccount } from "../runtime/serving-account.js";
 import { atEventBoundary, HEARTBEAT_COMMENT, withHeartbeat } from "./heartbeat.js";
 import { resolveMemoryScope } from "./memory-scope.js";
 import type { MessagesIdentity, PipelineRunResult, RouteError } from "./messages.js";
@@ -430,6 +431,7 @@ export function registerGeminiRoute(app: Hono<AppEnv>, deps: GeminiRouteDeps): v
           // backfill) already mutated result.decision. result.decision exists even
           // on a pre-stream failure, so a failed stream still records. Fail-open.
           if (deps.record) {
+            stampServingAccount(result.decision, result.servingAccount ?? null);
             await recordServed(
               deps.record,
               {
@@ -460,6 +462,7 @@ export function registerGeminiRoute(app: Hono<AppEnv>, deps: GeminiRouteDeps): v
       // /admin/requests. responseJson null = no body. result.decision exists even
       // on failure. Fail-open inside recordServed.
       if (deps.record) {
+        stampServingAccount(result.decision, result.servingAccount ?? null);
         await recordServed(
           deps.record,
           {
@@ -485,6 +488,7 @@ export function registerGeminiRoute(app: Hono<AppEnv>, deps: GeminiRouteDeps): v
     // Record the served (non-stream) request: telemetry row (→ /admin/requests) +
     // verbatim request/response body. Mirrors chat.ts. Fail-open inside recordServed.
     if (deps.record) {
+      stampServingAccount(result.decision, result.servingAccount ?? null);
       await recordServed(
         deps.record,
         {
diff --git a/apps/gateway/src/routes/image-telemetry.ts b/apps/gateway/src/routes/image-telemetry.ts
index 2522a7ee..45616075 100644
--- a/apps/gateway/src/routes/image-telemetry.ts
+++ b/apps/gateway/src/routes/image-telemetry.ts
@@ -95,5 +95,6 @@ export function buildImageDecision(p: {
             cache_creation_tokens: null,
           },
     generation_ms: null,
+    serving_account: null,
   };
 }
diff --git a/apps/gateway/src/routes/messages-pipeline.ts b/apps/gateway/src/routes/messages-pipeline.ts
index 55706717..0dae6250 100644
--- a/apps/gateway/src/routes/messages-pipeline.ts
+++ b/apps/gateway/src/routes/messages-pipeline.ts
@@ -862,6 +862,9 @@ export function createMessagesPipeline(
         // The exact body forwarded upstream for the served attempt (post inject +
         // translation). Exposed so the three faces capture it into the payload table.
         upstreamRequest: result.upstreamRequest,
+        // Concrete subscription account selected by the OAuth pool, or null for a
+        // configured/non-OAuth provider. Routes stamp it just before telemetry write.
+        servingAccount,
         async collect(): Promise<unknown> {
           if (failure !== null) throw failure;
           // Native passthrough (#217): the upstream's response is already in the
diff --git a/apps/gateway/src/routes/messages.ts b/apps/gateway/src/routes/messages.ts
index d0686441..ab6805f5 100644
--- a/apps/gateway/src/routes/messages.ts
+++ b/apps/gateway/src/routes/messages.ts
@@ -13,6 +13,7 @@ import type { ContentfulStatusCode } from "hono/utils/http-status";
 import type { AppEnv } from "../app.js";
 import { type ConcurrencyGatePort, concurrencyReleaseGuard } from "../middleware/concurrency.js";
 import { estimateRequestTokens } from "../middleware/estimate-tokens.js";
+import { type ServingAccount, stampServingAccount } from "../runtime/serving-account.js";
 import { atEventBoundary, HEARTBEAT_COMMENT, withHeartbeat } from "./heartbeat.js";
 import { type MemoryKeyDefaults, resolveMemoryScope } from "./memory-scope.js";
 import { PipelineError } from "./messages-pipeline.js";
@@ -100,6 +101,7 @@ export interface PipelineRunResult {
    *  Each face writes it verbatim into the captured payload so the admin sees what the
    *  model actually received. Null/absent on a routing failure (no served attempt). */
   readonly upstreamRequest?: string | null;
+  readonly servingAccount?: ServingAccount | null;
   /** Drain the full (non-stream) result into ONE IR response object. */
   collect(): Promise<unknown>;
   /** The outbound-protocol event stream: one object per wire event. For Anthropic
@@ -592,6 +594,7 @@ export function registerMessagesRoute(app: Hono<AppEnv>, deps: MessagesRouteDeps
           // backfill) already mutated result.decision. result.decision exists even
           // on a pre-stream failure, so a failed stream still records. Fail-open.
           if (deps.record) {
+            stampServingAccount(result.decision, result.servingAccount ?? null);
             await recordServed(
               deps.record,
               {
@@ -629,6 +632,7 @@ export function registerMessagesRoute(app: Hono<AppEnv>, deps: MessagesRouteDeps
       // /admin/requests. responseJson null = no body. result.decision exists even
       // on failure. Fail-open inside recordServed.
       if (deps.record) {
+        stampServingAccount(result.decision, result.servingAccount ?? null);
         await recordServed(
           deps.record,
           {
@@ -654,6 +658,7 @@ export function registerMessagesRoute(app: Hono<AppEnv>, deps: MessagesRouteDeps
     // Record the served (non-stream) request: telemetry row (→ /admin/requests) +
     // verbatim request/response body. Mirrors chat.ts. Fail-open inside recordServed.
     if (deps.record) {
+      stampServingAccount(result.decision, result.servingAccount ?? null);
       await recordServed(
         deps.record,
         {
diff --git a/apps/gateway/src/routes/payload-capture.test.ts b/apps/gateway/src/routes/payload-capture.test.ts
index 4b2f47c3..8e2b4c17 100644
--- a/apps/gateway/src/routes/payload-capture.test.ts
+++ b/apps/gateway/src/routes/payload-capture.test.ts
@@ -482,6 +482,7 @@ function decision(): DecisionRecord {
     cost_breakdown: { eval_usd: null, completion_usd: null, total_usd: null },
     usage: null,
     generation_ms: null,
+    serving_account: null,
     final: { status: "ok", model_alias: "openai/gpt" },
   } as unknown as DecisionRecord;
 }
diff --git a/apps/gateway/src/routes/responses.ts b/apps/gateway/src/routes/responses.ts
index d9fc3920..ea507a47 100644
--- a/apps/gateway/src/routes/responses.ts
+++ b/apps/gateway/src/routes/responses.ts
@@ -7,6 +7,7 @@ import type { AppEnv } from "../app.js";
 import { type ConcurrencyGatePort, concurrencyReleaseGuard } from "../middleware/concurrency.js";
 import { HelmHttpError } from "../middleware/error-handler.js";
 import { estimateRequestTokens } from "../middleware/estimate-tokens.js";
+import { stampServingAccount } from "../runtime/serving-account.js";
 import { atEventBoundary, HEARTBEAT_COMMENT, withHeartbeat } from "./heartbeat.js";
 import { resolveMemoryScope } from "./memory-scope.js";
 import type { MessagesIdentity, PipelineRunResult } from "./messages.js";
@@ -648,6 +649,7 @@ export function registerResponsesRoute(app: Hono<AppEnv>, deps: ResponsesRouteDe
           // result.decision exists even on a pre-stream failure, so a failed stream
           // still records. Fail-open inside recordServed.
           if (deps.record && result !== null) {
+            stampServingAccount(result.decision, result.servingAccount ?? null);
             await recordServed(
               deps.record,
               {
@@ -724,6 +726,7 @@ export function registerResponsesRoute(app: Hono<AppEnv>, deps: ResponsesRouteDe
       // still appears in /admin/requests. responseJson null = no body produced.
       // result.decision exists even on failure. Fail-open inside recordServed.
       if (deps.record) {
+        stampServingAccount(result.decision, result.servingAccount ?? null);
         await recordServed(
           deps.record,
           {
@@ -764,6 +767,7 @@ export function registerResponsesRoute(app: Hono<AppEnv>, deps: ResponsesRouteDe
     // Record the served (non-stream) request: telemetry row (→ /admin/requests) +
     // verbatim request/response body. Mirrors chat.ts. Fail-open inside recordServed.
     if (deps.record) {
+      stampServingAccount(result.decision, result.servingAccount ?? null);
       await recordServed(
         deps.record,
         {
diff --git a/apps/gateway/src/runtime/serving-account.test.ts b/apps/gateway/src/runtime/serving-account.test.ts
index 00eb2b9e..82bd2b39 100644
--- a/apps/gateway/src/runtime/serving-account.test.ts
+++ b/apps/gateway/src/runtime/serving-account.test.ts
@@ -1,7 +1,9 @@
+import type { DecisionRecord } from "@helm/shared";
 import { describe, expect, it } from "vitest";
 import {
   markServingAccount,
   servedByAccount,
+  stampServingAccount,
   withServingAccountCapture,
 } from "./serving-account.js";
 
@@ -52,3 +54,24 @@ describe("servedByAccount", () => {
     expect(servedByAccount(ACCT, "anthropic-mirror/x")).toBe(false);
   });
 });
+
+describe("stampServingAccount", () => {
+  function decision(modelAlias: string | null = "anthropic/claude-opus-4-8"): DecisionRecord {
+    return {
+      final: { model_alias: modelAlias },
+      serving_account: null,
+    } as unknown as DecisionRecord;
+  }
+
+  it("stamps the final subscription account only when the served alias belongs to it", () => {
+    const d = decision();
+    stampServingAccount(d, ACCT);
+    expect(d.serving_account).toEqual({ provider_id: "anthropic", account: "default" });
+  });
+
+  it("clears stale selections after fallback to another provider", () => {
+    const d = decision("openai/gpt-5");
+    stampServingAccount(d, ACCT);
+    expect(d.serving_account).toBeNull();
+  });
+});
diff --git a/apps/gateway/src/runtime/serving-account.ts b/apps/gateway/src/runtime/serving-account.ts
index aff14b84..c9bef2d2 100644
--- a/apps/gateway/src/runtime/serving-account.ts
+++ b/apps/gateway/src/runtime/serving-account.ts
@@ -1,4 +1,5 @@
 import { AsyncLocalStorage } from "node:async_hooks";
+import type { DecisionRecord } from "@helm/shared";
 
 // Per-request "which OAuth subscription served this call" holder (providers page
 // Tier 2 usage attribution). The pool's `onSelect(account)` fires DEEP inside
@@ -59,3 +60,17 @@ export function servedByAccount(
   if (!servingAccount || !servedAlias) return false;
   return servedAlias.startsWith(`${servingAccount.providerId}/`);
 }
+
+export function stampServingAccount(
+  decision: DecisionRecord,
+  servingAccount: ServingAccount | null,
+): void {
+  if (!servingAccount || !servedByAccount(servingAccount, decision.final.model_alias)) {
+    decision.serving_account = null;
+    return;
+  }
+  decision.serving_account = {
+    provider_id: servingAccount.providerId,
+    account: servingAccount.account,
+  };
+}
diff --git a/implementation-notes.md b/implementation-notes.md
index 2473f198..e165826c 100644
--- a/implementation-notes.md
+++ b/implementation-notes.md
@@ -7,6 +7,16 @@
 
 ---
 
+## 2026-07-04 · 请求记录最终订阅账号并重排请求列表字段（Telemetry / Admin requests，docs/04/07/11，原则 1/5/7）
+
+- **背景（Lukin）**：请求 telemetry 只能看到最终 model/provider alias，无法确认使用订阅 provider 时最后落到哪个具体订阅账号；排查成本、额度、限流和账号池调度时缺少每次请求的账号级事实。
+- **记录决策**：`DecisionRecord` 新增 `serving_account: { provider_id, account } | null`。OAuth pool 仍只在执行深处通过 ALS 标记候选账号；真正落库前由 gateway 在最终结果已知后 stamp，避免 core/provider 合同依赖 Hono 或 admin。
+- **防误记边界**：只有最终 `final.model_alias` 仍属于被标记账号的 provider 前缀时才记录账号；如果订阅账号候选失败后 fallback 到其它 provider，写入 `null`，避免把 stale selection 错记成实际服务账号。
+- **尝试元数据决策**：`provider_attempts` 保留 `provider_name`、`provider_model`、passthrough/protocol mutation 等执行层已有元数据。Admin 只展示已记录字段，不从 alias 重新推断业务事实。
+- **UI 决策**：请求列表字段按排障优先级重排为：Time、Status、Key、Provider、Subscription account、Served model、Requested model、Lane、Fallbacks、Latency、TPS、Tokens、Cost、Task、Complexity、Decided by、Error、Request ID。Trace ID 仍可点击但降到末尾；详情摘要同步露出 Provider / Subscription account，执行链路每个 attempt 显示 provider、账号、alias 与上游 wire model。
+- **安全/隐私**：只记录 provider id 与订阅账号显示名，不记录 OAuth token、明文 API key 或正文；key 仍只显示 prefix。旧记录和非订阅 provider 统一显示 `—`/`null`。
+- **验证计划**：覆盖 shared schema round-trip、telemetry builder 元数据保留、gateway stamp guard、admin mapper、请求列表/详情/执行链路渲染；再跑目标 Vitest、typecheck、lint、build。
+
 ## 2026-07-04 · cheap-model 短低风险当前轮不被长历史抬价（Classifier / routing，docs/03/04，原则 2/4/5）
 
 - **背景（Lukin）**：生产 `openclaw` 24h 数据显示，部分请求显式请求 `gpt-5.4-mini` / cheap alias，最后一条 user 只有约 200 字且是 read/check/status 类低风险动作，但因为完整 transcript 很大、tools 很多，被 Layer-1 长上下文/工具信号抬到 `balanced/coding`，最终使用 `gpt-5.5`。
@@ -88,27 +98,10 @@
 - **边界**：其它确定性 400/413/422（例如坏参数、图片过大）仍按 `invalid_request` 终止，因为换候选无法修复请求本身。这里只针对明确的 reasoning-history 缺失错误 fail-open。
 - **验证**：新增执行器回归覆盖跨协议剥离 thinking/reasoning 控制，以及 DeepSeek 类 400 后继续 fallback；目标 `execute.test.ts` 119/119 绿，`pnpm typecheck` 绿，`pnpm lint` 退出 0（仅既有 style info）。
 
-## 2026-07-04 · memory idle-flush 防饥饿与受控追赶（Memory worker / store，docs/08/12，原则 3/7）
-
-- **背景（Lukin）**：生产 `openclaw` key 已持续写入 raw memory，但 `/admin/memory` 看不到事实/反思；排查发现后台 worker 不是完全停了，而是大量旧项目候选反复进入 idle-flush 队列，导致新项目长期排不到。
-- **根因**：Observer 实际按 `message_index, created_at, id` 读取消息并生成 `source_message_range`；但 `listIdleFlushCandidates()` 用 `created_at/id` 判断 observation 覆盖范围。旧线程在 observer 顺序下已覆盖，却在候选 SQL 里被误判为未覆盖，形成永远不会消失的假候选。
-- **查询决策**：SQLite / Postgres 的 idle-flush 候选判断改用与 `listMessages()` 完全一致的 tuple order，并支持 range 两端反向的历史数据；这样同一线程被覆盖后会真正退出候选集。
-- **公平性决策**：候选排序增加 `ROW_NUMBER() OVER (PARTITION BY owner_id, project_id, resource_id ...)`，按 scope rank 交错输出，避免 `ww/luke/skillstore` 这类旧项目 backlog 独占整页，使 `openclaw` 这类新项目也能进入处理窗口。
-- **追赶决策**：worker 支持单次 tick/wake 连续 drain 多批，但仍是串行执行；gateway 默认 `batchSize=50`、`maxBatchesPerDrain=10`、`maxDrainMs=30000`，并在批次之间让出事件循环。也就是说默认最多 500 个任务/轮，明显快于旧的 10/轮，但不会引入并发洪峰；线上可通过环境变量逐步调大。
-- **索引/迁移决策**：新增 memory thread scope 索引与 message observer-order 索引；迁移只在目标表/列真实存在时建索引，兼容早期被部分标记为已迁移的自托管库和测试 fixture。
-- **验证**：新增 SQLite/Postgres observer-order 覆盖回归、project fair-interleaving 回归、scheduler 多批 drain 与时间上限回归；目标 memory/scheduler/migration 测试 83/83 绿。
-
-## 2026-07-03 · 策略级 reasoning_effort 覆盖 Lane 默认值（Routing policies / Admin policies，docs/04/11，原则 2/5/6）
-
-- **背景（Lukin）**：Lane 已支持 `reasoning_effort`，但 policy 命中后只能强制车道，不能针对某类任务把思考等级调高/调低；这导致同一 Lane 内的请求无法按策略更细粒度控制推理预算。
-- **语义决策**：Policy 新增可选 action `reasoning_effort`，与 Lane 使用同一严格枚举：`none|minimal|low|medium|high|xhigh|max`。最终优先级为 **policy > selected lane > client request**；其中 `none` 是显式覆盖，可关闭 Lane 的高思考等级。
-- **first-match 决策**：`reasoning_effort` 跟 `use_lane` 一样取第一条命中 policy 的值；`allowed_lanes` 仍保持原有“所有命中策略取交集”的 cap 语义。这样不会让靠后的 restrict-only rule 意外改写思考等级。
-- **执行路径**：`routeRequest` 在生成 `ExecutionPlan` 后用 `policy.reasoning_effort ?? lane.reasoning_effort` 覆盖 `req.reasoning_effort` 并设置 `reasoning_effort_forced=true`，复用现有 translated/native passthrough 出站改写链路，不触碰 provider 选择与 fallback 机制。
-- **Admin 决策**：Policies 页面新增“Forced reasoning effort”下拉，复用 LaneEditor 的同一组选项；API client round-trip `reasoning_effort`，gateway 仍由 `PoliciesConfigSchema` 对整个 policy 列表 fail-closed 校验。
-- **验证计划**：新增 shared/core policy schema、policy engine、routeRequest、admin API client、PolicyRow 回归测试；再跑目标 Vitest、typecheck/lint/build。
-
-## 历史条目摘要（最近 2 条）
+## 历史条目摘要（最近 4 条）
 
+- **2026-07-04 · memory idle-flush 防饥饿与受控追赶（Memory worker / store，docs/08/12，原则 3/7）**：idle-flush 候选判断改用 observer-order tuple，按 scope rank 交错输出，并支持受控多批 drain，避免旧 backlog 饿死新项目。
+- **2026-07-03 · 策略级 reasoning_effort 覆盖 Lane 默认值（Routing policies / Admin policies，docs/04/11，原则 2/5/6）**：Policy action 可强制 reasoning_effort，优先级为 policy > selected lane > client request，复用现有执行改写链路。
 - **2026-07-03 · cron monitor 自动化请求降到低成本规则（Classifier / routing，docs/03/04，原则 2/4）**：monitor/cron + no-reply 标记命中时降到 `simple/economy`，但保留显式 coding keyword 升级路径，避免自动化探针误打高价模型。
 - **2026-07-03 · 上下文窗口超限按候选跳过处理（执行 fallback / streaming telemetry，docs/04/07，原则 5/8）**：`context_length_exceeded` / prompt-too-long 类错误按候选 `context_too_small` 跳过并继续 fallback，不熔断 provider。
 
diff --git a/packages/core/src/executor/attempt-record.ts b/packages/core/src/executor/attempt-record.ts
index 1173eaf4..09050942 100644
--- a/packages/core/src/executor/attempt-record.ts
+++ b/packages/core/src/executor/attempt-record.ts
@@ -1,4 +1,9 @@
-import type { AttemptErrorDetail } from "@helm/shared";
+import type {
+  AttemptErrorDetail,
+  NativePassthroughMutationLedger,
+  Protocol,
+  TargetProviderProtocol,
+} from "@helm/shared";
 
 // One row of the decision record's `provider_attempts` (docs/02). Field-for-field
 // aligned with shared `ProviderAttemptSchema` so this array feeds it directly.
@@ -19,4 +24,14 @@ export interface AttemptRecord {
   // Upstream failure detail for THIS attempt (admin-debug-error-detail). Non-null
   // only for a genuine pre-first-chunk provider failure; null for ok/skipped rows.
   error_detail: AttemptErrorDetail | null;
+  passthrough_considered?: boolean;
+  passthrough_used?: boolean;
+  passthrough_disable_reason?: string | null;
+  source_protocol?: Protocol | null;
+  target_provider_protocol?: TargetProviderProtocol | null;
+  response_protocol?: Protocol | null;
+  provider_name?: string | null;
+  provider_model?: string | null;
+  passthrough_mutations?: NativePassthroughMutationLedger;
+  request_mutations?: NativePassthroughMutationLedger;
 }
diff --git a/packages/core/src/routing/route-request.ts b/packages/core/src/routing/route-request.ts
index de2339b9..15fdcd4a 100644
--- a/packages/core/src/routing/route-request.ts
+++ b/packages/core/src/routing/route-request.ts
@@ -802,6 +802,8 @@ export async function routeRequest(
       usage: null,
       // Served-stream generation window is gateway-timed/stamped — null here.
       generation_ms: null,
+      // Concrete subscription account is gateway-stamped after execution.
+      serving_account: null,
     };
     deps.log(decision);
     return {
@@ -884,6 +886,8 @@ export async function routeRequest(
     usage: null,
     // Served-stream generation window is gateway-timed/stamped — null here.
     generation_ms: null,
+    // Concrete subscription account is gateway-stamped after execution.
+    serving_account: null,
   };
 
   deps.log(decision);
diff --git a/packages/core/src/signals/aggregate.test.ts b/packages/core/src/signals/aggregate.test.ts
index 5e5fd17c..2431ff6c 100644
--- a/packages/core/src/signals/aggregate.test.ts
+++ b/packages/core/src/signals/aggregate.test.ts
@@ -54,6 +54,7 @@ function makeRecord(over: {
     memory: null,
     usage: null,
     generation_ms: null,
+    serving_account: null,
   };
 }
 
diff --git a/packages/core/src/signals/collector.test.ts b/packages/core/src/signals/collector.test.ts
index 63fe0b40..61f7a62a 100644
--- a/packages/core/src/signals/collector.test.ts
+++ b/packages/core/src/signals/collector.test.ts
@@ -43,6 +43,7 @@ function makeRecord(taskType: string, lane: string): DecisionRecord {
     memory: null,
     usage: null,
     generation_ms: null,
+    serving_account: null,
   };
 }
 
diff --git a/packages/core/src/store/ports.test.ts b/packages/core/src/store/ports.test.ts
index 0d922179..b669310e 100644
--- a/packages/core/src/store/ports.test.ts
+++ b/packages/core/src/store/ports.test.ts
@@ -362,6 +362,7 @@ describe("Store ports are implementable contracts", () => {
       memory: null,
       usage: null,
       generation_ms: null,
+      serving_account: null,
     } satisfies DecisionRecord;
     await store.insert({ decision, apiKeyId: "k1", createdAt: new Date() });
     expect(await store.queryRecent(10)).toHaveLength(1);
diff --git a/packages/core/src/store/sqlite/telemetry.test.ts b/packages/core/src/store/sqlite/telemetry.test.ts
index 7dbd0547..27a96da6 100644
--- a/packages/core/src/store/sqlite/telemetry.test.ts
+++ b/packages/core/src/store/sqlite/telemetry.test.ts
@@ -53,6 +53,7 @@ function decision(requestId: string, overrides: Partial<DecisionRecord> = {}): D
     memory: null,
     usage: null,
     generation_ms: null,
+    serving_account: null,
     ...overrides,
   };
 }
diff --git a/packages/core/src/store/store-contract.test.ts b/packages/core/src/store/store-contract.test.ts
index 3017f26b..ab5ca0c3 100644
--- a/packages/core/src/store/store-contract.test.ts
+++ b/packages/core/src/store/store-contract.test.ts
@@ -132,6 +132,7 @@ function decision(requestId: string, overrides: Partial<DecisionRecord> = {}): D
     memory: null,
     usage: null,
     generation_ms: null,
+    serving_account: null,
     ...overrides,
   };
 }
diff --git a/packages/core/src/telemetry/decision.test.ts b/packages/core/src/telemetry/decision.test.ts
index cf62c736..290be483 100644
--- a/packages/core/src/telemetry/decision.test.ts
+++ b/packages/core/src/telemetry/decision.test.ts
@@ -200,6 +200,29 @@ describe("buildDecisionRecord", () => {
     expect(record.provider_attempts[2]?.cost_usd).toBe(0.002);
   });
 
+  it("4b. preserves provider metadata on each attempt", () => {
+    const record = buildDecisionRecord(
+      parts({
+        attempts: [
+          {
+            ...okAttempt("anthropic/claude-opus-4-8"),
+            provider_name: "anthropic",
+            provider_model: "claude-opus-4-8",
+            target_provider_protocol: "anthropic_messages",
+          },
+        ],
+        final: finalOk({
+          model_alias: "anthropic/claude-opus-4-8",
+          provider_model: "claude-opus-4-8",
+        }),
+      }),
+    );
+
+    expect(record.provider_attempts[0]?.provider_name).toBe("anthropic");
+    expect(record.provider_attempts[0]?.provider_model).toBe("claude-opus-4-8");
+    expect(record.provider_attempts[0]?.target_provider_protocol).toBe("anthropic_messages");
+  });
+
   it("5. terminal failure: final.status error, error_reason is an error_class", () => {
     const record = buildDecisionRecord(
       parts({
diff --git a/packages/core/src/telemetry/decision.ts b/packages/core/src/telemetry/decision.ts
index c1fb0f18..b79957b8 100644
--- a/packages/core/src/telemetry/decision.ts
+++ b/packages/core/src/telemetry/decision.ts
@@ -109,6 +109,40 @@ function executionFallbackCount(attempts: AttemptRecord[]): number {
   return Math.max(0, served - 1);
 }
 
+function providerAttemptRecord(a: AttemptRecord): DecisionRecord["provider_attempts"][number] {
+  return {
+    alias: a.alias,
+    skipped: a.skipped,
+    skip_reason: a.skip_reason,
+    status: a.status,
+    error_class: a.error_class,
+    latency_ms: a.latency_ms,
+    cost_usd: a.cost_usd,
+    // Per-attempt upstream failure detail (admin-debug-error-detail). The whole
+    // record is run through `redact` below, so any key echoed in provider_raw
+    // is irreversibly fingerprinted before persistence (principle 7).
+    error_detail: a.error_detail,
+    ...(a.passthrough_considered !== undefined
+      ? { passthrough_considered: a.passthrough_considered }
+      : {}),
+    ...(a.passthrough_used !== undefined ? { passthrough_used: a.passthrough_used } : {}),
+    ...(a.passthrough_disable_reason !== undefined
+      ? { passthrough_disable_reason: a.passthrough_disable_reason }
+      : {}),
+    ...(a.source_protocol !== undefined ? { source_protocol: a.source_protocol } : {}),
+    ...(a.target_provider_protocol !== undefined
+      ? { target_provider_protocol: a.target_provider_protocol }
+      : {}),
+    ...(a.response_protocol !== undefined ? { response_protocol: a.response_protocol } : {}),
+    ...(a.provider_name !== undefined ? { provider_name: a.provider_name } : {}),
+    ...(a.provider_model !== undefined ? { provider_model: a.provider_model } : {}),
+    ...(a.passthrough_mutations !== undefined
+      ? { passthrough_mutations: a.passthrough_mutations }
+      : {}),
+    ...(a.request_mutations !== undefined ? { request_mutations: a.request_mutations } : {}),
+  };
+}
+
 // Assemble the complete DecisionRecord, every field filled (explicit null where
 // the spec requires null — never omitted). The request is the source of the
 // trace id (the pipeline uses request_id as the trace id). The whole record is
@@ -153,25 +187,14 @@ export function buildDecisionRecord(parts: DecisionParts): DecisionRecord {
     },
     // Field-for-field identical to ProviderAttemptSchema — copy verbatim, in
     // chain order (skipped candidates included with their skip_reason).
-    provider_attempts: attempts.map((a) => ({
-      alias: a.alias,
-      skipped: a.skipped,
-      skip_reason: a.skip_reason,
-      status: a.status,
-      error_class: a.error_class,
-      latency_ms: a.latency_ms,
-      cost_usd: a.cost_usd,
-      // Per-attempt upstream failure detail (admin-debug-error-detail). The whole
-      // record is run through `redact` below, so any key echoed in provider_raw
-      // is irreversibly fingerprinted before persistence (principle 7).
-      error_detail: a.error_detail,
-    })),
+    provider_attempts: attempts.map(providerAttemptRecord),
     final: {
       model_alias: final.model_alias,
       provider_model: final.provider_model,
       status: final.status,
       error_reason: final.error_reason,
     },
+    serving_account: null,
     latency_total_ms: attempts.reduce((acc, a) => acc + a.latency_ms, 0),
     fallback_count: executionFallbackCount(attempts),
     cost_breakdown: {
diff --git a/packages/shared/src/decision/schema.test.ts b/packages/shared/src/decision/schema.test.ts
index 4a76aa1f..e0ce0af5 100644
--- a/packages/shared/src/decision/schema.test.ts
+++ b/packages/shared/src/decision/schema.test.ts
@@ -227,6 +227,19 @@ describe("DecisionRecordSchema", () => {
     expect(parsed.cost_breakdown.eval_usd).toBeNull();
   });
 
+  it("defaults serving_account to null and round-trips the final subscription account", () => {
+    expect(DecisionRecordSchema.parse(fullRecord()).serving_account).toBeNull();
+
+    const parsed = DecisionRecordSchema.parse({
+      ...fullRecord(),
+      serving_account: { provider_id: "anthropic", account: "claude-team-a" },
+    });
+    expect(parsed.serving_account).toEqual({
+      provider_id: "anthropic",
+      account: "claude-team-a",
+    });
+  });
+
   it("rejects a negative fallback_count", () => {
     const r = { ...fullRecord(), fallback_count: -1 };
     expect(DecisionRecordSchema.safeParse(r).success).toBe(false);
diff --git a/packages/shared/src/decision/schema.ts b/packages/shared/src/decision/schema.ts
index 02d8ef10..4f81a6fc 100644
--- a/packages/shared/src/decision/schema.ts
+++ b/packages/shared/src/decision/schema.ts
@@ -123,6 +123,11 @@ export const FinalDecisionSchema = z.object({
   error_reason: z.string().nullable(),
 });
 
+export const ServingAccountDecisionSchema = z.object({
+  provider_id: z.string().min(1),
+  account: z.string().min(1),
+});
+
 // Cost split (docs/07 "cost split (including eval's own self-cost)"). `eval_usd` isolates the
 // Layer-2 small-model self-cost (non-null ONLY when eval actually ran; null when
 // eval was skipped/disabled) from `completion_usd` (Σ of the served provider
@@ -210,6 +215,10 @@ export const DecisionRecordSchema = z.object({
   lane: LaneDecisionSchema,
   provider_attempts: z.array(ProviderAttemptSchema),
   final: FinalDecisionSchema,
+  // Concrete subscription account that ultimately served the request. Null for
+  // non-subscription providers, legacy rows, failed requests, or stale selections
+  // that later fell back to a different provider. Body-free routing metadata only.
+  serving_account: ServingAccountDecisionSchema.nullable().default(null),
   // Total served latency = Σ provider_attempts.latency_ms (docs/07 latency).
   // `.default(0)` so legacy records validate; the builder computes the real sum.
   latency_total_ms: z.number().nonnegative().default(0),
@@ -255,6 +264,7 @@ export type LaneDecision = z.infer<typeof LaneDecisionSchema>;
 export type AttemptErrorDetail = z.infer<typeof AttemptErrorDetailSchema>;
 export type ProviderAttempt = z.infer<typeof ProviderAttemptSchema>;
 export type FinalDecision = z.infer<typeof FinalDecisionSchema>;
+export type ServingAccountDecision = z.infer<typeof ServingAccountDecisionSchema>;
 export type CostBreakdown = z.infer<typeof CostBreakdownSchema>;
 export type TokenUsageBreakdown = z.infer<typeof TokenUsageSchema>;
 export type MemoryDecision = z.infer<typeof MemoryDecisionSchema>;
diff --git a/packages/shared/src/index.ts b/packages/shared/src/index.ts
index ade17691..45cbd782 100644
--- a/packages/shared/src/index.ts
+++ b/packages/shared/src/index.ts
@@ -167,6 +167,8 @@ export {
   PolicyDecisionSchema,
   type ProviderAttempt,
   ProviderAttemptSchema,
+  type ServingAccountDecision,
+  ServingAccountDecisionSchema,
   type TokenUsageBreakdown,
   TokenUsageSchema,
 } from "./decision/schema.js";