From 41e322c53d7b4b796eb377d0df9c29ecd10ba431 Mon Sep 17 00:00:00 2001
From: Chia <Chia@93.nz>
Date: Thu, 6 Aug 2026 09:29:41 +1200
Subject: feat: complete commercial control plane, billing, auth, and model
 catalog

- add PostgreSQL control-plane persistence with Redis-degraded hot reload
- implement prepaid balance, usage ledger, Stripe top-up and reconciliation
- add registration, email verification, password reset, invitations and RBAC
- support TOTP, Passkey MFA, device sessions, quotas and rate limits
- add tenant billing profiles, audit logs and operational readiness checks
- build authenticated admin console, Quickstart, Playground and usage analytics
- add public model catalog with pricing, filtering and cost estimation
- support OpenAI Responses providers and provider health failover
- validate real upstream usage reporting and balance settlement
---
 internal/usage/observer.go | 52 +++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 44 insertions(+), 8 deletions(-)

(limited to 'internal/usage/observer.go')

diff --git a/internal/usage/observer.go b/internal/usage/observer.go
index 808781f..71c1379 100644
--- a/internal/usage/observer.go
+++ b/internal/usage/observer.go
@@ -95,18 +95,28 @@ func (o *Observer) parseSSELine(line []byte) {
 	o.parseJSON(payload)
 }
 
+type tokenDetails struct {
+	CachedTokens     *int64 `json:"cached_tokens"`
+	CacheWriteTokens *int64 `json:"cache_write_tokens"`
+}
+
 type usageFields struct {
-	PromptTokens             *int64 `json:"prompt_tokens"`
-	CompletionTokens         *int64 `json:"completion_tokens"`
-	TotalTokens              *int64 `json:"total_tokens"`
-	InputTokens              *int64 `json:"input_tokens"`
-	OutputTokens             *int64 `json:"output_tokens"`
-	CacheCreationInputTokens *int64 `json:"cache_creation_input_tokens"`
-	CacheReadInputTokens     *int64 `json:"cache_read_input_tokens"`
+	PromptTokens             *int64        `json:"prompt_tokens"`
+	CompletionTokens         *int64        `json:"completion_tokens"`
+	TotalTokens              *int64        `json:"total_tokens"`
+	InputTokens              *int64        `json:"input_tokens"`
+	OutputTokens             *int64        `json:"output_tokens"`
+	CacheCreationInputTokens *int64        `json:"cache_creation_input_tokens"`
+	CacheReadInputTokens     *int64        `json:"cache_read_input_tokens"`
+	PromptTokensDetails      *tokenDetails `json:"prompt_tokens_details"`
+	InputTokensDetails       *tokenDetails `json:"input_tokens_details"`
 }
 
 type responseEnvelope struct {
-	Usage   *usageFields `json:"usage"`
+	Usage    *usageFields `json:"usage"`
+	Response *struct {
+		Usage *usageFields `json:"usage"`
+	} `json:"response"`
 	Message *struct {
 		Usage *usageFields `json:"usage"`
 	} `json:"message"`
@@ -131,16 +141,22 @@ func (o *Observer) parseJSON(payload []byte) {
 	if envelope.Message != nil && envelope.Message.Usage != nil {
 		o.apply(envelope.Message.Usage)
 	}
+	if envelope.Response != nil && envelope.Response.Usage != nil {
+		o.apply(envelope.Response.Usage)
+	}
 }
 
 func (o *Observer) apply(fields *usageFields) {
+	inputReported := false
 	if fields.PromptTokens != nil {
 		o.usage.InputTokens = *fields.PromptTokens
 		o.found = true
+		inputReported = true
 	}
 	if fields.InputTokens != nil {
 		o.usage.InputTokens = *fields.InputTokens
 		o.found = true
+		inputReported = true
 	}
 	if fields.CompletionTokens != nil {
 		o.usage.OutputTokens = *fields.CompletionTokens
@@ -163,6 +179,26 @@ func (o *Observer) apply(fields *usageFields) {
 		o.usage.CacheReadInputTokens = *fields.CacheReadInputTokens
 		o.found = true
 	}
+	details := fields.InputTokensDetails
+	if details == nil {
+		details = fields.PromptTokensDetails
+	}
+	if details != nil {
+		if details.CachedTokens != nil {
+			o.usage.CacheReadInputTokens = *details.CachedTokens
+			o.found = true
+		}
+		if details.CacheWriteTokens != nil {
+			o.usage.CacheCreationInputTokens = *details.CacheWriteTokens
+			o.found = true
+		}
+		// OpenAI reports cached token details as subsets of prompt/input_tokens.
+		// Normalize them into mutually exclusive buckets before billing. Anthropic
+		// reports its top-level cache fields separately, so they are not adjusted.
+		if inputReported && (o.protocol == domain.ProtocolOpenAI || o.protocol == domain.ProtocolOpenAIResponses) {
+			o.usage.InputTokens = max(0, o.usage.InputTokens-o.usage.CacheReadInputTokens-o.usage.CacheCreationInputTokens)
+		}
+	}
 	if !o.explicitTotal && o.found {
 		o.usage.TotalTokens = o.usage.InputTokens + o.usage.OutputTokens
 	}
-- 
cgit v1.2.3