Fix codex token accounting and remove the two dead codex transports

agents/codelayer/README.md

@@ -39,7 +39,7 @@ CODELAYER_CODEX_MODEL=my-azure-deployment
 
 The selected CodeLayer model still controls prompts, reasoning, context, and cost data. Only the `model` value sent on the wire changes. Custom requests keep reasoning effort and summary, stateless `store: false`, encrypted reasoning content, and prompt caching. They omit fast mode and `service_tier`.
 
-Setting any optional override without both `CODELAYER_CODEX_BASE_URL` and `CODELAYER_CODEX_API_KEY` fails before CodeLayer reads Codex file auth or sends a request. Restart the Riptide daemon after changing any override value. When all override values are absent, CodeLayer keeps its current Codex file auth and `CODEX_PROVIDER=sse|websockets|aisdk_responses` behavior.
+Setting any optional override without both `CODELAYER_CODEX_BASE_URL` and `CODELAYER_CODEX_API_KEY` fails before CodeLayer reads Codex file auth or sends a request. Restart the Riptide daemon after changing any override value. When all override values are absent, CodeLayer keeps its current Codex file auth and `CODEX_PROVIDER=sse|websockets` behavior.
 
 Custom endpoint failures flow through the existing Codex diagnostics sink. The CLI writes them to its Codex diagnostics log; Riptide writes them to daemon logs and captures error events in Sentry. Diagnostic records omit API keys and response bodies.
 

agents/codelayer/src/providers.ts

@@ -6,12 +6,11 @@ import { createCopilotProvider } from '@humanlayer/agentlayer-provider-github-co
 import {
 	createCodexSseVendorProvider,
 	createCodexEffectProvider,
-	createCodexResponsesProvider,
 	type CodexDiagnosticsContext,
 	CODEX_DEFAULT_VERSION,
 } from '@humanlayer/agentlayer-provider-openai-codex'
 
-export type CodexProviderMode = 'sse' | 'aisdk_responses' | 'websockets'
+export type CodexProviderMode = 'sse' | 'websockets'
 
 export type ProviderType = 'anthropic' | 'openai' | 'codex' | 'copilot' | 'firepass'
 
@@ -79,12 +78,17 @@ export async function captureResponseUsage(response: Response, usage: RawCacheUs
 	if (!response.body || !response.ok) return response
 	const contentType = response.headers.get('content-type') ?? ''
 	if (contentType.includes('application/json')) {
+		// Read once and rebuild rather than clone(): the SDK (>= @ai-sdk/openai
+		// 3.0.96) reads the body in a way that races the clone's tee under Bun,
+		// surfacing as "JSON Parse error: Unexpected EOF" from a half-drained
+		// stream. A rebuilt Response hands it a fresh, fully-buffered body.
+		const text = await response.text()
 		try {
-			captureCacheUsage(await response.clone().json(), usage)
+			captureCacheUsage(JSON.parse(text), usage)
 		} catch {
-			// The SDK parses and reports malformed JSON from the original response.
+			// The SDK parses and reports malformed JSON itself.
 		}
-		return response
+		return new Response(text, { status: response.status, statusText: response.statusText, headers: response.headers })
 	}
 	if (!contentType.includes('text/event-stream')) return response
 
@@ -265,7 +269,7 @@ function reportCustomResponsesError(options: {
 		options.diagnostics.onEvent({
 			event: 'codex.provider.custom_responses.failed',
 			severity: 'error',
-			transport: 'aisdk_responses',
+			transport: 'custom_responses',
 			annotations: options.diagnostics.annotations,
 			metadata: {
 				error: safeMessage,
@@ -438,9 +442,16 @@ export async function resolveModel(
 			}
 
 			const authStore = await ensureFileAuthStore()
-			const codexMode = context?.codexProviderMode
-				?? (process.env.CODEX_PROVIDER as CodexProviderMode | undefined)
-				?? 'sse'
+			const requestedRaw = context?.codexProviderMode ?? (process.env.CODEX_PROVIDER as string | undefined)
+			// An empty env value (CODEX_PROVIDER= from a template) means unset, not unknown.
+			const requestedMode = requestedRaw === '' ? undefined : requestedRaw
+			// Unknown or retired CODEX_PROVIDER values fall back to the default
+			// transport instead of crashing a daemon that still carries the env var.
+			const codexMode: CodexProviderMode =
+				requestedMode === 'sse' || requestedMode === 'websockets' ? requestedMode : 'sse'
+			if (requestedMode !== undefined && requestedMode !== codexMode) {
+				console.error(`[codex-provider] unknown transport '${requestedMode}', falling back to 'sse'`)
+			}
 			const codexOpts = {
 				authStore,
 				version: CODEX_DEFAULT_VERSION,
@@ -450,8 +461,6 @@ export async function resolveModel(
 			}
 			console.error(`[codex-provider] using ${codexMode} transport for model ${modelId}`)
 			switch (codexMode) {
-				case 'aisdk_responses':
-					return createCodexResponsesProvider(codexOpts).languageModel(modelId) as LanguageModel
 				case 'websockets':
 					return createCodexEffectProvider(codexOpts).languageModel(modelId) as LanguageModel
 				case 'sse':

agents/codelayer/test/agent.test.ts

@@ -196,7 +196,6 @@ describe('provider resolution', () => {
 		process.env.CODELAYER_CODEX_MODEL = 'azure-coding-deployment'
 		process.env.CODEX_PROVIDER = 'websockets'
 		const sseSpy = spyOn(codexProvider, 'createCodexSseVendorProvider')
-		const responsesSpy = spyOn(codexProvider, 'createCodexResponsesProvider')
 		const websocketSpy = spyOn(codexProvider, 'createCodexEffectProvider')
 
 		const model = await resolveModel('codex', 'gpt-5.6-sol')
@@ -205,14 +204,12 @@ describe('provider resolution', () => {
 		expect((model as { modelId: string }).modelId).toBe('gpt-5.6-sol')
 		expect(providerAuth.ensureFileAuthStore).not.toHaveBeenCalled()
 		expect(sseSpy).not.toHaveBeenCalled()
-		expect(responsesSpy).not.toHaveBeenCalled()
 		expect(websocketSpy).not.toHaveBeenCalled()
 	})
 
 	test('rejects partial custom Codex settings before auth or private provider selection', async () => {
 		process.env.CODELAYER_CODEX_BASE_URL = 'https://example.test/openai/v1'
 		const sseSpy = spyOn(codexProvider, 'createCodexSseVendorProvider')
-		const responsesSpy = spyOn(codexProvider, 'createCodexResponsesProvider')
 		const websocketSpy = spyOn(codexProvider, 'createCodexEffectProvider')
 
 		await expect(resolveModel('codex', 'gpt-5.6-sol')).rejects.toThrow('CODELAYER_CODEX_API_KEY')
@@ -221,22 +218,18 @@ describe('provider resolution', () => {
 		await expect(resolveModel('codex', 'gpt-5.6-sol')).rejects.toThrow('CODELAYER_CODEX_BASE_URL')
 		expect(providerAuth.ensureFileAuthStore).not.toHaveBeenCalled()
 		expect(sseSpy).not.toHaveBeenCalled()
-		expect(responsesSpy).not.toHaveBeenCalled()
 		expect(websocketSpy).not.toHaveBeenCalled()
 	})
 
 	test('keeps every private Codex transport available when the override is absent', async () => {
 		const sseSpy = spyOn(codexProvider, 'createCodexSseVendorProvider')
-		const responsesSpy = spyOn(codexProvider, 'createCodexResponsesProvider')
 		const websocketSpy = spyOn(codexProvider, 'createCodexEffectProvider')
 
 		await resolveModel('codex', 'gpt-5.5', { codexProviderMode: 'sse' })
-		await resolveModel('codex', 'gpt-5.5', { codexProviderMode: 'aisdk_responses' })
 		await resolveModel('codex', 'gpt-5.5', { codexProviderMode: 'websockets' })
 
-		expect(providerAuth.ensureFileAuthStore).toHaveBeenCalledTimes(3)
+		expect(providerAuth.ensureFileAuthStore).toHaveBeenCalledTimes(2)
 		expect(sseSpy).toHaveBeenCalledTimes(1)
-		expect(responsesSpy).toHaveBeenCalledTimes(1)
 		expect(websocketSpy).toHaveBeenCalledTimes(1)
 	})
 
@@ -267,13 +260,17 @@ describe('provider resolution', () => {
 		)
 	})
 
-	test('respects explicit codex provider mode from caller context', async () => {
-		const providerSpy = spyOn(codexProvider, 'createCodexResponsesProvider')
+	test('falls back to the SSE transport when a retired mode is requested', async () => {
+		const sseSpy = spyOn(codexProvider, 'createCodexSseVendorProvider')
 
-		const model = await resolveModel('codex', 'gpt-5.5', { codexProviderMode: 'aisdk_responses' })
+		// 'aisdk_responses' was removed; a daemon still carrying the env var or a
+		// stale context value must degrade to the default transport, not crash.
+		const model = await resolveModel('codex', 'gpt-5.5', {
+			codexProviderMode: 'aisdk_responses' as never,
+		})
 
 		expect(model).toBeDefined()
-		expect(providerSpy).toHaveBeenCalled()
+		expect(sseSpy).toHaveBeenCalled()
 	})
 
 	test('respects CODEX_PROVIDER when caller context does not set a mode', async () => {

package.json

@@ -42,7 +42,7 @@
 	],
     "catalog": {
 		"@ai-sdk/anthropic": "3.0.82",
-		"@ai-sdk/openai": "3.0.69",
+		"@ai-sdk/openai": "3.0.96",
         "@ai-sdk/provider": "3.0.10",
         "@ai-sdk/provider-utils": "4.0.27",
         "@pulumi/aws": "7.26.0",

packages/agentlayer-core/src/agent.ts

@@ -42,7 +42,14 @@ import { sanitizeTextForModelState, sanitizeToolOutputForModelState } from './sa
 import type { AgentState, ApprovalDecision, ApprovalHistoryEntry, TerminalChildMap } from './state'
 import type { Step, StepToolResult, StopResult, StopTiming, StopWhen } from './stop-conditions'
 import { shouldStop } from './stop-conditions'
-import { extractUsage, getModelKey, type ModelTokenUsage, type TokenUsage, TokenUsageAccumulator } from './token-usage'
+import {
+	extractUsage,
+	getModelKey,
+	type ModelTokenUsage,
+	sumOrPoison,
+	type TokenUsage,
+	TokenUsageAccumulator,
+} from './token-usage'
 
 export type ProviderOptions = Parameters<typeof streamText>[0]['providerOptions']
 export type ProviderOptionsFactory = (ctx: { runId: string; promptCacheKey?: string }) => ProviderOptions
@@ -444,6 +451,7 @@ export class Agent<TTools extends Record<string, Tool<any, any>> = Record<string
 			cacheReadTokens: 0,
 			cacheWriteTokens: 0,
 			reasoningTokens: 0,
+			noCacheInputTokens: 0,
 		}
 		const inferenceMessages: ModelMessage[] = []
 		const summarize = async (requestText: string, turnPrefix: boolean): Promise<string> => {
@@ -478,6 +486,7 @@ export class Agent<TTools extends Record<string, Tool<any, any>> = Record<string
 			] as const) {
 				summaryUsage[key] += callUsage[key]
 			}
+			summaryUsage.noCacheInputTokens = sumOrPoison(summaryUsage.noCacheInputTokens, callUsage.noCacheInputTokens)
 			const responseMessages = response.messages.filter((message) => message.role !== 'tool')
 			const summary = responseMessages
 				.filter((message) => message.role === 'assistant')