@oh-my-pi/pi-ai 18.1.16 → 18.1.17

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## [Unreleased]
 
+## [18.1.17] - 2026-09-10
+
+### Fixed
+
+- Fixed transient Python HTTP/2 stream resets and HTTP/1.1 chunked response interruptions being treated as terminal errors when forwarded by a proxy ([#11160](https://github.com/can1357/oh-my-pi/pull/11160) by [@cyriusweng](https://github.com/cyriusweng)).
+- Ollama cache hits now populate cached-token usage: `prompt_eval_cached_count` from the `/api/chat` done chunk maps to `cacheRead`, with `input` reduced to the uncached portion, so status-line `cache_turn`/`cache_hit` segments and cache-prefix audits report real hit rates instead of false misses.
+- Fixed requests that run across a price change being costed at the newer rate; peak/off-peak estimates now use the rate in effect when the request started.
+- Fixed GitHub Copilot Business seats getting HTTP 403 on every model while the same token succeeds with a Chat client identity: chat and model-policy requests now identify as `copilot-chat`, denied requests retry once as the Copilot CLI (`copilot-developer-cli`), and `COPILOT_INTEGRATION_ID` pins the `Copilot-Integration-Id` header up front; model discovery keeps the CLI identity and the 403 message names the identity and the remedies ([#11372](https://github.com/can1357/oh-my-pi/issues/11372)).
+
 ## [18.1.16] - 2026-09-09
 
 ### Fixed

package/dist/types/error/flags.d.ts

@@ -28,6 +28,10 @@ export declare const Flag: {
 };
 export type Flag = (typeof Flag)[keyof typeof Flag];
 export declare const STREAM_READ_ERROR_PATTERN: RegExp;
+/** Python h2/httpx diagnostics forwarded through provider or proxy error events. */
+export declare const PYTHON_HTTP2_STREAM_RESET_PATTERN: RegExp;
+/** Python h11/httpx EOF while reading an HTTP/1.1 chunked response body. */
+export declare const PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN: RegExp;
 export declare const TRANSIENT_TRANSPORT_PATTERN: RegExp;
 /**
  * Local llama.cpp / Ollama deterministic tool-call argument JSON parse failure.

package/dist/types/providers/github-copilot-headers.d.ts

@@ -1,4 +1,4 @@
-import type { Message } from "../types.js";
+import type { FetchImpl, Message } from "../types.js";
 /**
  * Infer whether the current request to Copilot is user-initiated or agent-initiated.
  * Accepts `unknown[]` because providers may pass pre-converted message shapes.
@@ -11,6 +11,32 @@ export type CopilotDynamicHeaders = {
     premiumRequests: CopilotPremiumRequests;
 };
 export declare function resolveGitHubCopilotBaseUrl(baseUrl: string | undefined, apiKey: string | undefined): string | undefined;
+/**
+ * Opt-in `Copilot-Integration-Id` override for chat and model-policy requests.
+ * Reads `COPILOT_INTEGRATION_ID`; unset/invalid keeps the chat-surface default
+ * (`COPILOT_CHAT_INTEGRATION_ID`). Model discovery keeps the CLI identity: it
+ * unlocks enterprise/experimental models and listing models is not
+ * policy-gated the way chat completions are (#11372).
+ */
+export declare function resolveCopilotIntegrationIdOverride(env?: Record<string, string | undefined>): string | undefined;
+/**
+ * Effective identity before the chat-surface default: explicit value, then
+ * request headers, then `COPILOT_INTEGRATION_ID`. Pure given its inputs, so
+ * tests inject literals instead of mutating process state.
+ */
+export declare function resolveCopilotRequestIdentity(headers?: Record<string, string>, explicit?: unknown, env?: Record<string, string | undefined>): string | undefined;
+/**
+ * Reissue chat-surface Copilot 403s once as the Copilot CLI.
+ *
+ * Chat is the default surface (`COPILOT_CHAT_INTEGRATION_ID`) because Business
+ * organizations that gate premium models per client surface commonly allow
+ * chat while blocking CLI/agentic clients (issue #11372). The retry fires
+ * only for requests carrying the chat default and only when the caller
+ * resolved no explicit identity — an explicit choice is never second-guessed.
+ * The denied body is drained before reissuing, and the retry carries the CLI
+ * identity so the guard passes it through: at most two requests, never a loop.
+ */
+export declare function wrapFetchForCopilotFallback(base: FetchImpl | undefined, enabled: boolean, integrationId?: unknown): FetchImpl;
 export declare function inferCopilotInitiator(messages: unknown[]): CopilotInitiator;
 /** Check whether any message in the conversation contains image content. */
 export declare function hasCopilotVisionInput(messages: Message[]): boolean;
@@ -37,4 +63,6 @@ export declare function buildCopilotDynamicHeaders(params: {
     headers?: Record<string, string>;
     initiatorOverride?: CopilotInitiator;
     planTier?: string;
+    /** Raw explicit identity; validated here, chat default when absent/invalid. */
+    integrationId?: unknown;
 }): CopilotDynamicHeaders;

package/dist/types/providers/openai-completions.d.ts

@@ -42,5 +42,5 @@ export interface OpenAICompletionsOptions extends StreamOptions {
  * assistant output commits the attempt.
  */
 export declare const streamOpenAICompletions: StreamFunction<"openai-completions">;
-export declare function parseChunkUsage(rawUsage: object, model: Model<"openai-completions">, premiumRequests: number | undefined): AssistantMessage["usage"];
+export declare function parseChunkUsage(rawUsage: object, model: Model<"openai-completions">, premiumRequests: number | undefined, timestamp?: number): AssistantMessage["usage"];
 export declare function convertMessages(model: Model<"openai-completions">, context: Context, compat: ResolvedOpenAICompat): ChatCompletionMessageParam[];

package/dist/types/providers/vision-guard.d.ts

@@ -1,4 +1,4 @@
-import type { ImageContent, Model, TextContent } from "../types.js";
+import type { Api, ImageContent, Model, TextContent } from "../types.js";
 export declare const NON_VISION_IMAGE_PLACEHOLDER = "[image omitted: model does not support vision]";
 export declare function partitionVisionContent(content: ReadonlyArray<TextContent | ImageContent>, supportsImages: boolean): {
     textBlocks: TextContent[];
@@ -12,4 +12,18 @@ export declare function joinTextWithImagePlaceholder(text: string, omittedImages
  * misconfigured provider descriptors or user model entries (e.g. text-only
  * DashScope Qwen SKUs, DeepSeek models) whose endpoints reject `image_url`.
  */
-export declare function isOpenAICompletionsVisionSupported(model: Model<"openai-completions">): boolean;
+export declare function isOpenAICompletionsVisionSupported(model: Model<"openai-completions" | "openrouter">): boolean;
+/**
+ * Whether the transport that will carry `model` sends image content on the wire.
+ *
+ * The `pi-native` transport forwards the original context (images included) to
+ * the gateway, which resolves its own model server-side, so the Chat
+ * Completions guard below never runs client-side and the declared input
+ * applies. Otherwise the OpenAI Chat Completions path applies the text-only
+ * guard, as does the OpenRouter chat fallback (`PI_OPENROUTER_RESPONSES=0`,
+ * which dispatches `openrouter` models through `streamOpenAICompletions`);
+ * every other API ships the modalities the model declares. Callers that report
+ * or gate on the wire (for example the `omp models` table) read this
+ * predicate; declared capability reads `model.input`.
+ */
+export declare function sendsImageInputOnWire(model: Model<Api>): boolean;

package/dist/types/registry/oauth/github-copilot.d.ts

@@ -8,6 +8,7 @@ type GitHubCopilotLoginOptions = {
         allowEmpty?: boolean;
     }) => Promise<string>;
     onProgress?: (message: string) => void;
+    copilotIntegrationId?: unknown;
     signal?: AbortSignal;
     pollIntervalFloorMs?: number;
     pollIntervalScaleMs?: number;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-ai",
-	"version": "18.1.16",
+	"version": "18.1.17",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"keywords": [
 		"ai",
@@ -124,11 +124,11 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.1.16",
-		"@oh-my-pi/pi-catalog": "18.1.16",
-		"@oh-my-pi/pi-natives": "18.1.16",
-		"@oh-my-pi/pi-utils": "18.1.16",
-		"@oh-my-pi/pi-wire": "18.1.16"
+		"@oh-my-pi/omptype": "18.1.17",
+		"@oh-my-pi/pi-catalog": "18.1.17",
+		"@oh-my-pi/pi-natives": "18.1.17",
+		"@oh-my-pi/pi-utils": "18.1.17",
+		"@oh-my-pi/pi-wire": "18.1.17"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/error/flags.ts

@@ -156,6 +156,11 @@ const TIMEOUT_PATTERN = /\b(?:operation\s+)?timed?\s*out\b|\btimeout\b|\bstream
 const TRANSIENT_ENVELOPE_PATTERN = /anthropic stream envelope error:/i;
 const TRANSIENT_ENVELOPE_TRUNCATION_PATTERN = /before message_(?:start|stop)/i;
 export const STREAM_READ_ERROR_PATTERN = /stream[_ -]?read[_ -]?error/i;
+/** Python h2/httpx diagnostics forwarded through provider or proxy error events. */
+export const PYTHON_HTTP2_STREAM_RESET_PATTERN = /<StreamReset stream_id:\d+, error_code:(?:2|7), remote_reset:True>/;
+/** Python h11/httpx EOF while reading an HTTP/1.1 chunked response body. */
+export const PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN =
+	/peer closed connection without sending complete message body \(incomplete chunked read\)/;
 export const TRANSIENT_TRANSPORT_PATTERN =
 	/\b(?:no[_ -]?capacity|(?:high|peak)[ _-]?demand|(?:at|over|insufficient)[ _-]?capacity|capacity[ _-]?(?:exceeded|exhausted)|peak[ _-]?load)\b|overloaded|provider.?returned.?error|rate.?limit|too many requests|auth-gateway\s+5\d{2}(?=[:\s]|$)|\b(?:429|500|502|503|504)\b|service.?unavailable|server.?error|internal.?error|retry your request|network.?error|connection.?error|connection.?refused|unable.?to.?connect\.\s*is the computer able to access the url\?|other side closed|fetch failed|upstream.?connect|upstream.?request.?failed|reset before headers|socket hang up|timed? out|timeout|terminated|retry delay|stream stall|no error details in response|HTTP2(?:StreamReset|RefusedStream|EnhanceYourCalm)|nghttp2_(?:internal_error|refused_stream)|stream closed with error code nghttp2_(?:internal_error|refused_stream)|malformed.?function.?call/i;
 const AUTH_FAILURE_PATTERN =
@@ -396,6 +401,8 @@ function isTransientErrorText(text: string): boolean {
 	return (
 		isUnexpectedSocketCloseMessage(text) ||
 		isStreamReadErrorText(text) ||
+		PYTHON_HTTP2_STREAM_RESET_PATTERN.test(text) ||
+		PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN.test(text) ||
 		(TRANSIENT_ENVELOPE_PATTERN.test(text) && TRANSIENT_ENVELOPE_TRUNCATION_PATTERN.test(text)) ||
 		TRANSIENT_TRANSPORT_PATTERN.test(text)
 	);

package/src/providers/amazon-bedrock.ts

@@ -836,7 +836,7 @@ function handleMetadata(event: MetadataEvent, model: Model<"bedrock-converse-str
 		output.usage.cacheRead = event.usage.cacheReadInputTokens || 0;
 		output.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;
 		output.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;
-		calculateCost(model, output.usage);
+		calculateCost(model, output.usage, output.timestamp);
 	}
 }
 

package/src/providers/anthropic.ts

@@ -107,7 +107,9 @@ import {
 import {
 	buildCopilotDynamicHeaders,
 	hasCopilotVisionInput,
+	resolveCopilotRequestIdentity,
 	resolveGitHubCopilotBaseUrl,
+	wrapFetchForCopilotFallback,
 } from "./github-copilot-headers";
 import { getOpenAIPromptCacheKey } from "./openai-shared";
 import { applyInferenceHeaders } from "./inference-headers";
@@ -1805,6 +1807,7 @@ function calculateFallbackTurnCost(
 	requestModel: Model<"anthropic-messages">,
 	usage: Usage,
 	source: AnthropicWireUsage,
+	timestamp: number,
 ): boolean {
 	const iterations = source.iterations ?? [];
 	if (iterations.length === 0) return false;
@@ -1830,7 +1833,7 @@ function calculateFallbackTurnCost(
 		iterationUsage.cacheWrite = cacheWriteTokens;
 		iterationUsage.totalTokens =
 			iterationUsage.input + iterationUsage.output + iterationUsage.cacheRead + iterationUsage.cacheWrite;
-		calculateCost(resolveIterationModel(requestModel, iteration.model), iterationUsage);
+		calculateCost(resolveIterationModel(requestModel, iteration.model), iterationUsage, timestamp);
 		cost.input += iterationUsage.cost.input;
 		cost.output += iterationUsage.cost.output;
 		cost.cacheRead += iterationUsage.cost.cacheRead;
@@ -2025,6 +2028,7 @@ const streamAnthropicOnce = (
 							hasImages: hasCopilotVisionInput(context.messages),
 							premiumMultiplier: model.premiumMultiplier,
 							headers: { ...model.headers, ...options?.headers },
+							integrationId: resolveCopilotRequestIdentity(options?.headers),
 							initiatorOverride: options?.initiatorOverride,
 						})
 					: undefined;
@@ -2274,7 +2278,7 @@ const streamAnthropicOnce = (
 				applyAnthropicUsageExtras(output.usage, wireUsage);
 				output.usage.totalTokens =
 					output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
-				calculateCost(model, output.usage);
+				calculateCost(model, output.usage, output.timestamp);
 				output.duration = performance.now() - startTime;
 				stream.push({ type: "start", partial: output });
 				stream.push({ type: "done", reason: "stop", message: output });
@@ -2528,11 +2532,11 @@ const streamAnthropicOnce = (
 								if (serverSideFallback) {
 									const served = fallbackServedModelFromUsage(startUsage);
 									if (served) output.model = served;
-									if (!calculateFallbackTurnCost(model, output.usage, startUsage)) {
-										calculateCost(model, output.usage);
+									if (!calculateFallbackTurnCost(model, output.usage, startUsage, output.timestamp)) {
+										calculateCost(model, output.usage, output.timestamp);
 									}
 								} else {
-									calculateCost(model, output.usage);
+									calculateCost(model, output.usage, output.timestamp);
 								}
 							} else {
 								reportAnthropicEnvelopeAnomaly("message_start missing usage");
@@ -2860,11 +2864,11 @@ const streamAnthropicOnce = (
 								if (serverSideFallback) {
 									const served = fallbackServedModelFromUsage(deltaUsage);
 									if (served) output.model = served;
-									if (!calculateFallbackTurnCost(model, output.usage, deltaUsage)) {
-										calculateCost(model, output.usage);
+									if (!calculateFallbackTurnCost(model, output.usage, deltaUsage, output.timestamp)) {
+										calculateCost(model, output.usage, output.timestamp);
 									}
 								} else {
-									calculateCost(model, output.usage);
+									calculateCost(model, output.usage, output.timestamp);
 								}
 							}
 						} else if (event.type === "message_stop") {
@@ -3281,7 +3285,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 			maxRetries: 5,
 			maxRetryDelayMs,
 			defaultHeaders,
-			fetch: cchFetch,
+			fetch: wrapFetchForCopilotFallback(cchFetch, true, resolveCopilotRequestIdentity(headers)),
 			fetchOptions,
 		};
 	}

package/src/providers/cursor.ts

@@ -957,7 +957,7 @@ function streamCursorWithWireMode(
 			endCurrentThinkingBlock(output, stream, state);
 			flushOpenToolCalls(output, stream, state);
 
-			calculateCost(model, output.usage);
+			calculateCost(model, output.usage, output.timestamp);
 
 			output.duration = performance.now() - startTime;
 			if (firstTokenTime) output.ttft = firstTokenTime - startTime;

package/src/providers/devin.ts

@@ -450,7 +450,7 @@ export const streamDevin: StreamFunction<"devin-agent"> = (
 				toolBlocks.size > 0 ? "toolUse" : latestStopReason === StopReason.MAX_TOKENS ? "length" : "stop";
 			output.stopReason = doneReason;
 
-			calculateCost(model, output.usage);
+			calculateCost(model, output.usage, output.timestamp);
 			output.duration = performance.now() - startTime;
 			if (firstTokenTime) output.ttft = firstTokenTime - startTime;
 

package/src/providers/github-copilot-headers.ts

@@ -1,9 +1,12 @@
 import {
 	COPILOT_CAPI_IDENTITY_HEADERS,
+	COPILOT_CHAT_INTEGRATION_ID,
 	getGitHubCopilotBaseUrl,
+	normalizeCopilotIntegrationId,
 	parseGitHubCopilotApiKey,
 } from "@oh-my-pi/pi-catalog/wire/github-copilot";
-import type { Message } from "../types";
+import { $env, logger } from "@oh-my-pi/pi-utils";
+import type { FetchImpl, Message } from "../types";
 /**
  * Infer whether the current request to Copilot is user-initiated or agent-initiated.
  * Accepts `unknown[]` because providers may pass pre-converted message shapes.
@@ -26,6 +29,86 @@ export function resolveGitHubCopilotBaseUrl(
 	if (baseUrl && !baseUrl.includes("githubcopilot.com")) return baseUrl;
 	return getGitHubCopilotBaseUrl(enterpriseUrl);
 }
+
+/**
+ * Opt-in `Copilot-Integration-Id` override for chat and model-policy requests.
+ * Reads `COPILOT_INTEGRATION_ID`; unset/invalid keeps the chat-surface default
+ * (`COPILOT_CHAT_INTEGRATION_ID`). Model discovery keeps the CLI identity: it
+ * unlocks enterprise/experimental models and listing models is not
+ * policy-gated the way chat completions are (#11372).
+ */
+export function resolveCopilotIntegrationIdOverride(
+	env: Record<string, string | undefined> = $env,
+): string | undefined {
+	return normalizeCopilotIntegrationId(env.COPILOT_INTEGRATION_ID);
+}
+
+/**
+ * Explicit caller-supplied `Copilot-Integration-Id`, matched case-insensitively.
+ * Takes only caller layers (`extraHeaders` / `options.headers`) — never model
+ * catalog headers — so a catalog default can never masquerade as a choice.
+ */
+function explicitCopilotIntegrationId(headers: Record<string, string> | undefined): unknown {
+	if (!headers) return undefined;
+	for (const name of Object.keys(headers)) {
+		if (name.toLowerCase() === "copilot-integration-id") return headers[name];
+	}
+	return undefined;
+}
+
+/**
+ * Effective identity before the chat-surface default: explicit value, then
+ * request headers, then `COPILOT_INTEGRATION_ID`. Pure given its inputs, so
+ * tests inject literals instead of mutating process state.
+ */
+export function resolveCopilotRequestIdentity(
+	headers?: Record<string, string>,
+	explicit?: unknown,
+	env: Record<string, string | undefined> = $env,
+): string | undefined {
+	return (
+		normalizeCopilotIntegrationId(explicit) ??
+		normalizeCopilotIntegrationId(explicitCopilotIntegrationId(headers)) ??
+		resolveCopilotIntegrationIdOverride(env)
+	);
+}
+
+/**
+ * Reissue chat-surface Copilot 403s once as the Copilot CLI.
+ *
+ * Chat is the default surface (`COPILOT_CHAT_INTEGRATION_ID`) because Business
+ * organizations that gate premium models per client surface commonly allow
+ * chat while blocking CLI/agentic clients (issue #11372). The retry fires
+ * only for requests carrying the chat default and only when the caller
+ * resolved no explicit identity — an explicit choice is never second-guessed.
+ * The denied body is drained before reissuing, and the retry carries the CLI
+ * identity so the guard passes it through: at most two requests, never a loop.
+ */
+export function wrapFetchForCopilotFallback(
+	base: FetchImpl | undefined,
+	enabled: boolean,
+	integrationId?: unknown,
+): FetchImpl {
+	const inner = base ?? fetch;
+	if (!enabled) return inner;
+	return async (input, init) => {
+		const response = await inner(input, init);
+		if (response.status !== 403) return response;
+		if (input instanceof Request) return response;
+		if (normalizeCopilotIntegrationId(integrationId) !== undefined) return response;
+		const outgoing = new Headers(init?.headers);
+		if (outgoing.get("Copilot-Integration-Id") !== COPILOT_CHAT_INTEGRATION_ID) {
+			return response;
+		}
+		try {
+			await response.arrayBuffer();
+		} catch {}
+		logger.warn("GitHub Copilot chat identity denied (HTTP 403); retrying once as the Copilot CLI");
+		const retryHeaders = new Headers(outgoing);
+		retryHeaders.set("Copilot-Integration-Id", COPILOT_CAPI_IDENTITY_HEADERS["Copilot-Integration-Id"]);
+		return inner(input, { ...init, headers: retryHeaders });
+	};
+}
 export function inferCopilotInitiator(messages: unknown[]): CopilotInitiator {
 	if (messages.length === 0) return "user";
 
@@ -121,6 +204,8 @@ export function buildCopilotDynamicHeaders(params: {
 	headers?: Record<string, string>;
 	initiatorOverride?: CopilotInitiator;
 	planTier?: string;
+	/** Raw explicit identity; validated here, chat default when absent/invalid. */
+	integrationId?: unknown;
 }): CopilotDynamicHeaders {
 	const initiator =
 		params.initiatorOverride ?? getCopilotInitiatorOverride(params.headers) ?? inferCopilotInitiator(params.messages);
@@ -129,6 +214,8 @@ export function buildCopilotDynamicHeaders(params: {
 		"X-Initiator": initiator,
 		"X-Interaction-Type": `conversation-${initiator}`,
 	};
+	headers["Copilot-Integration-Id"] =
+		normalizeCopilotIntegrationId(params.integrationId) ?? COPILOT_CHAT_INTEGRATION_ID;
 
 	if (params.hasImages) {
 		headers["Copilot-Vision-Request"] = "true";

package/src/providers/google-gemini-cli.ts

@@ -894,7 +894,7 @@ export const streamGoogleGeminiCli: StreamFunction<"google-gemini-cli"> = (
 								total: 0,
 							},
 						};
-						calculateCost(model, output.usage);
+						calculateCost(model, output.usage, output.timestamp);
 					}
 				}
 

package/src/providers/google-shared.ts

@@ -755,7 +755,7 @@ export async function consumeGoogleStream<T extends GoogleApiType>(args: {
 					total: 0,
 				},
 			};
-			calculateCost(model, output.usage);
+			calculateCost(model, output.usage, output.timestamp);
 		}
 	}
 

package/src/providers/ollama.ts

@@ -82,6 +82,7 @@ type OllamaChatChunk = {
 	done?: boolean;
 	done_reason?: string;
 	prompt_eval_count?: number;
+	prompt_eval_cached_count?: number;
 	eval_count?: number;
 };
 
@@ -698,9 +699,17 @@ const streamOllamaOnce = (
 					if (healedToolCallEmitted && output.stopReason === "stop") {
 						output.stopReason = "toolUse";
 					}
-					output.usage.input = chunk.prompt_eval_count ?? 0;
+					// Ollama reports prompt cache splits as prompt_eval_cached_count
+					// (cached) vs prompt_eval_count (total; cached + uncached =
+					// total). Map cached → cacheRead, uncached → input so
+					// cache_turn/cache_hit status segments and the cache-prefix
+					// audit see real hit rates. Local Ollama omits the cached
+					// field; the ?? 0 fallbacks keep local behavior
+					// byte-identical to before.
+					output.usage.cacheRead = chunk.prompt_eval_cached_count ?? 0;
+					output.usage.input = (chunk.prompt_eval_count ?? 0) - output.usage.cacheRead;
 					output.usage.output = chunk.eval_count ?? 0;
-					output.usage.totalTokens = output.usage.input + output.usage.output;
+					output.usage.totalTokens = output.usage.input + output.usage.output + output.usage.cacheRead;
 				}
 			}
 			if (streamMarkupHealing) {

package/src/providers/openai-codex-responses.ts

@@ -2572,7 +2572,7 @@ class CodexStreamProcessor {
 			hasExecutableIncompleteResponsesToolCalls(output);
 		finalizePendingResponsesToolCalls(output);
 
-		calculateCost(model, output.usage);
+		calculateCost(model, output.usage, output.timestamp);
 		applyCodexServiceTierPricing(model, output.usage, serviceTier, runtime.requestBodyForState.service_tier);
 		output.stopReason = mapOpenAIResponsesStopReason(status);
 		promoteResponsesToolUseStopReason(
@@ -4837,7 +4837,12 @@ export function isRetryableCodexFailureEvent(rawEvent: Record<string, unknown>):
 		return true;
 	}
 	const message = error?.message ?? event.message ?? event.response?.message;
-	return !!message && CODEX_RETRYABLE_EVENT_MESSAGE.test(message);
+	return (
+		!!message &&
+		(CODEX_RETRYABLE_EVENT_MESSAGE.test(message) ||
+			AIError.PYTHON_HTTP2_STREAM_RESET_PATTERN.test(message) ||
+			AIError.PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN.test(message))
+	);
 }
 
 export function createCodexProviderStreamError(rawEvent: Record<string, unknown>): CodexProviderStreamError {

package/src/providers/openai-completions.ts

@@ -78,6 +78,7 @@ import {
 	rememberOpenAIReasoningEffortFallback,
 	resolveOpenAIReasoningEffortFallback,
 } from "./openai-reasoning-fallback";
+import { resolveCopilotRequestIdentity, wrapFetchForCopilotFallback } from "./github-copilot-headers";
 import {
 	applyChatCompletionsReasoningParams,
 	applyChatCompletionsToolStream,
@@ -796,7 +797,11 @@ const streamOpenAICompletionsOnce = (
 						headers: headersWithTimeout,
 						body: params,
 						signal: requestSignal,
-						fetch: options?.fetch,
+						fetch: wrapFetchForCopilotFallback(
+							options?.fetch,
+							model.provider === "github-copilot",
+							resolveCopilotRequestIdentity(options?.headers),
+						),
 						// Transient 408/429/5xx get Retry-After-aware transport retries.
 						// The first-event watchdog above aborts `requestSignal`, which
 						// bounds every attempt and backoff sleep — retries cannot
@@ -1122,7 +1127,7 @@ const streamOpenAICompletionsOnce = (
 			let sawUsagePayload = false;
 			let awaitTrailingUsageDetails = false;
 			const applyUsagePayload = (rawUsage: object): void => {
-				output.usage = parseChunkUsage(rawUsage, model, premiumRequestsTotal);
+				output.usage = parseChunkUsage(rawUsage, model, premiumRequestsTotal, output.timestamp);
 				sawUsagePayload = true;
 				awaitTrailingUsageDetails = !hasPositiveCacheReadTokenField(rawUsage);
 			};
@@ -1854,6 +1859,7 @@ export function parseChunkUsage(
 	rawUsage: object,
 	model: Model<"openai-completions">,
 	premiumRequests: number | undefined,
+	timestamp?: number,
 ): AssistantMessage["usage"] {
 	const usageLike = rawUsage as OpenAICompletionsUsageLike;
 	const rawPromptTokenDetails = usageLike.prompt_tokens_details;
@@ -1895,7 +1901,7 @@ export function parseChunkUsage(
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
 		...(premiumRequests !== undefined ? { premiumRequests } : {}),
 	};
-	calculateCost(model, usage);
+	calculateCost(model, usage, timestamp);
 	applyProviderReportedCost(model, usage, rawUsage);
 	return usage;
 }

package/src/providers/openai-responses.ts

@@ -64,6 +64,7 @@ import {
 	rememberOpenAIReasoningEffortFallback,
 	resolveOpenAIReasoningEffortFallback,
 } from "./openai-reasoning-fallback";
+import { resolveCopilotRequestIdentity, wrapFetchForCopilotFallback } from "./github-copilot-headers";
 import type {
 	Tool as OpenAITool,
 	ReasoningEffort,
@@ -567,7 +568,11 @@ const streamOpenAIResponsesOnce = (
 						headers: headersWithTimeout,
 						body: requestParams,
 						signal: requestSignal,
-						fetch: options?.fetch,
+						fetch: wrapFetchForCopilotFallback(
+							options?.fetch,
+							model.provider === "github-copilot",
+							resolveCopilotRequestIdentity(options?.headers),
+						),
 						// Transient 408/429/5xx get Retry-After-aware transport
 						// retries; the first-event watchdog aborts `requestSignal`,
 						// so retries cannot extend the caller's deadline.

package/src/providers/openai-shared.ts

@@ -59,6 +59,7 @@ import {
 	type ToolResultMessage,
 	type Usage,
 } from "../types";
+import { resolveCopilotRequestIdentity } from "./github-copilot-headers";
 
 export type { OpenAIPromptCacheOptions } from "../types";
 
@@ -259,6 +260,7 @@ export function resolveOpenAIRequestSetup(
 			premiumMultiplier: model.premiumMultiplier,
 			headers,
 			initiatorOverride: options.initiatorOverride,
+			integrationId: resolveCopilotRequestIdentity(options.extraHeaders),
 		});
 		Object.assign(headers, copilot.headers);
 		copilotPremiumRequests = copilot.premiumRequests;
@@ -3350,7 +3352,7 @@ export async function processResponsesStream<TApi extends Api>(
 				output.responseId = response.id;
 			}
 			populateResponsesUsageFromResponse(output, response?.usage);
-			calculateCost(model, output.usage);
+			calculateCost(model, output.usage, output.timestamp);
 			applyProviderReportedCost(model, output.usage, response?.usage);
 			applyOpenAIResponsesServiceTierCost(
 				model,

package/src/providers/vision-guard.ts

@@ -1,4 +1,5 @@
-import type { ImageContent, Model, TextContent } from "../types";
+import { $env } from "@oh-my-pi/pi-utils";
+import type { Api, ImageContent, Model, TextContent } from "../types";
 
 export const NON_VISION_IMAGE_PLACEHOLDER = "[image omitted: model does not support vision]";
 export function partitionVisionContent(
@@ -35,8 +36,33 @@ export function joinTextWithImagePlaceholder(text: string, omittedImages: boolea
  * misconfigured provider descriptors or user model entries (e.g. text-only
  * DashScope Qwen SKUs, DeepSeek models) whose endpoints reject `image_url`.
  */
-export function isOpenAICompletionsVisionSupported(model: Model<"openai-completions">): boolean {
+export function isOpenAICompletionsVisionSupported(model: Model<"openai-completions" | "openrouter">): boolean {
 	if (!model.input.includes("image")) return false;
 	if (model.compat.stripImageInput) return false;
 	return true;
 }
+
+/**
+ * Whether the transport that will carry `model` sends image content on the wire.
+ *
+ * The `pi-native` transport forwards the original context (images included) to
+ * the gateway, which resolves its own model server-side, so the Chat
+ * Completions guard below never runs client-side and the declared input
+ * applies. Otherwise the OpenAI Chat Completions path applies the text-only
+ * guard, as does the OpenRouter chat fallback (`PI_OPENROUTER_RESPONSES=0`,
+ * which dispatches `openrouter` models through `streamOpenAICompletions`);
+ * every other API ships the modalities the model declares. Callers that report
+ * or gate on the wire (for example the `omp models` table) read this
+ * predicate; declared capability reads `model.input`.
+ */
+export function sendsImageInputOnWire(model: Model<Api>): boolean {
+	if (model.transport === "pi-native") return model.input.includes("image");
+	if (isGuardedCompletionsTransport(model)) return isOpenAICompletionsVisionSupported(model);
+	return model.input.includes("image");
+}
+
+/** True for the transports that encode through the Chat Completions guard. */
+function isGuardedCompletionsTransport(model: Model<Api>): model is Model<"openai-completions" | "openrouter"> {
+	if (model.api === "openai-completions") return true;
+	return model.api === "openrouter" && $env.PI_OPENROUTER_RESPONSES === "0";
+}

package/src/registry/oauth/github-copilot.ts

@@ -15,12 +15,19 @@ import { scheduler } from "node:timers/promises";
 import { getBundledModels } from "@oh-my-pi/pi-catalog/models";
 import {
 	COPILOT_API_HEADERS,
+	COPILOT_CHAT_INTEGRATION_ID,
 	discoverGitHubCopilotApiEndpoint,
 	getGitHubCopilotBaseUrl,
 	isPublicGitHubHost,
+	normalizeCopilotIntegrationId,
 	normalizeDomain,
 	normalizeGitHubCopilotEnterpriseDomain,
 } from "@oh-my-pi/pi-catalog/wire/github-copilot";
+import { $env } from "@oh-my-pi/pi-utils";
+import {
+	resolveCopilotIntegrationIdOverride,
+	wrapFetchForCopilotFallback,
+} from "../../providers/github-copilot-headers";
 import * as AIError from "../../error";
 import type { FetchImpl } from "../../types";
 import type { OAuthController, OAuthCredentials } from "./types";
@@ -51,6 +58,7 @@ type GitHubCopilotLoginOptions = {
 	onAuth: (url: string, instructions?: string) => void;
 	onPrompt: (prompt: { message: string; placeholder?: string; allowEmpty?: boolean }) => Promise<string>;
 	onProgress?: (message: string) => void;
+	copilotIntegrationId?: unknown;
 	signal?: AbortSignal;
 	pollIntervalFloorMs?: number;
 	pollIntervalScaleMs?: number;
@@ -259,6 +267,7 @@ async function enableGitHubCopilotModel(
 	fetchImpl: FetchImpl,
 	enterpriseDomain: string | undefined,
 	apiEndpoint: string | undefined,
+	integrationId?: string,
 ): Promise<boolean> {
 	const baseUrl = apiEndpoint ?? getGitHubCopilotBaseUrl(enterpriseDomain);
 	const url = `${baseUrl}/models/${modelId}/policy`;
@@ -270,6 +279,7 @@ async function enableGitHubCopilotModel(
 				"Content-Type": "application/json",
 				Authorization: `Bearer ${token}`,
 				...COPILOT_API_HEADERS,
+				"Copilot-Integration-Id": integrationId ?? COPILOT_CHAT_INTEGRATION_ID,
 				"Openai-Intent": "chat-policy",
 				"X-Initiator": "user",
 				"X-Interaction-Type": "chat-policy",
@@ -292,16 +302,24 @@ async function enableAllGitHubCopilotModels(
 	apiEndpoint: string | undefined,
 	fetchImpl: FetchImpl,
 	onProgress?: (model: string, success: boolean) => void,
+	integrationId?: unknown,
 ): Promise<void> {
-	// Synthesized catalog variants (Copilot long-context `-1m` entries) share
-	// the upstream model id; enable each wire id exactly once.
 	const wireModelIds = [...new Set(getBundledModels("github-copilot").map(model => model.requestModelId ?? model.id))];
+	const resolvedId = normalizeCopilotIntegrationId(integrationId) ?? resolveCopilotIntegrationIdOverride();
+	const copilotFetch = wrapFetchForCopilotFallback(fetchImpl, true, resolvedId);
 	const BATCH_SIZE = 5;
 	for (let i = 0; i < wireModelIds.length; i += BATCH_SIZE) {
 		const batch = wireModelIds.slice(i, i + BATCH_SIZE);
 		await Promise.all(
 			batch.map(async modelId => {
-				const success = await enableGitHubCopilotModel(token, modelId, fetchImpl, enterpriseDomain, apiEndpoint);
+				const success = await enableGitHubCopilotModel(
+					token,
+					modelId,
+					copilotFetch,
+					enterpriseDomain,
+					apiEndpoint,
+					resolvedId,
+				);
 				onProgress?.(modelId, success);
 			}),
 		);
@@ -368,7 +386,14 @@ export async function loginGitHubCopilot(options: GitHubCopilotLoginOptions): Pr
 
 	// Enable all models after successful login
 	options.onProgress?.("Enabling models...");
-	await enableAllGitHubCopilotModels(githubAccessToken, enterpriseDomain ?? undefined, apiEndpoint, fetchImpl);
+	await enableAllGitHubCopilotModels(
+		githubAccessToken,
+		enterpriseDomain ?? undefined,
+		apiEndpoint,
+		fetchImpl,
+		undefined,
+		options.copilotIntegrationId,
+	);
 	return credentials;
 }
 

package/src/utils/http-inspector.ts

@@ -108,7 +108,7 @@ export function rewriteCopilotError(errorMessage: string, error: unknown, provid
 		return `GitHub Copilot authentication failed (HTTP 401). Your token may have been revoked. Please re-login with /login github-copilot`;
 	}
 	if (status === 403) {
-		return `GitHub Copilot access denied (HTTP 403). Your account may not have access to this model or feature. Check your Copilot plan or model policy settings.`;
+		return `GitHub Copilot access denied (HTTP 403). Your token is valid but the account may not have access to this model or feature. Check your Copilot plan or model policy settings. Business organizations can also restrict which clients may call the API: omp sends Copilot-Integration-Id copilot-chat by default (COPILOT_INTEGRATION_ID overrides it) and retries a denied default-identity request once as the Copilot CLI (copilot-developer-cli). If both identities are denied, ask your org admin to allow one of them; if you pinned an identity, try the other.`;
 	}
 	return errorMessage;
 }