feat: implemented anthropic keep-alive and migration support for updates

- Added Anthropic prompt-cache refresh scheduling and state management to keep prompts warm across idle sessions.
- Updated pricing models and database stats tracking to calculate and store cost-weighted cache savings.
- Integrated cache savings metrics and efficiency displays into the stats CLI, dashboard routes, and UI components.
- Added support for package renaming, manifest pointer tracking, and installation migration during CLI updates.

nix/dev-shell.nix

@@ -20,7 +20,6 @@ pkgs.mkShell (
         bun
         bun2nix
         rustToolchain
-        bazelisk
         cargo-nextest
         rustPlatform.bindgenHook
         nixfmt
@@ -61,6 +60,9 @@ pkgs.mkShell (
       ++ lib.optionals pkgs.stdenv.hostPlatform.isLinux linuxLibraries;
 
     CMAKE_POLICY_VERSION_MINIMUM = "3.5";
+    # Bazel's downloaded host tools assume an FHS loader; Cargo is the
+    # repository's supported local-iteration path inside the Nix shell.
+    OMP_NATIVE_BUILD_BACKEND = "cargo";
     PCRE2_SYS_STATIC = "1";
     RUST_SRC_PATH = "${rustToolchain}/lib/rustlib/src/rust/library";
   }

nix/package.nix

@@ -67,17 +67,6 @@ let
       runHook postInstall
     '';
   };
-  runtimeLibraries = [
-    libopus
-    openssl
-    pcre2
-  ]
-  ++ lib.optionals stdenv.hostPlatform.isLinux [
-    libpulseaudio
-    pipewire
-    stdenv.cc.cc.lib
-    zlib
-  ];
 in
 stdenv.mkDerivation (
   {
@@ -170,11 +159,18 @@ stdenv.mkDerivation (
 
       install -Dm755 packages/coding-agent/dist/omp "$out/libexec/omp/omp"
 
+      ${lib.optionalString stdenv.hostPlatform.isLinux ''
+        # The addon is gzip-compressed inside the compiled binary, so the Nix
+        # store RUNPATH autoPatchelf wrote into it is invisible to the output
+        # reference scanner. Record it in plain text to pin those libraries
+        # (pipewire, libopus, libgcc) into the runtime closure.
+        mkdir -p "$out/nix-support"
+        patchelf --print-rpath "packages/natives/native/${platform.addon}" \
+          > "$out/nix-support/embedded-addon-runpath"
+      ''}
+
       makeWrapper "$out/libexec/omp/omp" "$out/bin/omp" \
-        --set PI_SKIP_VERSION_CHECK 1 \
-        ${
-          if stdenv.hostPlatform.isLinux then "--prefix LD_LIBRARY_PATH" else "--prefix DYLD_LIBRARY_PATH"
-        } : "${lib.makeLibraryPath runtimeLibraries}"
+        --set PI_SKIP_VERSION_CHECK 1
 
       runHook postInstall
     '';

packages/ai/CHANGELOG.md

@@ -5,6 +5,7 @@
 ### Changed
 
 - OpenCode Go usage now comes from the official `GET /zen/go/v1/usage` endpoint (rolling 5h / weekly / monthly percent windows with server-computed resets) instead of synthesizing dollar estimates from OMP-observed request costs, so `/usage` reflects spend made outside OMP and the hardcoded $12/$30/$60 caps are gone. The usage probe now validates credentials (401 invalid key, 403 lapsed Go subscription), and a new ranking strategy routes multi-key pools by rolling/weekly headroom while keeping the monthly window display-only (an exhausted monthly can still serve requests via the console "Use balance" fallback) ([#8337](https://github.com/can1357/oh-my-pi/pull/8337) by [@will-bogusz](https://github.com/will-bogusz)).
+- Changed direct Anthropic prompt caching to two rolling 5-minute breakpoints on the latest real messages. Opt-in session owners can now keep the prefix warm through three bounded idle refreshes, using non-streaming `max_tokens: 0` when thinking is inactive and aborting immediately after generation begins when thinking is active.
 
 ### Fixed
 

packages/ai/src/providers/anthropic-client.ts

@@ -27,7 +27,7 @@ import { AnthropicApiError, AnthropicConnectionError, AnthropicConnectionTimeout
 export { AnthropicApiError, AnthropicConnectionError, AnthropicConnectionTimeoutError };
 
 import type { FetchImpl } from "../types";
-import type { MessageCreateParamsStreaming } from "./anthropic-wire";
+import type { MessageCreateParams } from "./anthropic-wire";
 
 /** Default pre-response timeout, matching the SDK's 10-minute default. */
 const DEFAULT_TIMEOUT_MS = 600_000;
@@ -173,7 +173,7 @@ export class AnthropicMessages {
 		this.#path = path;
 	}
 
-	create(params: MessageCreateParamsStreaming, options?: AnthropicRequestOptions): AnthropicApiRequest {
+	create(params: MessageCreateParams, options?: AnthropicRequestOptions): AnthropicApiRequest {
 		return this.#client.request(this.#path, params, options);
 	}
 }
@@ -184,8 +184,8 @@ export class AnthropicMessages {
  * alternative Messages-API client via `AnthropicOptions.client`.
  */
 export interface AnthropicMessagesClientLike {
-	messages: { create(params: MessageCreateParamsStreaming, options?: AnthropicRequestOptions): unknown };
-	beta?: { messages: { create(params: MessageCreateParamsStreaming, options?: AnthropicRequestOptions): unknown } };
+	messages: { create(params: MessageCreateParams, options?: AnthropicRequestOptions): unknown };
+	beta?: { messages: { create(params: MessageCreateParams, options?: AnthropicRequestOptions): unknown } };
 }
 
 export class AnthropicMessagesClient implements AnthropicMessagesClientLike {
@@ -199,7 +199,7 @@ export class AnthropicMessagesClient implements AnthropicMessagesClientLike {
 		this.beta = { messages: new AnthropicMessages(this, "/v1/messages?beta=true") };
 	}
 
-	request(path: string, params: MessageCreateParamsStreaming, options?: AnthropicRequestOptions): AnthropicApiRequest {
+	request(path: string, params: MessageCreateParams, options?: AnthropicRequestOptions): AnthropicApiRequest {
 		return new AnthropicApiRequest(() => this.#send(path, params, options));
 	}
 
@@ -218,11 +218,7 @@ export class AnthropicMessagesClient implements AnthropicMessagesClientLike {
 		return headers;
 	}
 
-	async #send(
-		path: string,
-		params: MessageCreateParamsStreaming,
-		options?: AnthropicRequestOptions,
-	): Promise<Response> {
+	async #send(path: string, params: MessageCreateParams, options?: AnthropicRequestOptions): Promise<Response> {
 		const opts = this.#options;
 		const fetchFn: FetchImpl = opts.fetch ?? fetch;
 		const callerSignal = options?.signal;

packages/ai/src/providers/anthropic.ts

@@ -80,6 +80,7 @@ import {
 	type ContentBlockParam,
 	type FallbackParam,
 	isAnthropicWebSearchHistoryBlock,
+	type MessageCreateParams,
 	type MessageCreateParamsStreaming,
 	type MessageParam,
 	type RawMessageStreamEvent,
@@ -482,17 +483,11 @@ function dropAnthropicStrictTools(params: MessageCreateParamsStreaming): void {
 function getCacheControl(
 	model: Model<"anthropic-messages">,
 	cacheRetention: CacheRetention | undefined,
-	isOAuthToken: boolean,
 ): { retention: CacheRetention; cacheControl?: AnthropicCacheControl } {
-	// OAuth mirrors Claude Code and always defaults to 1h retention. API-key
-	// requests also default to 1h where the endpoint supports it (canonical
-	// Anthropic API, `compat.supportsLongCacheRetention`): agent sessions
-	// routinely idle past 5 minutes waiting on background jobs, and a 5m
-	// breakpoint cold-misses the entire prefix on resume. PI_CACHE_RETENTION
-	// still overrides the API-key default in either direction.
-	const retention = isOAuthToken
-		? (cacheRetention ?? "long")
-		: resolveCacheRetention(cacheRetention, model.compat.supportsLongCacheRetention ? "long" : "short");
+	// Five-minute writes are the cheapest cache population strategy. Longer
+	// retention remains an explicit PI_CACHE_RETENTION/request override; idle
+	// sessions keep the short entry warm with bounded read-only refreshes.
+	const retention = resolveCacheRetention(cacheRetention, "short");
 	if (retention === "none") {
 		return { retention };
 	}
@@ -1653,6 +1648,31 @@ export function applyAnthropicUsageExtras(usage: Usage, source: AnthropicUsageLi
 	}
 }
 
+function parseAnthropicWireUsage(value: unknown): AnthropicWireUsage | undefined {
+	if (!isRecord(value)) return undefined;
+	const cacheCreation = isRecord(value.cache_creation)
+		? {
+				...(typeof value.cache_creation.ephemeral_5m_input_tokens === "number"
+					? { ephemeral_5m_input_tokens: value.cache_creation.ephemeral_5m_input_tokens }
+					: {}),
+				...(typeof value.cache_creation.ephemeral_1h_input_tokens === "number"
+					? { ephemeral_1h_input_tokens: value.cache_creation.ephemeral_1h_input_tokens }
+					: {}),
+			}
+		: undefined;
+	return {
+		...(typeof value.input_tokens === "number" ? { input_tokens: value.input_tokens } : {}),
+		...(typeof value.output_tokens === "number" ? { output_tokens: value.output_tokens } : {}),
+		...(typeof value.cache_read_input_tokens === "number"
+			? { cache_read_input_tokens: value.cache_read_input_tokens }
+			: {}),
+		...(typeof value.cache_creation_input_tokens === "number"
+			? { cache_creation_input_tokens: value.cache_creation_input_tokens }
+			: {}),
+		...(cacheCreation === undefined ? {} : { cache_creation: cacheCreation }),
+	};
+}
+
 function parseAnthropicFallbackWireBlock(value: unknown): AnthropicFallbackContent | undefined {
 	if (!isRecord(value) || value.type !== "fallback") return undefined;
 	const from = isRecord(value.from) && typeof value.from.model === "string" ? value.from.model : undefined;
@@ -1857,6 +1877,7 @@ const streamAnthropicOnce = (
 				});
 			}
 
+			const zeroOutputCacheRefresh = options?.anthropicCacheRefreshRequest === true;
 			let client: AnthropicMessagesClientLike;
 			let isOAuthToken: boolean;
 
@@ -1927,7 +1948,7 @@ const streamAnthropicOnce = (
 				// requests must not deviate from CC's header fingerprint.
 				if (
 					!(options?.isOAuth ?? isAnthropicOAuthToken(apiKey)) &&
-					getCacheControl(model, options?.cacheRetention, false).cacheControl?.ttl === "1h" &&
+					getCacheControl(model, options?.cacheRetention).cacheControl?.ttl === "1h" &&
 					!extraBetas.includes(extendedCacheTtlBeta)
 				) {
 					extraBetas.push(extendedCacheTtlBeta);
@@ -1958,7 +1979,7 @@ const streamAnthropicOnce = (
 					model,
 					apiKey,
 					extraBetas,
-					stream: true,
+					stream: !zeroOutputCacheRefresh,
 					interleavedThinking: options?.interleavedThinking ?? true,
 					headers: options?.headers,
 					dynamicHeaders: copilotDynamicHeaders?.headers,
@@ -2005,6 +2026,60 @@ const streamAnthropicOnce = (
 				return nextParams;
 			};
 			let params = await prepareParams();
+			const idleTimeoutMs = options?.streamIdleTimeoutMs ?? getStreamIdleTimeoutMs(model.compat.streamIdleTimeoutMs);
+			const firstEventTimeoutMs = options?.streamFirstEventTimeoutMs ?? getStreamFirstEventTimeoutMs(idleTimeoutMs);
+			const requestTimeoutMs =
+				firstEventTimeoutMs !== undefined && firstEventTimeoutMs > 0 ? firstEventTimeoutMs : undefined;
+
+			if (zeroOutputCacheRefresh) {
+				const refreshParams: MessageCreateParams = { ...params, max_tokens: 0, stream: false };
+				rawRequestDump = {
+					provider: model.provider,
+					api: output.api,
+					model: model.id,
+					method: "POST",
+					url: `${baseUrl}/v1/messages${isOAuthToken ? "?beta=true" : ""}`,
+					body: refreshParams,
+				};
+				const { requestSignal } = activeAbortTracker;
+				const requestOptions = {
+					...createSdkStreamRequestOptions(requestSignal, requestTimeoutMs),
+					maxRetries: 0,
+				};
+				const request: unknown =
+					isOAuthToken && client.beta
+						? client.beta.messages.create(refreshParams, requestOptions)
+						: client.messages.create(refreshParams, requestOptions);
+				if (!hasAnthropicRawResponseRequest(request)) {
+					throw new AIError.AnthropicStreamEnvelopeError(
+						"Anthropic cache refresh request did not expose a raw response",
+					);
+				}
+				const response = await request.asResponse();
+				await notifyProviderResponse(options, response, model, response.headers.get("request-id"));
+				const body: unknown = await response.json();
+				if (!isRecord(body)) {
+					throw new AIError.AnthropicStreamEnvelopeError("Anthropic cache refresh returned a malformed response");
+				}
+				const wireUsage = parseAnthropicWireUsage(body.usage);
+				if (!wireUsage) {
+					throw new AIError.AnthropicStreamEnvelopeError("Anthropic cache refresh response omitted usage");
+				}
+				if (typeof body.id === "string") output.responseId = body.id;
+				output.usage.input = wireUsage.input_tokens ?? 0;
+				output.usage.output = wireUsage.output_tokens ?? 0;
+				output.usage.cacheRead = wireUsage.cache_read_input_tokens ?? 0;
+				output.usage.cacheWrite = wireUsage.cache_creation_input_tokens ?? 0;
+				applyAnthropicUsageExtras(output.usage, wireUsage);
+				output.usage.totalTokens =
+					output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
+				calculateCost(model, output.usage);
+				output.duration = performance.now() - startTime;
+				stream.push({ type: "start", partial: output });
+				stream.push({ type: "done", reason: "stop", message: output });
+				stream.end();
+				return;
+			}
 
 			// Opt-in flag: the response parser only honors `fallback` content
 			// blocks and `usage.iterations` when the current request opted into
@@ -2019,10 +2094,6 @@ const streamAnthropicOnce = (
 				| (AnthropicServerToolContent & { [kStreamingPartialJson]?: string })
 				| (ToolCall & { [kStreamingPartialJson]: string; [kStreamingLastParseLen]?: number })
 			) & { [kStreamingBlockIndex]: number };
-			const idleTimeoutMs = options?.streamIdleTimeoutMs ?? getStreamIdleTimeoutMs(model.compat.streamIdleTimeoutMs);
-			const firstEventTimeoutMs = options?.streamFirstEventTimeoutMs ?? getStreamFirstEventTimeoutMs(idleTimeoutMs);
-			const requestTimeoutMs =
-				firstEventTimeoutMs !== undefined && firstEventTimeoutMs > 0 ? firstEventTimeoutMs : undefined;
 			const blocks = output.content as Block[];
 			const finalizeStreamBlock = (block: Block, contentIndex: number): void => {
 				if (block.type === "text") {
@@ -2807,65 +2878,19 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (model, con
 export type AnthropicSystemBlock = {
 	type: "text";
 	text: string;
-	cache_control?: AnthropicCacheControl;
 };
 type SystemBlockOptions = {
 	includeClaudeCodeInstruction?: boolean;
 	extraInstructions?: string[];
 	/** Text of the first user message — used as fingerprint seed for the billing header. */
 	firstUserMessageText?: string;
-	cacheControl?: AnthropicCacheControl;
 };
 
-/**
- * Place system-block cache breakpoints that survive volatile project context.
- *
- * omp normally appends its project footer (cwd, date, workspace tree) after the
- * stable system prefix. When cwd is outside a single direct child repository,
- * an active-repo context block follows that footer. Caching up to the last three
- * eligible blocks therefore covers both layouts:
- *
- * - stable prefix, project footer
- * - stable prefix, project footer, active-repo context
- *
- * A footer change can then fall back to the stable-prefix entry instead of
- * re-writing the entire system cache (issue #7324).
- *
- * @returns breakpoints placed, capped by `maxBreakpoints`.
- */
-function cacheSystemPrefixBreakpoints(
-	blocks: AnthropicSystemBlock[],
-	cacheControl: AnthropicCacheControl | undefined,
-	maxBreakpoints: number,
-	firstCacheableIndex: number,
-): number {
-	if (!cacheControl || maxBreakpoints <= 0) return 0;
-	let placed = 0;
-	for (let index = blocks.length - 1; index >= firstCacheableIndex && placed < maxBreakpoints; index--) {
-		if (blocks[index].cache_control != null) continue;
-		blocks[index] = { ...blocks[index], cache_control: cloneAnthropicCacheControl(cacheControl) };
-		placed++;
-	}
-	return placed;
-}
-
-/**
- * First system-block index that may carry a cache breakpoint. Skips the OAuth
- * cloak blocks that must stay uncached: the CC billing header (block 0, a
- * per-request fingerprint) and the Claude Code identity instruction (block 1).
- */
-function firstCacheableSystemIndex(blocks: readonly AnthropicSystemBlock[]): number {
-	let index = 0;
-	if (blocks[index]?.text?.startsWith(CLAUDE_BILLING_HEADER_PREFIX)) index++;
-	if (blocks[index]?.text === claudeCodeSystemInstruction) index++;
-	return index;
-}
-
 export function buildAnthropicSystemBlocks(
 	systemPrompt: readonly string[] | undefined,
 	options: SystemBlockOptions = {},
 ): AnthropicSystemBlock[] | undefined {
-	const { includeClaudeCodeInstruction = false, extraInstructions = [], firstUserMessageText, cacheControl } = options;
+	const { includeClaudeCodeInstruction = false, extraInstructions = [], firstUserMessageText } = options;
 	const sanitizedPrompts = normalizeSystemPrompts(systemPrompt);
 	const trimmedInstructions = extraInstructions.map(instruction => instruction.trim()).filter(Boolean);
 	const hasBillingHeader = sanitizedPrompts.some(prompt => prompt.startsWith(CLAUDE_BILLING_HEADER_PREFIX));
@@ -2882,7 +2907,6 @@ export function buildAnthropicSystemBlocks(
 		for (const prompt of sanitizedPrompts) {
 			blocks.push({ type: "text", text: prompt });
 		}
-		cacheSystemPrefixBreakpoints(blocks, cacheControl, 3, firstCacheableSystemIndex(blocks));
 
 		return blocks;
 	}
@@ -2894,10 +2918,6 @@ export function buildAnthropicSystemBlocks(
 	for (const prompt of sanitizedPrompts) {
 		blocks.push({ type: "text", text: prompt });
 	}
-	const lastIndex = blocks.length - 1;
-	if (cacheControl && lastIndex >= 0 && blocks[lastIndex].cache_control == null) {
-		blocks[lastIndex] = { ...blocks[lastIndex], cache_control: cloneAnthropicCacheControl(cacheControl) };
-	}
 	return blocks.length > 0 ? blocks : undefined;
 }
 
@@ -3149,29 +3169,17 @@ function ensureMaxTokensForThinking(params: MessageCreateParamsStreaming, maxAll
 	thinking.budget_tokens = clampedBudget;
 }
 
-type CacheControlBlock = {
-	cache_control?: AnthropicCacheControl | null;
-};
-
-function applyCacheControlToLastTextBlock(
-	blocks: Array<ContentBlockParam & CacheControlBlock>,
-	cacheControl: AnthropicCacheControl,
-): boolean {
-	if (blocks.length === 0) return false;
-	for (let i = blocks.length - 1; i >= 0; i--) {
-		if (blocks[i].type === "text") {
-			if (blocks[i].cache_control != null) return false;
-			blocks[i] = { ...blocks[i], cache_control: cloneAnthropicCacheControl(cacheControl) };
-			return true;
+function applyCacheControlToLastBlock(blocks: ContentBlockParam[], cacheControl: AnthropicCacheControl): boolean {
+	for (let index = blocks.length - 1; index >= 0; index--) {
+		const block = blocks[index];
+		// Anthropic rejects cache_control on generated reasoning and fallback
+		// boundary blocks. Preserve the requested trailing boundary on every
+		// ordinary content block, including tool use and tool results.
+		if (block.type === "thinking" || block.type === "redacted_thinking" || block.type === "fallback") {
+			continue;
 		}
-	}
-	// No text block — fall back to the last block that accepts cache_control;
-	// thinking/redacted_thinking blocks reject the field with a 400.
-	for (let i = blocks.length - 1; i >= 0; i--) {
-		const type = blocks[i].type;
-		if (type === "thinking" || type === "redacted_thinking") continue;
-		if (blocks[i].cache_control != null) return false;
-		blocks[i] = { ...blocks[i], cache_control: cloneAnthropicCacheControl(cacheControl) };
+		if ("cache_control" in block && block.cache_control != null) return false;
+		blocks[index] = { ...block, cache_control: cloneAnthropicCacheControl(cacheControl) };
 		return true;
 	}
 	return false;
@@ -3180,28 +3188,10 @@ function applyCacheControlToLastTextBlock(
 function applyPromptCaching(params: MessageCreateParamsStreaming, cacheControl?: AnthropicCacheControl): void {
 	if (!cacheControl) return;
 
-	const MAX_CACHE_BREAKPOINTS = 4;
-	let cacheBreakpointsUsed = countCacheControlBreakpoints(params);
-	if (cacheBreakpointsUsed >= MAX_CACHE_BREAKPOINTS) return;
-	let isCCLayout = false;
-
-	if (params.system && Array.isArray(params.system) && params.system.length > 0) {
-		isCCLayout = params.system[0]?.text?.startsWith(CLAUDE_BILLING_HEADER_PREFIX) === true;
-		const maxSystemBreakpoints = Math.min(3, MAX_CACHE_BREAKPOINTS - cacheBreakpointsUsed);
-		cacheBreakpointsUsed += cacheSystemPrefixBreakpoints(
-			params.system as AnthropicSystemBlock[],
-			cacheControl,
-			maxSystemBreakpoints,
-			isCCLayout ? firstCacheableSystemIndex(params.system as AnthropicSystemBlock[]) : 0,
-		);
-	}
-
-	if (cacheBreakpointsUsed >= MAX_CACHE_BREAKPOINTS) return;
-
 	// `convertAnthropicMessages` appends this neutral pad after a trailing
 	// assistant because Anthropic rejects assistant-prefill endings. It is absent
-	// from the next normal turn, so caching it wastes a scarce breakpoint; anchor
-	// the cache window on the preceding real assistant instead.
+	// from the next normal turn, so anchor the rolling window on the preceding
+	// real assistant instead.
 	const trailingIndex = params.messages.length - 1;
 	const trailingMessage = params.messages[trailingIndex];
 	const hasTrailingAssistantPad =
@@ -3209,158 +3199,18 @@ function applyPromptCaching(params: MessageCreateParamsStreaming, cacheControl?:
 		trailingMessage.content === "Continue." &&
 		params.messages[trailingIndex - 1]?.role === "assistant";
 	const messageEnd = hasTrailingAssistantPad ? trailingIndex - 1 : trailingIndex;
-	const messageWindowSize = isCCLayout ? 1 : 2;
-	const start = Math.max(0, messageEnd - messageWindowSize + 1);
-	for (let i = messageEnd; i >= start; i--) {
-		if (cacheBreakpointsUsed >= MAX_CACHE_BREAKPOINTS) break;
-		const message = params.messages[i];
+	const start = Math.max(0, messageEnd - 1);
+	for (let index = messageEnd; index >= start; index--) {
+		const message = params.messages[index];
 		if (!message) continue;
 		if (typeof message.content === "string") {
 			message.content = [
 				{ type: "text", text: message.content, cache_control: cloneAnthropicCacheControl(cacheControl) },
 			];
-			cacheBreakpointsUsed++;
-		} else if (Array.isArray(message.content) && message.content.length > 0) {
-			if (
-				applyCacheControlToLastTextBlock(
-					message.content as Array<ContentBlockParam & CacheControlBlock>,
-					cacheControl,
-				)
-			) {
-				cacheBreakpointsUsed++;
+		} else if (Array.isArray(message.content)) {
+			applyCacheControlToLastBlock(message.content, cacheControl);
 		}
 	}
-	}
-}
-
-function normalizeCacheControlBlockTtl(block: CacheControlBlock, seenFiveMinute: { value: boolean }): void {
-	const cacheControl = block.cache_control;
-	if (!cacheControl) return;
-	if (cacheControl.ttl !== "1h") {
-		seenFiveMinute.value = true;
-		return;
-	}
-	if (seenFiveMinute.value) {
-		const normalized = cloneAnthropicCacheControl(cacheControl);
-		delete normalized.ttl;
-		block.cache_control = normalized;
-	}
-}
-
-function normalizeCacheControlTtlOrdering(params: MessageCreateParamsStreaming): void {
-	const seenFiveMinute = { value: false };
-	if (params.tools) {
-		for (const tool of params.tools as Array<AnthropicWireTool & CacheControlBlock>) {
-			normalizeCacheControlBlockTtl(tool, seenFiveMinute);
-		}
-	}
-	if (params.system && Array.isArray(params.system)) {
-		for (const block of params.system as Array<AnthropicSystemBlock & CacheControlBlock>) {
-			normalizeCacheControlBlockTtl(block, seenFiveMinute);
-		}
-	}
-	for (const message of params.messages) {
-		if (!Array.isArray(message.content)) continue;
-		for (const block of message.content as Array<ContentBlockParam & CacheControlBlock>) {
-			normalizeCacheControlBlockTtl(block, seenFiveMinute);
-		}
-	}
-}
-
-function findLastCacheControlIndex<T extends CacheControlBlock>(blocks: T[]): number {
-	for (let index = blocks.length - 1; index >= 0; index--) {
-		if (blocks[index]?.cache_control != null) return index;
-	}
-	return -1;
-}
-
-function stripCacheControlExceptIndex<T extends CacheControlBlock>(
-	blocks: T[],
-	preserveIndex: number,
-	excessCounter: { value: number },
-): void {
-	for (let index = 0; index < blocks.length && excessCounter.value > 0; index++) {
-		if (index === preserveIndex) continue;
-		if (!blocks[index]?.cache_control) continue;
-		delete blocks[index].cache_control;
-		excessCounter.value--;
-	}
-}
-
-function stripAllCacheControl<T extends CacheControlBlock>(blocks: T[], excessCounter: { value: number }): void {
-	for (const block of blocks) {
-		if (excessCounter.value <= 0) return;
-		if (!block.cache_control) continue;
-		delete block.cache_control;
-		excessCounter.value--;
-	}
-}
-
-function stripMessageCacheControl(
-	messages: MessageCreateParamsStreaming["messages"],
-	excessCounter: { value: number },
-): void {
-	for (const message of messages) {
-		if (excessCounter.value <= 0) return;
-		if (!Array.isArray(message.content)) continue;
-		for (const block of message.content as Array<ContentBlockParam & CacheControlBlock>) {
-			if (excessCounter.value <= 0) return;
-			if (!block.cache_control) continue;
-			delete block.cache_control;
-			excessCounter.value--;
-		}
-	}
-}
-
-function countCacheControlBreakpoints(params: MessageCreateParamsStreaming): number {
-	let total = 0;
-	if (params.tools) {
-		for (const tool of params.tools as Array<AnthropicWireTool & CacheControlBlock>) {
-			if (tool.cache_control) total++;
-		}
-	}
-	if (params.system && Array.isArray(params.system)) {
-		for (const block of params.system as Array<AnthropicSystemBlock & CacheControlBlock>) {
-			if (block.cache_control) total++;
-		}
-	}
-	for (const message of params.messages) {
-		if (!Array.isArray(message.content)) continue;
-		for (const block of message.content as Array<ContentBlockParam & CacheControlBlock>) {
-			if (block.cache_control) total++;
-		}
-	}
-	return total;
-}
-
-function enforceCacheControlLimit(params: MessageCreateParamsStreaming, maxBreakpoints: number): void {
-	const total = countCacheControlBreakpoints(params);
-	if (total <= maxBreakpoints) return;
-	const excessCounter = { value: total - maxBreakpoints };
-	const systemBlocks =
-		params.system && Array.isArray(params.system)
-			? (params.system as Array<AnthropicSystemBlock & CacheControlBlock>)
-			: [];
-	const toolBlocks = (params.tools ?? []) as Array<AnthropicWireTool & CacheControlBlock>;
-	const lastSystemIndex = findLastCacheControlIndex(systemBlocks);
-	const lastToolIndex = findLastCacheControlIndex(toolBlocks);
-	if (systemBlocks.length > 0) {
-		stripCacheControlExceptIndex(systemBlocks, lastSystemIndex, excessCounter);
-	}
-	if (excessCounter.value <= 0) return;
-	if (toolBlocks.length > 0) {
-		stripCacheControlExceptIndex(toolBlocks, lastToolIndex, excessCounter);
-	}
-	if (excessCounter.value <= 0) return;
-	stripMessageCacheControl(params.messages, excessCounter);
-	if (excessCounter.value <= 0) return;
-	if (systemBlocks.length > 0) {
-		stripAllCacheControl(systemBlocks, excessCounter);
-	}
-	if (excessCounter.value <= 0) return;
-	if (toolBlocks.length > 0) {
-		stripAllCacheControl(toolBlocks, excessCounter);
-	}
 }
 
 function usesAdaptiveThinkingTagOnly(model: Model<"anthropic-messages">): boolean {
@@ -3445,7 +3295,7 @@ function buildParams(
 		forceDemoteUnsignedThinking && model.compat.replayUnsignedThinking
 			? { ...model, compat: { ...model.compat, replayUnsignedThinking: false } }
 			: model;
-	const { cacheControl } = getCacheControl(model, options?.cacheRetention, isOAuthToken);
+	const { cacheControl } = getCacheControl(model, options?.cacheRetention);
 
 	// Pre-compute system blocks so they occupy the right slot in the serialized body.
 	const shouldInjectClaudeCodeInstruction = isOAuthToken && !model.id.startsWith("claude-3-5-haiku");
@@ -3582,7 +3432,7 @@ function buildParams(
 		...(systemBlocks && { system: systemBlocks }),
 		...(tools !== undefined && { tools }),
 		...(metadata && { metadata }),
-		max_tokens: Math.min(maxOutputTokens, options?.maxTokens || modelMaxTokens),
+		max_tokens: Math.min(maxOutputTokens, options?.maxTokens ?? modelMaxTokens),
 		...(thinking && { thinking }),
 		...(contextManagement && { context_management: contextManagement }),
 		...(outputConfig && { output_config: outputConfig }),
@@ -3647,8 +3497,6 @@ function buildParams(
 	disableThinkingIfToolChoiceForced(params, model);
 	ensureMaxTokensForThinking(params, maxOutputTokens);
 	applyPromptCaching(params, cacheControl);
-	enforceCacheControlLimit(params, 4);
-	normalizeCacheControlTtlOrdering(params);
 
 	return params;
 }

packages/ai/src/stream.ts

@@ -24,6 +24,7 @@ import { isInvalidatedOAuthTokenError } from "./error/auth-classify";
 import { isConcurrencyCapExclusion, isUsageLimitOutcome } from "./error/rate-limit";
 import type { BedrockOptions } from "./providers/amazon-bedrock";
 import type { AnthropicOptions } from "./providers/anthropic";
+import type { MessageCreateParamsStreaming } from "./providers/anthropic-wire";
 import { coworkFetch } from "./providers/cowork-fetch";
 import type { CursorOptions } from "./providers/cursor";
 import type { DevinOptions } from "./providers/devin";
@@ -69,6 +70,7 @@ import type {
 	FetchImpl,
 	Model,
 	OptionsForApi,
+	ProviderSessionState,
 	SimpleStreamOptions,
 	StreamOptions,
 	ThinkingBudgets,
@@ -1118,10 +1120,296 @@ function emitBufferedEvents(stream: AssistantMessageEventStream, events: Assista
 	}
 }
 
+const ANTHROPIC_CACHE_TTL_MS = 5 * 60_000;
+const ANTHROPIC_CACHE_REFRESH_LEAD_MS = 15_000;
+const ANTHROPIC_CACHE_REFRESH_LIMIT = 3;
+const ANTHROPIC_CACHE_REFRESH_STATE_KEY = "anthropic-cache-refresh";
+
+interface AnthropicCacheRefreshPlan {
+	refresh(controller: AbortController): Promise<number | undefined>;
+}
+
+class AnthropicCacheRefreshState implements ProviderSessionState {
+	#controller: AbortController | undefined;
+	#generation = 0;
+	#plan: AnthropicCacheRefreshPlan | undefined;
+	#refreshesRemaining = 0;
+	#timer: NodeJS.Timeout | undefined;
+
+	cancel(): void {
+		this.#generation++;
+		if (this.#timer !== undefined) {
+			clearTimeout(this.#timer);
+			this.#timer = undefined;
+		}
+		this.#controller?.abort();
+		this.#controller = undefined;
+		this.#plan = undefined;
+		this.#refreshesRemaining = 0;
+	}
+
+	arm(plan: AnthropicCacheRefreshPlan, cacheTouchedAtMs: number): void {
+		this.cancel();
+		this.#plan = plan;
+		this.#refreshesRemaining = ANTHROPIC_CACHE_REFRESH_LIMIT;
+		this.#schedule(cacheTouchedAtMs, this.#generation);
+	}
+
+	close(): void {
+		this.cancel();
+	}
+
+	#schedule(cacheTouchedAtMs: number, generation: number): void {
+		const refreshAtMs = cacheTouchedAtMs + ANTHROPIC_CACHE_TTL_MS - ANTHROPIC_CACHE_REFRESH_LEAD_MS;
+		this.#timer = setTimeout(
+			() => {
+				this.#timer = undefined;
+				void this.#refresh(generation);
+			},
+			Math.max(0, refreshAtMs - Date.now()),
+		);
+		this.#timer.unref?.();
+	}
+
+	async #refresh(generation: number): Promise<void> {
+		const plan = this.#plan;
+		if (generation !== this.#generation || !plan || this.#refreshesRemaining <= 0) return;
+
+		const controller = new AbortController();
+		this.#controller = controller;
+		let cacheTouchedAtMs: number | undefined;
+		try {
+			cacheTouchedAtMs = await plan.refresh(controller);
+		} catch (error) {
+			if (generation === this.#generation && !controller.signal.aborted) {
+				logger.debug("Anthropic prompt-cache refresh failed", { error: String(error) });
+			}
+		}
+		if (generation !== this.#generation) return;
+
+		this.#controller = undefined;
+		if (cacheTouchedAtMs === undefined) {
+			this.#plan = undefined;
+			this.#refreshesRemaining = 0;
+			return;
+		}
+
+		this.#refreshesRemaining--;
+		if (this.#refreshesRemaining <= 0) {
+			this.#plan = undefined;
+			return;
+		}
+		this.#schedule(cacheTouchedAtMs, generation);
+	}
+}
+
+function supportsAnthropicCacheRefresh<TApi extends Api>(model: Model<TApi>): boolean {
+	return (
+		model.api === "anthropic-messages" &&
+		model.provider === "anthropic" &&
+		model.transport !== "pi-native" &&
+		isLeakedThinkingHealExempt(model)
+	);
+}
+
+function isAnthropicRefreshPayload(payload: unknown): payload is MessageCreateParamsStreaming {
+	return (
+		typeof payload === "object" &&
+		payload !== null &&
+		"messages" in payload &&
+		Array.isArray(payload.messages) &&
+		"max_tokens" in payload &&
+		typeof payload.max_tokens === "number"
+	);
+}
+
+function isShortAnthropicCacheControl(cacheControl: unknown): boolean {
+	return (
+		typeof cacheControl === "object" &&
+		cacheControl !== null &&
+		"type" in cacheControl &&
+		cacheControl.type === "ephemeral" &&
+		(!("ttl" in cacheControl) || cacheControl.ttl !== "1h")
+	);
+}
+
+function hasShortAnthropicMessageBreakpoint(payload: MessageCreateParamsStreaming): boolean {
+	for (const message of payload.messages) {
+		if (!Array.isArray(message.content)) continue;
+		for (const block of message.content) {
+			if ("cache_control" in block && isShortAnthropicCacheControl(block.cache_control)) return true;
+		}
+	}
+	return false;
+}
+
+function isAnthropicGenerationEvent(event: AssistantMessageEvent): boolean {
+	switch (event.type) {
+		case "text_start":
+		case "thinking_start":
+		case "toolcall_start":
+		case "image_end":
+			return true;
+		case "text_delta":
+		case "thinking_delta":
+		case "toolcall_delta":
+			return event.delta.length > 0;
+		default:
+			return false;
+	}
+}
+
+function isAnthropicThinkingActive(model: Model<Api>, payload: MessageCreateParamsStreaming): boolean {
+	if (payload.thinking) return payload.thinking.type !== "disabled";
+	return model.thinking?.mode === "anthropic-adaptive" && payload.output_config?.effort != null;
+}
+
+function createAnthropicCacheRefreshPlan<TApi extends Api>(
+	model: Model<TApi>,
+	context: Context,
+	options: SimpleStreamOptions | undefined,
+	payload: MessageCreateParamsStreaming,
+): AnthropicCacheRefreshPlan {
+	const thinkingEnabled = isAnthropicThinkingActive(model, payload);
+	return {
+		async refresh(controller) {
+			let cacheRead = 0;
+			let cacheWrite = 0;
+			let cacheTouchedAtMs: number | undefined;
+			let canceledAfterGenerationStarted = false;
+			const response = streamSimpleRequest(model, context, {
+				...options,
+				acceptEmptyResponse: true,
+				anthropicCacheRefreshRequest: !thinkingEnabled,
+				cacheRetention: "short",
+				maxTokens: thinkingEnabled ? options?.maxTokens : 0,
+				onPayload: () => ({
+					...payload,
+					max_tokens: thinkingEnabled ? payload.max_tokens : 0,
+				}),
+				onResponse: () => {
+					cacheTouchedAtMs = Date.now();
+				},
+				onSseEvent: undefined,
+				signal: controller.signal,
+			});
+
+			for await (const event of response) {
+				if ("partial" in event) {
+					cacheRead = event.partial.usage.cacheRead;
+					cacheWrite = event.partial.usage.cacheWrite;
+				}
+				if (event.type === "error") return undefined;
+				if (event.type === "done") {
+					cacheRead = event.message.usage.cacheRead;
+					cacheWrite = event.message.usage.cacheWrite;
+					return cacheTouchedAtMs !== undefined && cacheRead > 0 && cacheWrite === 0
+						? cacheTouchedAtMs
+						: undefined;
+				}
+				if (thinkingEnabled && isAnthropicGenerationEvent(event)) {
+					canceledAfterGenerationStarted = true;
+					controller.abort();
+					break;
+				}
+			}
+
+			if (canceledAfterGenerationStarted) {
+				try {
+					await response.result();
+				} catch (error) {
+					if (!controller.signal.aborted) throw error;
+				}
+			}
+			return cacheTouchedAtMs !== undefined && cacheRead > 0 && cacheWrite === 0 ? cacheTouchedAtMs : undefined;
+		},
+	};
+}
+
+function streamSimpleWithAnthropicCacheRefresh<TApi extends Api>(
+	model: Model<TApi>,
+	context: Context,
+	options: SimpleStreamOptions | undefined,
+): AssistantMessageEventStream {
+	const providerSessionState = options?.providerSessionState;
+	if (!options?.anthropicCacheRefresh || !providerSessionState) {
+		return streamSimpleRequest(model, context, options);
+	}
+
+	const existingState = providerSessionState.get(ANTHROPIC_CACHE_REFRESH_STATE_KEY);
+	if (existingState instanceof AnthropicCacheRefreshState) {
+		existingState.cancel();
+	} else if (existingState) {
+		return streamSimpleRequest(model, context, options);
+	}
+	if (!supportsAnthropicCacheRefresh(model) || resolveCacheRetention(options.cacheRetention) !== "short") {
+		return streamSimpleRequest(model, context, options);
+	}
+
+	const refreshState = existingState ?? new AnthropicCacheRefreshState();
+	if (!existingState) providerSessionState.set(ANTHROPIC_CACHE_REFRESH_STATE_KEY, refreshState);
+
+	let cacheTouchedAtMs: number | undefined;
+	let capturedPayload: MessageCreateParamsStreaming | undefined;
+	const inner = streamSimpleRequest(model, context, {
+		...options,
+		onPayload: async (payload, payloadModel) => {
+			const replacement = await options?.onPayload?.(payload, payloadModel);
+			const finalPayload = replacement ?? payload;
+			if (isAnthropicRefreshPayload(finalPayload)) capturedPayload = finalPayload;
+			return replacement;
+		},
+		onResponse: async (response, responseModel) => {
+			cacheTouchedAtMs = Date.now();
+			await options?.onResponse?.(response, responseModel);
+		},
+	});
+	const outer = new AssistantMessageEventStream();
+	const armRefresh = (message: AssistantMessage): void => {
+		if (
+			message.stopReason === "error" ||
+			message.stopReason === "aborted" ||
+			message.usage.cacheRead + message.usage.cacheWrite <= 0 ||
+			cacheTouchedAtMs === undefined ||
+			capturedPayload === undefined ||
+			!hasShortAnthropicMessageBreakpoint(capturedPayload)
+		) {
+			return;
+		}
+		refreshState.arm(createAnthropicCacheRefreshPlan(model, context, options, capturedPayload), cacheTouchedAtMs);
+	};
+
+	void (async () => {
+		try {
+			for await (const event of inner) {
+				if (event.type === "done") armRefresh(event.message);
+				outer.push(event);
+				if (outer.done) return;
+			}
+			if (!outer.done) {
+				const result = await inner.result();
+				armRefresh(result);
+				outer.end(result);
+			}
+		} catch (error) {
+			outer.fail(error);
+		}
+	})();
+	return outer;
+}
+
 export function streamSimple<TApi extends Api>(
 	model: Model<TApi>,
 	context: Context,
 	options?: SimpleStreamOptions,
+): AssistantMessageEventStream {
+	return streamSimpleWithAnthropicCacheRefresh(model, context, options);
+}
+
+function streamSimpleRequest<TApi extends Api>(
+	model: Model<TApi>,
+	context: Context,
+	options?: SimpleStreamOptions,
 ): AssistantMessageEventStream {
 	const inputOptions = (options || {}) as SimpleStreamOptions;
 	const baseOptions = { ...inputOptions, fetch: inputOptions.fetch ?? defaultFetchForModel(model) };
@@ -1147,7 +1435,7 @@ export function streamSimple<TApi extends Api>(
 			};
 
 			try {
-				const inner = streamSimple(model, context, { ...requestOptions, apiKey });
+				const inner = streamSimpleRequest(model, context, { ...requestOptions, apiKey });
 				for await (const event of inner) {
 					if (!emittedReplayUnsafeEvent && event.type === "start") {
 						bufferedEvents.push(event);
@@ -1613,6 +1901,7 @@ function mapOptionsForApi<TApi extends Api>(
 		fetch: options?.fetch,
 		fallbacks: options?.fallbacks,
 		acceptEmptyResponse: options?.acceptEmptyResponse,
+		anthropicCacheRefreshRequest: options?.anthropicCacheRefreshRequest,
 		...simpleProviderOptions,
 	};
 

packages/ai/src/types.ts

@@ -412,6 +412,16 @@ export interface StreamOptions {
 	signal?: AbortSignal;
 	apiKey?: string;
 	cacheRetention?: CacheRetention;
+	/**
+	 * Keep Anthropic's 5-minute prompt cache warm across bounded idle gaps.
+	 *
+	 * This is an ownership flag, not a general provider default: exactly one
+	 * primary agent loop sharing `providerSessionState` should enable it.
+	 * Side-channel and advisor requests must leave it unset.
+	 */
+	anthropicCacheRefresh?: boolean;
+	/** @internal Marks a replay-only Anthropic request that must use non-streaming `max_tokens: 0`. */
+	anthropicCacheRefreshRequest?: boolean;
 	/**
 	 * Additional headers to include in provider requests.
 	 * These are merged on top of model-defined headers.
@@ -559,10 +569,10 @@ export interface StreamOptions {
 	 */
 	providerRetryWait?: (delayMs: number, signal?: AbortSignal) => Promise<void>;
 	/**
-	 * Accept a Google `STOP` response with no visible text or tool call as a
-	 * successful completion. Passive callers such as advisors use this because
-	 * silence is a valid result; interactive agent turns retain empty-response
-	 * retries by default. Ignored by non-Google providers.
+	 * Accept a normal provider stop with no visible text or tool call as a
+	 * successful completion. Passive callers and zero-output cache refreshes use
+	 * this because silence is their expected result; interactive agent turns
+	 * retain empty-response retries by default.
 	 */
 	acceptEmptyResponse?: boolean;
 	/**

packages/ai/src/utils/empty-completion-retry.ts

@@ -63,6 +63,7 @@ function isMeaningfulCompletionEvent(event: AssistantMessageEvent): boolean {
 interface EmptyCompletionRetryOptions {
 	signal?: AbortSignal;
 	providerRetryWait?: (delayMs: number, signal?: AbortSignal) => Promise<void>;
+	acceptEmptyResponse?: boolean;
 }
 
 /**
@@ -82,7 +83,7 @@ export function withEmptyCompletionRetry<M, O extends EmptyCompletionRetryOption
 		for (let emptyAttempt = 0; ; emptyAttempt++) {
 			const inner = attempt(model, context, options);
 			const buffered: AssistantMessageEvent[] = [];
-			let committed = false;
+			let committed = options?.acceptEmptyResponse === true;
 			let terminal: AssistantMessageEvent | undefined;
 			const flush = (): void => {
 				for (const event of buffered) outer.push(event);
@@ -117,6 +118,7 @@ export function withEmptyCompletionRetry<M, O extends EmptyCompletionRetryOption
 			// one-token invisible stop is still the same empty-completion failure.
 			const message = terminal?.type === "done" ? terminal.message : undefined;
 			const isRetryableEmpty =
+				options?.acceptEmptyResponse !== true &&
 				!committed &&
 				message !== undefined &&
 				message.stopReason === "stop" &&

packages/ai/test/anthropic-alignment.test.ts

@@ -9,7 +9,6 @@ import {
 	applyClaudeToolPrefix,
 	buildAnthropicClientOptions,
 	buildAnthropicHeaders,
-	buildAnthropicSystemBlocks,
 	claudeCodeSystemInstruction,
 	claudeToolPrefix,
 	deriveClaudeDeviceId,
@@ -19,7 +18,7 @@ import {
 	streamAnthropic,
 	stripClaudeToolPrefix,
 } from "@oh-my-pi/pi-ai/providers/anthropic";
-import type { MessageCreateParamsStreaming } from "@oh-my-pi/pi-ai/providers/anthropic-wire";
+import type { MessageCreateParams } from "@oh-my-pi/pi-ai/providers/anthropic-wire";
 import { claudeCodeVersion } from "@oh-my-pi/pi-ai/providers/claude-code-fingerprint";
 import { getEnvApiKey, streamSimple } from "@oh-my-pi/pi-ai/stream";
 import type {
@@ -259,90 +258,7 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(options.defaultHeaders["anthropic-beta"]).not.toContain("context-1m-2025-08-07");
 	});
 
-	it("caches the stable prefix and the trailing block while leaving billing + CC identity uncached", () => {
-		const blocks = buildAnthropicSystemBlocks(["Stay concise."], {
-			includeClaudeCodeInstruction: true,
-			extraInstructions: ["Use citations when possible"],
-			cacheControl: { type: "ephemeral" },
-		});
-
-		expect(blocks).toHaveLength(4);
-		// OAuth cloak blocks stay uncached: the billing header is a per-request
-		// fingerprint and the CC identity block mimics Claude Code.
-		expect(blocks?.[0].text).toStartWith("x-anthropic-billing-header:");
-		expect(blocks?.[0].cache_control).toBeUndefined();
-		expect(blocks?.[1].text).toBe(claudeCodeSystemInstruction);
-		expect(blocks?.[1].cache_control).toBeUndefined();
-		// Stable-prefix breakpoint on the block before the trailing footer, plus a
-		// full-match breakpoint on the trailing block itself (#7324).
-		expect(blocks?.[2]).toEqual({
-			type: "text",
-			text: "Use citations when possible",
-			cache_control: { type: "ephemeral" },
-		});
-		expect(blocks?.[3]).toEqual({
-			type: "text",
-			text: "Stay concise.",
-			cache_control: { type: "ephemeral" },
-		});
-	});
-
-	it("keeps the stable-prefix breakpoint when the trailing project footer (cwd/date) changes (#7324)", () => {
-		const staticInstructions = "STATIC INSTRUCTIONS BLOCK";
-		const runA = buildAnthropicSystemBlocks([staticInstructions, "PROJECT\nToday is 2026-08-01, cwd '/tmp/a'."], {
-			includeClaudeCodeInstruction: true,
-			cacheControl: { type: "ephemeral" },
-		});
-		const runB = buildAnthropicSystemBlocks([staticInstructions, "PROJECT\nToday is 2026-08-02, cwd '/tmp/b'."], {
-			includeClaudeCodeInstruction: true,
-			cacheControl: { type: "ephemeral" },
-		});
-
-		for (const blocks of [runA, runB]) {
-			expect(blocks).toHaveLength(4);
-			expect(blocks?.[0].cache_control).toBeUndefined();
-			expect(blocks?.[1].cache_control).toBeUndefined();
-			// The static block carries a breakpoint whose prefix excludes the
-			// volatile footer, so a cwd/date change reuses it instead of
-			// re-writing the whole system cache.
-			expect(blocks?.[2].text).toBe(staticInstructions);
-			expect(blocks?.[2].cache_control).toEqual({ type: "ephemeral" });
-			expect(blocks?.[3].cache_control).toEqual({ type: "ephemeral" });
-		}
-		expect(runA?.[2].text).toBe(runB?.[2].text);
-	});
-
-	it("caches before the project footer when active-repo context follows it (#7324)", () => {
-		const staticInstructions = "STATIC INSTRUCTIONS BLOCK";
-		const projectFooter = "PROJECT\nToday is 2026-08-01, cwd '/tmp'.";
-		const activeRepoContext = "The active repository is './repo'.";
-		const blocks = buildAnthropicSystemBlocks([staticInstructions, projectFooter, activeRepoContext], {
-			includeClaudeCodeInstruction: true,
-			cacheControl: { type: "ephemeral" },
-		});
-
-		// blocks: [billing, CC identity, static, project footer, active-repo context]
-		expect(blocks).toHaveLength(5);
-		expect(blocks?.[0].cache_control).toBeUndefined();
-		expect(blocks?.[1].cache_control).toBeUndefined();
-		expect(blocks?.[2]).toEqual({
-			type: "text",
-			text: staticInstructions,
-			cache_control: { type: "ephemeral" },
-		});
-		expect(blocks?.[3]).toEqual({
-			type: "text",
-			text: projectFooter,
-			cache_control: { type: "ephemeral" },
-		});
-		expect(blocks?.[4]).toEqual({
-			type: "text",
-			text: activeRepoContext,
-			cache_control: { type: "ephemeral" },
-		});
-	});
-
-	it("caches Claude Code context and the last user block in OAuth request payloads", async () => {
+	it("places a short breakpoint only on the trailing message in a one-message OAuth request", async () => {
 		const payload = (await captureAnthropicPayload(ANTHROPIC_MODEL, {
 			systemPrompt: ["Stay concise."],
 			messages: [{ role: "user", content: "Hi", timestamp: Date.now() }],
@@ -355,12 +271,11 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(payload.system?.[0]?.cache_control).toBeUndefined();
 		expect(payload.system?.[1]?.text).toBe(claudeCodeSystemInstruction);
 		expect(payload.system?.[1]?.cache_control).toBeUndefined();
-		expect(payload.system?.[2]?.cache_control).toEqual({ type: "ephemeral", ttl: "1h" });
+		expect(payload.system?.[2]?.cache_control).toBeUndefined();
 		const content = payload.messages?.[0]?.content;
 		expect(Array.isArray(content)).toBe(true);
 		expect(Array.isArray(content) ? content[0]?.cache_control : undefined).toEqual({
 			type: "ephemeral",
-			ttl: "1h",
 		});
 	});
 
@@ -380,7 +295,6 @@ describe("Anthropic request fingerprint alignment", () => {
 		const content = payload.messages?.[0]?.content;
 		expect(Array.isArray(content) ? content[0]?.cache_control : undefined).toEqual({
 			type: "ephemeral",
-			ttl: "1h",
 		});
 	});
 
@@ -416,15 +330,23 @@ describe("Anthropic request fingerprint alignment", () => {
 					timestamp: Date.now(),
 				},
 			],
-		})) as { messages?: Array<{ content?: Array<{ type?: string; cache_control?: unknown }> | string }> };
+		})) as {
+			system?: Array<{ cache_control?: unknown }>;
+			messages?: Array<{ content?: Array<{ type?: string; cache_control?: unknown }> | string }>;
+		};
 
+		expect(payload.system?.some(block => block.cache_control != null)).toBe(false);
 		const messages = payload.messages ?? [];
-		const lastContent = messages[messages.length - 1]?.content;
-		expect(Array.isArray(lastContent)).toBe(true);
-		expect(Array.isArray(lastContent) ? lastContent[0]?.type : undefined).toBe("tool_result");
-		expect(Array.isArray(lastContent) ? lastContent[0]?.cache_control : undefined).toEqual({
+		expect(messages[0]?.content).toBe("Use the tool");
+		const assistantContent = messages.at(-2)?.content;
+		expect(Array.isArray(assistantContent) ? assistantContent.at(-1)?.type : undefined).toBe("tool_use");
+		expect(Array.isArray(assistantContent) ? assistantContent.at(-1)?.cache_control : undefined).toEqual({
+			type: "ephemeral",
+		});
+		const lastContent = messages.at(-1)?.content;
+		expect(Array.isArray(lastContent) ? lastContent.at(-1)?.type : undefined).toBe("tool_result");
+		expect(Array.isArray(lastContent) ? lastContent.at(-1)?.cache_control : undefined).toEqual({
 			type: "ephemeral",
-			ttl: "1h",
 		});
 	});
 
@@ -504,7 +426,7 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(last?.content).toBe("Continue.");
 	});
 
-	it("caches the real assistant before a synthetic Continue pad when the breakpoint budget is tight", async () => {
+	it("caches the last two real messages and ignores a synthetic Continue pad", async () => {
 		const assistant: AssistantMessage = {
 			role: "assistant",
 			content: [{ type: "text", text: "real assistant answer" }],
@@ -534,11 +456,14 @@ describe("Anthropic request fingerprint alignment", () => {
 			messages?: Array<{ role: string; content: string | Array<{ cache_control?: unknown }> }>;
 		};
 
-		expect(payload.system?.filter(block => block.cache_control != null)).toHaveLength(3);
+		expect(payload.system?.some(block => block.cache_control != null)).toBe(false);
+		const userContent = payload.messages?.[0]?.content;
+		expect(Array.isArray(userContent) ? userContent[0]?.cache_control : undefined).toEqual({
+			type: "ephemeral",
+		});
 		const assistantContent = payload.messages?.find(message => message.role === "assistant")?.content;
 		expect(Array.isArray(assistantContent) ? assistantContent[0]?.cache_control : undefined).toEqual({
 			type: "ephemeral",
-			ttl: "1h",
 		});
 		const pad = payload.messages?.at(-1);
 		expect(pad?.content).toBe("Continue.");
@@ -635,7 +560,7 @@ describe("Anthropic request fingerprint alignment", () => {
 				efforts: [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High, Effort.XHigh],
 			},
 		});
-		let capturedParams: MessageCreateParamsStreaming | undefined;
+		let capturedParams: MessageCreateParams | undefined;
 		let capturedOptions: { headers?: Record<string, string> } | undefined;
 		await streamAnthropic(
 			adaptiveModel,
@@ -673,7 +598,7 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(capturedOptions?.headers?.["anthropic-beta"] ?? "").toContain("effort-2025-11-24");
 	});
 
-	it("adds the extended-cache-ttl beta to API-key requests that default to 1h caching", async () => {
+	it("adds the extended-cache-ttl beta only when 1h caching is requested", async () => {
 		const captureBeta = () => {
 			let captured: string | undefined;
 			const fetchMock = (async (_input: string | URL | Request, init?: RequestInit) => {
@@ -690,18 +615,25 @@ describe("Anthropic request fingerprint alignment", () => {
 			messages: [{ role: "user", content: "Hi", timestamp: Date.now() }],
 		};
 
-		const canonical = captureBeta();
+		const short = captureBeta();
 		await streamAnthropic(ANTHROPIC_MODEL, cacheContext, {
 			apiKey: "sk-ant-api-test",
-			fetch: canonical.fetchMock,
+			fetch: short.fetchMock,
 		}).result();
-		expect(canonical.beta()).toContain("extended-cache-ttl-2025-04-11");
+		expect(short.beta()).not.toContain("extended-cache-ttl-2025-04-11");
+
+		const long = captureBeta();
+		await streamAnthropic(ANTHROPIC_MODEL, cacheContext, {
+			apiKey: "sk-ant-api-test",
+			cacheRetention: "long",
+			fetch: long.fetchMock,
+		}).result();
+		expect(long.beta()).toContain("extended-cache-ttl-2025-04-11");
 
-		// Endpoints without long-cache support never send `ttl: "1h"`, so the
-		// companion beta must stay off the wire too.
 		const proxy = captureBeta();
 		await streamAnthropic(UMANS_ANTHROPIC_MODEL, cacheContext, {
 			apiKey: "sk-umans-test",
+			cacheRetention: "long",
 			fetch: proxy.fetchMock,
 		}).result();
 		expect(proxy.beta()).not.toContain("extended-cache-ttl-2025-04-11");
@@ -832,7 +764,7 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(extractSuffix(billingWithDev)).toBe(extractSuffix(billingUserOnly));
 	});
 
-	it("caches the trailing and stable-prefix system blocks on API-key requests (#7324)", async () => {
+	it("leaves system blocks uncached on API-key requests", async () => {
 		const payload = (await captureAnthropicPayload(
 			ANTHROPIC_MODEL,
 			{
@@ -843,10 +775,8 @@ describe("Anthropic request fingerprint alignment", () => {
 		)) as { system?: Array<{ type: string; text?: string; cache_control?: unknown }> };
 
 		expect(payload.system).toEqual([
-			// Stable-prefix breakpoint: reused when the trailing block changes.
-			{ type: "text", text: "stable system", cache_control: { type: "ephemeral", ttl: "1h" } },
-			// Canonical Anthropic API-key requests default to the 1h breakpoint.
-			{ type: "text", text: "stable durable context", cache_control: { type: "ephemeral", ttl: "1h" } },
+			{ type: "text", text: "stable system" },
+			{ type: "text", text: "stable durable context" },
 		]);
 	});
 

packages/ai/test/anthropic-cache-refresh.test.ts

@@ -0,0 +1,287 @@
+import { afterEach, describe, expect, it, vi } from "bun:test";
+import { streamSimple } from "@oh-my-pi/pi-ai";
+import type { MessageCreateParams } from "@oh-my-pi/pi-ai/providers/anthropic-wire";
+import type { Context, FetchImpl, Model, ProviderSessionState } from "@oh-my-pi/pi-ai/types";
+import { buildModel } from "@oh-my-pi/pi-catalog/build";
+
+const CACHE_REFRESH_DELAY_MS = 5 * 60_000 - 15_000;
+const CACHE_TOKENS = 1_200;
+
+const model: Model<"anthropic-messages"> = buildModel({
+	id: "claude-sonnet-4-6",
+	name: "Claude Sonnet 4.6",
+	api: "anthropic-messages",
+	provider: "anthropic",
+	baseUrl: "https://api.anthropic.com",
+	reasoning: false,
+	input: ["text"],
+	cost: { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
+	contextWindow: 200_000,
+	maxTokens: 8_192,
+});
+
+const thinkingModel: Model<"anthropic-messages"> = buildModel({
+	...model,
+	reasoning: true,
+});
+
+const context: Context = {
+	messages: [{ role: "user", content: "Keep this prefix warm.", timestamp: 1 }],
+};
+
+type ResponseMode = "ordinary-write" | "ordinary-roll" | "refresh-read" | "thinking-refresh";
+
+interface FetchCapture {
+	bodies: MessageCreateParams[];
+	thinkingRefreshAborted: boolean;
+}
+
+const stateMaps: Array<Map<string, ProviderSessionState>> = [];
+
+function createProviderSessionState(): Map<string, ProviderSessionState> {
+	const states = new Map<string, ProviderSessionState>();
+	stateMaps.push(states);
+	return states;
+}
+
+function sseResponse(events: Array<Record<string, unknown>>): Response {
+	const body = `${events.map(event => `event: ${String(event.type)}\ndata: ${JSON.stringify(event)}`).join("\n\n")}\n\n`;
+	return new Response(body, {
+		status: 200,
+		headers: { "Content-Type": "text/event-stream", "request-id": "req_cache_refresh" },
+	});
+}
+
+function usage(cacheRead: number, cacheWrite: number, output: number): Record<string, unknown> {
+	return {
+		input_tokens: 0,
+		output_tokens: output,
+		cache_read_input_tokens: cacheRead,
+		cache_creation_input_tokens: cacheWrite,
+		cache_creation: {
+			ephemeral_5m_input_tokens: cacheWrite,
+			ephemeral_1h_input_tokens: 0,
+		},
+	};
+}
+
+function ordinaryResponse(mode: "ordinary-write" | "ordinary-roll"): Response {
+	const cacheRead = mode === "ordinary-roll" ? CACHE_TOKENS : 0;
+	return sseResponse([
+		{
+			type: "message_start",
+			message: {
+				id: "msg_ordinary",
+				usage: usage(cacheRead, CACHE_TOKENS, 0),
+			},
+		},
+		{ type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+		{ type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "ok" } },
+		{ type: "content_block_stop", index: 0 },
+		{
+			type: "message_delta",
+			delta: { stop_reason: "end_turn" },
+			usage: usage(cacheRead, CACHE_TOKENS, 1),
+		},
+		{ type: "message_stop" },
+	]);
+}
+
+function refreshResponse(): Response {
+	return new Response(
+		JSON.stringify({
+			id: "msg_refresh",
+			type: "message",
+			role: "assistant",
+			model: model.id,
+			content: [],
+			stop_reason: "end_turn",
+			usage: usage(CACHE_TOKENS, 0, 0),
+		}),
+		{
+			status: 200,
+			headers: { "Content-Type": "application/json", "request-id": "req_cache_refresh" },
+		},
+	);
+}
+
+function thinkingRefreshResponse(signal: AbortSignal | null | undefined, capture: FetchCapture): Response {
+	const encoder = new TextEncoder();
+	const body = new ReadableStream<Uint8Array>({
+		start(controller) {
+			const events = [
+				{
+					type: "message_start",
+					message: { id: "msg_thinking_refresh", usage: usage(CACHE_TOKENS, 0, 0) },
+				},
+				{
+					type: "content_block_start",
+					index: 0,
+					content_block: { type: "thinking", thinking: "", signature: "" },
+				},
+			];
+			controller.enqueue(
+				encoder.encode(
+					`${events.map(event => `event: ${event.type}\ndata: ${JSON.stringify(event)}`).join("\n\n")}\n\n`,
+				),
+			);
+			const closeOnAbort = () => {
+				capture.thinkingRefreshAborted = true;
+				controller.close();
+			};
+			if (signal?.aborted) closeOnAbort();
+			else signal?.addEventListener("abort", closeOnAbort, { once: true });
+		},
+		cancel() {
+			capture.thinkingRefreshAborted = true;
+		},
+	});
+	return new Response(body, {
+		status: 200,
+		headers: { "Content-Type": "text/event-stream", "request-id": "req_thinking_refresh" },
+	});
+}
+
+function createFetch(modes: ResponseMode[], capture: FetchCapture): FetchImpl {
+	return async (input, init) => {
+		const body: MessageCreateParams = JSON.parse(String(init?.body ?? "{}"));
+		capture.bodies.push(body);
+		const mode = modes[capture.bodies.length - 1];
+		switch (mode) {
+			case "ordinary-write":
+			case "ordinary-roll":
+				return ordinaryResponse(mode);
+			case "refresh-read":
+				return refreshResponse();
+			case "thinking-refresh":
+				return thinkingRefreshResponse(input instanceof Request ? input.signal : init?.signal, capture);
+		}
+	};
+}
+
+interface FinishRequestOptions {
+	anthropicCacheRefresh?: boolean;
+	model?: Model<"anthropic-messages">;
+	sessionId?: string;
+}
+
+async function finishRequest(
+	fetch: FetchImpl,
+	providerSessionState: Map<string, ProviderSessionState>,
+	options: FinishRequestOptions = {},
+): Promise<void> {
+	const requestModel = options.model ?? model;
+	const stream = streamSimple(requestModel, context, {
+		fetch,
+		anthropicCacheRefresh: options.anthropicCacheRefresh ?? true,
+		providerSessionState,
+		sessionId: options.sessionId ?? "cache-refresh-test-session",
+	});
+	for await (const _event of stream) {
+		// Drain the public response before the idle gap begins.
+	}
+	await stream.result();
+}
+
+async function drainUntil(predicate: () => boolean, message: string): Promise<void> {
+	for (let attempt = 0; attempt < 1_000; attempt++) {
+		if (predicate()) return;
+		await Promise.resolve();
+	}
+	throw new Error(message);
+}
+
+async function advanceToRefresh(capture: FetchCapture, expectedRequests: number): Promise<void> {
+	vi.advanceTimersByTime(CACHE_REFRESH_DELAY_MS);
+	await drainUntil(
+		() => capture.bodies.length >= expectedRequests,
+		`Expected ${expectedRequests} Anthropic requests, saw ${capture.bodies.length}`,
+	);
+}
+
+afterEach(() => {
+	for (const states of stateMaps.splice(0)) {
+		for (const state of states.values()) state.close();
+		states.clear();
+	}
+	vi.useRealTimers();
+	vi.restoreAllMocks();
+});
+
+describe("Anthropic prompt-cache refresh", () => {
+	it("replays max_tokens=0 once per interval and stops after three refreshes", async () => {
+		vi.useFakeTimers();
+		const capture: FetchCapture = { bodies: [], thinkingRefreshAborted: false };
+		const fetch = createFetch(["ordinary-write", "refresh-read", "refresh-read", "refresh-read"], capture);
+		const states = createProviderSessionState();
+
+		await finishRequest(fetch, states);
+		for (let requestCount = 2; requestCount <= 4; requestCount++) {
+			await advanceToRefresh(capture, requestCount);
+		}
+		vi.advanceTimersByTime(CACHE_REFRESH_DELAY_MS * 2);
+		await Promise.resolve();
+
+		expect(capture.bodies).toHaveLength(4);
+		for (const refresh of capture.bodies.slice(1)) {
+			expect(refresh.max_tokens).toBe(0);
+			expect(refresh.stream).toBe(false);
+		}
+	});
+
+	it("resets the idle gap when another normal request starts", async () => {
+		vi.useFakeTimers();
+		const capture: FetchCapture = { bodies: [], thinkingRefreshAborted: false };
+		const fetch = createFetch(["ordinary-write", "ordinary-roll", "refresh-read"], capture);
+		const states = createProviderSessionState();
+
+		await finishRequest(fetch, states);
+		vi.advanceTimersByTime(CACHE_REFRESH_DELAY_MS - 1);
+		await finishRequest(fetch, states);
+		vi.advanceTimersByTime(CACHE_REFRESH_DELAY_MS - 1);
+		await Promise.resolve();
+		expect(capture.bodies).toHaveLength(2);
+
+		vi.advanceTimersByTime(1);
+		await drainUntil(() => capture.bodies.length === 3, "Replacement idle timer did not refresh");
+		expect(capture.bodies).toHaveLength(3);
+	});
+
+	it("keeps refresh ownership with the main turn when a side request shares provider state", async () => {
+		vi.useFakeTimers();
+		const capture: FetchCapture = { bodies: [], thinkingRefreshAborted: false };
+		const fetch = createFetch(["ordinary-write", "ordinary-roll", "refresh-read"], capture);
+		const states = createProviderSessionState();
+		const halfInterval = Math.floor(CACHE_REFRESH_DELAY_MS / 2);
+
+		await finishRequest(fetch, states);
+		vi.advanceTimersByTime(halfInterval);
+		await finishRequest(fetch, states, {
+			anthropicCacheRefresh: false,
+			sessionId: "cache-refresh-test-session:side:1",
+		});
+		vi.advanceTimersByTime(CACHE_REFRESH_DELAY_MS - halfInterval);
+		await drainUntil(() => capture.bodies.length === 3, "Main idle timer did not refresh");
+
+		vi.advanceTimersByTime(halfInterval);
+		await Promise.resolve();
+		expect(capture.bodies).toHaveLength(3);
+	});
+
+	it("treats omitted adaptive thinking as active and aborts at the first generated block", async () => {
+		vi.useFakeTimers();
+		const capture: FetchCapture = { bodies: [], thinkingRefreshAborted: false };
+		const fetch = createFetch(["ordinary-write", "thinking-refresh"], capture);
+		const states = createProviderSessionState();
+
+		await finishRequest(fetch, states, { model: thinkingModel });
+		await advanceToRefresh(capture, 2);
+		await drainUntil(() => capture.thinkingRefreshAborted, "Thinking refresh was not aborted");
+
+		expect(capture.bodies[1]?.thinking).toBeUndefined();
+		expect(capture.bodies[1]?.output_config?.effort).toBe("low");
+		expect(capture.bodies[1]?.max_tokens).toBeGreaterThan(0);
+		expect(capture.bodies[1]?.stream).toBe(true);
+		expect(capture.thinkingRefreshAborted).toBe(true);
+	});
+});

packages/ai/test/anthropic-stream-envelope.test.ts

@@ -1735,7 +1735,7 @@ describe("anthropic stream envelope handling", () => {
 		expect(cacheControls[2]).toEqual({ type: "ephemeral" });
 	});
 
-	it("defaults API-key requests to 1h cache TTL where long retention is supported", async () => {
+	it("defaults Anthropic requests to 5m writes and keeps 1h retention opt-in", async () => {
 		type CapturedParams = { messages: Array<{ content: unknown }> };
 		const payloads: CapturedParams[] = [];
 		vi.spyOn(AnthropicMessages.prototype, "create").mockImplementation((params: unknown) => {
@@ -1758,7 +1758,7 @@ describe("anthropic stream envelope handling", () => {
 
 		await drain(model);
 		await drain(proxyModel);
-		await withEnv({ PI_CACHE_RETENTION: "short" }, () => drain(model));
+		await withEnv({ PI_CACHE_RETENTION: "long" }, () => drain(model));
 
 		const cacheControls = payloads.map(payload => {
 			const content = payload.messages.at(-1)?.content;
@@ -1766,13 +1766,8 @@ describe("anthropic stream envelope handling", () => {
 			const lastBlock: { cache_control?: { ttl?: string; type: string } } | undefined = content.at(-1);
 			return lastBlock?.cache_control;
 		});
-		// Agent sessions idle past 5 minutes on background jobs; the canonical
-		// Anthropic API defaults to the 1h breakpoint so resume doesn't cold-miss
-		// the whole prefix.
-		expect(cacheControls[0]).toEqual({ type: "ephemeral", ttl: "1h" });
-		// Endpoints without long-cache support keep the plain 5m breakpoint.
+		expect(cacheControls[0]).toEqual({ type: "ephemeral" });
 		expect(cacheControls[1]).toEqual({ type: "ephemeral" });
-		// PI_CACHE_RETENTION=short opts back out of the 1h default.
-		expect(cacheControls[2]).toEqual({ type: "ephemeral" });
+		expect(cacheControls[2]).toEqual({ type: "ephemeral", ttl: "1h" });
 	});
 });

packages/catalog/CHANGELOG.md

@@ -5,9 +5,11 @@
 ### Added
 
 - Added first-party OpenAI Daybreak Blue, Daybreak Red, and GPT-5.6 Cyber models with their documented API pricing, token limits, tools, and full `off`/`low`/`medium`/`high`/`xhigh`/`max` reasoning controls.
+- Added `calculateUncachedInputCost()` for pricing a prompt against its active context-length tier without prompt caching.
 
 ### Fixed
 
+- Fixed Anthropic cache-write pricing to honor mixed 5-minute and 1-hour TTL usage instead of charging every write at the 5-minute rate.
 - Fixed Ollama Cloud DeepSeek V4 Flash (including dated/preview tags like `deepseek-v4-flash:0731`) exposing the generic `minimal`/`low`/`medium`/`high`/`xhigh` effort ladder without `max`; the `ollama-chat` transport now applies the DeepSeek effort contract (Flash → `low`/`high`/`max`, V4 Pro and older reasoners → `high`/`max`), matching the direct API and every other host ([#8334](https://github.com/can1357/oh-my-pi/issues/8334)).
 - Bounded OpenAI-compatible model discovery with a default request timeout so a stalled provider `/models` endpoint can no longer hang startup indefinitely in `resolveModelDiscoveryFallback` ([#8315](https://github.com/can1357/oh-my-pi/issues/8315)).
 - Fixed Codex-discovered `gpt-daybreak-*` aliases being treated as unknown models, restoring the GPT-5.6 `low`/`medium`/`high`/`xhigh`/`max` effort ladder and its 372K fallback only when the Codex registry omits `context_window`.

packages/catalog/src/models.ts

@@ -42,18 +42,23 @@ export function getBundledModels(provider: GeneratedProvider): Model<Api>[] {
 	const models = getProviderModels(provider);
 	return models ? (Array.from(models.values()) as Model<Api>[]) : [];
 }
-function resolveTokenCost(cost: Model["cost"], usage: Usage): TokenCost {
+function resolveTokenCost(cost: Model["cost"], promptInputTokens: number): TokenCost {
 	const longContext = cost.longContext;
 	if (!longContext) return cost;
-	const orchestration = usage.orchestration;
-	const promptInputTokens =
-		usage.input + usage.cacheRead + usage.cacheWrite + (orchestration?.input ?? 0) + (orchestration?.cacheRead ?? 0);
 	return promptInputTokens > longContext.inputThreshold ? longContext : cost;
 }
 
+/** Price a prompt as fully uncached input under its active context-length tier. */
+export function calculateUncachedInputCost(cost: Model["cost"], promptInputTokens: number): number {
+	const rates = resolveTokenCost(cost, promptInputTokens);
+	return (rates.input / 1_000_000) * promptInputTokens;
+}
+
 export function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage["cost"] {
-	const rates = resolveTokenCost(model.cost, usage);
 	const orchestration = usage.orchestration;
+	const promptInputTokens =
+		usage.input + usage.cacheRead + usage.cacheWrite + (orchestration?.input ?? 0) + (orchestration?.cacheRead ?? 0);
+	const rates = resolveTokenCost(model.cost, promptInputTokens);
 	usage.cost.input = (rates.input / 1000000) * (usage.input + (orchestration?.input ?? 0));
 	usage.cost.output = (rates.output / 1000000) * (usage.output + (orchestration?.output ?? 0));
 	usage.cost.cacheRead = (rates.cacheRead / 1000000) * (usage.cacheRead + (orchestration?.cacheRead ?? 0));
@@ -66,11 +71,11 @@ export function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage
  * Price cache-write tokens, honoring the TTL breakdown when the provider reports one.
  *
  * `rates.cacheWrite` is the 5-minute write rate (Anthropic bills 5m writes at
- * 1.25x base input). When `usage.cttl` is present the write mixes 5m and 1h
- * breakpoints — omp defaults to 1h retention on first-party Anthropic, and 1h writes
- * bill at 2x base input — so each component is priced at its own rate instead of the
- * flat 5m rate. Deriving 1h from `input * 2` (Anthropic's published multiplier) is
- * model-independent and stays correct even for legacy entries whose stored
+ * 1.25x base input). When `usage.cttl` is present the write can mix 5m and 1h
+ * breakpoints, and 1h writes bill at 2x base input, so each component is
+ * priced at its own rate instead of the flat 5m rate. Deriving 1h from
+ * `input * 2` (Anthropic's published multiplier) is model-independent and
+ * stays correct even for legacy entries whose stored
  * `cacheWrite` scalar drifts from 1.25x input. Providers that omit `cttl`
  * (everyone but Anthropic) keep the flat-rate calculation.
  *

packages/coding-agent/CHANGELOG.md

@@ -6,11 +6,13 @@
 
 - Added Astral `ty` as a built-in Python primary LSP server (`ty server`), ordered behind `pyright`/`basedpyright`/`pylsp` so it becomes the primary Python LSP only when the existing servers are unavailable. `ruff` remains the Python linter and coexists alongside `ty` ([#4617](https://github.com/can1357/oh-my-pi/issues/4617)).
 - Added first-party Nix support with reproducible source builds for Linux and macOS on x86-64 and ARM64, a pinned development shell, an overlay, NixOS and Home Manager modules, offline Bun dependencies, and lightweight flake evaluation in CI. Nix-managed installs now direct updates back through Nix instead of replacing store-managed executables.
+- `omp update` and the startup version check now follow an `omp.rename` pointer in the published npm manifest, preparing existing installs for the upcoming npm package rename. Migration is transactional: the renamed agent/natives packages are installed first (npm uses `--force` to take over the `omp` bin), so an install failure leaves the old install untouched; the old-name globals are removed only afterwards, and a broken bin link is restored by re-running the idempotent install before verification decides the outcome.
 
 ### Changed
 
 - `/usage`, `omp usage`, and the status line now show authoritative OpenCode Go quota from the official `GET /zen/go/v1/usage` endpoint — including usage made outside OMP — instead of dollar estimates summed from OMP-observed request costs. The status line renders all three windows (`5h` / `7d` / `mo`), and the per-turn cost recording special case for `opencode-go` sessions is gone along with the "OMP-observed spend only" disclaimer ([#8337](https://github.com/can1357/oh-my-pi/pull/8337) by [@will-bogusz](https://github.com/will-bogusz)).
 - Clarified that the production collab relay source and binaries are not currently published, and documented the source-available local protocol relay ([#8165](https://github.com/can1357/oh-my-pi/issues/8165)).
+- Enabled bounded Anthropic prompt-cache refreshes for the main agent loop while keeping advisor and side-channel requests from taking over the shared refresh timer.
 - Fixed snapcompact compaction shipping its redundant frame archive out of `SessionMaintenance.compact()` on both the manual RPC response (which hard-failed protocol v1 with a transport error after the compaction had already persisted) and the `auto_compaction_end` event payload (which forced the shrink ladder on every unattended pass); the archive is now stripped from both exits while the persisted compaction entry keeps it ([#8168](https://github.com/can1357/oh-my-pi/issues/8168)).
 - Fixed the edit tool showing no diff preview in `apply_patch` mode: the built-in `edit` tool presents on the wire as `apply_patch`, but the renderer-provenance gate did not resolve that alias to its built-in owner, so the edit renderer was skipped ([#8184](https://github.com/can1357/oh-my-pi/issues/8184)).
 

packages/coding-agent/src/cli/stats-cli.ts

@@ -170,6 +170,7 @@ async function printStatsSummary(): Promise<void> {
 	console.log(`  Input Tokens: ${formatNumber(overall.totalInputTokens)}`);
 	console.log(`  Output Tokens: ${formatNumber(overall.totalOutputTokens)}`);
 	console.log(`  Cache Rate: ${formatPercent(overall.cacheRate)}`);
+	console.log(`  Cache Savings: ${formatPercent(overall.cacheSavings)}`);
 	console.log(`  Total Cost: ${formatCost(overall.totalCost)}`);
 	console.log(`  Premium Requests: ${formatNumber(normalizePremiumRequests(overall.totalPremiumRequests ?? 0))}`);
 	console.log(`  Avg Duration: ${overall.avgDuration !== null ? formatDuration(overall.avgDuration) : "-"}`);
@@ -182,7 +183,7 @@ async function printStatsSummary(): Promise<void> {
 		console.log(chalk.bold("\nBy Model:"));
 		for (const m of byModel.slice(0, 10)) {
 			console.log(
-				`  ${m.model}: ${formatNumber(m.totalRequests)} reqs, ${formatCost(m.totalCost)}, ${formatPercent(m.cacheRate)} cache`,
+				`  ${m.model}: ${formatNumber(m.totalRequests)} reqs, ${formatCost(m.totalCost)}, ${formatPercent(m.cacheRate)} cache rate, ${formatPercent(m.cacheSavings)} cache savings`,
 			);
 		}
 	}

packages/coding-agent/src/cli/update-cli.ts

@@ -67,11 +67,27 @@ function currentNativeTag(): string {
 /** Distribution channel advertised by a release's published npm manifest. */
 export type ReleaseDist = "npm" | "binary";
 
-interface ReleaseInfo {
+/** npm package names a release installs: the agent package and its natives companion. */
+export interface ReleasePackages {
+	pkg: string;
+	natives: string;
+}
+
+/** Parsed `omp.rename` pointer: the new agent package name and optional new natives name. */
+export interface ReleaseRename {
+	pkg: string;
+	natives?: string;
+}
+
+const CURRENT_PACKAGES: ReleasePackages = { pkg: PACKAGE, natives: NATIVES_PACKAGE };
+
+export interface ReleaseInfo {
 	tag: string;
 	version: string;
 	/** Parsed `omp.dist` from the registry manifest; undefined when absent. */
 	dist?: ReleaseDist;
+	/** npm names to install, resolved after following any `omp.rename` pointers. */
+	packages: ReleasePackages;
 }
 
 export interface ReleaseBinaryAsset {
@@ -103,6 +119,34 @@ export function resolveReleaseDist(manifest: unknown): ReleaseDist | undefined {
 	return dist === "npm" ? "npm" : "binary";
 }
 
+/**
+ * Parse the `omp.rename` pointer from a published package manifest.
+ *
+ * Forward-compatibility contract for renaming the npm package: the final
+ * version published under an old name is a stub whose manifest carries
+ * `"omp": { "rename": { "package": "<new-agent-pkg>", "natives": "<new-natives-pkg>" }, "dist": "binary" }`.
+ * Updaters that understand `rename` follow the pointer and resolve the
+ * release from the renamed package instead ({@link getLatestRelease});
+ * older deployed updaters ignore it and take the `dist: "binary"` escape
+ * hatch, replacing the install with the GitHub release binary rather than
+ * installing the stub via bun/npm.
+ *
+ * The renamed package's own manifest MUST declare `"dist": "npm"` (so
+ * package-manager installs stay package-managed across a major bump) and
+ * MUST continue the old version line (a version reset would compare as
+ * "already up to date" against the running build).
+ */
+export function resolveReleaseRename(manifest: unknown): ReleaseRename | undefined {
+	if (!isRecord(manifest) || !isRecord(manifest.omp)) return undefined;
+	const rename = manifest.omp.rename;
+	if (!isRecord(rename) || typeof rename.package !== "string" || rename.package.length === 0) return undefined;
+	const natives = rename.natives;
+	return {
+		pkg: rename.package,
+		natives: typeof natives === "string" && natives.length > 0 ? natives : undefined,
+	};
+}
+
 function majorVersion(version: string): number {
 	const major = Number.parseInt(version, 10);
 	return Number.isNaN(major) ? 0 : major;
@@ -529,36 +573,63 @@ async function resolveUpdateTarget(options: { allowPackageManagers: boolean }):
 	throw new Error(`Could not resolve ${APP_NAME} binary path in PATH`);
 }
 
-/**
- * Get the latest release info from the npm registry.
- * Uses npm instead of GitHub API to avoid unauthenticated rate limiting.
- */
-async function getLatestRelease(): Promise<ReleaseInfo> {
+/** Bound on `omp.rename` hops so a broken pointer chain cannot loop forever. */
+const MAX_RENAME_HOPS = 3;
+
+async function fetchLatestManifest(
+	pkg: string,
+	timeoutMs: number,
+): Promise<{ version: string; manifest: Record<string, unknown> }> {
 	let response: Response;
 	try {
-		response = await fetch(`${NPM_REGISTRY}${PACKAGE}/latest`, {
-			signal: withTimeoutSignal(RELEASE_METADATA_TIMEOUT_MS),
+		response = await fetch(`${NPM_REGISTRY}${pkg}/latest`, {
+			signal: withTimeoutSignal(timeoutMs),
 		});
 	} catch (err) {
 		if (isTimeoutError(err)) {
-			throw new Error("Timed out fetching release info after 30s", { cause: err });
+			throw new Error(`Timed out fetching release info for ${pkg} after ${Math.round(timeoutMs / 1000)}s`, {
+				cause: err,
+			});
 		}
 		throw err;
 	}
 	if (!response.ok) {
-		throw new Error(`Failed to fetch release info: ${response.statusText}`);
+		throw new Error(`Failed to fetch release info for ${pkg}: ${response.statusText}`);
 	}
 
 	const data: unknown = await response.json();
 	if (!isRecord(data) || typeof data.version !== "string") {
-		throw new Error("Malformed npm registry response: missing version");
+		throw new Error(`Malformed npm registry response for ${pkg}: missing version`);
+	}
+	return { version: data.version, manifest: data };
+}
+
+/**
+ * Get the latest release info from the npm registry, following `omp.rename`
+ * pointers ({@link resolveReleaseRename}) when the package has moved to a new
+ * npm name. Version, dist, and install names all come from the final manifest
+ * in the chain. Uses npm instead of GitHub API to avoid unauthenticated rate
+ * limiting.
+ */
+export async function getLatestRelease(options: { timeoutMs?: number } = {}): Promise<ReleaseInfo> {
+	const timeoutMs = options.timeoutMs ?? RELEASE_METADATA_TIMEOUT_MS;
+	const packages: ReleasePackages = { ...CURRENT_PACKAGES };
+	const visited = new Set([packages.pkg]);
+	let latest = await fetchLatestManifest(packages.pkg, timeoutMs);
+	for (let hop = 0; hop < MAX_RENAME_HOPS; hop++) {
+		const rename = resolveReleaseRename(latest.manifest);
+		if (!rename || visited.has(rename.pkg)) break;
+		visited.add(rename.pkg);
+		packages.pkg = rename.pkg;
+		if (rename.natives) packages.natives = rename.natives;
+		latest = await fetchLatestManifest(packages.pkg, timeoutMs);
 	}
-	const version = data.version;
 
 	return {
-		tag: `v${version}`,
-		version,
-		dist: resolveReleaseDist(data),
+		tag: `v${latest.version}`,
+		version: latest.version,
+		dist: resolveReleaseDist(latest.manifest),
+		packages,
 	};
 }
 
@@ -994,10 +1065,14 @@ export async function replaceBinaryForUpdate(options: BinaryReplacementOptions):
 	}
 }
 
-function buildVersionedPackageInstallArgs(expectedVersion: string, nativeTag: string): string[] {
-	const args = [`${PACKAGE}@${expectedVersion}`, `${NATIVES_PACKAGE}@${expectedVersion}`];
+function buildVersionedPackageInstallArgs(
+	expectedVersion: string,
+	nativeTag: string,
+	packages: ReleasePackages,
+): string[] {
+	const args = [`${packages.pkg}@${expectedVersion}`, `${packages.natives}@${expectedVersion}`];
 	if (SUPPORTED_NATIVE_TAGS.has(nativeTag)) {
-		args.push(`${NATIVES_PACKAGE}-${nativeTag}@${expectedVersion}`);
+		args.push(`${packages.natives}-${nativeTag}@${expectedVersion}`);
 	}
 	return args;
 }
@@ -1032,25 +1107,41 @@ function buildVersionedPackageInstallArgs(expectedVersion: string, nativeTag: st
  * the original "no matching version" message instead of `EBADPLATFORM`.
  * See #1824.
  */
-export function buildBunInstallArgs(expectedVersion: string, nativeTag: string = currentNativeTag()): string[] {
+export function buildBunInstallArgs(
+	expectedVersion: string,
+	nativeTag: string = currentNativeTag(),
+	packages: ReleasePackages = CURRENT_PACKAGES,
+): string[] {
 	return [
 		"install",
 		"-g",
 		"--no-cache",
 		`--registry=${NPM_REGISTRY}`,
-		...buildVersionedPackageInstallArgs(expectedVersion, nativeTag),
+		...buildVersionedPackageInstallArgs(expectedVersion, nativeTag, packages),
 	];
 }
 
-/** Build the npm argv used to update npm-managed global installs. */
-export function buildNpmInstallArgs(expectedVersion: string, nativeTag: string = currentNativeTag()): string[] {
-	const args = [
+/**
+ * Build the npm argv used to update npm-managed global installs.
+ *
+ * `force` is set only for rename migrations: npm refuses to write the `omp`
+ * bin while the old package still owns it (`EEXIST`), and the migration
+ * installs the new package BEFORE removing the old one so a failed install
+ * never leaves the user without a working `omp`.
+ */
+export function buildNpmInstallArgs(
+	expectedVersion: string,
+	nativeTag: string = currentNativeTag(),
+	packages: ReleasePackages = CURRENT_PACKAGES,
+	flags: { force?: boolean } = {},
+): string[] {
+	return [
 		"install",
 		"-g",
+		...(flags.force ? ["--force"] : []),
 		`--registry=${NPM_REGISTRY}`,
-		...buildVersionedPackageInstallArgs(expectedVersion, nativeTag),
+		...buildVersionedPackageInstallArgs(expectedVersion, nativeTag, packages),
 	];
-	return args;
 }
 
 export function buildHomebrewUpdateArgs(force: boolean): string[] {
@@ -1065,18 +1156,123 @@ export function buildMiseForceInstallArgs(expectedVersion: string): string[] {
 	return ["install", "--force", `${MISE_TOOL}@${expectedVersion}`];
 }
 
+/**
+ * Old-name globals a rename migration removes after the new install exists:
+ * the set difference between the old install's top-level globals
+ * ({@link buildVersionedPackageInstallArgs} installs the agent, natives core,
+ * and platform leaf explicitly) and the resolved install's. An agent-only
+ * rename keeps the natives names, and removing them would strip the addon
+ * the new install just pinned.
+ */
+export function buildRenameCleanupPackages(
+	packages: ReleasePackages,
+	nativeTag: string = currentNativeTag(),
+): string[] {
+	const old = [PACKAGE, NATIVES_PACKAGE];
+	if (SUPPORTED_NATIVE_TAGS.has(nativeTag)) {
+		old.push(`${NATIVES_PACKAGE}-${nativeTag}`);
+	}
+	const newLeaf = `${packages.natives}-${nativeTag}`;
+	return old.filter(name => name !== packages.pkg && name !== packages.natives && name !== newLeaf);
+}
+
+/** Injectable shell steps for {@link migrateRenamedInstall}; commands return process exit codes. */
+export interface RenameMigrationSteps {
+	/** Globally install the new package names. MUST be idempotent: re-running re-links the `omp` bin. */
+	install(): Promise<number>;
+	/** Remove the old-name globals. */
+	removeOld(): Promise<number>;
+	/** Check the PATH-resolved `omp` against the expected version. */
+	verify(): Promise<InstalledVersionVerification>;
+}
+
+/** Production {@link RenameMigrationSteps}: bun/npm global installs plus PATH verification. */
+function packageManagerMigrationSteps(manager: "bun" | "npm", release: ReleaseInfo): RenameMigrationSteps {
+	const nativeTag = currentNativeTag();
+	return {
+		async install() {
+			if (manager === "bun") {
+				const args = buildBunInstallArgs(release.version, nativeTag, release.packages);
+				return (await $`bun ${args}`.nothrow()).exitCode;
+			}
+			const args = buildNpmInstallArgs(release.version, nativeTag, release.packages, { force: true });
+			return (await $`npm ${args}`.nothrow()).exitCode;
+		},
+		async removeOld() {
+			// One invocation per package: a single batched remove fails wholesale
+			// when any name is absent (e.g. the platform leaf on an old install),
+			// which would skip the agent package that actually owns the bin.
+			let agentExit = 0;
+			for (const pkg of buildRenameCleanupPackages(release.packages, nativeTag)) {
+				const result =
+					manager === "bun"
+						? await $`bun remove -g ${pkg}`.quiet().nothrow()
+						: await $`npm uninstall -g ${pkg}`.quiet().nothrow();
+				if (pkg === PACKAGE) agentExit = result.exitCode;
+			}
+			return agentExit;
+		},
+		verify: () => verifyInstalledVersion(release.version),
+	};
+}
+
+/**
+ * Migrate a package-manager install across an `omp.rename` hop without a
+ * window where no working `omp` exists:
+ *
+ * 1. Install the new package FIRST. Nothing has been removed yet, so a
+ *    failure here leaves the old install fully functional.
+ * 2. Remove the old-name globals. Failure is non-fatal: a stale package
+ *    wastes disk, but the bin already points at the new install.
+ * 3. Verify the PATH-resolved `omp`. If the removal deleted the shared bin
+ *    link (manager-dependent), re-run the idempotent install to restore it
+ *    and verify again; only a repeated failure aborts, with a recovery hint.
+ */
+export async function migrateRenamedInstall(release: ReleaseInfo, steps: RenameMigrationSteps): Promise<void> {
+	console.log(chalk.dim(`npm package renamed to ${release.packages.pkg}; migrating this install.`));
+	const installExit = await steps.install();
+	if (installExit !== 0) {
+		throw new Error(
+			`install of ${release.packages.pkg} failed with exit code ${installExit}; the existing install was left untouched`,
+		);
+	}
+
+	const removeExit = await steps.removeOld();
+	if (removeExit !== 0) {
+		console.log(chalk.yellow(`Warning: could not remove the old ${PACKAGE} package; remove it manually later.`));
+	}
+
+	let verification = await steps.verify();
+	if (!verification.ok) {
+		// Removing the old package may have taken the shared bin link with it;
+		// reinstalling the new package restores the link.
+		if ((await steps.install()) === 0) {
+			verification = await steps.verify();
+		}
+	}
+	if (!verification.ok) {
+		throw new Error(
+			`${formatVerificationFailure(verification, release.version)}; reinstall with: curl -fsSL https://omp.sh/install | sh`,
+		);
+	}
+	printVerifiedVersion(release.version);
+}
+
 /**
  * Update via package manager.
  */
-async function updateViaBun(expectedVersion: string): Promise<void> {
+async function updateViaBun(release: ReleaseInfo): Promise<void> {
 	console.log(chalk.dim("Updating via bun..."));
-	const args = buildBunInstallArgs(expectedVersion);
+	if (release.packages.pkg !== PACKAGE) {
+		await migrateRenamedInstall(release, packageManagerMigrationSteps("bun", release));
+	} else {
+		const args = buildBunInstallArgs(release.version, currentNativeTag(), release.packages);
 		const result = await $`bun ${args}`.nothrow();
 		if (result.exitCode !== 0) {
 			throw new Error(`bun install failed with exit code ${result.exitCode}`);
 		}
-
-	await printVerification(expectedVersion);
+		await printVerification(release.version);
+	}
 	try {
 		const pruneResult = await pruneBunCacheAfterGlobalInstall();
 		if (pruneResult && pruneResult.removedEntries > 0) {
@@ -1087,15 +1283,19 @@ async function updateViaBun(expectedVersion: string): Promise<void> {
 	}
 }
 
-async function updateViaNpm(expectedVersion: string): Promise<void> {
+async function updateViaNpm(release: ReleaseInfo): Promise<void> {
 	console.log(chalk.dim("Updating via npm..."));
-	const args = buildNpmInstallArgs(expectedVersion);
+	if (release.packages.pkg !== PACKAGE) {
+		await migrateRenamedInstall(release, packageManagerMigrationSteps("npm", release));
+		return;
+	}
+	const args = buildNpmInstallArgs(release.version, currentNativeTag(), release.packages);
 	const result = await $`npm ${args}`.nothrow();
 	if (result.exitCode !== 0) {
 		throw new Error(`npm install failed with exit code ${result.exitCode}`);
 	}
 
-	await printVerification(expectedVersion);
+	await printVerification(release.version);
 }
 
 async function updateViaHomebrew(expectedVersion: string, force: boolean): Promise<void> {
@@ -1345,6 +1545,9 @@ export async function runUpdateCommand(opts: { force: boolean; check: boolean })
 	} else {
 		console.log(chalk.yellow(`Forcing reinstall of ${release.version}`));
 	}
+	if (release.packages.pkg !== PACKAGE) {
+		console.log(chalk.cyan(`The npm package moved to ${release.packages.pkg}; updating migrates this install.`));
+	}
 
 	if (opts.check) {
 		// Just check, don't install
@@ -1379,9 +1582,9 @@ export async function runUpdateCommand(opts: { force: boolean; check: boolean })
 					),
 				);
 			} else if (target.method === "bun") {
-				await updateViaBun(release.version);
+				await updateViaBun(release);
 			} else {
-				await updateViaNpm(release.version);
+				await updateViaNpm(release);
 			}
 		} else {
 			if (forceBinary && target.replacesSymlink) {

packages/coding-agent/src/main.ts

@@ -29,6 +29,7 @@ import { processFileArguments } from "./cli/file-processor";
 import { buildInitialMessage } from "./cli/initial-message";
 import { selectSession } from "./cli/session-picker";
 import { applyStartupCwd } from "./cli/startup-cwd";
+import { getLatestRelease } from "./cli/update-cli";
 import { findConfigFile } from "./config";
 import { ModelRegistry } from "./config/model-registry";
 import {
@@ -95,7 +96,6 @@ import { concreteThinkingLevel, parseConfiguredThinkingLevel } from "./thinking"
 import type { LspStartupServerInfo } from "./tools";
 import { getChangelogPath, resolveStartupChangelogForDisplay, type StartupChangelogSelection } from "./utils/changelog";
 import { EventBus } from "./utils/event-bus";
-import { withTimeoutSignal } from "./utils/fetch-timeout";
 
 type RunAcpMode = (createSession: AcpSessionFactory) => Promise<never>;
 type RunPrintMode = (session: AgentSession, options: PrintModeOptions) => Promise<void>;
@@ -115,19 +115,8 @@ async function checkForNewVersion(currentVersion: string): Promise<string | unde
 		return;
 	}
 	try {
-		const response = await fetch("https://registry.npmjs.org/@oh-my-pi/pi-coding-agent/latest", {
-			signal: withTimeoutSignal(5_000),
-		});
-		if (!response.ok) return undefined;
-
-		const data = (await response.json()) as { version?: string };
-		const latestVersion = data.version;
-
-		if (latestVersion && Bun.semver.order(latestVersion, currentVersion) > 0) {
-			return latestVersion;
-		}
-
-		return undefined;
+		const release = await getLatestRelease({ timeoutMs: 5_000 });
+		return Bun.semver.order(release.version, currentVersion) > 0 ? release.version : undefined;
 	} catch {
 		return undefined;
 	}

packages/coding-agent/src/sdk.ts

@@ -3247,6 +3247,7 @@ async function createAgentSessionScoped(options: CreateAgentSessionOptions): Pro
 					supportsExternalThinking(streamModel);
 				return settingsAwareStreamFn(streamModel, context, {
 					...streamOptions,
+					anthropicCacheRefresh: true,
 					forceReasoningOff: externalThinking || streamOptions?.forceReasoningOff,
 				});
 			},

packages/coding-agent/src/web/search/providers/anthropic.ts

@@ -130,7 +130,6 @@ function buildSystemBlocks(
 	return buildAnthropicSystemBlocks(systemPrompt ? [systemPrompt] : undefined, {
 		includeClaudeCodeInstruction: includeClaudeCode,
 		extraInstructions,
-		cacheControl: { type: "ephemeral" },
 	});
 }
 

packages/coding-agent/test/cli/update-rename-migration.integration.test.ts

@@ -0,0 +1,155 @@
+/**
+ * Real package-manager seam for `omp.rename` migrations.
+ *
+ * The unit tests in test/update-cli.test.ts prove the orchestration order of
+ * migrateRenamedInstall with injected steps; these fixtures prove the two
+ * empirical assumptions that orchestration stands on, against the actual
+ * package managers in isolated temp prefixes:
+ *
+ * - npm refuses to overwrite a bin owned by another package (EEXIST) and
+ *   `--force` takes ownership of it — and `npm uninstall -g <old>` deletes
+ *   the shared bin even though it points at the new package, so the repair
+ *   reinstall inside migrateRenamedInstall is the NORMAL npm path, not an
+ *   edge case.
+ * - bun clobbers the bin on install without force, and `bun remove -g <old>`
+ *   re-links the bin to the surviving package.
+ *
+ * Each scenario runs the full install-new/remove-old/verify transaction and
+ * asserts the resulting launcher executes the NEW version.
+ */
+import { afterEach, beforeAll, describe, expect, it, vi } from "bun:test";
+import * as fs from "node:fs/promises";
+import * as os from "node:os";
+import * as path from "node:path";
+import { $which, removeWithRetries } from "@oh-my-pi/pi-utils";
+import { $ } from "bun";
+import {
+	type InstalledVersionVerification,
+	migrateRenamedInstall,
+	type ReleaseInfo,
+	type RenameMigrationSteps,
+} from "../../src/cli/update-cli";
+import { initTheme } from "../../src/modes/theme/theme";
+
+const OLD_PKG = "omp-rename-fixture-old";
+const NEW_PKG = "omp-rename-fixture-new";
+const OLD_VERSION = "1.0.0";
+const NEW_VERSION = "2.0.0";
+
+const tempDirs: string[] = [];
+
+// printVerifiedVersion renders theme glyphs; the update command initializes
+// the theme before calling into update-cli, so the tests must too.
+beforeAll(async () => {
+	await initTheme();
+});
+
+afterEach(async () => {
+	vi.restoreAllMocks();
+	await Promise.all(tempDirs.splice(0).map(dir => removeWithRetries(dir)));
+});
+
+/** Two installable packages that both expose an `omp` bin, plus an empty prefix. */
+async function makeFixtures(): Promise<{ root: string; oldDir: string; newDir: string }> {
+	const root = await fs.mkdtemp(path.join(os.tmpdir(), "omp-rename-itest-"));
+	tempDirs.push(root);
+	const mkpkg = async (name: string, version: string): Promise<string> => {
+		const dir = path.join(root, name);
+		await Bun.write(path.join(dir, "package.json"), JSON.stringify({ name, version, bin: { omp: "cli.js" } }));
+		const cli = path.join(dir, "cli.js");
+		await Bun.write(cli, `#!/usr/bin/env bun\nconsole.log("omp/${version}");\n`);
+		await fs.chmod(cli, 0o755);
+		return dir;
+	};
+	return { root, oldDir: await mkpkg(OLD_PKG, OLD_VERSION), newDir: await mkpkg(NEW_PKG, NEW_VERSION) };
+}
+
+/** Run the installed launcher and parse its reported version, mirroring verifyBinaryAtPath. */
+async function verifyLauncher(binDir: string, expectedVersion: string): Promise<InstalledVersionVerification> {
+	const launcher = path.join(binDir, "omp");
+	const result = await $`${launcher}`.quiet().nothrow();
+	if (result.exitCode !== 0) return { ok: false, path: launcher };
+	const actual = result.text().match(/\/(\d+\.\d+\.\d+)/)?.[1];
+	return { ok: actual === expectedVersion, actual, path: launcher };
+}
+
+const RELEASE: ReleaseInfo = {
+	tag: `v${NEW_VERSION}`,
+	version: NEW_VERSION,
+	packages: { pkg: NEW_PKG, natives: "@oh-my-pi/pi-natives" },
+};
+
+describe.skipIf(process.platform === "win32" || !$which("npm"))("rename migration over real npm", () => {
+	it("takes bin ownership with --force, survives the uninstall deleting the bin, and lands on the new version", async () => {
+		vi.spyOn(console, "log").mockImplementation(() => {});
+		const { root, oldDir, newDir } = await makeFixtures();
+		const prefix = path.join(root, "npm-prefix");
+		const binDir = path.join(prefix, "bin");
+		const env = {
+			...process.env,
+			npm_config_cache: path.join(root, "npm-cache"),
+			npm_config_update_notifier: "false",
+			npm_config_fund: "false",
+			npm_config_audit: "false",
+		};
+
+		const seed = await $`npm install -g --prefix ${prefix} ${oldDir}`.env(env).quiet().nothrow();
+		expect(seed.exitCode).toBe(0);
+		expect(await verifyLauncher(binDir, OLD_VERSION)).toMatchObject({ ok: true, actual: OLD_VERSION });
+
+		// The load-bearing precondition for --force: while the old package owns
+		// the bin, a plain install of the new package fails instead of clobbering.
+		const plain = await $`npm install -g --prefix ${prefix} ${newDir}`.env(env).quiet().nothrow();
+		expect(plain.exitCode).not.toBe(0);
+		expect(await verifyLauncher(binDir, OLD_VERSION)).toMatchObject({ ok: true, actual: OLD_VERSION });
+
+		const steps: RenameMigrationSteps = {
+			async install() {
+				return (await $`npm install -g --force --prefix ${prefix} ${newDir}`.env(env).quiet().nothrow()).exitCode;
+			},
+			async removeOld() {
+				return (await $`npm uninstall -g --prefix ${prefix} ${OLD_PKG}`.env(env).quiet().nothrow()).exitCode;
+			},
+			verify: () => verifyLauncher(binDir, NEW_VERSION),
+		};
+		await migrateRenamedInstall(RELEASE, steps);
+
+		expect(await verifyLauncher(binDir, NEW_VERSION)).toMatchObject({ ok: true, actual: NEW_VERSION });
+		const globalPackages = await fs.readdir(path.join(prefix, "lib", "node_modules"));
+		expect(globalPackages).toContain(NEW_PKG);
+		expect(globalPackages).not.toContain(OLD_PKG);
+	}, 120_000);
+});
+
+describe.skipIf(process.platform === "win32")("rename migration over real bun", () => {
+	it("clobbers the old bin on install, survives removing the old package, and lands on the new version", async () => {
+		vi.spyOn(console, "log").mockImplementation(() => {});
+		const { root, oldDir, newDir } = await makeFixtures();
+		const binDir = path.join(root, "bun-bin");
+		await fs.mkdir(binDir, { recursive: true });
+		const env = {
+			...process.env,
+			BUN_INSTALL_GLOBAL_DIR: path.join(root, "bun-global"),
+			BUN_INSTALL_BIN: binDir,
+		};
+
+		const seed = await $`bun add -g file:${oldDir}`.env(env).quiet().nothrow();
+		expect(seed.exitCode).toBe(0);
+		expect(await verifyLauncher(binDir, OLD_VERSION)).toMatchObject({ ok: true, actual: OLD_VERSION });
+
+		const steps: RenameMigrationSteps = {
+			async install() {
+				return (await $`bun add -g file:${newDir}`.env(env).quiet().nothrow()).exitCode;
+			},
+			async removeOld() {
+				return (await $`bun remove -g ${OLD_PKG}`.env(env).quiet().nothrow()).exitCode;
+			},
+			verify: () => verifyLauncher(binDir, NEW_VERSION),
+		};
+		await migrateRenamedInstall(RELEASE, steps);
+
+		expect(await verifyLauncher(binDir, NEW_VERSION)).toMatchObject({ ok: true, actual: NEW_VERSION });
+		const globalManifest = await Bun.file(path.join(root, "bun-global", "package.json")).json();
+		expect(Object.keys(globalManifest.dependencies ?? {})).toEqual([NEW_PKG]);
+	}, 120_000);
+});

packages/coding-agent/test/update-cli.test.ts

@@ -291,13 +291,23 @@ describe("update-cli npm rename contract", () => {
 		expect(buildNpmInstallArgs("16.3.15", "win32-x64")).not.toContain("--force");
 	});
 
-	it("removes the old agent package and its natives companions, including the supported platform leaf", () => {
-		expect(buildRenameCleanupPackages("darwin-arm64")).toEqual([
+	it("removes the old agent package and its natives companions when both names moved", () => {
+		const packages = { pkg: "@new/omp", natives: "@new/natives" };
+		expect(buildRenameCleanupPackages(packages, "darwin-arm64")).toEqual([
 			"@oh-my-pi/pi-coding-agent",
 			"@oh-my-pi/pi-natives",
 			"@oh-my-pi/pi-natives-darwin-arm64",
 		]);
-		expect(buildRenameCleanupPackages("linux-arm")).toEqual(["@oh-my-pi/pi-coding-agent", "@oh-my-pi/pi-natives"]);
+		expect(buildRenameCleanupPackages(packages, "linux-arm")).toEqual([
+			"@oh-my-pi/pi-coding-agent",
+			"@oh-my-pi/pi-natives",
+		]);
+	});
+
+	it("keeps the natives packages on an agent-only rename so cleanup cannot strip the addon the new install pinned", () => {
+		const packages = { pkg: "@new/omp", natives: "@oh-my-pi/pi-natives" };
+		expect(buildRenameCleanupPackages(packages, "darwin-arm64")).toEqual(["@oh-my-pi/pi-coding-agent"]);
+		expect(buildRenameCleanupPackages(packages, "linux-arm")).toEqual(["@oh-my-pi/pi-coding-agent"]);
 	});
 });
 

packages/natives/scripts/build-bindings.ts

@@ -11,7 +11,7 @@ import * as fs from "node:fs/promises";
 import { createRequire } from "node:module";
 import * as path from "node:path";
 import { $ } from "bun";
-import { detectHostAvx2Support } from "../../../scripts/host-detect";
+import { detectHostAvx2Support, resolveLocalHostAddon } from "../../../scripts/host-detect";
 import { generateEnumExports } from "./gen-enums";
 
 // pcre2-sys prefers a system libpcre2 when pkg-config finds one. Keep the
@@ -61,10 +61,12 @@ const rustDir = path.join(repoRoot, "crates/pi-natives");
 const nativeDir = path.join(import.meta.dir, "../native");
 const packageJsonPath = path.join(import.meta.dir, "../package.json");
 
-type X64Variant = "modern" | "baseline";
-
-const effectiveVariant: X64Variant | null =
-	process.arch === "x64" ? (detectHostAvx2Support() ? "modern" : "baseline") : null;
+const localAddon = resolveLocalHostAddon({
+	platform: process.platform,
+	arch: process.arch,
+	avx2: detectHostAvx2Support(),
+});
+const effectiveVariant = localAddon.x64Variant;
 const variantSuffix = effectiveVariant ? `-${effectiveVariant}` : "";
 
 // Pin Rust target-cpu so x64 baseline/modern variants get a reproducible ISA floor
@@ -166,7 +168,7 @@ async function installGeneratedBindings(outputDir: string): Promise<void> {
 	}
 }
 
-const canonicalAddonFilename = `pi_natives.${process.platform}-${process.arch}${variantSuffix}.node`;
+const canonicalAddonFilename = localAddon.filename;
 const canonicalAddonPath = path.join(nativeDir, canonicalAddonFilename);
 
 console.log(`Building pi-natives bindings for ${process.platform}-${process.arch}${variantSuffix} (local)…`);

packages/stats/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added cost-weighted `cacheSavings` alongside `cacheRate`, including cache-read discounts and 5-minute/1-hour write premiums against the equivalent uncached prompt cost.
+
 ### Fixed
 
 - Made the embedded dashboard archive byte-reproducible by sorting entries and zeroing tar and gzip timestamps before it is compiled into OMP.

packages/stats/README.md

@@ -15,6 +15,7 @@ Local observability dashboard for AI usage statistics.
 |--------|-------------|
 | Tokens/s | `output_tokens / (duration / 1000)` |
 | Cache Rate | `cache_read / (input + cache_read) * 100` |
+| Cache Savings | `(uncached prompt cost - actual prompt cost) / uncached prompt cost * 100` |
 | Error Rate | `count(stopReason=error) / total_calls * 100` |
 | Total Cost | Sum of `usage.cost.total` |
 | Avg Latency | Mean of `duration` |
@@ -71,7 +72,7 @@ console.log(stats.byModel[0].avgTokensPerSecond);
 
 The web dashboard provides:
 
-- Overall metrics cards (requests, cost, cache rate, error rate, duration, tokens/s)
+- Overall metrics cards (requests, cost, cache rate, cache savings, error rate, duration, tokens/s)
 - Time series chart showing requests and errors over time
 - Per-model breakdown table
 - Per-folder breakdown table

packages/stats/src/client/routes/ModelsRoute.tsx

@@ -322,7 +322,7 @@ function ModelsTable({
 								<div className="grid gap-4" style={{ gridTemplateColumns: "200px 1fr" }}>
 									<div className="space-y-4 text-sm">
 										<div>
-											<div className="text-[var(--text-primary)] font-medium mb-2">Quality</div>
+											<div className="text-[var(--text-primary)] font-medium mb-2">Efficiency</div>
 											<div className="space-y-1 text-[var(--text-secondary)]">
 												<div className="flex items-center justify-between">
 													<span>Error rate</span>
@@ -336,8 +336,18 @@ function ModelsTable({
 												</div>
 												<div className="flex items-center justify-between">
 													<span>Cache rate</span>
-													<span className="text-[var(--accent-cyan)]">
-														{(model.cacheRate * 100).toFixed(1)}%
+													<span className="font-mono">{(model.cacheRate * 100).toFixed(1)}%</span>
+												</div>
+												<div className="flex items-center justify-between">
+													<span>Cache savings</span>
+													<span
+														className={
+															model.cacheSavings < 0
+																? "text-[var(--accent-red)]"
+																: "text-[var(--accent-green)]"
+														}
+													>
+														{(model.cacheSavings * 100).toFixed(1)}%
 													</span>
 												</div>
 											</div>

packages/stats/src/client/routes/ProjectsRoute.tsx

@@ -87,8 +87,16 @@ export function ProjectsRoute({ active, range, refreshTrigger }: ProjectsRoutePr
 				key: "cacheRate",
 				header: "Cache Rate",
 				numeric: true,
+				render: (item: FolderRowView) => <span className="font-mono">{formatPercent(item.cacheRate)}</span>,
+			},
+			{
+				key: "cacheSavings",
+				header: "Cache Savings",
+				numeric: true,
 				render: (item: FolderRowView) => (
-					<span className="stats-text-success font-medium">{formatPercent(item.cacheRate)}</span>
+					<span className={`${item.cacheSavings < 0 ? "stats-text-danger" : "stats-text-success"} font-medium`}>
+						{formatPercent(item.cacheSavings)}
+					</span>
 				),
 			},
 			{
@@ -129,9 +137,13 @@ export function ProjectsRoute({ active, range, refreshTrigger }: ProjectsRoutePr
 					<div className="stats-mobile-card-value font-mono">{formatCost(item.totalCost)}</div>
 				</div>
 				<div>
-					<div className="stats-mobile-card-label">Cache</div>
+					<div className="stats-mobile-card-label">Cache Rate</div>
 					<div className="stats-mobile-card-value">{formatPercent(item.cacheRate)}</div>
 				</div>
+				<div>
+					<div className="stats-mobile-card-label">Cache Savings</div>
+					<div className="stats-mobile-card-value">{formatPercent(item.cacheSavings)}</div>
+				</div>
 				<div>
 					<div className="stats-mobile-card-label">Duration</div>
 					<div className="stats-mobile-card-value">{formatDurationMs(item.avgDuration)}</div>

packages/stats/src/client/styles.css

@@ -563,7 +563,7 @@
 
 .stats-metric-primary-grid {
   display: grid;
-  grid-template-columns: repeat(4, 1fr);
+  grid-template-columns: repeat(5, minmax(0, 1fr));
   gap: 16px;
 }
 

packages/stats/src/client/ui/MetricCluster.tsx

@@ -29,7 +29,17 @@ export function MetricCluster({ stats }: MetricClusterProps) {
 					<div className="stats-metric-label">Requests</div>
 					<div className="stats-metric-value">{formatInteger(stats.totalRequests)}</div>
 				</div>
-				<div className="stats-metric-card primary">
+				<div
+					className="stats-metric-card primary"
+					title="Prompt-input cost saved versus billing the same tokens uncached; cache writes can make this negative"
+				>
+					<div className="stats-metric-label">Cache Savings</div>
+					<div className="stats-metric-value">{formatPercent(stats.cacheSavings)}</div>
+				</div>
+				<div
+					className="stats-metric-card primary"
+					title="Prompt input served from cache: cache reads / (uncached input + cache reads)"
+				>
 					<div className="stats-metric-label">Cache Rate</div>
 					<div className="stats-metric-value">{formatPercent(stats.cacheRate)}</div>
 				</div>

packages/stats/src/db.ts

@@ -2,7 +2,7 @@ import { Database } from "bun:sqlite";
 import * as fs from "node:fs/promises";
 import type { Usage } from "@oh-my-pi/pi-ai";
 import type { GeneratedProvider } from "@oh-my-pi/pi-catalog/models";
-import { getBundledModel } from "@oh-my-pi/pi-catalog/models";
+import { calculateUncachedInputCost, getBundledModel } from "@oh-my-pi/pi-catalog/models";
 import { getConfigRootDir, getStatsDbPath } from "@oh-my-pi/pi-utils";
 import { classifyAgentType } from "./parser";
 import type {
@@ -33,7 +33,7 @@ import type {
 
 type ModelCost = { input: number; output: number; cacheRead: number; cacheWrite: number };
 type UsageCost = Usage["cost"];
-type CostTokens = Pick<Usage, "input" | "output" | "cacheRead" | "cacheWrite">;
+type CostTokens = Pick<Usage, "input" | "output" | "cacheRead" | "cacheWrite" | "orchestration">;
 
 const ZERO_USAGE_COST: UsageCost = {
 	input: 0,
@@ -53,6 +53,42 @@ interface CostBackfillRow {
 	cache_write_tokens: number;
 }
 
+interface NoCacheInputCostBackfillRow {
+	id: number;
+	provider: string;
+	model: string;
+	input_tokens: number;
+	cache_read_tokens: number;
+	cache_write_tokens: number;
+}
+
+interface AggregatedStatsRow {
+	total_requests: number;
+	failed_requests: number | null;
+	total_input_tokens: number | null;
+	total_output_tokens: number | null;
+	total_cache_read_tokens: number | null;
+	total_cache_write_tokens: number | null;
+	total_premium_requests: number | null;
+	total_cost: number | null;
+	total_cached_prompt_cost: number | null;
+	total_no_cache_input_cost: number | null;
+	avg_duration: number | null;
+	avg_ttft: number | null;
+	avg_tokens_per_second: number | null;
+	first_timestamp: number | null;
+	last_timestamp: number | null;
+}
+
+interface ModelStatsRow extends AggregatedStatsRow {
+	model: string;
+	provider: string;
+}
+
+interface FolderStatsRow extends AggregatedStatsRow {
+	folder: string;
+}
+
 let db: Database | null = null;
 
 const BACKFILL_COMPLETE = "complete";
@@ -112,6 +148,7 @@ export async function initDb(): Promise<Database> {
 			cost_cache_read REAL NOT NULL,
 			cost_cache_write REAL NOT NULL,
 			cost_total REAL NOT NULL,
+			cost_no_cache_input REAL,
 			agent_type TEXT NOT NULL DEFAULT 'main',
 			UNIQUE(session_file, entry_id)
 		);
@@ -183,6 +220,9 @@ export async function initDb(): Promise<Database> {
 	if (!messageColumns.some(column => column.name === "premium_requests")) {
 		db.run("ALTER TABLE messages ADD COLUMN premium_requests REAL NOT NULL DEFAULT 0");
 	}
+	if (!messageColumns.some(column => column.name === "cost_no_cache_input")) {
+		db.run("ALTER TABLE messages ADD COLUMN cost_no_cache_input REAL");
+	}
 	db.run("UPDATE messages SET premium_requests = 0 WHERE premium_requests IS NULL");
 	// Token-usage-by-agent: each message is classified main / subagent / advisor
 	// from its transcript path. A brand-new table gets the column from CREATE
@@ -265,6 +305,7 @@ export async function initDb(): Promise<Database> {
 	backfillPriorityPremiumRequests(db);
 	backfillAgentType(db);
 	backfillMissingCatalogCosts(db);
+	backfillNoCacheInputCosts(db);
 	backfillForkDuplicates(db);
 	return db;
 }
@@ -323,6 +364,18 @@ function resolveStoredCost(stats: MessageStats): UsageCost {
 	return calculateCatalogCost(stats.provider, stats.model, stats.usage) ?? storedCost ?? ZERO_USAGE_COST;
 }
 
+function calculateNoCacheInputCost(provider: string, modelId: string, tokens: CostTokens): number | null {
+	const cost = getCatalogCost(provider, modelId);
+	if (!cost) return null;
+	const promptInputTokens =
+		tokens.input +
+		tokens.cacheRead +
+		tokens.cacheWrite +
+		(tokens.orchestration?.input ?? 0) +
+		(tokens.orchestration?.cacheRead ?? 0);
+	return calculateUncachedInputCost(cost, promptInputTokens);
+}
+
 function backfillMissingCatalogCosts(database: Database): void {
 	const rows = database
 		.prepare(`
@@ -358,6 +411,31 @@ function backfillMissingCatalogCosts(database: Database): void {
 	applyBackfill();
 }
 
+function backfillNoCacheInputCosts(database: Database): void {
+	const rows = database
+		.prepare(`
+			SELECT id, provider, model, input_tokens, cache_read_tokens, cache_write_tokens
+			FROM messages
+			WHERE cost_no_cache_input IS NULL
+		`)
+		.all() as NoCacheInputCostBackfillRow[];
+	if (rows.length === 0) return;
+
+	const update = database.prepare("UPDATE messages SET cost_no_cache_input = ? WHERE id = ?");
+	const applyBackfill = database.transaction(() => {
+		for (const row of rows) {
+			const cost = calculateNoCacheInputCost(row.provider, row.model, {
+				input: row.input_tokens,
+				output: 0,
+				cacheRead: row.cache_read_tokens,
+				cacheWrite: row.cache_write_tokens,
+			});
+			update.run(cost ?? 0, row.id);
+		}
+	});
+	applyBackfill();
+}
+
 /**
  * Get the stored offset for a session file.
  */
@@ -406,9 +484,9 @@ export function insertMessageStats(stats: MessageStats[]): number {
 			session_file, entry_id, folder, model, provider, api, timestamp,
 			duration, ttft, stop_reason, error_message,
 			input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, total_tokens, premium_requests,
-			cost_input, cost_output, cost_cache_read, cost_cache_write, cost_total, agent_type
+			cost_input, cost_output, cost_cache_read, cost_cache_write, cost_total, cost_no_cache_input, agent_type
 		)
-		SELECT ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?
+		SELECT ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?
 		WHERE NOT EXISTS (
 			SELECT 1 FROM messages
 			WHERE entry_id = ? AND timestamp = ? AND session_file <> ?
@@ -422,6 +500,7 @@ export function insertMessageStats(stats: MessageStats[]): number {
 	const insert = db.transaction(() => {
 		for (const s of stats) {
 			const cost = resolveStoredCost(s);
+			const noCacheInputCost = calculateNoCacheInputCost(s.provider, s.model, s.usage) ?? 0;
 			const result = stmt.run(
 				s.sessionFile,
 				s.entryId,
@@ -445,6 +524,7 @@ export function insertMessageStats(stats: MessageStats[]): number {
 				cost.cacheRead,
 				cost.cacheWrite,
 				cost.total,
+				noCacheInputCost,
 				s.agentType,
 				// `WHERE NOT EXISTS` binds: skip when a different session_file
 				// already holds this (entry_id, timestamp).
@@ -463,7 +543,7 @@ export function insertMessageStats(stats: MessageStats[]): number {
 /**
  * Build aggregated stats from query results.
  */
-function buildAggregatedStats(rows: any[]): AggregatedStats {
+function buildAggregatedStats(rows: AggregatedStatsRow[]): AggregatedStats {
 	if (rows.length === 0) {
 		return {
 			totalRequests: 0,
@@ -475,6 +555,7 @@ function buildAggregatedStats(rows: any[]): AggregatedStats {
 			totalCacheReadTokens: 0,
 			totalCacheWriteTokens: 0,
 			cacheRate: 0,
+			cacheSavings: 0,
 			totalCost: 0,
 			totalPremiumRequests: 0,
 			avgDuration: null,
@@ -492,6 +573,8 @@ function buildAggregatedStats(rows: any[]): AggregatedStats {
 	const totalInputTokens = row.total_input_tokens || 0;
 	const totalCacheReadTokens = row.total_cache_read_tokens || 0;
 	const totalPremiumRequests = row.total_premium_requests || 0;
+	const noCacheInputCost = row.total_no_cache_input_cost || 0;
+	const cachedPromptCost = row.total_cached_prompt_cost || 0;
 
 	return {
 		totalRequests,
@@ -506,6 +589,7 @@ function buildAggregatedStats(rows: any[]): AggregatedStats {
 			totalInputTokens + totalCacheReadTokens > 0
 				? totalCacheReadTokens / (totalInputTokens + totalCacheReadTokens)
 				: 0,
+		cacheSavings: noCacheInputCost > 0 ? (noCacheInputCost - cachedPromptCost) / noCacheInputCost : 0,
 		totalCost: row.total_cost || 0,
 		totalPremiumRequests,
 		avgDuration: row.avg_duration,
@@ -533,6 +617,10 @@ export function getOverallStats(cutoff?: number): AggregatedStats {
 			SUM(cache_write_tokens) as total_cache_write_tokens,
 			SUM(premium_requests) as total_premium_requests,
 			SUM(cost_total) as total_cost,
+			SUM(CASE WHEN cost_no_cache_input > 0
+				THEN cost_input + cost_cache_read + cost_cache_write
+				ELSE 0 END) as total_cached_prompt_cost,
+			SUM(cost_no_cache_input) as total_no_cache_input_cost,
 			AVG(duration) as avg_duration,
 			AVG(ttft) as avg_ttft,
 			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
@@ -543,7 +631,7 @@ export function getOverallStats(cutoff?: number): AggregatedStats {
 	`);
 
 	const rows = hasCutoff ? stmt.all(cutoff) : stmt.all();
-	return buildAggregatedStats(rows as any[]);
+	return buildAggregatedStats(rows as AggregatedStatsRow[]);
 }
 /**
  * Get stats grouped by model.
@@ -564,6 +652,10 @@ export function getStatsByModel(cutoff?: number): ModelStats[] {
 			SUM(cache_write_tokens) as total_cache_write_tokens,
 			SUM(premium_requests) as total_premium_requests,
 			SUM(cost_total) as total_cost,
+			SUM(CASE WHEN cost_no_cache_input > 0
+				THEN cost_input + cost_cache_read + cost_cache_write
+				ELSE 0 END) as total_cached_prompt_cost,
+			SUM(cost_no_cache_input) as total_no_cache_input_cost,
 			AVG(duration) as avg_duration,
 			AVG(ttft) as avg_ttft,
 			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
@@ -575,7 +667,7 @@ export function getStatsByModel(cutoff?: number): ModelStats[] {
 		ORDER BY total_requests DESC
 	`);
 
-	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as any[];
+	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as ModelStatsRow[];
 	return rows.map(row => ({
 		model: row.model,
 		provider: row.provider,
@@ -601,6 +693,10 @@ export function getStatsByFolder(cutoff?: number): FolderStats[] {
 			SUM(cache_write_tokens) as total_cache_write_tokens,
 			SUM(premium_requests) as total_premium_requests,
 			SUM(cost_total) as total_cost,
+			SUM(CASE WHEN cost_no_cache_input > 0
+				THEN cost_input + cost_cache_read + cost_cache_write
+				ELSE 0 END) as total_cached_prompt_cost,
+			SUM(cost_no_cache_input) as total_no_cache_input_cost,
 			AVG(duration) as avg_duration,
 			AVG(ttft) as avg_ttft,
 			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
@@ -612,7 +708,7 @@ export function getStatsByFolder(cutoff?: number): FolderStats[] {
 		ORDER BY total_requests DESC
 	`);
 
-	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as any[];
+	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as FolderStatsRow[];
 	return rows.map(row => ({
 		folder: row.folder,
 		...buildAggregatedStats([row]),

packages/stats/src/index.ts

@@ -68,6 +68,7 @@ async function printStats(): Promise<void> {
 	console.log(`  Input Tokens: ${formatNumber(overall.totalInputTokens)}`);
 	console.log(`  Output Tokens: ${formatNumber(overall.totalOutputTokens)}`);
 	console.log(`  Cache Rate: ${formatPercent(overall.cacheRate)}`);
+	console.log(`  Cache Savings: ${formatPercent(overall.cacheSavings)}`);
 	console.log(`  Total Cost: ${formatCost(overall.totalCost)}`);
 	console.log(`  Premium Requests: ${formatNumber(normalizePremiumRequests(overall.totalPremiumRequests ?? 0))}`);
 	console.log(`  Avg Duration: ${overall.avgDuration !== null ? formatDuration(overall.avgDuration) : "-"}`);
@@ -80,7 +81,7 @@ async function printStats(): Promise<void> {
 		console.log("\nBy Model:");
 		for (const m of byModel.slice(0, 10)) {
 			console.log(
-				`  ${m.model}: ${formatNumber(m.totalRequests)} reqs, ${formatCost(m.totalCost)}, ${formatPercent(m.cacheRate)} cache`,
+				`  ${m.model}: ${formatNumber(m.totalRequests)} reqs, ${formatCost(m.totalCost)}, ${formatPercent(m.cacheRate)} cache rate, ${formatPercent(m.cacheSavings)} cache savings`,
 			);
 		}
 	}

packages/stats/src/shared-types.ts

@@ -25,8 +25,13 @@ export interface AggregatedStats {
 	totalCacheReadTokens: number;
 	/** Total cache write tokens */
 	totalCacheWriteTokens: number;
-	/** Cache hit rate (0-1) */
+	/** Percentage of prompt input tokens served from cache (0-1). */
 	cacheRate: number;
+	/**
+	 * Prompt-input cost saved relative to billing the same tokens uncached
+	 * (0-1; negative when cache writes cost more than reads save).
+	 */
+	cacheSavings: number;
 	/** Total cost */
 	totalCost: number;
 	/** Total premium requests */

packages/stats/test/db-cost.test.ts

@@ -1,6 +1,6 @@
 import { Database } from "bun:sqlite";
 import { describe, expect, it } from "bun:test";
-import { closeDb, getRecentRequests, initDb, insertMessageStats } from "@oh-my-pi/omp-stats/db";
+import { closeDb, getOverallStats, getRecentRequests, initDb, insertMessageStats } from "@oh-my-pi/omp-stats/db";
 import type { MessageStats } from "@oh-my-pi/omp-stats/types";
 import { getBundledModel } from "@oh-my-pi/pi-catalog/models";
 import { getStatsDbPath } from "@oh-my-pi/pi-utils";
@@ -46,6 +46,32 @@ function expectedCodexGptCost() {
 	};
 }
 
+function createAnthropicCacheStats(entryId: string, cacheRead: number, cacheWrite: number): MessageStats {
+	const input = 1_000 - cacheRead - cacheWrite;
+	return {
+		sessionFile: "/tmp/anthropic-session.jsonl",
+		entryId,
+		folder: "/tmp/project",
+		model: "claude-sonnet-4-6",
+		provider: "anthropic",
+		api: "anthropic-messages",
+		timestamp: Date.now(),
+		duration: 1000,
+		ttft: 100,
+		stopReason: "stop",
+		errorMessage: null,
+		usage: {
+			input,
+			output: 0,
+			cacheRead,
+			cacheWrite,
+			totalTokens: 1_000,
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+		},
+		agentType: "main",
+	};
+}
+
 describe("stats GPT cost correction", () => {
 	it("stores catalog-derived cost when OpenAI Codex session usage has zero cost", async () => {
 		await initDb();
@@ -107,3 +133,55 @@ describe("stats GPT cost correction", () => {
 		expect(request?.usage.cost.total).toBeCloseTo(expectedCodexGptCost().total, 8);
 	});
 });
+
+describe("stats cache metrics", () => {
+	it("subtracts 5-minute writes from the savings produced by cache reads", async () => {
+		await initDb();
+		insertMessageStats([createAnthropicCacheStats("mixed-cache", 800, 100)]);
+
+		// 100 uncached + 800 reads at 0.1x + 100 writes at 1.25x = 305,
+		// versus 1,000 tokens at the uncached input rate.
+		expect(getOverallStats().cacheSavings).toBeCloseTo(0.695, 8);
+		expect(getOverallStats().cacheRate).toBeCloseTo(800 / 900, 8);
+	});
+
+	it("reports cache writes without reads as negative savings", async () => {
+		await initDb();
+		insertMessageStats([createAnthropicCacheStats("cache-write", 0, 1_000)]);
+
+		expect(getOverallStats().cacheSavings).toBeCloseTo(-0.25, 8);
+	});
+
+	it("charges 1-hour cache writes at their full overhead", async () => {
+		await initDb();
+		const stats = createAnthropicCacheStats("one-hour-write", 0, 1_000);
+		stats.usage.cost = {
+			input: 0,
+			output: 0,
+			cacheRead: 0,
+			cacheWrite: 0.006,
+			total: 0.006,
+		};
+		insertMessageStats([stats]);
+
+		expect(getOverallStats().cacheSavings).toBeCloseTo(-1, 8);
+	});
+
+	it("excludes unpriced custom models from the savings ratio", async () => {
+		await initDb();
+		const known = createAnthropicCacheStats("known", 800, 100);
+		const unpriced = createAnthropicCacheStats("unpriced", 0, 0);
+		unpriced.provider = "custom";
+		unpriced.model = "custom-model";
+		unpriced.usage.cost = {
+			input: 1,
+			output: 0,
+			cacheRead: 0,
+			cacheWrite: 0,
+			total: 1,
+		};
+		insertMessageStats([known, unpriced]);
+
+		expect(getOverallStats().cacheSavings).toBeCloseTo(0.695, 8);
+	});
+});

packages/stats/test/overview-token-labels.test.tsx

@@ -14,6 +14,7 @@ const stats: AggregatedStats = {
 	totalCacheReadTokens: 300,
 	totalCacheWriteTokens: 40,
 	cacheRate: 0.75,
+	cacheSavings: 0.695,
 	totalCost: 0,
 	totalPremiumRequests: 0,
 	avgDuration: 1000,
@@ -31,6 +32,11 @@ describe("overview token metrics", () => {
 		expect(html).toContain("Cache Read");
 		expect(html).toContain("Conversation Total");
 		expect(html).toContain("Uncached input + cache reads + cache writes + output");
+		expect(html).toContain("Cache Rate");
+		expect(html).toContain("Cache Savings");
+		expect(html).toContain("75.0%");
+		expect(html).toContain("69.5%");
+		expect(html).toContain("cache writes can make this negative");
 
 		const expectedTotal = formatCompact(
 			stats.totalInputTokens +

scripts/bazel-natives.ts

@@ -25,13 +25,16 @@
  * (packages/natives/scripts/build-bindings.ts) against the installed VS Build
  * Tools; every other target on a win32 host fails fast with guidance.
  *
+ * Set `OMP_NATIVE_BUILD_BACKEND=cargo` to route the host target through the
+ * same local N-API build on systems where Bazel's prebuilt host tools cannot run.
+ *
  * Note: musl addons intentionally reuse the plain linux-<arch> filenames, so a
  * `linux-all` copy overwrites the gnu addon with the musl one (and vice versa);
  * CI jobs that ship files always request an explicit disjoint target set.
  */
 import * as fs from "node:fs/promises";
 import * as path from "node:path";
-import { detectHostAvx2Support } from "./host-detect";
+import { detectHostAvx2Support, resolveLocalHostAddon } from "./host-detect";
 
 const repoRoot = path.join(import.meta.dir, "..");
 
@@ -221,14 +224,10 @@ async function installAddon(sourcePath: string, destPath: string): Promise<void>
 	}
 }
 
-/**
- * win32-host path for the `host` pseudo-target: the bazel msvc cross toolchain
- * cannot run here, but real MSVC can — build the addon via the napi local
- * build and install it into destDir like the bazel path would.
- */
-async function buildWindowsHostAddon(host: HostInfo, destDir: string): Promise<void> {
+/** Build and install the host addon through the local Cargo/N-API path. */
+async function buildLocalHostAddon(host: HostInfo, destDir: string): Promise<void> {
 	const script = path.join(repoRoot, "packages/natives/scripts/build-bindings.ts");
-	console.log(`win32 host: bazel msvc toolchain is linux/mac-only; building via ${path.relative(repoRoot, script)}`);
+	console.log(`local host build: using ${path.relative(repoRoot, script)}`);
 	const proc = Bun.spawn([process.execPath, script], {
 		cwd: repoRoot,
 		stdout: "inherit",
@@ -237,7 +236,7 @@ async function buildWindowsHostAddon(host: HostInfo, destDir: string): Promise<v
 	const exitCode = await proc.exited;
 	if (exitCode !== 0) process.exit(exitCode || 1);
 
-	const filename = `pi_natives.win32-x64-${host.avx2 ? "modern" : "baseline"}.node`;
+	const filename = resolveLocalHostAddon(host).filename;
 	const builtPath = path.join(repoRoot, "packages/natives/native", filename);
 	if (path.dirname(builtPath) !== destDir) {
 		await fs.mkdir(destDir, { recursive: true });
@@ -251,15 +250,18 @@ async function main(): Promise<void> {
 	const host: HostInfo = { platform: process.platform, arch: process.arch, avx2: detectHostAvx2Support() };
 	const destDir = options.dest ? path.resolve(options.dest) : path.join(repoRoot, "packages/natives/native");
 
-	if (host.platform === "win32" && !options.source) {
+	if ((host.platform === "win32" || Bun.env.OMP_NATIVE_BUILD_BACKEND === "cargo") && !options.source) {
 		if (options.targets.length !== 1 || options.targets[0] !== "host") {
+			if (host.platform === "win32") {
 				throw new Error(
 					`Cannot bazel-build [${options.targets.join(", ")}] on a Windows host: the msvc cross ` +
 						"toolchain (bazel/toolchains/msvc) only runs on linux/mac exec hosts. Use `host` here " +
 						"(local napi build via VS Build Tools), or run this script from WSL/linux for cross targets.",
 				);
 			}
-		await buildWindowsHostAddon(host, destDir);
+			throw new Error("OMP_NATIVE_BUILD_BACKEND=cargo supports only the host target");
+		}
+		await buildLocalHostAddon(host, destDir);
 		return;
 	}
 	let outputs: string[];

scripts/host-detect.ts

@@ -10,7 +10,27 @@ function runCommand(command: string, args: string[]): string | null {
 		return null;
 	}
 }
+/** Local N-API addon identity derived from a host platform and ISA. */
+export interface LocalHostAddon {
+	readonly filename: string;
+	readonly x64Variant: "modern" | "baseline" | null;
+}
 
+/** Resolve the exact filename and x86-64 ISA emitted by the local N-API build. */
+export function resolveLocalHostAddon(host: {
+	readonly platform: string;
+	readonly arch: string;
+	readonly avx2: boolean;
+}): LocalHostAddon {
+	const x64Variant = host.arch === "x64" ? (host.avx2 ? "modern" : "baseline") : null;
+	const variantSuffix = x64Variant ? `-${x64Variant}` : "";
+	return {
+		filename: `pi_natives.${host.platform}-${host.arch}${variantSuffix}.node`,
+		x64Variant,
+	};
+}
+
+/** Detect whether this x86-64 host can run the modern AVX2 addon. */
 export function detectHostAvx2Support(): boolean {
 	if (process.arch !== "x64") return false;