copilot-relay 0.4.6 → 0.4.7

package/README.md

@@ -55,7 +55,7 @@ running daemon. Add `--details` for safe failure evidence and private replay hin
 Use `copilot-relay status --deep` for daemon health, and `copilot-relay stop` when finished.
 
 Check how well prompt caching works per model and upstream route. The report reads
-only local logs and flags rows below a hit-rate goal, 95% by default:
+only local logs and shows a hit rate below the goal, 95% by default, in red:
 
 ```sh
 copilot-relay cache                  # last 24 hours

package/dist/main.js

@@ -220,7 +220,7 @@ const formatUpstreamProxyForDisplay = (upstreamProxy) => {
 * `status` prints config values directly to stdout. A config string carrying
 * raw control bytes could clear the line and paint status rows the relay never
 * produced. The config file is user-owned, so this is hardening rather than a
-* vulnerability — but it costs nothing to make the output non-forgeable. #47.
+* vulnerability — but it costs nothing to make the output non-forgeable.
 */
 const sanitizeTerminalString = (text$1) => text$1.replace(ansiEscapePattern, "").replace(controlBytePattern, "");
 /**
@@ -513,7 +513,7 @@ const cleanupLogsIfDue = async () => {
 * Each argument is rendered exactly once and scrubbed once, and the identical
 * array of strings goes to both the file and the console. Scrubbing only on the
 * way to disk would leave the original object on the console - the screen a
-* user screenshots into an issue - so the two sinks must not diverge. See #47.
+* user screenshots into an issue - so the two sinks must not diverge.
 *
 * Rendering is skipped only when neither sink would emit, so the cost matches
 * the old behavior at info level. The console gate is read from consola rather
@@ -595,7 +595,7 @@ var InvalidProxyEnvironmentError = class extends Error {
 /**
 * Builds the dispatcher for one resolved upstreamProxy value.
 *
-* Empty connects directly, as every release before #153 did. "env" hands HTTPS_PROXY, HTTP_PROXY
+* Empty connects directly. "env" hands HTTPS_PROXY, HTTP_PROXY
 * and NO_PROXY, in upper or lower case, to EnvHttpProxyAgent. Anything else is a proxy URL that
 * normalizeUpstreamProxy has already validated.
 */
@@ -1373,7 +1373,7 @@ const normalizeApiKey = (value) => {
 * through. None of them can be found again in arbitrary log text: practical URL
 * detection anchors on a literal "://", and loosening that to chase a bare
 * "https:" would start matching ordinary prose. Requiring the authority prefix
-* is the one boundary where this has a definite answer. See #47.
+* is the one boundary where this has a definite answer.
 */
 const conventionalHttpUrlPattern = /^https?:\/\//i;
 /**
@@ -1390,7 +1390,7 @@ const conventionalHttpUrlPattern = /^https?:\/\//i;
 *
 * Rejecting the raw byte costs nothing: every one of these has a
 * percent-encoded spelling that is unambiguous in log text and is accepted
-* here unchanged. See #47.
+* here unchanged.
 */
 const unsafeRawDelimiterPattern = /['"`<>\s\u0000-\u001F\u007F]/;
 /**
@@ -1404,7 +1404,7 @@ const copilotBaseUrlDelimiterMessage = "Invalid copilotBaseUrl: quotes, backtick
 /**
 * Validates copilotBaseUrl as a conventional HTTP(S) URL without credentials.
 *
-* Four rules, all learned from #47:
+* Four rules:
 *
 * The value is concatenated as `${base}${path}` for every upstream call, so it
 * must be something Undici will accept. A relative or non-HTTP value fails at
@@ -2146,36 +2146,6 @@ const dataColumns = [
 		alignRight: true,
 		cell: (row) => formatCount(row.requests)
 	},
-	{
-		header: "UNKNOWN",
-		alignRight: true,
-		cell: (row) => formatCount(row.unknownCacheRequests)
-	},
-	{
-		header: "0-READ",
-		alignRight: true,
-		cell: (row) => formatCount(row.zeroCacheReadRequests)
-	},
-	{
-		header: "INPUT",
-		alignRight: true,
-		cell: (row) => formatCount(row.totalInputTokens)
-	},
-	{
-		header: "CACHE READ",
-		alignRight: true,
-		cell: (row) => formatCount(row.cacheReadTokens)
-	},
-	{
-		header: "UNCACHED",
-		alignRight: true,
-		cell: (row) => formatCount(row.uncachedInputTokens)
-	},
-	{
-		header: "CACHE WRITE",
-		alignRight: true,
-		cell: (row) => row.cacheWriteTokens === null ? "-" : formatCount(row.cacheWriteTokens)
-	},
 	{
 		header: "HIT RATE",
 		alignRight: true,
@@ -2190,12 +2160,10 @@ const renderTable = (rows, view, color) => {
 	const pad = (text$1, index) => columns[index].alignRight ? text$1.padStart(widths[index]) : text$1.padEnd(widths[index]);
 	const header = columns.map((column, index) => pad(column.header, index)).join("  ");
 	const body = rows.map((row, rowIndex) => {
-		const rendered = columns.map((column, index) => {
+		return `  ${columns.map((column, index) => {
 			const text$1 = pad(cells[rowIndex][index], index);
 			return column.tone === void 0 ? text$1 : colorText(text$1, column.tone(row), color);
-		});
-		const marker = row.belowGoal ? `  ${colorText("below goal", "bad", color)}` : "";
-		return `  ${rendered.join("  ")}${marker}`;
+		}).join("  ")}`;
 	});
 	return [`  ${header}`, ...body];
 };
@@ -2208,7 +2176,7 @@ const formatLocalTime = (date) => `${formatLogDate(date)} ${twoDigits(date.getHo
 const describeScope = (options) => {
 	return `${options.since === void 0 ? "across all retained logs" : `since ${formatLocalTime(options.since)} local time`}${options.model === void 0 ? "" : `, for models matching "${terminalText(options.model)}"`}`;
 };
-const unknownNote = "UNKNOWN calls logged no cache_read_input_tokens; they are left out of the token columns and HIT RATE.";
+const unknownNote = (calls) => `HIT RATE leaves out ${formatCount(calls)} ${calls === 1 ? "call" : "calls"} that logged no cache_read_input_tokens.`;
 /** The report as text lines. Kept free of IO, so it is tested without a terminal. */
 const renderCacheReport = (rows, options, color = false) => {
 	if (rows.length === 0) return [
@@ -2221,7 +2189,8 @@ const renderCacheReport = (rows, options, color = false) => {
 		"",
 		...renderTable(rows, options.view, color)
 	];
-	if (rows.some((row) => row.unknownCacheRequests > 0)) lines.push("", unknownNote);
+	const unknownCalls = rows.reduce((total, row) => total + row.unknownCacheRequests, 0);
+	if (unknownCalls > 0) lines.push("", unknownNote(unknownCalls));
 	return lines;
 };
 
@@ -2255,7 +2224,7 @@ const cache = defineCommand({
 		},
 		goal: {
 			default: String(defaultGoal),
-			description: "Hit-rate goal in percent, with at most two decimals. Rows below it are flagged.",
+			description: "Hit-rate goal in percent, with at most two decimals. A hit rate below it is shown in red.",
 			type: "string"
 		}
 	},
@@ -3358,10 +3327,10 @@ const pinCopilotModelCatalog = (config) => {
 	pinnedCatalogs.set(config, config.modelCatalog);
 };
 const pendingCatalogs = /* @__PURE__ */ new WeakMap();
-const isRecord$4 = (value) => typeof value === "object" && value !== null && !Array.isArray(value);
+const isRecord$5 = (value) => typeof value === "object" && value !== null && !Array.isArray(value);
 const isPositiveInteger = (value) => typeof value === "number" && Number.isSafeInteger(value) && value > 0;
 function parseModelTokenLimits(value) {
-	if (!isRecord$4(value)) return;
+	if (!isRecord$5(value)) return;
 	const context = value.max_context_window_tokens;
 	const prompt = value.max_prompt_tokens;
 	const output = value.max_output_tokens;
@@ -3417,13 +3386,13 @@ async function loadCopilotModelCatalog(config) {
 		});
 		if (!response.ok) throw new HTTPError("Failed to validate upstream models", response);
 		const payload = await readCopilotJson(response, signal, config.upstreamTimeoutMs);
-		if (!isRecord$4(payload) || !Array.isArray(payload.data)) throw new Error("Copilot /models returned an invalid model catalog.");
+		if (!isRecord$5(payload) || !Array.isArray(payload.data)) throw new Error("Copilot /models returned an invalid model catalog.");
 		const models$1 = /* @__PURE__ */ new Map();
 		for (const model of payload.data) {
-			if (!isRecord$4(model) || typeof model.id !== "string" || !model.id) continue;
-			const capabilities = isRecord$4(model.capabilities) ? model.capabilities : void 0;
+			if (!isRecord$5(model) || typeof model.id !== "string" || !model.id) continue;
+			const capabilities = isRecord$5(model.capabilities) ? model.capabilities : void 0;
 			const limits = parseModelTokenLimits(capabilities?.limits);
-			const supports = isRecord$4(capabilities?.supports) ? capabilities.supports : void 0;
+			const supports = isRecord$5(capabilities?.supports) ? capabilities.supports : void 0;
 			const supportedEndpoints = parseStringArray(model.supported_endpoints);
 			const reasoningEfforts = parseReasoningEfforts(supports);
 			models$1.set(model.id, {
@@ -3671,11 +3640,15 @@ const probeHint = (code) => {
 //#endregion
 //#region src/claude/billing-line.ts
 const billingLinePrefix = "x-anthropic-billing-header:";
-const afterBillingLine = (text$1) => {
+const leadingBillingLine = (text$1) => {
 	if (!text$1.startsWith(billingLinePrefix)) return;
 	const lineEnd = text$1.indexOf("\n");
-	if (lineEnd === -1) return "";
-	return text$1.slice(lineEnd + 1).replace(/^(?:\r?\n)+/, "");
+	return lineEnd === -1 ? text$1 : text$1.slice(0, lineEnd);
+};
+const afterBillingLine = (text$1) => {
+	const line = leadingBillingLine(text$1);
+	if (line === void 0) return;
+	return text$1.slice(line.length).replace(/^(?:\r?\n)+/, "");
 };
 const isTextBlock = (block) => typeof block === "object" && block !== null && block.type === "text" && typeof block.text === "string";
 const withSystem = (payload, system) => {
@@ -3711,6 +3684,120 @@ const removeBillingLine = (payload) => {
 	if (!changed) return payload;
 	return withSystem(payload, kept.length > 0 ? kept : void 0);
 };
+const systemTexts$1 = (system) => {
+	if (typeof system === "string") return [system];
+	return Array.isArray(system) ? system.filter(isTextBlock).map((block) => block.text) : [];
+};
+const readBillingLineField = (payload, name) => {
+	if (typeof payload !== "object" || payload === null) return;
+	const line = systemTexts$1(payload.system).map((text$1) => leadingBillingLine(text$1)).find((text$1) => text$1 !== void 0);
+	if (line === void 0) return;
+	for (const field of line.slice(27).split(";")) {
+		const separator = field.indexOf("=");
+		if (separator !== -1 && field.slice(0, separator).trim() === name) return field.slice(separator + 1).trim();
+	}
+};
+
+//#endregion
+//#region src/claude/initiator.ts
+const ownRequestSystemPrompts = ["You are a helpful AI assistant tasked with summarizing conversations.", "You are naming a coding session"];
+const compactionGuard = "CRITICAL: Respond with TEXT ONLY.";
+const compactionInstruction = "Your task is to create a detailed summary of the conversation so far";
+const personDeliveries = [
+	"The user sent a new message while you were working:",
+	"A message arrived in the bound thread while you were working:",
+	"Messages arrived in the bound thread while you were working:"
+];
+const claudeCodeTriggers = [
+	"This session is being continued from a previous conversation",
+	"The summarized conversation included Artifact content",
+	"The previous response failed to produce a valid tool call.",
+	"Your tool call was malformed and could not be parsed.",
+	"[Your previous response had no visible output.",
+	"[structured-output-enforce]",
+	"[projects-reply-gate]",
+	"Output token limit hit",
+	"Your response above was cut off mid-stream",
+	"Your response above was stopped by a safety classifier",
+	"Activity in the bound conversation",
+	"Another Claude session sent a message",
+	"Your background observer",
+	"[MESSAGE FROM NON-USER SOURCE",
+	"[SCHEDULED TASK",
+	"<channel source=",
+	"<teammate-message"
+];
+const claudeCodeTriggerPatterns = [/^\S+ hook (?:feedback|blocking error from command):/, /^The \S+ plugin sent a message/];
+const claudeCodeContext = [
+	"[Request interrupted by user",
+	"[Tool call ",
+	"The user doesn't want to take this action right now.",
+	"Base directory for this skill:"
+];
+const claudeCodeContextPatterns = [/^\S+ hook additional context:/];
+const statusProbePrompt = "Reply with the single word: ok";
+const statusProbeMaxTokens = 16;
+const isRecord$4 = (value) => typeof value === "object" && value !== null && !Array.isArray(value);
+const innermostReminder = /<system-reminder>(?:(?!<system-reminder>)[\s\S])*?<\/system-reminder>/g;
+const withoutReminders = (text$1) => {
+	let previous = text$1;
+	let remaining = text$1.replace(innermostReminder, "");
+	while (remaining !== previous) {
+		previous = remaining;
+		remaining = remaining.replace(innermostReminder, "");
+	}
+	return remaining.trim();
+};
+const startsWithAny = (text$1, prefixes) => prefixes.some((prefix) => text$1.startsWith(prefix));
+const matchesAny = (text$1, prefixes, patterns) => startsWithAny(text$1, prefixes) || patterns.some((pattern) => pattern.test(text$1));
+const contentBlocks = (content) => {
+	if (typeof content === "string") return [{
+		type: "text",
+		text: content
+	}];
+	return Array.isArray(content) ? content.filter(isRecord$4) : [];
+};
+const originatingTurn = (messages) => {
+	const blocks = [];
+	let sawUser = false;
+	for (let index = messages.length - 1; index >= 0; index--) {
+		const message = messages[index];
+		if (!isRecord$4(message) || message.role === "system") continue;
+		if (message.role !== "user") break;
+		sawUser = true;
+		blocks.unshift(...contentBlocks(message.content));
+	}
+	return sawUser ? blocks : void 0;
+};
+const blockOrigin = (block) => {
+	if (block.type === "tool_result") return "agent";
+	if (block.type === "image" || block.type === "document") return "user";
+	if (block.type !== "text" || typeof block.text !== "string") return;
+	const text$1 = withoutReminders(block.text);
+	if (text$1 === "" || matchesAny(text$1, claudeCodeContext, claudeCodeContextPatterns)) return;
+	if (startsWithAny(text$1, personDeliveries)) return "user";
+	if (text$1.includes(compactionGuard) && text$1.includes(compactionInstruction) || matchesAny(text$1, claudeCodeTriggers, claudeCodeTriggerPatterns)) return "agent";
+	return "user";
+};
+const isStatusProbe = (body) => {
+	if (Object.keys(body).sort().join(",") !== "max_tokens,messages,model" || body.max_tokens !== statusProbeMaxTokens) return false;
+	const messages = body.messages;
+	return Array.isArray(messages) && messages.length === 1 && isRecord$4(messages[0]) && messages[0].role === "user" && messages[0].content === statusProbePrompt;
+};
+const systemTexts = (system) => contentBlocks(system).flatMap((block) => block.type === "text" && typeof block.text === "string" ? [block.text] : []);
+const getClaudeRequestInitiator = (body) => {
+	if (!isRecord$4(body) || !Array.isArray(body.messages) || isStatusProbe(body)) return "agent";
+	const payload = body;
+	if (readBillingLineField(payload, "cc_is_subagent") === "true" || readBillingLineField(payload, "cc_workload") === "cron") return "agent";
+	if (systemTexts(removeBillingLine(payload).system).some((text$1) => startsWithAny(text$1.trimStart(), ownRequestSystemPrompts))) return "agent";
+	const turn = originatingTurn(body.messages);
+	if (turn === void 0) return "agent";
+	for (let index = turn.length - 1; index >= 0; index--) {
+		const origin = blockOrigin(turn[index]);
+		if (origin !== void 0) return origin;
+	}
+	return "agent";
+};
 
 //#endregion
 //#region src/claude/utils.ts
@@ -3741,7 +3828,7 @@ function mapOpenAIStopReasonToClaude(finishReason) {
 }
 /**
 * Upstream sends an empty `arguments` string for tools that take no
-* parameters (#114). Blank text is the empty object; anything else must be a
+* parameters. Blank text is the empty object; anything else must be a
 * JSON object, and a failure names the tool rather than surfacing a bare
 * SyntaxError. Argument text is never echoed: it can carry user data.
 */
@@ -5690,7 +5777,6 @@ const normalizeFinalAssistantPrefill = (payload) => {
 		messages
 	};
 };
-const isAgentInitiator = (messages) => messages.some((message) => message.role === "assistant" || message.role === "tool") ? "agent" : "user";
 const holdsImage = (message) => Array.isArray(message.content) && message.content.some((part) => part.type === "image_url");
 const messagesIncludeImage = (messages) => messages.some(holdsImage);
 const isToolImageMessage = (message) => message.role === "tool" && holdsImage(message);
@@ -5767,7 +5853,7 @@ const createChatCompletions = async (config, payload, options = {}) => {
 	const compatiblePayload = useResponsesApi ? upstreamPayload : normalizeFinalAssistantPrefill(moveToolImagesToUserMessages(upstreamPayload));
 	const provider = getCopilotProviderContext(config);
 	const enableVision = messagesIncludeImage(compatiblePayload.messages);
-	const initiator = isAgentInitiator(compatiblePayload.messages);
+	const initiator = options.initiator ?? "agent";
 	const requestPayload = buildRequestPayload(compatiblePayload);
 	const effectiveEffortLabel = requestPayload.reasoning_effort ?? "omitted";
 	log.info(sanitizeTerminalString([
@@ -5778,7 +5864,8 @@ const createChatCompletions = async (config, payload, options = {}) => {
 		`upstream_model=${compatiblePayload.model}`,
 		`requested_think_effort=${requestedThinkEffort}`,
 		`requested_thinking=${requestedThinking}`,
-		`effective_think_effort=${effectiveEffortLabel}`
+		`effective_think_effort=${effectiveEffortLabel}`,
+		`initiator=${initiator}`
 	].join(" ")));
 	if (useResponsesApi) return completeResponse(await createResponses(provider, compatiblePayload, {
 		vision: enableVision,
@@ -6012,7 +6099,8 @@ async function createNativeMessages(config, payload, options) {
 	const beta = options.headers?.get("anthropic-beta");
 	if (beta) headers["anthropic-beta"] = beta;
 	const effortLabel = effort === void 0 ? "omitted" : getClaudeTurnEffort(body).effective;
-	log.info(sanitizeTerminalString(`request_id=${options.requestId} Model request client=claude requested_model=${payload.model} upstream_model=${payload.model} upstream_api=messages effective_think_effort=${effortLabel}`));
+	const initiator = options.initiator ?? "agent";
+	log.info(sanitizeTerminalString(`request_id=${options.requestId} Model request client=claude requested_model=${payload.model} upstream_model=${payload.model} upstream_api=messages effective_think_effort=${effortLabel} initiator=${initiator}`));
 	const response = await fetchCopilot(getCopilotProviderContext(config), "/v1/messages", {
 		method: "POST",
 		headers,
@@ -6021,7 +6109,7 @@ async function createNativeMessages(config, payload, options) {
 		requestId: options.requestId,
 		signal: options.signal,
 		timeoutMs: config.upstreamTimeoutMs,
-		initiator: payload.messages.some((message) => message.role === "assistant") ? "agent" : "user"
+		initiator
 	});
 	if (!response.ok) {
 		const detail = await readCopilotText(response, options.signal, config.upstreamTimeoutMs);
@@ -6294,7 +6382,8 @@ async function handleNativeMessages(config, payload, options, write) {
 		]
 	}, {
 		...options,
-		signal
+		signal,
+		initiator: "agent"
 	});
 	const offset = combined.length;
 	let finalHeldFrom;
@@ -6497,13 +6586,14 @@ const translationEndpoint = (config, model) => {
 	const selection = selectCopilotEndpoint(config, normalizeCopilotModelId(model));
 	return selection.endpoint && !selection.responsesFallback ? selection.endpoint : void 0;
 };
-const handleClaudeMessageRequest = async (config, claudePayload, requestSignal, requestId, writeEvent, requestHeaders) => {
+const handleClaudeMessageRequest = async (config, claudePayload, initiator, requestSignal, requestId, writeEvent, requestHeaders) => {
 	const upstreamModel = translateModelName(claudePayload.model);
 	if (shouldUseNativeMessages(config, upstreamModel)) return handleNativeMessages(config, {
 		...claudePayload,
 		model: upstreamModel
 	}, {
 		requestId,
+		initiator,
 		signal: requestSignal,
 		headers: requestHeaders
 	}, writeEvent);
@@ -6515,6 +6605,7 @@ const handleClaudeMessageRequest = async (config, claudePayload, requestSignal,
 	if (shouldLetModelDecideWebSearch && !canStreamWebSearchDecision) openAIPayload.stream = false;
 	const response = await createChatCompletions(config, openAIPayload, {
 		client: "claude",
+		initiator,
 		requestedModel: claudePayload.model,
 		requestedThinkEffort: getClaudeRequestedThinkEffort(claudePayload),
 		requestedThinking: getClaudeRequestedThinking(claudePayload),
@@ -6588,6 +6679,7 @@ const handleClaudeMessageRequest = async (config, claudePayload, requestSignal,
 			else {
 				const finalResponse = await createChatCompletions(config, createFinalWebSearchPayload(openAIPayload, search, toolNameMapper), {
 					client: "claude",
+					initiator: "agent",
 					requestedModel: claudePayload.model,
 					requestedThinkEffort: getClaudeRequestedThinkEffort(claudePayload),
 					requestedThinking: getClaudeRequestedThinking(claudePayload),
@@ -6665,7 +6757,9 @@ claudeRoutes.get("/models", (c) => c.json({
 claudeRoutes.post("/messages", async (c) => {
 	const config = c.get("config");
 	const requestId = c.get("requestId");
-	const claudePayload = removeBillingLine(await c.req.json());
+	const receivedPayload = await c.req.json();
+	const initiator = c.env?.relayProbe === true ? "agent" : getClaudeRequestInitiator(receivedPayload);
+	const claudePayload = removeBillingLine(receivedPayload);
 	const requestSignal = createCopilotRequestSignal(c.req.raw.signal, config.upstreamTimeoutMs);
 	try {
 		getRequestReasoningEffort(claudePayload);
@@ -6716,7 +6810,7 @@ claudeRoutes.post("/messages", async (c) => {
 				data: JSON.stringify(event)
 			}));
 			try {
-				await handleClaudeMessageRequest(config, claudePayload, requestSignal, requestId, writeEvent, c.req.raw.headers);
+				await handleClaudeMessageRequest(config, claudePayload, initiator, requestSignal, requestId, writeEvent, c.req.raw.headers);
 			} catch (error) {
 				log.error(`request_id=${requestId} Error during Claude stream request:`, error);
 				if (error instanceof UpstreamToolInputError) trace?.recordFailure("invalid-tool-input");
@@ -6730,7 +6824,7 @@ claudeRoutes.post("/messages", async (c) => {
 		});
 	}
 	try {
-		return c.json(await handleClaudeMessageRequest(config, claudePayload, requestSignal, requestId, void 0, c.req.raw.headers));
+		return c.json(await handleClaudeMessageRequest(config, claudePayload, initiator, requestSignal, requestId, void 0, c.req.raw.headers));
 	} catch (error) {
 		if (error instanceof ProxyNotImplementedError) {
 			c.set("requestErrorMessage", error.message);
@@ -7140,7 +7234,7 @@ async function probeModels(config, entries, options) {
 								content: "Reply with OK only."
 							}]
 						})
-					}));
+					}), { relayProbe: true });
 					return {
 						response,
 						body: await response.json().catch(() => void 0)
@@ -8820,7 +8914,7 @@ const restart = defineCommand({
 * for different responses - restart the service versus re-authenticate.
 *
 * A failed health probe is a 2, not a 0. Printing FAILED while exiting 0 would
-* make every scripted caller treat an unreachable relay as fine. See #34.
+* make every scripted caller treat an unreachable relay as fine.
 */
 const exitCodes = {
 	notRunning: 1,
@@ -8852,9 +8946,9 @@ const hasVersionMismatch = (status$1) => status$1.running && status$1.daemonVers
 *
 * 0 requires a live process *and* a passing health probe. Printing FAILED while
 * exiting 0 would make every scripted caller treat an unreachable relay as
-* fine. See #34.
+* fine.
 *
-* A version mismatch stays 0 on purpose (#43). The relay works; it is just not
+* A version mismatch stays 0 on purpose. The relay works; it is just not
 * the build that was installed. Mapping it to 2 would overstate it and break
 * every scripted caller that treats non-zero as broken. It is surfaced in the
 * text instead.
@@ -8994,7 +9088,7 @@ const describeError = (error) => {
 * The version rides along on a probe `status` already makes, so asking what is
 * running costs nothing extra. Absent for a daemon older than v0.3.1, which is
 * itself informative, so it stays optional rather than defaulting to the
-* caller's version - that substitution is the bug. See #43.
+* caller's version - that substitution is the bug.
 */
 const checkHealth = async (baseUrl) => {
 	try {
@@ -9037,9 +9131,9 @@ const checkDeep = async (baseUrl, model, apiKey) => {
 	try {
 		const result = await probe(`${baseUrl}/v1/messages`, {
 			body: JSON.stringify({
-				max_tokens: 16,
+				max_tokens: statusProbeMaxTokens,
 				messages: [{
-					content: "Reply with the single word: ok",
+					content: statusProbePrompt,
 					role: "user"
 				}],
 				model