github-router 0.3.319 → 0.3.321

package/dist/{server-setup-DPay_uYp.js → server-setup-CXXKmPIE.js}

@@ -1,4 +1,4 @@
-import { $ as EFFORT_ORDER, $t as parseJsonOrDiagnose, An as oneMContextDisabled, B as resolveAdvisorEffort, Bn as CHEAP_PROFILE_MODELS, Bt as getTextTokenCount, Cn as isBudgetClaudeLead, D as toolbeltEnabled, Dn as upstreamMaxConnections, En as upstreamAllowH2, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as CHEAPEST_PROFILE_NATIVE_AGENT_NAMES, G as repairKnownThinkingHistory, Gn as withInstallLock, H as resolveCheapestAdvisorModel, Ht as getTokenizerFromModel, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, J as buildOpenAIErrorEvent, Jt as resolveMcpToolTimeoutMs, K as repairRejectedThinkingHistory, Kt as assembleResponsesPayload, L as buildAdvisorStream, Lt as shimDefaultsToXhigh, Mn as withOneMSuffixForLead, N as searchWeb, Nn as CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL, On as classifyMessagesRoute, P as ADVISOR_INTERNAL_TOOL_NAME, Pn as CHEAPEST_PROFILE_MODELS, Q as relayAnthropicStream, Qt as readResponseBodyCapped, R as injectAdvisorTool, Rn as CHEAPEST_PROFILE_SUBAGENT_CONTEXT_TOKENS, Rt as countTokens, Sn as generateRandomPort, U as formatThinkingRepairDecline, Ut as findLaunchBySecret, V as resolveAdvisorModel, Vn as CHEAP_PROFILE_NATIVE_AGENT_NAMES, Vt as getTokenCount, W as rememberThinkingHistoryRepair, Wn as CHEAP_PROFILE_SUBAGENT_CONTEXT_TOKENS, X as logStreamError, Xt as createChatCompletions, Y as isControllerClosedError, Yt as createResponses, Z as readIteratorWithTimeout, Zt as MAX_RESPONSE_BODY_BYTES, at as agentToolsEnabled, bn as UPSTREAM_FETCH_TIMEOUT_MS, cn as provisionTreeSitterAssets, en as normalizeOpenAIUsage, et as UNKNOWN_EFFORT_ANCHOR, hn as BUDGET_SMALL_FAST_SLUG, i as assertMcpToolSurfaceConsistent, it as handleMcpPost, jn as withOneMSuffix, kn as catalogAdvertises1M, nt as clampEffort, pn as toolbeltPathOverride, q as buildAnthropicErrorEvent, qt as warnOnTokenPriceDrift, rt as handleMcpDelete, tt as bucketEffort, xn as UPSTREAM_INACTIVITY_TIMEOUT_MS, z as isAdvisorRequested, zn as CHEAP_PROFILE_ADVISOR_CLIENT_MODEL, zt as createMessages } from "./peer-mcp-personas-CRMVuMVF.js";
+import { $ as EFFORT_ORDER, $t as parseJsonOrDiagnose, An as oneMContextDisabled, B as resolveAdvisorEffort, Bn as CHEAP_PROFILE_MODELS, Bt as getTextTokenCount, Cn as isBudgetClaudeLead, D as toolbeltEnabled, Dn as upstreamMaxConnections, En as upstreamAllowH2, F as ADVISOR_TOOL_INSTRUCTIONS, Fn as CHEAPEST_PROFILE_NATIVE_AGENT_NAMES, G as repairKnownThinkingHistory, Gn as withInstallLock, H as resolveCheapestAdvisorModel, Ht as getTokenizerFromModel, I as FAST_ADVISOR_TOOL_INSTRUCTIONS, J as buildOpenAIErrorEvent, Jt as resolveMcpToolTimeoutMs, K as repairRejectedThinkingHistory, Kt as assembleResponsesPayload, L as buildAdvisorStream, Lt as shimDefaultsToXhigh, Mn as withOneMSuffixForLead, N as searchWeb, Nn as CHEAPEST_PROFILE_ADVISOR_CLIENT_MODEL, On as classifyMessagesRoute, P as ADVISOR_INTERNAL_TOOL_NAME, Pn as CHEAPEST_PROFILE_MODELS, Q as relayAnthropicStream, Qt as readResponseBodyCapped, R as injectAdvisorTool, Rn as CHEAPEST_PROFILE_SUBAGENT_CONTEXT_TOKENS, Rt as countTokens, Sn as generateRandomPort, U as formatThinkingRepairDecline, Ut as findLaunchBySecret, V as resolveAdvisorModel, Vn as CHEAP_PROFILE_NATIVE_AGENT_NAMES, Vt as getTokenCount, W as rememberThinkingHistoryRepair, Wn as CHEAP_PROFILE_SUBAGENT_CONTEXT_TOKENS, X as logStreamError, Xt as createChatCompletions, Y as isControllerClosedError, Yt as createResponses, Z as readIteratorWithTimeout, Zt as MAX_RESPONSE_BODY_BYTES, at as agentToolsEnabled, bn as UPSTREAM_FETCH_TIMEOUT_MS, cn as provisionTreeSitterAssets, en as normalizeOpenAIUsage, et as UNKNOWN_EFFORT_ANCHOR, hn as BUDGET_SMALL_FAST_SLUG, i as assertMcpToolSurfaceConsistent, it as handleMcpPost, jn as withOneMSuffix, kn as catalogAdvertises1M, nt as clampEffort, pn as toolbeltPathOverride, q as buildAnthropicErrorEvent, qt as warnOnTokenPriceDrift, rt as handleMcpDelete, tt as bucketEffort, xn as UPSTREAM_INACTIVITY_TIMEOUT_MS, z as isAdvisorRequested, zn as CHEAP_PROFILE_ADVISOR_CLIENT_MODEL, zt as createMessages } from "./peer-mcp-personas-BerVEkZ9.js";
 import { t as getPackageVersion } from "./version-C8x2hQrZ.js";
 import { i as ensurePaths, t as PATHS } from "./paths-CHBAj_t9.js";
 import { t as state } from "./state-BlMpEFAt.js";
@@ -7,7 +7,7 @@ import { t as getCopilotUsage } from "./get-copilot-usage-BWCAzmSc.js";
 import { c as runCommandVoid, i as parseBoolEnv, o as resolveExecutable, r as killManagedTree, s as runCommandCapture } from "./exec-COUT4QsS.js";
 import { c as FAST_PROFILE_MODELS, l as FAST_PROFILE_NATIVE_AGENT_NAMES, r as FAST_PROFILE_ADVISOR_MODEL } from "./fast-profile-contract-_DhIWank.js";
 import { A as normalizeTrailingOneMSuffix, a as MAX_PROFILE_NATIVE_AGENT_NAMES, d as maxAdvisorPinIsValid, h as maxOpusModel, i as MAX_PROFILE_MODELS, j as stripTrailingOneMSuffix, k as fastEndpointForModel, l as maxAdvisorEffortForModel, n as MAX_PROFILE_ALLOWED_LEAD_MODEL_IDS, t as MAX_PROFILE_ADVISOR_INSTRUCTIONS, u as maxAdvisorModelFromPin, y as validateMaxProfilePrerequisites } from "./max-profile-contract-Bl7I_EOC.js";
-import { c as recordAic, d as extractCopilotUsage, f as formatAic, i as extractAndRecordAic, p as nanoAiuToCredits } from "./aic-ledger-DY03LbBM.js";
+import { a as extractAndRecordPricedAic, f as extractCopilotUsage, i as extractAndRecordAic, l as recordAic, m as nanoAiuToCredits, p as formatAic } from "./aic-ledger-BrDqsuge.js";
 import consola from "consola";
 import fs, { existsSync } from "node:fs";
 import path from "node:path";
@@ -2171,10 +2171,10 @@ function formatTokens(n) {
 * Build a context window summary: "in:1.2K out:50 ctx:1.2K/1M (0.1%)"
 */
 function formatTokenInfo(inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, model, aicNano) {
-	if (inputTokens === void 0) return void 0;
+	if (inputTokens === void 0 && (aicNano ?? 0) <= 0) return void 0;
 	const parts = [];
 	const maxPrompt = model?.capabilities?.limits?.max_prompt_tokens;
-	if (maxPrompt) {
+	if (inputTokens !== void 0) if (maxPrompt) {
 		const pct = (inputTokens / maxPrompt * 100).toFixed(1);
 		parts.push(`in:${formatTokens(inputTokens)}/${formatTokens(maxPrompt)} (${pct}%)`);
 	} else parts.push(`in:${formatTokens(inputTokens)}`);
@@ -2291,7 +2291,7 @@ function collectToolFieldKeys(body) {
 //#endregion
 //#region package.json
 var name = "github-router";
-var version = "0.3.319";
+var version = "0.3.321";
 //#endregion
 //#region src/lib/approval.ts
 const awaitApproval = async () => {
@@ -3011,13 +3011,33 @@ async function handleCompletion$1(c) {
 	let upstreamFinished = firstResult.done;
 	let consumerCancelled = false;
 	let aicRecorded = false;
+	let streamAicNano;
+	let completionLogged = false;
 	const tapAic = (chunk) => {
 		if (aicRecorded || !chunk.data || chunk.data === "[DONE]") return;
 		try {
 			const parsed = JSON.parse(chunk.data);
-			if (extractAndRecordAic(resolvedModel, parsed) !== void 0) aicRecorded = true;
+			const nano = extractAndRecordPricedAic(resolvedModel, parsed);
+			if (nano !== void 0) {
+				aicRecorded = true;
+				streamAicNano = nano;
+			}
 		} catch {}
 	};
+	const logStreamCompletion = () => {
+		if (completionLogged || consumerCancelled || streamAicNano === void 0) return;
+		completionLogged = true;
+		logRequest({
+			method: "POST",
+			path: c.req.path,
+			model: originalModel,
+			resolvedModel,
+			aicNano: streamAicNano,
+			status: 200,
+			streaming: true,
+			streamMs: Date.now() - startTime
+		}, selectedModel, startTime);
+	};
 	const safeClose = (controller) => {
 		try {
 			controller.close();
@@ -3061,6 +3081,7 @@ async function handleCompletion$1(c) {
 				}
 				if (result.done) {
 					upstreamFinished = true;
+					logStreamCompletion();
 					safeClose(controller);
 					return;
 				}
@@ -3490,7 +3511,10 @@ function sanitizeAnthropicBody(rawBody) {
 * Verified live: `copilot_usage` rides inside the terminal `message_delta`
 * event beside `usage`. Chunks are arbitrary byte splits, so the tap buffers
 * text, splits on blank-line event boundaries, and keeps the tail. Recording
-* fires at most once per tap instance (one terminal event per stream).
+* fires at most once per tap instance (one terminal event per stream), and
+* only for a priced reading: a zero-nano frame is skipped so an interim
+* free frame can never latch ahead of the real terminal value (same
+* contract as `extractAndRecordPricedAic`).
 *
 * Total and never throws — a malformed frame is skipped silently. The relay
 * path additionally wraps the call in try/catch, so a tap bug can never break
@@ -3513,7 +3537,7 @@ function createAnthropicAicTap(record) {
 		try {
 			const parsed = JSON.parse(dataLines.join("\n"));
 			const usage = extractCopilotUsage(parsed.copilot_usage);
-			if (usage) {
+			if (usage && usage.totalNanoAiu > 0) {
 				recorded = true;
 				try {
 					record(usage);
@@ -4200,6 +4224,72 @@ function parseOutputConfigEffort(body, model) {
 	return clampEffort(known ? effort : UNKNOWN_EFFORT_ANCHOR, supported);
 }
 /**
+* True for a `user` message carrying ONLY image parts.
+*
+* These are the synthetic follow-ups `anthropicMessageToNeutral` emits for
+* images found inside `tool_result` blocks (a `function_call_output` / chat
+* `tool` item cannot carry pixels). A genuine image-only user turn also
+* matches, but such a turn never appears mid-tool-run from Claude Code — and
+* even if it did, deferring it past the tool block is still the valid shape.
+*/
+function isImageOnlyUserMessage(m) {
+	return m.role === "user" && Array.isArray(m.content) && m.content.length > 0 && m.content.every((c) => c.type === "image");
+}
+/**
+* Hoist `tool_result`-extracted images to the END of each contiguous run of
+* tool results.
+*
+* A tool-output wire item cannot carry an image on any endpoint, so images
+* have to ride in a separate user message. Fanning each `toolResult` out to
+* `[tool, user]` independently is WRONG for parallel tool calls: it yields
+* `tool A, user A, tool B, user B`, and providers require the tool messages
+* answering one assistant turn to be CONTIGUOUS — the interjected user message
+* orphans the following tool message and the request is rejected outright
+* (verified live on `gemini-3.8-flash` `/chat/completions`: interleaved →
+* `400 invalid_request_body`, batched → `200`; see also
+* `github/copilot-sdk#1922`). Mirrors `translateMessages` in
+* `src/lib/worker-agent/stream-fn.ts`, which implements the same invariant
+* for the worker path.
+*
+* So images accumulate across the run and flush once, after the last tool
+* message of that run: `tool A, tool B, user(imgA, imgB)`. Any other message
+* ends the run and the pending images flush before it. Runs at parse time
+* over the FULL neutral sequence (a run routinely spans several Anthropic
+* messages — one `tool_result` per message), never written back to state.
+* Single-image and image-free inputs are returned unchanged in shape.
+*/
+function groupToolResultImages(messages) {
+	const out = [];
+	let pending = [];
+	let inToolRun = false;
+	const flushImages = () => {
+		if (pending.length === 0) return;
+		out.push({
+			role: "user",
+			content: pending
+		});
+		pending = [];
+	};
+	for (const m of messages) {
+		if (m.role === "toolResult") {
+			out.push(m);
+			inToolRun = true;
+			continue;
+		}
+		if (inToolRun && isImageOnlyUserMessage(m)) {
+			pending.push(...m.content);
+			continue;
+		}
+		if (inToolRun) {
+			flushImages();
+			inToolRun = false;
+		}
+		out.push(m);
+	}
+	flushImages();
+	return out;
+}
+/**
 * Parse an already-JSON-parsed Anthropic Messages body into the neutral shape.
 * `resolvedModel` is the catalog id the request will run on; `model` its
 * catalog entry (for the reasoning-effort allowlist).
@@ -4218,7 +4308,7 @@ function parseAnthropicRequest(body, resolvedModel, model) {
 		model: resolvedModel,
 		instructions: system.stable,
 		dynamicInstructions: system.dynamic,
-		messages,
+		messages: groupToolResultImages(messages),
 		tools,
 		toolChoice: parseToolChoice(body.tool_choice),
 		parallelToolCalls: parseDisableParallelToolUse(body.tool_choice),
@@ -4550,7 +4640,7 @@ async function* synthAnthropicFromChat(upstream, opts) {
 			continue;
 		}
 		if (!aicRecorded) {
-			if (extractAndRecordAic(opts.modelId, chunk) !== void 0) aicRecorded = true;
+			if (extractAndRecordPricedAic(opts.modelId, chunk) !== void 0) aicRecorded = true;
 		}
 		if (chunk.usage) {
 			usageIn = Math.max(usageIn, chunk.usage.prompt_tokens ?? 0);
@@ -4963,7 +5053,7 @@ async function* synthAnthropicFromResponses(upstream, opts) {
 			case "response.incomplete": {
 				sawTerminal = true;
 				if (!aicRecorded) {
-					if (extractAndRecordAic(opts.modelId, ev) !== void 0) aicRecorded = true;
+					if (extractAndRecordPricedAic(opts.modelId, ev) !== void 0) aicRecorded = true;
 				}
 				const u = ev.response?.usage;
 				if (u) {
@@ -6857,13 +6947,33 @@ async function handleResponses(c) {
 	let pendingFirstChunk = firstChunk;
 	let consumerCancelled = false;
 	let aicRecorded = false;
+	let streamAicNano;
+	let completionLogged = false;
 	const tapAic = (chunk) => {
 		if (aicRecorded || !chunk.data || chunk.data === "[DONE]") return;
 		try {
 			const parsed = JSON.parse(chunk.data);
-			if (extractAndRecordAic(resolvedModel, parsed) !== void 0) aicRecorded = true;
+			const nano = extractAndRecordPricedAic(resolvedModel, parsed);
+			if (nano !== void 0) {
+				aicRecorded = true;
+				streamAicNano = nano;
+			}
 		} catch {}
 	};
+	const logStreamCompletion = () => {
+		if (completionLogged || consumerCancelled || streamAicNano === void 0) return;
+		completionLogged = true;
+		logRequest({
+			method: "POST",
+			path: c.req.path,
+			model: originalModel,
+			resolvedModel,
+			aicNano: streamAicNano,
+			status: 200,
+			streaming: true,
+			streamMs: Date.now() - startTime
+		}, selectedModel, startTime);
+	};
 	if (pendingFirstChunk !== void 0) tapAic(pendingFirstChunk);
 	const safeClose = (controller) => {
 		try {
@@ -6908,12 +7018,14 @@ async function handleResponses(c) {
 				}
 				if (result.done) {
 					upstreamFinished = true;
+					logStreamCompletion();
 					safeClose(controller);
 					return;
 				}
 				if (result.value === void 0 || result.value === null) return;
 				if (result.value.data === "[DONE]") {
 					upstreamFinished = true;
+					logStreamCompletion();
 					safeClose(controller);
 					return;
 				}
@@ -7012,12 +7124,14 @@ async function handleResponsesCompact(c) {
 	});
 	const response = await fetchWithTransientRetry(() => tryRefreshAndRetry(doFetch, "/responses/compact"), { label: "/responses/compact" });
 	if (response.ok) {
+		const json = await response.json();
+		extractAndRecordAic(resolveModel(body.model), json);
 		logRequest({
 			method: "POST",
 			path: c.req.path,
 			status: 200
 		}, void 0, startTime);
-		return c.json(await response.json());
+		return c.json(json);
 	}
 	if (response.status === 404) {
 		consola.debug("Copilot API does not support /responses/compact, using synthetic compaction");
@@ -7064,6 +7178,7 @@ async function syntheticCompact(c, body, startTime) {
 		}, void 0, startTime);
 		throw error;
 	}
+	extractAndRecordAic(resolveModel(body.model), result);
 	logRequest({
 		method: "POST",
 		path: c.req.path,
@@ -7786,4 +7901,4 @@ function getCodexEnvVars(serverUrl) {
 //#endregion
 export { validateCheapProfilePrerequisites as A, formatCheap1mPrerequisiteFailure as C, profileDescriptor as D, formatFastPrerequisiteFailure as E, checkClaudeVersion as F, updateClaude as I, validateFastProfilePrerequisites as M, validateMaxProfileLaunch as N, resolveLaunchProfile as O, runSelfUpdate as P, LUNA_SCOUT_ALIAS_ID as S, formatCheapestPrerequisiteFailure as T, CHEAP_EXPLORE_ALIAS_ID as _, sharedServerArgs as a, CHEAP_PLAN_ALIAS_ID as b, stopKeepAwake as c, enableFileLogging as d, CHEAPEST_EXPLORE_ALIAS_ID as f, CHEAPEST_REVIEWER_ALIAS_ID as g, CHEAPEST_PLAN_ALIAS_ID as h, setupAndServe as i, validateCheapestProfilePrerequisites as j, validateCheap1mProfilePrerequisites as k, injectModelPickerSettingsFile as l, CHEAPEST_IMPLEMENTER_ALIAS_ID as m, getCodexEnvVars as n, LAUNCH_SECRET_HEADER as o, CHEAPEST_GENERAL_PURPOSE_ALIAS_ID as p, parseSharedArgs as r, startKeepAwake as s, getClaudeCodeEnvVars as t, listModelsForEndpoint as u, CHEAP_GENERAL_PURPOSE_ALIAS_ID as v, formatCheapPrerequisiteFailure as w, CHEAP_REVIEWER_ALIAS_ID as x, CHEAP_IMPLEMENTER_ALIAS_ID as y };
 
-//# sourceMappingURL=server-setup-DPay_uYp.js.map
+//# sourceMappingURL=server-setup-CXXKmPIE.js.map