crosscheck-mcp 0.2.21 → 0.2.22

package/dist/node-stdio.js

@@ -1176,7 +1176,27 @@ var PROVIDER_CAPS = {
     // temperature, so the prefix is the fix.
     reasoning_prefixes: ["gpt-5", "gpt-6", "o1", "o3", "o4"]
   },
-  xai: { family: "openai_chat", system_role: "inline", supports_temperature: true },
+  xai: {
+    family: "openai_chat",
+    system_role: "inline",
+    // Kept `true`, unlike other reasoning models: grok-4 accepts temperature
+    // (verified 200 at 0.4 and 1.0) and a panel wants the variance. Dropping
+    // it would buy nothing and cost diversity.
+    supports_temperature: true,
+    // grok-4 IS reasoning-class — it emitted 1,432 reasoning tokens against 64
+    // visible on one call. Saying so strips the "think step by step" preambles
+    // it does not need, and lifts it off the non-reasoning ceiling (1500) onto
+    // the reasoning-safe one (2048).
+    //
+    // NOT the reason grok answers short, which was the initial guess and was
+    // wrong: at caps of 1500, 2048 and 6144 it returned 445, 404 and 461
+    // visible tokens, finish_reason=stop every time. The cap was never
+    // binding; grok is simply terse. The ceiling change removes a latent
+    // constraint, it does not make grok say more.
+    reasoning_prefixes: ["grok-4"],
+    // Measured: xAI's cap bounds visible output only. See the field docs.
+    reasoning_shares_output_budget: false
+  },
   mistral: { family: "openai_chat", system_role: "inline", supports_temperature: true },
   groq: { family: "openai_chat", system_role: "inline", supports_temperature: true },
   deepseek: { family: "openai_chat", system_role: "inline", supports_temperature: true },
@@ -1215,6 +1235,10 @@ function supportsTemperature(provider, model) {
   if (caps.supports_temperature === "model") return !isReasoningModel(provider, model);
   return Boolean(caps.supports_temperature);
 }
+function reasoningSharesOutputBudget(provider) {
+  const caps = PROVIDER_CAPS[provider.toLowerCase()];
+  return caps?.reasoning_shares_output_budget ?? true;
+}
 
 // src/providers/types.ts
 init_esm_shims();
@@ -1467,6 +1491,23 @@ async function acquireRateLimit(provider, deps) {
 var ANTHROPIC_API_URL = "https://api.anthropic.com/v1/messages";
 var ANTHROPIC_VERSION_HEADER = "2023-06-01";
 var ANTHROPIC_STRUCTURED_TOOL_NAME = "structured_output";
+function applyPromptCaching(body, env = process.env) {
+  const flag = (env["CROSSCHECK_ANTHROPIC_PROMPT_CACHE"] ?? "").trim();
+  if (flag === "0" || flag.toLowerCase() === "false") return;
+  if (typeof body.system === "string" && body.system !== "") {
+    body.system = [{ type: "text", text: body.system, cache_control: { type: "ephemeral" } }];
+  }
+  if (body.messages.length >= 2) {
+    const idx = body.messages.length - 2;
+    const m = body.messages[idx];
+    if (typeof m.content === "string" && m.content !== "") {
+      body.messages[idx] = {
+        role: m.role,
+        content: [{ type: "text", text: m.content, cache_control: { type: "ephemeral" } }]
+      };
+    }
+  }
+}
 function buildAnthropicRequest(opts) {
   let system;
   const convo = [];
@@ -1491,6 +1532,7 @@ function buildAnthropicRequest(opts) {
   if (system !== void 0) {
     body.system = system;
   }
+  applyPromptCaching(body);
   if (isReasoningModel("anthropic", opts.model)) {
     if (opts.model.toLowerCase().startsWith("claude-fable-5")) {
       body.thinking = { type: "adaptive" };
@@ -1550,6 +1592,7 @@ function parseAnthropicResponse(opts) {
   const prompt = Math.trunc(Number(u["input_tokens"] ?? 0)) || 0;
   const cached = Math.trunc(Number(u["cache_read_input_tokens"] ?? 0)) || 0;
   const completion = Math.trunc(Number(u["output_tokens"] ?? 0)) || 0;
+  const cacheWrite = Math.trunc(Number(u["cache_creation_input_tokens"] ?? 0)) || 0;
   const usage = {
     provider: "anthropic",
     model: opts.model,
@@ -1557,7 +1600,7 @@ function parseAnthropicResponse(opts) {
     // production helper folds them in so prompt_tokens is the FULL
     // input volume; calculateCost then bills the cached subset at the
     // cached rate (cached <= prompt_tokens, since prompt = input+cached).
-    prompt_tokens: prompt + cached,
+    prompt_tokens: prompt + cached + cacheWrite,
     completion_tokens: completion,
     cached_tokens: cached,
     total_tokens: 0,
@@ -1909,16 +1952,21 @@ function parseOpenAICompatibleResponse(opts) {
   const u = r["usage"] ?? {};
   const details = u["prompt_tokens_details"] ?? {};
   const cached = Math.trunc(Number(details["cached_tokens"] ?? 0)) || 0;
+  const promptTokens = Math.trunc(Number(u["prompt_tokens"] ?? 0)) || 0;
+  const reportedCompletion = Math.trunc(Number(u["completion_tokens"] ?? 0)) || 0;
+  const reportedTotal = Math.trunc(Number(u["total_tokens"] ?? 0)) || 0;
+  const derivedCompletion = reportedTotal > promptTokens ? reportedTotal - promptTokens : reportedCompletion;
+  const completionTokens = Math.max(reportedCompletion, derivedCompletion);
   const usage = {
     provider: opts.provider,
     model: opts.model,
-    prompt_tokens: Math.trunc(Number(u["prompt_tokens"] ?? 0)) || 0,
-    completion_tokens: Math.trunc(Number(u["completion_tokens"] ?? 0)) || 0,
+    prompt_tokens: promptTokens,
+    completion_tokens: completionTokens,
     cached_tokens: cached,
     // Use the response's reported total_tokens directly. Python's
     // `Usage.to_dict()` falls back to prompt+completion only when
     // total_tokens is 0/missing; mirror that.
-    total_tokens: Math.trunc(Number(u["total_tokens"] ?? 0)) || 0,
+    total_tokens: reportedTotal,
     cost_usd: 0,
     estimated: Object.keys(u).length === 0,
     purpose: opts.purpose
@@ -1960,7 +2008,11 @@ async function sendOpenAICompatible(args) {
       body.reasoning_effort = effort;
     }
   }
-  if (isReasoningModel(args.provider, args.model) && typeof body.max_completion_tokens === "number") {
+  if (isReasoningModel(args.provider, args.model) && // Only where the cap is a SHARED budget. On a provider that bounds visible
+  // output only (xAI), headroom cannot protect the answer from being crowded
+  // out — nothing is crowding it — so it would merely authorise a
+  // 25,000-token reply and the bill that comes with it.
+  reasoningSharesOutputBudget(args.provider) && typeof body.max_completion_tokens === "number") {
     const raw = Number(process.env["CROSSCHECK_OPENAI_REASONING_HEADROOM_TOKENS"]);
     const headroom = Number.isFinite(raw) && raw >= 0 ? Math.trunc(raw) : 25e3;
     body.max_completion_tokens += headroom;
@@ -2279,7 +2331,7 @@ import { z } from "zod";
 // src/server-meta.ts
 init_esm_shims();
 var SERVER_NAME = "crosscheck-agent";
-var SERVER_VERSION = true ? "0.2.21" : "0.0.0-dev";
+var SERVER_VERSION = true ? "0.2.22" : "0.0.0-dev";
 
 // src/tools/audit.ts
 init_esm_shims();
@@ -12788,7 +12840,7 @@ var CHECK_INTERVAL_SECONDS = 3 * 24 * 60 * 60;
 var DEFAULT_PACKAGE = "crosscheck-cli";
 var FETCH_TIMEOUT_MS = 3e3;
 function engineVersion() {
-  return true ? "0.2.21" : "0.0.0-dev";
+  return true ? "0.2.22" : "0.0.0-dev";
 }
 function defaultUpdateCachePath() {
   const base = process.env["CROSSCHECK_DATA_DIR"] || path10.join(os.homedir() || os.tmpdir(), ".crosscheck");
@@ -14832,9 +14884,7 @@ function usageEventsFromEnvelope(out, toolName) {
   const timing = out["timing"];
   const timingByCall = timing && typeof timing === "object" ? timing["by_call"] : void 0;
   const allTiming = Array.isArray(timingByCall) ? timingByCall : [];
-  const latencies = allTiming.filter(
-    (t) => !t?.["error_kind"]
-  );
+  const latencies = allTiming;
   const aligned = latencies.length === byCall.length;
   const pattern = toolToPattern(toolName);
   const events = [];
@@ -14844,13 +14894,18 @@ function usageEventsFromEnvelope(out, toolName) {
     if (!provider) continue;
     const model = String(u.model ?? "").slice(0, 128) || "unknown";
     const cost = Math.max(0, Number(u.cost_usd) || 0);
-    const latencyMs = aligned ? intNonNeg(latencies[i]?.["wall_ms"]) : 0;
+    const timingRow = aligned ? latencies[i] : void 0;
+    if (timingRow?.["error_kind"]) continue;
+    const promptTokens = intNonNeg(u.prompt_tokens);
+    const completionTokens = intNonNeg(u.completion_tokens);
+    if (promptTokens === 0 && completionTokens === 0) continue;
+    const latencyMs = intNonNeg(timingRow?.["wall_ms"]);
     events.push({
       provider,
       model,
       pattern,
-      promptTokens: intNonNeg(u.prompt_tokens),
-      completionTokens: intNonNeg(u.completion_tokens),
+      promptTokens,
+      completionTokens,
       costUsdEstimate: cost,
       latencyMs,
       status: "ok",