crosscheck-mcp 0.2.19 → 0.2.21

package/dist/browser-ext.js

@@ -11,8 +11,8 @@ var pricing_default = {
   _meta: {
     currency: "USD",
     unit: "per_1k_tokens",
-    updated_at: "2026-08-04",
-    notes: "Prices reflect public list pricing per provider. cached_per_1k applies to prompt tokens served from prompt-cache (Anthropic) or input-cache (OpenAI) when supported. Update when providers change rates. Missing models fall back to cost=0 with estimated=true. 2026-06-07: removed decommissioned gemini-2.0-flash / gemini-2.0-flash-lite (Google returns 404) and replaced the low-tier lite model with gemini-2.5-flash-lite. 2026-06-08: bumped the default Anthropic model to claude-opus-4-8. Kept gemini-2.5-pro as the Gemini default (gemini-3.1-pro-preview measured ~2.2x the cost at $2/$12 per M \u2014 output-token dominated by ~1.5k thinking tokens \u2014 so not worth it for a 1-of-4 panel voice). 2026-06-09: corrected rates verified against live provider pricing pages \u2014 gemini-2.5-pro output $5->$10/M; gemini-2.5-flash $0.15/$0.60 -> $0.30/$2.50; gpt-5 $5/$15 -> $1.25/$10 (cached $0.125); grok-4-latest input $5->$3/M (legacy grok-4 tier). Anthropic opus/sonnet/haiku rates confirmed unchanged. Older/unused models (gpt-4o, o1, o3, grok-3) not all re-verified. 2026-07-02: added claude-fable-5 as the optional reasoning-upgrade model (used only for the lead reasoning seat when the operator bumps a heavy task). 2026-07-04: set claude-fable-5 to its real pay-as-you-go API price effective 2026-07-07 \u2014 $10/M input ($0.01/1k), $50/M output ($0.05/1k), cached at 10% of input ($0.001/1k). Dropped the estimated flag. Note: Fable 5 is cheaper than opus-4-8 on both input and output. 2026-07-09: Google retired gemini-2.5-pro AND gemini-3-pro-preview (both now 404 'no longer available'). Moved the Gemini default to gemini-3.1-pro-preview ($2/M in, $12/M out, cache $0.20/M \u2014 the newest working pro; note it's a preview and may need another bump) and added gemini-3.5-flash ($1.50/M in, $9/M out, cache $0.15/M \u2014 GA) to the low/cheap tier. Removed the dead gemini-2.5-pro rate. IMPORTANT: gemini-3.x are heavy thinking models \u2014 thoughtsTokenCount (billed as output) is now folded into completion_tokens in the adapter, so cost reflects reasoning tokens. 2026-07-13: bumped the default OpenAI model to gpt-5.5 ($5/M in, $30/M out, cached $0.50/M \u2014 verified live). Note gpt-5.5 is markedly pricier than gpt-5 ($1.25/$10): ~4x input, 3x output. Kept gpt-5 in the table for cost lookups on pinned installs. 2026-07-16: added gpt-5.6 at the Sol (flagship) tier \u2014 $5/M in, $30/M out, cached $0.50/M (cache-write $6.25/M not modeled). gpt-5.6 is the second premium voice for the max-reasoning co-reasoner (runs alongside claude-fable-5 when an OpenAI key is set). Terra/Luna tiers not added (unused). 2026-07-17: added Moonshot AI (Kimi) provider with kimi-k3, base URL https://api.moonshot.ai/v1. Confirmed real rates \u2014 $3.00/M input (cache miss, $0.003/1k), $15.00/M output ($0.015/1k), $0.30/M cache-hit input ($0.0003/1k); 1M-token context. Kimi K3 is reasoning-class (temperature must be 1; uses max_completion_tokens). 2026-07-24: added claude-opus-5 ($5/M in, $25/M out, cached $0.50/M -- 10% of input, matching this table's existing cache-discount ratio) and made it the default anthropic model (was claude-opus-4-8). Fable 5 remains the reasoning-upgrade/super-mode target, unchanged. 2026-08-04: added Alibaba Cloud Qwen provider (OpenAI-compatible, DashScope international endpoint https://dashscope-intl.aliyuncs.com/compatible-mode/v1) with default model qwen3.8-max -- confirmed business pricing $2.00/M input ($0.002/1k), $6.00/M output ($0.006/1k), cache estimated at 10% of input ($0.0002/1k, not independently confirmed). Also added qwen3.7-max as a fallback tier at $1.25/M in ($0.00125/1k) / $3.75/M out ($0.00375/1k). Note: the model string initially configured (qwen-v3.8 / qwen-v3.7) does not exist on DashScope and 404s -- live-probed the real key against qwen3.8-max, qwen3.7-max, qwen3-max, qwen-max, and qwen-plus, all of which work; corrected the default and pricing keys to qwen3.8-max/qwen3.7-max and fixed QWEN_MODEL in both .env files to match. Qwen's Max tier is a standard instruct model, not reasoning-class -- normal temperature/max_tokens handling."
+    updated_at: "2026-09-06",
+    notes: "Prices reflect public list pricing per provider. cached_per_1k applies to prompt tokens served from prompt-cache (Anthropic) or input-cache (OpenAI) when supported. Update when providers change rates. Missing models fall back to cost=0 with estimated=true. 2026-06-07: removed decommissioned gemini-2.0-flash / gemini-2.0-flash-lite (Google returns 404) and replaced the low-tier lite model with gemini-2.5-flash-lite. 2026-06-08: bumped the default Anthropic model to claude-opus-4-8. Kept gemini-2.5-pro as the Gemini default (gemini-3.1-pro-preview measured ~2.2x the cost at $2/$12 per M \u2014 output-token dominated by ~1.5k thinking tokens \u2014 so not worth it for a 1-of-4 panel voice). 2026-06-09: corrected rates verified against live provider pricing pages \u2014 gemini-2.5-pro output $5->$10/M; gemini-2.5-flash $0.15/$0.60 -> $0.30/$2.50; gpt-5 $5/$15 -> $1.25/$10 (cached $0.125); grok-4-latest input $5->$3/M (legacy grok-4 tier). Anthropic opus/sonnet/haiku rates confirmed unchanged. Older/unused models (gpt-4o, o1, o3, grok-3) not all re-verified. 2026-07-02: added claude-fable-5 as the optional reasoning-upgrade model (used only for the lead reasoning seat when the operator bumps a heavy task). 2026-07-04: set claude-fable-5 to its real pay-as-you-go API price effective 2026-07-07 \u2014 $10/M input ($0.01/1k), $50/M output ($0.05/1k), cached at 10% of input ($0.001/1k). Dropped the estimated flag. Note: Fable 5 is cheaper than opus-4-8 on both input and output. 2026-07-09: Google retired gemini-2.5-pro AND gemini-3-pro-preview (both now 404 'no longer available'). Moved the Gemini default to gemini-3.1-pro-preview ($2/M in, $12/M out, cache $0.20/M \u2014 the newest working pro; note it's a preview and may need another bump) and added gemini-3.5-flash ($1.50/M in, $9/M out, cache $0.15/M \u2014 GA) to the low/cheap tier. Removed the dead gemini-2.5-pro rate. IMPORTANT: gemini-3.x are heavy thinking models \u2014 thoughtsTokenCount (billed as output) is now folded into completion_tokens in the adapter, so cost reflects reasoning tokens. 2026-07-13: bumped the default OpenAI model to gpt-5.5 ($5/M in, $30/M out, cached $0.50/M \u2014 verified live). Note gpt-5.5 is markedly pricier than gpt-5 ($1.25/$10): ~4x input, 3x output. Kept gpt-5 in the table for cost lookups on pinned installs. 2026-07-16: added gpt-5.6 at the Sol (flagship) tier \u2014 $5/M in, $30/M out, cached $0.50/M (cache-write $6.25/M not modeled). gpt-5.6 is the second premium voice for the max-reasoning co-reasoner (runs alongside claude-fable-5 when an OpenAI key is set). Terra/Luna tiers not added (unused). 2026-07-17: added Moonshot AI (Kimi) provider with kimi-k3, base URL https://api.moonshot.ai/v1. Confirmed real rates \u2014 $3.00/M input (cache miss, $0.003/1k), $15.00/M output ($0.015/1k), $0.30/M cache-hit input ($0.0003/1k); 1M-token context. Kimi K3 is reasoning-class (temperature must be 1; uses max_completion_tokens). 2026-07-24: added claude-opus-5 ($5/M in, $25/M out, cached $0.50/M -- 10% of input, matching this table's existing cache-discount ratio) and made it the default anthropic model (was claude-opus-4-8). Fable 5 remains the reasoning-upgrade/super-mode target, unchanged. 2026-08-04: added Alibaba Cloud Qwen provider (OpenAI-compatible, DashScope international endpoint https://dashscope-intl.aliyuncs.com/compatible-mode/v1) with default model qwen3.8-max -- confirmed business pricing $2.00/M input ($0.002/1k), $6.00/M output ($0.006/1k), cache estimated at 10% of input ($0.0002/1k, not independently confirmed). Also added qwen3.7-max as a fallback tier at $1.25/M in ($0.00125/1k) / $3.75/M out ($0.00375/1k). Note: the model string initially configured (qwen-v3.8 / qwen-v3.7) does not exist on DashScope and 404s -- live-probed the real key against qwen3.8-max, qwen3.7-max, qwen3-max, qwen-max, and qwen-plus, all of which work; corrected the default and pricing keys to qwen3.8-max/qwen3.7-max and fixed QWEN_MODEL in both .env files to match. Qwen's Max tier is a standard instruct model, not reasoning-class -- normal temperature/max_tokens handling. 2026-09-06: added gpt-6-astra at published list rates ($10/$50 per Mtok, $1/Mtok cache read; the $12.50/Mtok cache-WRITE rate has no field in this schema and is not modelled). gemini-3.8-flash now carries Google's published paid-tier rates ($0.75/$3.75/$0.075 per Mtok). THESE DOUBLE ON 2027-01-01 ($1.50/$7.50/$0.15) and this schema has no effective-date support, so they must be updated by hand on that date or every Gemini cost will read half of actual. claude-fable-5-1 now carries published rates ($10/$50/$0.25 per Mtok). Note its cache-read rate is 4x cheaper than claude-fable-5's, so the two must not be assumed to track each other. No inferred rows remain. All three were previously absent, which silently recorded their cost as $0."
   },
   anthropic: {
     "claude-opus-4-5": {
@@ -64,6 +64,11 @@ var pricing_default = {
       prompt_per_1k: 8e-4,
       completion_per_1k: 4e-3,
       cached_per_1k: 8e-5
+    },
+    "claude-fable-5-1": {
+      prompt_per_1k: 0.01,
+      completion_per_1k: 0.05,
+      cached_per_1k: 25e-5
     }
   },
   openai: {
@@ -116,6 +121,11 @@ var pricing_default = {
       prompt_per_1k: 11e-4,
       completion_per_1k: 44e-4,
       cached_per_1k: 55e-5
+    },
+    "gpt-6-astra": {
+      prompt_per_1k: 0.01,
+      completion_per_1k: 0.05,
+      cached_per_1k: 1e-3
     }
   },
   xai: {
@@ -155,6 +165,11 @@ var pricing_default = {
       prompt_per_1k: 1e-4,
       completion_per_1k: 4e-4,
       cached_per_1k: 25e-6
+    },
+    "gemini-3.8-flash": {
+      prompt_per_1k: 75e-5,
+      completion_per_1k: 375e-5,
+      cached_per_1k: 75e-6
     }
   },
   mistral: {
@@ -390,7 +405,13 @@ var PROVIDER_CAPS = {
     family: "openai_chat",
     system_role: "inline",
     supports_temperature: "model",
-    reasoning_prefixes: ["gpt-5", "o1", "o3", "o4"]
+    // gpt-6 added 2026-09-06: GPT-6 Astra rejects `max_tokens` outright
+    // ("Unsupported parameter ... use max_completion_tokens"), which is a 400,
+    // NOT a model-access failure — so the fallback chain correctly declines to
+    // fire and the provider silently drops out of the panel instead. Reasoning
+    // -class routing is what selects max_completion_tokens and omits
+    // 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 },
   mistral: { family: "openai_chat", system_role: "inline", supports_temperature: true },
@@ -445,12 +466,20 @@ var ProviderError = class extends Error {
    *  actually fix. Distinct from `transient`: this is never worth
    *  retrying the SAME model, but IS worth trying a different one. */
   modelAccessFailure;
+  /** True for "the key works, the account is out of credit / over a spend
+   *  limit." Rides alongside kind "auth" rather than replacing it, because
+   *  every retry and fallback decision downstream is already tuned to that
+   *  kind — but the two are completely different things to a human, and
+   *  reporting "API key rejected" for an unpaid invoice sends people to
+   *  regenerate a key that was never the problem. */
+  billing;
   constructor(kind, message, opts) {
     super(message);
     this.kind = kind;
     this.status = opts?.status;
     this.transient = opts?.transient ?? defaultTransient(kind);
     this.modelAccessFailure = opts?.modelAccessFailure ?? false;
+    this.billing = opts?.billing ?? false;
     if (opts?.retryAfterS !== void 0) {
       this.retryAfterS = opts.retryAfterS;
     }
@@ -488,7 +517,7 @@ function httpFailureToProviderError(provider, status, bodyText, retryAfterS, mod
     return new ProviderError(
       "auth",
       `${provider}: out of credits or billing isn't set up (HTTP ${status}). Your API key is reaching ${provider}, but the account has no usable balance \u2014 add credits / enable billing in your ${provider} console, then retry. Detail: ${detail}`,
-      { status }
+      { status, billing: true }
     );
   }
   if (status === 401 || status === 403) {
@@ -1330,7 +1359,7 @@ import { z } from "zod";
 
 // src/server-meta.ts
 var SERVER_NAME = "crosscheck-agent";
-var SERVER_VERSION = true ? "0.2.19" : "0.0.0-dev";
+var SERVER_VERSION = true ? "0.2.21" : "0.0.0-dev";
 
 // src/tools/audit.ts
 import { readdirSync, readFileSync as readFileSync3, statSync } from "fs";
@@ -1836,7 +1865,9 @@ async function attachUsageBlock(result, answers, opts) {
       purpose: a.usage?.purpose ?? null,
       wall_ms: a.elapsed_ms ?? 0,
       cpu_ms: a.cpu_ms ?? 0,
-      cache_hit: Boolean(a.cache_hit)
+      cache_hit: Boolean(a.cache_hit),
+      ...a.error_kind ? { error_kind: a.error_kind } : {},
+      ...a.error_billing ? { error_billing: true } : {}
     }))
   };
   result["timing"] = timing;
@@ -2137,11 +2168,13 @@ async function askOne(provider, messages, opts) {
     const cpuMs = Math.trunc((cpu.user + cpu.system) / 1e3);
     const kind = e instanceof ProviderError ? e.kind : "other";
     const msg = e instanceof Error ? e.message : String(e);
+    const billing = e instanceof ProviderError && e.billing;
     return {
       provider: provider.name,
       model: provider.model,
       error: msg,
       error_kind: kind,
+      ...billing ? { error_billing: true } : {},
       attempts: 0,
       usage: emptyUsage(provider.name, provider.model, opts.purpose),
       cache_hit: false,
@@ -11749,7 +11782,7 @@ var CHECK_INTERVAL_SECONDS = 3 * 24 * 60 * 60;
 var DEFAULT_PACKAGE = "crosscheck-cli";
 var FETCH_TIMEOUT_MS = 3e3;
 function engineVersion() {
-  return true ? "0.2.19" : "0.0.0-dev";
+  return true ? "0.2.21" : "0.0.0-dev";
 }
 function defaultUpdateCachePath() {
   const base = process.env["CROSSCHECK_DATA_DIR"] || path9.join(os.homedir() || os.tmpdir(), ".crosscheck");