crosscheck-mcp 0.2.10 → 0.2.12

package/dist/browser-ext.cjs

@@ -44,8 +44,8 @@ var pricing_default = {
   _meta: {
     currency: "USD",
     unit: "per_1k_tokens",
-    updated_at: "2026-06-07",
-    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."
+    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."
   },
   anthropic: {
     "claude-opus-4-5": {
@@ -233,6 +233,18 @@ var pricing_default = {
       cached_per_1k: 3e-4
     }
   },
+  qwen: {
+    "qwen3.8-max": {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 6e-3,
+      cached_per_1k: 2e-4
+    },
+    "qwen3.7-max": {
+      prompt_per_1k: 125e-5,
+      completion_per_1k: 375e-5,
+      cached_per_1k: 125e-6
+    }
+  },
   _tiers: {
     low: {
       description: "Cheap, fast models for simple subtasks (extraction, formatting, short summaries).",
@@ -430,7 +442,11 @@ var PROVIDER_CAPS = {
     system_role: "separate",
     supports_temperature: true,
     reasoning_prefixes: ["gemini-3.1-pro", "gemini-3.5-flash", "gemini-2.5-pro"]
-  }
+  },
+  // Qwen's "Max" tier (qwen-v3.8, qwen-v3.7) is a standard instruct/chat
+  // model, not a visible-reasoning model like QwQ — normal temperature and
+  // max_tokens handling, same as xai/mistral/groq/deepseek.
+  qwen: { family: "openai_chat", system_role: "inline", supports_temperature: true }
 };
 function isReasoningModel(provider, model) {
   const caps = PROVIDER_CAPS[provider];
@@ -1001,7 +1017,12 @@ var OPENAI_COMPAT_DEFAULT_URLS = {
   groq: "https://api.groq.com/openai/v1/chat/completions",
   deepseek: "https://api.deepseek.com/v1/chat/completions",
   // Moonshot AI (Kimi). OpenAI-compatible chat completions endpoint.
-  kimi: "https://api.moonshot.ai/v1/chat/completions"
+  kimi: "https://api.moonshot.ai/v1/chat/completions",
+  // Alibaba Cloud DashScope (Qwen). International endpoint — the mainland
+  // China endpoint (dashscope.aliyuncs.com) 403s/404s from outside China in
+  // a way that can look like a bad key; this is the deliberate default for
+  // a US-based deployment.
+  qwen: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions"
 };
 var OPENAI_COMPAT_NATIVE_STRUCTURED = /* @__PURE__ */ new Set([
   "openai"
@@ -1170,7 +1191,8 @@ var DEFAULT_MODELS = {
   groq: "llama-3.3-70b-versatile",
   deepseek: "deepseek-chat",
   gemini: "gemini-3.1-pro-preview",
-  kimi: "kimi-k3"
+  kimi: "kimi-k3",
+  qwen: "qwen3.8-max"
 };
 function buildProviders(opts) {
   const out = {};
@@ -1185,7 +1207,8 @@ function buildProviders(opts) {
     { name: "mistral", keyEnv: "MISTRAL_API_KEY", modelEnv: "MISTRAL_MODEL", defaultModel: DEFAULT_MODELS["mistral"] },
     { name: "groq", keyEnv: "GROQ_API_KEY", modelEnv: "GROQ_MODEL", defaultModel: DEFAULT_MODELS["groq"] },
     { name: "deepseek", keyEnv: "DEEPSEEK_API_KEY", modelEnv: "DEEPSEEK_MODEL", defaultModel: DEFAULT_MODELS["deepseek"] },
-    { name: "kimi", keyEnv: "KIMI_API_KEY", modelEnv: "KIMI_MODEL", defaultModel: DEFAULT_MODELS["kimi"] }
+    { name: "kimi", keyEnv: "KIMI_API_KEY", modelEnv: "KIMI_MODEL", defaultModel: DEFAULT_MODELS["kimi"] },
+    { name: "qwen", keyEnv: "QWEN_API_KEY", modelEnv: "QWEN_MODEL", defaultModel: DEFAULT_MODELS["qwen"] }
   ];
   for (const s of openAiCompatSpec) {
     const apiKey = opts.env[s.keyEnv];
@@ -1257,7 +1280,7 @@ var import_zod = require("zod");
 
 // src/server-meta.ts
 var SERVER_NAME = "crosscheck-agent";
-var SERVER_VERSION = true ? "0.2.10" : "0.0.0-dev";
+var SERVER_VERSION = true ? "0.2.12" : "0.0.0-dev";
 
 // src/tools/audit.ts
 var import_node_fs4 = require("fs");
@@ -4106,7 +4129,9 @@ var SUPER_MODELS = {
   // already the default — no-op retarget
   gemini: DEFAULT_MODELS["gemini"],
   // already the default — no-op retarget
-  kimi: DEFAULT_MODELS["kimi"]
+  kimi: DEFAULT_MODELS["kimi"],
+  // already the default — no-op retarget
+  qwen: DEFAULT_MODELS["qwen"]
   // already the default — no-op retarget
 };
 var SUPER_PROVIDER_NAMES = Object.keys(SUPER_MODELS);
@@ -9892,7 +9917,8 @@ var KNOWN_PROVIDERS6 = [
   "mistral",
   "groq",
   "deepseek",
-  "kimi"
+  "kimi",
+  "qwen"
 ];
 var USAGE_HINT = "Pass a 'providers' array to confer/debate/plan/review to pick an ad-hoc subset, e.g. providers=['openai','gemini']. Omit the field to use the configured active set.";
 function runListProviders(args, opts) {
@@ -11464,7 +11490,7 @@ var CHECK_INTERVAL_SECONDS = 3 * 24 * 60 * 60;
 var DEFAULT_PACKAGE = "crosscheck-cli";
 var FETCH_TIMEOUT_MS = 3e3;
 function engineVersion() {
-  return true ? "0.2.10" : "0.0.0-dev";
+  return true ? "0.2.12" : "0.0.0-dev";
 }
 function defaultUpdateCachePath() {
   const base = process.env["CROSSCHECK_DATA_DIR"] || import_node_path12.default.join(import_node_os2.default.homedir() || import_node_os2.default.tmpdir(), ".crosscheck");
@@ -12288,8 +12314,24 @@ function registerCoreTools(opts = {}) {
       o.repoRoot,
       o.workerToolsCfg
     ),
-    triangulateTool(o.providers ?? {}, o.providerAllowlist ?? null, o.bridge),
-    planTool(o.providers ?? {}, o.providerAllowlist ?? null, o.bridge),
+    triangulateTool(
+      o.providers ?? {},
+      o.providerAllowlist ?? null,
+      o.bridge,
+      o.storage,
+      o.circuitBreakers,
+      o.transcriptsDir,
+      o.repoRoot
+    ),
+    planTool(
+      o.providers ?? {},
+      o.providerAllowlist ?? null,
+      o.bridge,
+      o.storage,
+      o.circuitBreakers,
+      o.transcriptsDir,
+      o.repoRoot
+    ),
     critiqueTool(
       o.providers ?? {},
       o.providerAllowlist ?? null,
@@ -12299,7 +12341,15 @@ function registerCoreTools(opts = {}) {
       o.transcriptsDir,
       o.repoRoot
     ),
-    reviewTool(o.providers ?? {}, o.providerAllowlist ?? null, o.bridge),
+    reviewTool(
+      o.providers ?? {},
+      o.providerAllowlist ?? null,
+      o.bridge,
+      o.storage,
+      o.circuitBreakers,
+      o.transcriptsDir,
+      o.repoRoot
+    ),
     listProvidersTool(o.providers ?? {}, o.activeProviders ?? null, o.moderatorDefault ?? "anthropic"),
     recallTool(o.storage, o.bridge),
     sessionMemoryTool(o.storage, o.bridge),
@@ -12716,7 +12766,7 @@ function listProvidersTool(providers, activeProviders, moderatorDefault) {
     })
   };
 }
-function reviewTool(providers, allowlist, bridge) {
+function reviewTool(providers, allowlist, bridge, storage, breakers, transcriptsDir, repoRoot) {
   return {
     name: "review",
     description: "Have an LLM panel peer-review a code or proposal snippet. Returns the confer envelope (one answer per provider).",
@@ -12735,7 +12785,11 @@ function reviewTool(providers, allowlist, bridge) {
     handler: (args) => runReview(args, {
       providers,
       allowlist,
-      ...bridge ? { bridge } : {}
+      ...bridge ? { bridge } : {},
+      ...storage ? { storage } : {},
+      ...breakers ? { breakers } : {},
+      ...transcriptsDir ? { transcriptsDir } : {},
+      ...repoRoot ? { repoRoot } : {}
     })
   };
 }
@@ -12767,7 +12821,7 @@ function critiqueTool(providers, allowlist, bridge, storage, breakers, transcrip
     })
   };
 }
-function planTool(providers, allowlist, bridge) {
+function planTool(providers, allowlist, bridge, storage, breakers, transcriptsDir, repoRoot) {
   return {
     name: "plan",
     description: 'Have an LLM panel debate a step-by-step plan for the stated goal under the given constraints. Returns the debate envelope with a moderator-synthesised plan. Defaults to `mode: "fast"` (2 rounds + early_stop) \u2014 set `mode: "thorough"` for 5 rounds without early_stop. Explicit `max_rounds` / `early_stop` / `early_stop_threshold` always override the mode preset. Use `structured: true` for schema-validated synthesis.',
@@ -12794,11 +12848,15 @@ function planTool(providers, allowlist, bridge) {
     handler: (args) => runPlan(args, {
       providers,
       allowlist,
-      ...bridge ? { bridge } : {}
+      ...bridge ? { bridge } : {},
+      ...storage ? { storage } : {},
+      ...breakers ? { breakers } : {},
+      ...transcriptsDir ? { transcriptsDir } : {},
+      ...repoRoot ? { repoRoot } : {}
     })
   };
 }
-function triangulateTool(providers, allowlist, bridge) {
+function triangulateTool(providers, allowlist, bridge, storage, breakers, transcriptsDir, repoRoot) {
   return {
     name: "triangulate",
     description: "Run a coordinate flow and reshape the output as a consensus + minority report with per-provider weights. v1 uses 1.0 weights (matches a fresh provider_stats DB); future versions thread real win-rate weights when the DB layer ports.",
@@ -12819,7 +12877,11 @@ function triangulateTool(providers, allowlist, bridge) {
     handler: (args) => runTriangulate(args, {
       providers,
       allowlist,
-      ...bridge ? { bridge } : {}
+      ...bridge ? { bridge } : {},
+      ...storage ? { storage } : {},
+      ...breakers ? { breakers } : {},
+      ...transcriptsDir ? { transcriptsDir } : {},
+      ...repoRoot ? { repoRoot } : {}
     })
   };
 }
@@ -13085,7 +13147,8 @@ var KEY_ENV = {
   groq: "GROQ_API_KEY",
   deepseek: "DEEPSEEK_API_KEY",
   mistral: "MISTRAL_API_KEY",
-  kimi: "KIMI_API_KEY"
+  kimi: "KIMI_API_KEY",
+  qwen: "QWEN_API_KEY"
 };
 function createCrosscheck(opts) {
   const env = {};