crosscheck-mcp 0.2.30 → 0.2.31

package/dist/browser-ext.cjs

@@ -39,426 +39,35 @@ module.exports = __toCommonJS(browser_ext_exports);
 var getImportMetaUrl = () => typeof document === "undefined" ? new URL(`file:${__filename}`).href : document.currentScript && document.currentScript.tagName.toUpperCase() === "SCRIPT" ? document.currentScript.src : new URL("main.js", document.baseURI).href;
 var importMetaUrl = /* @__PURE__ */ getImportMetaUrl();
 
-// ../../config/pricing.json
-var pricing_default = {
-  _meta: {
-    currency: "USD",
-    unit: "per_1k_tokens",
-    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: {
-    _source: {
-      url: "https://platform.claude.com/docs/en/about-claude/pricing",
-      verified_at: "2026-09-27"
-    },
-    "claude-opus-4-5": {
-      prompt_per_1k: 5e-3,
-      completion_per_1k: 0.025,
-      cached_per_1k: 5e-4
-    },
-    "claude-opus-4-7": {
-      prompt_per_1k: 5e-3,
-      completion_per_1k: 0.025,
-      cached_per_1k: 5e-4
-    },
-    "claude-opus-4-8": {
-      prompt_per_1k: 5e-3,
-      completion_per_1k: 0.025,
-      cached_per_1k: 5e-4
-    },
-    "claude-opus-5-5": {
-      prompt_per_1k: 4e-3,
-      completion_per_1k: 0.02,
-      cached_per_1k: 2e-4
-    },
-    "claude-opus-5": {
-      prompt_per_1k: 5e-3,
-      completion_per_1k: 0.025,
-      cached_per_1k: 5e-4
-    },
-    "claude-fable-5": {
-      prompt_per_1k: 0.01,
-      completion_per_1k: 0.05,
-      cached_per_1k: 1e-3
-    },
-    "claude-sonnet-4-5": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      cached_per_1k: 3e-4
-    },
-    "claude-sonnet-4-6": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      cached_per_1k: 3e-4
-    },
-    "claude-haiku-4-5": {
-      prompt_per_1k: 1e-3,
-      completion_per_1k: 5e-3,
-      cached_per_1k: 1e-4
-    },
-    "claude-3-5-sonnet": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      cached_per_1k: 3e-4
-    },
-    "claude-3-5-haiku": {
-      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: {
-    _source: {
-      url: "https://developers.openai.com/api/docs/pricing",
-      verified_at: "2026-09-27"
-    },
-    "gpt-6-astra": {
-      prompt_per_1k: 0.01,
-      completion_per_1k: 0.05,
-      cached_per_1k: 1e-3
-    },
-    "gpt-6-sol": {
-      prompt_per_1k: 2e-3,
-      completion_per_1k: 0.01,
-      cached_per_1k: 2e-4
-    },
-    "gpt-6-luna": {
-      prompt_per_1k: 1e-4,
-      completion_per_1k: 5e-4,
-      cached_per_1k: 1e-5
-    },
-    "gpt-5.6-sol": {
-      prompt_per_1k: 4e-3,
-      completion_per_1k: 0.02,
-      cached_per_1k: 4e-4
-    },
-    "gpt-5.6-terra": {
-      prompt_per_1k: 2e-3,
-      completion_per_1k: 0.012,
-      cached_per_1k: 2e-4
-    },
-    "gpt-5.6-luna": {
-      prompt_per_1k: 2e-4,
-      completion_per_1k: 12e-4,
-      cached_per_1k: 2e-5
-    },
-    "gpt-5.6": {
-      prompt_per_1k: 4e-3,
-      completion_per_1k: 0.02,
-      cached_per_1k: 4e-4
-    },
-    "gpt-5.5": {
-      prompt_per_1k: 5e-3,
-      completion_per_1k: 0.03,
-      cached_per_1k: 5e-4
-    },
-    "gpt-5": {
-      prompt_per_1k: 125e-5,
-      completion_per_1k: 0.01,
-      cached_per_1k: 125e-6
-    },
-    "gpt-5-mini": {
-      prompt_per_1k: 25e-5,
-      completion_per_1k: 2e-3,
-      cached_per_1k: 25e-6
-    },
-    "gpt-5-nano": {
-      prompt_per_1k: 5e-5,
-      completion_per_1k: 4e-4,
-      cached_per_1k: 5e-6
-    },
-    "gpt-4o": {
-      prompt_per_1k: 25e-4,
-      completion_per_1k: 0.01,
-      cached_per_1k: 125e-5
-    },
-    "gpt-4o-mini": {
-      prompt_per_1k: 15e-5,
-      completion_per_1k: 6e-4,
-      cached_per_1k: 75e-6
-    },
-    o1: {
-      prompt_per_1k: 0.015,
-      completion_per_1k: 0.06,
-      cached_per_1k: 75e-4
-    },
-    "o1-mini": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.012,
-      cached_per_1k: 15e-4
-    },
-    o3: {
-      prompt_per_1k: 2e-3,
-      completion_per_1k: 8e-3,
-      cached_per_1k: 5e-4
-    },
-    "o3-mini": {
-      prompt_per_1k: 11e-4,
-      completion_per_1k: 44e-4,
-      cached_per_1k: 55e-5
-    },
-    "gpt-5-pro": {
-      prompt_per_1k: 0.015,
-      completion_per_1k: 0.12,
-      cached_per_1k: 0
-    },
-    "gpt-5.2-pro": {
-      prompt_per_1k: 0.021,
-      completion_per_1k: 0.168,
-      cached_per_1k: 0
-    },
-    "gpt-5.4-pro": {
-      prompt_per_1k: 0.03,
-      completion_per_1k: 0.18,
-      cached_per_1k: 0
-    },
-    "gpt-5.5-pro": {
-      prompt_per_1k: 0.03,
-      completion_per_1k: 0.18,
-      cached_per_1k: 0
-    },
-    "o1-pro": {
-      prompt_per_1k: 0.15,
-      completion_per_1k: 0.6,
-      cached_per_1k: 0
-    },
-    "o3-pro": {
-      prompt_per_1k: 0.02,
-      completion_per_1k: 0.08,
-      cached_per_1k: 0
-    }
-  },
-  xai: {
-    "grok-4-latest": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      cached_per_1k: 75e-5
-    },
-    "grok-3": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      cached_per_1k: 15e-4
-    },
-    "grok-3-mini": {
-      prompt_per_1k: 3e-4,
-      completion_per_1k: 5e-4,
-      cached_per_1k: 15e-5
-    }
-  },
-  gemini: {
-    "gemini-3.1-pro-preview": {
-      prompt_per_1k: 2e-3,
-      completion_per_1k: 0.012,
-      cached_per_1k: 2e-4
-    },
-    "gemini-3.5-flash": {
-      prompt_per_1k: 15e-4,
-      completion_per_1k: 9e-3,
-      cached_per_1k: 15e-5
-    },
-    "gemini-2.5-flash": {
-      prompt_per_1k: 3e-4,
-      completion_per_1k: 25e-4,
-      cached_per_1k: 75e-6
-    },
-    "gemini-2.5-flash-lite": {
-      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: {
-    "mistral-large-latest": {
-      prompt_per_1k: 2e-3,
-      completion_per_1k: 6e-3,
-      cached_per_1k: 2e-4
-    },
-    "mistral-small-latest": {
-      prompt_per_1k: 2e-4,
-      completion_per_1k: 6e-4,
-      cached_per_1k: 2e-5
-    }
-  },
-  groq: {
-    "llama-3.3-70b-versatile": {
-      prompt_per_1k: 59e-5,
-      completion_per_1k: 79e-5,
-      cached_per_1k: 295e-6
-    },
-    "llama-3.1-8b-instant": {
-      prompt_per_1k: 5e-5,
-      completion_per_1k: 8e-5,
-      cached_per_1k: 25e-6
-    }
-  },
-  deepseek: {
-    "deepseek-chat": {
-      prompt_per_1k: 27e-5,
-      completion_per_1k: 11e-4,
-      cached_per_1k: 7e-5
-    },
-    "deepseek-reasoner": {
-      prompt_per_1k: 55e-5,
-      completion_per_1k: 22e-4,
-      cached_per_1k: 14e-5
-    }
-  },
-  kimi: {
-    "kimi-k3": {
-      prompt_per_1k: 3e-3,
-      completion_per_1k: 0.015,
-      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).",
-      models: [
-        {
-          provider: "openai",
-          model: "gpt-4o-mini"
-        },
-        {
-          provider: "openai",
-          model: "gpt-5-mini"
-        },
-        {
-          provider: "anthropic",
-          model: "claude-haiku-4-5"
-        },
-        {
-          provider: "anthropic",
-          model: "claude-3-5-haiku"
-        },
-        {
-          provider: "gemini",
-          model: "gemini-3.5-flash"
-        },
-        {
-          provider: "gemini",
-          model: "gemini-2.5-flash"
-        },
-        {
-          provider: "gemini",
-          model: "gemini-2.5-flash-lite"
-        },
-        {
-          provider: "xai",
-          model: "grok-3-mini"
-        },
-        {
-          provider: "groq",
-          model: "llama-3.1-8b-instant"
-        },
-        {
-          provider: "deepseek",
-          model: "deepseek-chat"
-        }
-      ]
-    },
-    med: {
-      description: "Mid-tier models for reasoning, code, multi-step thinking.",
-      models: [
-        {
-          provider: "anthropic",
-          model: "claude-sonnet-4-5"
-        },
-        {
-          provider: "anthropic",
-          model: "claude-sonnet-4-6"
-        },
-        {
-          provider: "openai",
-          model: "gpt-4o"
-        },
-        {
-          provider: "openai",
-          model: "o3-mini"
-        },
-        {
-          provider: "gemini",
-          model: "gemini-3.1-pro-preview"
-        },
-        {
-          provider: "xai",
-          model: "grok-3"
-        },
-        {
-          provider: "mistral",
-          model: "mistral-large-latest"
-        },
-        {
-          provider: "deepseek",
-          model: "deepseek-reasoner"
-        },
-        {
-          provider: "groq",
-          model: "llama-3.3-70b-versatile"
-        }
-      ]
-    },
-    high: {
-      description: "Most capable models for hard reasoning, planning, audits.",
-      models: [
-        {
-          provider: "anthropic",
-          model: "claude-opus-4-5"
-        },
-        {
-          provider: "anthropic",
-          model: "claude-opus-4-7"
-        },
-        {
-          provider: "anthropic",
-          model: "claude-fable-5"
-        },
-        {
-          provider: "openai",
-          model: "gpt-5.5"
-        },
-        {
-          provider: "openai",
-          model: "gpt-5"
-        },
-        {
-          provider: "openai",
-          model: "o1"
-        },
-        {
-          provider: "openai",
-          model: "o3"
-        },
-        {
-          provider: "xai",
-          model: "grok-4-latest"
-        }
-      ]
+// src/core/retarget.ts
+function retargetProvider(p, newModel, fallbackModels) {
+  const chain = fallbackModels && fallbackModels.length > 0 ? fallbackModels : void 0;
+  if (p.model === newModel && chain === void 0) return p;
+  return {
+    name: p.name,
+    model: newModel,
+    ...chain ? { fallbackModels: chain } : {},
+    send: (args) => p.send({ ...args, modelOverride: newModel })
+  };
+}
+async function loadProviderWeights(storage, names) {
+  const out = {};
+  for (const raw of names) {
+    const name = raw.toLowerCase();
+    if (name in out) continue;
+    const row = await storage.getProviderStats(name);
+    if (!row) {
+      out[name] = 0.5;
+      continue;
     }
+    const total = row.wins + row.losses + row.abstains;
+    out[name] = total <= 0 ? 0.5 : (row.wins + 0.5 * row.abstains) / total;
   }
-};
+  return out;
+}
+
+// src/core/super-mode.ts
+var import_node_async_hooks = require("async_hooks");
 
 // src/core/pricing.ts
 var import_node_fs = require("fs");
@@ -1349,7 +958,208 @@ async function sendGemini(args) {
   const doFetch = args.fetchImpl ?? globalThis.fetch;
   const init = {
     method: "POST",
-    headers: { ...headers, "content-type": "application/json" },
+    headers: { ...headers, "content-type": "application/json" },
+    body: JSON.stringify(body)
+  };
+  if (args.signal) init.signal = args.signal;
+  const attemptOnce = async () => {
+    let respLike;
+    try {
+      respLike = await doFetch(url, init);
+    } catch (e) {
+      const d = diagnoseNetworkError(e);
+      if (!looksLikeNetworkFailure(d)) throw e;
+      throw new ProviderError(d.kind, networkErrorMessage("gemini", d), {
+        transient: d.transient
+      });
+    }
+    const status = respLike.status;
+    if (status >= 200 && status < 300) {
+      let parsed;
+      try {
+        parsed = await respLike.json();
+      } catch (e) {
+        throw new ProviderError("parse", `gemini: response body not JSON: ${e.message}`);
+      }
+      const { text, usage, finishReason } = parseGeminiResponse({
+        resp: parsed,
+        model: args.model,
+        purpose: args.purpose ?? "worker"
+      });
+      if (text.trim() === "" && finishReason === "MAX_TOKENS") {
+        throw new ProviderError(
+          "parse",
+          `gemini: the answer was cut off before any visible text \u2014 thinking used the whole output budget (${usage.completion_tokens} tokens). Raise CROSSCHECK_GEMINI_THINKING_HEADROOM_TOKENS or the purpose's budget.`,
+          { transient: false }
+        );
+      }
+      return { text, attempts: 1, usage: applyPricing2(usage, args.pricing) };
+    }
+    const bodyText = await respLike.text().catch(() => "");
+    throw httpFailureToProviderError("gemini", status, bodyText, parseRetryAfter(respLike), args.model);
+  };
+  await acquireRateLimit("gemini", args.signal ? { signal: args.signal } : void 0);
+  return sendWithRetry(attemptOnce, resolveRetryConfig(), args.signal);
+}
+
+// src/providers/openai-compatible.ts
+var OPENAI_COMPAT_DEFAULT_URLS = {
+  openai: "https://api.openai.com/v1/chat/completions",
+  xai: "https://api.x.ai/v1/chat/completions",
+  mistral: "https://api.mistral.ai/v1/chat/completions",
+  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",
+  // 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"
+]);
+function supportsNativeJsonSchema(provider, model) {
+  if (!OPENAI_COMPAT_NATIVE_STRUCTURED.has(provider.toLowerCase())) {
+    return false;
+  }
+  const m = model.toLowerCase();
+  if (/^o1(?:-|$)/.test(m)) return false;
+  return true;
+}
+function buildOpenAICompatibleRequest(opts) {
+  const msgs = opts.messages.map((m) => {
+    const text = typeof m.content === "string" ? m.content : "";
+    const images = imagesOf(m);
+    return images.length ? { role: m.role, content: [{ type: "text", text }, ...images.map((i) => ({ type: "image_url", image_url: { url: `data:${i.mediaType};base64,${i.data}` } }))] } : { role: m.role, content: text };
+  });
+  const body = {
+    model: opts.model,
+    messages: msgs
+  };
+  if (supportsTemperature(opts.provider, opts.model)) {
+    body.max_tokens = opts.maxTokens;
+    body.temperature = opts.temperature;
+  } else {
+    body.max_completion_tokens = opts.maxTokens;
+  }
+  if (opts.jsonSchema && supportsNativeJsonSchema(opts.provider, opts.model)) {
+    body.response_format = {
+      type: "json_schema",
+      json_schema: {
+        name: "structured_output",
+        schema: opts.jsonSchema,
+        // NOT strict. Strict mode requires `additionalProperties: false` on
+        // every object and every property listed in `required` — the
+        // engine's schemas don't meet that, so every structured OpenAI call
+        // (planner, judges, synthesis) was a 400, unnoticed until the canary
+        // grew a structured check (2026-10-11). Non-strict still shapes the
+        // reply to the schema, and structured.ts validates it against the
+        // full schema regardless, so strictness added a failure mode and no
+        // safety.
+        strict: false
+      }
+    };
+  }
+  const headers = {
+    "content-type": "application/json",
+    Authorization: `Bearer ${opts.apiKey}`
+  };
+  return { url: opts.url, headers, body };
+}
+function parseOpenAICompatibleResponse(opts) {
+  const r = opts.resp ?? {};
+  let text;
+  try {
+    const choices = r["choices"];
+    if (!Array.isArray(choices) || choices.length === 0) {
+      throw new Error("choices missing or empty");
+    }
+    const first = choices[0];
+    const message = first["message"];
+    if (!message || typeof message["content"] !== "string") {
+      throw new Error("message.content missing or not string");
+    }
+    text = message["content"];
+  } catch (e) {
+    throw new ProviderError(
+      "parse",
+      `${opts.provider}: unexpected response shape: ${JSON.stringify(r).slice(0, 200)}`
+    );
+  }
+  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: 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: reportedTotal,
+    cost_usd: 0,
+    estimated: Object.keys(u).length === 0,
+    purpose: opts.purpose
+  };
+  if (usage.total_tokens === 0) {
+    usage.total_tokens = usage.prompt_tokens + usage.completion_tokens;
+  }
+  return { text, usage };
+}
+function applyPricing3(usage, pricing) {
+  const { cost_usd, estimated } = calculateCost(
+    pricing,
+    usage.provider,
+    usage.model,
+    usage.prompt_tokens,
+    usage.completion_tokens,
+    usage.cached_tokens
+  );
+  return {
+    ...usage,
+    cost_usd,
+    estimated: usage.estimated || estimated
+  };
+}
+async function sendOpenAICompatible(args) {
+  const { url, headers, body } = buildOpenAICompatibleRequest({
+    provider: args.provider,
+    model: args.model,
+    apiKey: args.apiKey,
+    url: args.url,
+    messages: args.messages,
+    maxTokens: args.maxTokens,
+    temperature: args.temperature,
+    ...args.jsonSchema ? { jsonSchema: args.jsonSchema } : {}
+  });
+  if (args.provider === "openai" && isReasoningModel("openai", args.model) && !/^o1(?:-|$)/i.test(args.model)) {
+    const effort = (process.env["CROSSCHECK_OPENAI_REASONING_EFFORT"] ?? "low").toLowerCase();
+    if (effort === "minimal" || effort === "low" || effort === "medium" || effort === "high") {
+      body.reasoning_effort = effort;
+    }
+  }
+  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;
+  }
+  const doFetch = args.fetchImpl ?? globalThis.fetch;
+  const init = {
+    method: "POST",
+    headers,
     body: JSON.stringify(body)
   };
   if (args.signal) init.signal = args.signal;
@@ -1360,7 +1170,7 @@ async function sendGemini(args) {
     } catch (e) {
       const d = diagnoseNetworkError(e);
       if (!looksLikeNetworkFailure(d)) throw e;
-      throw new ProviderError(d.kind, networkErrorMessage("gemini", d), {
+      throw new ProviderError(d.kind, networkErrorMessage(args.provider, d), {
         transient: d.transient
       });
     }
@@ -1370,374 +1180,808 @@ async function sendGemini(args) {
       try {
         parsed = await respLike.json();
       } catch (e) {
-        throw new ProviderError("parse", `gemini: response body not JSON: ${e.message}`);
+        throw new ProviderError("parse", `${args.provider}: response body not JSON: ${e.message}`);
       }
-      const { text, usage, finishReason } = parseGeminiResponse({
+      const { text, usage } = parseOpenAICompatibleResponse({
         resp: parsed,
+        provider: args.provider,
         model: args.model,
         purpose: args.purpose ?? "worker"
       });
-      if (text.trim() === "" && finishReason === "MAX_TOKENS") {
-        throw new ProviderError(
-          "parse",
-          `gemini: the answer was cut off before any visible text \u2014 thinking used the whole output budget (${usage.completion_tokens} tokens). Raise CROSSCHECK_GEMINI_THINKING_HEADROOM_TOKENS or the purpose's budget.`,
-          { transient: false }
-        );
-      }
-      return { text, attempts: 1, usage: applyPricing2(usage, args.pricing) };
+      return { text, attempts: 1, usage: applyPricing3(usage, args.pricing) };
     }
     const bodyText = await respLike.text().catch(() => "");
-    throw httpFailureToProviderError("gemini", status, bodyText, parseRetryAfter(respLike), args.model);
+    throw httpFailureToProviderError(args.provider, status, bodyText, parseRetryAfter(respLike), args.model);
   };
-  await acquireRateLimit("gemini", args.signal ? { signal: args.signal } : void 0);
+  await acquireRateLimit(args.provider, args.signal ? { signal: args.signal } : void 0);
   return sendWithRetry(attemptOnce, resolveRetryConfig(), args.signal);
 }
 
-// src/providers/openai-compatible.ts
-var OPENAI_COMPAT_DEFAULT_URLS = {
-  openai: "https://api.openai.com/v1/chat/completions",
-  xai: "https://api.x.ai/v1/chat/completions",
-  mistral: "https://api.mistral.ai/v1/chat/completions",
-  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",
-  // 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"
+// src/providers/registry.ts
+var DEFAULT_MODELS = {
+  anthropic: "claude-opus-5-5",
+  openai: "gpt-6-sol",
+  xai: "grok-4-latest",
+  mistral: "mistral-large-latest",
+  groq: "llama-3.3-70b-versatile",
+  deepseek: "deepseek-chat",
+  gemini: "gemini-3.8-flash",
+  kimi: "kimi-k3",
+  qwen: "qwen3.8-max"
+};
+var DEFAULT_FALLBACKS = {
+  // gpt-5.6 requires a billing-enabled account and is excluded from the free
+  // tier, so gpt-5 remains the broadly-available floor. One hop, not a
+  // ladder — an intermediate 5.5 step would just be another model to reason
+  // about.
+  openai: ["gpt-5"],
+  // 3.8 Flash is recent enough that some projects won't have it yet.
+  gemini: ["gemini-3.1-pro-preview"],
+  // Opus 5.5 is new enough that an account could still be waiting on it.
+  // Opus 5 is the same family one step back — and this chain is consulted
+  // ONLY for a model-access failure, never for a bad key or a rate limit,
+  // so it can't quietly downgrade anyone who simply hit a 429.
+  anthropic: ["claude-opus-5"]
+};
+var MODEL_ENV_VARS = {
+  anthropic: "ANTHROPIC_MODEL",
+  openai: "OPENAI_MODEL",
+  xai: "XAI_MODEL",
+  mistral: "MISTRAL_MODEL",
+  groq: "GROQ_MODEL",
+  deepseek: "DEEPSEEK_MODEL",
+  gemini: "GEMINI_MODEL",
+  kimi: "KIMI_MODEL",
+  qwen: "QWEN_MODEL"
+};
+function parseFallbackModels(raw) {
+  if (!raw) return [];
+  return raw.split(",").map((s) => s.trim()).filter(Boolean);
+}
+function fallbackModelsFor(env, provider) {
+  const envVar = MODEL_ENV_VARS[provider];
+  const fromEnv = envVar ? parseFallbackModels(env[`${envVar}_FALLBACKS`]) : [];
+  if (fromEnv.length > 0) return fromEnv;
+  const pinned = envVar ? env[envVar] : void 0;
+  if (pinned && pinned !== DEFAULT_MODELS[provider]) return [];
+  return DEFAULT_FALLBACKS[provider] ?? [];
+}
+var DEFAULT_EXCLUDED_PROVIDERS = ["mistral", "groq"];
+function excludedProviders(env) {
+  const parse = (v) => (v ?? "").split(",").map((x) => x.trim().toLowerCase()).filter((x) => x !== "");
+  const excluded = new Set(
+    parse(env["CROSSCHECK_EXCLUDE_PROVIDERS"]).length > 0 ? parse(env["CROSSCHECK_EXCLUDE_PROVIDERS"]) : DEFAULT_EXCLUDED_PROVIDERS
+  );
+  for (const name of parse(env["CROSSCHECK_ENABLE_PROVIDERS"])) excluded.delete(name);
+  return excluded;
+}
+function buildProviders(opts) {
+  const out = {};
+  const excluded = excludedProviders(opts.env);
+  const anthropicKey = opts.env["ANTHROPIC_API_KEY"];
+  if (anthropicKey && !excluded.has("anthropic")) {
+    const model = opts.env["ANTHROPIC_MODEL"] ?? DEFAULT_MODELS.anthropic;
+    out["anthropic"] = makeAnthropicProvider(model, anthropicKey, opts, fallbackModelsFor(opts.env, "anthropic"));
+  }
+  const openAiCompatSpec = [
+    { name: "openai", keyEnv: "OPENAI_API_KEY", modelEnv: "OPENAI_MODEL", defaultModel: DEFAULT_MODELS["openai"] },
+    { name: "xai", keyEnv: "XAI_API_KEY", modelEnv: "XAI_MODEL", defaultModel: DEFAULT_MODELS["xai"] },
+    { 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: "qwen", keyEnv: "QWEN_API_KEY", modelEnv: "QWEN_MODEL", defaultModel: DEFAULT_MODELS["qwen"] }
+  ];
+  for (const s of openAiCompatSpec) {
+    const apiKey = opts.env[s.keyEnv];
+    if (!apiKey) continue;
+    if (excluded.has(s.name)) continue;
+    const model = opts.env[s.modelEnv] ?? s.defaultModel;
+    out[s.name] = makeOpenAICompatibleProvider(s.name, model, apiKey, opts, fallbackModelsFor(opts.env, s.name));
+  }
+  const geminiKey = opts.env["GEMINI_API_KEY"];
+  if (geminiKey && !excluded.has("gemini")) {
+    const model = opts.env["GEMINI_MODEL"] ?? DEFAULT_MODELS.gemini;
+    out["gemini"] = makeGeminiProvider(model, geminiKey, opts, fallbackModelsFor(opts.env, "gemini"));
+  }
+  return out;
+}
+function makeAnthropicProvider(model, apiKey, opts, fallbackModels = []) {
+  return {
+    name: "anthropic",
+    model,
+    fallbackModels,
+    send: async (args) => {
+      const sendOpts = {
+        ...args,
+        apiKey,
+        model: args.modelOverride ?? model,
+        pricing: opts.pricing
+      };
+      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
+      return sendAnthropic(sendOpts);
+    }
+  };
+}
+function makeOpenAICompatibleProvider(name, model, apiKey, opts, fallbackModels = []) {
+  const url = OPENAI_COMPAT_DEFAULT_URLS[name];
+  return {
+    name,
+    model,
+    fallbackModels,
+    send: async (args) => {
+      const sendOpts = {
+        ...args,
+        provider: name,
+        apiKey,
+        model: args.modelOverride ?? model,
+        url,
+        pricing: opts.pricing
+      };
+      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
+      return sendOpenAICompatible(sendOpts);
+    }
+  };
+}
+function makeGeminiProvider(model, apiKey, opts, fallbackModels = []) {
+  return {
+    name: "gemini",
+    model,
+    fallbackModels,
+    send: async (args) => {
+      const sendOpts = {
+        ...args,
+        apiKey,
+        model: args.modelOverride ?? model,
+        pricing: opts.pricing
+      };
+      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
+      return sendGemini(sendOpts);
+    }
+  };
+}
+
+// src/core/reasoning-upgrade.ts
+var UPGRADE_MODEL = "claude-fable-5";
+var UPGRADE_LABEL = "Fable 5";
+function isUpgradeRequested(args) {
+  if (args["upgrade"] === true) return true;
+  const reasoning = args["reasoning"];
+  if (typeof reasoning === "string" && reasoning.trim().toLowerCase() === "max") {
+    return true;
+  }
+  const model = args["model"];
+  if (typeof model === "string") {
+    const m = model.trim().toLowerCase();
+    if (m === "fable5" || m === "fable-5" || m === "fable" || m === UPGRADE_MODEL) {
+      return true;
+    }
+  }
+  return false;
+}
+var HEAVY_TERMS = [
+  /\bprov(e|ing|en)\b|\bproof\b|\btheorem\b|\bderiv(e|ation)\b/i,
+  /\barchitect(ure|ing)?\b|\bdesign\s+(a|the|an)\b|\bhigh[-\s]?level\s+design\b/i,
+  /\btrade[-\s]?offs?\b|\bpros?\s+and\s+cons?\b|\bweigh\b/i,
+  /\boptimi[sz]e\b|\bcomplexity\b|\bO\(|\balgorithm(ic)?\b/i,
+  /\brace\s+condition\b|\bconcurren(t|cy)\b|\bdeadlock\b|\bdistributed\b|\bconsensus\b/i,
+  /\broot[-\s]?cause\b|\bdebug(ging)?\b|\bwhy\s+(does|is|would|might)\b/i,
+  /\bsecurity\b|\bthreat\s+model\b|\bexploit\b|\bvulnerabilit(y|ies)\b/i,
+  /\bmigrat(e|ion)\b|\brefactor\b|\bre[-\s]?architect\b/i,
+  /\binvariant\b|\bcorrectness\b|\bedge\s+cases?\b|\bformal(ly)?\b/i,
+  /\bstrateg(y|ic|ies)\b|\bplan\s+(the|a|an|out)\b|\broadmap\b/i,
+  /\breason\s+(through|about)\b|\bstep[-\s]?by[-\s]?step\b|\bthink\s+deeply\b/i,
+  /\bnp[-\s]?(hard|complete)\b|\bgraph\b|\bstate\s+machine\b|\bconstraint\b/i
+];
+var DIFFICULTY_TERMS = /\b(hard|tricky|subtle|complex|complicated|non[-\s]?trivial|thorny|gnarly|difficult|deep)\b/i;
+function assessReasoning(text) {
+  const s = typeof text === "string" ? text : "";
+  const signals = [];
+  let score = 0;
+  let termHits = 0;
+  for (const re of HEAVY_TERMS) {
+    if (re.test(s)) termHits += 1;
+  }
+  if (termHits > 0) {
+    const add = Math.min(termHits, 4);
+    score += add;
+    signals.push(`${termHits} reasoning cue${termHits === 1 ? "" : "s"}`);
+  }
+  if (DIFFICULTY_TERMS.test(s)) {
+    score += 1;
+    signals.push("explicit difficulty marker");
+  }
+  const words = s.split(/\s+/).filter(Boolean).length;
+  if (words >= 120) {
+    score += 2;
+    signals.push("long, detailed prompt");
+  } else if (words >= 50) {
+    score += 1;
+    signals.push("substantial prompt");
+  }
+  const constraints = (s.match(/\b(must|should|need to|ensure|require[ds]?|constraint)\b/gi) ?? []).length + (s.match(/(^|\n)\s*(?:[-*]|\d+[.)])\s+/g) ?? []).length;
+  if (constraints >= 3) {
+    score += 1;
+    signals.push(`${constraints} constraints/steps`);
+  }
+  const heavy = score >= 3;
+  const reason = heavy ? `Task looks reasoning-heavy (${signals.join(", ") || "score " + score}).` : `Task looks routine (score ${score}).`;
+  return { heavy, score, reason, signals };
+}
+function buildSuggestedUpgrade(toolName, assessment) {
+  return {
+    model: UPGRADE_MODEL,
+    label: UPGRADE_LABEL,
+    reason: assessment.reason,
+    how: `re-run \`${toolName}\` with reasoning:"max" (or upgrade:true) to route the lead reasoning seat to ${UPGRADE_LABEL}, and \u2014 when an OpenAI key is configured \u2014 add an OpenAI second-opinion pass that ${UPGRADE_LABEL} weighs before finalizing (the result names which model ran). Other panel seats stay on their normal models.`
+  };
+}
+
+// src/core/super-mode.ts
+var SUPER_MODELS = {
+  // Fable 5.1 verified present on the account; 5 as the fallback for accounts
+  // that haven't been granted 5.1 yet.
+  anthropic: ["claude-fable-5-1", UPGRADE_MODEL],
+  // Astra is live (verified against the account's model list 2026-09-23).
+  // The fallback is named explicitly rather than reusing CO_REASON_MODEL:
+  // that constant means "OpenAI's top reasoning model" for the max-reasoning
+  // co-reasoner, which is a different question from "what should super fall
+  // back to". Sharing the symbol silently coupled two unrelated decisions.
+  openai: ["gpt-6-astra", "gpt-6-sol"],
+  // Explicitly left as-is per product decision. Note grok-4-latest is not in
+  // xAI's published model list, but it resolves in practice — real panel calls
+  // on it succeeded as recently as 2026-08-30 — so it is an unlisted alias
+  // rather than a dead id.
+  xai: [DEFAULT_MODELS["xai"]],
+  // 3.8 Flash is now BOTH the everyday default and the super pick for Gemini,
+  // so this can no longer borrow DEFAULT_MODELS as its fallback — that would
+  // repeat the same model and retry one that just failed. Named explicitly.
+  gemini: ["gemini-3.8-flash", "gemini-3.1-pro-preview"],
+  kimi: [DEFAULT_MODELS["kimi"]],
+  qwen: [DEFAULT_MODELS["qwen"]]
 };
-var OPENAI_COMPAT_NATIVE_STRUCTURED = /* @__PURE__ */ new Set([
-  "openai"
-]);
-function supportsNativeJsonSchema(provider, model) {
-  if (!OPENAI_COMPAT_NATIVE_STRUCTURED.has(provider.toLowerCase())) {
-    return false;
+var SUPER_ENV_VARS = {
+  anthropic: "ANTHROPIC_SUPER_MODEL",
+  openai: "OPENAI_SUPER_MODEL",
+  xai: "XAI_SUPER_MODEL",
+  gemini: "GEMINI_SUPER_MODEL",
+  kimi: "KIMI_SUPER_MODEL",
+  qwen: "QWEN_SUPER_MODEL",
+  mistral: "MISTRAL_SUPER_MODEL",
+  groq: "GROQ_SUPER_MODEL",
+  deepseek: "DEEPSEEK_SUPER_MODEL"
+};
+function parseChain(v) {
+  if (!v) return [];
+  const out = [];
+  for (const raw of v.split(",")) {
+    const m = raw.trim();
+    if (m && !out.includes(m)) out.push(m);
   }
-  const m = model.toLowerCase();
-  if (/^o1(?:-|$)/.test(m)) return false;
-  return true;
+  return out;
 }
-function buildOpenAICompatibleRequest(opts) {
-  const msgs = opts.messages.map((m) => {
-    const text = typeof m.content === "string" ? m.content : "";
-    const images = imagesOf(m);
-    return images.length ? { role: m.role, content: [{ type: "text", text }, ...images.map((i) => ({ type: "image_url", image_url: { url: `data:${i.mediaType};base64,${i.data}` } }))] } : { role: m.role, content: text };
-  });
-  const body = {
-    model: opts.model,
-    messages: msgs
+var superEnvStore = new import_node_async_hooks.AsyncLocalStorage();
+var currentEnv = () => superEnvStore.getStore() ?? process.env;
+function superChainFor(provider, env = currentEnv()) {
+  const varName = SUPER_ENV_VARS[provider];
+  const user = varName ? parseChain(env[varName]) : [];
+  if (user.length > 0) {
+    const fallbacks = varName ? parseChain(env[`${varName}_FALLBACKS`]) : [];
+    return [...user, ...fallbacks.filter((m) => !user.includes(m))];
+  }
+  return SUPER_MODELS[provider] ?? [];
+}
+function superPrimary(provider, env = currentEnv()) {
+  return superChainFor(provider, env)[0];
+}
+var SUPER_PROVIDER_NAMES = Object.keys(SUPER_MODELS);
+function isSuperRequested(args) {
+  return args["super"] === true;
+}
+function retargetForSuper(p, env = currentEnv()) {
+  const chain = superChainFor(p.name, env);
+  if (chain.length === 0) return p;
+  const [primary, ...fallbacks] = chain;
+  return retargetProvider(p, primary, fallbacks);
+}
+var SUPER_BUDGET_MULTIPLIER = 3;
+var SUPER_BUDGET_CEILING = 16e3;
+function superBudget(base) {
+  return Math.min(base * SUPER_BUDGET_MULTIPLIER, SUPER_BUDGET_CEILING);
+}
+
+// src/core/moderator.ts
+var DEFAULT_MODERATOR = "anthropic";
+var PROVIDER_ALIASES = {
+  grok: "xai",
+  claude: "anthropic",
+  google: "gemini",
+  chatgpt: "openai",
+  gpt: "openai",
+  moonshot: "kimi"
+};
+function parseModeratorSpec(raw) {
+  if (typeof raw !== "string") return null;
+  const s = raw.trim();
+  if (!s) return null;
+  const i = s.indexOf(":");
+  const providerPart = (i < 0 ? s : s.slice(0, i)).trim().toLowerCase();
+  const modelPart = i < 0 ? "" : s.slice(i + 1).trim();
+  if (!providerPart) return null;
+  return {
+    provider: PROVIDER_ALIASES[providerPart] ?? providerPart,
+    model: modelPart || null
   };
-  if (supportsTemperature(opts.provider, opts.model)) {
-    body.max_tokens = opts.maxTokens;
-    body.temperature = opts.temperature;
-  } else {
-    body.max_completion_tokens = opts.maxTokens;
+}
+function pickModeratorSpec(...candidates) {
+  for (const c of candidates) {
+    const s = parseModeratorSpec(c);
+    if (s) return s;
+  }
+  return { provider: DEFAULT_MODERATOR, model: null };
+}
+function resolveModerator(providers, ...candidates) {
+  const spec = pickModeratorSpec(...candidates);
+  const base = providers[spec.provider];
+  if (!base) return { provider: void 0, spec, pinned: false };
+  if (!spec.model) return { provider: base, spec, pinned: false };
+  return { provider: retargetProvider(base, spec.model), spec, pinned: true };
+}
+function synthesisSeat(o) {
+  const { moderator, pinned, superRequested, upgradeRequested } = o;
+  const wantsMore = superRequested || upgradeRequested;
+  if (!wantsMore) return { provider: moderator, upgraded: false, label: null, skipReason: null };
+  if (pinned) {
+    return {
+      provider: moderator,
+      upgraded: false,
+      label: null,
+      skipReason: `the moderator is pinned to ${moderator.name}:${moderator.model}; a pinned model is used as-is.`
+    };
   }
-  if (opts.jsonSchema && supportsNativeJsonSchema(opts.provider, opts.model)) {
-    body.response_format = {
-      type: "json_schema",
-      json_schema: {
-        name: "structured_output",
-        schema: opts.jsonSchema,
-        // NOT strict. Strict mode requires `additionalProperties: false` on
-        // every object and every property listed in `required` — the
-        // engine's schemas don't meet that, so every structured OpenAI call
-        // (planner, judges, synthesis) was a 400, unnoticed until the canary
-        // grew a structured check (2026-10-11). Non-strict still shapes the
-        // reply to the schema, and structured.ts validates it against the
-        // full schema regardless, so strictness added a failure mode and no
-        // safety.
-        strict: false
-      }
+  if (moderator.name === "anthropic" && !superRequested) {
+    return { provider: retargetProvider(moderator, UPGRADE_MODEL), upgraded: true, label: UPGRADE_LABEL, skipReason: null };
+  }
+  if (superChainFor(moderator.name).length === 0) {
+    return {
+      provider: moderator,
+      upgraded: false,
+      label: null,
+      skipReason: `no flagship model is known for ${moderator.name}; the moderator ran at ${moderator.model}.`
     };
   }
-  const headers = {
-    "content-type": "application/json",
-    Authorization: `Bearer ${opts.apiKey}`
-  };
-  return { url: opts.url, headers, body };
+  const seat = retargetForSuper(moderator);
+  const label = moderator.name === "anthropic" ? UPGRADE_LABEL : seat.model;
+  return { provider: seat, upgraded: true, label, skipReason: null };
 }
-function parseOpenAICompatibleResponse(opts) {
-  const r = opts.resp ?? {};
-  let text;
-  try {
-    const choices = r["choices"];
-    if (!Array.isArray(choices) || choices.length === 0) {
-      throw new Error("choices missing or empty");
+
+// ../../config/pricing.json
+var pricing_default = {
+  _meta: {
+    currency: "USD",
+    unit: "per_1k_tokens",
+    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: {
+    _source: {
+      url: "https://platform.claude.com/docs/en/about-claude/pricing",
+      verified_at: "2026-09-27"
+    },
+    "claude-opus-4-5": {
+      prompt_per_1k: 5e-3,
+      completion_per_1k: 0.025,
+      cached_per_1k: 5e-4
+    },
+    "claude-opus-4-7": {
+      prompt_per_1k: 5e-3,
+      completion_per_1k: 0.025,
+      cached_per_1k: 5e-4
+    },
+    "claude-opus-4-8": {
+      prompt_per_1k: 5e-3,
+      completion_per_1k: 0.025,
+      cached_per_1k: 5e-4
+    },
+    "claude-opus-5-5": {
+      prompt_per_1k: 4e-3,
+      completion_per_1k: 0.02,
+      cached_per_1k: 2e-4
+    },
+    "claude-opus-5": {
+      prompt_per_1k: 5e-3,
+      completion_per_1k: 0.025,
+      cached_per_1k: 5e-4
+    },
+    "claude-fable-5": {
+      prompt_per_1k: 0.01,
+      completion_per_1k: 0.05,
+      cached_per_1k: 1e-3
+    },
+    "claude-sonnet-4-5": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 3e-4
+    },
+    "claude-sonnet-4-6": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 3e-4
+    },
+    "claude-haiku-4-5": {
+      prompt_per_1k: 1e-3,
+      completion_per_1k: 5e-3,
+      cached_per_1k: 1e-4
+    },
+    "claude-3-5-sonnet": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 3e-4
+    },
+    "claude-3-5-haiku": {
+      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: {
+    _source: {
+      url: "https://developers.openai.com/api/docs/pricing",
+      verified_at: "2026-09-27"
+    },
+    "gpt-6-astra": {
+      prompt_per_1k: 0.01,
+      completion_per_1k: 0.05,
+      cached_per_1k: 1e-3
+    },
+    "gpt-6-sol": {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 0.01,
+      cached_per_1k: 2e-4
+    },
+    "gpt-6-luna": {
+      prompt_per_1k: 1e-4,
+      completion_per_1k: 5e-4,
+      cached_per_1k: 1e-5
+    },
+    "gpt-5.6-sol": {
+      prompt_per_1k: 4e-3,
+      completion_per_1k: 0.02,
+      cached_per_1k: 4e-4
+    },
+    "gpt-5.6-terra": {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 0.012,
+      cached_per_1k: 2e-4
+    },
+    "gpt-5.6-luna": {
+      prompt_per_1k: 2e-4,
+      completion_per_1k: 12e-4,
+      cached_per_1k: 2e-5
+    },
+    "gpt-5.6": {
+      prompt_per_1k: 4e-3,
+      completion_per_1k: 0.02,
+      cached_per_1k: 4e-4
+    },
+    "gpt-5.5": {
+      prompt_per_1k: 5e-3,
+      completion_per_1k: 0.03,
+      cached_per_1k: 5e-4
+    },
+    "gpt-5": {
+      prompt_per_1k: 125e-5,
+      completion_per_1k: 0.01,
+      cached_per_1k: 125e-6
+    },
+    "gpt-5-mini": {
+      prompt_per_1k: 25e-5,
+      completion_per_1k: 2e-3,
+      cached_per_1k: 25e-6
+    },
+    "gpt-5-nano": {
+      prompt_per_1k: 5e-5,
+      completion_per_1k: 4e-4,
+      cached_per_1k: 5e-6
+    },
+    "gpt-4o": {
+      prompt_per_1k: 25e-4,
+      completion_per_1k: 0.01,
+      cached_per_1k: 125e-5
+    },
+    "gpt-4o-mini": {
+      prompt_per_1k: 15e-5,
+      completion_per_1k: 6e-4,
+      cached_per_1k: 75e-6
+    },
+    o1: {
+      prompt_per_1k: 0.015,
+      completion_per_1k: 0.06,
+      cached_per_1k: 75e-4
+    },
+    "o1-mini": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.012,
+      cached_per_1k: 15e-4
+    },
+    o3: {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 8e-3,
+      cached_per_1k: 5e-4
+    },
+    "o3-mini": {
+      prompt_per_1k: 11e-4,
+      completion_per_1k: 44e-4,
+      cached_per_1k: 55e-5
+    },
+    "gpt-5-pro": {
+      prompt_per_1k: 0.015,
+      completion_per_1k: 0.12,
+      cached_per_1k: 0
+    },
+    "gpt-5.2-pro": {
+      prompt_per_1k: 0.021,
+      completion_per_1k: 0.168,
+      cached_per_1k: 0
+    },
+    "gpt-5.4-pro": {
+      prompt_per_1k: 0.03,
+      completion_per_1k: 0.18,
+      cached_per_1k: 0
+    },
+    "gpt-5.5-pro": {
+      prompt_per_1k: 0.03,
+      completion_per_1k: 0.18,
+      cached_per_1k: 0
+    },
+    "o1-pro": {
+      prompt_per_1k: 0.15,
+      completion_per_1k: 0.6,
+      cached_per_1k: 0
+    },
+    "o3-pro": {
+      prompt_per_1k: 0.02,
+      completion_per_1k: 0.08,
+      cached_per_1k: 0
     }
-    const first = choices[0];
-    const message = first["message"];
-    if (!message || typeof message["content"] !== "string") {
-      throw new Error("message.content missing or not string");
+  },
+  xai: {
+    "grok-4-latest": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 75e-5
+    },
+    "grok-3": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 15e-4
+    },
+    "grok-3-mini": {
+      prompt_per_1k: 3e-4,
+      completion_per_1k: 5e-4,
+      cached_per_1k: 15e-5
     }
-    text = message["content"];
-  } catch (e) {
-    throw new ProviderError(
-      "parse",
-      `${opts.provider}: unexpected response shape: ${JSON.stringify(r).slice(0, 200)}`
-    );
-  }
-  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: 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: reportedTotal,
-    cost_usd: 0,
-    estimated: Object.keys(u).length === 0,
-    purpose: opts.purpose
-  };
-  if (usage.total_tokens === 0) {
-    usage.total_tokens = usage.prompt_tokens + usage.completion_tokens;
-  }
-  return { text, usage };
-}
-function applyPricing3(usage, pricing) {
-  const { cost_usd, estimated } = calculateCost(
-    pricing,
-    usage.provider,
-    usage.model,
-    usage.prompt_tokens,
-    usage.completion_tokens,
-    usage.cached_tokens
-  );
-  return {
-    ...usage,
-    cost_usd,
-    estimated: usage.estimated || estimated
-  };
-}
-async function sendOpenAICompatible(args) {
-  const { url, headers, body } = buildOpenAICompatibleRequest({
-    provider: args.provider,
-    model: args.model,
-    apiKey: args.apiKey,
-    url: args.url,
-    messages: args.messages,
-    maxTokens: args.maxTokens,
-    temperature: args.temperature,
-    ...args.jsonSchema ? { jsonSchema: args.jsonSchema } : {}
-  });
-  if (args.provider === "openai" && isReasoningModel("openai", args.model) && !/^o1(?:-|$)/i.test(args.model)) {
-    const effort = (process.env["CROSSCHECK_OPENAI_REASONING_EFFORT"] ?? "low").toLowerCase();
-    if (effort === "minimal" || effort === "low" || effort === "medium" || effort === "high") {
-      body.reasoning_effort = effort;
+  },
+  gemini: {
+    "gemini-3.1-pro-preview": {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 0.012,
+      cached_per_1k: 2e-4
+    },
+    "gemini-3.5-flash": {
+      prompt_per_1k: 15e-4,
+      completion_per_1k: 9e-3,
+      cached_per_1k: 15e-5
+    },
+    "gemini-2.5-flash": {
+      prompt_per_1k: 3e-4,
+      completion_per_1k: 25e-4,
+      cached_per_1k: 75e-6
+    },
+    "gemini-2.5-flash-lite": {
+      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
     }
-  }
-  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;
-  }
-  const doFetch = args.fetchImpl ?? globalThis.fetch;
-  const init = {
-    method: "POST",
-    headers,
-    body: JSON.stringify(body)
-  };
-  if (args.signal) init.signal = args.signal;
-  const attemptOnce = async () => {
-    let respLike;
-    try {
-      respLike = await doFetch(url, init);
-    } catch (e) {
-      const d = diagnoseNetworkError(e);
-      if (!looksLikeNetworkFailure(d)) throw e;
-      throw new ProviderError(d.kind, networkErrorMessage(args.provider, d), {
-        transient: d.transient
-      });
+  },
+  mistral: {
+    "mistral-large-latest": {
+      prompt_per_1k: 2e-3,
+      completion_per_1k: 6e-3,
+      cached_per_1k: 2e-4
+    },
+    "mistral-small-latest": {
+      prompt_per_1k: 2e-4,
+      completion_per_1k: 6e-4,
+      cached_per_1k: 2e-5
     }
-    const status = respLike.status;
-    if (status >= 200 && status < 300) {
-      let parsed;
-      try {
-        parsed = await respLike.json();
-      } catch (e) {
-        throw new ProviderError("parse", `${args.provider}: response body not JSON: ${e.message}`);
-      }
-      const { text, usage } = parseOpenAICompatibleResponse({
-        resp: parsed,
-        provider: args.provider,
-        model: args.model,
-        purpose: args.purpose ?? "worker"
-      });
-      return { text, attempts: 1, usage: applyPricing3(usage, args.pricing) };
+  },
+  groq: {
+    "llama-3.3-70b-versatile": {
+      prompt_per_1k: 59e-5,
+      completion_per_1k: 79e-5,
+      cached_per_1k: 295e-6
+    },
+    "llama-3.1-8b-instant": {
+      prompt_per_1k: 5e-5,
+      completion_per_1k: 8e-5,
+      cached_per_1k: 25e-6
     }
-    const bodyText = await respLike.text().catch(() => "");
-    throw httpFailureToProviderError(args.provider, status, bodyText, parseRetryAfter(respLike), args.model);
-  };
-  await acquireRateLimit(args.provider, args.signal ? { signal: args.signal } : void 0);
-  return sendWithRetry(attemptOnce, resolveRetryConfig(), args.signal);
-}
-
-// src/providers/registry.ts
-var DEFAULT_MODELS = {
-  anthropic: "claude-opus-5-5",
-  openai: "gpt-6-sol",
-  xai: "grok-4-latest",
-  mistral: "mistral-large-latest",
-  groq: "llama-3.3-70b-versatile",
-  deepseek: "deepseek-chat",
-  gemini: "gemini-3.8-flash",
-  kimi: "kimi-k3",
-  qwen: "qwen3.8-max"
-};
-var DEFAULT_FALLBACKS = {
-  // gpt-5.6 requires a billing-enabled account and is excluded from the free
-  // tier, so gpt-5 remains the broadly-available floor. One hop, not a
-  // ladder — an intermediate 5.5 step would just be another model to reason
-  // about.
-  openai: ["gpt-5"],
-  // 3.8 Flash is recent enough that some projects won't have it yet.
-  gemini: ["gemini-3.1-pro-preview"],
-  // Opus 5.5 is new enough that an account could still be waiting on it.
-  // Opus 5 is the same family one step back — and this chain is consulted
-  // ONLY for a model-access failure, never for a bad key or a rate limit,
-  // so it can't quietly downgrade anyone who simply hit a 429.
-  anthropic: ["claude-opus-5"]
-};
-var MODEL_ENV_VARS = {
-  anthropic: "ANTHROPIC_MODEL",
-  openai: "OPENAI_MODEL",
-  xai: "XAI_MODEL",
-  mistral: "MISTRAL_MODEL",
-  groq: "GROQ_MODEL",
-  deepseek: "DEEPSEEK_MODEL",
-  gemini: "GEMINI_MODEL",
-  kimi: "KIMI_MODEL",
-  qwen: "QWEN_MODEL"
-};
-function parseFallbackModels(raw) {
-  if (!raw) return [];
-  return raw.split(",").map((s) => s.trim()).filter(Boolean);
-}
-function fallbackModelsFor(env, provider) {
-  const envVar = MODEL_ENV_VARS[provider];
-  const fromEnv = envVar ? parseFallbackModels(env[`${envVar}_FALLBACKS`]) : [];
-  if (fromEnv.length > 0) return fromEnv;
-  const pinned = envVar ? env[envVar] : void 0;
-  if (pinned && pinned !== DEFAULT_MODELS[provider]) return [];
-  return DEFAULT_FALLBACKS[provider] ?? [];
-}
-var DEFAULT_EXCLUDED_PROVIDERS = ["mistral", "groq"];
-function excludedProviders(env) {
-  const parse = (v) => (v ?? "").split(",").map((x) => x.trim().toLowerCase()).filter((x) => x !== "");
-  const excluded = new Set(
-    parse(env["CROSSCHECK_EXCLUDE_PROVIDERS"]).length > 0 ? parse(env["CROSSCHECK_EXCLUDE_PROVIDERS"]) : DEFAULT_EXCLUDED_PROVIDERS
-  );
-  for (const name of parse(env["CROSSCHECK_ENABLE_PROVIDERS"])) excluded.delete(name);
-  return excluded;
-}
-function buildProviders(opts) {
-  const out = {};
-  const excluded = excludedProviders(opts.env);
-  const anthropicKey = opts.env["ANTHROPIC_API_KEY"];
-  if (anthropicKey && !excluded.has("anthropic")) {
-    const model = opts.env["ANTHROPIC_MODEL"] ?? DEFAULT_MODELS.anthropic;
-    out["anthropic"] = makeAnthropicProvider(model, anthropicKey, opts, fallbackModelsFor(opts.env, "anthropic"));
-  }
-  const openAiCompatSpec = [
-    { name: "openai", keyEnv: "OPENAI_API_KEY", modelEnv: "OPENAI_MODEL", defaultModel: DEFAULT_MODELS["openai"] },
-    { name: "xai", keyEnv: "XAI_API_KEY", modelEnv: "XAI_MODEL", defaultModel: DEFAULT_MODELS["xai"] },
-    { 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: "qwen", keyEnv: "QWEN_API_KEY", modelEnv: "QWEN_MODEL", defaultModel: DEFAULT_MODELS["qwen"] }
-  ];
-  for (const s of openAiCompatSpec) {
-    const apiKey = opts.env[s.keyEnv];
-    if (!apiKey) continue;
-    if (excluded.has(s.name)) continue;
-    const model = opts.env[s.modelEnv] ?? s.defaultModel;
-    out[s.name] = makeOpenAICompatibleProvider(s.name, model, apiKey, opts, fallbackModelsFor(opts.env, s.name));
-  }
-  const geminiKey = opts.env["GEMINI_API_KEY"];
-  if (geminiKey && !excluded.has("gemini")) {
-    const model = opts.env["GEMINI_MODEL"] ?? DEFAULT_MODELS.gemini;
-    out["gemini"] = makeGeminiProvider(model, geminiKey, opts, fallbackModelsFor(opts.env, "gemini"));
-  }
-  return out;
-}
-function makeAnthropicProvider(model, apiKey, opts, fallbackModels = []) {
-  return {
-    name: "anthropic",
-    model,
-    fallbackModels,
-    send: async (args) => {
-      const sendOpts = {
-        ...args,
-        apiKey,
-        model: args.modelOverride ?? model,
-        pricing: opts.pricing
-      };
-      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
-      return sendAnthropic(sendOpts);
+  },
+  deepseek: {
+    "deepseek-chat": {
+      prompt_per_1k: 27e-5,
+      completion_per_1k: 11e-4,
+      cached_per_1k: 7e-5
+    },
+    "deepseek-reasoner": {
+      prompt_per_1k: 55e-5,
+      completion_per_1k: 22e-4,
+      cached_per_1k: 14e-5
     }
-  };
-}
-function makeOpenAICompatibleProvider(name, model, apiKey, opts, fallbackModels = []) {
-  const url = OPENAI_COMPAT_DEFAULT_URLS[name];
-  return {
-    name,
-    model,
-    fallbackModels,
-    send: async (args) => {
-      const sendOpts = {
-        ...args,
-        provider: name,
-        apiKey,
-        model: args.modelOverride ?? model,
-        url,
-        pricing: opts.pricing
-      };
-      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
-      return sendOpenAICompatible(sendOpts);
+  },
+  kimi: {
+    "kimi-k3": {
+      prompt_per_1k: 3e-3,
+      completion_per_1k: 0.015,
+      cached_per_1k: 3e-4
     }
-  };
-}
-function makeGeminiProvider(model, apiKey, opts, fallbackModels = []) {
-  return {
-    name: "gemini",
-    model,
-    fallbackModels,
-    send: async (args) => {
-      const sendOpts = {
-        ...args,
-        apiKey,
-        model: args.modelOverride ?? model,
-        pricing: opts.pricing
-      };
-      if (opts.fetchImpl) sendOpts.fetchImpl = opts.fetchImpl;
-      return sendGemini(sendOpts);
+  },
+  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).",
+      models: [
+        {
+          provider: "openai",
+          model: "gpt-4o-mini"
+        },
+        {
+          provider: "openai",
+          model: "gpt-5-mini"
+        },
+        {
+          provider: "anthropic",
+          model: "claude-haiku-4-5"
+        },
+        {
+          provider: "anthropic",
+          model: "claude-3-5-haiku"
+        },
+        {
+          provider: "gemini",
+          model: "gemini-3.5-flash"
+        },
+        {
+          provider: "gemini",
+          model: "gemini-2.5-flash"
+        },
+        {
+          provider: "gemini",
+          model: "gemini-2.5-flash-lite"
+        },
+        {
+          provider: "xai",
+          model: "grok-3-mini"
+        },
+        {
+          provider: "groq",
+          model: "llama-3.1-8b-instant"
+        },
+        {
+          provider: "deepseek",
+          model: "deepseek-chat"
+        }
+      ]
+    },
+    med: {
+      description: "Mid-tier models for reasoning, code, multi-step thinking.",
+      models: [
+        {
+          provider: "anthropic",
+          model: "claude-sonnet-4-5"
+        },
+        {
+          provider: "anthropic",
+          model: "claude-sonnet-4-6"
+        },
+        {
+          provider: "openai",
+          model: "gpt-4o"
+        },
+        {
+          provider: "openai",
+          model: "o3-mini"
+        },
+        {
+          provider: "gemini",
+          model: "gemini-3.1-pro-preview"
+        },
+        {
+          provider: "xai",
+          model: "grok-3"
+        },
+        {
+          provider: "mistral",
+          model: "mistral-large-latest"
+        },
+        {
+          provider: "deepseek",
+          model: "deepseek-reasoner"
+        },
+        {
+          provider: "groq",
+          model: "llama-3.3-70b-versatile"
+        }
+      ]
+    },
+    high: {
+      description: "Most capable models for hard reasoning, planning, audits.",
+      models: [
+        {
+          provider: "anthropic",
+          model: "claude-opus-4-5"
+        },
+        {
+          provider: "anthropic",
+          model: "claude-opus-4-7"
+        },
+        {
+          provider: "anthropic",
+          model: "claude-fable-5"
+        },
+        {
+          provider: "openai",
+          model: "gpt-5.5"
+        },
+        {
+          provider: "openai",
+          model: "gpt-5"
+        },
+        {
+          provider: "openai",
+          model: "o1"
+        },
+        {
+          provider: "openai",
+          model: "o3"
+        },
+        {
+          provider: "xai",
+          model: "grok-4-latest"
+        }
+      ]
     }
-  };
-}
+  }
+};
 
 // src/tools/index.ts
 var import_zod = require("zod");
 
 // src/server-meta.ts
 var SERVER_NAME = "crosscheck-agent";
-var SERVER_VERSION = true ? "0.2.30" : "0.0.0-dev";
+var SERVER_VERSION = true ? "0.2.31" : "0.0.0-dev";
 
 // src/tools/audit.ts
 var import_node_fs4 = require("fs");
@@ -2420,33 +2664,6 @@ function operatorCeiling(purpose, provider) {
   return null;
 }
 
-// src/core/retarget.ts
-function retargetProvider(p, newModel, fallbackModels) {
-  const chain = fallbackModels && fallbackModels.length > 0 ? fallbackModels : void 0;
-  if (p.model === newModel && chain === void 0) return p;
-  return {
-    name: p.name,
-    model: newModel,
-    ...chain ? { fallbackModels: chain } : {},
-    send: (args) => p.send({ ...args, modelOverride: newModel })
-  };
-}
-async function loadProviderWeights(storage, names) {
-  const out = {};
-  for (const raw of names) {
-    const name = raw.toLowerCase();
-    if (name in out) continue;
-    const row = await storage.getProviderStats(name);
-    if (!row) {
-      out[name] = 0.5;
-      continue;
-    }
-    const total = row.wins + row.losses + row.abstains;
-    out[name] = total <= 0 ? 0.5 : (row.wins + 0.5 * row.abstains) / total;
-  }
-  return out;
-}
-
 // src/core/model-fallback.ts
 async function sendWithModelFallback(provider, args) {
   const chain = [provider.model, ...provider.fallbackModels ?? []];
@@ -3513,82 +3730,6 @@ function selectForDifficulty(args) {
   };
 }
 
-// src/core/reasoning-upgrade.ts
-var UPGRADE_MODEL = "claude-fable-5";
-var UPGRADE_LABEL = "Fable 5";
-function isUpgradeRequested(args) {
-  if (args["upgrade"] === true) return true;
-  const reasoning = args["reasoning"];
-  if (typeof reasoning === "string" && reasoning.trim().toLowerCase() === "max") {
-    return true;
-  }
-  const model = args["model"];
-  if (typeof model === "string") {
-    const m = model.trim().toLowerCase();
-    if (m === "fable5" || m === "fable-5" || m === "fable" || m === UPGRADE_MODEL) {
-      return true;
-    }
-  }
-  return false;
-}
-var HEAVY_TERMS = [
-  /\bprov(e|ing|en)\b|\bproof\b|\btheorem\b|\bderiv(e|ation)\b/i,
-  /\barchitect(ure|ing)?\b|\bdesign\s+(a|the|an)\b|\bhigh[-\s]?level\s+design\b/i,
-  /\btrade[-\s]?offs?\b|\bpros?\s+and\s+cons?\b|\bweigh\b/i,
-  /\boptimi[sz]e\b|\bcomplexity\b|\bO\(|\balgorithm(ic)?\b/i,
-  /\brace\s+condition\b|\bconcurren(t|cy)\b|\bdeadlock\b|\bdistributed\b|\bconsensus\b/i,
-  /\broot[-\s]?cause\b|\bdebug(ging)?\b|\bwhy\s+(does|is|would|might)\b/i,
-  /\bsecurity\b|\bthreat\s+model\b|\bexploit\b|\bvulnerabilit(y|ies)\b/i,
-  /\bmigrat(e|ion)\b|\brefactor\b|\bre[-\s]?architect\b/i,
-  /\binvariant\b|\bcorrectness\b|\bedge\s+cases?\b|\bformal(ly)?\b/i,
-  /\bstrateg(y|ic|ies)\b|\bplan\s+(the|a|an|out)\b|\broadmap\b/i,
-  /\breason\s+(through|about)\b|\bstep[-\s]?by[-\s]?step\b|\bthink\s+deeply\b/i,
-  /\bnp[-\s]?(hard|complete)\b|\bgraph\b|\bstate\s+machine\b|\bconstraint\b/i
-];
-var DIFFICULTY_TERMS = /\b(hard|tricky|subtle|complex|complicated|non[-\s]?trivial|thorny|gnarly|difficult|deep)\b/i;
-function assessReasoning(text) {
-  const s = typeof text === "string" ? text : "";
-  const signals = [];
-  let score = 0;
-  let termHits = 0;
-  for (const re of HEAVY_TERMS) {
-    if (re.test(s)) termHits += 1;
-  }
-  if (termHits > 0) {
-    const add = Math.min(termHits, 4);
-    score += add;
-    signals.push(`${termHits} reasoning cue${termHits === 1 ? "" : "s"}`);
-  }
-  if (DIFFICULTY_TERMS.test(s)) {
-    score += 1;
-    signals.push("explicit difficulty marker");
-  }
-  const words = s.split(/\s+/).filter(Boolean).length;
-  if (words >= 120) {
-    score += 2;
-    signals.push("long, detailed prompt");
-  } else if (words >= 50) {
-    score += 1;
-    signals.push("substantial prompt");
-  }
-  const constraints = (s.match(/\b(must|should|need to|ensure|require[ds]?|constraint)\b/gi) ?? []).length + (s.match(/(^|\n)\s*(?:[-*]|\d+[.)])\s+/g) ?? []).length;
-  if (constraints >= 3) {
-    score += 1;
-    signals.push(`${constraints} constraints/steps`);
-  }
-  const heavy = score >= 3;
-  const reason = heavy ? `Task looks reasoning-heavy (${signals.join(", ") || "score " + score}).` : `Task looks routine (score ${score}).`;
-  return { heavy, score, reason, signals };
-}
-function buildSuggestedUpgrade(toolName, assessment) {
-  return {
-    model: UPGRADE_MODEL,
-    label: UPGRADE_LABEL,
-    reason: assessment.reason,
-    how: `re-run \`${toolName}\` with reasoning:"max" (or upgrade:true) to route the lead reasoning seat to ${UPGRADE_LABEL}, and \u2014 when an OpenAI key is configured \u2014 add an OpenAI second-opinion pass that ${UPGRADE_LABEL} weighs before finalizing (the result names which model ran). Other panel seats stay on their normal models.`
-  };
-}
-
 // src/tools/audit.ts
 var AUDIT_OBVIOUS_FAILURE_HIGH = 0.3;
 var AUDIT_OBVIOUS_FAILURE_MED = 0.2;
@@ -3886,9 +4027,14 @@ ${rubricText}`;
     );
   }
   const upgradeRequested = isUpgradeRequested(args);
-  const upgraded = upgradeRequested && opts.providers["anthropic"] !== void 0;
-  const auditorForCall = upgraded ? retargetProvider(opts.providers["anthropic"], UPGRADE_MODEL) : auditor;
-  const coProvider = pickCoReasoner(opts.providers, upgradeRequested);
+  const seat = synthesisSeat({ moderator: auditor, pinned: false, superRequested: false, upgradeRequested });
+  const upgraded = seat.upgraded;
+  const auditorForCall = seat.provider;
+  const coProvider = pickCoReasoner(
+    opts.providers,
+    upgradeRequested,
+    auditorForCall.name === "openai" ? auditorForCall.model : void 0
+  );
   const coReason = await coReasonThenInject({
     provider: coProvider,
     messages: msgs,
@@ -3971,15 +4117,16 @@ ${rubricText}`;
   if (upgraded) {
     result["reasoning_upgrade"] = {
       applied: true,
-      model: UPGRADE_MODEL,
-      label: UPGRADE_LABEL,
+      provider: auditorForCall.name,
+      model: auditorForCall.model,
+      label: seat.label ?? auditorForCall.model,
       seat: "auditor",
       ...coReasoned ? { co_reasoner: { model: coProvider.model, label: coReasonLabel(coProvider.model) } } : {}
     };
   } else if (upgradeRequested) {
     result["reasoning_upgrade"] = {
       applied: false,
-      reason: `no Anthropic key available \u2014 the ${UPGRADE_LABEL} upgrade needs it.`
+      reason: seat.skipReason ?? "the auditor couldn't be upgraded."
     };
   } else {
     const assessment = assessReasoning(
@@ -4062,7 +4209,7 @@ function pickAuditor(providers, exclude, explicit, moderatorName, allowlist, che
       }
     }
   }
-  const mod = providers[moderatorName.toLowerCase()];
+  const mod = resolveModerator(providers, moderatorName).provider;
   if (mod !== void 0 && !exclude.has(mod.name) && (allowlist === null || allowlist.includes(mod.name))) {
     return { auditor: mod, reason: null };
   }
@@ -4381,9 +4528,7 @@ function pickPanelJudge(providers, moderatorName, pricing) {
       return entry.model && entry.model !== base.model ? retargetProvider(base, entry.model) : base;
     }
   }
-  const mod = providers[moderatorName.toLowerCase()];
-  if (mod !== void 0) return mod;
-  return null;
+  return resolveModerator(providers, moderatorName).provider ?? null;
 }
 var AGREEMENT_CHECK_SCHEMA = {
   type: "object",
@@ -4700,80 +4845,6 @@ async function pickAutoPanel(storage, purpose, n, providers, allowlist) {
   return { providers: ranked.slice(0, Math.max(1, n)), coldStart: true };
 }
 
-// src/core/super-mode.ts
-var import_node_async_hooks = require("async_hooks");
-var SUPER_MODELS = {
-  // Fable 5.1 verified present on the account; 5 as the fallback for accounts
-  // that haven't been granted 5.1 yet.
-  anthropic: ["claude-fable-5-1", UPGRADE_MODEL],
-  // Astra is live (verified against the account's model list 2026-09-23).
-  // The fallback is named explicitly rather than reusing CO_REASON_MODEL:
-  // that constant means "OpenAI's top reasoning model" for the max-reasoning
-  // co-reasoner, which is a different question from "what should super fall
-  // back to". Sharing the symbol silently coupled two unrelated decisions.
-  openai: ["gpt-6-astra", "gpt-6-sol"],
-  // Explicitly left as-is per product decision. Note grok-4-latest is not in
-  // xAI's published model list, but it resolves in practice — real panel calls
-  // on it succeeded as recently as 2026-08-30 — so it is an unlisted alias
-  // rather than a dead id.
-  xai: [DEFAULT_MODELS["xai"]],
-  // 3.8 Flash is now BOTH the everyday default and the super pick for Gemini,
-  // so this can no longer borrow DEFAULT_MODELS as its fallback — that would
-  // repeat the same model and retry one that just failed. Named explicitly.
-  gemini: ["gemini-3.8-flash", "gemini-3.1-pro-preview"],
-  kimi: [DEFAULT_MODELS["kimi"]],
-  qwen: [DEFAULT_MODELS["qwen"]]
-};
-var SUPER_ENV_VARS = {
-  anthropic: "ANTHROPIC_SUPER_MODEL",
-  openai: "OPENAI_SUPER_MODEL",
-  xai: "XAI_SUPER_MODEL",
-  gemini: "GEMINI_SUPER_MODEL",
-  kimi: "KIMI_SUPER_MODEL",
-  qwen: "QWEN_SUPER_MODEL",
-  mistral: "MISTRAL_SUPER_MODEL",
-  groq: "GROQ_SUPER_MODEL",
-  deepseek: "DEEPSEEK_SUPER_MODEL"
-};
-function parseChain(v) {
-  if (!v) return [];
-  const out = [];
-  for (const raw of v.split(",")) {
-    const m = raw.trim();
-    if (m && !out.includes(m)) out.push(m);
-  }
-  return out;
-}
-var superEnvStore = new import_node_async_hooks.AsyncLocalStorage();
-var currentEnv = () => superEnvStore.getStore() ?? process.env;
-function superChainFor(provider, env = currentEnv()) {
-  const varName = SUPER_ENV_VARS[provider];
-  const user = varName ? parseChain(env[varName]) : [];
-  if (user.length > 0) {
-    const fallbacks = varName ? parseChain(env[`${varName}_FALLBACKS`]) : [];
-    return [...user, ...fallbacks.filter((m) => !user.includes(m))];
-  }
-  return SUPER_MODELS[provider] ?? [];
-}
-function superPrimary(provider, env = currentEnv()) {
-  return superChainFor(provider, env)[0];
-}
-var SUPER_PROVIDER_NAMES = Object.keys(SUPER_MODELS);
-function isSuperRequested(args) {
-  return args["super"] === true;
-}
-function retargetForSuper(p, env = currentEnv()) {
-  const chain = superChainFor(p.name, env);
-  if (chain.length === 0) return p;
-  const [primary, ...fallbacks] = chain;
-  return retargetProvider(p, primary, fallbacks);
-}
-var SUPER_BUDGET_MULTIPLIER = 3;
-var SUPER_BUDGET_CEILING = 16e3;
-function superBudget(base) {
-  return Math.min(base * SUPER_BUDGET_MULTIPLIER, SUPER_BUDGET_CEILING);
-}
-
 // src/core/session-memory.ts
 var SESSION_MEMORY_DEFAULT_LIMIT = 50;
 var SESSION_MEMORY_INJECT_BUDGET_CHARS = 4e3;
@@ -7124,8 +7195,9 @@ async function runOrchestrate(args, opts) {
       blocked
     };
   }
-  const moderatorName = typeof args["moderator"] === "string" && args["moderator"] ? args["moderator"] : opts.moderator ?? "anthropic";
-  const moderator = opts.providers[moderatorName.toLowerCase()] ?? selected[0];
+  const resolvedModerator = resolveModerator(opts.providers, args["moderator"], opts.moderator);
+  const moderator = resolvedModerator.provider ?? selected[0];
+  const moderatorChosen = resolvedModerator.provider !== void 0 && (resolvedModerator.pinned || parseModeratorSpec(args["moderator"]) !== null);
   const failFast = boolArg2(args["fail_fast"], false);
   const planOnly = boolArg2(args["plan_only"], false);
   const cheapMode = boolArg2(args["cheap_mode"], opts.ctx?.cheap_mode ?? false);
@@ -7609,11 +7681,17 @@ ${recombineSys}` : recombineSys },
   let synthProvider = moderator;
   let synthModel = null;
   const upgradeRequested = isUpgradeRequested(args);
-  const upgraded = upgradeRequested && opts.providers["anthropic"] !== void 0;
+  const seat = synthesisSeat({
+    moderator,
+    pinned: resolvedModerator.provider !== void 0 && resolvedModerator.pinned,
+    superRequested: false,
+    upgradeRequested
+  });
+  const upgraded = seat.upgraded;
   if (upgraded) {
-    synthProvider = retargetProvider(opts.providers["anthropic"], UPGRADE_MODEL);
-    synthModel = `anthropic/${UPGRADE_MODEL}`;
-  } else if (cheapMode && opts.pricing) {
+    synthProvider = seat.provider;
+    synthModel = `${seat.provider.name}/${seat.provider.model}`;
+  } else if (cheapMode && opts.pricing && !moderatorChosen) {
     const pick = selectForDifficulty({
       pricing: opts.pricing,
       tier: "med",
@@ -7639,7 +7717,11 @@ ${recombineSys}` : recombineSys },
       provider: synthProvider.name
     } }
   );
-  const coProvider = pickCoReasoner(opts.providers, upgradeRequested);
+  const coProvider = pickCoReasoner(
+    opts.providers,
+    upgradeRequested,
+    synthProvider.name === "openai" ? synthProvider.model : void 0
+  );
   const coReason = await coReasonThenInject({
     provider: coProvider,
     messages: recMsgs,
@@ -7689,7 +7771,8 @@ ${recombineSys}` : recombineSys },
     missing,
     partial,
     fail_fast: failFast,
-    cheap_mode: cheapMode
+    cheap_mode: cheapMode,
+    moderator: { provider: moderator.name, model: moderator.model }
   };
   if (personaMeta.used !== null) result["persona"] = personaMeta;
   if (synthModel) result["synth_model"] = synthModel;
@@ -7697,15 +7780,16 @@ ${recombineSys}` : recombineSys },
   if (upgraded) {
     result["reasoning_upgrade"] = {
       applied: true,
-      model: UPGRADE_MODEL,
-      label: UPGRADE_LABEL,
+      provider: seat.provider.name,
+      model: seat.provider.model,
+      label: seat.label ?? seat.provider.model,
       seat: "recombine/synthesis",
       ...coReasoned ? { co_reasoner: { model: coProvider.model, label: coReasonLabel(coProvider.model) } } : {}
     };
   } else if (upgradeRequested) {
     result["reasoning_upgrade"] = {
       applied: false,
-      reason: `no Anthropic key available \u2014 the ${UPGRADE_LABEL} upgrade needs it.`
+      reason: seat.skipReason ?? "the moderator couldn't be upgraded."
     };
   } else if (goal) {
     const assessment = assessReasoning(goal);
@@ -8887,13 +8971,24 @@ async function runCoordinate(args, opts) {
     };
   }
   const proposerName = typeof args["proposer"] === "string" && args["proposer"] ? args["proposer"] : selected[0].name;
-  const synthName = typeof args["synthesizer"] === "string" && args["synthesizer"] ? args["synthesizer"] : typeof args["moderator"] === "string" && args["moderator"] ? args["moderator"] : opts.moderator ?? "anthropic";
+  const resolvedSynth = resolveModerator(
+    opts.providers,
+    args["synthesizer"],
+    args["moderator"],
+    opts.moderator
+  );
   const proposer = opts.providers[proposerName.toLowerCase()] ?? selected[0];
-  const synth = opts.providers[synthName.toLowerCase()] ?? proposer;
+  const synth = resolvedSynth.provider ?? proposer;
   const proposerForCall = superRequested ? retargetForSuper(proposer) : proposer;
   const upgradeRequested = isUpgradeRequested(args) || superRequested;
-  const upgraded = upgradeRequested && opts.providers["anthropic"] !== void 0;
-  const synthForCall = superRequested && opts.providers["anthropic"] ? retargetForSuper(opts.providers["anthropic"]) : upgraded ? retargetProvider(opts.providers["anthropic"], UPGRADE_MODEL) : synth;
+  const seat = synthesisSeat({
+    moderator: synth,
+    pinned: resolvedSynth.provider !== void 0 && resolvedSynth.pinned,
+    superRequested,
+    upgradeRequested
+  });
+  const upgraded = seat.upgraded;
+  const synthForCall = seat.provider;
   let critics = [];
   if (Array.isArray(args["critics"])) {
     for (const n of args["critics"]) {
@@ -9044,7 +9139,11 @@ CRITIQUES:
 ${critiqueBlock}`
     }
   ];
-  const coProvider = pickCoReasoner(opts.providers, upgradeRequested);
+  const coProvider = pickCoReasoner(
+    opts.providers,
+    upgradeRequested,
+    synthForCall.name === "openai" ? synthForCall.model : void 0
+  );
   const coReason = await coReasonThenInject({
     provider: coProvider,
     messages: synthMessages,
@@ -9089,7 +9188,8 @@ ${critiqueBlock}`
     critique_answers: scannedCritiques,
     critique_structured: critiqueStructured,
     synthesis_answer: synthAns,
-    synthesis_structured: synthObj
+    synthesis_structured: synthObj,
+    moderator: { provider: synthForCall.name, model: synthForCall.model }
   };
   if (synthErrs.length > 0) result["synthesis_errors"] = synthErrs;
   if (canaryLeaks.length > 0) result["canary_leaks"] = canaryLeaks;
@@ -9107,15 +9207,16 @@ ${critiqueBlock}`
     result["reasoning_upgrade"] = {
       applied: true,
       // The model that actually ran the seat, not the one we'd have picked.
-      model: superRequested ? superPrimary("anthropic") ?? UPGRADE_MODEL : UPGRADE_MODEL,
-      label: UPGRADE_LABEL,
+      provider: synthForCall.name,
+      model: synthForCall.model,
+      label: seat.label ?? synthForCall.model,
       seat: "synthesizer",
       ...coReasonAns ? { co_reasoner: { model: coProvider.model, label: coReasonLabel(coProvider.model) } } : {}
     };
   } else if (upgradeRequested) {
     result["reasoning_upgrade"] = {
       applied: false,
-      reason: `no Anthropic key available \u2014 the ${UPGRADE_LABEL} upgrade needs it.`
+      reason: seat.skipReason ?? "the synthesizer couldn't be upgraded."
     };
   } else {
     const assessment = assessReasoning(topic);
@@ -9721,10 +9822,12 @@ async function runDebate(args, opts) {
   const earlyStopThreshold = Number.isFinite(earlyStopThresholdRaw) ? earlyStopThresholdRaw : 0.7;
   let earlyStoppedRound = null;
   let agreementBlock = null;
-  const moderatorName = typeof args["moderator"] === "string" && args["moderator"] ? args["moderator"] : opts.moderator ?? "anthropic";
-  const moderator = opts.providers[moderatorName.toLowerCase()] ?? selected[0];
+  const resolvedModerator = resolveModerator(opts.providers, args["moderator"], opts.moderator);
+  const moderatorName = resolvedModerator.spec.provider;
+  const moderator = resolvedModerator.provider ?? selected[0];
   const upgradeRequested = isUpgradeRequested(args) || superRequested;
-  const upgraded = upgradeRequested && opts.providers["anthropic"] !== void 0;
+  const seat = moderator ? synthesisSeat({ moderator, pinned: resolvedModerator.pinned, superRequested, upgradeRequested }) : null;
+  const upgraded = seat?.upgraded ?? false;
   for (let rnd = 1; rnd <= maxRounds; rnd++) {
     const roundMessages = [
       {
@@ -9809,8 +9912,10 @@ ${prior.body}` });
   let personaMeta = null;
   let coReasoned = false;
   let coReasonerModel = null;
+  let moderatorUsed = null;
   if (moderator) {
-    const synthProvider = superRequested && opts.providers["anthropic"] ? retargetForSuper(opts.providers["anthropic"]) : upgraded ? retargetProvider(opts.providers["anthropic"], UPGRADE_MODEL) : moderator;
+    const synthProvider = seat?.provider ?? moderator;
+    moderatorUsed = { provider: synthProvider.name, model: synthProvider.model };
     const condensed = buildModeratorTranscript(transcript);
     const persona = buildPersonaInjection({
       toolName: "debate",
@@ -9836,7 +9941,7 @@ TRANSCRIPT:
 ${condensed}`
       }
     ];
-    const openAiSeat = selected.find((p) => p.name === "openai")?.model;
+    const openAiSeat = synthProvider.name === "openai" ? synthProvider.model : selected.find((p) => p.name === "openai")?.model;
     const coProvider = pickCoReasoner(opts.providers, upgradeRequested, openAiSeat);
     const coReason = await coReasonThenInject({
       provider: coProvider,
@@ -9896,20 +10001,22 @@ ${condensed}`
     transcript,
     synthesis
   };
+  if (moderatorUsed) result["moderator"] = moderatorUsed;
   if (personaMeta && personaMeta.used !== null) result["persona"] = personaMeta;
   if (compactionMeta) result["round_compaction"] = compactionMeta;
   if (upgraded) {
     result["reasoning_upgrade"] = {
       applied: true,
-      model: superRequested ? superPrimary("anthropic") ?? UPGRADE_MODEL : UPGRADE_MODEL,
-      label: UPGRADE_LABEL,
+      provider: seat.provider.name,
+      model: seat.provider.model,
+      label: seat.label ?? seat.provider.model,
       seat: "moderator synthesis",
       ...coReasoned && coReasonerModel ? { co_reasoner: { model: coReasonerModel, label: coReasonLabel(coReasonerModel) } } : {}
     };
   } else if (upgradeRequested) {
     result["reasoning_upgrade"] = {
       applied: false,
-      reason: `no Anthropic key available \u2014 the ${UPGRADE_LABEL} upgrade needs it.`
+      reason: seat?.skipReason ?? "no moderator was available for the upgrade."
     };
   } else {
     const assessment = assessReasoning(topic);
@@ -10538,12 +10645,25 @@ function runListProviders(args, opts) {
   }
   const result = {
     providers,
-    moderator_default: opts.moderatorDefault ?? "anthropic",
+    moderator_default: opts.moderatorDefault ?? DEFAULT_MODERATOR,
+    // Who will actually moderate right now: the configured spec if that
+    // provider has a key, else the first available (what the tools fall to).
+    moderator: describeModerator(opts.providers ?? {}, opts.moderatorDefault),
     usage_hint: USAGE_HINT
   };
   if (superRequested) result["super_mode"] = true;
   return result;
 }
+function describeModerator(providers, spec) {
+  const r = resolveModerator(providers, spec);
+  const p = r.provider ?? Object.values(providers)[0];
+  return {
+    spec: spec ?? DEFAULT_MODERATOR,
+    provider: p?.name ?? null,
+    model: p?.model ?? null,
+    available: r.provider !== void 0
+  };
+}
 
 // src/tools/models.ts
 var KNOWN_PROVIDERS7 = [
@@ -12093,7 +12213,7 @@ var CHECK_INTERVAL_SECONDS = 3 * 24 * 60 * 60;
 var DEFAULT_PACKAGE = "crosscheck-cli";
 var FETCH_TIMEOUT_MS = 3e3;
 function engineVersion() {
-  return true ? "0.2.30" : "0.0.0-dev";
+  return true ? "0.2.31" : "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");
@@ -12538,7 +12658,7 @@ var HELP_CATALOG = [
     primaryArg: "topic",
     summary: "Bounded multi-round debate between models, ending in a synthesized recommendation.",
     example: "xc debate should we adopt a monorepo?",
-    detail: `The panel argues across a few rounds \u2014 each model sees the others' prior turns, concedes or pushes back \u2014 then a moderator synthesizes a single grounded recommendation. Supports reasoning:"max" to bump the moderator synthesis seat to Fable 5. Great for contentious or trade-off-heavy calls.`
+    detail: `The panel argues across a few rounds \u2014 each model sees the others' prior turns, concedes or pushes back \u2014 then a moderator synthesizes a single grounded recommendation. Any model can moderate \u2014 pass moderator:"openai" or an exact model (moderator:"openai:gpt-6-sol"), or set a default with \`crosscheck moderator set\`. reasoning:"max" bumps the moderator's synthesis to its provider's flagship (Fable 5 for Anthropic). Great for contentious or trade-off-heavy calls.`
   },
   {
     name: "coordinate",
@@ -12781,6 +12901,10 @@ function renderHelpCommand(entry) {
 }
 
 // src/tools/index.ts
+var MODERATOR_PARAM = {
+  type: "string",
+  description: 'Who moderates (synthesises) this run: a provider ("openai") or an exact model ("openai:gpt-6-sol"). Any configured provider works. Defaults to the configured moderator (CROSSCHECK_MODERATOR / config `moderator`).'
+};
 function defineTool(opts) {
   return {
     name: opts.name,
@@ -13015,7 +13139,7 @@ function createTool(o, toolName, cheapDefault) {
         instruction: { type: "string" },
         documents: { type: "array", items: { type: "string" } },
         providers: { type: "array", items: { type: "string" } },
-        moderator: { type: "string" },
+        moderator: MODERATOR_PARAM,
         constraints: { type: "string" },
         target_path: { type: "string" },
         session_id: { type: "string" },
@@ -13056,7 +13180,7 @@ function orchestrateTool(providers, allowlist, moderator, storage, pricing, brea
         dag: { type: "object", additionalProperties: true },
         context: { type: "string" },
         providers: { type: "array", items: { type: "string" } },
-        moderator: { type: "string" },
+        moderator: MODERATOR_PARAM,
         fail_fast: { type: "boolean" },
         plan_only: { type: "boolean" },
         cheap_mode: { type: "boolean" },
@@ -13435,7 +13559,7 @@ function planTool(providers, allowlist, storage, breakers, transcriptsDir, repoR
         constraints: { type: "string" },
         context: { type: "string" },
         providers: { type: "array", items: { type: "string" } },
-        moderator: { type: "string" },
+        moderator: MODERATOR_PARAM,
         session_id: { type: "string" },
         structured: { type: "boolean" },
         mode: { type: "string", enum: ["fast", "thorough"] },
@@ -13497,7 +13621,7 @@ function coordinateTool(providers, allowlist, innerCallers, storage, breakers, t
         providers: { type: "array", items: { type: "string" } },
         proposer: { type: "string" },
         synthesizer: { type: "string" },
-        moderator: { type: "string" },
+        moderator: MODERATOR_PARAM,
         critics: { type: "array", items: { type: "string" } },
         session_id: { type: "string" },
         untrusted_input: { type: "boolean" },
@@ -13530,7 +13654,7 @@ function debateTool(providers, allowlist, innerCallers, storage, pricing, breake
         context: { type: "string" },
         max_rounds: { type: "integer", minimum: 1 },
         providers: { type: "array", items: { type: "string" } },
-        moderator: { type: "string" },
+        moderator: MODERATOR_PARAM,
         session_id: { type: "string" },
         reasoning: { type: "string", enum: ["default", "max"], description: '"max" bumps the moderator-synthesis seat to Fable 5 (premium thinking model); the debaters stay on their normal models. Omit for the Opus 4.8 default.' },
         structured: { type: "boolean" },
@@ -13747,6 +13871,7 @@ var KEY_ENV = {
   qwen: "QWEN_API_KEY"
 };
 function createCrosscheck(opts) {
+  const moderator = opts.moderator ?? DEFAULT_MODERATOR;
   const env = {};
   for (const [provider, key] of Object.entries(opts.keys)) {
     const varName = KEY_ENV[provider.trim().toLowerCase()];
@@ -13754,7 +13879,7 @@ function createCrosscheck(opts) {
   }
   const pricingDoc = pricing_default;
   const providers = buildProviders({ env, pricing: pricingDoc });
-  const tools = registerCoreTools({ providers, pricing: pricingDoc });
+  const tools = registerCoreTools({ providers, pricing: pricingDoc, moderatorDefault: moderator });
   return {
     listTools: () => [...tools.keys()],
     call: (name, args) => {
@@ -13767,11 +13892,11 @@ function createCrosscheck(opts) {
     // same opts the registry wires for confer in a browser build: providers
     // + pricing + default moderator; storage/innerCallers aren't
     // available here, matching the non-streaming path.
-    conferStream: (args, onAnswer) => runConfer(args, { providers, pricing: pricingDoc, moderator: "anthropic", onAnswer }),
+    conferStream: (args, onAnswer) => runConfer(args, { providers, pricing: pricingDoc, moderator, onAnswer }),
     debateStream: (args, { onTurn, onSynthesis } = {}) => runDebate(args, {
       providers,
       pricing: pricingDoc,
-      moderator: "anthropic",
+      moderator,
       ...onTurn ? { onTurn: (entry, round, provider) => onTurn(entry, round, provider) } : {},
       ...onSynthesis ? { onSynthesis: (s) => onSynthesis(s) } : {}
     })