@iowarp/clio-coder 0.4.1 → 0.4.3

package/dist/{chunk-UEDMSP56.js → chunk-XPWWI35G.js}

@@ -3,65 +3,62 @@ import {
   createSentinelStripper,
   stripTokenizerSentinels
 } from "./chunk-CFGTUFWB.js";
-import {
-  credentialsPresent
-} from "./chunk-72GZI5EV.js";
 import {
   ceilChars,
   estimateAgentMessageTokens,
   toolSchemaChars
-} from "./chunk-CWVRRIEI.js";
-import {
-  getSharedBus
-} from "./chunk-PMZVFWSR.js";
+} from "./chunk-NZMNUPZZ.js";
 import {
-  BusChannels
-} from "./chunk-HLW2MRKE.js";
+  credentialsPresent
+} from "./chunk-2JH2WHGE.js";
 import {
   readSettings
-} from "./chunk-EIMVLWB3.js";
+} from "./chunk-GHS5EBTQ.js";
 import {
   warnLegacyNaming
 } from "./chunk-D7E7WCDF.js";
 import {
-  sleep
-} from "./chunk-EBY63B7G.js";
+  getSharedBus
+} from "./chunk-PMZVFWSR.js";
+import {
+  BusChannels
+} from "./chunk-F4EKGO4N.js";
 import {
   CLIO_CONTEXT_WINDOW_WARN_BELOW,
   CLIO_MIN_CONTEXT_WINDOW,
   CLIO_MIN_MAX_OUTPUT_TOKENS,
+  EMPTY_CAPABILITIES,
+  VALID_THINKING_LEVELS,
+  availableThinkingLevels,
   extractLocalModelQuirks,
+  inferLocalModelFamily,
+  isHarmonyModelId,
   listLmStudioModels,
   lmStudioReasoningEffort,
   lmStudioRootUrl,
   registerBuiltinRuntimes,
   requestLmStudioJson,
   resolveLmStudioInstance
-} from "./chunk-QKIFBZKT.js";
+} from "./chunk-VW6DOEDG.js";
+import {
+  sleep
+} from "./chunk-EBY63B7G.js";
 import {
   listKnownModelsForRuntime
-} from "./chunk-CDNVLKUX.js";
+} from "./chunk-LLDJM5XK.js";
 import {
   authNotRequiredStatus,
   openAuthStorage,
   registerClioOAuthProviders,
   resolveAuthTarget,
   targetRequiresAuth
-} from "./chunk-WNP7O5WZ.js";
+} from "./chunk-ID64D7PE.js";
 import {
   withStateFileLock
 } from "./chunk-IKCO5N3L.js";
 import {
   require_dist
 } from "./chunk-APJ265NV.js";
-import {
-  clioCacheDir,
-  clioConfigDir,
-  clioStateDir,
-  clioStatePath,
-  resolveClioDirs,
-  stateRootRemoved
-} from "./chunk-BYP5D4HI.js";
 import {
   safeResourceWrite
 } from "./chunk-WEPFGWHJ.js";
@@ -70,17 +67,24 @@ import {
 } from "./chunk-NUGM5KR6.js";
 import {
   capabilitiesFromCatalogModel,
-  catalogThinkingLevelsForRuntime,
   getCatalogModelForRuntime,
   mergeCapabilities,
   resolveCostProvenance
-} from "./chunk-O42A54GG.js";
+} from "./chunk-5OIVVPHF.js";
 import {
   activateExternalPluginApiBridge,
   calculateEngineCost,
   ensurePiAiRegistered,
   registerEngineApiProvider
-} from "./chunk-VXMFAE2W.js";
+} from "./chunk-YPC6ZR5L.js";
+import {
+  clioCacheDir,
+  clioConfigDir,
+  clioStateDir,
+  clioStatePath,
+  resolveClioDirs,
+  stateRootRemoved
+} from "./chunk-BYP5D4HI.js";
 import {
   __toESM,
   init_esm_shims
@@ -430,7 +434,7 @@ function evictionNotice(facts, entry) {
     facts,
     "swap",
     "warning",
-    `swapping resident '${entry.modelId}' for requested '${facts.keepModelId}' on '${facts.targetId}' (Clio did not load it; recorded transition instead of a silent unload; set CLIO_CODER_RESIDENCY=observe or lifecycle: user-managed to forbid swaps).`,
+    `swapping resident '${entry.modelId}' for requested '${facts.keepModelId}' on '${facts.targetId}' (Clio did not load it; recorded transition instead of a silent unload; set lifecycle: user-managed to forbid swaps).`,
     {
       ...entry.sizeVramBytes !== void 0 ? { freedVramBytes: entry.sizeVramBytes } : {},
       swappedOut: entry.modelId
@@ -733,16 +737,8 @@ async function reconcileResidency(adapter) {
   }
   return plan;
 }
-function residencyManaged(env = process.env) {
-  const opt = (env.CLIO_CODER_RESIDENCY ?? "").trim().toLowerCase();
-  if (opt === "observe" || opt === "off" || opt === "0" || opt === "false" || opt === "user" || opt === "user-managed") {
-    return false;
-  }
-  return true;
-}
-function residencyManagedFor(lifecycle, env = process.env) {
-  if (lifecycle === "user-managed") return false;
-  return residencyManaged(env);
+function residencyManagedFor(lifecycle) {
+  return lifecycle !== "user-managed";
 }
 
 // src/core/run-overrides.ts
@@ -790,97 +786,6 @@ async function withRunOverrides(overrides, fn) {
   }
 }
 
-// src/domains/providers/types/capability-flags.ts
-init_esm_shims();
-
-// src/domains/providers/model-family.ts
-init_esm_shims();
-function normalizeModelIdForFamily(modelId) {
-  return (modelId ?? "").trim().toLowerCase().replace(/_/g, "-");
-}
-function isHarmonyModelId(modelId) {
-  return normalizeModelIdForFamily(modelId).includes("gpt-oss");
-}
-function inferLocalModelFamily(modelId) {
-  const normalized = normalizeModelIdForFamily(modelId);
-  if (normalized.length === 0) return "unknown";
-  if (normalized.includes("gpt-oss")) return "openai-gpt-oss";
-  if (normalized.includes("nemotron-cascade-2") || normalized.includes("nemotron-cascade2")) {
-    return "nemotron-cascade-2";
-  }
-  if (normalized.includes("nemotron-3-nano-omni") || normalized.includes("nemotron-nano-omni")) {
-    return "nemotron-3-nano-omni";
-  }
-  if (normalized.includes("qwen3")) return "qwen3";
-  if (normalized.includes("gemma-4") || normalized.includes("gemma4") || normalized.includes("gemopus")) {
-    return "gemma-4";
-  }
-  return "unknown";
-}
-
-// src/domains/providers/types/capability-flags.ts
-var EMPTY_CAPABILITIES = {
-  chat: false,
-  tools: false,
-  reasoning: false,
-  vision: false,
-  audio: false,
-  embeddings: false,
-  rerank: false,
-  fim: false,
-  contextWindow: 0,
-  maxTokens: 0
-};
-var VALID_THINKING_LEVELS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"];
-var THINKING_LEVELS_WITHOUT_XHIGH = ["off", "minimal", "low", "medium", "high"];
-var THINKING_LEVELS_OPENAI_5_1_MINI = ["off", "minimal", "low", "medium", "high"];
-var THINKING_LEVELS_OPENAI_5_2_PLUS = [
-  "off",
-  "minimal",
-  "low",
-  "medium",
-  "high",
-  "xhigh"
-];
-var THINKING_LEVELS_OPENAI_5_6 = [...THINKING_LEVELS_OPENAI_5_2_PLUS, "max"];
-var THINKING_LEVELS_HARMONY = ["low", "medium", "high"];
-function normalizeModelId(modelId) {
-  if (!modelId) return void 0;
-  const trimmed = modelId.trim();
-  if (trimmed.length === 0) return void 0;
-  return trimmed.includes("/") ? trimmed.split("/").pop() ?? trimmed : trimmed;
-}
-function availableOpenAICodexThinkingLevels(modelId) {
-  const normalized = normalizeModelId(modelId);
-  if (!normalized) return VALID_THINKING_LEVELS;
-  if (normalized === "gpt-5.1-codex-mini") return THINKING_LEVELS_OPENAI_5_1_MINI;
-  if (normalized.startsWith("gpt-5.2") || normalized.startsWith("gpt-5.3") || normalized.startsWith("gpt-5.4") || normalized.startsWith("gpt-5.5")) {
-    return THINKING_LEVELS_OPENAI_5_2_PLUS;
-  }
-  if (normalized.startsWith("gpt-5.6")) {
-    return THINKING_LEVELS_OPENAI_5_6;
-  }
-  return THINKING_LEVELS_WITHOUT_XHIGH;
-}
-function availableThinkingLevels(caps, options) {
-  if (!caps.reasoning) return ["off"];
-  if (caps.thinkingFormat === "harmony" || isHarmonyModelId(options?.modelId)) {
-    return THINKING_LEVELS_HARMONY;
-  }
-  const catalogLevels = options?.runtimeId && options.modelId ? catalogThinkingLevelsForRuntime(options.runtimeId, options.modelId) : void 0;
-  if (catalogLevels) {
-    if (!catalogLevels.includes("xhigh")) return catalogLevels.filter((level) => level !== "max");
-    return catalogLevels;
-  }
-  if (caps.thinkingFormat === "openai-codex" || options?.runtimeId === "openai-codex") {
-    return availableOpenAICodexThinkingLevels(options?.modelId);
-  }
-  if (caps.thinkingFormat === "anthropic-extended") {
-    return THINKING_LEVELS_WITHOUT_XHIGH;
-  }
-  return THINKING_LEVELS_WITHOUT_XHIGH;
-}
-
 // src/domains/providers/model-capabilities.ts
 init_esm_shims();
 function normalizedModelId(wireModelId) {
@@ -892,6 +797,12 @@ function applyModelCapabilityPatch(model, caps) {
   if (typeof caps.contextWindow === "number") model.contextWindow = caps.contextWindow;
   if (typeof caps.maxTokens === "number") model.maxTokens = caps.maxTokens;
   if (typeof caps.reasoning === "boolean") model.reasoning = caps.reasoning;
+  if (model.clioCoder || caps.thinkingControlRuntime !== void 0) {
+    const metadata2 = { ...model.clioCoder };
+    delete metadata2.thinkingControlRuntime;
+    if (caps.thinkingControlRuntime !== void 0) metadata2.thinkingControlRuntime = caps.thinkingControlRuntime;
+    model.clioCoder = metadata2;
+  }
   return model;
 }
 function applyReasoningResolution(caps, kbHit, detectedReasoning) {
@@ -1140,14 +1051,15 @@ function supportedBudgetLevels(baseLevels, quirks) {
 function supportedEffortLevels(baseLevels, quirks, harmony) {
   if (harmony) return HARMONY_LEVELS;
   const efforts = quirks?.thinking?.effortByLevel;
-  if (!efforts) return baseLevels;
+  const byLevelValues = (quirks?.chatTemplateKwargs ?? quirks?.thinking?.chatTemplateKwargs)?.byLevel?.values;
+  if (!efforts && !byLevelValues) return baseLevels;
   const out = [];
   if (baseLevels.includes("off")) out.push("off");
-  if (efforts.minimal !== void 0) out.push("minimal");
-  if (efforts.low !== void 0) out.push("low");
-  if (efforts.medium !== void 0) out.push("medium");
-  if (efforts.high !== void 0) out.push("high");
-  if (efforts.xhigh !== void 0) out.push("xhigh");
+  if (efforts?.minimal !== void 0 || byLevelValues?.minimal !== void 0) out.push("minimal");
+  if (efforts?.low !== void 0 || byLevelValues?.low !== void 0) out.push("low");
+  if (efforts?.medium !== void 0 || byLevelValues?.medium !== void 0) out.push("medium");
+  if (efforts?.high !== void 0 || byLevelValues?.high !== void 0) out.push("high");
+  if (efforts?.xhigh !== void 0 || byLevelValues?.xhigh !== void 0) out.push("xhigh");
   return out.length > 0 ? out : baseLevels;
 }
 function restrictThinkingLevelsByMechanism(levels, mechanism, quirks, options) {
@@ -1162,7 +1074,8 @@ function effectiveThinkingLevel(configured, available) {
   const fallback = available[0] ?? "off";
   if (!configured) return fallback;
   if (available.includes(configured)) return configured;
-  if (configured === "max" && available.includes("xhigh")) return "xhigh";
+  if ((configured === "max" || configured === "high" && !available.includes("high")) && available.includes("xhigh"))
+    return "xhigh";
   if ((configured === "high" || configured === "xhigh" || configured === "max") && available.includes("high")) {
     return "high";
   }
@@ -1283,12 +1196,45 @@ function resolveThinkingCapability(input, quirks, parser) {
     budgetEnforcement
   };
 }
-var REASONING_EFFORT_ON_OFF_RUNTIMES = /* @__PURE__ */ new Set(["lmstudio"]);
+var REASONING_EFFORT_ONLY_RUNTIMES = /* @__PURE__ */ new Set(["lmstudio"]);
 function onOffReasoningEffort(thinkingActive) {
   return thinkingActive ? "low" : "none";
 }
-function resolveRequestCapability(thinking, parser, runtimeId) {
+function resolveFamilyChatTemplateKwargs(quirks, effectiveLevel, configuredLevel) {
+  const block = quirks?.chatTemplateKwargs ?? quirks?.thinking?.chatTemplateKwargs;
+  if (!block) return void 0;
+  const out = {};
+  if (block.static) {
+    for (const [k, v] of Object.entries(block.static)) {
+      out[k] = v;
+    }
+  }
+  if (block.byLevel) {
+    const byLevel = block.byLevel;
+    const val = byLevel.values[effectiveLevel] ?? byLevel.values[configuredLevel];
+    if (val !== void 0) {
+      out[byLevel.key] = val;
+    }
+  }
+  return Object.keys(out).length > 0 ? out : void 0;
+}
+function familyKwargsDeclaredUnsupported(quirks, runtimeId) {
+  if (runtimeId !== "lmstudio") return false;
+  const block = quirks?.chatTemplateKwargs ?? quirks?.thinking?.chatTemplateKwargs;
+  return block?.lmstudio === "unsupported" || block?.byLevel?.lmstudio === "unsupported";
+}
+function resolveRequestCapability(thinking, parser, runtimeId, quirks) {
   const request = { budgetEnforcement: thinking.budgetEnforcement };
+  const effortOnly = REASONING_EFFORT_ONLY_RUNTIMES.has(runtimeId);
+  const familyKwargs = resolveFamilyChatTemplateKwargs(quirks, thinking.effectiveLevel, thinking.configuredLevel);
+  if (familyKwargs && effortOnly) {
+    request.undeliverableChatTemplateKwargs = {
+      keys: Object.keys(familyKwargs),
+      declaredUnsupported: familyKwargsDeclaredUnsupported(quirks, runtimeId)
+    };
+  } else if (familyKwargs) {
+    request.chatTemplateKwargs = { ...familyKwargs };
+  }
   if (thinking.mechanism === "effort-levels" && thinking.effort) {
     request.reasoningEffort = thinking.effort;
   }
@@ -1299,15 +1245,16 @@ function resolveRequestCapability(thinking, parser, runtimeId) {
     request.budgetTokens = thinking.budgetTokens;
   }
   if (thinking.mechanism === "on-off" && thinking.chatTemplateKwargs) {
-    request.chatTemplateKwargs = { ...thinking.chatTemplateKwargs };
-    if (REASONING_EFFORT_ON_OFF_RUNTIMES.has(runtimeId)) {
-      request.reasoningEffort = onOffReasoningEffort(thinking.thinkingActive);
-    }
+    request.chatTemplateKwargs = { ...request.chatTemplateKwargs ?? {}, ...thinking.chatTemplateKwargs };
   }
   if (parser === "harmony" && thinking.effort) {
     request.reasoningEffort = thinking.effort;
     request.chatTemplateKwargs = { ...request.chatTemplateKwargs ?? {}, reasoning_effort: thinking.effort };
   }
+  if (effortOnly && thinking.mechanism !== "none" && thinking.mechanism !== "always-on") {
+    if (!thinking.thinkingActive) request.reasoningEffort = onOffReasoningEffort(false);
+    else if (thinking.mechanism === "on-off") request.reasoningEffort = onOffReasoningEffort(true);
+  }
   return request;
 }
 function resolveModelRuntimeCapabilities(input) {
@@ -1323,7 +1270,12 @@ function resolveModelRuntimeCapabilities(input) {
     family,
     capabilities: input.capabilities,
     thinking,
-    request: resolveRequestCapability(thinking, parser, input.runtimeId),
+    request: resolveRequestCapability(
+      thinking,
+      parser,
+      input.runtimeId === "litellm" ? input.capabilities.thinkingControlRuntime ?? input.runtimeId : input.runtimeId,
+      quirks
+    ),
     response: {
       parser,
       stripTokenizerSentinels: true
@@ -1389,6 +1341,7 @@ function thinkingFormatFromModelApi(api) {
 }
 function capabilitiesFromModel(model) {
   const format = model.compat?.thinkingFormat ?? thinkingFormatFromModelApi(model.api);
+  const controlRuntime = (model.clioCoder ?? model.clio)?.thinkingControlRuntime;
   const caps = {
     chat: true,
     tools: true,
@@ -1399,7 +1352,8 @@ function capabilitiesFromModel(model) {
     rerank: false,
     fim: false,
     contextWindow: model.contextWindow,
-    maxTokens: model.maxTokens
+    maxTokens: model.maxTokens,
+    ...controlRuntime ? { thinkingControlRuntime: controlRuntime } : {}
   };
   if (format === "qwen-chat-template" || format === "openrouter" || format === "zai" || format === "anthropic-extended" || format === "deepseek-r1" || format === "openai-codex" || format === "harmony") {
     caps.thinkingFormat = format;
@@ -1471,6 +1425,21 @@ function normalizedModels(models) {
   }
   return out;
 }
+function normalizedLabels(models, labels) {
+  if (typeof labels !== "object" || labels === null || Array.isArray(labels)) return void 0;
+  const source = labels;
+  const out = {};
+  for (const id of models) {
+    const label = source[id];
+    if (typeof label !== "string") continue;
+    const trimmed = label.trim();
+    if (trimmed.length === 0 || trimmed.length > TARGET_MODEL_CACHE_MAX_MODEL_ID_CHARS || /[\r\n\0]/u.test(trimmed)) {
+      continue;
+    }
+    out[id] = trimmed;
+  }
+  return Object.keys(out).length > 0 ? out : void 0;
+}
 function parseSnapshot(value) {
   if (typeof value !== "object" || value === null || Array.isArray(value)) return null;
   const row = value;
@@ -1482,7 +1451,7 @@ function parseSnapshot(value) {
   if (typeof row.observedAt !== "string" || !Number.isFinite(Date.parse(row.observedAt))) return null;
   const models = normalizedModels(row.models);
   if (models.length !== row.models.length) return null;
-  return {
+  const snapshot = {
     version: TARGET_MODEL_CACHE_VERSION,
     targetId: row.targetId,
     runtimeId: row.runtimeId,
@@ -1490,6 +1459,9 @@ function parseSnapshot(value) {
     models,
     observedAt: row.observedAt
   };
+  const modelLabels = normalizedLabels(models, row.modelLabels);
+  if (modelLabels !== void 0) snapshot.modelLabels = modelLabels;
+  return snapshot;
 }
 function readTargetModelSnapshot(target, options = {}) {
   const path = targetModelSnapshotPath(target.id, options.cacheDir);
@@ -1518,6 +1490,8 @@ function recordTargetModelSnapshot(target, models, options = {}) {
     models: normalizedModels(models),
     observedAt: new Date(options.nowMs ?? Date.now()).toISOString()
   };
+  const modelLabels = normalizedLabels(snapshot.models, options.modelLabels);
+  if (modelLabels !== void 0) snapshot.modelLabels = modelLabels;
   try {
     safeResourceWrite(targetModelSnapshotPath(target.id, options.cacheDir), `${JSON.stringify(snapshot, null, 2)}
 `, {
@@ -1591,6 +1565,7 @@ function modelCandidatesForStatus(status) {
     const state = status.discoveredModelStates?.[trimmed];
     out.push({
       id: trimmed,
+      ...status.discoveredModelLabels?.[trimmed] ? { label: status.discoveredModelLabels[trimmed] } : {},
       source,
       ...state ? { loadState: state.state } : {},
       ...state?.detail ? { loadStateDetail: state.detail } : {}
@@ -1821,6 +1796,20 @@ function appendThinkingDiagnostics(diagnostics, resolved, requested) {
   const severity = thinking.noticeKind === "unsupported" || thinking.noticeKind === "always-on" || thinking.noticeKind === "ignored-on-off" ? "warning" : "info";
   diagnostics.push(diagnostic(severity, `thinking-${thinking.noticeKind}`, thinking.notice));
 }
+var CHAT_TEMPLATE_KWARGS_UNDELIVERABLE = "chat-template-kwargs-undeliverable";
+function appendChatTemplateKwargsDiagnostics(diagnostics, resolved) {
+  const undeliverable = resolved.request.undeliverableChatTemplateKwargs;
+  if (!undeliverable || undeliverable.keys.length === 0) return;
+  const keys = undeliverable.keys.join(", ");
+  const declared = undeliverable.declaredUnsupported ? "its family entry marks them unsupported there" : "its family entry does not say so";
+  diagnostics.push(
+    diagnostic(
+      "warning",
+      CHAT_TEMPLATE_KWARGS_UNDELIVERABLE,
+      `chat-template kwargs ${keys} cannot reach ${resolved.runtimeId}, which ignores chat_template_kwargs; ${resolved.modelId} runs without them and ${declared}`
+    )
+  );
+}
 function nonNegativeFiniteNumber(value) {
   return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
 }
@@ -1948,6 +1937,7 @@ function resolveRuntimeTarget(providers, input) {
   const decisions = capabilityDecisions(runtime, capabilities);
   appendCapabilityDiagnostics(diagnostics, input, capabilities, decisions, targetId);
   appendThinkingDiagnostics(diagnostics, modelRuntime, requestedThinkingLevel);
+  appendChatTemplateKwargsDiagnostics(diagnostics, modelRuntime);
   if (hasError(diagnostics)) return { ok: false, diagnostics };
   const resolved = {
     targetId: target.id,
@@ -1987,6 +1977,7 @@ function modelHintPatch(target, model) {
 function withoutStaleRuntimeDiagnostics(diagnostics, decisions) {
   return diagnostics.filter((entry) => {
     if (entry.code.startsWith("thinking-")) return false;
+    if (entry.code === CHAT_TEMPLATE_KWARGS_UNDELIVERABLE) return false;
     if (entry.code === "output-budget-unknown" && decisions.maxTokens > 0) return false;
     if (entry.code === "tools-unsupported" && decisions.tools) return false;
     return true;
@@ -2025,6 +2016,7 @@ function refineRuntimeTargetWithModelHints(target, model, knowledgeBase) {
   const decisions = capabilityDecisions(target.runtime, capabilities);
   const diagnostics = withoutStaleRuntimeDiagnostics(target.diagnostics, decisions);
   appendThinkingDiagnostics(diagnostics, modelRuntime, target.requestedThinkingLevel);
+  appendChatTemplateKwargsDiagnostics(diagnostics, modelRuntime);
   return {
     ...target,
     capabilities,
@@ -2152,6 +2144,15 @@ init_esm_shims();
 function normalizeCostProvenance(value) {
   return value ?? "unknown";
 }
+var COST_PROVENANCE_VALUES = /* @__PURE__ */ new Set([
+  "known",
+  "known_free",
+  "estimated",
+  "unknown"
+]);
+function resolveCostProvenance2(value, fallback) {
+  return typeof value === "string" && COST_PROVENANCE_VALUES.has(value) ? value : fallback;
+}
 
 // src/domains/providers/index.ts
 init_esm_shims();
@@ -2185,6 +2186,9 @@ function residencyTargetKey(runtimeId, baseUrl) {
 // src/engine/gemma-channel-filter.ts
 init_esm_shims();
 import { createAssistantMessageEventStream } from "@earendil-works/pi-ai";
+function usesGemmaChannelMarkers(modelId) {
+  return inferLocalModelFamily(modelId) === "gemma-4";
+}
 var THOUGHT_START = /<\|channel>[^\n]*\n/;
 var BARE_THOUGHT_START = /^\s*(?:thought|own[- ]?(?:thought|think))\s*\n/i;
 var BARE_THOUGHT_ONLY = /^\s*(?:thought|own[- ]?(?:thought|think))\s*$/i;
@@ -2384,7 +2388,6 @@ var globalDefaultMaxOutputTokens = 0;
 function setGlobalDefaultMaxOutputTokens(value) {
   globalDefaultMaxOutputTokens = Number.isFinite(value) && value > 0 ? Math.floor(value) : 0;
 }
-var LOCAL_TOOL_TURN_MAX_OUTPUT_TOKENS = CLIO_MIN_MAX_OUTPUT_TOKENS;
 function resolveReservedOutputTokens(maxOutputTokens) {
   const limit = typeof maxOutputTokens === "number" && Number.isFinite(maxOutputTokens) && maxOutputTokens > 0 ? maxOutputTokens : DEFAULT_MAX_OUTPUT_TOKENS;
   return Math.min(limit, DEFAULT_MAX_OUTPUT_TOKENS);
@@ -2403,7 +2406,7 @@ function remainingContextMaxTokens(model, context, options, limits) {
   const contextWindow = Math.min(configuredContextWindow, loadedContextWindow);
   const budget = Number.isFinite(contextWindow) ? Math.max(1, contextWindow - inputTokens - safety) : Number.POSITIVE_INFINITY;
   const modelLimit = model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY;
-  const defaultLimit = limits?.maxOutputTokens !== void 0 && limits.maxOutputTokens > 0 ? limits.maxOutputTokens : globalDefaultMaxOutputTokens > 0 ? globalDefaultMaxOutputTokens : modelLimit;
+  const defaultLimit = limits?.maxOutputTokens !== void 0 && limits.maxOutputTokens > 0 ? limits.maxOutputTokens : globalDefaultMaxOutputTokens > 0 ? globalDefaultMaxOutputTokens : model.maxTokens > 0 ? modelLimit : DEFAULT_MAX_OUTPUT_TOKENS;
   const requested = options?.maxTokens ?? defaultLimit;
   const resolved = Math.min(requested, modelLimit, budget);
   return Number.isFinite(resolved) ? resolved : DEFAULT_MAX_OUTPUT_TOKENS;
@@ -2441,15 +2444,15 @@ function mergeSamplingOverride(profile, override = runSamplingOverrides()) {
 // src/engine/apis/ollama-native.ts
 var REASONING_CHARS_PER_TOKEN = 4;
 function clioQuirks(model) {
-  return model.clio?.quirks;
+  return model.clioCoder?.quirks;
 }
 function ollamaTargetId(model) {
-  return model.clio?.targetId ?? model.provider;
+  return model.clioCoder?.targetId ?? model.provider;
 }
 async function reconcileOllamaResidency(model, headers) {
   const baseUrl = model.baseUrl;
   if (!baseUrl) return;
-  const metadata2 = model.clio;
+  const metadata2 = model.clioCoder;
   const adapter = {
     targetKey: residencyTargetKey("ollama-native", baseUrl),
     targetId: ollamaTargetId(model),
@@ -2681,7 +2684,7 @@ function runStream(model, context, options, thinkingLevel) {
       let doneReason;
       const sentinelStripper = createSentinelStripper();
       const resolved = resolveModelRuntimeCapabilitiesForModel(model, thinkingLevel);
-      const channelFilter = resolved.family === "gemma-4" ? createGemmaChannelFilter() : null;
+      const channelFilter = usesGemmaChannelMarkers(resolved.modelId) ? createGemmaChannelFilter() : null;
       const closeActiveThinking = () => {
         if (!activeThinking) return;
         stream.push({
@@ -2828,6 +2831,115 @@ import {
 } from "@earendil-works/pi-ai";
 import { openAICompletionsApi } from "@earendil-works/pi-ai/api/openai-completions.lazy";
 
+// src/core/gateway-routing.ts
+init_esm_shims();
+function liteLLMRouteFailureMessage(errorMessage, targetId, modelId) {
+  const providerMessage = errorMessage?.trim() || "LiteLLM request failed.";
+  const advice = `LiteLLM route '${modelId}' on target '${targetId}' failed. Clio did not retry or substitute another model; select a different route with /model, then resend.`;
+  return providerMessage.includes(advice) ? providerMessage : `${providerMessage}
+
+${advice}`;
+}
+function nonEmptyString(value) {
+  return typeof value === "string" && value.trim().length > 0 ? value.trim() : void 0;
+}
+function nonnegativeNumber(value) {
+  if (typeof value === "number" && Number.isFinite(value) && value >= 0) return value;
+  if (typeof value !== "string" || value.trim().length === 0) return void 0;
+  const parsed = Number(value);
+  return Number.isFinite(parsed) && parsed >= 0 ? parsed : void 0;
+}
+function nonnegativeInteger(value) {
+  const parsed = nonnegativeNumber(value);
+  return parsed === void 0 ? void 0 : Math.floor(parsed);
+}
+function apiBaseHost(value) {
+  const raw = nonEmptyString(value);
+  if (raw === void 0) return void 0;
+  try {
+    return new URL(raw).host || void 0;
+  } catch {
+    return void 0;
+  }
+}
+function storedApiBaseHost(value) {
+  const raw = nonEmptyString(value);
+  if (raw === void 0) return void 0;
+  try {
+    return new URL(raw.includes("://") ? raw : `http://${raw}`).host || void 0;
+  } catch {
+    return void 0;
+  }
+}
+function liteLLMGatewayRoutingFromHeaders(headers) {
+  const callId = nonEmptyString(headers.get("x-litellm-call-id"));
+  const modelGroup = nonEmptyString(headers.get("x-litellm-model-group"));
+  const modelName = nonEmptyString(headers.get("x-litellm-model-name"));
+  const host = apiBaseHost(headers.get("x-litellm-model-api-base"));
+  const deploymentId = nonEmptyString(headers.get("x-litellm-model-id"));
+  const attemptedFallbacks = nonnegativeInteger(headers.get("x-litellm-attempted-fallbacks"));
+  const attemptedRetries = nonnegativeInteger(headers.get("x-litellm-attempted-retries"));
+  const overheadMs = nonnegativeNumber(headers.get("x-litellm-overhead-duration-ms"));
+  const responseDurationMs = nonnegativeNumber(headers.get("x-litellm-response-duration-ms"));
+  const version = nonEmptyString(headers.get("x-litellm-version"));
+  const observation = {
+    gateway: "litellm",
+    ...callId !== void 0 ? { callId } : {},
+    ...modelGroup !== void 0 ? { modelGroup } : {},
+    ...modelName !== void 0 ? { modelName } : {},
+    ...host !== void 0 ? { apiBaseHost: host } : {},
+    ...deploymentId !== void 0 ? { deploymentId } : {},
+    ...attemptedFallbacks !== void 0 ? { attemptedFallbacks } : {},
+    ...attemptedRetries !== void 0 ? { attemptedRetries } : {},
+    ...overheadMs !== void 0 ? { overheadMs } : {},
+    ...responseDurationMs !== void 0 ? { responseDurationMs } : {},
+    ...version !== void 0 ? { version } : {}
+  };
+  return Object.keys(observation).length > 1 ? observation : null;
+}
+function gatewayRoutingObservationFromRecord(record) {
+  const value = record.gatewayRouting;
+  if (value === null || typeof value !== "object" || Array.isArray(value)) return null;
+  const raw = value;
+  if (raw.gateway !== "litellm") return null;
+  const callId = nonEmptyString(raw.callId);
+  const modelGroup = nonEmptyString(raw.modelGroup);
+  const modelName = nonEmptyString(raw.modelName);
+  const host = storedApiBaseHost(raw.apiBaseHost);
+  const deploymentId = nonEmptyString(raw.deploymentId);
+  const attemptedFallbacks = nonnegativeInteger(raw.attemptedFallbacks);
+  const attemptedRetries = nonnegativeInteger(raw.attemptedRetries);
+  const overheadMs = nonnegativeNumber(raw.overheadMs);
+  const responseDurationMs = nonnegativeNumber(raw.responseDurationMs);
+  const version = nonEmptyString(raw.version);
+  const observation = {
+    gateway: "litellm",
+    ...callId !== void 0 ? { callId } : {},
+    ...modelGroup !== void 0 ? { modelGroup } : {},
+    ...modelName !== void 0 ? { modelName } : {},
+    ...host !== void 0 ? { apiBaseHost: host } : {},
+    ...deploymentId !== void 0 ? { deploymentId } : {},
+    ...attemptedFallbacks !== void 0 ? { attemptedFallbacks } : {},
+    ...attemptedRetries !== void 0 ? { attemptedRetries } : {},
+    ...overheadMs !== void 0 ? { overheadMs } : {},
+    ...responseDurationMs !== void 0 ? { responseDurationMs } : {},
+    ...version !== void 0 ? { version } : {}
+  };
+  return Object.keys(observation).length > 1 ? observation : null;
+}
+function gatewayRoutingObservationLabel(observation) {
+  const route = [observation.modelGroup, observation.modelName].filter(Boolean).join("->");
+  const destination = [route, observation.apiBaseHost ? `@${observation.apiBaseHost}` : ""].filter(Boolean).join("");
+  const parts = [
+    "gateway=litellm",
+    ...destination ? [`route=${destination}`] : [],
+    ...observation.attemptedFallbacks !== void 0 ? [`fallbacks=${observation.attemptedFallbacks}`] : [],
+    ...observation.attemptedRetries !== void 0 ? [`retries=${observation.attemptedRetries}`] : [],
+    ...observation.overheadMs !== void 0 ? [`overhead_ms=${observation.overheadMs}`] : []
+  ];
+  return parts.join(" ");
+}
+
 // src/engine/harmony-response.ts
 init_esm_shims();
 var START = "<|start|>";
@@ -3113,7 +3225,7 @@ async function ensureLlamaCppResidency(input) {
     targetId: input.targetId,
     runtimeId: input.runtimeId,
     keepModelId: input.keepModelId,
-    managed: input.managed ?? residencyManaged(),
+    managed: input.managed ?? true,
     strategy: "router",
     listResident: async () => {
       const models = await fetchRouterModels(input, fetchImpl);
@@ -3193,7 +3305,7 @@ function responseInstanceId(data) {
   return typeof id === "string" && id.trim().length > 0 ? id.trim() : void 0;
 }
 function metadata(model) {
-  return model.clio;
+  return model.clioCoder;
 }
 function targetForModel(model) {
   const info = metadata(model);
@@ -3390,10 +3502,10 @@ async function ensureLmStudioResidency(model, options = {}) {
 var piOpenAICompletions = openAICompletionsApi();
 var REASONING_CHARS_PER_TOKEN2 = 4;
 function chatTemplateKwargsUnsupported(model) {
-  return model.clio?.chatTemplateKwargsUnsupported === true;
+  return model.clioCoder?.chatTemplateKwargsUnsupported === true;
 }
 function clioQuirks2(model) {
-  return model.clio?.quirks;
+  return model.clioCoder?.quirks;
 }
 function pickSamplingProfile2(quirks, thinkingActive) {
   const sampling = quirks?.sampling;
@@ -3424,7 +3536,7 @@ function backendCompletionTimings(value, source) {
   };
 }
 function backendTimingsSourceForModel(model) {
-  const metadata2 = model.clio;
+  const metadata2 = model.clioCoder;
   if (model.provider === "llamacpp" && metadata2?.runtimeId === "llamacpp") return "llamacpp-timings";
   if (model.provider === "lmstudio" && metadata2?.runtimeId === "lmstudio") return "lmstudio-timings";
   return null;
@@ -3475,7 +3587,10 @@ function observeResponseModelIdBytes(chunk, capture, flush = false) {
   }
   if (captureCanStopEarly(capture) || flush) capture.decoder = null;
 }
-function captureResponseModelId(response, capture) {
+function captureResponseModelId(response, capture, model) {
+  if (runtimeMetadata(model)?.runtimeId === "litellm") {
+    capture.gatewayRouting = liteLLMGatewayRoutingFromHeaders(response.headers);
+  }
   if (!response.body || !response.headers.get("content-type")?.toLowerCase().includes("text/event-stream")) {
     return response;
   }
@@ -3505,13 +3620,14 @@ function withResponseModelIdCapture(model, options, sourceFactory) {
     modelIdDone: false,
     backendTimings: null,
     backendTimingsSource: backendTimingsSourceForModel(model),
+    gatewayRouting: null,
     buffer: "",
     decoder: new TextDecoder()
   };
   const fetchImpl = options.fetch ?? ((input, init) => globalThis.fetch(input, init));
   const capturedOptions = {
     ...options,
-    fetch: async (input, init) => captureResponseModelId(await fetchImpl(input, init), capture)
+    fetch: async (input, init) => captureResponseModelId(await fetchImpl(input, init), capture, model)
   };
   const source = sourceFactory(capturedOptions);
   const annotated = createAssistantMessageEventStream3();
@@ -3521,9 +3637,11 @@ function withResponseModelIdCapture(model, options, sourceFactory) {
         const observation = capture.observed ? capture.reportedModelId === null ? { state: "not-reported" } : { state: "reported", reportedModelId: capture.reportedModelId } : { state: "not-observed" };
         if (event.type === "done") {
           event.message.responseModelIdObservation = observation;
+          if (capture.gatewayRouting !== null) event.message.gatewayRouting = capture.gatewayRouting;
           if (capture.backendTimings !== null) event.message.backendTimings = capture.backendTimings;
         } else if (event.type === "error") {
           event.error.responseModelIdObservation = observation;
+          if (capture.gatewayRouting !== null) event.error.gatewayRouting = capture.gatewayRouting;
           if (capture.backendTimings !== null) event.error.backendTimings = capture.backendTimings;
         }
         annotated.push(event);
@@ -3536,6 +3654,26 @@ function withResponseModelIdCapture(model, options, sourceFactory) {
   })();
   return annotated;
 }
+function withLiteLLMRouteFailureAdvice(model, source) {
+  const metadata2 = runtimeMetadata(model);
+  if (metadata2?.runtimeId !== "litellm") return source;
+  const advised = createAssistantMessageEventStream3();
+  (async () => {
+    try {
+      for await (const event of source) {
+        if (event.type === "error") {
+          event.error.errorMessage = liteLLMRouteFailureMessage(event.error.errorMessage, metadata2.targetId, model.id);
+        }
+        advised.push(event);
+      }
+      advised.end();
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      throw new Error(liteLLMRouteFailureMessage(message, metadata2.targetId, model.id));
+    }
+  })();
+  return advised;
+}
 function samplingParamsFromProfile(profile) {
   return {
     ...profile.topP !== void 0 ? { top_p: profile.topP } : {},
@@ -3602,12 +3740,31 @@ function withStrippedPartial(event) {
   return { ...event, partial: stripThinkingFromMessage(event.partial) };
 }
 function applyThinkingPayload(payload, applied, resolved, model) {
-  if (applied.mechanism === "none") return stripThinkingRequestFields(payload);
-  if (applied.mechanism === "always-on") return payload;
+  if (applied.mechanism === "none") {
+    const next2 = stripThinkingRequestFields(payload);
+    if (resolved.request.chatTemplateKwargs && !chatTemplateKwargsUnsupported(model)) {
+      const existing = isPlainRecord(next2.chat_template_kwargs) ? next2.chat_template_kwargs : {};
+      next2.chat_template_kwargs = { ...existing, ...resolved.request.chatTemplateKwargs };
+    }
+    return next2;
+  }
+  if (applied.mechanism === "always-on") {
+    if (resolved.request.chatTemplateKwargs && !chatTemplateKwargsUnsupported(model)) {
+      const next2 = { ...payload };
+      const existing = isPlainRecord(next2.chat_template_kwargs) ? next2.chat_template_kwargs : {};
+      next2.chat_template_kwargs = { ...existing, ...resolved.request.chatTemplateKwargs };
+      return next2;
+    }
+    return payload;
+  }
   const next = { ...payload };
   if (resolved.request.reasoningEffort && (next.reasoning_effort === void 0 || resolved.response.parser === "harmony")) {
     next.reasoning_effort = resolved.request.reasoningEffort;
   }
+  if (resolved.runtimeId === "litellm" && resolved.request.reasoningEffort && next.reasoning_effort === resolved.request.reasoningEffort) {
+    const allowed = Array.isArray(next.allowed_openai_params) ? next.allowed_openai_params : [];
+    next.allowed_openai_params = [.../* @__PURE__ */ new Set([...allowed, "reasoning_effort"])];
+  }
   if (resolved.request.chatTemplateKwargs && !chatTemplateKwargsUnsupported(model)) {
     const existing = isPlainRecord(next.chat_template_kwargs) ? next.chat_template_kwargs : {};
     next.chat_template_kwargs = { ...existing, ...resolved.request.chatTemplateKwargs };
@@ -3629,13 +3786,13 @@ function applyLmStudioPayload(payload, model, resolved) {
   if (request?.ttlSeconds !== void 0) next.ttl = request.ttlSeconds;
   if (request?.draftModel !== void 0) next.draft_model = request.draftModel;
   if (resolved.thinking.mechanism === "none" || resolved.thinking.mechanism === "always-on") return next;
-  const resolvedEffort = resolved.request.reasoningEffort ?? lmStudioReasoningEffort(resolved.thinking.effectiveLevel, runtimeMetadata(model)?.lmstudioReasoningOptions);
+  const resolvedEffort = runtimeMetadata(model)?.lmstudioReasoningOptions ? lmStudioReasoningEffort(resolved.thinking.effectiveLevel, runtimeMetadata(model)?.lmstudioReasoningOptions) : resolved.request.reasoningEffort ?? lmStudioReasoningEffort(resolved.thinking.effectiveLevel);
   switch (request?.reasoning) {
     case "off":
-      next.reasoning_effort = "none";
+      next.reasoning_effort = lmStudioReasoningEffort("off", runtimeMetadata(model)?.lmstudioReasoningOptions);
       break;
     case "on":
-      next.reasoning_effort = "low";
+      next.reasoning_effort = lmStudioReasoningEffort("low", runtimeMetadata(model)?.lmstudioReasoningOptions);
       break;
     case "low":
     case "medium":
@@ -3734,16 +3891,39 @@ function thinkingLevelFromSimple2(options) {
   return reasoning;
 }
 function withRemainingContextBudget(model, context, options) {
-  const metadata2 = runtimeMetadata(model);
-  const localToolOutputLimit = options?.maxTokens === void 0 && (context.tools?.length ?? 0) > 0 && (model.provider === "llamacpp" || metadata2?.runtimeId === "llamacpp") ? LOCAL_TOOL_TURN_MAX_OUTPUT_TOKENS : void 0;
   return {
     ...options ?? {},
-    maxTokens: remainingContextMaxTokens(
-      model,
-      context,
-      options,
-      localToolOutputLimit === void 0 ? void 0 : { maxOutputTokens: localToolOutputLimit }
-    )
+    maxTokens: remainingContextMaxTokens(model, context, options)
+  };
+}
+function hasHeader(headers, name) {
+  if (!headers) return false;
+  const wanted = name.toLowerCase();
+  return Object.keys(headers).some((key) => key.toLowerCase() === wanted);
+}
+function withLiteLLMRequestOptions(model, options) {
+  const metadata2 = runtimeMetadata(model);
+  if (metadata2?.runtimeId !== "litellm") return options;
+  const request = metadata2.litellm?.request;
+  const modelHeaders = model.headers;
+  const optionHeaders = options.headers;
+  const headers = { ...options.headers ?? {} };
+  const setDefault = (name, value) => {
+    if (value === void 0 || hasHeader(modelHeaders, name) || hasHeader(optionHeaders, name)) return;
+    headers[name] = value;
+  };
+  setDefault("x-litellm-tags", ["clio-coder", ...request?.tags ?? []].join(","));
+  if (request?.sendSessionId !== false) setDefault("x-litellm-session-id", options.sessionId);
+  setDefault("x-litellm-timeout", request?.timeoutSeconds?.toString());
+  setDefault("x-litellm-stream-timeout", request?.streamTimeoutSeconds?.toString());
+  setDefault("x-litellm-num-retries", request?.numRetries?.toString());
+  return {
+    ...options,
+    headers,
+    // Never hide duplicate attempts inside the OpenAI SDK beneath Clio's
+    // operator-visible failure. The gateway may still be configured to retry,
+    // but physical-routing deployments should keep that policy at zero too.
+    maxRetries: options.maxRetries ?? 0
   };
 }
 function requiredToolArguments(tool) {
@@ -3757,7 +3937,7 @@ function hasEmptyArguments(args) {
   return Object.keys(args).length === 0;
 }
 function runtimeMetadata(model) {
-  return model.clio;
+  return model.clioCoder;
 }
 function emptyErrorMessage(model, message) {
   return {
@@ -4014,24 +4194,31 @@ var openAICompletionsApiProvider = {
     const resolved = resolvedCapabilitiesForModel(model, defaultLevel);
     const effectiveContext = stripsThinking(resolved) ? stripThinkingFromContext(context) : context;
     const withSamplers = withSamplingOverrides(model, options, resolved);
+    const withGatewayOptions = withLiteLLMRequestOptions(
+      model,
+      withRemainingContextBudget(model, effectiveContext, withSamplers)
+    );
     return guardMalformedToolCalls(
       withReasoningTokenEstimate(
         stripSentinelsFromStream(
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
-              withResponseModelIdCapture(
+              withLiteLLMRouteFailureAdvice(
                 model,
-                withRemainingContextBudget(model, effectiveContext, withSamplers),
-                (capturedOptions) => withLocalResidency(
+                withResponseModelIdCapture(
                   model,
-                  {
-                    ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
-                    ...options?.signal !== void 0 ? { signal: options.signal } : {}
-                  },
-                  (requestModel) => piOpenAICompletions.stream(requestModel, effectiveContext, capturedOptions)
+                  withGatewayOptions,
+                  (capturedOptions) => withLocalResidency(
+                    model,
+                    {
+                      ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
+                      ...options?.signal !== void 0 ? { signal: options.signal } : {}
+                    },
+                    (requestModel) => piOpenAICompletions.stream(requestModel, effectiveContext, capturedOptions)
+                  )
                 )
               ),
-              resolved.family === "gemma-4"
+              usesGemmaChannelMarkers(resolved.modelId)
             ),
             resolved
           ),
@@ -4046,24 +4233,31 @@ var openAICompletionsApiProvider = {
     const resolved = resolvedCapabilitiesForModel(model, thinkingLevelFromSimple2(options));
     const effectiveContext = stripsThinking(resolved) ? stripThinkingFromContext(context) : context;
     const withSamplers = withSamplingOverrides(model, options, resolved);
+    const withGatewayOptions = withLiteLLMRequestOptions(
+      model,
+      withRemainingContextBudget(model, effectiveContext, withSamplers)
+    );
     return guardMalformedToolCalls(
       withReasoningTokenEstimate(
         stripSentinelsFromStream(
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
-              withResponseModelIdCapture(
+              withLiteLLMRouteFailureAdvice(
                 model,
-                withRemainingContextBudget(model, effectiveContext, withSamplers),
-                (capturedOptions) => withLocalResidency(
+                withResponseModelIdCapture(
                   model,
-                  {
-                    ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
-                    ...options?.signal !== void 0 ? { signal: options.signal } : {}
-                  },
-                  (requestModel) => piOpenAICompletions.streamSimple(requestModel, effectiveContext, capturedOptions)
+                  withGatewayOptions,
+                  (capturedOptions) => withLocalResidency(
+                    model,
+                    {
+                      ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
+                      ...options?.signal !== void 0 ? { signal: options.signal } : {}
+                    },
+                    (requestModel) => piOpenAICompletions.streamSimple(requestModel, effectiveContext, capturedOptions)
+                  )
                 )
               ),
-              resolved.family === "gemma-4"
+              usesGemmaChannelMarkers(resolved.modelId)
             ),
             resolved
           ),
@@ -4400,7 +4594,7 @@ function discoveredModelsSource(probe, preservePreviousProbe, previous, cachedMo
 function sameProbeIdentity(previous, next) {
   return previous.id === next.id && previous.runtime === next.runtime && (previous.url ?? "") === (next.url ?? "") && (previous.defaultModel ?? "") === (next.defaultModel ?? "");
 }
-function mergeProbeResult(desc, target, probe, previous, cachedModels = []) {
+function mergeProbeResult(desc, target, probe, previous, cachedModels = [], cachedModelLabels = {}) {
   const probeSucceeded = probe?.ok ?? false;
   const preservePrevious = !probeSucceeded && previous !== void 0 && sameProbeIdentity(previous.target, target);
   const probeCapabilities = probe?.discoveredCapabilities ?? (preservePrevious ? previous.probeCapabilities : null) ?? null;
@@ -4412,8 +4606,10 @@ function mergeProbeResult(desc, target, probe, previous, cachedModels = []) {
     probe?.models ?? (preservePrevious ? previous.discoveredModels : void 0) ?? (cachedModels.length > 0 ? cachedModels : void 0) ?? desc.knownModels ?? []
   );
   const discoveredModelStates = probe?.modelStates ?? (preservePrevious ? previous.discoveredModelStates : null) ?? null;
+  const discoveredModelLabels = probe?.modelLabels ?? (preservePrevious ? previous.discoveredModelLabels : void 0) ?? (cachedModels.length > 0 ? cachedModelLabels : void 0) ?? {};
   const merge = {
     discoveredModels,
+    discoveredModelLabels,
     discoveredModelsSource: discoveredModelsSource(probe, preservePrevious, previous, cachedModels, desc),
     discoveredModelStates,
     probeCapabilities,
@@ -4428,7 +4624,9 @@ function unservedDefaultModelReason(desc, target, merge) {
   const model = target.defaultModel;
   if (!model) return null;
   if (merge.discoveredModelsSource !== "probe" || merge.discoveredModels.length === 0) return null;
-  if (listKnownModelsForRuntime(desc.id).length > 0) return null;
+  if (listKnownModelsForRuntime(desc.id).length > 0 && desc.externalAgentLoop?.modelCatalog !== "live-authoritative") {
+    return null;
+  }
   if (merge.discoveredModels.includes(model)) return null;
   if (merge.discoveredModelStates && model in merge.discoveredModelStates) return null;
   return `default model '${model}' is not advertised by the target`;
@@ -4448,14 +4646,18 @@ function createProvidersBundle(context) {
     if (config) return config.get();
     return readSettings();
   }
-  async function buildProbeContextForTarget(target, desc) {
+  async function buildProbeContextForTarget(target, desc, signal) {
     const probeCtx = {
       credentialsPresent: credentialsPresent(),
-      httpTimeoutMs: DEFAULT_PROBE_TIMEOUT_MS
+      httpTimeoutMs: DEFAULT_PROBE_TIMEOUT_MS,
+      ...signal ? { signal } : {}
     };
     if (!targetRequiresAuth(target, desc)) return probeCtx;
     try {
-      const resolution = await authStore.resolveForTarget(resolveAuthTarget(target, desc), { includeFallback: false });
+      const resolution = await authStore.resolveForTarget(resolveAuthTarget(target, desc), {
+        includeFallback: false,
+        ...signal ? { signal } : {}
+      });
       if (resolution.apiKey) probeCtx.authToken = resolution.apiKey;
     } catch {
     }
@@ -4507,8 +4709,9 @@ function createProvidersBundle(context) {
       return out2;
     }
     const availability = availabilityFor(desc, target, authStatusFor);
-    const cachedModels = readTargetModelSnapshot(target)?.models ?? [];
-    const merge = mergeProbeResult(desc, target, probe, previous, cachedModels);
+    const cachedSnapshot = readTargetModelSnapshot(target);
+    const cachedModels = cachedSnapshot?.models ?? [];
+    const merge = mergeProbeResult(desc, target, probe, previous, cachedModels, cachedSnapshot?.modelLabels ?? {});
     const { capabilities, contextWindowProvenance } = capabilitiesFor(desc, target, merge, kb);
     const healthy = probe !== null ? probe.ok : null;
     const unservedDefault = probe?.ok ? unservedDefaultModelReason(desc, target, merge) : null;
@@ -4532,6 +4735,7 @@ function createProvidersBundle(context) {
       probeModelCapabilities: merge.probeModelCapabilities,
       probeModelId: merge.probeModelId,
       discoveredModels: merge.discoveredModels,
+      discoveredModelLabels: merge.discoveredModelLabels,
       discoveredModelsSource: merge.discoveredModelsSource,
       discoveredModelStates: merge.discoveredModelStates
     };
@@ -4540,6 +4744,7 @@ function createProvidersBundle(context) {
     return out;
   }
   async function probeTargetInternal(target, live, options) {
+    options?.signal?.throwIfAborted();
     const previous = statuses.get(target.id);
     const desc = registry.get(target.runtime);
     if (!desc) {
@@ -4553,13 +4758,15 @@ function createProvidersBundle(context) {
       statuses.set(target.id, status2);
       return status2;
     }
-    const probeCtx = await buildProbeContextForTarget(target, desc);
+    const probeCtx = await buildProbeContextForTarget(target, desc, options?.signal);
+    options?.signal?.throwIfAborted();
     let probeResult;
     try {
       probeResult = await desc.probe(target, probeCtx);
     } catch (err) {
       probeResult = { ok: false, error: err instanceof Error ? err.message : String(err) };
     }
+    options?.signal?.throwIfAborted();
     if (probeResult.ok && typeof desc.probeModels === "function" && !probeResult.models) {
       try {
         const ids = await desc.probeModels(target, probeCtx);
@@ -4567,6 +4774,7 @@ function createProvidersBundle(context) {
       } catch {
       }
     }
+    options?.signal?.throwIfAborted();
     if (probeResult.ok && options?.reasoning !== false && typeof desc.probeReasoning === "function") {
       const settings = readConfig();
       const orchestratorTarget = settings.chat.target === target.id ? settings.chat.model : null;
@@ -4574,6 +4782,7 @@ function createProvidersBundle(context) {
       if (candidateModelId) {
         try {
           const result = await desc.probeReasoning(target, candidateModelId, probeCtx);
+          options?.signal?.throwIfAborted();
           reasoningCache.set(reasoningCacheKey(target.id, candidateModelId), result.reasoning);
           const capabilityModelId = probeResult.capabilityModelId ?? null;
           if (capabilityModelId === null || capabilityModelId === candidateModelId) {
@@ -4594,10 +4803,15 @@ function createProvidersBundle(context) {
         }
       }
     }
+    options?.signal?.throwIfAborted();
     const status = buildStatus(target, desc, probeResult, previous);
     statuses.set(target.id, status);
     if (probeResult.ok && probeResult.models !== void 0) {
-      recordTargetModelSnapshot(target, probeResult.models);
+      recordTargetModelSnapshot(
+        target,
+        probeResult.models,
+        probeResult.modelLabels ? { modelLabels: probeResult.modelLabels } : {}
+      );
     }
     void recordEndpointSlotsFromStatus(status);
     context.bus.emit(BusChannels.ProviderHealth, { id: target.id, status });
@@ -4783,8 +4997,8 @@ export {
   setResidencyNoticeSink,
   setProtectedModelsProvider,
   withRunOverrides,
-  EMPTY_CAPABILITIES,
-  VALID_THINKING_LEVELS,
+  gatewayRoutingObservationFromRecord,
+  gatewayRoutingObservationLabel,
   applyModelCapabilityPatch,
   probeCapabilitiesForModel,
   resolveModelCapabilities,
@@ -4802,6 +5016,7 @@ export {
   isOrchestratorEligibleRuntime,
   isDispatchEligibleRuntime,
   modelResidencyForStatus,
+  hasLiveModelCatalog,
   modelCandidatesForStatus,
   modelIdsForStatus,
   canonicalizeWireModelId,
@@ -4813,6 +5028,7 @@ export {
   runtimeResolutionWarnings,
   runtimeResolutionWarningsBesideThinkingNotice,
   normalizeCostProvenance,
+  resolveCostProvenance2 as resolveCostProvenance,
   ProvidersDomainModule
 };
-//# sourceMappingURL=chunk-UEDMSP56.js.map
+//# sourceMappingURL=chunk-XPWWI35G.js.map