atlarix 15.8.2 → 15.9.0

package/dist-headless/atlarix-headless.mjs

@@ -9587,6 +9587,10 @@ var init_schema = __esm({
         name: text("name").notNull(),
         api_base_url: text("api_base_url").notNull(),
         env_var_name: text("env_var_name"),
+        /** Wire format: null / "openai" = OpenAI-compatible, "anthropic" = Messages API. */
+        api_format: text("api_format"),
+        /** How the key is sent: null = the format's usual header, or "x-api-key" / "bearer". */
+        auth_header: text("auth_header"),
         createdAt: integer("created_at", { mode: "timestamp" }).notNull().default(sql`(unixepoch())`),
         updatedAt: integer("updated_at", { mode: "timestamp" }).notNull().default(sql`(unixepoch())`)
       }
@@ -10911,6 +10915,14 @@ var init_usage_stats = __esm({
   }
 });
 
+// raw-file:/home/runner/work/Atlarix/Atlarix/drizzle/0080_custom_provider_api_format.sql
+var custom_provider_api_format_default;
+var init_custom_provider_api_format = __esm({
+  "raw-file:/home/runner/work/Atlarix/Atlarix/drizzle/0080_custom_provider_api_format.sql"() {
+    custom_provider_api_format_default = "-- The wire format a provider the user added speaks, and how it takes the key.\n--\n-- A custom provider could only be OpenAI-compatible. Many endpoints serve the\n-- Anthropic Messages API instead (Bedrock's /anthropic, gateways, proxies), and\n-- those take the key either as `x-api-key` or as `Authorization: Bearer`.\n--\n-- Additive only (two nullable columns, null = OpenAI-compatible with its usual\n-- Bearer key): an older reader ignores them, so min_reader stays where it is.\nALTER TABLE `language_model_providers` ADD `api_format` text;--> statement-breakpoint\nALTER TABLE `language_model_providers` ADD `auth_header` text;\n";
+  }
+});
+
 // src/db/migrations/bundled.generated.ts
 var BUNDLED_MIGRATIONS, DRIZZLE_HASH_TO_NAME;
 var init_bundled_generated = __esm({
@@ -10995,6 +11007,7 @@ var init_bundled_generated = __esm({
     init_automation_leases();
     init_mcp_headers();
     init_usage_stats();
+    init_custom_provider_api_format();
     BUNDLED_MIGRATIONS = [
       { name: "0000_nebulous_proemial_gods", sql: nebulous_proemial_gods_default, drizzleHash: "2fd0348e72ac1429b9246311e8db17a34078771470c622a13ab09afdee0a0d31", when: 1744692127560 },
       { name: "0001_hesitant_roland_deschain", sql: hesitant_roland_deschain_default, drizzleHash: "b9b37284db688cf3aaaa2719dda788238fd216933d3756c0482f26b5343d0059", when: 1744999922420 },
@@ -11075,7 +11088,8 @@ var init_bundled_generated = __esm({
       { name: "0076_drop_installed_skills", sql: drop_installed_skills_default, drizzleHash: "b361a2b664ce74a3efeaa2f73cdf6de582509c6569a1aab54febbef507b98b3c", when: 17877024e5 },
       { name: "0077_automation_leases", sql: automation_leases_default, drizzleHash: "22f09427b49eb0fcdd2b62cd472735a88e5dd60974af0d0c008f6e6fddfdf545", when: 17877888e5 },
       { name: "0078_mcp_headers", sql: mcp_headers_default, drizzleHash: "2e2b040eac31e336e8e7662555fd8526e8cdfd59929671e9ca6e063a3f2d4004", when: 17878752e5 },
-      { name: "0079_usage_stats", sql: usage_stats_default, drizzleHash: "8ea9653e8339a4f4c4f0fb140766daace073443ed9e6efae3d4597b2fc6bd60d", when: 17879616e5 }
+      { name: "0079_usage_stats", sql: usage_stats_default, drizzleHash: "8ea9653e8339a4f4c4f0fb140766daace073443ed9e6efae3d4597b2fc6bd60d", when: 17879616e5 },
+      { name: "0080_custom_provider_api_format", sql: custom_provider_api_format_default, drizzleHash: "0c66aed1a57dea6a1e9925a9a74dc1ca4f181db3de0c8e446a860a9f3ef7b950", when: 1788048e6 }
     ];
     DRIZZLE_HASH_TO_NAME = {
       "2fd0348e72ac1429b9246311e8db17a34078771470c622a13ab09afdee0a0d31": "0000_nebulous_proemial_gods",
@@ -11157,7 +11171,8 @@ var init_bundled_generated = __esm({
       "b361a2b664ce74a3efeaa2f73cdf6de582509c6569a1aab54febbef507b98b3c": "0076_drop_installed_skills",
       "22f09427b49eb0fcdd2b62cd472735a88e5dd60974af0d0c008f6e6fddfdf545": "0077_automation_leases",
       "2e2b040eac31e336e8e7662555fd8526e8cdfd59929671e9ca6e063a3f2d4004": "0078_mcp_headers",
-      "8ea9653e8339a4f4c4f0fb140766daace073443ed9e6efae3d4597b2fc6bd60d": "0079_usage_stats"
+      "8ea9653e8339a4f4c4f0fb140766daace073443ed9e6efae3d4597b2fc6bd60d": "0079_usage_stats",
+      "0c66aed1a57dea6a1e9925a9a74dc1ca4f181db3de0c8e446a860a9f3ef7b950": "0080_custom_provider_api_format"
     };
   }
 });
@@ -31811,7 +31826,26 @@ function renderModeSystemBlock(activeMode) {
   lines.push("</mode_system>");
   return lines.join("\n");
 }
-async function buildSystemPromptWithProjectInstructions(params) {
+function joinSystemPromptParts({ stableHeader, dynamicTail }) {
+  return `${stableHeader}
+
+${dynamicTail}
+`;
+}
+function splitDynamicTail(systemPrompt, dynamicTail) {
+  if (!dynamicTail?.trim()) return systemPrompt;
+  const joined = `
+
+${dynamicTail}
+`;
+  const at2 = systemPrompt.indexOf(joined);
+  if (at2 < 0) return systemPrompt;
+  return {
+    systemPromptOverride: systemPrompt.slice(0, at2) + systemPrompt.slice(at2 + joined.length),
+    dynamicSystemTail: dynamicTail
+  };
+}
+async function buildSystemPromptPartsWithProjectInstructions(params) {
   const {
     workspacePath,
     workspacePromptContent,
@@ -31840,7 +31874,7 @@ async function buildSystemPromptWithProjectInstructions(params) {
     buildInstructionsFromLayers(layers),
     projectDocsHint
   ].filter((s3) => s3.trim().length > 0).join("\n\n");
-  return constructSystemPrompt({
+  return constructSystemPromptParts({
     contextSize,
     chatMode,
     effectiveWorkMode,
@@ -31866,7 +31900,7 @@ async function buildSystemPromptWithProjectInstructions(params) {
     registeredToolNames
   });
 }
-var import_electron_log10, logger8, cachedPromptShell, AGENTS_MD, loadAgentsFile, FALLBACK_AGENTS, getAgentsMdForContextSize, getAgentsMdForWorkMode, MODE_SYSTEM_DESCRIPTIONS, constructSystemPromptParts, constructSystemPrompt;
+var import_electron_log10, logger8, cachedPromptShell, AGENTS_MD, loadAgentsFile, FALLBACK_AGENTS, getAgentsMdForContextSize, getAgentsMdForWorkMode, MODE_SYSTEM_DESCRIPTIONS, constructSystemPromptParts;
 var init_system_prompt = __esm({
   "src/prompts/system_prompt.ts"() {
     import_electron_log10 = __toESM(require_src());
@@ -32047,13 +32081,6 @@ When you need clarification or a decision from the user (architecture choices, d
       }
       return { stableHeader, dynamicTail: tail };
     };
-    constructSystemPrompt = (params) => {
-      const { stableHeader, dynamicTail } = constructSystemPromptParts(params);
-      return `${stableHeader}
-
-${dynamicTail}
-`;
-    };
   }
 });
 
@@ -67768,7 +67795,11 @@ var init_provider_sdk_factories = __esm({
         const provider = createOpenAI({ apiKey, ...baseURL ? { baseURL } : {} });
         return baseURL ? provider(modelId) : provider.responses(modelId);
       },
-      "@ai-sdk/anthropic": ({ modelId, apiKey, baseURL }) => createAnthropic({ apiKey, ...baseURL ? { baseURL } : {} })(modelId),
+      // Exactly one of apiKey / authToken: the SDK refuses both.
+      "@ai-sdk/anthropic": ({ modelId, apiKey, baseURL, authHeader }) => createAnthropic({
+        ...authHeader === "bearer" ? { authToken: apiKey } : { apiKey },
+        ...baseURL ? { baseURL } : {}
+      })(modelId),
       "@ai-sdk/google": ({ modelId, apiKey, baseURL }) => createGoogle({ apiKey, ...baseURL ? { baseURL } : {} })(modelId),
       "@ai-sdk/xai": ({ modelId, apiKey, baseURL }) => createXai({ apiKey, ...baseURL ? { baseURL } : {} })(modelId)
     };
@@ -68427,7 +68458,7 @@ var init_model_reasoning_profile = __esm({
     temperatureById = /* @__PURE__ */ new Map();
     effortOverrides = /* @__PURE__ */ new Map();
     BUDGET_LEVELS = ["low", "medium", "high"];
-    BUDGET_HINT = { low: "medium", medium: "high", high: "max" };
+    BUDGET_HINT = { low: "low", medium: "medium", high: "max" };
     NO_REASONING = /* @__PURE__ */ new Set(["none", "minimal"]);
     AUTO_EFFORT_LEVELS = ["low", "medium", "max"];
     AUTO_EFFORT_CHOICES = { kind: "effort", levels: [...AUTO_EFFORT_LEVELS], defaultLevel: "medium" };
@@ -68963,6 +68994,18 @@ var init_models_dev = __esm({
   }
 });
 
+// src/lib/provider-key-check.ts
+function anthropicHeaders(apiKey, authHeader) {
+  return {
+    ...authHeader === "bearer" ? { Authorization: `Bearer ${apiKey}` } : { "x-api-key": apiKey },
+    "anthropic-version": "2023-06-01"
+  };
+}
+var init_provider_key_check = __esm({
+  "src/lib/provider-key-check.ts"() {
+  }
+});
+
 // node_modules/isexe/windows.js
 var require_windows = __commonJS({
   "node_modules/isexe/windows.js"(exports2, module2) {
@@ -71123,21 +71166,22 @@ async function fetchModelsFromOpenAI(apiKey) {
     return [];
   }
 }
-async function fetchModelsFromAnthropic(apiKey) {
-  const cached4 = fromCache("anthropic");
+async function fetchModelsFromAnthropic(apiKey, endpoint2) {
+  const cacheId = endpoint2?.providerId ?? "anthropic";
+  const label = endpoint2 ? endpoint2.providerId : "Anthropic";
+  const cached4 = fromCache(cacheId);
   if (cached4) return cached4;
   try {
     const items = [];
     let after2;
     for (let page = 0; page < MAX_PAGES; page++) {
-      const url2 = new URL("https://api.anthropic.com/v1/models");
+      const url2 = new URL(
+        endpoint2 ? `${endpoint2.baseUrl.replace(/\/+$/, "")}/models` : "https://api.anthropic.com/v1/models"
+      );
       url2.searchParams.set("limit", "1000");
       if (after2) url2.searchParams.set("after_id", after2);
       const res = await fetch(url2, {
-        headers: {
-          "x-api-key": apiKey,
-          "anthropic-version": "2023-06-01"
-        },
+        headers: anthropicHeaders(apiKey, endpoint2?.authHeader),
         signal: AbortSignal.timeout(15e3)
       });
       if (!res.ok) throw new Error(`HTTP ${res.status}`);
@@ -71149,18 +71193,18 @@ async function fetchModelsFromAnthropic(apiKey) {
     const models = items.map((m3) => ({
       apiName: m3.id,
       displayName: m3.display_name ?? m3.id,
-      description: "Claude model with tool support",
+      description: endpoint2 ? "Chat model with tool support" : "Claude model with tool support",
       ...m3.max_input_tokens ? { contextWindow: m3.max_input_tokens } : {},
       ...m3.max_tokens ? { maxOutputTokens: m3.max_tokens } : {},
       type: "cloud",
       supportsToolChoice: true
     }));
     sortModels(models);
-    toCache("anthropic", models);
-    logger10.info(`[Anthropic] Fetched ${models.length} models`);
+    toCache(cacheId, models);
+    logger10.info(`[${label}] Fetched ${models.length} models`);
     return models;
   } catch (e3) {
-    logger10.warn("[Anthropic] Fetch failed", e3);
+    logger10.warn(`[${label}] Fetch failed`, e3);
     return [];
   }
 }
@@ -71363,7 +71407,14 @@ async function fetchModelsFromOpenRouter(apiKey) {
     return [];
   }
 }
-async function fetchCustomProviderModels(providerId, baseUrl, apiKey) {
+async function fetchCustomProviderModels(providerId, baseUrl, apiKey, wire = {}) {
+  if (wire.apiFormat === "anthropic") {
+    return fetchModelsFromAnthropic(apiKey?.trim() || "none", {
+      providerId,
+      baseUrl,
+      authHeader: wire.authHeader
+    });
+  }
   return fetchOpenAICompatibleModels({
     providerId,
     url: `${baseUrl.replace(/\/+$/, "")}/models`,
@@ -71434,6 +71485,7 @@ var init_provider_model_fetchers = __esm({
     import_electron_log12 = __toESM(require_src());
     init_openrouter_app_headers();
     init_language_model_constants();
+    init_provider_key_check();
     logger10 = import_electron_log12.default.scope("provider_fetchers");
     CACHE_MS = 5 * 60 * 1e3;
     cache = /* @__PURE__ */ new Map();
@@ -71560,6 +71612,29 @@ var init_provider_api_keys = __esm({
   }
 });
 
+// src/lib/routing/custom-provider-format.ts
+function rememberCustomProviderFormats(providers) {
+  anthropicFormat.clear();
+  for (const p2 of providers) {
+    if (p2.apiFormat === "anthropic") anthropicFormat.add(p2.id);
+  }
+}
+function isCustomAnthropicFormat(providerId) {
+  return Boolean(providerId && anthropicFormat.has(providerId));
+}
+function anthropicFormatCapabilityId(providerId) {
+  return isCustomAnthropicFormat(providerId) ? "anthropic" : providerId;
+}
+function anthropicCatalogModelId(modelId) {
+  return modelId.trim().replace(/^(?:[a-z]{2,6}\.)?anthropic\./i, "").replace(/-v\d+(?::\d+)?$/i, "");
+}
+var anthropicFormat;
+var init_custom_provider_format = __esm({
+  "src/lib/routing/custom-provider-format.ts"() {
+    anthropicFormat = /* @__PURE__ */ new Set();
+  }
+});
+
 // src/ipc/utils/thinking_utils.ts
 function reasoningCeiling(providerId) {
   return providerId && PROVIDERS_THAT_SUPPORT_THINKING.includes(providerId) ? "max" : "high";
@@ -71583,7 +71658,7 @@ function reasoningBudgetTokens(effort, maxOutputTokens) {
   const max4 = Math.max(1, Math.floor(maxOutputTokens));
   const reserve = Math.max(Math.round(max4 * 0.15), 2048);
   const ceil = Math.max(1024, max4 - reserve);
-  const frac = effort === "max" ? 0.85 : effort === "xhigh" ? 0.7 : effort === "high" ? 0.5 : 0.33;
+  const frac = effort === "max" ? 0.85 : effort === "xhigh" ? 0.7 : effort === "high" ? 0.5 : effort === "low" ? 0.15 : 0.33;
   return Math.min(ceil, Math.max(1024, Math.round(max4 * frac)));
 }
 function getExtraProviderOptions(providerId, settings, includeExtendedThinking = false, maxOutputTokens, modelId) {
@@ -71667,6 +71742,9 @@ function resolveReasoningCatalogModelId(effectiveModel) {
   if (effectiveModel.provider === "compass") {
     return getCoreModelsDevTargetForTier(name29)?.modelId ?? getCoreModelIdForCoreTier(name29) ?? name29;
   }
+  if (isCustomAnthropicFormat(effectiveModel.provider)) {
+    return anthropicCatalogModelId(name29);
+  }
   return name29;
 }
 function resolveMediaModelId(effectiveModel) {
@@ -71677,13 +71755,13 @@ function resolveMediaModelId(effectiveModel) {
 }
 function resolveCapabilityProviderId(effectiveModel) {
   const provider = (effectiveModel.provider ?? "").trim();
-  if (provider !== "compass") return provider;
+  if (provider !== "compass") return anthropicFormatCapabilityId(provider);
   return getCoreModelsDevTargetForTier(effectiveModel.name ?? "")?.providerId ?? provider;
 }
 function detectReasoningCapability(modelApiId, builtinProviderId) {
   const id = modelApiId.trim();
   const lower2 = id.toLowerCase();
-  const pid = builtinProviderId ?? "";
+  const pid = anthropicFormatCapabilityId(builtinProviderId) ?? "";
   let useOpenRouterReasoning = false;
   if (pid === "openrouter" || pid === "compass") {
     const row = id.includes("/") ? getCachedOpenRouterModelRow(id) : void 0;
@@ -71737,7 +71815,7 @@ function isReasoningCapableModel(model) {
   return offersAChannel(capabilityProvider) || offersAChannel(model.provider ?? "");
 }
 function resolveDispatchCapabilityProviderId(effectiveModel, builtinProviderId) {
-  if (builtinProviderId !== "compass") return builtinProviderId;
+  if (builtinProviderId !== "compass") return anthropicFormatCapabilityId(builtinProviderId);
   if (effectiveModel.provider !== "compass") return builtinProviderId;
   return resolveCapabilityProviderId(effectiveModel);
 }
@@ -71833,6 +71911,7 @@ var init_reasoning_detector = __esm({
     init_model_capability_profile();
     init_language_model_constants();
     init_model_reasoning_profile();
+    init_custom_provider_format();
     init_thinking_utils();
     GEMINI_LEVELS = /* @__PURE__ */ new Set(["minimal", "low", "medium", "high"]);
   }
@@ -72009,11 +72088,16 @@ async function getLanguageModelProviders() {
       name: cp.name,
       apiBaseUrl: cp.api_base_url,
       envVarName: cp.env_var_name ?? void 0,
+      // The Anthropic format dispatches through the catalog's own Anthropic SDK
+      // factory (provider-sdk-factories), like any models.dev provider on it.
+      ...cp.api_format === "anthropic" ? { apiFormat: "anthropic", npm: ANTHROPIC_NPM } : {},
+      ...cp.auth_header === "x-api-key" || cp.auth_header === "bearer" ? { authHeader: cp.auth_header } : {},
       type: "custom"
       // hasFreeTier, websiteUrl, gatewayPrefix are not in the custom DB schema
       // They will be undefined unless overridden by hardcoded values if IDs match
     });
   }
+  rememberCustomProviderFormats([...customProvidersMap.values()]);
   const catalogById = /* @__PURE__ */ new Map();
   try {
     for (const c2 of await getModelsDevProviderCatalog())
@@ -72100,15 +72184,15 @@ function providerDisplayName(providerId) {
   if (!providerId) return void 0;
   return providerNames.get(providerId) ?? providerId.replace(/^custom::/, "");
 }
-async function liveCustomModels(providerId, baseUrl, apiKey, now2 = Date.now()) {
+async function liveCustomModels(providerId, baseUrl, apiKey, now2 = Date.now(), wire = {}) {
   const { createHash: createHash10 } = await import("node:crypto");
   const keyTag = apiKey?.trim() ? createHash10("sha256").update(apiKey.trim()).digest("hex").slice(0, 12) : "none";
-  const cacheKey = `${providerId}|${baseUrl}|${keyTag}`;
+  const cacheKey = `${providerId}|${baseUrl}|${keyTag}|${wire.apiFormat ?? "openai"}|${wire.authHeader ?? ""}`;
   const hit = liveCache.get(cacheKey);
   if (hit && now2 - hit.at < LIVE_TTL_MS) return hit.models;
   const { fetchCustomProviderModels: fetchCustomProviderModels2 } = await Promise.resolve().then(() => (init_provider_model_fetchers(), provider_model_fetchers_exports));
   const models = Promise.race([
-    fetchCustomProviderModels2(providerId, baseUrl, apiKey).catch(
+    fetchCustomProviderModels2(providerId, baseUrl, apiKey, wire).catch(
       () => []
     ),
     new Promise(
@@ -72141,7 +72225,10 @@ async function getLanguageModels({
       maxOutputTokens: language_models.max_output_tokens,
       contextWindow: language_models.context_window
     }).from(language_models).where(
-      isCustomProvider({ providerId }) ? eq(language_models.customProviderId, providerId) : eq(language_models.builtinProviderId, providerId)
+      // A provider from the providers table, whatever its id: the CLI's dev
+      // endpoint is `atlarix-cli-dev`, with no `custom::`, and its model row
+      // was looked up as a built-in's and never found (QA B3).
+      provider.type === "custom" || isCustomProvider({ providerId }) ? eq(language_models.customProviderId, providerId) : eq(language_models.builtinProviderId, providerId)
     );
     customModels = customModelsDb.map((model) => ({
       ...model,
@@ -72202,7 +72289,10 @@ async function getLanguageModels({
       key = readSettings2().providerSettings?.[providerId]?.apiKey?.value;
     } catch {
     }
-    const listed = await liveCustomModels(providerId, provider.apiBaseUrl, key);
+    const listed = await liveCustomModels(providerId, provider.apiBaseUrl, key, Date.now(), {
+      apiFormat: provider.apiFormat,
+      authHeader: provider.authHeader
+    });
     const typed = new Set(customModels.map((m3) => m3.apiName));
     return [...customModels, ...listed.filter((m3) => !typed.has(m3.apiName))];
   }
@@ -72313,7 +72403,7 @@ function stampDeepThinkingSupport(modelsByProvider) {
 function isCustomProvider({ providerId }) {
   return providerId.startsWith(CUSTOM_PROVIDER_PREFIX);
 }
-var providerNames, LIVE_TTL_MS, liveCache, CUSTOM_PROVIDER_PREFIX;
+var ANTHROPIC_NPM, providerNames, LIVE_TTL_MS, liveCache, CUSTOM_PROVIDER_PREFIX;
 var init_language_model_helpers = __esm({
   "src/ipc/shared/language_model_helpers.ts"() {
     init_db2();
@@ -72328,6 +72418,8 @@ var init_language_model_helpers = __esm({
     init_compass_constants();
     init_reasoning_detector();
     init_model_capability_profile();
+    init_custom_provider_format();
+    ANTHROPIC_NPM = "@ai-sdk/anthropic";
     providerNames = /* @__PURE__ */ new Map();
     LIVE_TTL_MS = 6e4;
     liveCache = /* @__PURE__ */ new Map();
@@ -95745,7 +95837,8 @@ async function resolveModelContextLimit(model, settings) {
   if (pid.startsWith("custom::")) {
     try {
       const { getLanguageModelProviders: getLanguageModelProviders2 } = await Promise.resolve().then(() => (init_language_model_helpers(), language_model_helpers_exports));
-      const base = (await getLanguageModelProviders2()).find((p2) => p2.id.toLowerCase() === pid)?.apiBaseUrl;
+      const custom2 = (await getLanguageModelProviders2()).find((p2) => p2.id.toLowerCase() === pid);
+      const base = custom2?.apiFormat === "anthropic" ? void 0 : custom2?.apiBaseUrl;
       if (base) {
         const { isOllamaEndpoint: isOllamaEndpoint2, ollamaOrigin: ollamaOrigin2, ollamaNumCtx: ollamaNumCtx2 } = await Promise.resolve().then(() => (init_ollama_native_fetch(), ollama_native_fetch_exports));
         if (await isOllamaEndpoint2(base)) {
@@ -95768,10 +95861,7 @@ async function resolveModelContextLimit(model, settings) {
   }
   try {
     const { getModelsDevLimits: getModelsDevLimits2 } = await Promise.resolve().then(() => (init_models_dev(), models_dev_exports));
-    const limits = await getModelsDevLimits2(
-      model.provider ?? "",
-      model.name ?? ""
-    );
+    const limits = isCustomAnthropicFormat(model.provider) ? await getModelsDevLimits2("anthropic", anthropicCatalogModelId(model.name ?? "")) : await getModelsDevLimits2(model.provider ?? "", model.name ?? "");
     if (limits?.context && limits.context > 0) {
       return {
         tierLimit: floorSafe(limits.context * clamp),
@@ -95793,6 +95883,7 @@ var init_resolve_model_context_limit = __esm({
     init_compass_constants();
     init_findLanguageModel();
     import_electron_log16 = __toESM(require_src());
+    init_custom_provider_format();
     logger14 = import_electron_log16.default.scope("context-limit");
     SAFETY_FACTOR = 0.8;
     localContextCache = /* @__PURE__ */ new Map();
@@ -215041,7 +215132,7 @@ async function getRegularModelClient(model, settings, providerConfig, options) {
       };
     }
     default: {
-      if (providerConfig.type === "custom" && providerConfig.apiBaseUrl) {
+      if (providerConfig.type === "custom" && providerConfig.apiBaseUrl && providerConfig.apiFormat !== "anthropic") {
         const { isOllamaEndpoint: isOllamaEndpoint2, ollamaOrigin: ollamaOrigin2 } = await Promise.resolve().then(() => (init_ollama_native_fetch(), ollama_native_fetch_exports));
         if (await isOllamaEndpoint2(providerConfig.apiBaseUrl)) {
           return {
@@ -215061,7 +215152,8 @@ async function getRegularModelClient(model, settings, providerConfig, options) {
               modelId: model.name,
               apiKey,
               baseURL: providerConfig.apiBaseUrl,
-              npm: providerConfig.npm
+              npm: providerConfig.npm,
+              authHeader: providerConfig.authHeader
             }),
             builtinProviderId: providerConfig.id
           },
@@ -241873,6 +241965,7 @@ var init_dist21 = __esm({
 function resolveCacheMarkerPlacement(builtinProviderId, modelName) {
   if (builtinProviderId === "anthropic") return "message";
   if (!/claude|anthropic/i.test(modelName ?? "")) return "none";
+  if (isCustomAnthropicFormat(builtinProviderId)) return "message";
   if (builtinProviderId === "bedrock" || builtinProviderId === "vertex") {
     return "message";
   }
@@ -241984,6 +242077,7 @@ function advanceCacheBreakpointsForStep(messages, placement) {
 var CACHE_MARKER, CACHE_MARKER_HEADER_1H, CACHE_KEY_BY_PROVIDER;
 var init_anthropic_cache_breakpoints = __esm({
   "src/lib/ai/anthropic-cache-breakpoints.ts"() {
+    init_custom_provider_format();
     CACHE_MARKER = {
       anthropic: { cacheControl: { type: "ephemeral" } },
       openrouter: { cacheControl: { type: "ephemeral" } },
@@ -338396,7 +338490,8 @@ async function simpleStreamText(ctx, deps, {
         ...reasoningPatch.anthropic
       };
     }
-    const handledNatively = offByBudget || Boolean(reasoningPatch.openrouter) || Boolean(reasoningPatch.google) || Boolean(reasoningPatch.openai) || Boolean(reasoningPatch.anthropic);
+    const handledNatively = offByBudget || // The Anthropic SDK reads only its own bucket, in its own shape.
+    isCustomAnthropicFormat(modelClient.builtinProviderId) || Boolean(reasoningPatch.openrouter) || Boolean(reasoningPatch.google) || Boolean(reasoningPatch.openai) || Boolean(reasoningPatch.anthropic);
     if (modelClient.builtinProviderId && !handledNatively) {
       const compatibleReasoning = getExtraProviderOptions(
         capabilityProviderId,
@@ -338415,6 +338510,7 @@ async function simpleStreamText(ctx, deps, {
     }
   }
   const isAnthropic = modelClient.builtinProviderId === "anthropic";
+  const speaksAnthropic = isAnthropic || isCustomAnthropicFormat(modelClient.builtinProviderId);
   const cacheMarkerPlacement = resolveCacheMarkerPlacement(
     modelClient.builtinProviderId,
     effectiveModel?.name
@@ -338486,7 +338582,7 @@ async function simpleStreamText(ctx, deps, {
     history,
     placement: cacheMarkerPlacement,
     // v13.3: header prefix survives >5min between-turn gaps.
-    headerTtl1h: isAnthropic
+    headerTtl1h: speaksAnthropic
   });
   const completedStepMessages = [];
   const recordStepMessages = (msgs) => {
@@ -338935,6 +339031,7 @@ var init_simple_stream_text = __esm({
     init_perf_marks();
     init_byok_pricing();
     init_reasoning_detector();
+    init_custom_provider_format();
     init_system_reminder();
     init_step_messages();
     init_texts();
@@ -343033,6 +343130,7 @@ async function runBuildBranch(ctx, deps) {
     modelContextResolved,
     placeholderAssistantMessage,
     systemPrompt,
+    dynamicTail,
     taskLedgerBlock,
     simpleStreamTextDeps,
     streamTurnId,
@@ -344280,7 +344378,9 @@ async function runBuildBranch(ctx, deps) {
           escalationMode: effectiveWorkMode === "plan" ? "plan" : effectiveWorkMode === "build" ? "build" : "agent",
           tools,
           repetition: agenticRepetition,
-          buildRoundSystemPrompt: () => systemPrompt,
+          // The per-turn tail in its own system message, after the cached
+          // header, so the next turn reads the cache instead of rewriting it.
+          buildRoundSystemPrompt: () => splitDynamicTail(systemPrompt, dynamicTail),
           detectTextToolCalls: (fr2) => useTextToolFallback && containsTextToolCall2(fr2) ? parseTextToolCalls2(fr2) : null,
           executeTextToolCalls: (calls) => execBuildTextToolCalls(
             calls
@@ -344423,6 +344523,7 @@ var init_build_branch = __esm({
     init_dist21();
     import_electron_log102 = __toESM(require_src());
     init_edit_stats2();
+    init_system_prompt();
     logger99 = import_electron_log102.default.scope("chat_stream_handlers");
   }
 });
@@ -346470,11 +346571,11 @@ function registerChatStreamHandlers() {
               providerBrand: modelIdentity.providerBrand,
               labelWithProvider: modelIdentity.labelWithProvider
             };
-            let baseSystemPrompt;
+            let baseParts;
             if (hasWorkspace && appPath && effectiveAppId != null) {
-              baseSystemPrompt = await preflight.time(
+              baseParts = await preflight.time(
                 "system-prompt",
-                () => buildSystemPromptWithProjectInstructions({
+                () => buildSystemPromptPartsWithProjectInstructions({
                   workspacePath: appPath,
                   workspacePromptContent: null,
                   contextSize,
@@ -346487,7 +346588,7 @@ function registerChatStreamHandlers() {
                 })
               );
             } else {
-              baseSystemPrompt = constructSystemPrompt({
+              baseParts = constructSystemPromptParts({
                 chatMode,
                 effectiveWorkMode: modeBlockKey,
                 contextSize,
@@ -346497,6 +346598,8 @@ function registerChatStreamHandlers() {
                 ...promptIdentityOpts
               });
             }
+            let baseSystemPrompt = joinSystemPromptParts(baseParts);
+            const baseDynamicTail = baseParts.dynamicTail;
             if (activeUserAgent) {
               const { renderUserAgentSystemBlock: renderUserAgentSystemBlock2 } = await Promise.resolve().then(() => (init_user_agents(), user_agents_exports));
               baseSystemPrompt = `${renderUserAgentSystemBlock2(activeUserAgent)}
@@ -347006,6 +347109,7 @@ ${baseSystemPrompt}`;
                 modelContextResolved,
                 placeholderAssistantMessage,
                 systemPrompt,
+                dynamicTail: baseDynamicTail,
                 taskLedgerBlock,
                 simpleStreamTextDeps,
                 streamTurnId,
@@ -347939,7 +348043,7 @@ ${err2.stack}` : "";
 }
 
 // src/headless/atlarix-headless.mts
-var VERSION15 = false ? "dev" : "15.8.2";
+var VERSION15 = false ? "dev" : "15.9.0";
 if (process.argv[2] === "run") process.argv.splice(2, 1);
 var RUN_HELP = `atlarix run: one turn without the UI, for scripts and CI
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "atlarix",
-  "version": "15.8.2",
+  "version": "15.9.0",
   "description": "Atlarix CLI: an AI coding agent in your terminal. Works with any model (Atlarix Auto, your own API key, or a local model) and runs unattended with `atlarix run` for CI.",
   "license": "PROPRIETARY",
   "author": {
@@ -65,12 +65,12 @@
     "provenance": false
   },
   "optionalDependencies": {
-    "atlarix-darwin-arm64": "15.8.2",
-    "atlarix-darwin-x64": "15.8.2",
-    "atlarix-linux-x64": "15.8.2",
-    "atlarix-linux-arm64": "15.8.2",
-    "atlarix-linux-x64-musl": "15.8.2",
-    "atlarix-linux-arm64-musl": "15.8.2",
-    "atlarix-windows-x64": "15.8.2"
+    "atlarix-darwin-arm64": "15.9.0",
+    "atlarix-darwin-x64": "15.9.0",
+    "atlarix-linux-x64": "15.9.0",
+    "atlarix-linux-arm64": "15.9.0",
+    "atlarix-linux-x64-musl": "15.9.0",
+    "atlarix-linux-arm64-musl": "15.9.0",
+    "atlarix-windows-x64": "15.9.0"
   }
 }