@axlsdk/axl 0.25.1 → 0.25.3

package/dist/index.d.cts

@@ -7378,8 +7378,8 @@ type DirectOpenAIPricingContext = {
     };
 };
 /**
- * Estimate OpenAI call cost from token usage. Exact match first, then
- * longest-prefix match for versioned models (e.g. `gpt-4o-2024-05-13`).
+ * Estimate OpenAI call cost from token usage using exact catalog IDs only.
+ * Explicitly listed snapshots (e.g. `gpt-4o-2024-05-13`) have their own rows.
  *
  * Returns `undefined` when the model is not in the table — callers must treat
  * that as "unknown cost", never as free (a silent `0` would break
@@ -7409,8 +7409,8 @@ declare class OpenAIProvider extends OpenAICompatibleProvider {
         dangerouslyAllowInsecureHttp?: boolean;
         rateLimit?: RateLimitConfig;
     });
-    /** Report a clamped `effort` — Chat Completions caps `'max'` at `'xhigh'` and
-     *  rejects `'none'` before gpt-5.1 (clamped to `'minimal'`). */
+    /** Report a clamped `effort` — Chat caps `'max'` at `'xhigh'` on GPT-5.6 and
+     *  GPT-6.1 Sol, and maps `'none'` to the model's supported floor. */
     effortResolution(options: Pick<ChatOptions, 'model' | 'effort' | 'thinkingBudget' | 'includeThoughts' | 'providerOptions'>): EffortResolution | undefined;
     protected validateFinalRequestBody(body: Record<string, unknown>): void;
     protected computeCost(model: string, usage: ProviderResponse['usage'], _reportedCost: number | undefined, context?: DirectOpenAIPricingContext): number | undefined;
@@ -7534,7 +7534,19 @@ declare class AnthropicProvider implements Provider {
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
     private buildHeaders;
+    /**
+     * Whether the final body needs the binding beta header. Runs after
+     * `validateAndBindFinalBody`, which has already injected `block_binding`
+     * wherever it applies, so the header follows the final body rather than
+     * re-deriving the injection rule.
+     */
     private requiresBindingBeta;
+    /**
+     * Validate the final body (after `providerOptions`) against the effective
+     * model's capability, then inject `block_binding` where it applies. Every
+     * rule is gated on its own capability field and is a no-op when unset, so a
+     * model only gets the guards its provider contract calls for.
+     */
     private validateAndBindFinalBody;
     /** Extract a human-readable message from an API error response body. */
     private extractErrorMessage;

package/dist/index.d.ts

@@ -7378,8 +7378,8 @@ type DirectOpenAIPricingContext = {
     };
 };
 /**
- * Estimate OpenAI call cost from token usage. Exact match first, then
- * longest-prefix match for versioned models (e.g. `gpt-4o-2024-05-13`).
+ * Estimate OpenAI call cost from token usage using exact catalog IDs only.
+ * Explicitly listed snapshots (e.g. `gpt-4o-2024-05-13`) have their own rows.
  *
  * Returns `undefined` when the model is not in the table — callers must treat
  * that as "unknown cost", never as free (a silent `0` would break
@@ -7409,8 +7409,8 @@ declare class OpenAIProvider extends OpenAICompatibleProvider {
         dangerouslyAllowInsecureHttp?: boolean;
         rateLimit?: RateLimitConfig;
     });
-    /** Report a clamped `effort` — Chat Completions caps `'max'` at `'xhigh'` and
-     *  rejects `'none'` before gpt-5.1 (clamped to `'minimal'`). */
+    /** Report a clamped `effort` — Chat caps `'max'` at `'xhigh'` on GPT-5.6 and
+     *  GPT-6.1 Sol, and maps `'none'` to the model's supported floor. */
     effortResolution(options: Pick<ChatOptions, 'model' | 'effort' | 'thinkingBudget' | 'includeThoughts' | 'providerOptions'>): EffortResolution | undefined;
     protected validateFinalRequestBody(body: Record<string, unknown>): void;
     protected computeCost(model: string, usage: ProviderResponse['usage'], _reportedCost: number | undefined, context?: DirectOpenAIPricingContext): number | undefined;
@@ -7534,7 +7534,19 @@ declare class AnthropicProvider implements Provider {
     chat(messages: ChatMessage[], options: ChatOptions): Promise<ProviderResponse>;
     stream(messages: ChatMessage[], options: ChatOptions): AsyncGenerator<StreamChunk>;
     private buildHeaders;
+    /**
+     * Whether the final body needs the binding beta header. Runs after
+     * `validateAndBindFinalBody`, which has already injected `block_binding`
+     * wherever it applies, so the header follows the final body rather than
+     * re-deriving the injection rule.
+     */
     private requiresBindingBeta;
+    /**
+     * Validate the final body (after `providerOptions`) against the effective
+     * model's capability, then inject `block_binding` where it applies. Every
+     * rule is gated on its own capability field and is a no-op when unset, so a
+     * model only gets the guards its provider contract calls for.
+     */
     private validateAndBindFinalBody;
     /** Extract a human-readable message from an API error response body. */
     private extractErrorMessage;

package/dist/index.js

@@ -348,7 +348,7 @@ var sharedGuard = typeof installedGuard === "object" && installedGuard !== null
 var buildPath = typeof __AXL_BUILD_PATH__ === "string" ? __AXL_BUILD_PATH__ : "<source>";
 var thisCopy = {
   path: buildPath.startsWith("file:") ? fileURLToPath(buildPath) : buildPath,
-  version: true ? "0.25.1" : "unknown"
+  version: true ? "0.25.3" : "unknown"
 };
 function warnCrossCopy(other, joined) {
   if (other === thisCopy) return;
@@ -3617,7 +3617,7 @@ var OPENAI_PRICING = {
   "gpt-4.1-mini": [4e-7, 16e-7, 0.25],
   "gpt-4.1-nano": [1e-7, 4e-7, 0.25],
   o3: [2e-6, 8e-6, 0.25],
-  "o3-mini": [11e-7, 44e-7, 0.25],
+  "o3-mini": [11e-7, 44e-7, 0.5],
   "o4-mini": [11e-7, 44e-7, 0.25],
   "gpt-5": [125e-8, 1e-5, 0.1],
   "gpt-5-mini": [25e-8, 2e-6, 0.1],
@@ -3640,6 +3640,13 @@ var DIRECT_OPENAI_CATALOG = [
     long: { input: 20 / M, cachedInput: 2 / M, cacheWrite: 25 / M, output: 75 / M },
     contextBoundary: LONG_CONTEXT_BOUNDARY
   },
+  {
+    aliases: ["gpt-6.1-sol"],
+    // Rechecked 2026-10-15 against https://developers.openai.com/api/docs/pricing.
+    short: { input: 2 / M, cachedInput: 0.1 / M, cacheWrite: 2.5 / M, output: 10 / M },
+    long: { input: 4 / M, cachedInput: 0.2 / M, cacheWrite: 5 / M, output: 15 / M },
+    contextBoundary: LONG_CONTEXT_BOUNDARY
+  },
   {
     aliases: ["gpt-6-sol"],
     short: { input: 2 / M, cachedInput: 0.2 / M, cacheWrite: 2.5 / M, output: 10 / M },
@@ -4019,12 +4026,12 @@ var BASE_EFFORTS = ["minimal", "low", "medium", "high"];
 function bothEndpoints(value) {
   return { chat: value, responses: value };
 }
-function gpt6(id, efforts, chatTools) {
+function gpt6(id, responsesEfforts, chatTools, chatEfforts = responsesEfforts) {
   return {
     id,
     family: "gpt-6",
     reasoning: true,
-    efforts: bothEndpoints(efforts),
+    efforts: { chat: chatEfforts, responses: responsesEfforts },
     defaultEffort: "medium",
     // Same policy as GPT-5.x: a portable `temperature` is dropped while
     // reasoning is active (including the active default). A raw
@@ -4052,6 +4059,13 @@ function gpt56(id) {
 var OPENAI_EXACT_MODELS = new Map(
   [
     gpt6("gpt-6-astra", ["low", "medium", "high", "xhigh", "max"], "never"),
+    // Live Chat Completions rejects max; Responses accepts it (2026-10-15).
+    gpt6("gpt-6.1-sol", ["low", "medium", "high", "xhigh", "max"], "never", [
+      "low",
+      "medium",
+      "high",
+      "xhigh"
+    ]),
     gpt6("gpt-6-sol", ["none", "low", "medium", "high", "xhigh", "max"], "effort-none-only"),
     gpt6("gpt-6-luna", ["none", "low", "medium", "high", "xhigh", "max"], "effort-none-only"),
     gpt56("gpt-5.6"),
@@ -4195,15 +4209,15 @@ function validateOpenAIFinalRequestBody(body, endpoint) {
     }
   }
   if (activeReasoning) {
+    const remediation = (key) => descriptor.efforts[endpoint].includes("none") ? `Remove ${key} or set reasoning effort to 'none'.` : `Remove ${key}; ${model} does not support reasoning effort 'none'.`;
     const forbidden = endpoint === "chat" ? ["temperature", "top_p", "top_logprobs", "logprobs"] : ["temperature", "top_p", "top_logprobs"];
     for (const key of forbidden) {
-      if (body[key] !== void 0)
-        fail(key, `Remove ${key} or set reasoning effort to 'none' on Sol or Luna.`);
+      if (body[key] !== void 0) fail(key, remediation(key));
     }
     if (endpoint === "responses" && Array.isArray(body.include) && body.include.includes("message.output_text.logprobs")) {
       fail(
         "message.output_text.logprobs",
-        "Remove it from include or set reasoning effort to none on Sol or Luna."
+        remediation("message.output_text.logprobs from include")
       );
     }
   }
@@ -4256,8 +4270,8 @@ var OpenAIProvider = class extends OpenAICompatibleProvider {
   constructor(options = {}) {
     super({ profile: OPENAI_PROFILE, ...options });
   }
-  /** Report a clamped `effort` — Chat Completions caps `'max'` at `'xhigh'` and
-   *  rejects `'none'` before gpt-5.1 (clamped to `'minimal'`). */
+  /** Report a clamped `effort` — Chat caps `'max'` at `'xhigh'` on GPT-5.6 and
+   *  GPT-6.1 Sol, and maps `'none'` to the model's supported floor. */
   effortResolution(options) {
     return resolveOpenAIEffortResolution(
       options,
@@ -4913,6 +4927,17 @@ var ANTHROPIC_THINKING_BINDING_BETA = "thinking-binding-controls-2026-08-01";
 function isRecord(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
+function isFinalThinkingAdaptive(capability, body) {
+  const thinking = body.thinking;
+  if (thinking === void 0) {
+    return capability.thinking === "adaptive-default-on" || capability.thinking === "adaptive-always-on";
+  }
+  return isRecord(thinking) && thinking.type === "adaptive";
+}
+var NATIVE_CLAUDE_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
+function isNativeClaudeEffort(value) {
+  return NATIVE_CLAUDE_EFFORTS.includes(value);
+}
 function hasReplayedThinking(body) {
   return Array.isArray(body.messages) && body.messages.some(
     (message) => isRecord(message) && Array.isArray(message.content) && message.content.some(
@@ -4947,11 +4972,40 @@ function hasAnthropicProviderFileImage(messages) {
   );
 }
 var CLAUDE_CAPABILITIES = {
+  // Thinking off is `between_tools` (progress updates between tool calls);
+  // `disabled` is a provider 400. Forced tool choice is a 400.
+  "claude-sonnet-5-5": {
+    thinking: "adaptive-default-on",
+    effortLevels: ["low", "medium", "high", "xhigh", "max"],
+    manualBudget: false,
+    stripTemperature: true,
+    disableAt: "high",
+    offShape: "between_tools",
+    forcedToolChoice: "rejected",
+    defaultOnlySampling: true,
+    strictThinking: true,
+    thinkingBinding: true
+  },
+  // Thinking off is `disabled`, only at effort high or below. Forced tool
+  // choice is accepted (the provider skips thinking for that turn).
+  "claude-haiku-5-5": {
+    thinking: "adaptive-default-on",
+    effortLevels: ["low", "medium", "high", "xhigh", "max"],
+    manualBudget: false,
+    stripTemperature: true,
+    disableAt: "high",
+    defaultOnlySampling: true,
+    strictThinking: true,
+    thinkingBinding: true
+  },
   "claude-opus-5-5": {
     thinking: "adaptive-always-on",
     effortLevels: ["low", "medium", "high", "xhigh", "max"],
     manualBudget: false,
     stripTemperature: true,
+    forcedToolChoice: "rejected",
+    defaultOnlySampling: true,
+    strictThinking: true,
     thinkingBinding: true
   },
   "claude-fable-5-1": {
@@ -4959,6 +5013,9 @@ var CLAUDE_CAPABILITIES = {
     effortLevels: ["low", "medium", "high", "xhigh", "max"],
     manualBudget: false,
     stripTemperature: true,
+    forcedToolChoice: "rejected",
+    defaultOnlySampling: true,
+    strictThinking: true,
     thinkingBinding: true
   },
   "claude-fable-5": {
@@ -5053,6 +5110,15 @@ function resolveClaudeCapability(model) {
 }
 var DEFAULT_CACHE_READ_MULTIPLIER = 0.1;
 var ANTHROPIC_RATES = {
+  // Sonnet 5.5 cache reads are 0.05x ($0.10/MTok): the pricing table, its
+  // footnote, the models overview, and the prompt-caching guide agree
+  // (re-checked 2026-10-07).
+  "claude-sonnet-5-5": { input: 2e-6, output: 1e-5, cacheReadMultiplier: 0.05 },
+  "claude-haiku-5-5": {
+    input: 1e-7,
+    output: 5e-7,
+    longPrompt: { aboveTokens: 1e5, card: { input: 5e-7, output: 25e-7 } }
+  },
   "claude-opus-5-5": { input: 4e-6, output: 2e-5, cacheReadMultiplier: 0.05 },
   "claude-fable-5-1": { input: 1e-5, output: 5e-5, cacheReadMultiplier: 0.025 },
   "claude-fable-5": { input: 1e-5, output: 5e-5 },
@@ -5115,14 +5181,16 @@ function estimateAnthropicCost(model, usage) {
     return void 0;
   }
   const cacheWriteTotal = safeTokenSum(cacheWrite5m, cacheWrite1h);
-  if (cacheWriteTotal === void 0 || safeTokenSum(usage.inputTokens, cacheRead, cacheWriteTotal) === void 0) {
+  const promptTokens = cacheWriteTotal === void 0 ? void 0 : safeTokenSum(usage.inputTokens, cacheRead, cacheWriteTotal);
+  if (cacheWriteTotal === void 0 || promptTokens === void 0) {
     return void 0;
   }
   if (usage.aggregateCacheWriteTokens !== void 0 && (!isValidTokenCount(usage.aggregateCacheWriteTokens) || usage.aggregateCacheWriteTokens !== cacheWriteTotal)) {
     return void 0;
   }
-  const { input, output } = rate;
-  const cacheReadRate = input * (rate.cacheReadMultiplier ?? DEFAULT_CACHE_READ_MULTIPLIER);
+  const card = rate.longPrompt && promptTokens > rate.longPrompt.aboveTokens ? rate.longPrompt.card : rate;
+  const { input, output } = card;
+  const cacheReadRate = input * (card.cacheReadMultiplier ?? DEFAULT_CACHE_READ_MULTIPLIER);
   return usage.inputTokens * input + cacheRead * cacheReadRate + cacheWrite5m * input * 1.25 + cacheWrite1h * input * 2 + usage.outputTokens * output;
 }
 var THINKING_BUDGETS = {
@@ -5184,7 +5252,7 @@ function resolveClaudeThinking(model, resolved) {
         };
       }
       return {
-        thinking: { type: "disabled" },
+        thinking: { type: capability.offShape ?? "disabled" },
         outputConfig: activeEffort ? { effort: activeEffort } : void 0,
         stripTemperature: true
       };
@@ -5590,50 +5658,106 @@ var AnthropicProvider = class {
       ...betas.length > 0 ? { "anthropic-beta": betas.join(",") } : {}
     };
   }
+  /**
+   * Whether the final body needs the binding beta header. Runs after
+   * `validateAndBindFinalBody`, which has already injected `block_binding`
+   * wherever it applies, so the header follows the final body rather than
+   * re-deriving the injection rule.
+   */
   requiresBindingBeta(body) {
+    const capability = resolveClaudeCapability(String(body.model));
     const thinking = body.thinking;
-    return resolveClaudeCapability(String(body.model))?.thinkingBinding === true && (hasReplayedThinking(body) || isRecord(thinking) && thinking.block_binding !== void 0);
+    return capability?.thinkingBinding === true && isRecord(thinking) && thinking.type === "adaptive" && thinking.block_binding !== void 0;
   }
+  /**
+   * Validate the final body (after `providerOptions`) against the effective
+   * model's capability, then inject `block_binding` where it applies. Every
+   * rule is gated on its own capability field and is a no-op when unset, so a
+   * model only gets the guards its provider contract calls for.
+   */
   validateAndBindFinalBody(body) {
     const model = typeof body.model === "string" ? body.model : "";
-    if (resolveClaudeCapability(model)?.thinkingBinding !== true) return;
+    const capability = resolveClaudeCapability(model);
+    if (!capability) return;
     const reject = (option, remediation) => {
       throw new UnsupportedModelOptionError({ provider: this.name, model, option, remediation });
     };
-    const toolChoice = body.tool_choice;
-    if (isRecord(toolChoice) && (toolChoice.type === "any" || toolChoice.type === "tool")) {
-      reject("forced tool choice", "Use toolChoice: 'auto' or 'none' on this model.");
-    }
-    if (body.thinking !== void 0 && (!isRecord(body.thinking) || body.thinking.type !== "adaptive" || body.thinking.budget_tokens !== void 0)) {
-      reject(
-        "manual or disabled thinking",
-        "Use thinking: { type: 'adaptive' } and output_config.effort."
-      );
-    }
-    const outputConfig = body.output_config;
-    if (outputConfig !== void 0 && (!isRecord(outputConfig) || outputConfig.effort !== void 0 && !["low", "medium", "high", "xhigh", "max"].includes(String(outputConfig.effort)))) {
-      reject("output_config.effort", "Use low, medium, high, xhigh, or max effort.");
-    }
-    if (body.temperature !== void 0 && body.temperature !== 1) {
-      reject("temperature", "Omit temperature or use the default value 1.");
-    }
-    if (body.top_p !== void 0 && (typeof body.top_p !== "number" || body.top_p < 0.99 || body.top_p > 1)) {
-      reject("top_p", "Omit top_p or use a value from 0.99 to 1.");
+    if (capability.forcedToolChoice === "rejected") {
+      const toolChoice = body.tool_choice;
+      if (isRecord(toolChoice) && (toolChoice.type === "any" || toolChoice.type === "tool")) {
+        reject("forced tool choice", "Use toolChoice: 'auto' or 'none' on this model.");
+      }
+    }
+    if (capability.strictThinking) {
+      const offShape = capability.thinking === "adaptive-always-on" ? void 0 : capability.offShape ?? "disabled";
+      const thinking = body.thinking;
+      const outputConfig = body.output_config;
+      const effort = isRecord(outputConfig) ? outputConfig.effort : void 0;
+      if (thinking !== void 0 && offShape === void 0) {
+        if (!isRecord(thinking) || thinking.type !== "adaptive" || thinking.budget_tokens !== void 0) {
+          reject(
+            "manual or disabled thinking",
+            "Use thinking: { type: 'adaptive' } and output_config.effort."
+          );
+        }
+      } else if (thinking !== void 0) {
+        if (!isRecord(thinking) || thinking.type !== "adaptive" && thinking.type !== offShape) {
+          reject(
+            "thinking.type",
+            `Use thinking: { type: 'adaptive' }, or { type: '${offShape}' } to turn thinking off.`
+          );
+        }
+        const record = thinking;
+        if (record.type === "adaptive" && record.budget_tokens !== void 0) {
+          reject("thinking.budget_tokens", "Use output_config.effort instead of budget_tokens.");
+        }
+        if (record.type === offShape) {
+          const extra = Object.keys(record).filter((key) => key !== "type");
+          if (extra.length > 0) {
+            reject(
+              `thinking.${offShape}.${extra[0]}`,
+              `Send thinking: { type: '${offShape}' } with no other fields.`
+            );
+          }
+        }
+      }
+      if (outputConfig !== void 0 && (!isRecord(outputConfig) || effort !== void 0 && !isNativeClaudeEffort(effort))) {
+        reject("output_config.effort", "Use low, medium, high, xhigh, or max effort.");
+      }
+      if (offShape !== void 0 && isRecord(thinking) && thinking.type === offShape && effort !== void 0 && capability.disableAt && !isEffortAtOrBelow(effort, capability.disableAt)) {
+        reject(
+          "output_config.effort",
+          `Thinking can be turned off only at effort '${capability.disableAt}' or below; lower the effort or use adaptive thinking.`
+        );
+      }
     }
-    if (body.top_k !== void 0) {
-      reject("top_k", "Omit top_k on this model.");
+    if (capability.defaultOnlySampling) {
+      if (body.temperature !== void 0 && body.temperature !== 1) {
+        reject("temperature", "Omit temperature or use the default value 1.");
+      }
+      if (body.top_p !== void 0 && body.top_p !== 0.99) {
+        reject("top_p", "Omit top_p or use the default value 0.99.");
+      }
+      if (body.temperature !== void 0 && body.top_p !== void 0) {
+        reject("temperature and top_p", "Send at most one of temperature and top_p.");
+      }
+      if (body.top_k !== void 0) {
+        reject("top_k", "Omit top_k on this model.");
+      }
     }
-    const replayed = hasReplayedThinking(body);
-    if (!replayed) return;
-    const thinking = isRecord(body.thinking) ? body.thinking : { type: "adaptive" };
-    const binding = thinking.block_binding;
-    if (binding !== void 0 && (!isRecord(binding) || binding.prefix_mismatch_behavior !== "error" && binding.prefix_mismatch_behavior !== "drop_block")) {
-      reject("thinking.block_binding", "Use prefix_mismatch_behavior: 'error' or 'drop_block'.");
+    if (capability.thinkingBinding === true && isFinalThinkingAdaptive(capability, body)) {
+      const thinking = isRecord(body.thinking) ? body.thinking : { type: "adaptive" };
+      const binding = thinking.block_binding;
+      if (binding !== void 0 && (!isRecord(binding) || binding.prefix_mismatch_behavior !== "error" && binding.prefix_mismatch_behavior !== "drop_block")) {
+        reject("thinking.block_binding", "Use prefix_mismatch_behavior: 'error' or 'drop_block'.");
+      }
+      if (hasReplayedThinking(body)) {
+        body.thinking = {
+          ...thinking,
+          block_binding: binding ?? { prefix_mismatch_behavior: "drop_block" }
+        };
+      }
     }
-    body.thinking = {
-      ...thinking,
-      block_binding: binding ?? { prefix_mismatch_behavior: "drop_block" }
-    };
   }
   /** Extract a human-readable message from an API error response body. */
   extractErrorMessage(body, status) {