@cognipeer/agent-sdk 0.20.0 → 0.22.0

package/dist/model-rules.mjs

@@ -18,6 +18,14 @@ var model_rules_default = {
     xhigh: 24576,
     max: 32768
   },
+  adaptiveMaxTokens: {
+    minimal: 4096,
+    low: 8192,
+    medium: 16384,
+    high: 32768,
+    xhigh: 64e3,
+    max: 64e3
+  },
   geminiBudgets: {
     minimal: 512,
     low: 1024,
@@ -304,7 +312,8 @@ var model_rules_default = {
           "xhigh",
           "max"
         ],
-        off: "disabled"
+        off: "disabled",
+        defaultOn: true
       },
       sampling: {
         temperature: "unsupported",
@@ -318,7 +327,7 @@ var model_rules_default = {
         reasoning: "any"
       },
       source: "verified-2026-10-14",
-      note: "Claude Haiku 5.x: thinking:{type:adaptive} + output_config.effort; thinking:{type:enabled,budget_tokens} is rejected. Off = thinking:{type:disabled}. Sampling params are rejected."
+      note: "Claude Haiku 5.x: thinking:{type:adaptive} + output_config.effort; thinking:{type:enabled,budget_tokens} is rejected. Off = thinking:{type:disabled}. Sampling params are rejected. Thinks by default with display omitted (empty text): includeThoughts alone sends adaptive + display summarized."
     },
     {
       id: "claude-sonnet-5",
@@ -333,7 +342,8 @@ var model_rules_default = {
           "xhigh",
           "max"
         ],
-        off: "between_tools"
+        off: "between_tools",
+        defaultOn: true
       },
       sampling: {
         temperature: "unsupported",
@@ -347,7 +357,7 @@ var model_rules_default = {
         reasoning: "any"
       },
       source: "verified-2026-10-14",
-      note: "Claude Sonnet 5.x: adaptive + output_config.effort. thinking:{type:disabled} is rejected; off = thinking:{type:between_tools}. Sampling params are rejected."
+      note: "Claude Sonnet 5.x: adaptive + output_config.effort. thinking:{type:disabled} is rejected; off = thinking:{type:between_tools}. Sampling params are rejected. Thinks by default with display omitted (empty text): includeThoughts alone sends adaptive + display summarized."
     },
     {
       id: "claude-opus-5",
@@ -362,7 +372,8 @@ var model_rules_default = {
           "xhigh",
           "max"
         ],
-        off: null
+        off: null,
+        defaultOn: true
       },
       sampling: {
         temperature: "unsupported",
@@ -376,12 +387,12 @@ var model_rules_default = {
         reasoning: "any"
       },
       source: "verified-2026-10-14",
-      note: "Claude Opus 5.x: adaptive + output_config.effort; thinking cannot be turned off (none maps to the lowest effort). Sampling params are rejected."
+      note: "Claude Opus 5.x: adaptive + output_config.effort; thinking cannot be turned off (none maps to the lowest effort). Sampling params are rejected. Thinks by default with display omitted (empty text): includeThoughts alone sends adaptive + display summarized."
     },
     {
       id: "claude-fable",
       family: "anthropic",
-      match: "^claude-fable",
+      match: "^claude-(?:fable|mythos-5)",
       reasoning: {
         style: "adaptive",
         efforts: [
@@ -391,7 +402,8 @@ var model_rules_default = {
           "xhigh",
           "max"
         ],
-        off: null
+        off: null,
+        defaultOn: true
       },
       sampling: {
         temperature: "unsupported",
@@ -405,7 +417,37 @@ var model_rules_default = {
         reasoning: "any"
       },
       source: "docs",
-      note: "Claude Fable line: treated like the 5.x adaptive models."
+      note: "Claude Fable and Mythos 5 lines (5, 5.1): thinking is always on (disabled and budget_tokens are 400s), so treated like the 5.x adaptive models: effort low..max, sampling params rejected, display omitted by default (includeThoughts alone sends adaptive + display summarized)."
+    },
+    {
+      id: "claude-mythos-preview",
+      family: "anthropic",
+      match: "^claude-mythos-preview",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "thinking-requires-1",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Claude Mythos Preview: the one 4.7+ model that still takes extended thinking (budget_tokens). Thinking cannot be disabled (type disabled is a 400) and defaults to adaptive when a request says nothing, so effort none sends no thinking field at all."
     },
     {
       id: "claude-4-7",
@@ -526,6 +568,223 @@ var model_rules_default = {
       source: "docs",
       note: "Permissive default for unknown Anthropic-wire models: budget style (the long-standing form); the provider flips the thinking form when the API says so."
     },
+    {
+      id: "gemini-3-pro-image",
+      family: "gemini",
+      match: "^gemini-3(?:\\.\\d+)?-pro-image",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3 Pro Image accepts only thinkingLevel high. thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-3-image",
+      family: "gemini",
+      match: "^gemini-3(?:\\.\\d+)?-.*image",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "minimal",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3.x Flash / Flash-Lite image models accept thinkingLevel minimal or high. thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-3-pro",
+      family: "gemini",
+      match: "^gemini-3(?:\\.0)?-pro",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "low",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3 Pro accepts thinkingLevel low or high (no minimal, no medium; default high). thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-3x-pro",
+      family: "gemini",
+      match: "^gemini-3\\.\\d+-pro",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3.x Pro (3.1 and later) accepts thinkingLevel low, medium or high (no minimal; default high). thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-3x-flash-no-minimal",
+      family: "gemini",
+      match: "^gemini-3\\.[78]-flash",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3.7 / 3.8 Flash accept thinkingLevel low, medium or high (no minimal). thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-3-level",
+      family: "gemini",
+      match: "^gemini-3",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "minimal",
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 3 Flash / Flash-Lite and other gemini-3 models accept thinkingLevel minimal, low, medium or high. thinkingLevel replaces thinkingBudget (a request with both is a 400); thinking cannot be switched off, so effort none maps to the lowest level the model accepts."
+    },
+    {
+      id: "gemini-4-level",
+      family: "gemini",
+      match: "^gemini-[4-9]",
+      reasoning: {
+        style: "level",
+        efforts: [
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 4 and later accept thinkingLevel low, medium or high; MINIMAL is not supported on Gemini 4 models. Setting thinkingBudget on Gemini 4 returns a 400 INVALID_ARGUMENT, so no budget is ever derived (a numeric budgetTokens picks the nearest level; only an explicit providerExtras.thinkingBudget sends one); thinking cannot be switched off, so effort none maps to the lowest level."
+    },
+    {
+      id: "gemini-2.5-pro",
+      family: "gemini",
+      match: "^gemini-2\\.5-pro",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: null,
+        minBudget: 128
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 2.5 Pro: thinkingBudget 128-32768 and thinking cannot be turned off, so effort none maps to minBudget (128) instead of 0."
+    },
     {
       id: "gemini-budget",
       family: "gemini",
@@ -682,6 +941,19 @@ var learned = /* @__PURE__ */ new Map();
 function learnedKey(provider, model, scope) {
   return `${(provider ?? "").toLowerCase()}@${scope ?? ""}::${(model ?? "").trim().toLowerCase()}`;
 }
+var BEDROCK_CLAUDE = /(?:^|[./])anthropic\.claude/;
+function isClaudeModelId(model) {
+  const id = String(model ?? "").trim().toLowerCase();
+  if (!id) return false;
+  return BEDROCK_CLAUDE.test(id) || modelIdCandidates(id).some((c) => c.startsWith("claude"));
+}
+function bedrockModelFamily(model) {
+  const id = String(model ?? "").trim().toLowerCase();
+  if (!id) return "unknown";
+  if (isClaudeModelId(id)) return "claude";
+  if (id.startsWith("arn:") && !(id.split("/").pop() ?? "").includes(".")) return "unknown";
+  return "other";
+}
 function resolveModelRule(opts) {
   const families = familiesFor(opts.provider);
   const ids = modelIdCandidates(opts.model);
@@ -693,7 +965,8 @@ function resolveModelRule(opts) {
   let isDefault = false;
   if (!rule) {
     isDefault = true;
-    rule = matchRule(families[0], ids, true) ?? MODEL_RULES.rules[MODEL_RULES.rules.length - 1];
+    const notClaude = (opts.provider ?? "").toLowerCase() === "bedrock" && bedrockModelFamily(opts.model) === "other";
+    rule = matchRule(notClaude ? "other" : families[0], ids, true) ?? MODEL_RULES.rules[MODEL_RULES.rules.length - 1];
   }
   let effective = rule;
   const known = opts.model ? learned.get(learnedKey(opts.provider, opts.model, opts.scope))?.patch : void 0;