@cognipeer/agent-sdk 0.20.0 → 0.22.1

package/dist/model-rules.json

@@ -17,6 +17,14 @@
     "xhigh": 24576,
     "max": 32768
   },
+  "adaptiveMaxTokens": {
+    "minimal": 4096,
+    "low": 8192,
+    "medium": 16384,
+    "high": 32768,
+    "xhigh": 64000,
+    "max": 64000
+  },
   "geminiBudgets": {
     "minimal": 512,
     "low": 1024,
@@ -303,7 +311,8 @@
           "xhigh",
           "max"
         ],
-        "off": "disabled"
+        "off": "disabled",
+        "defaultOn": true
       },
       "sampling": {
         "temperature": "unsupported",
@@ -317,7 +326,7 @@
         "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",
@@ -332,7 +341,8 @@
           "xhigh",
           "max"
         ],
-        "off": "between_tools"
+        "off": "between_tools",
+        "defaultOn": true
       },
       "sampling": {
         "temperature": "unsupported",
@@ -346,7 +356,7 @@
         "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",
@@ -361,7 +371,8 @@
           "xhigh",
           "max"
         ],
-        "off": null
+        "off": null,
+        "defaultOn": true
       },
       "sampling": {
         "temperature": "unsupported",
@@ -375,12 +386,12 @@
         "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": [
@@ -390,7 +401,8 @@
           "xhigh",
           "max"
         ],
-        "off": null
+        "off": null,
+        "defaultOn": true
       },
       "sampling": {
         "temperature": "unsupported",
@@ -404,7 +416,37 @@
         "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",
@@ -525,6 +567,223 @@
       "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",