@cognipeer/agent-sdk 0.10.6 → 0.13.0

package/dist/model-rules.mjs

@@ -0,0 +1,756 @@
+// src/models/model-rules.json
+var model_rules_default = {
+  version: 1,
+  effortLadder: [
+    "none",
+    "minimal",
+    "low",
+    "medium",
+    "high",
+    "xhigh",
+    "max"
+  ],
+  budgets: {
+    minimal: 1024,
+    low: 2048,
+    medium: 8192,
+    high: 16384,
+    xhigh: 24576,
+    max: 32768
+  },
+  geminiBudgets: {
+    minimal: 512,
+    low: 1024,
+    medium: 4096,
+    high: 12e3,
+    xhigh: 24576,
+    max: 24576
+  },
+  rules: [
+    {
+      id: "openai-chat-models",
+      family: "openai",
+      match: "^(?:chatgpt|gpt-[\\d.]+[a-z0-9.-]*?-chat)",
+      reasoning: {
+        style: "none",
+        efforts: [],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_completion_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Chat-tuned *-chat variants are not reasoning models: ordinary sampling, no reasoning field."
+    },
+    {
+      id: "openai-gpt-6-plus",
+      family: "openai",
+      match: "^gpt-?(?:[6-9]|(?!35(?!\\d))[1-9]\\d+)(?:[.-]|$)",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "none",
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        off: "none",
+        responsesEfforts: [
+          "none",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ]
+      },
+      sampling: {
+        temperature: "default-only",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_completion_tokens",
+      chatTools: {
+        reasoning: "force-none"
+      },
+      source: "verified-2026-10-14",
+      note: "gpt-N, N>=6. Chat Completions takes none/low/medium/high/xhigh (minimal and max are 400s); Responses also takes max. Function tools with an effort other than none are a 400 on Chat Completions and the default effort is not none, so none is sent explicitly. Rejects max_tokens, temperature != 1, top_p, penalties, logprobs."
+    },
+    {
+      id: "openai-gpt-5x",
+      family: "openai",
+      match: "^gpt-?5\\.[1-9]",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "none",
+          "low",
+          "medium",
+          "high",
+          "xhigh"
+        ],
+        off: "none",
+        responsesEfforts: [
+          "none",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ]
+      },
+      sampling: {
+        temperature: "default-only",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_completion_tokens",
+      chatTools: {
+        reasoning: "force-none"
+      },
+      source: "verified-2026-10-14",
+      note: "gpt-5.1 and newer 5.x. Same wire rules as gpt-6: no minimal/max on Chat Completions (Responses takes max), function tools need reasoning_effort none on Chat Completions."
+    },
+    {
+      id: "openai-gpt-5-base",
+      family: "openai",
+      match: "^gpt-?5(?![.\\d])",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "minimal",
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "default-only",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_completion_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "gpt-5, -mini, -nano, -codex: minimal/low/medium/high, reasoning cannot be turned off (none maps to minimal)."
+    },
+    {
+      id: "openai-o-series",
+      family: "openai",
+      match: "^o[1-9]\\d?(?![a-z0-9])",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "low",
+          "medium",
+          "high"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "default-only",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_completion_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "o1/o3/o4 family: low/medium/high, no sampling overrides."
+    },
+    {
+      id: "openai-gpt-oss",
+      family: "openai",
+      match: "^gpt-?oss",
+      reasoning: {
+        style: "effort",
+        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: "Open-weight gpt-oss: low/medium/high; sampling and max_tokens behave like an ordinary chat model."
+    },
+    {
+      id: "openai-legacy-gpt",
+      family: "openai",
+      match: "^(?:gpt-?(?:[0-4]|35)(?![0-9])|text-|davinci|babbage)",
+      reasoning: {
+        style: "none",
+        efforts: [],
+        off: null
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "gpt-4o, gpt-4.1, gpt-3.5 and older: no reasoning, ordinary sampling."
+    },
+    {
+      id: "openai-compatible-toggle",
+      family: "openai",
+      match: "^(?:deepseek-(?:r\\d|reasoner)|qwq|qwen-?3)",
+      reasoning: {
+        style: "toggle",
+        efforts: [
+          "none",
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "none"
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "DeepSeek R/reasoner, QwQ, Qwen3: on/off style hosts. The effort is forwarded as-is (including none, which switches thinking off on thinking-by-default models)."
+    },
+    {
+      id: "openai-default",
+      family: "openai",
+      match: ".*",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "none",
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "none"
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Permissive default for unknown OpenAI-wire models: pass what the caller asked; the provider learns from 400s."
+    },
+    {
+      id: "claude-haiku-5",
+      family: "anthropic",
+      match: "^claude-(?:haiku-5|5(?:-\\d+)?-haiku)(?!\\d)",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "disabled"
+      },
+      sampling: {
+        temperature: "unsupported",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        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."
+    },
+    {
+      id: "claude-sonnet-5",
+      family: "anthropic",
+      match: "^claude-(?:sonnet-5|5(?:-\\d+)?-sonnet)(?!\\d)",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "between_tools"
+      },
+      sampling: {
+        temperature: "unsupported",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        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."
+    },
+    {
+      id: "claude-opus-5",
+      family: "anthropic",
+      match: "^claude-(?:opus-5|5(?:-\\d+)?-opus)(?!\\d)",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "unsupported",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        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."
+    },
+    {
+      id: "claude-fable",
+      family: "anthropic",
+      match: "^claude-fable",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: null
+      },
+      sampling: {
+        temperature: "unsupported",
+        topP: false,
+        topK: false,
+        penalties: false,
+        logprobs: false
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Claude Fable line: treated like the 5.x adaptive models."
+    },
+    {
+      id: "claude-4-7",
+      family: "anthropic",
+      match: "^claude-(?:(?:opus|sonnet|haiku)-4[-.]7|4[-.]7-(?:opus|sonnet|haiku))(?!\\d)",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "disabled"
+      },
+      sampling: {
+        temperature: "thinking-requires-1",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Claude 4.7: adaptive thinking, low/medium/high/xhigh/max."
+    },
+    {
+      id: "claude-4-6",
+      family: "anthropic",
+      match: "^claude-(?:(?:opus|sonnet|haiku)-4[-.]6|4[-.]6-(?:opus|sonnet|haiku))(?!\\d)",
+      reasoning: {
+        style: "adaptive",
+        efforts: [
+          "low",
+          "medium",
+          "high",
+          "max"
+        ],
+        off: "disabled"
+      },
+      sampling: {
+        temperature: "thinking-requires-1",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Claude 4.6: adaptive thinking, low/medium/high/max (no xhigh)."
+    },
+    {
+      id: "claude-legacy-budget",
+      family: "anthropic",
+      match: "^claude-(?:3|4-|(?:opus|sonnet|haiku)-4)",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "none",
+          "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 3.7 and 4.0-4.5: thinking:{type:enabled,budget_tokens}; every effort maps to a budget; sampling overrides are not allowed while thinking."
+    },
+    {
+      id: "anthropic-default",
+      family: "anthropic",
+      match: ".*",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "none",
+          "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: "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-budget",
+      family: "gemini",
+      match: "^gemini",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "none",
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "none"
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Gemini 2.5-style thinkingBudget; effort maps to a token budget, none = budget 0."
+    },
+    {
+      id: "gemini-default",
+      family: "gemini",
+      match: ".*",
+      reasoning: {
+        style: "budget",
+        efforts: [
+          "none",
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "none"
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Permissive default for unknown Gemini-wire models."
+    },
+    {
+      id: "other-default",
+      family: "other",
+      match: ".*",
+      reasoning: {
+        style: "effort",
+        efforts: [
+          "none",
+          "minimal",
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        off: "none"
+      },
+      sampling: {
+        temperature: "any",
+        topP: true,
+        topK: true,
+        penalties: true,
+        logprobs: true
+      },
+      maxTokensParam: "max_tokens",
+      chatTools: {
+        reasoning: "any"
+      },
+      source: "docs",
+      note: "Permissive default for model families the SDK has no rules for."
+    }
+  ]
+};
+
+// src/models/rules.ts
+var MODEL_RULES = model_rules_default;
+var LADDER = MODEL_RULES.effortLadder;
+var compiled = /* @__PURE__ */ new Map();
+function regexFor(rule) {
+  let re = compiled.get(rule.id);
+  if (!re) {
+    re = new RegExp(rule.match, "i");
+    compiled.set(rule.id, re);
+  }
+  return re;
+}
+var GATEWAY_PREFIX = /^(?:(?:global|us|eu|apac|jp|au|ca|us-gov)\.)?(?:(?:openai|anthropic|google|meta|amazon|deepseek|qwen|mistral|cohere|ai21)\.)?/;
+function modelIdCandidates(model) {
+  if (!model) return [];
+  const lower = String(model).trim().toLowerCase();
+  const last = lower.split("/").pop() ?? lower;
+  const out = [];
+  for (const id of [lower, last]) {
+    for (const c of [id, id.replace(GATEWAY_PREFIX, "")]) {
+      if (c && !out.includes(c)) out.push(c);
+    }
+  }
+  return out;
+}
+var OPENAI_WIRE = /* @__PURE__ */ new Set(["openai", "azure", "openai-compatible", "together", "openrouter", "groq", "fireworks", "ollama", "vllm", "litellm", "mistral", "deepseek", "xai", "jev"]);
+function familiesFor(provider) {
+  const p = (provider ?? "").toLowerCase();
+  if (OPENAI_WIRE.has(p)) return ["openai"];
+  if (p === "anthropic") return ["anthropic"];
+  if (p === "bedrock") return ["anthropic", "openai", "other"];
+  if (p === "vertex" || p === "gemini" || p === "google") return ["gemini", "anthropic"];
+  return ["openai", "anthropic", "gemini", "other"];
+}
+function matchRule(family, ids, allowDefault) {
+  for (const rule of MODEL_RULES.rules) {
+    if (rule.family !== family) continue;
+    if (rule.match === ".*") {
+      if (allowDefault) return rule;
+      continue;
+    }
+    const re = regexFor(rule);
+    if (ids.some((id) => re.test(id))) return rule;
+  }
+  return void 0;
+}
+function mergeRulePatch(base, patch) {
+  if (!patch) return base;
+  const out = { ...base };
+  if (patch.reasoning) out.reasoning = { ...base.reasoning ?? {}, ...patch.reasoning };
+  if (patch.sampling) out.sampling = { ...base.sampling ?? {}, ...patch.sampling };
+  if (patch.chatTools) out.chatTools = { ...base.chatTools ?? {}, ...patch.chatTools };
+  if (patch.maxTokensParam) out.maxTokensParam = patch.maxTokensParam;
+  if (patch.dropParams) out.dropParams = [.../* @__PURE__ */ new Set([...base.dropParams ?? [], ...patch.dropParams])];
+  return out;
+}
+var learned = /* @__PURE__ */ new Map();
+function learnedKey(provider, model, scope) {
+  return `${(provider ?? "").toLowerCase()}@${scope ?? ""}::${(model ?? "").trim().toLowerCase()}`;
+}
+function resolveModelRule(opts) {
+  const families = familiesFor(opts.provider);
+  const ids = modelIdCandidates(opts.model);
+  let rule;
+  for (const family of families) {
+    rule = matchRule(family, ids, false);
+    if (rule) break;
+  }
+  let isDefault = false;
+  if (!rule) {
+    isDefault = true;
+    rule = matchRule(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;
+  if (known) effective = mergeRulePatch(effective, known);
+  if (opts.patch) effective = mergeRulePatch(effective, opts.patch);
+  return { ...effective, matchedRuleId: rule.id, isDefault, learned: !!known };
+}
+function mapReasoningEffort(rule, effort, endpoint = "chat") {
+  const efforts = endpoint === "responses" && rule.reasoning.responsesEfforts || rule.reasoning.efforts;
+  if (rule.reasoning.style === "none" || efforts.length === 0) return { mapped: true };
+  const supported = LADDER.filter((e) => efforts.includes(e));
+  if (supported.includes(effort)) return { effort, mapped: false };
+  if (effort === "none") {
+    const off = rule.reasoning.off;
+    if (off && off !== "none") return { off, mapped: true };
+    const lowest = supported.find((e) => e !== "none");
+    return lowest ? { effort: lowest, mapped: true } : { mapped: true };
+  }
+  const at = LADDER.indexOf(effort);
+  if (at < 0) return { mapped: true };
+  for (let i = at - 1; i >= 1; i--) if (supported.includes(LADDER[i])) return { effort: LADDER[i], mapped: true };
+  for (let i = at + 1; i < LADDER.length; i++) if (supported.includes(LADDER[i])) return { effort: LADDER[i], mapped: true };
+  return supported.includes("none") ? { effort: "none", mapped: true } : { mapped: true };
+}
+function allowsSampling(rule, field, value, thinkingOn = false) {
+  const s = rule.sampling;
+  switch (field) {
+    case "temperature":
+      switch (s.temperature) {
+        case "any":
+          return true;
+        case "unsupported":
+          return false;
+        // The default is what the server applies anyway, so it is never sent.
+        case "default-only":
+          return false;
+        case "thinking-requires-1":
+          return !thinkingOn || value === 1;
+      }
+      return true;
+    // Anthropic's thinking restriction covers top_p / top_k as well as temperature.
+    case "topP":
+      return s.topP && !(thinkingOn && s.temperature === "thinking-requires-1");
+    case "topK":
+      return s.topK && !(thinkingOn && s.temperature === "thinking-requires-1");
+    case "presencePenalty":
+    case "frequencyPenalty":
+      return s.penalties;
+    case "logprobs":
+      return s.logprobs;
+  }
+}
+function effortBudget(effort, table = "anthropic") {
+  if (effort === "none") return 0;
+  return (table === "gemini" ? MODEL_RULES.geminiBudgets : MODEL_RULES.budgets)[effort] ?? 0;
+}
+
+export { MODEL_RULES, allowsSampling, effortBudget, mapReasoningEffort, modelIdCandidates, resolveModelRule };
+//# sourceMappingURL=model-rules.mjs.map
+//# sourceMappingURL=model-rules.mjs.map