bizrouter 0.3.0 → 0.3.1

package/AGENT_MANUAL.md

@@ -228,5 +228,7 @@ user; when `calls` is empty, say the answer came from the model alone.
 - Model calls return 400 `model not found` → the key's allowed-model list or
   the organization's model policy excludes it (`bizrouter models` shows what the
   key may use).
-- Hermes + model with output cap under 65536 → pick another model
+- Hermes + model with output cap under 65536 → the gateway lowers Hermes'
+  fixed `max_tokens` to the cap, so the run works but one answer cannot exceed
+  the cap; pick a model with a higher cap only when long answers matter
   (`bizrouter hermes --model anthropic/claude-sonnet-5`).

package/dist/commands/misc.js

@@ -2,7 +2,7 @@ import { spawn } from 'node:child_process';
 import { ENV_KEY_NAME } from '../config.js';
 import { c, info, print } from '../ui.js';
 import { requireApiKey } from './shared.js';
-export const VERSION = '0.3.0';
+export const VERSION = '0.3.1';
 /** `eval "$(bizrouter env)"` exports the saved key for tools configured by `bizrouter setup`. */
 export function envCommand() {
     const key = requireApiKey();

package/dist/harness/hermes.js

@@ -1,14 +1,15 @@
 /** Hermes sends this max_tokens to custom OpenAI-compatible endpoints regardless of the model. */
 export const HERMES_FIXED_MAX_TOKENS = 65536;
 /**
- * Models whose output cap is below Hermes' fixed max_tokens get rejected by the
- * gateway (`max_tokens: 65536 > 64000 …`) and Hermes misreads that as a context
- * overflow, so warn before launching.
+ * The gateway lowers an over-cap max_tokens to the model's output cap instead
+ * of rejecting the call, so a model whose cap is below Hermes' fixed value
+ * still works; its answers just cannot run past the cap. Say so before
+ * launching so a truncated long answer is not mistaken for a gateway fault.
  */
 export function hermesOutputCapWarning(model, maxOutputTokens) {
     if (maxOutputTokens === undefined || maxOutputTokens >= HERMES_FIXED_MAX_TOKENS)
         return undefined;
-    return `Hermes는 max_tokens ${HERMES_FIXED_MAX_TOKENS.toLocaleString('en-US')}을 고정으로 보내는데 ${model}의 출력 한도는 ${maxOutputTokens.toLocaleString('en-US')} 입니다. 요청이 400으로 거절되고 Hermes가 이를 컨텍스트 초과로 오해합니다. 출력 한도가 65,536 이상인 모델(예: anthropic/claude-sonnet-5)을 쓰세요.`;
+    return `Hermes는 max_tokens ${HERMES_FIXED_MAX_TOKENS.toLocaleString('en-US')}을 고정으로 보내는데 ${model}의 출력 한도는 ${maxOutputTokens.toLocaleString('en-US')} 입니다. BizRouter가 요청을 ${maxOutputTokens.toLocaleString('en-US')}으로 잘라 보내므로 실행은 되지만, 한 번의 답은 그 길이를 넘지 못합니다.`;
 }
 const HERMES_SUBCOMMANDS = new Set(['chat']);
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bizrouter",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "description": "BizRouter CLI - run Claude Code, Codex, OpenCode, and Hermes through BizRouter, operate the BizRouter console (API keys, usage, policies, audit logs) from the terminal or as an MCP server, and ask BizRouter Search for answers grounded in 50+ live data sources",
   "license": "UNLICENSED",
   "type": "module",