更新

server/config/console.php

@@ -4,7 +4,9 @@
 // +----------------------------------------------------------------------
 return [
     // 指令定义
-    'commands' => [
+    'commands' => [
+        'prescription-ai:work' => 'app\\command\\PrescriptionAiWork',
+        'prescription-ai:backfill' => 'app\\command\\PrescriptionAiBackfill',
         // 定时任务
         'crontab' => 'app\common\command\Crontab',
         // 退款查询

server/config/prescription_ai.php

@@ -22,15 +22,58 @@ return [
      * 单次请求可随附的附件总数上限。Dify 应用的 file_upload.number_limits 超限时
      * 直接返回 400 invalid_param 拒绝整单，患者纵向资料的附件数量又不可控，
      * 因此这里必须与上游应用配置保持一致（默认 3），超出的附件改以清单形式送达。
+     * 各应用的上限不同时用下面 models 里的 max_files 覆盖；核对方式是只读调用
+     * 该应用的 /parameters，读取 file_upload.number_limits。
      */
     'max_files' => (int) env(
         'prescription_ai.MAX_FILES',
         env('prescription_ai.max_files', 3)
     ),
+    // Per-model staged analysis limits. Oversized semantic units and exhausted call budgets
+    // are explicit task errors; they never silently remove patient evidence or attachments.
+    'manual_analysis' => [
+        /**
+         * 单次提示词的字节上限（UTF-8 字节是 token 的保守上界）。调大可减少分片与压缩
+         * 轮次、显著缩短一份报告的总耗时；上游应用限制更严时会返回可重试的拒绝，
+         * 此时调小本值。
+         */
+        'input_token_budget' => (int) env(
+            'prescription_ai.MANUAL_INPUT_TOKEN_BUDGET',
+            env('prescription_ai.manual_input_token_budget', 48000)
+        ),
+        'max_calls_per_model' => (int) env(
+            'prescription_ai.MANUAL_MAX_CALLS_PER_MODEL',
+            env('prescription_ai.manual_max_calls_per_model', 128)
+        ),
+        /**
+         * 医学研究对照模式：无论资料是否完整，两个模型都必须先各自独立开出候选处方，
+         * 再由服务端与人工方比较。缺口、假设与复核要求写入候选方的说明与风险提示，
+         * 不再以资料不足为由返回空方案。候选方仍不写回正式处方、审核或订单。
+         */
+        'require_candidate' => filter_var(
+            env('prescription_ai.MANUAL_REQUIRE_CANDIDATE', true),
+            FILTER_VALIDATE_BOOLEAN
+        ),
+        /**
+         * 后台分阶段分析的单次请求超时（秒，上限 300）。同步页面仍使用上面的 timeout；
+         * 该值必须明显小于任务租约 prescription_analysis.lease_seconds。
+         */
+        'request_timeout' => (int) env(
+            'prescription_ai.MANUAL_REQUEST_TIMEOUT',
+            240
+        ),
+        // 模型仍拒绝开方时，携带其拒绝理由重新追问的次数。
+        'candidate_insist_rounds' => (int) env(
+            'prescription_ai.MANUAL_CANDIDATE_INSIST_ROUNDS',
+            2
+        ),
+    ],
     'models' => [
         'qwen' => [
             'name' => 'qwen3.6-35b',
             'label' => '千问',
+            // 该应用 file_upload.number_limits = 3
+            'max_files' => (int) env('prescription_ai.QWEN_MAX_FILES', 3),
             'api_key' => (string) env(
                 'prescription_ai.QWEN_API_KEY',
                 env('prescription_ai.qwen_api_key', '')
@@ -39,6 +82,8 @@ return [
         'openai' => [
             'name' => 'gpt-5.6-sol',
             'label' => 'OpenAI',
+            // 该应用 file_upload.number_limits = 10：一次多带附件可显著减少往返次数
+            'max_files' => (int) env('prescription_ai.OPENAI_MAX_FILES', 10),
             'api_key' => (string) env(
                 'prescription_ai.OPENAI_API_KEY',
                 env('prescription_ai.openai_api_key', '')

server/config/prescription_analysis.php

@@ -0,0 +1,22 @@
+<?php
+
+/** Async clinician-facing analysis. Enable only after migration and workers are ready. */
+return [
+    'enabled' => filter_var(env('prescription_analysis.ENABLED', false), FILTER_VALIDATE_BOOLEAN),
+    'encryption_key' => (string) env('prescription_analysis.ENCRYPTION_KEY', ''),
+    'start_at' => (int) env('prescription_analysis.START_AT', 0),
+    'transcript_wait_seconds' => 300,
+    'debounce_seconds' => 60,
+    'lease_seconds' => 600,
+    'max_attempts' => 3,
+    'max_manual_retries' => 2,
+    'daily_model_tasks' => 200,
+    'daily_patient_batches' => 10,
+    // Concurrent tasks allowed per model. It only takes effect when that many consumer
+    // processes run for the lane (php think prescription-ai:work --lane=<model>).
+    'max_parallel_per_model' => 2,
+    'refresh_recent_days' => 7,
+    'auto_refresh_sources' => true,
+    'auto_refresh_prescription' => true,
+    'max_context_bytes' => 8000000,
+];