refactor: remove configurable toolcall policy and fix to feature matching with high-confidence early emit

This commit is contained in:
CJACK
2026-03-30 01:56:25 +08:00
parent 822b14ed6b
commit a7c9dfd7c0
20 changed files with 38 additions and 179 deletions

View File

@@ -270,10 +270,6 @@ cp opencode.json.example opencode.json
"compat": {
"wide_input_strict_output": true
},
"toolcall": {
"mode": "feature_match",
"early_emit_confidence": "high"
},
"responses": {
"store_ttl_seconds": 900
},
@@ -304,7 +300,7 @@ cp opencode.json.example opencode.json
- `token`:配置文件中即使填写也会在加载时被清空(不会从 `config.json` 读取 token实际 token 仅在运行时内存中维护并自动刷新
- `model_aliases`:常见模型名(如 GPT/Codex/Claude到 DeepSeek 模型的映射
- `compat.wide_input_strict_output`:建议保持 `true`(当前实现默认宽进严出)
- `toolcall`固定采用特征匹配 + 高置信早发策略
- `toolcall`策略已固定为特征匹配 + 高置信早发,不再作为可配置项
- `responses.store_ttl_seconds``/v1/responses/{id}` 的内存缓存 TTL
- `embeddings.provider`embedding 提供方(当前内置 `deterministic/mock/builtin`
- `claude_mapping`:字典中 `fast`/`slow` 后缀映射到对应 DeepSeek 模型(兼容读取 `claude_model_mapping`