liuyu/nofx
1
0
Fork 0
mirror of https://github.com/NoFxAiOS/nofx.git synced 2026-07-16 01:06:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Split strategy config by strategy type

Browse Source
This commit is contained in:
lky-spec
2026-04-28 20:19:24 +08:00
parent 2d45e7ab15
commit 25d0b30ea9
11 changed files with 502 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
agent/central_brain.go
View File

@@ -605,6 +605,8 @@ Rules:
- For strategy_management:create: when the user asks you to design/recommend settings, think as the strategy designer, produce a concrete recommended config in your reply, and also put the same structured config into extracted_data.config_patch. Do not ask the user to fill fields you can reasonably choose for them.
- For strategy_management:create: once the structured config is sufficient to create, ask for one final confirmation and set extracted_data.awaiting_final_confirmation=true. Do not execute create in that same turn.
- For strategy_management:create: choose execute_skill only when awaiting_final_confirmation is already true and the current user message confirms the final summary. If the user changes a number, update config_patch and ask for final confirmation again.
- For strategy_management:create: if the previous assistant reply said the strategy was not actually created yet and that the next step is to call the structured create tool, then a user request to continue/proceed means execute the current skill when the structured config is ready. Do not answer with another promise such as "I will create it now"; choose execute_skill.
- For any mutating task, a reply that only promises future execution ("now I will create/update/start it", "result soon") is not a valid finish_task or ask_user outcome. If execution is the next step, choose execute_skill.
- Never choose finish_task for an unfinished mutating active task by claiming it was created/updated/deleted/started/stopped. Only a real skill/tool execution outcome can support that claim.
- If the user says they do not understand the current form, choices, or required information, choose "ask_user" and explain the current pending question in plain language before asking the next easiest question. Cover the relevant concepts from the previous assistant reply; do not collapse the answer to only the first missing field.
- For beginner/confusion replies, give a safe recommended path when the domain supports one, but do not execute or create anything unless the user confirms after the explanation.
@@ -616,7 +618,7 @@ Rules:
- For trader bindings, exchange/model/strategy must resolve to an ID from Relevant disclosed resources before execution. Never invent a resource name or use a generic venue type like Binance/OKX as the bound exchange unless it appears as an actual disclosed resource.
- For strategy_management:create, do not ask for exchange accounts or model bindings. Strategy templates are independent drafts/configs; exchange/model are only needed when creating, deploying, or starting a trader.
- Strategy templates should be visible in the strategy list/page after creation. Do not bring up trader/model/exchange binding unless the user asks to run or deploy.
- For strategy_management:create or strategy_management:update_config, when the user describes strategy intent, output config_patch as a partial StrategyConfig JSON object instead of leaving the default template unchanged. Example: "BTC趋势做空" should set coin_source to static BTCUSDT and add prompt/risk/entry rules for BTC trend-following short bias.
- For strategy_management:create or strategy_management:update_config, when the user describes strategy intent, output config_patch as a partial StrategyConfig JSON object instead of leaving the default template unchanged. The product schema is type-isolated: grid strategies use only top-level strategy_type + grid_config + publish_config; AI strategies use only top-level strategy_type + ai_config + publish_config. For example, "BTC趋势做空" as an AI strategy should set ai_config.coin_source to static BTCUSDT and add ai_config prompt/risk/entry rules.
- If there are multiple targets and the user did not disambiguate, ask a natural question with the available names.
- If the current user message answers a missing field directly, extract it and continue.
- extracted_data must use only canonical keys from Allowed field spec JSON. Never output aliases, translated labels, or raw user wording as keys.