Load-balance weights and distribution by speed and reliability, slow-start switch, upstream concurrency, model specs, key usage limits, per-turn WebSocket requests

bin/twcore/src/main.rs

@@ -820,10 +820,16 @@ fn cmd_serve(path: &Path, port: Option<u16>, safe: bool, parent: Option<u32>) ->
             state.pricing.clone(),
             body_rx,
             run_rx,
+            tw_control::key_limits::settle_hook(&state),
         );
-        if store.is_some() {
+        if let Some(rec) = &store {
             state.set_body_sink(body_tx);
             state.set_plugin_sink(run_tx);
+            // 密钥的用量上限：这一天、这一周、这个月已经用了多少，从记录里加回来。**在第一个
+            // 请求之前**（网关还没开始听）
+            tw_control::key_limits::rebuild(&state, rec.lock().await.db());
+            // 之后一期的开头变了（到了下一期、换了时区），从记录里把新的那一期加起来
+            tw_control::key_limits::follow(&state, rec.clone());
         }
 
         /*
@@ -1004,6 +1010,8 @@ fn build_store(
     pricing: tw_pricing::Shared,
     bodies: tokio::sync::mpsc::Receiver<tw_gateway::bodies::BodyRecord>,
     runs: tokio::sync::mpsc::Receiver<tw_gateway::plugin::RunRecord>,
+    // 每记下一行请求，交给网关的密钥用量上限结算（见 `tw_control::key_limits`）
+    settled: tw_store::SettleHook,
 ) -> Option<std::sync::Arc<tokio::sync::Mutex<tw_store::Recorder>>> {
     let events = bus.subscribe();
     let (db, blobs) = match tw_store::open(dir) {
@@ -1078,7 +1086,9 @@ fn build_store(
     let recorder = tw_store::task::spawn(
         // 算完价钱往回报一条 —— 见 `Event::RequestPriced`。这里是唯一
         // 同时看得见总线和存储层的地方，所以接线在这儿完成。
-        tw_store::Recorder::new(db, blobs, pricing).reporting_to(bus),
+        tw_store::Recorder::new(db, blobs, pricing)
+            .reporting_to(bus)
+            .settling_to(settled),
         events,
         rx,
     );

crates/tw-api/msg-codes.txt

@@ -67,12 +67,24 @@ config.edit.unwritable
 config.empty_key
 config.empty_models_only
 config.failover_range
+config.key_limit_cache_reads
+config.key_limit_cost_too_small
+config.key_limit_duplicate
+config.key_limit_empty
+config.key_limit_not_positive
+config.key_limit_retention
+config.key_limit_two_measures
+config.model_spec_blank_model
+config.model_spec_empty
+config.model_spec_wildcard
+config.model_spec_zero
 config.name_collision
 config.no_clients
 config.plugin.bad_id
 config.plugin.duplicate
 config.plugin.file
 config.plugin.sha256
+config.provider_concurrency_range
 config.rejected
 config.rejected_at
 config.remote_port_is_gateway
@@ -90,6 +102,7 @@ config.schema_too_new
 config.secret.empty_name
 config.secret.env_missing
 config.secret.unterminated
+config.slow_start_too_short
 config.store.conflict
 config.store.missing
 config.store.read_failed
@@ -126,11 +139,15 @@ control.device_code_unavailable
 control.device_code_unreadable
 control.disabled_key_cannot_be_default
 control.dryrun_needs_target
+control.group.balance_not_load_balance
 control.group.empty
 control.group.name_is_upstream
 control.group.no_such_upstream
 control.group.preferred_not_member
 control.group.upstream_twice
+control.group.weight_not_load_balance
+control.group.weight_not_member
+control.group.weight_out_of_range
 control.group_in_use
 control.header_no_value
 control.internal_error
@@ -228,8 +245,11 @@ engine.compare.no_operator
 engine.duplicate_group
 engine.duplicate_route
 engine.empty_group
+engine.group_balance_not_load_balance
 engine.group_unknown_upstream
 engine.group_upstream_twice
+engine.group_weight_not_load_balance
+engine.group_weight_out_of_range
 engine.no_action
 engine.no_match
 engine.phase_two_with_to
@@ -247,6 +267,8 @@ gw.auth.key_invalid
 gw.auth.no_key
 gw.auth.source_not_allowed
 gw.auth.source_not_allowed_hint
+gw.busy_all
+gw.busy_upstream
 gw.chatgpt.token_missing
 gw.chatgpt.token_not_json
 gw.chatgpt.token_status
@@ -266,6 +288,12 @@ gw.convert.tools_unsendable
 gw.count_tokens.bedrock_upstream
 gw.files.unsupported
 gw.internal
+gw.key_limit.cost_per_period
+gw.key_limit.cost_rolling
+gw.key_limit.requests_per_period
+gw.key_limit.requests_rolling
+gw.key_limit.tokens_per_period
+gw.key_limit.tokens_rolling
 gw.listen.addr_unavailable
 gw.listen.bind_failed
 gw.listen.denied
@@ -348,6 +376,7 @@ gw.route.protocol_mismatch
 gw.route.rule_failed
 gw.route.selected_upstream_missing
 gw.route.upstream_missing
+gw.slow_start
 gw.toolcall.connection_cut
 gw.toolcall.response_cut
 gw.toolcall.response_withheld
@@ -374,6 +403,7 @@ gw.ws.connect_failed
 gw.ws.proxy_unsupported
 gw.ws.send_failed
 gw.ws.upstream_broke
+gw.ws.upstream_closed
 l1.config.bad_url
 l1.config.no_host
 l1.config.proxy_addr_form

crates/tw-api/src/ep.rs

@@ -95,6 +95,9 @@ endpoints! {
     UpdateProvider: PUT "/providers/{name}" [name], api::ProviderSave => api::ConfigWritten;
     DeleteProvider: DELETE "/providers/{name}" [name], api::BaseVersion => api::ConfigWritten;
     ProviderModels: GET "/providers/{name}/models" [name], () => api::ProviderModelsView;
+    /// 手写一家上游的一个模型的上下文窗口、输出上限，优先于价目表；两项都空就删掉。
+    /// **不在 `/providers/` 底下**：写死的一段会盖住 `/providers/{name}`
+    SetModelSpec: PUT "/provider-model-spec", api::ModelSpecSave => api::ConfigWritten;
     RefreshProviderModels: POST "/providers/{name}/models/refresh" [name], () => api::ProviderModelsView;
     RefreshStaleModels: POST "/models/refresh", () => api::ModelsRefreshing;
     CreateProxy: POST "/proxies", api::ProxySave => api::ConfigWritten;