fix: plugin effort compress

server/src/control/service.rs

@@ -392,7 +392,6 @@ impl ControlService {
         if model_hash.starts_with(crate::plugin::ADAPTER_ID_PREFIX) {
             let descriptor = self.plugins.model_descriptor(model_hash).await?;
             model.display_name = Some(descriptor.display_name);
-            model.context_window_tokens = descriptor.context_window_tokens;
             model.max_output_tokens = Some(descriptor.max_output_tokens.unwrap_or(65_536));
         } else {
             let configured = self

server/src/cursor/services/model_catalog.rs

@@ -580,14 +580,9 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
     let tooltip = TooltipData {
         markdown_content: model.description.clone(),
     };
-    let contexts = context_options(model.context_window_tokens);
-    let variants = model_variants(
-        &model.id,
-        &model.display_name,
-        &tooltip,
-        &contexts,
-        model.thinking,
-    );
+    // Effort 与上下文档位由宿主统一提供,与内置模型一致;插件不再声明这两项。
+    let contexts = context_options(None);
+    let variants = model_variants(&model.id, &model.display_name, &tooltip, &contexts, true);
     let legacy_slugs = variants
         .iter()
         .filter_map(|variant| variant.legacy_slug.clone())
@@ -598,7 +593,7 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
         supports_agent: Some(true),
         degradation_status: Some(0),
         tooltip_data: Some(tooltip.clone()),
-        supports_thinking: Some(model.thinking),
+        supports_thinking: Some(true),
         supports_images: Some(model.images),
         supports_max_mode: Some(false),
         client_display_name: Some(model.display_name.clone()),
@@ -610,7 +605,7 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
         inputbox_short_model_name: Some(model.display_name.clone()),
         supports_sandboxing: Some(true),
         supports_cmd_k: Some(false),
-        parameter_definitions: model_parameters(&contexts, model.thinking),
+        parameter_definitions: model_parameters(&contexts, true),
         variants,
         legacy_slugs,
         named_model_section_index: Some(1),
@@ -633,7 +628,7 @@ fn usable_plugin_model(model: &PluginModelDescriptor) -> agent::ModelDetails {
         display_model_id: model.id.clone(),
         display_name: model.display_name.clone(),
         display_name_short: model.display_name.clone(),
-        thinking_details: model.thinking.then(agent::ThinkingDetails::default),
+        thinking_details: Some(agent::ThinkingDetails::default()),
         ..Default::default()
     }
 }

server/src/plugin/descriptor.rs

@@ -107,9 +107,7 @@ pub struct PluginModelDescriptor {
     pub description: Option<String>,
     pub icon: String,
     pub provider_type: String,
-    pub context_window_tokens: Option<u64>,
     pub max_output_tokens: Option<u64>,
-    pub thinking: bool,
     pub images: bool,
 }
 
@@ -209,9 +207,7 @@ impl PluginModelDescriptor {
             description: model.description.clone(),
             icon: icon.to_owned(),
             provider_type: provider.provider_type.clone(),
-            context_window_tokens: model.context_window_tokens,
             max_output_tokens: model.max_output_tokens,
-            thinking: model.thinking,
             images: model.images,
         }
     }

server/src/plugin/sdk/model.ts

@@ -2,7 +2,6 @@ import type { JsonValue, PluginContext } from "./plugin.ts";
 import type { ResourceSnapshot } from "./resource.ts";
 
 export type ModelCapabilities = {
-  thinking?: boolean;
   images?: boolean;
 };
 
@@ -10,7 +9,6 @@ export type ModelDefinition = {
   id: string;
   displayName: string;
   description?: string;
-  contextWindowTokens?: number;
   maxOutputTokens?: number;
   capabilities?: ModelCapabilities;
   /** 之后的调用原样传回;永远不会展示给用户。 */

server/src/plugin/state.rs

@@ -135,12 +135,8 @@ pub struct StoredModel {
     #[serde(default)]
     pub description: Option<String>,
     #[serde(default)]
-    pub context_window_tokens: Option<u64>,
-    #[serde(default)]
     pub max_output_tokens: Option<u64>,
     #[serde(default)]
-    pub thinking: bool,
-    #[serde(default)]
     pub images: bool,
     #[serde(default)]
     pub private_data: serde_json::Value,
@@ -179,13 +175,9 @@ impl StoredModel {
                 .get("description")
                 .and_then(serde_json::Value::as_str)
                 .map(str::to_owned),
-            context_window_tokens: object
-                .get("contextWindowTokens")
-                .and_then(serde_json::Value::as_u64),
             max_output_tokens: object
                 .get("maxOutputTokens")
                 .and_then(serde_json::Value::as_u64),
-            thinking: capability("thinking"),
             images: capability("images"),
             private_data: object
                 .get("privateData")
@@ -200,9 +192,8 @@ impl StoredModel {
             "id": self.id,
             "displayName": self.display_name,
             "description": self.description,
-            "contextWindowTokens": self.context_window_tokens,
             "maxOutputTokens": self.max_output_tokens,
-            "capabilities": { "thinking": self.thinking, "images": self.images },
+            "capabilities": { "images": self.images },
             "privateData": self.private_data,
         })
     }
@@ -450,7 +441,7 @@ mod tests {
         let model = StoredModel::from_definition(&serde_json::json!({
             "id": "gpt-test",
             "displayName": "GPT Test",
-            "capabilities": {"thinking": true},
+            "capabilities": {"images": true},
             "privateData": {"reasoningEfforts": ["low"]},
         }))
         .unwrap();
@@ -460,7 +451,7 @@ mod tests {
             .unwrap();
         let models = store.models("dev.example", "codex").await.unwrap();
         assert_eq!(models.len(), 1);
-        assert!(models[0].thinking);
+        assert!(models[0].images);
         assert_eq!(models[0].private_data["reasoningEfforts"][0], "low");
     }
 }

server/src/provider/router.rs

@@ -67,9 +67,6 @@ impl Provider for ProviderRouter {
                     recorder.request(serde_json::json!({}), &crate::plugin::plugin_llm_request(&invocation)?).await?;
                     let mut routed = invocation.clone();
                     routed.request.model.display_name = Some(plan.model.display_name.clone());
-                    if let Some(tokens) = plan.model.context_window_tokens {
-                        routed.request.model.context_window_tokens.get_or_insert(tokens);
-                    }
                     if let Some(tokens) = plan.model.max_output_tokens {
                         routed.request.model.max_output_tokens.get_or_insert(tokens);
                     }

server/src/run/compaction.rs

@@ -2,9 +2,7 @@
 
 use std::collections::HashSet;
 
-use crate::model::{
-    CanonicalMessage, LlmCallUsageAnchor, PreparedRun, ProjectedMessage, RunAction,
-};
+use crate::model::{CanonicalMessage, LlmCallUsageAnchor, PreparedRun, ProjectedMessage};
 
 const FALLBACK_CHARS: usize = 12_000;
 
@@ -34,9 +32,6 @@ pub(super) fn should_compact(
     projected_messages: &[ProjectedMessage],
     anchor: Option<ContextUsageAnchor>,
 ) -> bool {
-    if prepared.action != RunAction::Start {
-        return false;
-    }
     let Some(context_window) = prepared.model.context_window_tokens else {
         return false;
     };
@@ -113,15 +108,15 @@ fn estimate_serialized_tokens(serialized: &str) -> u64 {
 mod tests {
     use super::*;
     use crate::model::{
-        project_messages, CheckpointId, ConversationId, ModelSpec, Origin, PromptSpec, Role, RunId,
-        RunKind,
+        project_messages, CheckpointId, ConversationId, ModelSpec, Origin, PromptSpec, Role,
+        RunAction, RunId, RunKind,
     };
 
     #[test]
-    fn automatic_compaction_starts_only_after_the_context_window_is_exceeded() {
+    fn automatic_compaction_runs_for_start_and_resume_actions_after_the_limit() {
         let mut model = ModelSpec::new("model");
         model.context_window_tokens = Some(200_000);
-        let prepared = PreparedRun {
+        let mut prepared = PreparedRun {
             run_id: RunId::new("run"),
             cursor_request_id: None,
             conversation_id: ConversationId::new("conversation"),
@@ -169,5 +164,15 @@ mod tests {
             &projected,
             anchor(200_001)
         ));
+
+        prepared.action = RunAction::Resume {
+            pending_tool_round: None,
+        };
+        assert!(should_compact(
+            &prepared,
+            &messages,
+            &projected,
+            anchor(200_001)
+        ));
     }
 }

server/src/run/engine.rs

@@ -170,7 +170,7 @@ impl RunEngine {
                 Ok(messages) => messages,
                 Err(error) => return (RunOutcome::Failed(error.into()), usage),
             };
-            let context_anchor = if !auto_compacted && prepared.action == RunAction::Start {
+            let context_anchor = if !auto_compacted {
                 match self
                     .store
                     .latest_llm_call_usage_anchor(