feat(agent): strip trailing assistant prefill for proxy provider compat

.gitignore

@@ -46,3 +46,6 @@ k8s-*/*
 tests/**/creds.json
 ui/simple-saas
 *.tar
+# AI tool config directories (user-specific)
+.gemini/
+.opencode/

internal/agent/loop_types.go

@@ -119,6 +119,9 @@ type Loop struct {
 	skillEvolve        bool
 	skillNudgeInterval int // nudge every N tool calls (0 = disabled, 15 = default)
 
+	// Strip trailing assistant messages before LLM call (for proxy providers)
+	stripAssistantPrefill bool
+
 	// Config permission store for group file writer checks
 	configPermStore store.ConfigPermissionStore
 
@@ -229,6 +232,9 @@ type LoopConfig struct {
 	// Thinking level: "off", "low", "medium", "high" (from agent other_config)
 	ThinkingLevel string
 
+	// Strip trailing assistant messages before LLM call (for proxy providers)
+	StripAssistantPrefill bool
+
 	// Self-evolve: predefined agents can update SOUL.md (style/tone) through chat
 	SelfEvolve bool
 
@@ -335,6 +341,7 @@ func NewLoop(cfg LoopConfig) *Loop {
 		selfEvolve:             cfg.SelfEvolve,
 		skillEvolve:            cfg.SkillEvolve,
 		skillNudgeInterval:     cfg.SkillNudgeInterval,
+		stripAssistantPrefill:  cfg.StripAssistantPrefill,
 		configPermStore:        cfg.ConfigPermStore,
 		teamStore:              cfg.TeamStore,
 		secureCLIStore:         cfg.SecureCLIStore,

internal/agent/resolver.go

@@ -347,6 +347,7 @@ func NewManagedResolver(deps ResolverDeps) ResolverFunc {
 			SelfEvolve:             ag.ParseSelfEvolve(),
 			SkillEvolve:            ag.AgentType == store.AgentTypePredefined && ag.ParseSkillEvolve(),
 			SkillNudgeInterval:     ag.ParseSkillNudgeInterval(),
+			StripAssistantPrefill:  ag.ParseStripAssistantPrefill(),
 			WorkspaceSharing:       ag.ParseWorkspaceSharing(),
 			ShellDenyGroups:        ag.ParseShellDenyGroups(),
 			ConfigPermStore:        deps.ConfigPermStore,

internal/store/agent_store.go

@@ -204,6 +204,21 @@ func (a *AgentData) ParseSkillNudgeInterval() int {
 	return *cfg.SkillNudgeInterval
 }
 
+// ParseStripAssistantPrefill extracts strip_assistant_prefill from other_config JSONB.
+// When true, trailing assistant messages are removed before LLM call (for proxy providers).
+func (a *AgentData) ParseStripAssistantPrefill() bool {
+	if len(a.OtherConfig) == 0 {
+		return false
+	}
+	var cfg struct {
+		StripAssistantPrefill bool `json:"strip_assistant_prefill"`
+	}
+	if json.Unmarshal(a.OtherConfig, &cfg) != nil {
+		return false
+	}
+	return cfg.StripAssistantPrefill
+}
+
 // WorkspaceSharingConfig controls per-user workspace isolation.
 // When shared_dm/shared_group is true, users share the base workspace directory
 // instead of each getting an isolated subfolder.

ui/web/src/i18n/locales/en/agents.json

@@ -147,7 +147,9 @@
     "contextWindow": "Context Window",
     "contextWindowHint": "Token limit for the model context.",
     "maxToolIterations": "Max Tool Iterations",
-    "maxToolIterationsHint": "Max tool calls per request."
+    "maxToolIterationsHint": "Max tool calls per request.",
+    "stripAssistantPrefill": "Strip Assistant Prefill",
+    "stripAssistantPrefillHint": "Remove trailing assistant messages before LLM call. Enable for proxy providers that reject prefill."
   },
   "workspace": {
     "title": "Workspace",

ui/web/src/i18n/locales/vi/agents.json

@@ -147,7 +147,9 @@
     "contextWindow": "Cửa sổ ngữ cảnh",
     "contextWindowHint": "Giới hạn token cho ngữ cảnh model.",
     "maxToolIterations": "Số lần lặp công cụ tối đa",
-    "maxToolIterationsHint": "Số lần gọi công cụ tối đa mỗi yêu cầu."
+    "maxToolIterationsHint": "Số lần gọi công cụ tối đa mỗi yêu cầu.",
+    "stripAssistantPrefill": "Loại bỏ Assistant Prefill",
+    "stripAssistantPrefillHint": "Xóa tin nhắn assistant cuối cùng trước khi gọi LLM. Bật cho proxy provider không hỗ trợ prefill."
   },
   "workspace": {
     "title": "Workspace",

ui/web/src/i18n/locales/zh/agents.json

@@ -147,7 +147,9 @@
     "contextWindow": "上下文窗口",
     "contextWindowHint": "模型上下文的令牌限制。",
     "maxToolIterations": "最大工具迭代次数",
-    "maxToolIterationsHint": "每次请求最大工具调用次数。"
+    "maxToolIterationsHint": "每次请求最大工具调用次数。",
+    "stripAssistantPrefill": "移除助手预填充",
+    "stripAssistantPrefillHint": "在调用LLM前移除末尾的助手消息。为不支持预填充的代理提供商启用。"
   },
   "workspace": {
     "title": "工作区",

ui/web/src/pages/agents/agent-detail/agent-overview-tab.tsx

@@ -38,6 +38,8 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
   const [budgetDollars, setBudgetDollars] = useState(
     agent.budget_monthly_cents ? String(agent.budget_monthly_cents / 100) : "",
   );
+  const [stripAssistantPrefill, setStripAssistantPrefill] = useState(Boolean(otherCfg.strip_assistant_prefill));
+
   // Evolution (predefined only)
   const [selfEvolve, setSelfEvolve] = useState(Boolean(otherCfg.self_evolve));
   const [skillEvolve, setSkillEvolve] = useState(Boolean(otherCfg.skill_evolve));
@@ -70,6 +72,7 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
         self_evolve: selfEvolve,
         skill_evolve: skillEvolve,
         skill_nudge_interval: skillEvolve ? skillNudgeInterval : undefined,
+        strip_assistant_prefill: stripAssistantPrefill || undefined,
       };
       const budgetCents = budgetDollars ? Math.round(parseFloat(budgetDollars) * 100) : null;
       await onUpdate({
@@ -128,6 +131,8 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
         budgetDollars={budgetDollars}
         onBudgetDollarsChange={setBudgetDollars}
         onSaveBlockedChange={setLlmSaveBlocked}
+        stripAssistantPrefill={stripAssistantPrefill}
+        onStripAssistantPrefillChange={setStripAssistantPrefill}
       />
 
       <HeartbeatCard heartbeat={heartbeat} />

ui/web/src/pages/agents/agent-detail/overview-sections/model-budget-section.tsx

@@ -3,6 +3,7 @@ import { useTranslation } from "react-i18next";
 import { DollarSign } from "lucide-react";
 import { Input } from "@/components/ui/input";
 import { Label } from "@/components/ui/label";
+import { Switch } from "@/components/ui/switch";
 import { ProviderModelSelect } from "@/components/shared/provider-model-select";
 
 interface ModelBudgetSectionProps {
@@ -19,6 +20,8 @@ interface ModelBudgetSectionProps {
   budgetDollars: string;
   onBudgetDollarsChange: (v: string) => void;
   onSaveBlockedChange?: (blocked: boolean) => void;
+  stripAssistantPrefill?: boolean;
+  onStripAssistantPrefillChange?: (v: boolean) => void;
 }
 
 export function ModelBudgetSection({
@@ -28,6 +31,8 @@ export function ModelBudgetSection({
   savedProvider, savedModel,
   budgetDollars, onBudgetDollarsChange,
   onSaveBlockedChange,
+  stripAssistantPrefill,
+  onStripAssistantPrefillChange,
 }: ModelBudgetSectionProps) {
   const { t } = useTranslation("agents");
 
@@ -100,6 +105,20 @@ export function ModelBudgetSection({
         </div>
         <p className="text-xs text-muted-foreground">{t("general.budgetHint")}</p>
       </div>
+
+      {onStripAssistantPrefillChange && (
+        <div className="flex items-center justify-between gap-4 rounded-md border px-3 py-2">
+          <div className="space-y-0.5">
+            <Label htmlFor="strip-prefill" className="text-xs">{t("llmConfig.stripAssistantPrefill")}</Label>
+            <p className="text-xs text-muted-foreground">{t("llmConfig.stripAssistantPrefillHint")}</p>
+          </div>
+          <Switch
+            id="strip-prefill"
+            checked={stripAssistantPrefill ?? false}
+            onCheckedChange={onStripAssistantPrefillChange}
+          />
+        </div>
+      )}
     </section>
   );
 }