feat(agent): 支持为每个 agent 绑定 LLM 配置(agent 级默认模型)

按优先级 Agent 绑定 > 任务/会话指定 > 全局激活 解析每个 agent 运行的模型,
使 planner/worker/goals/mainagent/自定义 agent 可各自走不同 profile(如 plan 用
强模型、work 用弱模型)。未绑定时行为与原先完全一致。

- db: Agent 增 llm_profile_id + SetAgentLLMProfile;schema 幂等兜底 ALTER
- server: 新增 effectiveProfileForAgent/providerForProfile/providerForAgent;
  buildPlannerWorker 让 planner/worker 各自解析 provider;provider 按 profile id
  缓存以共享限速器;goals/会话链同步走绑定;绑定变更热重载
- api: agent 配置接口收 llm_profile_id(三态) + 返回可选 profile 列表
- web: agent 编辑器「配置与提示词」页新增「默认模型」下拉

db/config.go

@@ -123,6 +123,7 @@ type Agent struct {
 	Role             string `json:"role"`
 	Builtin          bool   `json:"builtin"`
 	Enabled          bool   `json:"enabled"`
+	LLMProfileID     *int64 `json:"llm_profile_id"`    // 绑定的 LLM 配置;nil=跟随任务/会话 pin,再回退全局激活
 	MaxTurns         int    `json:"max_turns"`         // 单次运行最大轮次;0=不限制
 	RunSecs          int    `json:"run_seconds"`       // worker 单次运行墙钟上限(秒);0=不限制
 	WebSearch        bool   `json:"web_search"`        // 是否启用网络搜索(受系统全局开关门控)
@@ -141,11 +142,16 @@ type Agent struct {
 	TriggerMaxParallel int    `json:"trigger_max_parallel"`
 }
 
-const agentCols = `id,key,name,COALESCE(description,''),role,builtin,enabled,COALESCE(max_turns,0),COALESCE(run_seconds,600),COALESCE(web_search,false),COALESCE(interactive_shell,false),COALESCE(wrapup_prompt,''),COALESCE(wrapup_max_turns,0),COALESCE(task_timeout_wrapup_prompt,''),COALESCE(task_timeout_wrapup_max_turns,0),COALESCE(trigger_run_mode,'serial'),COALESCE(trigger_merge_mode,'by_task'),COALESCE(trigger_max_parallel,5)`
+const agentCols = `id,key,name,COALESCE(description,''),role,builtin,enabled,COALESCE(max_turns,0),COALESCE(run_seconds,600),COALESCE(web_search,false),COALESCE(interactive_shell,false),COALESCE(wrapup_prompt,''),COALESCE(wrapup_max_turns,0),COALESCE(task_timeout_wrapup_prompt,''),COALESCE(task_timeout_wrapup_max_turns,0),COALESCE(trigger_run_mode,'serial'),COALESCE(trigger_merge_mode,'by_task'),COALESCE(trigger_max_parallel,5),llm_profile_id`
 
 func scanAgent(sc interface{ Scan(...any) error }) (*Agent, error) {
 	var a Agent
-	err := sc.Scan(&a.ID, &a.Key, &a.Name, &a.Description, &a.Role, &a.Builtin, &a.Enabled, &a.MaxTurns, &a.RunSecs, &a.WebSearch, &a.InteractiveShell, &a.WrapupPrompt, &a.WrapupMaxTurns, &a.TaskTimeoutWrapupPrompt, &a.TaskTimeoutWrapupMaxTurns, &a.TriggerRunMode, &a.TriggerMergeMode, &a.TriggerMaxParallel)
+	var prof sql.NullInt64 // llm_profile_id 可空:未绑定时为 NULL
+	err := sc.Scan(&a.ID, &a.Key, &a.Name, &a.Description, &a.Role, &a.Builtin, &a.Enabled, &a.MaxTurns, &a.RunSecs, &a.WebSearch, &a.InteractiveShell, &a.WrapupPrompt, &a.WrapupMaxTurns, &a.TaskTimeoutWrapupPrompt, &a.TaskTimeoutWrapupMaxTurns, &a.TriggerRunMode, &a.TriggerMergeMode, &a.TriggerMaxParallel, &prof)
+	if err == nil && prof.Valid {
+		v := prof.Int64
+		a.LLMProfileID = &v
+	}
 	return &a, err
 }
 
@@ -262,6 +268,14 @@ func (d *DB) SetAgentMaxTurns(key string, maxTurns int) error {
 	return err
 }
 
+// SetAgentLLMProfile binds an agent to a specific LLM profile (id != nil), or clears
+// the binding (id == nil) so the agent follows the task/conversation pin, else the
+// global active profile. Precedence at runtime: agent binding → task/conv pin → active.
+func (d *DB) SetAgentLLMProfile(key string, id *int64) error {
+	_, err := d.Exec(`UPDATE agents SET llm_profile_id=$1 WHERE key=$2`, id, key)
+	return err
+}
+
 // SetAgentWebSearch toggles whether an agent uses network search (still gated by
 // the global web-search master switch + backend/key config).
 func (d *DB) SetAgentWebSearch(key string, on bool) error {

db/schema.sql

@@ -307,6 +307,8 @@ CREATE TABLE IF NOT EXISTS agents (
 ALTER TABLE agents ADD COLUMN IF NOT EXISTS trigger_run_mode     TEXT    NOT NULL DEFAULT 'serial';
 ALTER TABLE agents ADD COLUMN IF NOT EXISTS trigger_merge_mode   TEXT    NOT NULL DEFAULT 'by_task';
 ALTER TABLE agents ADD COLUMN IF NOT EXISTS trigger_max_parallel INTEGER NOT NULL DEFAULT 5;
+-- per-agent LLM 绑定(agent 级默认模型):列自初版即在上方 CREATE 中,此 ALTER 仅为极旧库兜底(幂等)。
+ALTER TABLE agents ADD COLUMN IF NOT EXISTS llm_profile_id BIGINT REFERENCES llm_profiles(id) ON DELETE SET NULL;
 DROP TRIGGER IF EXISTS trg_agents_upd ON agents;
 CREATE TRIGGER trg_agents_upd BEFORE UPDATE ON agents
     FOR EACH ROW EXECUTE FUNCTION set_updated_at();

server/conversations.go

@@ -321,9 +321,10 @@ func (s *Server) runConversationSync(c *db.Conversation, msg, busyKey string) {
 		delete(s.chatCancel, busyKey)
 		s.chatMu.Unlock()
 	}()
+	// precedence: the conversation agent's own binding → this conversation's pin → global.
 	ca := s.chatAgentRef()
-	if c.LLMProfileID != nil {
-		if pa := s.chatAgentForProfile(*c.LLMProfileID); pa != nil {
+	if eff := s.effectiveProfileForAgent(c.AgentKey, c.LLMProfileID); eff != nil {
+		if pa := s.chatAgentForProfile(*eff); pa != nil {
 			ca = pa
 		}
 	}

server/dto.go

@@ -304,6 +304,7 @@ type AgentDTO struct {
 	RunSecs          int    `json:"run_seconds"`
 	WebSearch        bool   `json:"web_search"`
 	InteractiveShell bool   `json:"interactive_shell"`
+	LLMProfileID     *int64 `json:"llm_profile_id"` // 绑定的 LLM 配置;null=跟随任务/全局
 	// P3 触发后处理策略(仅自定义 agent 有意义)。
 	TriggerRunMode     string `json:"trigger_run_mode"`
 	TriggerMergeMode   string `json:"trigger_merge_mode"`
@@ -327,6 +328,7 @@ func agentDTO(a *db.Agent) AgentDTO {
 		RunSecs:            a.RunSecs,
 		WebSearch:          a.WebSearch,
 		InteractiveShell:   a.InteractiveShell,
+		LLMProfileID:       a.LLMProfileID,
 		TriggerRunMode:     a.TriggerRunMode,
 		TriggerMergeMode:   a.TriggerMergeMode,
 		TriggerMaxParallel: a.TriggerMaxParallel,

server/goals.go

@@ -62,11 +62,16 @@ func (s *Server) createGoals(ctx context.Context, t *Task, emit func(db.Activity
 	return specs
 }
 
-// goalsLLMConfig resolves the LLM config for goal decomposition: the task's pinned
-// profile if any, else the global active profile (same source the engine runs on).
+// goalsLLMConfig resolves the LLM config for goal decomposition by the standard
+// precedence: the goals agent's own binding → the task's pinned profile → the global
+// active profile (same source the engine runs on).
 func (s *Server) goalsLLMConfig(t *Task) (agent.Config, bool) {
-	if t != nil && t.LLMProfileID != nil {
-		if cfg, ok := s.loadProfileConfig(*t.LLMProfileID); ok {
+	var pin *int64
+	if t != nil {
+		pin = t.LLMProfileID
+	}
+	if eff := s.effectiveProfileForAgent("goals", pin); eff != nil {
+		if cfg, ok := s.loadProfileConfig(*eff); ok {
 			return cfg, true
 		}
 	}

server/server.go

@@ -78,6 +78,15 @@ type Server struct {
 	profMu         sync.Mutex
 	profAgents     map[int64]*profBundle
 	profChatAgents map[int64]*agent.ChatAgent // per-profile ChatAgent cache (chat page)
+
+	// provByProfile caches ONE provider per LLM profile id so every agent bound or
+	// pinned to the same profile shares a single provider instance — hence one rate
+	// limiter. Separate from profAgents (whole planner/worker pairs). The globally-active
+	// provider (built in applyLLM) lives OUTSIDE this cache, so binding an agent to the
+	// currently-active profile builds a second instance — harmless (that binding equals
+	// leaving it unset), just not deduped. Cleared alongside profAgents on profile edits.
+	provCacheMu   sync.Mutex
+	provByProfile map[int64]*provEntry
 }
 
 // profBundle is a planner/worker pair built from one LLM profile.
@@ -86,6 +95,12 @@ type profBundle struct {
 	wk *agent.Worker
 }
 
+// provEntry is a cached provider + its config for one LLM profile id.
+type provEntry struct {
+	prov llm.Provider
+	cfg  agent.Config
+}
+
 // triggeredRun is one queued P3 trigger fire awaiting its turn for an agent.
 // taskID + mergeable let the drainer coalesce several event triggers (finding/goal)
 // from the SAME task into one conversation before it starts (interval fires don't merge).
@@ -105,7 +120,8 @@ func New(ctx context.Context, m *Manager, skillDir string, dataDir string, keyDi
 	s := &Server{m: m, engine: NewEngine(m), ctx: ctx, skillDir: skillDir, jwtKey: key, chatBusy: map[string]bool{},
 		chatCancel: map[string]context.CancelFunc{}, triggerQ: map[string][]triggeredRun{},
 		triggerActive: map[string]int{}, triggerCfg: map[string]triggerBehavior{},
-		profAgents: map[int64]*profBundle{}, profChatAgents: map[int64]*agent.ChatAgent{}}
+		profAgents: map[int64]*profBundle{}, profChatAgents: map[int64]*agent.ChatAgent{},
+		provByProfile: map[int64]*provEntry{}}
 	// per-task LLM: a task pinned to a specific profile runs on that profile's
 	// dedicated planner/worker; unpinned tasks fall back to the global active pair.
 	s.engine.SetAgentResolver(func(t *Task) (*agent.Planner, *agent.Worker) {
@@ -307,21 +323,24 @@ func (s *Server) webSearchFor(key string) agent.WebSearchOpts {
 	return o
 }
 
-// buildPlannerWorker builds a planner+worker pair on an already-constructed provider
-// + cfg, with all engine callbacks / proxy / web-search / memory wiring. Shared by the
-// global apply path (applyLLM) and the per-profile path (agentsForProfile), so a task
-// pinned to a specific profile behaves identically to the active one — just a different LLM.
-func (s *Server) buildPlannerWorker(prov llm.Provider, cfg agent.Config) (*agent.Planner, *agent.Worker) {
+// buildPlannerWorker builds a planner+worker pair. Each agent resolves its OWN LLM by
+// precedence agent-binding → pin → the passed global fallback (gProv/gCfg), so planner
+// and worker can run on different models (e.g. a stronger planner, a cheaper worker).
+// pinID is the task's pinned profile (nil on the global active path). With no agent
+// binding and no pin, both fall back to gProv/gCfg — identical to the previous single-
+// provider behavior. Shared by applyLLM (global) and agentsForProfile (per-task pin).
+func (s *Server) buildPlannerWorker(pinID *int64, gProv llm.Provider, gCfg agent.Config) (*agent.Planner, *agent.Worker) {
 	tx := transcript.NewStore(filepath.Join(s.m.dir, "transcripts")) // raw LLM conversation logs
-	win := cfg.CompactionWindow()                                    // context window in tokens (compaction)
 	// traffic host tools flow through ToolAugment for every agent and are filtered by
 	// the tools-table binding (default = worker), so worker behavior is unchanged.
-	wk := agent.NewWorker(prov, cfg.Model, s.m.dir, tx, win, s.agentMaxTurns("worker"))
+	wProv, wCfg := s.providerForAgent("worker", pinID, gProv, gCfg)
+	wk := agent.NewWorker(wProv, wCfg.Model, s.m.dir, tx, wCfg.CompactionWindow(), s.agentMaxTurns("worker"))
 	wk.SetRunTimeout(time.Duration(s.agentRunSeconds("worker")) * time.Second)
 	wk.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
 	wk.SetMemory(memory.NewStore(filepath.Join(s.m.dir, "memory")))
 	wk.SetWebSearch(s.webSearchFor("worker"))
-	pl := agent.NewPlanner(prov, cfg.Model, s.m.dir, tx, win, s.agentMaxTurns("planner"))
+	pProv, pCfg := s.providerForAgent("planner", pinID, gProv, gCfg)
+	pl := agent.NewPlanner(pProv, pCfg.Model, s.m.dir, tx, pCfg.CompactionWindow(), s.agentMaxTurns("planner"))
 	pl.SetKillWork(s.engine.KillWork)               // planner kill_work → terminate a running work
 	pl.SetSteerWork(s.engine.SteerWork)             // planner steer_work → inject mid-run course-correction
 	pl.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert()) // WebFetch through the recording proxy
@@ -341,13 +360,17 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	profName := s.llmProf
 	s.cfgMu.Unlock()
 	prov = llmrec.Wrap(prov, s.m.PG(), cfg.Model, profName, s.m.LLMRecordEnabled)
-	pl, wk := s.buildPlannerWorker(prov, cfg)
+	pl, wk := s.buildPlannerWorker(nil, prov, cfg)
 	s.engine.UseLLM(pl, wk)
 
 	tx := transcript.NewStore(filepath.Join(s.m.dir, "transcripts"))
 	win := cfg.CompactionWindow()
+	// mainagent resolves its own binding (→ global fallback); chat stays the GLOBAL
+	// fallback since one ChatAgent serves many agent keys — its per-agent binding is
+	// resolved at Chat time (runConversationSync → chatAgentForProfile).
+	mProv, mCfg := s.providerForAgent("mainagent", nil, prov, cfg)
 	s.cfgMu.Lock()
-	s.mainAgent = agent.NewMainAgent(prov, cfg.Model, s.m.dir, tx, win, s.agentMaxTurns("mainagent"))
+	s.mainAgent = agent.NewMainAgent(mProv, mCfg.Model, s.m.dir, tx, mCfg.CompactionWindow(), s.agentMaxTurns("mainagent"))
 	s.mainAgent.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert()) // WebFetch through the recording proxy
 	s.mainAgent.SetWebSearch(s.webSearchFor("mainagent"))
 	// chat agent serves MANY custom agents by key → it holds the GLOBAL opts
@@ -384,31 +407,88 @@ func (s *Server) loadProfileConfig(id int64) (agent.Config, bool) {
 	return cfg, true
 }
 
-// agentsForProfile returns the dedicated planner/worker for a specific LLM profile,
-// built + cached on first use (tasks on the same profile share one provider + limiter).
-// nil,nil when the profile is invalid → the caller falls back to the global active pair.
-func (s *Server) agentsForProfile(id int64) (*agent.Planner, *agent.Worker) {
-	s.profMu.Lock()
-	defer s.profMu.Unlock()
-	if b := s.profAgents[id]; b != nil {
-		return b.pl, b.wk
+// effectiveProfileForAgent resolves the LLM profile id an agent should run on, by
+// precedence: agent binding (agents.llm_profile_id) → pin (task/conversation) → nil
+// (caller falls back to the global active profile). A binding to a deleted profile
+// can't happen (FK ON DELETE SET NULL); an otherwise-invalid one is dropped downstream
+// by loadProfileConfig, letting the caller fall back.
+func (s *Server) effectiveProfileForAgent(agentKey string, pinID *int64) *int64 {
+	if s.m.pg != nil && agentKey != "" {
+		if a, _ := s.m.pg.GetAgentByKey(agentKey); a != nil && a.LLMProfileID != nil {
+			return a.LLMProfileID
+		}
 	}
+	return pinID
+}
+
+// providerForProfile returns a cached provider+cfg for a profile id, so every agent
+// bound/pinned to the same profile shares one provider instance (one rate limiter).
+// ok=false when the profile is missing/invalid → caller falls back to the global pair.
+func (s *Server) providerForProfile(id int64) (llm.Provider, agent.Config, bool) {
+	s.provCacheMu.Lock()
+	if e := s.provByProfile[id]; e != nil {
+		s.provCacheMu.Unlock()
+		return e.prov, e.cfg, true
+	}
+	s.provCacheMu.Unlock()
 	cfg, ok := s.loadProfileConfig(id)
 	if !ok {
-		return nil, nil
+		return nil, agent.Config{}, false
 	}
 	prov, err := cfg.NewProvider()
 	if err != nil {
 		log.Printf("[engine] build provider for LLM profile %d failed: %v", id, err)
-		return nil, nil
+		return nil, agent.Config{}, false
 	}
 	// Wrap with recorder, tagged with this profile's name.
 	if p, _ := s.m.pg.ProfileByID(id); p != nil {
 		prov = llmrec.Wrap(prov, s.m.PG(), cfg.Model, p.Name, s.m.LLMRecordEnabled)
 	}
-	pl, wk := s.buildPlannerWorker(prov, cfg)
-	s.profAgents[id] = &profBundle{pl: pl, wk: wk}
-	log.Printf("[engine] built dedicated planner/worker for LLM profile %d (%s / %s)", id, cfg.Provider(), cfg.Model)
+	s.provCacheMu.Lock()
+	if e := s.provByProfile[id]; e != nil { // lost the race → keep the winner
+		prov, cfg = e.prov, e.cfg
+	} else {
+		s.provByProfile[id] = &provEntry{prov: prov, cfg: cfg}
+		log.Printf("[engine] built provider for LLM profile %d (%s / %s)", id, cfg.Provider(), cfg.Model)
+	}
+	s.provCacheMu.Unlock()
+	return prov, cfg, true
+}
+
+// providerForAgent returns the provider+cfg one agent should run on: its bound/pinned
+// profile if that resolves, else the passed global fallback (gProv/gCfg).
+func (s *Server) providerForAgent(agentKey string, pinID *int64, gProv llm.Provider, gCfg agent.Config) (llm.Provider, agent.Config) {
+	if eff := s.effectiveProfileForAgent(agentKey, pinID); eff != nil {
+		if prov, cfg, ok := s.providerForProfile(*eff); ok {
+			return prov, cfg
+		}
+	}
+	return gProv, gCfg
+}
+
+// agentsForProfile returns the dedicated planner/worker for a task pinned to a specific
+// LLM profile, built + cached on first use (tasks on the same pin share one pair). Each
+// agent still honors its own binding first (via buildPlannerWorker), falling back to this
+// pinned profile. nil,nil when the profile is invalid → caller uses the global active pair.
+func (s *Server) agentsForProfile(id int64) (*agent.Planner, *agent.Worker) {
+	s.profMu.Lock()
+	b := s.profAgents[id]
+	s.profMu.Unlock()
+	if b != nil {
+		return b.pl, b.wk
+	}
+	prov, cfg, ok := s.providerForProfile(id)
+	if !ok {
+		return nil, nil
+	}
+	pl, wk := s.buildPlannerWorker(&id, prov, cfg)
+	s.profMu.Lock()
+	if ex := s.profAgents[id]; ex != nil { // lost the race → keep the winner
+		pl, wk = ex.pl, ex.wk
+	} else {
+		s.profAgents[id] = &profBundle{pl: pl, wk: wk}
+	}
+	s.profMu.Unlock()
 	return pl, wk
 }
 
@@ -416,40 +496,41 @@ func (s *Server) agentsForProfile(id int64) (*agent.Planner, *agent.Worker) {
 // per profile id. Returns nil if the profile is missing or has no API key.
 func (s *Server) chatAgentForProfile(id int64) *agent.ChatAgent {
 	s.profMu.Lock()
-	defer s.profMu.Unlock()
-	if ca := s.profChatAgents[id]; ca != nil {
-		return ca
+	cached := s.profChatAgents[id]
+	s.profMu.Unlock()
+	if cached != nil {
+		return cached
 	}
-	cfg, ok := s.loadProfileConfig(id)
+	prov, cfg, ok := s.providerForProfile(id)
 	if !ok {
 		return nil
 	}
-	prov, err := cfg.NewProvider()
-	if err != nil {
-		log.Printf("[chat] build provider for LLM profile %d failed: %v", id, err)
-		return nil
-	}
-	// Wrap with recorder, tagged with this profile's name.
-	if p, _ := s.m.pg.ProfileByID(id); p != nil {
-		prov = llmrec.Wrap(prov, s.m.PG(), cfg.Model, p.Name, s.m.LLMRecordEnabled)
-	}
 	tx := transcript.NewStore(filepath.Join(s.m.dir, "transcripts"))
-	win := cfg.CompactionWindow()
-	ca := agent.NewChatAgent(prov, cfg.Model, s.m.dir, tx, win)
+	ca := agent.NewChatAgent(prov, cfg.Model, s.m.dir, tx, cfg.CompactionWindow())
 	ca.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
 	ca.SetWebSearch(s.m.WebSearchOpts())
 	ca.SetGuard(s.chatGuard())
-	s.profChatAgents[id] = ca
+	s.profMu.Lock()
+	if ex := s.profChatAgents[id]; ex != nil { // lost the race → keep the winner
+		ca = ex
+	} else {
+		s.profChatAgents[id] = ca
+	}
+	s.profMu.Unlock()
 	return ca
 }
 
-// invalidateProfileAgents drops the per-profile agent cache so a profile save/activate/
-// delete rebuilds pinned tasks' planner/worker on their next round.
+// invalidateProfileAgents drops the per-profile agent + provider caches so a profile
+// save/activate/delete — or an agent's binding change — rebuilds pinned tasks' planner/
+// worker (and re-resolves each agent's bound model) on their next round.
 func (s *Server) invalidateProfileAgents() {
 	s.profMu.Lock()
 	s.profAgents = map[int64]*profBundle{}
 	s.profChatAgents = map[int64]*agent.ChatAgent{}
 	s.profMu.Unlock()
+	s.provCacheMu.Lock()
+	s.provByProfile = map[int64]*provEntry{}
+	s.provCacheMu.Unlock()
 }
 
 func (s *Server) mainAgentRef() *agent.MainAgent {

server/server_mgmt.go

@@ -220,6 +220,8 @@ func (s *Server) pgSaveAgentConfig(w http.ResponseWriter, r *http.Request) {
 		RunSeconds       *int  `json:"run_seconds"`
 		WebSearch        *bool `json:"web_search"`
 		InteractiveShell *bool `json:"interactive_shell"`
+		// llm_profile_id 三态:字段缺省=不动;显式 null=解绑(跟随任务/全局);数字=绑定该 profile。
+		LLMProfileID json.RawMessage `json:"llm_profile_id"`
 		// P3 触发后处理策略(三者一起可选,提供任一即整体写入;未提供则不动)。
 		TriggerRunMode     *string `json:"trigger_run_mode"`
 		TriggerMergeMode   *string `json:"trigger_merge_mode"`
@@ -229,6 +231,25 @@ func (s *Server) pgSaveAgentConfig(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, 400, err.Error())
 		return
 	}
+	profileChanged := false
+	if req.LLMProfileID != nil { // key present (数字 或 null)
+		var id *int64
+		if err := json.Unmarshal(req.LLMProfileID, &id); err != nil {
+			writeErr(w, 400, "llm_profile_id 格式错误")
+			return
+		}
+		if id != nil { // 绑定:校验目标 profile 有效
+			if _, ok := s.loadProfileConfig(*id); !ok {
+				writeErr(w, 400, "指定的 LLM 配置不存在或无效")
+				return
+			}
+		}
+		if err := pg.SetAgentLLMProfile(a.Key, id); err != nil {
+			writeErr(w, 500, err.Error())
+			return
+		}
+		profileChanged = true
+	}
 	if req.MaxTurns != nil {
 		mt := *req.MaxTurns
 		if mt < 0 {
@@ -279,7 +300,12 @@ func (s *Server) pgSaveAgentConfig(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 	}
-	// rebuild the live agents so the new max_turns/run_seconds apply without a restart.
+	// an agent's LLM binding change also invalidates pinned-task / per-profile caches
+	// so their next round re-resolves each agent's bound model.
+	if profileChanged {
+		s.invalidateProfileAgents()
+	}
+	// rebuild the live agents so the new max_turns/run_seconds/binding apply without a restart.
 	s.cfgMu.Lock()
 	cfg, on := s.llmCfg, s.llmOn
 	s.cfgMu.Unlock()
@@ -313,9 +339,19 @@ func (s *Server) pgGetAgent(w http.ResponseWriter, r *http.Request) {
 	if sk == nil {
 		sk = []string{}
 	}
+	// 可选 LLM 配置列表(id/name/model/是否默认),供前端渲染 "默认模型" 下拉;当前绑定见 agent.llm_profile_id。
+	profs, _ := pg.ListProfiles()
+	llmProfiles := make([]map[string]any, 0, len(profs))
+	for _, p := range profs {
+		llmProfiles = append(llmProfiles, map[string]any{
+			"id": p.ID, "name": p.Name, "model": p.Model, "is_default": p.IsDefault,
+		})
+	}
 	writeJSON(w, 200, map[string]any{
 		"agent": agentDTO(a), "prompt": cur, "variables": vars, "versions": vers,
 		"visibility":               map[string]any{"mcp": mcp, "skill": sk},
+		"llm_profiles":             llmProfiles, // 可绑定的 LLM 配置候选
+
 		"wrapup_prompt":            a.WrapupPrompt,                  // 已保存的收尾提示词(空=用内置默认)
 		"wrapup_default":           agent.WrapupDefault(a.Key),      // 内置默认(供占位/恢复默认)
 		"wrapup_max_turns":         a.WrapupMaxTurns,                // 已保存的收尾轮数(0=用内置默认)

web/src/components/agent-editor.tsx

@@ -51,6 +51,9 @@ export function AgentEditor({ agentKey, onSaved }: { agentKey: string; onSaved?:
   const [preview, setPreview] = React.useState("");
   const [maxTurns, setMaxTurns] = React.useState("0");
   const [runSecs, setRunSecs] = React.useState("600");
+  // "" = 跟随(未绑定)；否则为 profile id 字符串
+  const [llmProfileId, setLlmProfileId] = React.useState("");
+  const [llmProfiles, setLlmProfiles] = React.useState<NonNullable<AgentDetail["llm_profiles"]>>([]);
   const [webSearch, setWebSearch] = React.useState(false);
   const [interactiveShell, setInteractiveShell] = React.useState(false);
   const [wrapup, setWrapup] = React.useState("");
@@ -87,6 +90,8 @@ export function AgentEditor({ agentKey, onSaved }: { agentKey: string; onSaved?:
         setSkillVisible(d.visibility?.skill ?? []);
         setMaxTurns(String(d.agent?.max_turns ?? 0));
         setRunSecs(String(d.agent?.run_seconds ?? 600));
+        setLlmProfileId(d.agent?.llm_profile_id != null ? String(d.agent.llm_profile_id) : "");
+        setLlmProfiles(d.llm_profiles ?? []);
         setWebSearch(!!d.agent?.web_search);
         setInteractiveShell(!!d.agent?.interactive_shell);
         setWrapup(d.wrapup_prompt ?? "");
@@ -173,9 +178,17 @@ export function AgentEditor({ agentKey, onSaved }: { agentKey: string; onSaved?:
   }
   async function saveConfig() {
     try {
-      const n = Math.max(0, Math.floor(Number(maxTurns) || 0));
-      const rs = Math.max(0, Math.floor(Number(runSecs) || 0));
-      await api.saveAgentConfig(agentKey, { max_turns: n, run_seconds: rs, web_search: webSearch, interactive_shell: interactiveShell });
+      // 只提交本 agent 实际展示的字段，避免把未显示项(如 goals 的 max_turns)覆盖成默认。
+      const patch: Parameters<typeof api.saveAgentConfig>[1] = {
+        llm_profile_id: llmProfileId === "" ? null : Number(llmProfileId),
+      };
+      if (showConfig) {
+        patch.max_turns = Math.max(0, Math.floor(Number(maxTurns) || 0));
+        patch.run_seconds = Math.max(0, Math.floor(Number(runSecs) || 0));
+      }
+      if (showWebSearch) patch.web_search = webSearch;
+      if (showInteractiveShell) patch.interactive_shell = interactiveShell;
+      await api.saveAgentConfig(agentKey, patch);
       toast.success("已保存运行配置（立即生效）");
       reload();
     } catch (e) {
@@ -248,6 +261,8 @@ export function AgentEditor({ agentKey, onSaved }: { agentKey: string; onSaved?:
   const showWebSearch = agentKey !== "goals";
   // interactive shell (持久 PTY 会话工具族) 同样对除 goals 外的 agent 开放;无全局门控。
   const showInteractiveShell = agentKey !== "goals";
+  // 每个 agent(含 goals/mainagent)都跑在某个 LLM 上,故「默认模型」绑定对所有 agent 开放。
+  const showLLM = true;
   // triggers (P3) only attach to custom agents.
   const isCustom = !!detail && !detail.agent?.builtin;
 
@@ -265,8 +280,31 @@ export function AgentEditor({ agentKey, onSaved }: { agentKey: string; onSaved?:
       {/* 配置 + 提示词 */}
       <TabsContent value="prompt" className="min-h-0 flex-1 overflow-y-auto px-4 pb-4">
         <div className="grid gap-4">
-          {(showConfig || showWebSearch || showInteractiveShell) && (
+          {(showLLM || showConfig || showWebSearch || showInteractiveShell) && (
             <div className="grid gap-3 rounded-md border p-3">
+              {showLLM && (
+                <div className="grid gap-1.5">
+                  <Label htmlFor="llm-profile" className="text-xs">默认模型（LLM 配置）</Label>
+                  <div className="flex flex-wrap items-center gap-3">
+                    <Select value={llmProfileId || "__follow__"} onValueChange={(v) => setLlmProfileId(v === "__follow__" ? "" : v)}>
+                      <SelectTrigger id="llm-profile" className="h-8 w-72">
+                        <SelectValue />
+                      </SelectTrigger>
+                      <SelectContent>
+                        <SelectItem value="__follow__">跟随任务 / 全局激活配置</SelectItem>
+                        {llmProfiles.map((p) => (
+                          <SelectItem key={p.id} value={String(p.id)}>
+                            {p.name}（{p.model}）{p.is_default ? " · 默认" : ""}
+                          </SelectItem>
+                        ))}
+                      </SelectContent>
+                    </Select>
+                    <span className="text-muted-foreground max-w-md text-xs">
+                      为该 Agent 绑定固定 LLM 配置（点「保存配置」生效）。优先级：Agent 绑定 &gt; 任务/会话指定 &gt; 全局激活。
+                    </span>
+                  </div>
+                </div>
+              )}
               <div className="flex flex-wrap items-end gap-3">
                 {showConfig && (
                   <>

web/src/lib/api.ts

@@ -421,6 +421,7 @@ export const api = {
   saveAgentConfig: (
     key: string,
     patch: {
+      llm_profile_id?: number | null; // number=绑定；null=解绑(跟随任务/全局)；缺省=不动
       max_turns?: number;
       run_seconds?: number;
       web_search?: boolean;

web/src/lib/types.ts

@@ -456,6 +456,7 @@ export interface Agent {
   role: string;
   builtin: boolean;
   enabled: boolean;
+  llm_profile_id?: number | null; // 绑定的 LLM 配置；null/absent = 跟随任务/会话/全局
   max_turns?: number; // 0 = 无限制
   run_seconds?: number; // worker 单次运行墙钟上限(秒)；0 = 无限制
   web_search?: boolean; // 是否启用网络搜索(受系统全局开关门控)
@@ -490,6 +491,8 @@ export interface AgentDetail {
   variables: PromptVar[];
   versions: PromptVersion[];
   visibility: { mcp: number[]; skill: string[] };
+  // 可绑定的 LLM 配置候选(供「默认模型」下拉)；当前绑定见 agent.llm_profile_id
+  llm_profiles?: { id: number; name: string; model: string; is_default: boolean }[];
   wrapup_prompt?: string; // 已保存的收尾提示词(空=用内置默认)
   wrapup_default?: string; // 内置默认收尾提示词(占位/恢复默认)
   wrapup_max_turns?: number; // 已保存的收尾轮数(0=用内置默认)