Merge branch 'feat/noa-experimental'：noa 实验性上下文压缩

agent/chat.go

@@ -33,6 +33,7 @@ type ChatAgent struct {
 	webSearch      WebSearchOpts
 	guard          *guard.Guard // optional; nil disables intercept hooks for chat
 	nonStreamingFn func() bool  // resolver: use non-streaming (Complete) path? (nil = streaming)
+	noaEnabledFn   func() bool  // resolver: use experimental noa compaction? (nil = off)
 	maxTokensFn    func() int   // resolver: per-reply output cap (nil/0 = send no cap)
 }
 
@@ -46,6 +47,11 @@ func (c *ChatAgent) SetNonStreaming(fn func() bool) { c.nonStreamingFn = fn }
 
 func (c *ChatAgent) nonStreaming() bool { return c.nonStreamingFn != nil && c.nonStreamingFn() }
 
+// SetNoaEnabled wires a resolver deciding whether chat runs use the experimental
+// noa context-compression mechanism. nil/unset = off (built-in compaction). Read
+// per run so the settings toggle takes effect without rebuilding the agent.
+func (c *ChatAgent) SetNoaEnabled(fn func() bool) { c.noaEnabledFn = fn }
+
 // SetMaxTokens wires a resolver for the per-reply output cap. nil/unset or 0 =
 // send no cap and let the endpoint decide. Read per run, like nonStreaming.
 func (c *ChatAgent) SetMaxTokens(fn func() int) { c.maxTokensFn = fn }
@@ -161,6 +167,8 @@ func (c *ChatAgent) Chat(ctx context.Context, agentKey, sessionID, message strin
 		opts.Transcript = c.tx
 		opts.SessionID = sessionID
 	}
+	// 实验功能:开启后由 noa 接管上下文压缩(归档集中在 <workDir>/noa/<SessionID> 下,持久)。
+	enableNoa(&opts, c.noaEnabledFn, c.workDir, "chat-"+sessionID, noaWarn("chat-"+sessionID))
 	ctx = attachSideCapture(ctx, &opts)
 	s := agentcore.NewSession(opts)
 	defer s.Close()

agent/mainagent.go

@@ -31,9 +31,15 @@ type MainAgent struct {
 	workDir         string                                 // shared work dir (surfaced in prompt as artifact-output target)
 	steerWork       func(intentID int64, msg string) error // engine callback: steer a running work (nil = off)
 	nonStreamingFn  func() bool                            // resolver: use non-streaming (Complete) path? (nil = streaming)
+	noaEnabledFn    func() bool                            // resolver: use experimental noa compaction? (nil = off)
 	maxTokensFn     func() int                             // resolver: per-reply output cap (nil/0 = send no cap)
 }
 
+// SetNoaEnabled wires a resolver deciding whether runs use the experimental noa
+// context-compression mechanism. nil/unset = off (built-in compaction). Read per
+// run so the settings toggle takes effect without rebuilding the agent.
+func (m *MainAgent) SetNoaEnabled(fn func() bool) { m.noaEnabledFn = fn }
+
 // SetNonStreaming wires a resolver deciding whether runs use the non-streaming
 // model path (true = non-streaming). nil/unset = streaming (default).
 func (m *MainAgent) SetNonStreaming(fn func() bool) { m.nonStreamingFn = fn }
@@ -167,6 +173,13 @@ func (m *MainAgent) Chat(ctx context.Context, taskID int64, mainSeg int, as *db.
 			opts.SessionID = fmt.Sprintf("exp%d-main-s%d", ts.ID(), mainSeg)
 		}
 	}
+	// 实验功能:开启后由 noa 接管上下文压缩(归档集中在 <workDir>/noa/<SessionID> 下,持久)。
+	// session id 与 transcript 同规则(分段感知),使归档与恢复对齐。
+	noaSession := fmt.Sprintf("exp%d-main", ts.ID())
+	if mainSeg > 0 {
+		noaSession = fmt.Sprintf("exp%d-main-s%d", ts.ID(), mainSeg)
+	}
+	enableNoa(&opts, m.noaEnabledFn, m.workDir, noaSession, noaWarn(noaSession))
 	ctx = attachSideCapture(ctx, &opts)
 	s := agentcore.NewSession(opts)
 	defer s.Close()

agent/noa.go

@@ -0,0 +1,48 @@
+package agent
+
+import (
+	"log"
+	"path/filepath"
+
+	"github.com/Autumn-27/norma/agentcore"
+	"github.com/Autumn-27/norma/noaadapter"
+)
+
+// noaWarn returns a diagnostics sink tagging non-fatal noa messages with the
+// session, routed through the package logger (agents have no per-instance one).
+func noaWarn(session string) func(string) {
+	return func(msg string) { log.Printf("[noa] %s: %s", session, msg) }
+}
+
+// noa 是 norma v0.4.0 引入的「模型驱动上下文压缩」机制,作为平台实验功能由用户在
+// 系统设置中开关。它与内置 compaction 互斥:noaadapter.Enable 是唯一入口,一次挂上
+// 上下文接管器(Compactor)、Compress 工具与三段常驻提示词,不调用 Enable 即为关闭
+// (内置 compaction 照常工作)。开关由每个 agent 注入的 noaEnabledFn 解析,每 run 读
+// 一次,故切换只影响之后启动的 run,无需重建 agent。
+
+// enableNoa 在解析器报告开启时把 noa 接入 opts。archiveRoot 是压缩原文的持久化基目录
+// (取全局 workDir,各 agent 统一落在 <workDir>/noa 下,不随任务/意图目录分散),sessionID
+// 命名其下的归档子目录(全局唯一,故同一基目录内不冲突)。
+//
+// noa 是实验功能:接入失败不得中断真实任务。发生错误时经 onWarn 上报并回退内置压缩。
+// 启用成功时清掉 opts.Compaction,避免 agentcore 因「两个上下文管理器同时设置」告警。
+func enableNoa(opts *agentcore.Options, enabled func() bool, archiveRoot, sessionID string, onWarn func(string)) {
+	if enabled == nil || !enabled() {
+		return
+	}
+	if opts.OnWarn == nil {
+		opts.OnWarn = onWarn
+	}
+	if err := noaadapter.Enable(opts, noaadapter.Options{
+		ArchiveBaseDir: filepath.Join(archiveRoot, "noa"),
+		SessionID:      sessionID,
+		OnWarn:         onWarn,
+	}); err != nil {
+		if onWarn != nil {
+			onWarn("noa 压缩启用失败,回退内置压缩:" + err.Error())
+		}
+		return
+	}
+	// Compactor 覆盖 Compaction,但两者并存时 agentcore 每次会告警;明确清掉。
+	opts.Compaction = nil
+}

agent/planner.go

@@ -36,6 +36,7 @@ type Planner struct {
 	workDir           string                                 // shared work dir (surfaced in prompt as artifact-output target)
 	injectConstraints func() bool                            // resolver: inject task operation constraints into system prompt? (nil = yes)
 	nonStreamingFn    func() bool                            // resolver: use non-streaming (Complete) path? (nil = streaming)
+	noaEnabledFn      func() bool                            // resolver: use experimental noa compaction? (nil = off)
 	maxTokensFn       func() int                             // resolver: per-reply output cap (nil/0 = send no cap)
 	compactor         *Compactor                             // cold-node compaction (§7); nil = disabled
 
@@ -64,6 +65,11 @@ func (p *Planner) SetNonStreaming(fn func() bool) { p.nonStreamingFn = fn }
 
 func (p *Planner) nonStreaming() bool { return p.nonStreamingFn != nil && p.nonStreamingFn() }
 
+// SetNoaEnabled wires a resolver deciding whether runs use the experimental noa
+// context-compression mechanism. nil/unset = off (built-in compaction). Read per
+// run so the settings toggle takes effect without rebuilding the agent.
+func (p *Planner) SetNoaEnabled(fn func() bool) { p.noaEnabledFn = fn }
+
 // SetMaxTokens wires a resolver for the per-reply output cap. nil/unset or 0 =
 // send no cap and let the endpoint decide. Read per run, like nonStreaming.
 func (p *Planner) SetMaxTokens(fn func() int) { p.maxTokensFn = fn }
@@ -430,6 +436,9 @@ func (p *Planner) Plan(ctx context.Context, taskID int64, as *db.AssetStore, ts
 		opts.Transcript = p.tx
 		opts.SessionID = fmt.Sprintf("exp%d-planner", ts.ID())
 	}
+	// 实验功能:开启后由 noa 接管上下文压缩(归档集中在 <workDir>/noa/<SessionID> 下,持久)。
+	noaSession := fmt.Sprintf("exp%d-planner", ts.ID())
+	enableNoa(&opts, p.noaEnabledFn, p.workDir, noaSession, noaWarn(noaSession))
 	// 态势（刚完成的意图 + 完整图）现在拼进本轮 user 输入（见下方 input）。user 里还有
 	// 指令 + 跨唤醒待办（todo 是模型自己的规划便签，可再生，放 user 即可）。
 	// 开场白按「本轮有无具体变动」分两种：有变动 → 指向下方【实际变动】块；无变动

agent/worker.go

@@ -78,6 +78,10 @@ type Worker struct {
 	// path. Read per run so a profile/task toggle takes effect without rebuilding
 	// the agent. nil = streaming (default).
 	nonStreamingFn func() bool
+	// noaEnabledFn resolves whether this run uses the experimental noa context-
+	// compression mechanism. Read per run, like nonStreaming. nil = off (built-in
+	// compaction).
+	noaEnabledFn func() bool
 	// maxTokensFn resolves the per-reply output cap in tokens, on the same
 	// per-run basis. nil or 0 = send no cap and let the endpoint decide.
 	maxTokensFn func() int
@@ -114,6 +118,11 @@ func (w *Worker) SetNonStreaming(fn func() bool) { w.nonStreamingFn = fn }
 
 func (w *Worker) nonStreaming() bool { return w.nonStreamingFn != nil && w.nonStreamingFn() }
 
+// SetNoaEnabled wires a resolver deciding whether runs use the experimental noa
+// context-compression mechanism. nil/unset = off (built-in compaction). Read per
+// run so the settings toggle takes effect without rebuilding the agent.
+func (w *Worker) SetNoaEnabled(fn func() bool) { w.noaEnabledFn = fn }
+
 // SetMaxTokens wires a resolver for the per-reply output cap. nil/unset or 0 =
 // send no cap and let the endpoint decide. Read per run, like nonStreaming.
 func (w *Worker) SetMaxTokens(fn func() int) { w.maxTokensFn = fn }
@@ -481,6 +490,9 @@ func (w *Worker) execute(ctx context.Context, name string, taskID int64, as *db.
 		input = "开始执行 system 里领到的意图：只做它、只产生事实、assets、finding、做完即停。"
 	}
 
+	// 实验功能:开启后由 noa 接管上下文压缩(归档集中在 <workDir>/noa/<SessionID> 下,持久)。
+	noaSession := WorkerSessionID(ts.ID(), intent.ID)
+	enableNoa(&opts, w.noaEnabledFn, w.workDir, noaSession, noaWarn(noaSession))
 	ctx = attachSideCapture(ctx, &opts)
 	s := agentcore.NewSession(opts)
 	defer s.Close() // release the session's background-task manager (temp dir + processes)

go.mod

@@ -3,7 +3,7 @@ module github.com/Autumn-27/artex
 go 1.26.3
 
 require (
-	github.com/Autumn-27/norma v0.3.8
+	github.com/Autumn-27/norma v0.4.0
 	github.com/golang-jwt/jwt/v5 v5.3.1
 	github.com/google/uuid v1.6.0
 	github.com/jackc/pgx/v5 v5.10.0

go.sum

@@ -1,7 +1,5 @@
-github.com/Autumn-27/norma v0.3.7 h1:BdzVV3LUqN3zzaoL5cqY9aq0l/ucOYonSewQHhoL8Wo=
+github.com/Autumn-27/norma v0.4.0 h1:kJkFu6z8kyLV8/eo0cuPIqCTCCVBDrSETe8JppeSf/g=
-github.com/Autumn-27/norma v0.3.7/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
+github.com/Autumn-27/norma v0.4.0/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
-github.com/Autumn-27/norma v0.3.8 h1:MOz9bekMyQ29cgt728nZ4gq+HjpeDgQwCTd/7dtBvuc=
-github.com/Autumn-27/norma v0.3.8/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
 github.com/andybalholm/brotli v1.2.1 h1:R+f5xP285VArJDRgowrfb9DqL18yVK0gKAW/F+eTWro=
 github.com/andybalholm/brotli v1.2.1/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=
 github.com/asaskevich/govalidator v0.0.0-20230301143203-a9d515a09cc2 h1:DklsrG3dyBCFEj5IhUbnKptjxatkF07cF2ak3yi77so=

server/manager.go

@@ -271,6 +271,10 @@ const (
 	// 任务并发上限:开关 + 上限数。默认关闭;开启后默认上限 5(见 defaultConcurrencyLimit)。
 	settingConcurrencyOn    = "task_concurrency_enabled"
 	settingConcurrencyLimit = "task_concurrency_limit"
+	// 实验功能:noa 模型驱动上下文压缩(norma v0.4.0)。默认关闭——开启后平台接入的四类
+	// agent(planner/worker/主 agent/对话)由 noa 接管上下文压缩,取代内置 compaction。
+	// 每 run 读一次,切换只影响之后启动的 run。
+	settingNoaCompaction = "noa_compaction"
 	// defaultWebSearchBackend is used when web search is on but no backend was picked.
 	defaultWebSearchBackend = "ddgs"
 	// deepSeekWebSearchBackend borrows the active LLM profile instead of its own
@@ -488,6 +492,19 @@ func (m *Manager) SetLLMRecordEnabled(on bool) error {
 	return nil
 }
 
+// NoaCompactionEnabled reports whether the experimental noa context-compression
+// mechanism is on (默认关；settings.noa_compaction). Read per agent run via the
+// injected resolver, so a toggle takes effect on the next run without rebuild.
+func (m *Manager) NoaCompactionEnabled() bool {
+	return m.pg.GetBool(settingNoaCompaction, false)
+}
+
+// SetNoaCompaction persists the noa toggle. Effective on the next agent run —
+// the resolver reads it per run, so no rebuild is needed.
+func (m *Manager) SetNoaCompaction(on bool) error {
+	return m.pg.SetBool(settingNoaCompaction, on)
+}
+
 // LLMPoolEnabled reports whether LLM failover ("轮询") is on (默认关；
 // settings.llm_pool_enabled). Read when the provider chain is built (applyLLM),
 // so a change requires a rebuild — putSettings does that.

server/server.go

@@ -428,6 +428,7 @@ func (s *Server) buildPlannerWorker(pinID *int64, gProv llm.Provider, gCfg agent
 	wk.SetConstraintInject(s.constraintInjectWorker) // 操作约束注入 worker(可配置,默认开;每轮读)
 	wk.SetNonStreaming(nonStreamingResolver(wCfg))   // 该 profile 选非流式时走 Provider.Complete
 	wk.SetMaxTokens(maxTokensResolver(wCfg))         // 单次回复输出上限(0 = 不发)
+	wk.SetNoaEnabled(s.m.NoaCompactionEnabled)       // 实验功能:noa 上下文压缩(每 run 读)
 	pProv, pCfg := s.providerForAgent("planner", pinID, gProv, gCfg)
 	pl := agent.NewPlanner(pProv, pCfg.Model, s.m.dir, tx, pCfg.CompactionWindow(), s.agentMaxTurns("planner"))
 	pl.SetFindingRecorder(s.evidenceStore())
@@ -438,6 +439,7 @@ func (s *Server) buildPlannerWorker(pinID *int64, gProv llm.Provider, gCfg agent
 	pl.SetConstraintInject(s.constraintInjectPlanner) // 操作约束注入 planner(可配置,默认开;每轮读)
 	pl.SetNonStreaming(nonStreamingResolver(pCfg))    // 该 profile 选非流式时走 Provider.Complete
 	pl.SetMaxTokens(maxTokensResolver(pCfg))          // 单次回复输出上限(0 = 不发)
+	pl.SetNoaEnabled(s.m.NoaCompactionEnabled)        // 实验功能:noa 上下文压缩(每 run 读)
 	// cold-digest §7: background cold-node compaction, on the planner's provider/
 	// model (§4 uses the running agent's model). Uses Complete (non-streaming) for
 	// the one-shot body summarization.
@@ -509,6 +511,7 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	s.mainAgent.SetSteerWork(s.engine.SteerWork) // steer_work：人对运行中 work 实时纠偏
 	s.mainAgent.SetNonStreaming(nonStreamingResolver(mCfg))
 	s.mainAgent.SetMaxTokens(maxTokensResolver(mCfg))
+	s.mainAgent.SetNoaEnabled(s.m.NoaCompactionEnabled) // 实验功能:noa 上下文压缩(每 run 读)
 	// chat agent serves MANY custom agents by key → it holds the GLOBAL opts
 	// (backend/key) and gates Enabled per-conversation-agent at Chat time. 对话始终用激活配置。
 	s.chatAgent = agent.NewChatAgent(prov, cfg.Model, s.m.dir, tx, win) // chat page runner
@@ -517,6 +520,7 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	s.chatAgent.SetGuard(s.chatGuard())
 	s.chatAgent.SetNonStreaming(nonStreamingResolver(cfg))
 	s.chatAgent.SetMaxTokens(maxTokensResolver(cfg))
+	s.chatAgent.SetNoaEnabled(s.m.NoaCompactionEnabled) // 实验功能:noa 上下文压缩(每 run 读)
 	s.llmProv = prov
 	s.llmCfg = cfg
 	s.llmOn = true
@@ -700,6 +704,7 @@ func (s *Server) chatAgentForProfile(id int64) *agent.ChatAgent {
 	ca.SetGuard(s.chatGuard())
 	ca.SetNonStreaming(nonStreamingResolver(cfg))
 	ca.SetMaxTokens(maxTokensResolver(cfg))
+	ca.SetNoaEnabled(s.m.NoaCompactionEnabled) // 实验功能:noa 上下文压缩(每 run 读)
 	s.profMu.Lock()
 	if ex := s.profChatAgents[id]; ex != nil { // lost the race → keep the winner
 		ca = ex
@@ -3212,6 +3217,9 @@ func (s *Server) settingsPayload() map[string]any {
 		// 操作约束注入范围(默认都开):把本任务的 allow/deny 约束拼进对应 agent 的系统提示。
 		"constraints_inject_planner": s.constraintInjectPlanner(),
 		"constraints_inject_worker":  s.constraintInjectWorker(),
+		// 实验功能:noa 模型驱动上下文压缩(默认关)。开启后平台接入的四类 agent 由 noa
+		// 接管上下文压缩,取代内置 compaction;每 run 读一次,对之后启动的 run 生效。
+		"noa_compaction": s.m.NoaCompactionEnabled(),
 	}
 }
 
@@ -3257,6 +3265,8 @@ func (s *Server) putSettings(w http.ResponseWriter, r *http.Request) {
 		// 操作约束注入范围开关(默认都开);即时生效(planner/worker 每轮读),无需重建 agent。
 		ConstraintsInjectPlanner *bool `json:"constraints_inject_planner"`
 		ConstraintsInjectWorker  *bool `json:"constraints_inject_worker"`
+		// 实验功能:noa 上下文压缩开关(默认关);每 run 读,对之后启动的 run 生效,无需重建 agent。
+		NoaCompaction *bool `json:"noa_compaction"`
 	}
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
 		writeErr(w, 400, err.Error())
@@ -3274,6 +3284,13 @@ func (s *Server) putSettings(w http.ResponseWriter, r *http.Request) {
 			return
 		}
 	}
+	if req.NoaCompaction != nil {
+		// 每 run 读的解析器,切换即时对之后启动的 run 生效,无需 applyLLM 重建。
+		if err := s.m.SetNoaCompaction(*req.NoaCompaction); err != nil {
+			writeErr(w, 500, err.Error())
+			return
+		}
+	}
 	if req.Workers != nil {
 		if err := s.m.SetWorkers(*req.Workers); err != nil {
 			writeErr(w, 400, err.Error())

server/task_llm.go

@@ -566,6 +566,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	wk.SetCompactionWindowResolver(workerRuntime.CompactionWindow)
 	wk.SetNonStreaming(workerRuntime.nonStreaming) // 按任务当前激活 profile 的流式开关(每轮读)
 	wk.SetMaxTokens(workerRuntime.maxTokens)       // 同上,输出上限也跟随当前激活 profile
+	wk.SetNoaEnabled(s.m.NoaCompactionEnabled)     // 实验功能:noa 上下文压缩(平台级开关,每 run 读)
 	wk.SetRunTimeout(time.Duration(s.agentRunSeconds("worker")) * time.Second)
 	wk.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
 	wk.SetWebSearch(s.webSearchFor("worker"))
@@ -575,6 +576,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	pl.SetCompactionWindowResolver(plannerRuntime.CompactionWindow)
 	pl.SetNonStreaming(plannerRuntime.nonStreaming)
 	pl.SetMaxTokens(plannerRuntime.maxTokens)
+	pl.SetNoaEnabled(s.m.NoaCompactionEnabled) // 实验功能:noa 上下文压缩(平台级开关,每 run 读)
 	pl.SetKillWork(s.engine.KillWork)
 	pl.SetSteerWork(s.engine.SteerWork)
 	pl.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
@@ -585,6 +587,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	main.SetCompactionWindowResolver(mainRuntime.CompactionWindow)
 	main.SetNonStreaming(mainRuntime.nonStreaming)
 	main.SetMaxTokens(mainRuntime.maxTokens)
+	main.SetNoaEnabled(s.m.NoaCompactionEnabled) // 实验功能:noa 上下文压缩(平台级开关,每 run 读)
 	main.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
 	main.SetWebSearch(s.webSearchFor("mainagent"))
 	main.SetSteerWork(s.engine.SteerWork) // steer_work：人对运行中 work 实时纠偏

web/src/app/(main)/system/settings/page.tsx

@@ -2,7 +2,7 @@
 
 import * as React from "react";
 
-import { CpuIcon, KeyboardIcon, RadioTowerIcon, SearchIcon, ShieldAlertIcon } from "lucide-react";
+import { CpuIcon, FlaskConicalIcon, KeyboardIcon, RadioTowerIcon, SearchIcon, ShieldAlertIcon } from "lucide-react";
 import { toast } from "sonner";
 
 import { Button } from "@/components/ui/button";
@@ -41,6 +41,8 @@ export default function SystemSettingsPage() {
   // 操作约束注入范围(默认都开)。
   const [injectPlanner, setInjectPlanner] = React.useState(true);
   const [injectWorker, setInjectWorker] = React.useState(true);
+  // 实验功能:noa 上下文压缩(默认关)。
+  const [noaCompaction, setNoaCompaction] = React.useState(false);
   // 纯前端偏好：不走 /api/settings，直接读写 localStorage。
   const sendMode = useChatSendMode();
 
@@ -57,6 +59,7 @@ export default function SystemSettingsPage() {
     setWorkers(String(s.workers ?? 3));
     setInjectPlanner(s.constraints_inject_planner !== false);
     setInjectWorker(s.constraints_inject_worker !== false);
+    setNoaCompaction(!!s.noa_compaction);
   }, []);
 
   const saveWorkers = () => {
@@ -146,6 +149,20 @@ export default function SystemSettingsPage() {
       .catch(() => setInjectWorker(!v)); // revert on failure
   };
 
+  const toggleNoaCompaction = (v: boolean) => {
+    setNoaCompaction(v); // optimistic
+    api
+      .setSettings({ noa_compaction: v })
+      .then((s) => {
+        apply(s);
+        toast.success(v ? "已开启 noa 上下文压缩（对之后启动的运行生效）" : "已关闭 noa 上下文压缩（恢复内置压缩）");
+      })
+      .catch((e) => {
+        setNoaCompaction(!v); // revert on failure
+        toast.error(`保存失败：${(e as Error).message}`);
+      });
+  };
+
   // Persist a web-search patch (enable and/or backend). Optimistic with refetch.
   const saveWebSearch = (patch: Partial<Settings>) => {
     setSaving(true);
@@ -387,6 +404,34 @@ export default function SystemSettingsPage() {
           </CardContent>
         </Card>
 
+        <Card className="mb-4 break-inside-avoid md:mb-6">
+          <CardHeader>
+            <CardTitle className="flex items-center gap-2 text-base">
+              <FlaskConicalIcon className="size-4" />
+              实验功能
+            </CardTitle>
+            <CardDescription>
+              尚在验证中的机制，默认关闭。可能改变 Agent 行为或影响稳定性，请在了解影响后启用。
+              <br />
+              <b>noa 上下文压缩</b>：由模型主动压缩长对话历史（norma v0.4.0）。开启后平台接入的四类 Agent（
+              <b>规划者 / 执行者 / 主 Agent / 对话</b>）改用 noa 接管上下文，取代内置压缩，
+              压缩原文会归档到任务工作目录下便于回溯。切换即时生效（对之后启动的运行生效），无需重建 Agent；
+              关闭后立即恢复内置压缩。
+            </CardDescription>
+          </CardHeader>
+          <CardContent className="flex items-center justify-between gap-4">
+            <Label htmlFor="noa-compaction" className="text-sm font-normal text-muted-foreground">
+              noa 上下文压缩{noaCompaction ? " · 已开启" : " · 已关闭"}
+            </Label>
+            <Switch
+              id="noa-compaction"
+              checked={noaCompaction}
+              disabled={!loaded}
+              onCheckedChange={toggleNoaCompaction}
+            />
+          </CardContent>
+        </Card>
+
         <Card className="mb-4 break-inside-avoid md:mb-6">
           <CardHeader>
             <CardTitle className="flex items-center gap-2 text-base">

web/src/lib/mock/data.ts

@@ -3326,6 +3326,7 @@ export const settings: Settings = {
   workers: 3,
   llm_pool_enabled: true,
   llm_pool_bind_fallback: false,
+  noa_compaction: false,
 };
 
 // ── LLM 轮询（故障转移）──────────────────────────────────────────────────────

web/src/lib/types.ts

@@ -913,6 +913,10 @@ export interface Settings {
   // 操作约束注入范围(默认都开):把任务的 allow/deny 约束拼进对应 agent 的系统提示。
   constraints_inject_planner?: boolean;
   constraints_inject_worker?: boolean;
+  // 实验功能:noa 模型驱动上下文压缩(默认关)。开启后平台接入的四类 agent(planner/
+  // worker/主 agent/对话)由 noa 接管上下文压缩,取代内置 compaction;每 run 读一次,对
+  // 之后启动的 run 生效。
+  noa_compaction?: boolean;
 }
 
 // ---- LLM config ----