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)