feat(llm): 模型配置新增输出上限,支持 max_completion_tokens 字段名

llm_profiles 加两列:max_tokens(单次回复输出上限,0=不发送该字段,由服务端
默认值决定)与 max_tokens_field(上限发哪个键)。后者仅对 format=openai 生效:
空(默认)= max_tokens;'max_completion_tokens' = 新键。OpenAI 推理模型
(o 系列/GPT-5)拒收 max_tokens、只认新键,而多数兼容网关反过来只认旧键,
故做成显式开关而非自动推断。anthropic 与 openai-responses 自带字段名,
存下来只会误导,保存时一并清空。

顺带补上一处从未接通的链路:planner/worker/chat/mainagent/goals 五个
agentcore.Options 都没设过 MaxTokens,恒为 0——openai 系因 omitempty 压根
不发上限,anthropic 走 norma 的 8192 兜底。光加配置项而不接这条线,用户在
后台填了数字也不会生效,故本次一并接通,取值与 streaming 一样按 profile
每轮解析,故障转移切换后下一轮即生效。

task_llm.go 里 nonStreaming() 的来源优先级(agent 绑定 → 任务链激活 profile
→ 全局)抽成共用的 activeCfg(),避免 maxTokens() 再抄一份同样的三级回退。

norma 升到 v0.3.3(该版本提供 Config.MaxTokensField)。go.mod 只动 norma 这
一行:仓库此前未 tidy 过,跑 go mod tidy 会顺带升 x/crypto、x/net 等一批无关
依赖,不混进本次改动。

迁移:新列走 ADD COLUMN IF NOT EXISTS + 默认 0/'',旧库升级后行为完全不变;
CHECK 约束限定取值,每次启动先删后建保持幂等。

agent/chat.go

@@ -32,6 +32,7 @@ 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)
+	maxTokensFn    func() int   // resolver: per-reply output cap (nil/0 = send no cap)
 }
 
 func NewChatAgent(prov llm.Provider, model, workDir string, tx *transcript.Store, window int) *ChatAgent {
@@ -44,6 +45,17 @@ func (c *ChatAgent) SetNonStreaming(fn func() bool) { c.nonStreamingFn = fn }
 
 func (c *ChatAgent) nonStreaming() bool { return c.nonStreamingFn != nil && c.nonStreamingFn() }
 
+// 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 }
+
+func (c *ChatAgent) maxTokens() int {
+	if c.maxTokensFn == nil {
+		return 0
+	}
+	return c.maxTokensFn()
+}
+
 // SetProxy points the chat agent's WebFetch/Bash at the recording proxy plus the
 // CA cert it trusts (empty addr = direct). Kept for parity with the other agents.
 func (c *ChatAgent) SetProxy(addr, caCert string) { c.proxyAddr, c.proxyCACert = addr, caCert }
@@ -135,6 +147,7 @@ func (c *ChatAgent) Chat(ctx context.Context, agentKey, sessionID, message strin
 		// (自定义 agent 各自一份;留空/0 用通用默认:10 轮)。
 		Settlement:   wrapupSettlement(agentKey, nil),
 		NonStreaming: c.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:    c.maxTokens(),   // 0 = 不发上限,由服务端默认值决定
 	}
 	if c.guard != nil {
 		opts.Hooks = c.guard.Hooks()

agent/goals.go

@@ -98,13 +98,14 @@ func DecomposeGoals(ctx context.Context, prov llm.Provider, dataDir, goalText, d
 	if prov == nil {
 		return nil
 	}
-	return DecomposeGoalsWithProvider(ctx, prov, dataDir, goalText, desc, as, ts, taskID, false, emit)
+	return DecomposeGoalsWithProvider(ctx, prov, dataDir, goalText, desc, as, ts, taskID, false, 0, emit)
 }
 
 // DecomposeGoalsWithProvider is the task-runtime variant used when a task has an
 // ordered provider chain. It preserves the same tools and write behavior while
-// letting the caller own provider selection/failover.
-func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, nonStreaming bool, emit func(db.Activity)) []GoalSpec {
+// letting the caller own provider selection/failover. maxTokens is the profile's
+// per-reply output cap (0 = send none).
+func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, nonStreaming bool, maxTokens int, emit func(db.Activity)) []GoalSpec {
 	if prov == nil {
 		return nil
 	}
@@ -147,6 +148,7 @@ func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir,
 		// 3 步(抽约束 → 登记范围 → 拆目标)各需一次工具调用,给足回合避免收尾前漏调 set_goals。
 		MaxTurns:     8,
 		NonStreaming: nonStreaming, // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:    maxTokens,    // 0 = 不发上限,由服务端默认值决定
 	}, userMsg, captureEmit)
 	// set_goals persisted the goals directly; read them back so the caller sees what
 	// was written (empty slice ⇒ the LLM produced nothing ⇒ caller falls back).

agent/mainagent.go

@@ -29,6 +29,7 @@ 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)
+	maxTokensFn    func() int                             // resolver: per-reply output cap (nil/0 = send no cap)
 }
 
 // SetNonStreaming wires a resolver deciding whether runs use the non-streaming
@@ -37,6 +38,17 @@ func (m *MainAgent) SetNonStreaming(fn func() bool) { m.nonStreamingFn = fn }
 
 func (m *MainAgent) nonStreaming() bool { return m.nonStreamingFn != nil && m.nonStreamingFn() }
 
+// 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 (m *MainAgent) SetMaxTokens(fn func() int) { m.maxTokensFn = fn }
+
+func (m *MainAgent) maxTokens() int {
+	if m.maxTokensFn == nil {
+		return 0
+	}
+	return m.maxTokensFn()
+}
+
 func NewMainAgent(prov llm.Provider, model, workDir string, tx *transcript.Store, window, maxTurns int) *MainAgent {
 	return &MainAgent{prov: prov, model: model, workDir: workDir, tx: tx, window: window, maxTurns: maxTurns}
 }
@@ -137,6 +149,7 @@ func (m *MainAgent) Chat(ctx context.Context, taskID int64, as *db.AssetStore, t
 		// 命中预算(步数)→ SDK 跑收尾:向用户输出一句进展总结。Prompt 与收尾轮数可后台编辑(默认 10 轮)。
 		Settlement:   wrapupSettlement("mainagent", nil),
 		NonStreaming: m.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:    m.maxTokens(),   // 0 = 不发上限,由服务端默认值决定
 	}
 	if m.tx != nil { // persist raw human↔AI conversation; one accumulating file per task
 		opts.Transcript = m.tx

agent/planner.go

@@ -34,6 +34,7 @@ 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)
+	maxTokensFn       func() int                             // resolver: per-reply output cap (nil/0 = send no cap)
 
 	// todos keeps ONE plan-scratchpad per task (keyed by exploration id) so the
 	// planner's multi-step plan survives across wake-ups — each Plan() is a fresh
@@ -55,6 +56,17 @@ func (p *Planner) SetNonStreaming(fn func() bool) { p.nonStreamingFn = fn }
 
 func (p *Planner) nonStreaming() bool { return p.nonStreamingFn != nil && p.nonStreamingFn() }
 
+// 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 }
+
+func (p *Planner) maxTokens() int {
+	if p.maxTokensFn == nil {
+		return 0
+	}
+	return p.maxTokensFn()
+}
+
 func (p *Planner) compactionWindow() int {
 	if p.windowFn != nil {
 		return p.windowFn()
@@ -366,6 +378,7 @@ func (p *Planner) Plan(ctx context.Context, taskID int64, as *db.AssetStore, ts
 		// clamped(被任务 deadline 夹逼)时改用 PromptByReason(见 wrapupSettlementForTask)。
 		Settlement:   settle,
 		NonStreaming: p.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:    p.maxTokens(),   // 0 = 不发上限,由服务端默认值决定
 	}
 	if p.tx != nil { // persist raw LLM conversation; one accumulating file per task's planner
 		opts.Transcript = p.tx

agent/provider.go

@@ -58,6 +58,15 @@ type Config struct {
 	// 某些网关糟糕的 SSE 实现(空帧、思考字段丢帧),代价是失去运行中的实时进度/实时
 	// token 计数。映射为 agentcore.Options.NonStreaming = !Stream。
 	Stream bool
+	// MaxTokens 是单次回复的输出上限(token)。0 = 不发送该字段,由服务端默认值决定
+	// (历史行为)。与 ContextWindowK 不同:后者是模型总容量,只在本地用来算压缩阈值,
+	// 不出现在请求里;本值随每次请求发出。映射为 agentcore.Options.MaxTokens。
+	MaxTokens int
+	// MaxTokensField 选择 MaxTokens 用哪个请求字段名,仅对 format=openai 生效:
+	//   "" = max_tokens(默认); "max_completion_tokens" = 新字段。
+	// OpenAI 推理模型(o 系列/GPT-5)只认后者,收到 max_tokens 会直接报
+	// unsupported_parameter;而多数兼容网关只认前者,故不做自动推断,交由用户按端点选。
+	MaxTokensField string
 }
 
 // compaction window resolution bounds (in K tokens). Below the floor the
@@ -230,6 +239,9 @@ func (c Config) NewProvider() (llm.Provider, error) {
 	// 可只发 thinking.type、只发 effort、都发、或都不发。
 	lc.ThinkingType = c.ThinkingType
 	lc.ReasoningEffort = c.ReasoningEffort
+	// 输出上限的字段名选择(空 = 用 max_tokens)。上限的「值」不在这里:它每轮随
+	// agentcore.Options.MaxTokens 走,provider 只决定把它塞进哪个键。
+	lc.MaxTokensField = c.MaxTokensField
 	if c.RatePerSecond > 0 || c.RatePerMinute > 0 {
 		lc.RateLimit = &llm.RateLimit{PerSecond: c.RatePerSecond, PerMinute: c.RatePerMinute}
 	}

agent/worker.go

@@ -67,6 +67,9 @@ type Worker struct {
 	// path. Read per run so a profile/task toggle takes effect without rebuilding
 	// the agent. nil = streaming (default).
 	nonStreamingFn 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
 }
 
 // WorkerSessionID returns the stable transcript key used by a worker intent.
@@ -100,6 +103,17 @@ func (w *Worker) SetNonStreaming(fn func() bool) { w.nonStreamingFn = fn }
 
 func (w *Worker) nonStreaming() bool { return w.nonStreamingFn != nil && w.nonStreamingFn() }
 
+// 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 }
+
+func (w *Worker) maxTokens() int {
+	if w.maxTokensFn == nil {
+		return 0
+	}
+	return w.maxTokensFn()
+}
+
 // SetConstraintInject wires a resolver deciding whether this task's operation
 // constraints get injected into the worker system prompt. nil = inject (default).
 func (w *Worker) SetConstraintInject(fn func() bool) { w.injectConstraints = fn }
@@ -386,6 +400,7 @@ func (w *Worker) execute(ctx context.Context, name string, taskID int64, as *db.
 		Compaction:    compactionConfig(w.compactionWindow()), // long tool-heavy runs stay within the window
 		Todos:         actool.NewTodoStore(),                  // 会话级临时待办（TodoWrite），纯规划用，退出即丢
 		NonStreaming:  w.nonStreaming(),                       // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:     w.maxTokens(),                          // 0 = 不发上限,由服务端默认值决定
 	}
 	if hooks != nil { // typed-nil guard: only set when concrete (avoids harness panic)
 		opts.Hooks = hooks

db/config.go

@@ -44,18 +44,30 @@ type LLMProfile struct {
 	// Provider.Complete). Non-streaming sidesteps flaky gateway SSE at the cost of
 	// live in-run progress. Maps to agent.Config.Stream.
 	Streaming bool `json:"streaming"`
+	// MaxTokens caps a single reply's output in tokens. 0 = send no cap and let
+	// the endpoint's own default apply (the historical behaviour). Unlike
+	// ContextWindowK — the model's total capacity, used locally to size compaction
+	// — this value travels with every request.
+	MaxTokens int `json:"max_tokens"`
+	// MaxTokensField picks the request key carrying MaxTokens, for format
+	// "openai" only: "" = max_tokens (default); "max_completion_tokens" = the
+	// newer key, which OpenAI's reasoning models require and whose budget covers
+	// reasoning tokens plus visible output. Ignored by anthropic and
+	// openai-responses, which name the field themselves.
+	MaxTokensField string `json:"max_tokens_field"`
 }
 
 // profileCols is the read column list (hint variant, no api key) shared by the
 // list query; profileColsKey is the same with api_key for the single-row loads.
-const profileCols = `id,name,format,COALESCE(base_url,''),COALESCE(proxy,''),model,COALESCE(api_key_hint,''),rate_per_second,rate_per_minute,context_window_k,COALESCE(reasoning_effort,''),is_default,priority,pool_exclude,COALESCE(thinking_type,''),COALESCE(streaming,true)`
-const profileColsKey = `id,name,format,COALESCE(base_url,''),COALESCE(proxy,''),model,COALESCE(api_key,''),rate_per_second,rate_per_minute,context_window_k,COALESCE(reasoning_effort,''),is_default,priority,pool_exclude,COALESCE(thinking_type,''),COALESCE(streaming,true)`
+const profileCols = `id,name,format,COALESCE(base_url,''),COALESCE(proxy,''),model,COALESCE(api_key_hint,''),rate_per_second,rate_per_minute,context_window_k,COALESCE(reasoning_effort,''),is_default,priority,pool_exclude,COALESCE(thinking_type,''),COALESCE(streaming,true),COALESCE(max_tokens,0),COALESCE(max_tokens_field,'')`
+const profileColsKey = `id,name,format,COALESCE(base_url,''),COALESCE(proxy,''),model,COALESCE(api_key,''),rate_per_second,rate_per_minute,context_window_k,COALESCE(reasoning_effort,''),is_default,priority,pool_exclude,COALESCE(thinking_type,''),COALESCE(streaming,true),COALESCE(max_tokens,0),COALESCE(max_tokens_field,'')`
 
 // scanProfile reads one row in the profileCols / profileColsKey column order. The
 // 7th column lands in APIKeyHint or APIKey depending on which list the caller used.
 func scanProfile(sc interface{ Scan(...any) error }, into *string, p *LLMProfile) error {
 	return sc.Scan(&p.ID, &p.Name, &p.Format, &p.BaseURL, &p.Proxy, &p.Model, into,
-		&p.RatePerSecond, &p.RatePerMinute, &p.ContextWindowK, &p.ReasoningEffort, &p.IsDefault, &p.Priority, &p.PoolExclude, &p.ThinkingType, &p.Streaming)
+		&p.RatePerSecond, &p.RatePerMinute, &p.ContextWindowK, &p.ReasoningEffort, &p.IsDefault, &p.Priority, &p.PoolExclude, &p.ThinkingType, &p.Streaming,
+		&p.MaxTokens, &p.MaxTokensField)
 }
 
 func (d *DB) ListProfiles() ([]*LLMProfile, error) {
@@ -128,18 +140,18 @@ func (d *DB) SaveProfile(p *LLMProfile) (int64, error) {
 	}
 	if p.ID == 0 {
 		var id int64
-		err := d.QueryRow(`INSERT INTO llm_profiles(name,format,base_url,proxy,model,api_key,api_key_hint,rate_per_second,rate_per_minute,context_window_k,reasoning_effort,priority,pool_exclude,thinking_type,streaming)
-VALUES ($1,$2,NULLIF($3,''),NULLIF($4,''),$5,NULLIF($6,''),NULLIF($7,''),$8,$9,$10,$11,$12,$13,$14,$15) RETURNING id`,
-			p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.APIKey, hint, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming).Scan(&id)
+		err := d.QueryRow(`INSERT INTO llm_profiles(name,format,base_url,proxy,model,api_key,api_key_hint,rate_per_second,rate_per_minute,context_window_k,reasoning_effort,priority,pool_exclude,thinking_type,streaming,max_tokens,max_tokens_field)
+VALUES ($1,$2,NULLIF($3,''),NULLIF($4,''),$5,NULLIF($6,''),NULLIF($7,''),$8,$9,$10,$11,$12,$13,$14,$15,$16,$17) RETURNING id`,
+			p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.APIKey, hint, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming, p.MaxTokens, p.MaxTokensField).Scan(&id)
 		return id, err
 	}
 	if p.APIKey == "" {
-		_, err := d.Exec(`UPDATE llm_profiles SET name=$1,format=$2,base_url=NULLIF($3,''),proxy=NULLIF($4,''),model=$5,rate_per_second=$6,rate_per_minute=$7,context_window_k=$8,reasoning_effort=$9,priority=$10,pool_exclude=$11,thinking_type=$12,streaming=$13 WHERE id=$14`,
-			p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming, p.ID)
+		_, err := d.Exec(`UPDATE llm_profiles SET name=$1,format=$2,base_url=NULLIF($3,''),proxy=NULLIF($4,''),model=$5,rate_per_second=$6,rate_per_minute=$7,context_window_k=$8,reasoning_effort=$9,priority=$10,pool_exclude=$11,thinking_type=$12,streaming=$13,max_tokens=$14,max_tokens_field=$15 WHERE id=$16`,
+			p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming, p.MaxTokens, p.MaxTokensField, p.ID)
 		return p.ID, err
 	}
-	_, err := d.Exec(`UPDATE llm_profiles SET name=$1,format=$2,base_url=NULLIF($3,''),proxy=NULLIF($4,''),model=$5,api_key=$6,api_key_hint=$7,rate_per_second=$8,rate_per_minute=$9,context_window_k=$10,reasoning_effort=$11,priority=$12,pool_exclude=$13,thinking_type=$14,streaming=$15 WHERE id=$16`,
-		p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.APIKey, hint, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming, p.ID)
+	_, err := d.Exec(`UPDATE llm_profiles SET name=$1,format=$2,base_url=NULLIF($3,''),proxy=NULLIF($4,''),model=$5,api_key=$6,api_key_hint=$7,rate_per_second=$8,rate_per_minute=$9,context_window_k=$10,reasoning_effort=$11,priority=$12,pool_exclude=$13,thinking_type=$14,streaming=$15,max_tokens=$16,max_tokens_field=$17 WHERE id=$18`,
+		p.Name, p.Format, p.BaseURL, p.Proxy, p.Model, p.APIKey, hint, p.RatePerSecond, p.RatePerMinute, p.ContextWindowK, p.ReasoningEffort, p.Priority, p.PoolExclude, p.ThinkingType, p.Streaming, p.MaxTokens, p.MaxTokensField, p.ID)
 	return p.ID, err
 }
 

db/config_max_tokens_test.go

@@ -0,0 +1,100 @@
+package db
+
+import "testing"
+
+// TestProfileMaxTokensRoundTrip pins the output-cap columns through the whole
+// read/write surface: both column lists (list vs keyed loads) must carry them,
+// and an update must not drop them. A miscounted column list silently shifts
+// every later Scan target, so this is the test that catches it.
+func TestProfileMaxTokensRoundTrip(t *testing.T) {
+	d, err := Open(testDSN(t))
+	if err != nil {
+		t.Skipf("postgres unavailable (%v) — skipping", err)
+	}
+	// Close via Cleanup, registered FIRST so it runs LAST: cleanups are LIFO, and a
+	// plain `defer d.Close()` would fire before them — the row-deleting cleanups
+	// would then run against a closed pool and silently leave test rows behind.
+	t.Cleanup(func() { d.Close() })
+
+	id, err := d.SaveProfile(&LLMProfile{
+		Name: "t-maxtok", Format: "openai", Model: "m", APIKey: "k",
+		MaxTokens: 8192, MaxTokensField: "max_completion_tokens",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { d.Exec(`DELETE FROM llm_profiles WHERE id=$1`, id) })
+
+	// Keyed single-row load (profileColsKey).
+	p, err := d.ProfileByID(id)
+	if err != nil || p == nil {
+		t.Fatalf("ProfileByID: %v, p=%v", err, p)
+	}
+	if p.MaxTokens != 8192 || p.MaxTokensField != "max_completion_tokens" {
+		t.Fatalf("keyed load: max_tokens=%d field=%q", p.MaxTokens, p.MaxTokensField)
+	}
+
+	// List load (profileCols — the hint variant, a separate column list).
+	ps, err := d.ListProfiles()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var found *LLMProfile
+	for _, x := range ps {
+		if x.ID == id {
+			found = x
+		}
+	}
+	if found == nil {
+		t.Fatal("profile missing from ListProfiles")
+	}
+	if found.MaxTokens != 8192 || found.MaxTokensField != "max_completion_tokens" {
+		t.Fatalf("list load: max_tokens=%d field=%q", found.MaxTokens, found.MaxTokensField)
+	}
+
+	// Update with a blank key takes the "keep existing key" UPDATE branch, which
+	// has its own column list and is the easiest one to forget.
+	p.APIKey = ""
+	p.MaxTokens = 4096
+	p.MaxTokensField = ""
+	if _, err := d.SaveProfile(p); err != nil {
+		t.Fatal(err)
+	}
+	after, err := d.ProfileByID(id)
+	if err != nil || after == nil {
+		t.Fatalf("reload: %v", err)
+	}
+	if after.MaxTokens != 4096 || after.MaxTokensField != "" {
+		t.Fatalf("after update: max_tokens=%d field=%q", after.MaxTokens, after.MaxTokensField)
+	}
+	if after.APIKey != "k" {
+		t.Fatalf("blank key on update must keep the stored one, got %q", after.APIKey)
+	}
+}
+
+// A profile saved without touching the new fields must read back as "no cap,
+// classic field name" — the pre-feature behaviour old rows also get.
+func TestProfileMaxTokensDefaults(t *testing.T) {
+	d, err := Open(testDSN(t))
+	if err != nil {
+		t.Skipf("postgres unavailable (%v) — skipping", err)
+	}
+	// Close via Cleanup, registered FIRST so it runs LAST: cleanups are LIFO, and a
+	// plain `defer d.Close()` would fire before them — the row-deleting cleanups
+	// would then run against a closed pool and silently leave test rows behind.
+	t.Cleanup(func() { d.Close() })
+
+	id, err := d.SaveProfile(&LLMProfile{Name: "t-maxtok-default", Format: "openai", Model: "m", APIKey: "k"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { d.Exec(`DELETE FROM llm_profiles WHERE id=$1`, id) })
+
+	p, err := d.ProfileByID(id)
+	if err != nil || p == nil {
+		t.Fatalf("ProfileByID: %v", err)
+	}
+	if p.MaxTokens != 0 || p.MaxTokensField != "" {
+		t.Fatalf("defaults: max_tokens=%d field=%q, want 0 and \"\"", p.MaxTokens, p.MaxTokensField)
+	}
+}

db/schema.sql

@@ -312,6 +312,15 @@ CREATE TABLE IF NOT EXISTS llm_profiles (
     pool_exclude     BOOLEAN NOT NULL DEFAULT false,
     -- streaming=true(默认)走流式 SSE；false 走真·非流式(stream:false，一次性 JSON)。
     streaming        BOOLEAN NOT NULL DEFAULT true,
+    -- 单次回复的输出上限(token)。0=不发送该字段，由服务端默认值决定——保持既有行为。
+    -- 与 context_window_k(模型总容量，仅本地用于压缩阈值)是两回事：本值会随请求发出。
+    max_tokens       INTEGER NOT NULL DEFAULT 0,
+    -- 输出上限用哪个请求字段名，仅对 format='openai' 生效：
+    --   ''                      = max_tokens(默认，兼容绝大多数网关)
+    --   'max_completion_tokens' = 新字段；OpenAI 推理模型(o 系列/GPT-5)只认它，
+    --                             发 max_tokens 会被 unsupported_parameter 拒绝。
+    -- anthropic(max_tokens 必填)与 openai-responses(max_output_tokens)自带字段名，不受此值影响。
+    max_tokens_field TEXT NOT NULL DEFAULT '',
     created_at       TIMESTAMPTZ NOT NULL DEFAULT now(),
     updated_at       TIMESTAMPTZ NOT NULL DEFAULT now()
 );
@@ -330,6 +339,18 @@ ALTER TABLE llm_profiles DROP CONSTRAINT IF EXISTS llm_profiles_format_check;
 ALTER TABLE llm_profiles ADD  CONSTRAINT llm_profiles_format_check
     CHECK (format IN ('openai','anthropic','openai-responses'));
 
+-- 输出上限及其字段名；补旧库。默认 0 / '' = 不发送上限、沿用 max_tokens 字段名，
+-- 旧配置行为完全不变。
+ALTER TABLE llm_profiles ADD COLUMN IF NOT EXISTS max_tokens       INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE llm_profiles ADD COLUMN IF NOT EXISTS max_tokens_field TEXT    NOT NULL DEFAULT '';
+-- 同 format：先删再建，保证每次启动幂等。
+ALTER TABLE llm_profiles DROP CONSTRAINT IF EXISTS llm_profiles_max_tokens_field_check;
+ALTER TABLE llm_profiles ADD  CONSTRAINT llm_profiles_max_tokens_field_check
+    CHECK (max_tokens_field IN ('','max_completion_tokens'));
+ALTER TABLE llm_profiles DROP CONSTRAINT IF EXISTS llm_profiles_max_tokens_check;
+ALTER TABLE llm_profiles ADD  CONSTRAINT llm_profiles_max_tokens_check
+    CHECK (max_tokens >= 0);
+
 -- 思考开关字段 thinking_type，从旧的单一 reasoning_effort 语义一次性拆分而来。
 -- schema.sql 每次启动都执行，故迁移必须只跑一次：仅当该列尚不存在时才回填，
 -- 否则每次启动都会把用户后来手动设的组合覆盖回去。旧 reasoning_effort 语义：

go.mod

@@ -3,7 +3,7 @@ module github.com/Autumn-27/artex
 go 1.26.3
 
 require (
-	github.com/Autumn-27/norma v0.3.2
+	github.com/Autumn-27/norma v0.3.3
 	github.com/golang-jwt/jwt/v5 v5.3.1
 	github.com/jackc/pgx/v5 v5.10.0
 	github.com/klauspost/compress v1.18.6
@@ -59,4 +59,4 @@ require (
 	modernc.org/libc v1.72.3 // indirect
 	modernc.org/mathutil v1.7.1 // indirect
 	modernc.org/memory v1.11.0 // indirect
-)
+)

go.sum

@@ -1,5 +1,7 @@
 github.com/Autumn-27/norma v0.3.2 h1:3gW84SbSrz8luW8kbi2tv3O7CLdRMQvyylCHj1QUdNk=
 github.com/Autumn-27/norma v0.3.2/go.mod h1:BudWrf7iHl96FgHqFr/HwN3mKYmGjJDt+xL2ZHXfFZw=
+github.com/Autumn-27/norma v0.3.3 h1:nppsOFjg+h7vj1H+IH67wj378Spj32jSAqxnG+XqLts=
+github.com/Autumn-27/norma v0.3.3/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/dto.go

@@ -583,6 +583,10 @@ type LLMProfileDTO struct {
 	// 收发模式：true=流式(SSE) | false=非流式。没有 omitempty —— false 必须出现在
 	// 响应里，否则前端读不到「非流式」，开关会回落成默认的流式。
 	Streaming bool `json:"streaming"`
+	// 单次回复输出上限(0=不发送，由服务端默认值决定)，以及它用哪个请求字段名
+	// (''=max_tokens | 'max_completion_tokens'，仅 openai 格式有意义)。
+	MaxTokens      int    `json:"max_tokens"`
+	MaxTokensField string `json:"max_tokens_field"`
 }
 
 func llmProfileDTO(p *db.LLMProfile) LLMProfileDTO {
@@ -603,6 +607,8 @@ func llmProfileDTO(p *db.LLMProfile) LLMProfileDTO {
 		Priority:        p.Priority,
 		PoolExclude:     p.PoolExclude,
 		Streaming:       p.Streaming,
+		MaxTokens:       p.MaxTokens,
+		MaxTokensField:  p.MaxTokensField,
 	}
 }
 

server/goals.go

@@ -384,7 +384,7 @@ func (s *Server) createGoals(ctx context.Context, t *Task, emit func(db.Activity
 	taskID, _ := strconv.ParseInt(t.ID, 10, 64)
 	s.engine.BeginLLMCall(t.ID)
 	goalRuntime := s.agentsForTask(t).runtime
-	decomposed := agent.DecomposeGoalsWithProvider(ctx, goalRuntime, s.m.dir, t.Goal, t.Description, as, t.Store, taskID, goalRuntime.nonStreaming(), emit)
+	decomposed := agent.DecomposeGoalsWithProvider(ctx, goalRuntime, s.m.dir, t.Goal, t.Description, as, t.Store, taskID, goalRuntime.nonStreaming(), goalRuntime.maxTokens(), emit)
 	s.engine.EndLLMCall(t.ID)
 	for _, g := range decomposed {
 		if strings.TrimSpace(g.Text) != "" {

server/server.go

@@ -312,6 +312,7 @@ func (s *Server) loadLLMConfig() (agent.Config, bool) {
 	cfg.ThinkingType = p.ThinkingType
 	cfg.ReasoningEffort = p.ReasoningEffort
 	cfg.Stream = p.Streaming
+	cfg.MaxTokens, cfg.MaxTokensField = p.MaxTokens, p.MaxTokensField
 	if cfg.APIKey == "" {
 		return cfg, false
 	}
@@ -327,16 +328,19 @@ func (s *Server) saveLLMConfig(cfg agent.Config) error {
 	// (anthropic / openai / openai-responses), matching the DB CHECK constraint.
 	format := cfg.Provider()
 	var id int64
-	// agent.Config carries no failover fields, so carry the stored ones forward —
-	// otherwise this legacy endpoint would silently reset the profile's priority,
-	// pool_exclude, and streaming to zero values on every save.
+	// 这个 legacy 端点的请求体不含轮询/收发/输出上限参数,故把库里已存的值原样带回 —
+	// 否则每次保存都会把 profile 的 priority、pool_exclude、streaming 以及输出上限
+	// (max_tokens / max_tokens_field)悄悄重置成零值。
 	var priority int
 	var poolExclude bool
 	streaming := true // 旧库/新建默认流式
+	var maxTokens int
+	var maxTokensField string
 	if profs, _ := s.m.pg.ListProfiles(); profs != nil {
 		for _, p := range profs {
 			if p.Name == "default" {
 				id, priority, poolExclude, streaming = p.ID, p.Priority, p.PoolExclude, p.Streaming
+				maxTokens, maxTokensField = p.MaxTokens, p.MaxTokensField
 				break
 			}
 		}
@@ -346,6 +350,7 @@ func (s *Server) saveLLMConfig(cfg agent.Config) error {
 		APIKey: cfg.APIKey, RatePerSecond: cfg.RatePerSecond, RatePerMinute: cfg.RatePerMinute,
 		ContextWindowK: cfg.ContextWindowK, ThinkingType: cfg.ThinkingType, ReasoningEffort: cfg.ReasoningEffort, IsDefault: true,
 		Priority: priority, PoolExclude: poolExclude, Streaming: streaming,
+		MaxTokens: maxTokens, MaxTokensField: maxTokensField,
 	})
 	if err != nil {
 		return err
@@ -396,6 +401,7 @@ func (s *Server) buildPlannerWorker(pinID *int64, gProv llm.Provider, gCfg agent
 	wk.SetWebSearch(s.webSearchFor("worker"))
 	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)
 	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
@@ -404,6 +410,7 @@ func (s *Server) buildPlannerWorker(pinID *int64, gProv llm.Provider, gCfg agent
 	pl.SetWebSearch(s.webSearchFor("planner"))
 	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)
 	return pl, wk
 }
 
@@ -415,6 +422,12 @@ func nonStreamingResolver(cfg agent.Config) func() bool {
 	return func() bool { return nonStreaming }
 }
 
+// maxTokensResolver mirrors nonStreamingResolver for the per-reply output cap.
+func maxTokensResolver(cfg agent.Config) func() int {
+	maxTokens := cfg.MaxTokens
+	return func() int { return maxTokens }
+}
+
 // applyLLM (re)builds the planner/worker/main-agent from cfg and installs them on
 // the running engine as the GLOBAL active pair. Safe to call at runtime (UI configures LLM).
 func (s *Server) applyLLM(cfg agent.Config) error {
@@ -459,6 +472,7 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	s.mainAgent.SetWebSearch(s.webSearchFor("mainagent"))
 	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))
 	// 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
@@ -466,6 +480,7 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	s.chatAgent.SetWebSearch(s.m.WebSearchOpts())
 	s.chatAgent.SetGuard(s.chatGuard())
 	s.chatAgent.SetNonStreaming(nonStreamingResolver(cfg))
+	s.chatAgent.SetMaxTokens(maxTokensResolver(cfg))
 	s.llmProv = prov
 	s.llmCfg = cfg
 	s.llmOn = true
@@ -492,6 +507,7 @@ func (s *Server) loadProfileConfig(id int64) (agent.Config, bool) {
 	cfg.ThinkingType = p.ThinkingType
 	cfg.ReasoningEffort = p.ReasoningEffort
 	cfg.Stream = p.Streaming
+	cfg.MaxTokens, cfg.MaxTokensField = p.MaxTokens, p.MaxTokensField
 	if cfg.APIKey == "" {
 		return cfg, false
 	}
@@ -644,6 +660,7 @@ func (s *Server) chatAgentForProfile(id int64) *agent.ChatAgent {
 	ca.SetWebSearch(s.m.WebSearchOpts())
 	ca.SetGuard(s.chatGuard())
 	ca.SetNonStreaming(nonStreamingResolver(cfg))
+	ca.SetMaxTokens(maxTokensResolver(cfg))
 	s.profMu.Lock()
 	if ex := s.profChatAgents[id]; ex != nil { // lost the race → keep the winner
 		ca = ex

server/server_mgmt.go

@@ -25,6 +25,7 @@ import (
 
 	"github.com/Autumn-27/artex/agent"
 	"github.com/Autumn-27/artex/db"
+	"github.com/Autumn-27/norma/llm"
 	"github.com/Autumn-27/norma/skill"
 )
 
@@ -1804,6 +1805,15 @@ func (s *Server) pgSaveProfile(w http.ResponseWriter, r *http.Request) {
 	p := body.LLMProfile
 	p.APIKey = body.APIKey
 	p.Streaming = body.Streaming == nil || *body.Streaming
+	// 输出上限:负数无意义,归零(= 不发送该字段)。字段名开关只有 Chat Completions
+	// 用得上——anthropic 与 openai-responses 各自定死了字段名,存下来只会误导后续读者,
+	// 故非 openai 格式一律清空。未知取值同样清空,避免把 DB CHECK 的报错甩给用户。
+	if p.MaxTokens < 0 {
+		p.MaxTokens = 0
+	}
+	if p.Format != "openai" || p.MaxTokensField != llm.MaxTokensFieldCompletion {
+		p.MaxTokensField = ""
+	}
 	id, err := pg.SaveProfile(&p)
 	if err != nil {
 		writeErr(w, 500, err.Error())

server/task_llm.go

@@ -139,34 +139,46 @@ func (r *taskLLMRuntime) current() (*Task, int64, int64, llm.Provider, error) {
 	return t, 0, pt.LLMChainRevision, prov, nil
 }
 
-// nonStreaming reports whether the task's currently-active LLM source is set to
-// non-streaming. It mirrors current()'s source precedence (agent binding →
-// active chain profile → global) but reads only the streaming flag. Read-only
-// and best-effort: any resolution error falls back to streaming (false), the
-// safe default. If failover switches to a profile with a different streaming
-// setting, the change takes effect on the next agent run (a fresh Session is
-// built per run in captureRun).
-func (r *taskLLMRuntime) nonStreaming() bool {
+// activeCfg resolves the task's currently-active LLM config, mirroring current()'s
+// source precedence (agent binding → active chain profile → global). Read-only and
+// best-effort: ok=false when nothing resolves, leaving the per-setting fallback to
+// the caller. If failover switches profiles, the change takes effect on the next
+// agent run (a fresh Session is built per run in captureRun).
+func (r *taskLLMRuntime) activeCfg() (agent.Config, bool) {
 	taskNum, err := parseTaskID(r.taskID)
 	if err != nil {
-		return false
+		return agent.Config{}, false
 	}
 	pt, err := r.s.m.pg.GetTask(taskNum)
 	if err != nil || pt == nil {
-		return false
+		return agent.Config{}, false
 	}
 	if _, cfg, ok := r.s.agentBindingProvider(r.agentKey); ok {
-		return !cfg.Stream
+		return cfg, true
 	}
 	if len(pt.LLMProfileIDs) > 0 && pt.ActiveLLMProfileID != nil {
 		if _, cfg, ok := r.s.providerForProfile(*pt.ActiveLLMProfileID); ok {
-			return !cfg.Stream
+			return cfg, true
 		}
 	}
 	if _, cfg, ok := r.s.globalProvider(); ok {
-		return !cfg.Stream
+		return cfg, true
 	}
-	return false
+	return agent.Config{}, false
+}
+
+// nonStreaming reports whether the task's currently-active LLM source is set to
+// non-streaming. Unresolvable → streaming (false), the safe default.
+func (r *taskLLMRuntime) nonStreaming() bool {
+	cfg, ok := r.activeCfg()
+	return ok && !cfg.Stream
+}
+
+// maxTokens returns the currently-active source's per-reply output cap.
+// Unresolvable → 0, i.e. send no cap, matching the pre-setting behaviour.
+func (r *taskLLMRuntime) maxTokens() int {
+	cfg, _ := r.activeCfg() // 未解析出配置时是零值 0
+	return cfg.MaxTokens
 }
 
 func parseTaskID(id string) (int64, error) {
@@ -531,6 +543,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	wk := agent.NewWorker(workerRuntime, "task-router", s.m.dir, tx, window, s.agentMaxTurns("worker"))
 	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.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")))
@@ -539,6 +552,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	pl := agent.NewPlanner(plannerRuntime, "task-router", s.m.dir, tx, plannerRuntime.CompactionWindow(), s.agentMaxTurns("planner"))
 	pl.SetCompactionWindowResolver(plannerRuntime.CompactionWindow)
 	pl.SetNonStreaming(plannerRuntime.nonStreaming)
+	pl.SetMaxTokens(plannerRuntime.maxTokens)
 	pl.SetKillWork(s.engine.KillWork)
 	pl.SetSteerWork(s.engine.SteerWork)
 	pl.SetProxy(s.m.ProxyAddr(), s.m.ProxyCACert())
@@ -547,6 +561,7 @@ func (s *Server) agentsForTask(t *Task) *taskAgentBundle {
 	main := agent.NewMainAgent(mainRuntime, "task-router", s.m.dir, tx, mainRuntime.CompactionWindow(), s.agentMaxTurns("mainagent"))
 	main.SetCompactionWindowResolver(mainRuntime.CompactionWindow)
 	main.SetNonStreaming(mainRuntime.nonStreaming)
+	main.SetMaxTokens(mainRuntime.maxTokens)
 	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/llm/page.tsx

@@ -41,6 +41,18 @@ const THINKING_TYPES: { value: string; label: string }[] = [
   { value: "disabled", label: "关闭" },
   { value: "enabled", label: "开启" },
 ];
+// 输出上限用哪个请求字段名（仅 openai 格式有意义）。NONE ↔ "" 走同一套哨兵转换。
+const MAX_TOKENS_FIELDS: { value: string; label: string }[] = [
+  { value: NONE, label: "max_tokens（默认）" },
+  { value: "max_completion_tokens", label: "max_completion_tokens" },
+];
+// 另外两种格式各自定死了字段名，选项对它们无意义，说明文案里直接讲清楚。
+const MAX_TOKENS_FIELD_HINTS: Record<string, string> = {
+  openai:
+    "上限发哪个键。max_tokens 是默认，绝大多数兼容网关只认它；OpenAI 官方推理模型（o 系列 / GPT-5）反过来只认 max_completion_tokens，收到 max_tokens 会直接报 unsupported_parameter。",
+  anthropic: "仅 openai 格式可选。Anthropic 的字段名固定为 max_tokens。",
+  "openai-responses": "仅 openai 格式可选。Responses API 的字段名固定为 max_output_tokens。",
+};
 const EFFORT_LEVELS: { value: string; label: string }[] = [
   { value: NONE, label: "不发送（默认）" },
   { value: "low", label: "low" },
@@ -310,6 +322,8 @@ function ProfileSheet({
   const [priority, setPriority] = React.useState("0"); // 轮询顺位;越大越先
   const [poolExclude, setPoolExclude] = React.useState(false);
   const [streaming, setStreaming] = React.useState(true); // true=流式(默认);false=非流式
+  const [maxTokens, setMaxTokens] = React.useState("0"); // 单次回复输出上限;0=不发送
+  const [maxTokensField, setMaxTokensField] = React.useState(NONE); // 上限用哪个字段名;NONE=max_tokens
   const [testing, setTesting] = React.useState(false);
   const [saving, setSaving] = React.useState(false);
   const [models, setModels] = React.useState<string[]>([]);
@@ -335,6 +349,8 @@ function ProfileSheet({
     setPriority(String(profile?.priority ?? 0));
     setPoolExclude(profile?.pool_exclude ?? false);
     setStreaming(profile?.streaming ?? true);
+    setMaxTokens(String(profile?.max_tokens ?? 0));
+    setMaxTokensField(fromStore(profile?.max_tokens_field));
     setApiKey("");
     setKeyHint(profile?.api_key_hint ?? "");
     setModels([]);
@@ -417,6 +433,10 @@ function ProfileSheet({
         priority: Number(priority) || 0,
         pool_exclude: poolExclude,
         streaming,
+        max_tokens: Math.max(0, Number(maxTokens) || 0),
+        // 字段名开关只对 openai(Chat Completions) 有意义，其它格式一律回落到默认；
+        // 后端也会再做一次同样的归一化，这里只是别让 UI 送出自相矛盾的值。
+        max_tokens_field: format === "openai" ? toStore(maxTokensField) : "",
       });
       if (isNew) toast.success(`已新建：${name.trim()}（在卡片上「设为激活」以启用）`);
       else toast.success(profile?.is_default ? "已保存，激活配置即时生效，无需重启" : "已保存");
@@ -629,6 +649,52 @@ function ProfileSheet({
             </div>
           </div>
 
+          <div className="grid gap-3 rounded-lg border p-3">
+            <div className="flex items-center justify-between gap-4">
+              <div className="grid gap-0.5">
+                <Label htmlFor="p-max-tokens" className="text-sm">
+                  输出上限 · max tokens
+                </Label>
+                <p className="text-muted-foreground text-xs">
+                  单次回复最多生成多少 token，随每次请求发出。0（默认）= 不发送该字段，由服务端默认值决定。
+                  这与上面的「上下文窗口」是两回事：那是模型总容量，只在本地用来算压缩阈值。
+                  设太小会让推理模型在思考阶段就被截断，一个字答案都出不来。
+                </p>
+              </div>
+              <Input
+                id="p-max-tokens"
+                type="number"
+                min={0}
+                className="w-28 shrink-0"
+                value={maxTokens}
+                onChange={(e) => setMaxTokens(e.target.value)}
+                placeholder="0"
+              />
+            </div>
+            <div className="flex items-center justify-between gap-4 border-t pt-3">
+              <div className="grid gap-0.5">
+                <Label className="text-sm">上限字段名</Label>
+                <p className="text-muted-foreground text-xs">{MAX_TOKENS_FIELD_HINTS[format]}</p>
+              </div>
+              <Select
+                value={format === "openai" ? maxTokensField : NONE}
+                onValueChange={setMaxTokensField}
+                disabled={format !== "openai"}
+              >
+                <SelectTrigger className="w-56 shrink-0">
+                  <SelectValue />
+                </SelectTrigger>
+                <SelectContent>
+                  {MAX_TOKENS_FIELDS.map((o) => (
+                    <SelectItem key={o.value} value={o.value}>
+                      {o.label}
+                    </SelectItem>
+                  ))}
+                </SelectContent>
+              </Select>
+            </div>
+          </div>
+
           <div className="grid gap-3 rounded-lg border p-3">
             <div className="flex items-center justify-between gap-4">
               <div className="grid gap-0.5">

web/src/lib/api.ts

@@ -721,6 +721,8 @@ export const api = {
     priority?: number; // 轮询顺位，越大越先用
     pool_exclude?: boolean; // true=不作为故障转移目标
     streaming?: boolean; // true(默认)=流式 | false=非流式
+    max_tokens?: number; // 单次回复输出上限；0=不发送，由服务端默认值决定
+    max_tokens_field?: string; // ""=max_tokens(默认) | "max_completion_tokens"（仅 openai 格式）
   }) => post<{ id: number }>("/llm/profiles", p),
   deleteLLMProfile: (id: string) => del<{ deleted: number }>(`/llm/profiles/${id}`),
   activateLLMProfile: (id: string) => post<{ ok: boolean }>("/llm/profiles/active", { id: Number(id) }),

web/src/lib/mock/data.ts

@@ -3017,6 +3017,9 @@ export const llmProfiles: LLMProfile[] = [
     is_default: true,
     priority: 0,
     pool_exclude: false,
+    // anthropic 的字段名固定，故 max_tokens_field 恒为空。
+    max_tokens: 0,
+    max_tokens_field: "",
   },
   {
     id: "2",
@@ -3033,6 +3036,9 @@ export const llmProfiles: LLMProfile[] = [
     is_default: false,
     priority: 10,
     pool_exclude: false,
+    // openai 格式 + 推理模型：上限走 max_completion_tokens。
+    max_tokens: 8192,
+    max_tokens_field: "max_completion_tokens",
   },
 ];
 

web/src/lib/types.ts

@@ -904,6 +904,12 @@ export interface LLMProfile {
   pool_exclude?: boolean;
   // true（默认）= 流式(SSE) | false = 真·非流式(stream:false，一次性返回)。
   streaming?: boolean;
+  // 单次回复的输出上限(token)。0 = 不发送该字段，由服务端默认值决定。
+  // 注意与 context_window_k 区分：后者是模型总容量，只在本地用于压缩阈值。
+  max_tokens?: number;
+  // 上限用哪个请求字段名，仅 format="openai" 有意义：
+  // ""=max_tokens(默认) | "max_completion_tokens"(OpenAI 推理模型只认它)
+  max_tokens_field?: string;
 }
 
 // ---- LLM 轮询（故障转移）----