feat(llm): 每个 LLM 配置支持自定义会话头

新增 session_header_key:非空时每次 LLM 请求带上该 HTTP 头,头值为当前
运行的 session id(chat 会话=conv-<id>,worker=exp<x>-worker-i<intent>
等)。用于某些按 session-id 头做提示缓存/粘性路由的网关。

实现:在已有的 quotaAwareTransport 里从请求 context 读取 session id
(transcript.WithSessionID 挂上的)并注入,无需改 norma;同一共享 provider
也能按会话发出不同头值。DB 加 session_header_key 列(带旧库迁移),前端
LLM 配置弹窗新增输入框。

agent/provider.go

@@ -25,6 +25,7 @@ import (
 	"github.com/Autumn-27/norma/compaction"
 	"github.com/Autumn-27/norma/llm"
 	acperm "github.com/Autumn-27/norma/permission"
+	"github.com/Autumn-27/norma/transcript"
 )
 
 // Config describes the LLM backend resolved from the environment.
@@ -67,6 +68,12 @@ type Config struct {
 	// 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
+	// SessionHeaderKey,非空时,让每次 LLM 请求带上一个自定义 HTTP 头,头名为该值、
+	// 头值为【当前会话的 session id】(chat 会话=conv-<id>,worker=exp<x>-worker-i<intent>
+	// 等,见 WorkerSessionID)。用于某些按 session-id 头做提示缓存/粘性路由的网关。
+	// 空 = 不发送。值由 transcript.WithSessionID 挂在请求 context 上,由 RoundTripper
+	// 读取填入,因此同一共享 provider 也能按会话发出不同的头值。
+	SessionHeaderKey string
 }
 
 // compaction window resolution bounds (in K tokens). Below the floor the
@@ -224,7 +231,7 @@ func (c Config) Provider() string {
 // limiter lives on the single provider instance — so planner + all workers +
 // main agent (which share this provider) are bounded by one shared rate limit.
 func (c Config) NewProvider() (llm.Provider, error) {
-	client, err := quotaAwareHTTPClient(c.Proxy)
+	client, err := quotaAwareHTTPClient(c.Proxy, c.SessionHeaderKey)
 	if err != nil {
 		return nil, err
 	}
@@ -292,9 +299,24 @@ func IsQuotaExhaustedMessage(message string) bool {
 // retry loop treats every 429 as transient; normalizing only that response to
 // 402 lets a task router fail over immediately while retaining the original
 // response body for provider-specific classification and audit logs.
-type quotaAwareTransport struct{ base http.RoundTripper }
+type quotaAwareTransport struct {
+	base http.RoundTripper
+	// sessionHeaderKey, when non-empty, is the HTTP header name each request
+	// carries; its value is the session id read from the request context. Empty
+	// disables it. See Config.SessionHeaderKey.
+	sessionHeaderKey string
+}
 
 func (t quotaAwareTransport) RoundTrip(req *http.Request) (*http.Response, error) {
+	// Custom session-id header: name is user-configured, value is THIS run's
+	// session id (norma stashes it on the context via transcript.WithSessionID).
+	// Stable across a session's turns and distinct across sessions — exactly what
+	// a session-keyed prompt cache wants. Skipped when no session id is present.
+	if t.sessionHeaderKey != "" {
+		if sid := transcript.SessionIDFrom(req.Context()); sid != "" {
+			req.Header.Set(t.sessionHeaderKey, sid)
+		}
+	}
 	// When LLM recording is on, the Recorder puts a Capture on the context so the
 	// raw wire bodies can be persisted. This is the only layer that still sees
 	// them: norma builds the request body internally and decodes the SSE response
@@ -347,7 +369,7 @@ func requestBodySnapshot(req *http.Request) string {
 	return string(b)
 }
 
-func quotaAwareHTTPClient(proxy string) (*http.Client, error) {
+func quotaAwareHTTPClient(proxy, sessionHeaderKey string) (*http.Client, error) {
 	transport := http.DefaultTransport.(*http.Transport).Clone()
 	proxy = strings.TrimSpace(proxy)
 	if proxy == "" {
@@ -366,7 +388,7 @@ func quotaAwareHTTPClient(proxy string) (*http.Client, error) {
 		}
 		transport.Proxy = http.ProxyURL(proxyURL)
 	}
-	return &http.Client{Transport: quotaAwareTransport{base: transport}}, nil
+	return &http.Client{Transport: quotaAwareTransport{base: transport, sessionHeaderKey: strings.TrimSpace(sessionHeaderKey)}}, nil
 }
 
 // logTestConnection prints the raw HTTP status code(s) and response body of a

agent/provider_capture_test.go

@@ -21,7 +21,7 @@ func TestRoundTripCapturesRawBodies(t *testing.T) {
 	}))
 	defer srv.Close()
 
-	client, err := quotaAwareHTTPClient("")
+	client, err := quotaAwareHTTPClient("", "")
 	if err != nil {
 		t.Fatalf("client: %v", err)
 	}
@@ -65,7 +65,7 @@ func TestRoundTripCaptures429BodyAlongsideQuotaRewrite(t *testing.T) {
 	}))
 	defer srv.Close()
 
-	client, err := quotaAwareHTTPClient("")
+	client, err := quotaAwareHTTPClient("", "")
 	if err != nil {
 		t.Fatalf("client: %v", err)
 	}
@@ -106,7 +106,7 @@ func TestRoundTripWithoutCaptureIsUnchanged(t *testing.T) {
 	}))
 	defer srv.Close()
 
-	client, err := quotaAwareHTTPClient("")
+	client, err := quotaAwareHTTPClient("", "")
 	if err != nil {
 		t.Fatalf("client: %v", err)
 	}

agent/session_header_test.go

@@ -0,0 +1,66 @@
+package agent
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"github.com/Autumn-27/norma/transcript"
+)
+
+// fakeRT records the request it saw and returns a minimal 200 response.
+type fakeRT struct{ seen *http.Request }
+
+func (f *fakeRT) RoundTrip(req *http.Request) (*http.Response, error) {
+	f.seen = req
+	return &http.Response{
+		StatusCode: 200,
+		Header:     make(http.Header),
+		Body:       io.NopCloser(strings.NewReader("")),
+	}, nil
+}
+
+func newReq(ctx context.Context) *http.Request {
+	req, _ := http.NewRequestWithContext(ctx, "POST", "https://api.example.com/v1/messages", strings.NewReader("{}"))
+	return req
+}
+
+func TestSessionHeaderInjectedFromContext(t *testing.T) {
+	base := &fakeRT{}
+	rt := quotaAwareTransport{base: base, sessionHeaderKey: "x-session-id"}
+	ctx := transcript.WithSessionID(context.Background(), "conv-42")
+	if _, err := rt.RoundTrip(newReq(ctx)); err != nil {
+		t.Fatalf("RoundTrip: %v", err)
+	}
+	if got := base.seen.Header.Get("x-session-id"); got != "conv-42" {
+		t.Fatalf("x-session-id = %q, want conv-42", got)
+	}
+}
+
+func TestSessionHeaderSkippedWhenKeyEmpty(t *testing.T) {
+	base := &fakeRT{}
+	rt := quotaAwareTransport{base: base} // no key configured
+	ctx := transcript.WithSessionID(context.Background(), "conv-42")
+	if _, err := rt.RoundTrip(newReq(ctx)); err != nil {
+		t.Fatalf("RoundTrip: %v", err)
+	}
+	// The header name is whatever the user would have set; with no key, nothing
+	// session-related is added. Assert the common key stays absent.
+	if got := base.seen.Header.Get("x-session-id"); got != "" {
+		t.Fatalf("unexpected session header %q with empty key", got)
+	}
+}
+
+func TestSessionHeaderSkippedWhenNoSessionID(t *testing.T) {
+	base := &fakeRT{}
+	rt := quotaAwareTransport{base: base, sessionHeaderKey: "x-session-id"}
+	// Context carries no session id (transcript persistence off).
+	if _, err := rt.RoundTrip(newReq(context.Background())); err != nil {
+		t.Fatalf("RoundTrip: %v", err)
+	}
+	if got := base.seen.Header.Get("x-session-id"); got != "" {
+		t.Fatalf("x-session-id = %q, want empty when no session id on context", got)
+	}
+}

db/config.go

@@ -55,19 +55,25 @@ type LLMProfile struct {
 	// reasoning tokens plus visible output. Ignored by anthropic and
 	// openai-responses, which name the field themselves.
 	MaxTokensField string `json:"max_tokens_field"`
+	// SessionHeaderKey, when non-empty, names a custom HTTP header sent on every
+	// request built from this profile; its value is the current run's session id
+	// (chat conversation / worker intent). For gateways that key prompt caching
+	// or sticky routing off a session-id header. "" = not sent. Maps to
+	// agent.Config.SessionHeaderKey.
+	SessionHeaderKey string `json:"session_header_key"`
 }
 
 // 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),COALESCE(max_tokens,0),COALESCE(max_tokens_field,'')`
+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,''),COALESCE(session_header_key,'')`
-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,'')`
+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,''),COALESCE(session_header_key,'')`
 
 // 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.MaxTokens, &p.MaxTokensField)
+		&p.MaxTokens, &p.MaxTokensField, &p.SessionHeaderKey)
 }
 
 func (d *DB) ListProfiles() ([]*LLMProfile, error) {
@@ -140,18 +146,18 @@ 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,max_tokens,max_tokens_field)
+		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,session_header_key)
-VALUES ($1,$2,NULLIF($3,''),NULLIF($4,''),$5,NULLIF($6,''),NULLIF($7,''),$8,$9,$10,$11,$12,$13,$14,$15,$16,$17) RETURNING id`,
+VALUES ($1,$2,NULLIF($3,''),NULLIF($4,''),$5,NULLIF($6,''),NULLIF($7,''),$8,$9,$10,$11,$12,$13,$14,$15,$16,$17,NULLIF($18,'')) 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)
+			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.SessionHeaderKey).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,max_tokens=$14,max_tokens_field=$15 WHERE id=$16`,
+		_, 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,session_header_key=NULLIF($16,'') WHERE id=$17`,
-			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)
+			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.SessionHeaderKey, 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,max_tokens=$16,max_tokens_field=$17 WHERE id=$18`,
+	_, 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,session_header_key=NULLIF($18,'') WHERE id=$19`,
-		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)
+		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.SessionHeaderKey, p.ID)
 	return p.ID, err
 }
 

db/schema.sql

@@ -321,6 +321,9 @@ CREATE TABLE IF NOT EXISTS llm_profiles (
     --                             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 '',
+    -- 自定义会话头：非空时每次请求带一个该名字的 HTTP 头，头值=当前运行的 session id
+    -- (chat 会话/worker 意图)。用于某些按 session-id 头做提示缓存/粘性路由的网关。''=不发送。
+    session_header_key TEXT NOT NULL DEFAULT '',
     created_at       TIMESTAMPTZ NOT NULL DEFAULT now(),
     updated_at       TIMESTAMPTZ NOT NULL DEFAULT now()
 );
@@ -350,6 +353,8 @@ ALTER TABLE llm_profiles ADD  CONSTRAINT llm_profiles_max_tokens_field_check
 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);
+-- 自定义会话头名；补旧库。默认 '' = 不发送，旧配置行为不变。
+ALTER TABLE llm_profiles ADD COLUMN IF NOT EXISTS session_header_key TEXT NOT NULL DEFAULT '';
 
 -- 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，

server/dto.go

@@ -587,28 +587,32 @@ type LLMProfileDTO struct {
 	// (''=max_tokens | 'max_completion_tokens'，仅 openai 格式有意义)。
 	MaxTokens      int    `json:"max_tokens"`
 	MaxTokensField string `json:"max_tokens_field"`
+	// 自定义会话头名：非空时每次请求带该 HTTP 头，头值=当前会话/意图的 session id。
+	// ''=不发送。用于按 session-id 头做提示缓存/粘性路由的网关。
+	SessionHeaderKey string `json:"session_header_key"`
 }
 
 func llmProfileDTO(p *db.LLMProfile) LLMProfileDTO {
 	return LLMProfileDTO{
-		ID:              i64s(p.ID),
+		ID:               i64s(p.ID),
-		Name:            p.Name,
+		Name:             p.Name,
-		Format:          p.Format,
+		Format:           p.Format,
-		BaseURL:         p.BaseURL,
+		BaseURL:          p.BaseURL,
-		Proxy:           p.Proxy,
+		Proxy:            p.Proxy,
-		Model:           p.Model,
+		Model:            p.Model,
-		APIKeyHint:      p.APIKeyHint,
+		APIKeyHint:       p.APIKeyHint,
-		RatePerSecond:   p.RatePerSecond,
+		RatePerSecond:    p.RatePerSecond,
-		RatePerMinute:   p.RatePerMinute,
+		RatePerMinute:    p.RatePerMinute,
-		ContextWindowK:  p.ContextWindowK,
+		ContextWindowK:   p.ContextWindowK,
-		ThinkingType:    p.ThinkingType,
+		ThinkingType:     p.ThinkingType,
-		ReasoningEffort: p.ReasoningEffort,
+		ReasoningEffort:  p.ReasoningEffort,
-		IsDefault:       p.IsDefault,
+		IsDefault:        p.IsDefault,
-		Priority:        p.Priority,
+		Priority:         p.Priority,
-		PoolExclude:     p.PoolExclude,
+		PoolExclude:      p.PoolExclude,
-		Streaming:       p.Streaming,
+		Streaming:        p.Streaming,
-		MaxTokens:       p.MaxTokens,
+		MaxTokens:        p.MaxTokens,
-		MaxTokensField:  p.MaxTokensField,
+		MaxTokensField:   p.MaxTokensField,
+		SessionHeaderKey: p.SessionHeaderKey,
 	}
 }
 

server/server.go

@@ -313,6 +313,7 @@ func (s *Server) loadLLMConfig() (agent.Config, bool) {
 	cfg.ReasoningEffort = p.ReasoningEffort
 	cfg.Stream = p.Streaming
 	cfg.MaxTokens, cfg.MaxTokensField = p.MaxTokens, p.MaxTokensField
+	cfg.SessionHeaderKey = p.SessionHeaderKey
 	if cfg.APIKey == "" {
 		return cfg, false
 	}
@@ -336,11 +337,13 @@ func (s *Server) saveLLMConfig(cfg agent.Config) error {
 	streaming := true // 旧库/新建默认流式
 	var maxTokens int
 	var maxTokensField string
+	var sessionHeaderKey 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
+				sessionHeaderKey = p.SessionHeaderKey
 				break
 			}
 		}
@@ -350,7 +353,7 @@ 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,
+		MaxTokens: maxTokens, MaxTokensField: maxTokensField, SessionHeaderKey: sessionHeaderKey,
 	})
 	if err != nil {
 		return err
@@ -508,6 +511,7 @@ func (s *Server) loadProfileConfig(id int64) (agent.Config, bool) {
 	cfg.ReasoningEffort = p.ReasoningEffort
 	cfg.Stream = p.Streaming
 	cfg.MaxTokens, cfg.MaxTokensField = p.MaxTokens, p.MaxTokensField
+	cfg.SessionHeaderKey = p.SessionHeaderKey
 	if cfg.APIKey == "" {
 		return cfg, false
 	}

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

@@ -324,6 +324,7 @@ function ProfileSheet({
   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 [sessionHeaderKey, setSessionHeaderKey] = React.useState(""); // 自定义会话头名;空=不发送
   const [testing, setTesting] = React.useState(false);
   const [saving, setSaving] = React.useState(false);
   const [models, setModels] = React.useState<string[]>([]);
@@ -351,6 +352,7 @@ function ProfileSheet({
     setStreaming(profile?.streaming ?? true);
     setMaxTokens(String(profile?.max_tokens ?? 0));
     setMaxTokensField(fromStore(profile?.max_tokens_field));
+    setSessionHeaderKey(profile?.session_header_key ?? "");
     setApiKey("");
     setKeyHint(profile?.api_key_hint ?? "");
     setModels([]);
@@ -437,6 +439,7 @@ function ProfileSheet({
         // 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) : "",
+        session_header_key: sessionHeaderKey.trim(),
       });
       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");
@@ -570,6 +573,22 @@ function ProfileSheet({
             </p>
           </div>
 
+          <div className="grid gap-2">
+            <Label htmlFor="p-session-header">自定义会话头（可选）</Label>
+            <Input
+              id="p-session-header"
+              className="font-mono"
+              placeholder="如 x-session-id（留空=不发送）"
+              value={sessionHeaderKey}
+              onChange={(e) => setSessionHeaderKey(e.target.value)}
+            />
+            <p className="text-muted-foreground text-xs">
+              填写头名后，每次请求都会带上这个 HTTP 头，头值自动填为{" "}
+              <b>当前会话的 session id</b>（chat 会话如 conv-12、worker 如 exp3-worker-i87）。用于按 session-id
+              头做提示缓存 / 粘性路由的网关；同一会话多轮稳定、不同会话互不相同。留空则不发送。
+            </p>
+          </div>
+
           <div className="grid gap-2">
             <Label htmlFor="p-api-key">API Key</Label>
             <Input

web/src/lib/api.ts

@@ -723,6 +723,7 @@ export const api = {
     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）
+    session_header_key?: string; // 非空=每次请求带该 HTTP 头，头值=当前会话 session id；""=不发送
   }) => 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/types.ts

@@ -913,6 +913,9 @@ export interface LLMProfile {
   // 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;
+  // 自定义会话头名：非空时每次请求带该 HTTP 头，头值=当前会话/意图的 session id。
+  // ""=不发送。用于按 session-id 头做提示缓存/粘性路由的网关。
+  session_header_key?: string;
 }
 
 // ---- LLM 轮询（故障转移）----