admin/ARTEX
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(llm): 每个 LLM 配置支持流式/非流式切换(默认流式)

Browse Source 
依赖 norma v0.3.0 的 Provider.Complete。为每个 llm_profile 增加 streaming
开关:开(默认)走流式 SSE;关走真·非流式(stream:false、一次性返回完整
JSON),可绕开部分网关糟糕的 SSE 实现(空帧、思考字段丢帧),代价是失去
运行中的实时进度与实时 token 计数。

- agent.Config.Stream → agentcore.Options.NonStreaming = !Stream;worker /
  planner / mainagent / chat / goals 经 SetNonStreaming resolver 注入,任务
  路由按当前激活 profile 动态取值
- llmpool.Pool / llmrec.Recorder / taskLLMRuntime 三个 Provider 包装均实现
  Complete(分别带故障转移 / 录制 / 同 provider 重试与额度 failover),使
  输出、token 统计、操作统计、LLM 录制四条链路在非流式下保持兼容
- db: llm_profiles 新增 streaming 列 + ALTER 补旧库(默认 true,旧配置无感)
- 系统 → LLM 配置表单增加「流式输出」开关
This commit is contained in:
Autumn-27 committed 2026-08-25 10:49:47 -04:00
1 parent 19a7206d12
commit 587ba422b4
22 files changed
+409 -75
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

agent/capture_usage_test.go
+11
View File
@@ -19,6 +19,17 @@ func (p captureUsageProvider) Stream(ctx context.Context, _ llm.CompletionReques
return func(yield func(llm.StreamEvent, error) bool) { p.stream(ctx, yield) }
}
func (p captureUsageProvider) Complete(ctx context.Context, req llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
acc := llm.NewAccumulator()
for ev, err := range p.Stream(ctx, req) {
if err != nil {
return llm.Message{}, "", llm.Usage{}, err
}
acc.Add(ev)
}
return acc.Message(), acc.StopReason, acc.Usage, nil
}
func TestCaptureRunPersistsUsageOnProviderFailure(t *testing.T) {
wantErr := errors.New("provider failed after reporting usage")
provider := captureUsageProvider{stream: func(_ context.Context, yield func(llm.StreamEvent, error) bool) {
agent/chat.go
+26 -18
View File
@@ -6,13 +6,13 @@ import (
"path/filepath"
"time"
"github.com/Autumn-27/artex/db"
"github.com/Autumn-27/artex/guard"
"github.com/Autumn-27/norma/agentcore"
"github.com/Autumn-27/norma/llm"
"github.com/Autumn-27/norma/permission"
actool "github.com/Autumn-27/norma/tool"
"github.com/Autumn-27/norma/transcript"
"github.com/Autumn-27/artex/db"
"github.com/Autumn-27/artex/guard"
)
// ChatAgent is the generic, task-independent conversational runner behind the chat
@@ -22,21 +22,28 @@ import (
// LS/Glob/Grep) plus whatever skills/MCP the key is made visible; NO pentest
// graph/task context is injected (that stays exclusive to MainAgent).
type ChatAgent struct {
prov llm.Provider
model string
workDir string
tx *transcript.Store
window int
proxyAddr string
proxyCACert string
webSearch WebSearchOpts
guard *guard.Guard // optional; nil disables intercept hooks for chat
prov llm.Provider
model string
workDir string
tx *transcript.Store
window int
proxyAddr string
proxyCACert string
webSearch WebSearchOpts
guard *guard.Guard // optional; nil disables intercept hooks for chat
nonStreamingFn func() bool // resolver: use non-streaming (Complete) path? (nil = streaming)
}
func NewChatAgent(prov llm.Provider, model, workDir string, tx *transcript.Store, window int) *ChatAgent {
return &ChatAgent{prov: prov, model: model, workDir: workDir, tx: tx, window: window}
}
// SetNonStreaming wires a resolver deciding whether chat runs use the
// non-streaming model path (true = non-streaming). nil/unset = streaming.
func (c *ChatAgent) SetNonStreaming(fn func() bool) { c.nonStreamingFn = fn }
func (c *ChatAgent) nonStreaming() bool { return c.nonStreamingFn != nil && c.nonStreamingFn() }
// 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 }
@@ -115,18 +122,19 @@ func (c *ChatAgent) Chat(ctx context.Context, agentKey, sessionID, message strin
BraveSearchAPIKey: ws.BraveKey,
TavilySearchAPIKey: ws.TavilyKey,
WebSearchProxy: ws.Proxy,
BashEnv: proxyEnv(c.proxyAddr, c.proxyCACert), // Bash 子命令默认走代理+信任 CA
WorkingDir: sessionWorkDir,
MaxTurns: maxTurns,
MaxDuration: maxDuration,
Compaction: compactionConfig(c.window),
Todos: actool.NewTodoStore(),
BashEnv: proxyEnv(c.proxyAddr, c.proxyCACert), // Bash 子命令默认走代理+信任 CA
WorkingDir: sessionWorkDir,
MaxTurns: maxTurns,
MaxDuration: maxDuration,
Compaction: compactionConfig(c.window),
Todos: actool.NewTodoStore(),
// large tool output spills to cmd-output/ under the session dir.
// 截断上限用 SDK 默认(tool.Capture 的 30000 字符)。
ToolOutputDir: filepath.Join(sessionWorkDir, "cmd-output"),
// 命中预算(步数)→ SDK 跑收尾:输出一句总结。Prompt 与收尾轮数按本 agent key 后台可编辑
// (自定义 agent 各自一份;留空/0 用通用默认:10 轮)。
Settlement: wrapupSettlement(agentKey, nil),
Settlement: wrapupSettlement(agentKey, nil),
NonStreaming: c.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
}
if c.guard != nil {
opts.Hooks = c.guard.Hooks()
agent/goals.go
+4 -3
View File
@@ -98,13 +98,13 @@ 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, emit)
return DecomposeGoalsWithProvider(ctx, prov, dataDir, goalText, desc, as, ts, taskID, false, 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, emit func(db.Activity)) []GoalSpec {
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 {
if prov == nil {
return nil
}
@@ -145,7 +145,8 @@ func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir,
PermissionMode: acperm.ModeBypass,
DisableBackgroundTasks: true,
// 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,
MaxTurns: 8,
NonStreaming: nonStreaming, // 该 profile 选非流式时走 Provider.Complete
}, 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
+20 -12
View File
@@ -17,19 +17,26 @@ import (
// (→planner) or direct high-priority intents (→frontier). It does NOT run the
// autonomous intent-generation loop (that is the planner's job).
type MainAgent struct {
prov llm.Provider
model string
tx *transcript.Store // raw LLM conversation persistence (nil = off)
window int // context window in tokens (for compaction)
windowFn func() int // optional dynamic task-chain minimum
maxTurns int // max agent turns per run (0 = unlimited)
proxyAddr string // recording proxy for WebFetch (empty = direct)
proxyCACert string // recording proxy's CA cert path (HTTPS verify)
webSearch WebSearchOpts // web_search tool backend selection (off by default)
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)
prov llm.Provider
model string
tx *transcript.Store // raw LLM conversation persistence (nil = off)
window int // context window in tokens (for compaction)
windowFn func() int // optional dynamic task-chain minimum
maxTurns int // max agent turns per run (0 = unlimited)
proxyAddr string // recording proxy for WebFetch (empty = direct)
proxyCACert string // recording proxy's CA cert path (HTTPS verify)
webSearch WebSearchOpts // web_search tool backend selection (off by default)
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)
}
// 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 }
func (m *MainAgent) nonStreaming() bool { return m.nonStreamingFn != nil && m.nonStreamingFn() }
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}
}
@@ -128,7 +135,8 @@ func (m *MainAgent) Chat(ctx context.Context, taskID int64, as *db.AssetStore, t
Compaction: compactionConfig(m.compactionWindow()), // long chats stay within the window
Todos: actool.NewTodoStore(), // 会话级临时待办（TodoWrite），纯规划用，退出即丢
// 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),
Settlement: wrapupSettlement("mainagent", nil),
NonStreaming: m.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
}
if m.tx != nil { // persist raw human↔AI conversation; one accumulating file per task
opts.Transcript = m.tx
agent/planner.go
+22 -14
View File
@@ -20,19 +20,20 @@ import (
// assets, judges whether the task goal is met, and emits 0..N exploration intents
// into the frontier. It is the sole intent generator.
type Planner struct {
prov llm.Provider
model string
tx *transcript.Store // raw LLM conversation persistence (nil = off)
window int // context window in tokens (for compaction)
windowFn func() int // optional dynamic task-chain minimum
maxTurns int // max agent turns per run (0 = unlimited)
killWork func(intentID int64) error // engine callback to terminate a running work (nil = off)
steerWork func(intentID int64, msg string) error // engine callback to steer a running work mid-run (nil = off)
proxyAddr string // recording proxy for WebFetch (empty = direct)
proxyCACert string // recording proxy's CA cert path (HTTPS verify)
webSearch WebSearchOpts // web_search tool backend selection (off by default)
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)
prov llm.Provider
model string
tx *transcript.Store // raw LLM conversation persistence (nil = off)
window int // context window in tokens (for compaction)
windowFn func() int // optional dynamic task-chain minimum
maxTurns int // max agent turns per run (0 = unlimited)
killWork func(intentID int64) error // engine callback to terminate a running work (nil = off)
steerWork func(intentID int64, msg string) error // engine callback to steer a running work mid-run (nil = off)
proxyAddr string // recording proxy for WebFetch (empty = direct)
proxyCACert string // recording proxy's CA cert path (HTTPS verify)
webSearch WebSearchOpts // web_search tool backend selection (off by default)
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)
// 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
@@ -48,6 +49,12 @@ func NewPlanner(prov llm.Provider, model, workDir string, tx *transcript.Store,
func (p *Planner) SetCompactionWindowResolver(fn func() int) { p.windowFn = fn }
// SetNonStreaming wires a resolver deciding whether runs use the non-streaming
// model path (true = non-streaming). nil/unset = streaming (default).
func (p *Planner) SetNonStreaming(fn func() bool) { p.nonStreamingFn = fn }
func (p *Planner) nonStreaming() bool { return p.nonStreamingFn != nil && p.nonStreamingFn() }
func (p *Planner) compactionWindow() int {
if p.windowFn != nil {
return p.windowFn()
@@ -353,7 +360,8 @@ func (p *Planner) Plan(ctx context.Context, taskID int64, as *db.AssetStore, ts
// 命中【本轮】步数预算→ SDK 跑收尾:把本轮已想清楚的结论落地(该派的 add_intent、
// 能证的 prove_goal、串行链记 TodoWrite),而非停止规划——planner 之后仍会被反复唤醒。
// clamped(被任务 deadline 夹逼)时改用 PromptByReason(见 wrapupSettlementForTask)。
Settlement: settle,
Settlement: settle,
NonStreaming: p.nonStreaming(), // 该 profile 选非流式时走 Provider.Complete
}
if p.tx != nil { // persist raw LLM conversation; one accumulating file per task's planner
opts.Transcript = p.tx
agent/provider.go
+19
View File
@@ -52,6 +52,11 @@ type Config struct {
// "" = 不发送(默认); "low"/"medium"/"high"/"xhigh"/"max" = 对应强度.
// OpenAI 映射为顶层 reasoning_effort;Anthropic 映射为 output_config.effort.
ReasoningEffort string
// Stream 控制该 profile 是否使用流式(SSE)接口。true(默认)= 流式;false = 真·
// 非流式(发 stream:false,一次性拿完整 JSON,走 Provider.Complete)。非流式可绕开
// 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
}
// compaction window resolution bounds (in K tokens). Below the floor the
@@ -117,6 +122,8 @@ func FromEnv() (Config, bool) {
BaseURL: os.Getenv("ARTEX_LLM_BASE_URL"),
Model: os.Getenv("ARTEX_LLM_MODEL"),
Proxy: strings.TrimSpace(os.Getenv("ARTEX_LLM_PROXY")),
// 默认流式;ARTEX_LLM_STREAM=false/0/off 显式关闭走非流式。
Stream: !isFalsy(os.Getenv("ARTEX_LLM_STREAM")),
}
switch prov {
case "openai":
@@ -148,6 +155,7 @@ func ConfigFrom(provider, model, baseURL, apiKey, proxy string) Config {
BaseURL: strings.TrimRight(strings.TrimSpace(baseURL), "/"),
APIKey: strings.TrimSpace(apiKey),
Proxy: strings.TrimSpace(proxy),
Stream: true, // 默认流式;调用方按 profile 覆盖
}
switch strings.TrimSpace(provider) {
case "openai":
@@ -168,6 +176,16 @@ func ConfigFrom(provider, model, baseURL, apiKey, proxy string) Config {
return c
}
// isFalsy reports whether an env-var string explicitly requests "off". Empty or
// unrecognized → false (so an unset var keeps the streaming default).
func isFalsy(s string) bool {
switch strings.ToLower(strings.TrimSpace(s)) {
case "0", "false", "off", "no":
return true
}
return false
}
// Provider returns the short provider name ("anthropic"/"openai").
func (c Config) Provider() string {
if c.Format == llm.FormatOpenAI {
@@ -343,6 +361,7 @@ func TestConnection(ctx context.Context, c Config) (time.Duration, error) {
PermissionMode: acperm.ModeBypass,
MaxTurns: 1,
MaxTokens: 8192,
NonStreaming: !c.Stream, // 用该 profile 的真实收发模式做连接测试
}, "ping")
return time.Since(start), err
}
agent/worker.go
+12
View File
@@ -62,8 +62,19 @@ type Worker struct {
// injected into the worker system prompt. Read per run so the settings toggle
// takes effect without rebuilding the agent. nil = inject (default).
injectConstraints func() bool
// nonStreamingFn resolves whether this run uses the non-streaming (Complete)
// path. Read per run so a profile/task toggle takes effect without rebuilding
// the agent. nil = streaming (default).
nonStreamingFn func() bool
}
// SetNonStreaming wires a resolver deciding whether runs use the non-streaming
// model path (true = non-streaming). nil/unset = streaming (default). Read per
// run so a profile or task-chain toggle takes effect without rebuilding.
func (w *Worker) SetNonStreaming(fn func() bool) { w.nonStreamingFn = fn }
func (w *Worker) nonStreaming() bool { return w.nonStreamingFn != nil && w.nonStreamingFn() }
// 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 }
@@ -337,6 +348,7 @@ func (w *Worker) Execute(ctx context.Context, name string, taskID int64, as *db.
ToolOutputDir: cmdOutDir(runDir),
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
}
if hooks != nil { // typed-nil guard: only set when concrete (avoids harness panic)
opts.Hooks = hooks