mirror of
https://gitcode.com/JianFeeeee/HomeAgent.git
synced 2026-10-03 15:53:56 +00:00
refactor(scheduler): M1 把 process() 拆成 step 状态机 + TaskFrame(行为等价)
设计依据 docs/zh/input-scheduler-design.md §14 M1。 - 新增 task.go:Step 游标、TaskFrame,以及 stepPrepare/stepLLM/ stepToolBegin/stepToolExec/stepToolAfter/stepTurnEnd 六个 step; 原函数内嵌的 provider 回退/重试抽为 resolveProviders + callLLMWithFallback - process() 改为驱动状态机的薄壳:签名不变,调用方(processInput/ processConsolidation/测试)零改动 - StepToolExec 显式标注为临界区:工具副作用不可回滚,执行中不是安全点 - 步数上限护栏:转移缺失时以错误退出而非死循环 - 新增 task_test.go:R3(工具往返结果与配对正确)、X3(多轮必然终止)、 批中途中断必须放弃剩余工具、未知 step 必须失败退出 - 验收:既有 agent 全量测试通过;-race 通过(TMPDIR 指向真实磁盘, /tmp tmpfs 已 98% 满会导致链接失败,与代码无关)
This commit is contained in:
@ -7,12 +7,10 @@ import (
|
||||
"fmt"
|
||||
"log"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
agentAPI "gitcode.com/JianFeeeee/HomeAgent/internal/agent/api"
|
||||
"gitcode.com/JianFeeeee/HomeAgent/internal/events"
|
||||
sdk "gitcode.com/JianFeeeee/HomeAgent/internal/sdk"
|
||||
pubsdk "gitcode.com/JianFeeeee/homeagent-sdk/sdk"
|
||||
)
|
||||
|
||||
// continuationPlaceholder 是工具轮之后补的 user 占位内容。
|
||||
@ -99,420 +97,6 @@ func dropContinuationPlaceholders(msgs []agentAPI.Message) []agentAPI.Message {
|
||||
return out
|
||||
}
|
||||
|
||||
func (a *Agent) process(input string, stageCtx *sdk.StageContext) (response string, toolsUsed []string, toolResults []ToolResultItem, err error) {
|
||||
a.mu.Lock()
|
||||
defer a.mu.Unlock()
|
||||
|
||||
if a.provider == nil {
|
||||
return "", nil, nil, fmt.Errorf("agent: no LLM provider configured")
|
||||
}
|
||||
|
||||
budget := ComputeTokenBudget(a.provider, a.systemPrompt)
|
||||
|
||||
memContext := a.buildMemoryContext(input, budget.MemoryTokens)
|
||||
sysPrompt := a.buildSystemPrompt(memContext, input)
|
||||
tools := a.buildToolDefs()
|
||||
|
||||
msgs := a.buildMessages(sysPrompt, input, budget.ContextTokens)
|
||||
// 工具提醒(interrupt):以 system 角色注入,不让模型误认为用户发言
|
||||
if a.interruptInput {
|
||||
last := msgs[len(msgs)-1]
|
||||
last.Role = "system"
|
||||
last.Content = "[中断消息] " + last.Content
|
||||
msgs[len(msgs)-1] = last
|
||||
a.interruptInput = false
|
||||
}
|
||||
if blocks, ok := stageCtx.Extra["media_blocks"].([]agentAPI.ContentBlock); ok && len(blocks) > 0 {
|
||||
if len(msgs) > 0 {
|
||||
msgs[len(msgs)-1].Blocks = blocks
|
||||
}
|
||||
}
|
||||
|
||||
log.Printf("[agent] tool call loop start, max_ctx=%d target=%d fixed=%d mem=%d ctx=%d %d tools, %d events, personality=%t, docs=%d",
|
||||
budget.MaxContext, budget.TargetUsage, budget.FixedTokens, budget.MemoryTokens, budget.ContextTokens,
|
||||
len(tools), a.context.Len(),
|
||||
a.personality != nil && a.personality.Content != "",
|
||||
a.docStoreSize())
|
||||
|
||||
if a.runStage(sdk.StagePreAction, stageCtx) {
|
||||
return *stageCtx.Response, toolsUsed, toolResults, nil
|
||||
}
|
||||
if len(stageCtx.ContextMsgs) > 0 {
|
||||
for _, m := range stageCtx.ContextMsgs {
|
||||
role, _ := m["role"].(string)
|
||||
content, _ := m["content"].(string)
|
||||
if role != "" {
|
||||
msgs = append(msgs, agentAPI.Message{Role: role, Content: content})
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// lastBatchReplyOnly 记录上一批工具调用是否全部是输出通道发送。
|
||||
lastBatchReplyOnly := false
|
||||
|
||||
for turn := 0; ; turn++ {
|
||||
for _, interrupt := range a.drainInterrupts() {
|
||||
msgs = append(msgs, agentAPI.Message{
|
||||
Role: "system",
|
||||
Content: "[中断消息] " + interrupt,
|
||||
})
|
||||
}
|
||||
|
||||
// zen 兼容网关要求请求的最后一条消息必须是 user(thinking 续写模式校验),
|
||||
// 工具轮产出的 tool/assistant 消息作结尾会被 400 拒绝,故补一条 user 占位。
|
||||
// 注意:仅当尾部确为工具轮产物(assistant/tool)时才补位;首轮 system 上下文结尾不补,
|
||||
// 否则会错误覆盖实际用户输入(如 injectSourceContext 追加的 system 说明)。
|
||||
//
|
||||
// 补位前先移除前面轮次插入的同类占位,保证占位**不随轮次线性累积**——
|
||||
// 占位是核心插的传输层附加物,不是用户发言,不该在 prompt 里叠成 N 条。
|
||||
//
|
||||
// 文案分情况:上一批全是 output_send__* 时不能说“继续”,详见
|
||||
// replyDeliveredPlaceholder 的说明。
|
||||
msgs = dropContinuationPlaceholders(msgs)
|
||||
if last := msgs[len(msgs)-1]; last.Role == "assistant" || last.Role == "tool" {
|
||||
msgs = append(msgs, agentAPI.Message{
|
||||
Role: "user",
|
||||
Content: continuationFor(lastBatchReplyOnly),
|
||||
})
|
||||
}
|
||||
|
||||
req := &agentAPI.CompletionRequest{
|
||||
Messages: msgs,
|
||||
MaxTokens: 4096,
|
||||
Tools: tools,
|
||||
ToolChoice: "auto",
|
||||
DisableThinking: !a.thinkingEnabled,
|
||||
}
|
||||
|
||||
var providers []agentAPI.Provider
|
||||
if a.providerManager != nil {
|
||||
// 精确模型名走 byModel 路由;AUTO/空走优先级链
|
||||
var allProviders []agentAPI.Provider
|
||||
if req.Model != "" && !strings.EqualFold(req.Model, "AUTO") {
|
||||
allProviders = a.providerManager.ResolveForModel(req.Model)
|
||||
} else {
|
||||
allProviders = a.providerManager.OrderedProviders()
|
||||
}
|
||||
providers = make([]agentAPI.Provider, 0, len(allProviders))
|
||||
for _, p := range allProviders {
|
||||
if a.providerManager.IsAvailable(p.Name()) {
|
||||
providers = append(providers, p)
|
||||
}
|
||||
}
|
||||
}
|
||||
if len(providers) == 0 {
|
||||
providers = []agentAPI.Provider{a.provider}
|
||||
}
|
||||
var resp *agentAPI.CompletionResponse
|
||||
var llmErr error
|
||||
|
||||
for pi, fbProvider := range providers {
|
||||
if pi > 0 {
|
||||
log.Printf("[agent] LLM fallback: trying provider %q (fallback #%d/%d)",
|
||||
fbProvider.Name(), pi, len(providers)-1)
|
||||
}
|
||||
|
||||
// 同源瞬时错误重试:网关瞬断(502/503/504/429/网络抖动)通常秒级恢复,
|
||||
// 直接跳下一个 provider(或直接报错)会丢掉本可成功的请求。
|
||||
// 凭证错误(401/403)与用户中断不重试。
|
||||
const maxAttempts = 2
|
||||
for attempt := 1; attempt <= maxAttempts; attempt++ {
|
||||
if attempt > 1 {
|
||||
log.Printf("[agent] provider %q transient failure, retry %d/%d in 2s: %v",
|
||||
fbProvider.Name(), attempt, maxAttempts, llmErr)
|
||||
select {
|
||||
case <-time.After(2 * time.Second):
|
||||
case <-a.ctx.Done():
|
||||
llmErr = a.ctx.Err()
|
||||
}
|
||||
if llmErr == nil || errors.Is(llmErr, context.Canceled) || errors.Is(llmErr, context.DeadlineExceeded) {
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
fCtx, fCancel := context.WithCancel(a.ctx)
|
||||
a.llmMu.Lock()
|
||||
a.cancelLLM = fCancel
|
||||
a.llmMu.Unlock()
|
||||
|
||||
resp, llmErr = chatStreamWithFallback(fCtx, fbProvider, req, a)
|
||||
|
||||
a.llmMu.Lock()
|
||||
a.cancelLLM = nil
|
||||
a.llmMu.Unlock()
|
||||
fCancel()
|
||||
|
||||
if llmErr == nil {
|
||||
a.providerManager.ResetAvailability(fbProvider.Name())
|
||||
if fbProvider != a.provider {
|
||||
a.provider = fbProvider
|
||||
log.Printf("[agent] switched active provider to %q after fallback",
|
||||
fbProvider.Name())
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
// 用户中断:立即终止,不重试也不换 provider
|
||||
if errors.Is(llmErr, context.Canceled) {
|
||||
break
|
||||
}
|
||||
// 凭证错误:重试无意义,跳出重试循环进入 provider 标记/切换
|
||||
var pe *agentAPI.ProviderError
|
||||
if errors.As(llmErr, &pe) && (pe.StatusCode == 401 || pe.StatusCode == 403) {
|
||||
break
|
||||
}
|
||||
// 其余错误(含 5xx/429/网络):还有重试机会则继续,否则跳出
|
||||
}
|
||||
|
||||
if llmErr == nil {
|
||||
break
|
||||
}
|
||||
if errors.Is(llmErr, context.Canceled) {
|
||||
break
|
||||
}
|
||||
var pe *agentAPI.ProviderError
|
||||
if errors.As(llmErr, &pe) && (pe.StatusCode == 401 || pe.StatusCode == 403) {
|
||||
a.providerManager.ReportStatus(fbProvider.Name(), pe.StatusCode)
|
||||
log.Printf("[agent] provider %q marked unavailable (HTTP %d)", fbProvider.Name(), pe.StatusCode)
|
||||
} else {
|
||||
a.providerManager.MarkUnavailable(fbProvider.Name())
|
||||
}
|
||||
log.Printf("[agent] provider %q failed: %v", fbProvider.Name(), llmErr)
|
||||
}
|
||||
|
||||
if llmErr != nil {
|
||||
if errors.Is(llmErr, context.Canceled) && a.ctx.Err() == nil {
|
||||
if a.currentOutputChannel == "_consolidation_" {
|
||||
return "", toolsUsed, toolResults, fmt.Errorf("interrupted by user input")
|
||||
}
|
||||
continue
|
||||
}
|
||||
return "", toolsUsed, toolResults, fmt.Errorf("all %d providers failed, last error: %w",
|
||||
len(providers), llmErr)
|
||||
}
|
||||
|
||||
stageCtx.LLMText = resp.Content
|
||||
stageCtx.ReasoningContent = resp.ReasoningContent
|
||||
stageCtx.TokenUsage = map[string]int{
|
||||
"prompt_tokens": resp.TokenUsage.Prompt,
|
||||
"completion_tokens": resp.TokenUsage.Completion,
|
||||
"total_tokens": resp.TokenUsage.Total,
|
||||
}
|
||||
stageCtx.ToolCalls = convertToolCalls(resp.ToolCalls)
|
||||
for i := range stageCtx.ToolCalls {
|
||||
if stageCtx.ToolCalls[i].Plugin == "" {
|
||||
stageCtx.ToolCalls[i].Plugin = a.resolveToolPlugin(stageCtx.ToolCalls[i].Name)
|
||||
}
|
||||
}
|
||||
if a.runStage(sdk.StagePostAction, stageCtx) {
|
||||
return *stageCtx.Response, toolsUsed, toolResults, nil
|
||||
}
|
||||
resp.Content = stageCtx.LLMText
|
||||
resp.ToolCalls = convertBackToolCalls(stageCtx.ToolCalls)
|
||||
|
||||
chainPayload := map[string]interface{}{
|
||||
"content": resp.Content,
|
||||
"reasoning": resp.ReasoningContent,
|
||||
"tool_calls": resp.ToolCalls,
|
||||
"phase": "intermediate",
|
||||
"turn": turn,
|
||||
}
|
||||
if resp.TokenUsage.Total > 0 {
|
||||
chainPayload["usage"] = map[string]int{
|
||||
"prompt": resp.TokenUsage.Prompt,
|
||||
"completion": resp.TokenUsage.Completion,
|
||||
"total": resp.TokenUsage.Total,
|
||||
}
|
||||
}
|
||||
a.publishEvent(events.EventAgentLLMChain, chainPayload)
|
||||
|
||||
if resp.ReasoningContent != "" {
|
||||
a.publishEvent(events.EventReasoning, map[string]interface{}{
|
||||
"content": resp.ReasoningContent,
|
||||
"channel": a.currentOutputChannel,
|
||||
})
|
||||
}
|
||||
|
||||
if len(resp.ToolCalls) == 0 {
|
||||
return resp.Content, toolsUsed, toolResults, nil
|
||||
}
|
||||
|
||||
// 本批是否全部是输出通道发送(=模型刚交付了给用户的回复)。
|
||||
// 必须在执行前判定:执行过程中的中断/拒绝分支会 continue/break,
|
||||
// 放在循环里统计会漏。
|
||||
replyOnly := true
|
||||
for _, tc := range resp.ToolCalls {
|
||||
if !isOutputDeliveryTool(tc.Name) {
|
||||
replyOnly = false
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
contentOnce := true
|
||||
for _, tc := range resp.ToolCalls {
|
||||
if len(a.interceptCh) > 0 {
|
||||
for _, interrupt := range a.drainInterrupts() {
|
||||
msgs = append(msgs, agentAPI.Message{Role: "system", Content: "[中断消息] " + interrupt})
|
||||
}
|
||||
a.publishEvent(events.EventToolCall, map[string]interface{}{
|
||||
"tool": tc.Name,
|
||||
"plugin": a.resolveToolPlugin(tc.Name),
|
||||
"args": tc.Arguments,
|
||||
"status": "interrupted",
|
||||
"reason": "user interrupt before execution",
|
||||
"channel": a.currentOutputChannel,
|
||||
})
|
||||
break
|
||||
}
|
||||
|
||||
toolsUsed = append(toolsUsed, tc.Name)
|
||||
pluginName := a.resolveToolPlugin(tc.Name)
|
||||
log.Printf("[agent] executing tool: %s (plugin=%s, id=%s)", tc.Name, pluginName, tc.ID)
|
||||
if tc.RawArguments != "" {
|
||||
log.Printf("[agent] tool %s raw_arguments: %s", tc.Name, truncateStr(tc.RawArguments, 300))
|
||||
}
|
||||
|
||||
sdkTC := sdk.ToolCall{ID: tc.ID, Name: tc.Name, Plugin: pluginName, Arguments: tc.Arguments}
|
||||
stageCtx.ToolCalls = []sdk.ToolCall{sdkTC}
|
||||
stageCtx.ToolResults = nil
|
||||
if a.runStage(sdk.StageBeforeToolcall, stageCtx) {
|
||||
result := fmt.Sprintf("工具 %s 已被插件拒绝", tc.Name)
|
||||
msgs = append(msgs, agentAPI.Message{Role: "assistant", ToolCalls: []agentAPI.ToolCall{tc}})
|
||||
msgs = append(msgs, agentAPI.Message{Role: "tool", ToolCallID: tc.ID, Content: result})
|
||||
a.publishEvent(events.EventToolCall, map[string]interface{}{
|
||||
"tool": tc.Name,
|
||||
"plugin": pluginName,
|
||||
"args": tc.Arguments,
|
||||
"result": result,
|
||||
"status": "denied",
|
||||
"channel": a.currentOutputChannel,
|
||||
})
|
||||
continue
|
||||
}
|
||||
tc.Arguments = stageCtx.ToolCalls[0].Arguments
|
||||
|
||||
if pluginName != "" && !a.pluginHealth.isHealthy(pluginName) {
|
||||
result := fmt.Sprintf("插件 %s 处于崩溃状态,已跳过执行,等待自动恢复重载", pluginName)
|
||||
log.Printf("[agent] skip tool %s: plugin %s unhealthy", tc.Name, pluginName)
|
||||
msgs = append(msgs, agentAPI.Message{Role: "assistant", ToolCalls: []agentAPI.ToolCall{tc}})
|
||||
msgs = append(msgs, agentAPI.Message{Role: "tool", ToolCallID: tc.ID, Content: result})
|
||||
continue
|
||||
}
|
||||
|
||||
result := a.executeToolCall(tc)
|
||||
toolResults = append(toolResults, ToolResultItem{Name: tc.Name, Output: result})
|
||||
log.Printf("[agent] tool %s result: %s", tc.Name, truncateStr(result, 100))
|
||||
|
||||
stageCtx.ToolResults = []sdk.ToolResult{{CallID: tc.ID, Name: tc.Name, Plugin: pluginName, Success: true, Result: result}}
|
||||
a.runStage(sdk.StageAfterToolcall, stageCtx)
|
||||
if len(stageCtx.ToolResults) > 0 {
|
||||
if r, ok := stageCtx.ToolResults[0].Result.(string); ok {
|
||||
result = r
|
||||
}
|
||||
}
|
||||
// ContextPolicy: prune 工具调用后执行上下文裁剪(§13.8)
|
||||
if def := a.stageHost.ToolDef(tc.Name); def != nil && def.ContextPolicy == "prune" {
|
||||
if a.context != nil {
|
||||
topK := a.maxContextSize - 1
|
||||
if topK < 1 {
|
||||
topK = 1
|
||||
}
|
||||
// 查询向量取**清洗后**的有效内容,否则噪声(ANSI/base64/JSON
|
||||
// 包装)会把相关性打分带偏,裁掉本该保留的事件。
|
||||
a.context.Prune(a.toolOutputForQuery(tc.Name, result), topK, a.docStore)
|
||||
}
|
||||
}
|
||||
|
||||
msgContent := ""
|
||||
if contentOnce {
|
||||
msgContent = resp.Content
|
||||
contentOnce = false
|
||||
}
|
||||
msgs = append(msgs, agentAPI.Message{Role: "assistant", Content: msgContent, ReasoningContent: resp.ReasoningContent, ToolCalls: []agentAPI.ToolCall{tc}})
|
||||
|
||||
// 多模态工具结果:插件通过 SDK.SetToolBlocks 注入 image_url/audio_url block。
|
||||
//
|
||||
// 媒体不挂在 tool message 上,而是另起一条紧随其后的 user message——
|
||||
// 这也是插件文案一直在说的「注入后续对话」。
|
||||
// 为何不能挂 tool message:同一张图、同一模型、三轮实测——
|
||||
// 图在 user message → 3/3 读到
|
||||
// 图在 tool message → 0/3(模型答「没能读到这张图」)
|
||||
// tool 纯文本 + 后接 user → 3/3 读到
|
||||
// tool message 那轮 prompt_tokens 反而更高(7967 vs 7089),base64 确实
|
||||
// 进了上游,但 role=tool 上的多模态 content 数组不被当作可视内容。
|
||||
//
|
||||
// 主模型不支持该模态时更不能直接塞:网关会把 image_url 静默剥离后仍
|
||||
// 返回 200,模型回答「我没有看到图片」而内核以为注入成功。改走回退链。
|
||||
toolMsg := agentAPI.Message{Role: "tool", ToolCallID: tc.ID, Content: result}
|
||||
var mediaMsg *agentAPI.Message
|
||||
if rawBlocks := a.io.ConsumeToolBlocks(); len(rawBlocks) > 0 {
|
||||
var blocks []agentAPI.ContentBlock
|
||||
for _, b := range rawBlocks {
|
||||
if cb, ok := b.(pubsdk.ContentBlock); ok {
|
||||
// 跨包类型拷贝(pubsdk.ContentBlock → agentAPI.ContentBlock)
|
||||
block := agentAPI.ContentBlock{Type: cb.Type, Text: cb.Text}
|
||||
if cb.ImageURL != nil {
|
||||
block.ImageURL = &agentAPI.ImageURL{URL: cb.ImageURL.URL, Detail: cb.ImageURL.Detail}
|
||||
}
|
||||
if cb.AudioURL != nil {
|
||||
block.AudioURL = &agentAPI.AudioURL{URL: cb.AudioURL.URL}
|
||||
}
|
||||
blocks = append(blocks, block)
|
||||
}
|
||||
}
|
||||
if len(blocks) > 0 {
|
||||
// 先落进 CAS:无论下面走直视还是回退转写,媒体本体都该进记忆。
|
||||
// 不存的后果是 ToolResultItem.Output 只剩那句
|
||||
// "[已将图片注入后续对话] /tmp/x.png",文件一删线索就断了。
|
||||
a.stageMediaDigests(a.captureBlockMedia(blocks, tc.Name)...)
|
||||
|
||||
if native, fallbackText := a.prepareToolBlocks(blocks); len(native) > 0 {
|
||||
// 能直视:另起一条 user message 承载媒体,并补一句来源说明,
|
||||
// 否则模型会把它当成用户新发的图而不是工具拉回来的。
|
||||
mediaBlocks := append([]agentAPI.ContentBlock{{
|
||||
Type: "text",
|
||||
Text: fmt.Sprintf("[以下是 %s 注入的媒体内容]", tc.Name),
|
||||
}}, native...)
|
||||
mediaMsg = &agentAPI.Message{Role: "user", Blocks: mediaBlocks}
|
||||
} else if fallbackText != "" {
|
||||
// 回退链已把媒体转写成文字:并进 tool message 的纯文本 content,
|
||||
// 不再另起消息(文字在 tool message 里本来就能被读到)。
|
||||
toolMsg.Content = result + "\n\n" + fallbackText
|
||||
result = toolMsg.Content
|
||||
if len(toolResults) > 0 {
|
||||
toolResults[len(toolResults)-1].Output = result
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
msgs = append(msgs, toolMsg)
|
||||
if mediaMsg != nil {
|
||||
// 必须紧跟在 toolMsg 之后:中间插入其他消息会让 tool_call_id 配对断开。
|
||||
msgs = append(msgs, *mediaMsg)
|
||||
}
|
||||
|
||||
a.publishEvent(events.EventToolCall, map[string]interface{}{
|
||||
"tool": tc.Name,
|
||||
"plugin": pluginName,
|
||||
"args": tc.Arguments,
|
||||
"result": result,
|
||||
"status": "ok",
|
||||
"channel": a.currentOutputChannel,
|
||||
})
|
||||
|
||||
if len(a.interceptCh) > 0 {
|
||||
for _, interrupt := range a.drainInterrupts() {
|
||||
msgs = append(msgs, agentAPI.Message{Role: "system", Content: "[中断消息] " + interrupt})
|
||||
}
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
// 供下一轮顶部选择补位文案。
|
||||
lastBatchReplyOnly = replyOnly
|
||||
}
|
||||
}
|
||||
|
||||
// chatStreamWithFallback 优先流式调用 provider,失败时回退非流式 Chat()。
|
||||
//
|
||||
// 流式路径:ChatStream 拿到 chunk channel,逐块累积 content/reasoning_content,
|
||||
|
||||
Reference in New Issue
Block a user