ci / vet-and-test (pull_request) Successful in 38s
Why: User rulings 4 plus the P3 live evidence. The base library taught layout standards but not how to work: no planning discipline (plan tool used in 2 of 31 tool runs), no retry-escalation rule (same failing call retried verbatim), no check-API-first rule (phantom API calls), and nothing about edit granularity - the P3 after-run showed one turn of 27 single-line edits plus a self-review spiral (117 tools). The final-reply rewrite layer sees only 160-char summaries and fabricated chapters/promised unfinished work; with the transcript now authoritative the working model already has full evidence, so per ruling 4 the rewrite defaults off (kept as fallback when a work loop ends with no user-facing text at all, plus opt-in setting). Changes: new base.method module (plan-first for 3+ steps, switch approach after two failures, check API signatures before scripting, batch edits with multi_edit, persist to completion - never end on a promise, trace numbers/ dates/names back to source material or mark speculation, honest final report); tool-description overrides for read_file/edit_file/grep/glob/ run_python/run_shell (dedicated tools over cat/sed, build pipeline for Word); plan reminder now also nudges once when NO plan exists after several tool turns (ignorable, once per run); runtime.composeFinal default true->false with empty-body fallback. Evidence: SELF - TestPlanlessNudgeFiresOnce (fires once, never repeats), TestResolveSystemDefaultBase updated for the ninth base block; go vet + go test ./... pass; live before/after on the golden task set recorded on the phase04 task card per the policy clause. Impact: Behavior-surface change only; no contract or storage changes.
189 lines
6.5 KiB
Go
189 lines
6.5 KiB
Go
package runtime
|
||
|
||
import (
|
||
"fmt"
|
||
"strings"
|
||
"time"
|
||
|
||
"agent-native/runtime/gates"
|
||
agentllm "agent-native/runtime/llm"
|
||
core "agent-system-core"
|
||
)
|
||
|
||
const verifyCapNotice = "本轮回复未通过终轮检查。"
|
||
|
||
const verifierUnavailableNotice = "本轮回复未通过独立验证(验证器不可用)"
|
||
|
||
const verifyFallbackPrompt = `你是独立验证器,不调用任何工具。根据用户要求、终轮回复、工具证据和工件清单判断这一轮是否已经完成。
|
||
同一对象的后一次检查覆盖前一次,以最后一次为准。内置质量检查已通过时,不要用更早的工具输出推翻它。
|
||
拿不准时 passed 必须为 false。只输出一个 JSON 对象,不要写其他文字:{"passed":false,"reason":"为什么没通过","nextAction":"主模型下一步该做什么"}。`
|
||
|
||
// systemCheckReason 把终轮检查写成系统提醒。内核会把它当成一条用户消息送回模型,
|
||
// 不用这层包裹的话,模型会以为是用户自己提出的意见。
|
||
func systemCheckReason(body string) string {
|
||
body = strings.TrimSpace(body)
|
||
return "<system-reminder>\n这是系统的终轮自动检查,不是用户发言。先用工具核实并修复指出的问题。检查带位置时先修那一处;确认检查本身有误时,在最终答复里用一句话说明原因(不要引用检查原文),然后如实收尾——既不要默默无视未通过的检查,也不要在没核实的情况下反复空转。\n" + body + "\n</system-reminder>"
|
||
}
|
||
|
||
func (l *LLMRunner) runVerifier(rc RunCtx, body string, evidence []toolEvidence, refs []*core.ArtifactRef) (gates.Verdict, error) {
|
||
route := ResolveLightweightModel(rc.Profile(), rc.Settings())
|
||
stream := l.VerifyStream
|
||
if stream == nil {
|
||
if route.Model == "" || route.Base == "" {
|
||
return gates.Verdict{}, fmt.Errorf("验证器调用失败")
|
||
}
|
||
client := openLLMClient(route.Base, route.Key)
|
||
// 验证器不经内核循环:显式接重试(WP6 同映射)与 5xx 回退链。
|
||
stream = agentllm.WithFallback(
|
||
agentllm.WithRetry(client.Stream, retryAttemptsOf(rc.Settings()), nil),
|
||
ChatFallbacks(rc.Settings()), nil)
|
||
}
|
||
model := route.Model
|
||
if model == "" {
|
||
model = "verify"
|
||
}
|
||
msgs := []agentllm.Message{
|
||
{Role: agentllm.RoleSystem, Content: l.verifyPrompt(rc.Profile().Owner)},
|
||
{Role: agentllm.RoleUser, Content: evidencePrompt(currentUserText(rc), body, evidence, refs)},
|
||
}
|
||
var b strings.Builder
|
||
for ev := range stream(rc.Context(), model, msgs, nil, agentllm.StreamOptions{}) {
|
||
switch ev.Kind {
|
||
case agentllm.KindTextDelta:
|
||
b.WriteString(ev.Text)
|
||
case agentllm.KindError:
|
||
return gates.Verdict{}, fmt.Errorf("验证器调用失败")
|
||
case agentllm.KindMessageEnd:
|
||
if b.Len() == 0 {
|
||
b.WriteString(ev.Message.Content)
|
||
}
|
||
}
|
||
}
|
||
verdict, err := gates.ParseVerdict(b.String())
|
||
if err != nil {
|
||
return gates.Verdict{}, fmt.Errorf("验证器返回无法解析")
|
||
}
|
||
return verdict, nil
|
||
}
|
||
|
||
func (l *LLMRunner) verifyPrompt(owner string) string {
|
||
if l.Prompts != nil {
|
||
if body, ok, err := l.Prompts.ModuleBody(owner, "aux.verify.final"); err == nil && ok && strings.TrimSpace(body) != "" {
|
||
return body
|
||
}
|
||
}
|
||
return verifyFallbackPrompt
|
||
}
|
||
|
||
func evidencePrompt(user, body string, evidence []toolEvidence, refs []*core.ArtifactRef) string {
|
||
var b strings.Builder
|
||
b.WriteString("用户要求:\n")
|
||
if strings.TrimSpace(user) == "" {
|
||
b.WriteString("(无)\n")
|
||
} else {
|
||
b.WriteString(user)
|
||
b.WriteString("\n")
|
||
}
|
||
b.WriteString("\n终轮回复:\n")
|
||
b.WriteString(body)
|
||
b.WriteString("\n\n本轮工具(按时间编号。同一对象的后一次检查覆盖前一次,以最后一次为准):\n")
|
||
if len(evidence) == 0 {
|
||
b.WriteString("(无)\n")
|
||
}
|
||
for i, ev := range evidence {
|
||
text := ev.VerifyText
|
||
if text == "" {
|
||
text = ev.Summary
|
||
}
|
||
fmt.Fprintf(&b, "#%d %s %s %s\n", i+1, ev.Name, ev.Status, text)
|
||
}
|
||
b.WriteString("\n本轮工件:\n")
|
||
if len(refs) == 0 {
|
||
b.WriteString("(无)\n")
|
||
}
|
||
for _, ref := range refs {
|
||
if ref == nil {
|
||
continue
|
||
}
|
||
fmt.Fprintf(&b, "- %s %s\n", ref.ArtifactID, ref.Name)
|
||
}
|
||
b.WriteString("\n内置质量检查:上面列出的每个文件的最新版本均已检查:存在且非空。")
|
||
if len(refs) > 0 {
|
||
b.WriteString("文本类工件(md、txt、docx)的版式与内容评分也已通过;若某项检查未实际运行,会在结论里单独标注「自检未运行」。这是强证据,不要用更早的工具输出推翻它。")
|
||
}
|
||
b.WriteString("\n")
|
||
return b.String()
|
||
}
|
||
|
||
func (st *directRun) loadChecklist() {
|
||
if src, ok := st.rc.(interface{ Todos() (core.TodoList, bool) }); ok {
|
||
if list, ok := src.Todos(); ok {
|
||
st.todos = list
|
||
}
|
||
}
|
||
if src, ok := st.rc.(interface {
|
||
Plan() (core.PlanSnapshot, bool)
|
||
}); ok {
|
||
if p, ok := src.Plan(); ok {
|
||
st.plan = p
|
||
}
|
||
}
|
||
}
|
||
|
||
func (st *directRun) reminders(turn int) []string {
|
||
if st.closedToolTurn {
|
||
if st.checklistTouched {
|
||
st.toolTurnsSince = 0
|
||
} else {
|
||
st.toolTurnsSince++
|
||
}
|
||
st.closedToolTurn = false
|
||
st.checklistTouched = false
|
||
}
|
||
if st.toolTurnsSince < gates.ReminderToolGap {
|
||
return nil
|
||
}
|
||
todos, plans := st.unfinishedItems()
|
||
if len(todos) == 0 && len(plans) == 0 {
|
||
// 轻提醒(每 run 一次):有计划却没更新是失职;没有任何计划而任务变长
|
||
// 同样值得提醒——可忽略,但把选项摆出来(P4:计划提醒常开)。
|
||
if !st.nudgedPlanless {
|
||
st.nudgedPlanless = true
|
||
st.toolTurnsSince = 0
|
||
return []string{"本次任务已连续多个工具轮没有任何计划。如果剩余工作不止一步,用 update_plan 列出步骤再继续;确实只剩简单收尾就忽略本条。"}
|
||
}
|
||
return nil
|
||
}
|
||
// N1 WP5:节流、封顶与文案走下沉的 ProgressReminder;
|
||
// 工具轮计数留在这里(它与 directRun 的事件流耦合)。
|
||
if st.progress == nil {
|
||
st.progress = &gates.ProgressReminder{}
|
||
}
|
||
got := st.progress.Remind(todos, plans, time.Now())
|
||
if got != nil {
|
||
st.toolTurnsSince = 0
|
||
}
|
||
return got
|
||
}
|
||
|
||
// unfinishedItems 返回计划与待办的未完成项正文(状态过滤留在产品侧)。
|
||
func (st *directRun) unfinishedItems() (todos, plans []string) {
|
||
for _, item := range st.todos.Items {
|
||
if item.Status == "done" {
|
||
continue
|
||
}
|
||
if text := strings.TrimSpace(item.Text); text != "" {
|
||
todos = append(todos, text)
|
||
}
|
||
}
|
||
for _, step := range st.plan.Steps {
|
||
if step.Status == "done" || step.Status == "skipped" {
|
||
continue
|
||
}
|
||
if text := strings.TrimSpace(step.Text); text != "" {
|
||
plans = append(plans, text)
|
||
}
|
||
}
|
||
return todos, plans
|
||
}
|