Sakurasan
d5dac5a6fc
llm（三协议 Provider）/ session（JSONL 持久化）/ tools（Registry、围栏、 builtin、computer）/ cdp（browser）/ skills / mcp / agent（循环、两级压缩、 编排、审批门控）+ 底座 jsonl/textx/execx/diag/scratch/spill/imagecache。 产品装配代码（config 桥接、schedule 工具、memory/plan/retro/prompt） 留在 AutoGo，宿主接入方式见 MIGRATION.md。
259 lines
8.8 KiB
Go
259 lines
8.8 KiB
Go
package agent
|
||
|
||
import (
|
||
"context"
|
||
"errors"
|
||
"fmt"
|
||
"io"
|
||
"regexp"
|
||
"strings"
|
||
"testing"
|
||
|
||
"git.gopher.ink/mirrors2/AutoZero/llm"
|
||
"git.gopher.ink/mirrors2/AutoZero/session"
|
||
)
|
||
|
||
// TestCheckpointUpdatedInPlace：第二次压缩要把旧 checkpoint + 新消息交给更新
|
||
// prompt，而不是把整段历史（含旧摘要）重新摘要一遍；sidecar 同步更新。
|
||
func TestCheckpointUpdatedInPlace(t *testing.T) {
|
||
var inputs []string
|
||
var summaryCalls int
|
||
prov := &scriptedProvider{fn: func(_ int, system string, msgs []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
|
||
if strings.HasPrefix(system, "You are a context compaction") {
|
||
summaryCalls++
|
||
inputs = append(inputs, msgs[0].Content)
|
||
return llm.Message{Role: llm.RoleAssistant, Content: fmt.Sprintf("CHECKPOINT-V%d", summaryCalls)}, nil
|
||
}
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "ok"}, nil
|
||
}}
|
||
a, sess := newTestAgent(t)
|
||
a.Provider = prov
|
||
a.MinCompactTokens = 1
|
||
a.Opt.Stderr = io.Discard
|
||
|
||
if err := sess.Append(llm.Message{Role: llm.RoleUser, Content: "first task"}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if err := sess.Append(llm.Message{Role: llm.RoleAssistant, Content: "working"}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if _, err := a.CompactNow(context.Background()); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if a.checkpoint != "CHECKPOINT-V1" || a.checkpointCovered != 2 {
|
||
t.Fatalf("first checkpoint = %q covered=%d", a.checkpoint, a.checkpointCovered)
|
||
}
|
||
if len(a.compacted) != 1 || !strings.Contains(a.compacted[0].Content, "CHECKPOINT-V1") {
|
||
t.Fatalf("working set = %+v", a.compacted)
|
||
}
|
||
|
||
if err := sess.Append(llm.Message{Role: llm.RoleUser, Content: "more context"}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if err := sess.Append(llm.Message{Role: llm.RoleAssistant, Content: "answer"}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if _, err := a.CompactNow(context.Background()); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if a.checkpoint != "CHECKPOINT-V2" {
|
||
t.Fatalf("second checkpoint = %q", a.checkpoint)
|
||
}
|
||
if a.checkpointCovered != 4 {
|
||
t.Fatalf("covered = %d, want 4", a.checkpointCovered)
|
||
}
|
||
if len(inputs) != 2 {
|
||
t.Fatalf("summary calls = %d", len(inputs))
|
||
}
|
||
second := inputs[1]
|
||
if !strings.Contains(second, "Existing checkpoint") || !strings.Contains(second, "CHECKPOINT-V1") {
|
||
t.Fatalf("update prompt did not carry the prior checkpoint: %q", second)
|
||
}
|
||
if !strings.Contains(second, "more context") {
|
||
t.Fatalf("update prompt missing newer history: %q", second)
|
||
}
|
||
if strings.Contains(second, "first task") {
|
||
t.Fatalf("update prompt re-summarized already-covered history: %q", second)
|
||
}
|
||
cp, ok := session.LoadCheckpoint(a.sessionID())
|
||
if !ok || cp.Text != "CHECKPOINT-V2" || cp.Covered != 4 {
|
||
t.Fatalf("sidecar = %+v ok=%v", cp, ok)
|
||
}
|
||
}
|
||
|
||
// TestLoadCheckpointRestoresAndValidates：新回合能从 sidecar 直接接上
|
||
// checkpoint + 未覆盖消息；transcript 被改写后指纹失效，退回完整历史。
|
||
func TestLoadCheckpointRestoresAndValidates(t *testing.T) {
|
||
prov := &scriptedProvider{fn: func(_ int, system string, _ []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
|
||
if strings.HasPrefix(system, "You are a context compaction") {
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "CHECKPOINT-X"}, nil
|
||
}
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "ok"}, nil
|
||
}}
|
||
a, sess := newTestAgent(t)
|
||
a.Provider = prov
|
||
a.MinCompactTokens = 1
|
||
a.Opt.Stderr = io.Discard
|
||
for _, content := range []string{"u1", "a1", "u2", "a2"} {
|
||
role := llm.RoleUser
|
||
if strings.HasPrefix(content, "a") {
|
||
role = llm.RoleAssistant
|
||
}
|
||
if err := sess.Append(llm.Message{Role: role, Content: content}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
}
|
||
if _, err := a.CompactNow(context.Background()); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
|
||
b := &Agent{Provider: prov, Tools: a.Tools, Session: sess, System: "test",
|
||
Opt: Options{Stdout: io.Discard, Stderr: io.Discard}}
|
||
if !b.LoadCheckpoint() {
|
||
t.Fatal("valid checkpoint not restored")
|
||
}
|
||
if b.checkpoint != "CHECKPOINT-X" || b.compactFrom != 4 {
|
||
t.Fatalf("restored state: checkpoint=%q from=%d", b.checkpoint, b.compactFrom)
|
||
}
|
||
if err := sess.Append(llm.Message{Role: llm.RoleUser, Content: "next"}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
b.ContextWindow = 1 << 30
|
||
b.CompactAt = 1
|
||
b.MinCompactTokens = 0
|
||
view, err := b.contextView(context.Background())
|
||
if err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if len(view) != 2 || !strings.Contains(view[0].Content, "CHECKPOINT-X") || view[1].Content != "next" {
|
||
t.Fatalf("restored view = %+v", view)
|
||
}
|
||
|
||
// 改写已覆盖的历史：指纹失效，不能继续用旧 checkpoint。
|
||
sess.Messages[3].Content = "tampered"
|
||
c := &Agent{Provider: prov, Tools: a.Tools, Session: sess, System: "test",
|
||
Opt: Options{Stdout: io.Discard, Stderr: io.Discard}}
|
||
if c.LoadCheckpoint() {
|
||
t.Fatal("stale checkpoint must be rejected after transcript changes")
|
||
}
|
||
}
|
||
|
||
// TestObservedTokensPrefersActualUsage：压缩判据用网关回报的真实 prompt tokens，
|
||
// 不再被 bytes/3 的估算单独带偏；请求之后新增的消息仍按估算补上。
|
||
func TestObservedTokensPrefersActualUsage(t *testing.T) {
|
||
a := &Agent{
|
||
System: strings.Repeat("s", 300),
|
||
ContextWindow: 10000,
|
||
CompactAt: 0.5,
|
||
MinCompactTokens: 1,
|
||
lastInputTokens: 0,
|
||
}
|
||
view := []llm.Message{
|
||
{Role: llm.RoleUser, Content: "short"},
|
||
{Role: llm.RoleAssistant, Content: "ok"},
|
||
}
|
||
a.noteRequestUsage(&llm.Usage{InputTokens: 6000}, view)
|
||
if !a.shouldCompact(view) {
|
||
t.Fatal("actual 6000 tokens should trigger compaction despite a low estimate")
|
||
}
|
||
|
||
big := []llm.Message{
|
||
{Role: llm.RoleUser, Content: strings.Repeat("a", 30000)},
|
||
{Role: llm.RoleAssistant, Content: "ok"},
|
||
}
|
||
a.noteRequestUsage(&llm.Usage{InputTokens: 1000}, big)
|
||
if a.shouldCompact(big) {
|
||
t.Fatal("actual 1000 tokens should not be overridden by the estimate")
|
||
}
|
||
|
||
grown := append(append([]llm.Message{}, big...),
|
||
llm.Message{Role: llm.RoleUser, Content: strings.Repeat("b", 15000)})
|
||
if !a.shouldCompact(grown) {
|
||
t.Fatal("messages added after the measured request must count toward the trigger")
|
||
}
|
||
}
|
||
|
||
// TestForceCompactTargetHint：收缩阶梯要把目标 token 数写进摘要提示词。
|
||
func TestForceCompactTargetHint(t *testing.T) {
|
||
var inputs []string
|
||
prov := &scriptedProvider{fn: func(_ int, system string, msgs []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
|
||
if strings.HasPrefix(system, "You are a context compaction") {
|
||
inputs = append(inputs, msgs[0].Content)
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "CP"}, nil
|
||
}
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "ok"}, nil
|
||
}}
|
||
a, sess := newTestAgent(t)
|
||
a.Provider = prov
|
||
a.MinCompactTokens = 1
|
||
a.Opt.Stderr = io.Discard
|
||
for _, content := range []string{"u1", "a1"} {
|
||
if err := sess.Append(llm.Message{Role: llm.RoleUser, Content: content}); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
}
|
||
view := append([]llm.Message(nil), sess.Messages...)
|
||
a.noteRequestUsage(&llm.Usage{InputTokens: 10000}, view)
|
||
if err := a.forceCompactForSize(context.Background(), view, 0.5); err != nil {
|
||
t.Fatal(err)
|
||
}
|
||
if len(inputs) != 1 || !strings.Contains(inputs[0], "under about 5000 tokens") {
|
||
t.Fatalf("target hint missing: %v", inputs)
|
||
}
|
||
}
|
||
|
||
// TestOversizedShrinkLadder：请求过大被拒后按 0.7 / 0.5 / 0.35 逐级收缩，
|
||
// 每次重发都基于刚压出来的新视图；压到网关肯收为止。
|
||
func TestOversizedShrinkLadder(t *testing.T) {
|
||
realCalls, summaryCalls := 0, 0
|
||
var targets []int
|
||
prov := &scriptedProvider{fn: func(_ int, system string, msgs []llm.Message, _ []llm.ToolDef) (llm.Message, error) {
|
||
if strings.HasPrefix(system, "You are a context compaction") {
|
||
summaryCalls++
|
||
m := regexp.MustCompile(`under about (\d+) tokens`).FindStringSubmatch(msgs[0].Content)
|
||
if m == nil {
|
||
t.Fatalf("summary input missing target: %q", msgs[0].Content)
|
||
}
|
||
target := atoiTest(m[1])
|
||
targets = append(targets, target)
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "CP:" + strings.Repeat("x", target*3)}, nil
|
||
}
|
||
realCalls++
|
||
if realCalls <= 3 {
|
||
return llm.Message{}, errors.New("HTTP 400: context length too large")
|
||
}
|
||
return llm.Message{Role: llm.RoleAssistant, Content: "done"}, nil
|
||
}}
|
||
a, sess := newTestAgent(t)
|
||
a.Provider = prov
|
||
a.Opt.Stderr = io.Discard
|
||
a.ContextWindow = 1 << 30
|
||
a.CompactAt = 1
|
||
a.MinCompactTokens = 1
|
||
sess.Messages = []llm.Message{
|
||
{Role: llm.RoleUser, Content: strings.Repeat("a", 30000)},
|
||
{Role: llm.RoleAssistant, Content: strings.Repeat("b", 30000)},
|
||
}
|
||
|
||
if err := a.Run(context.Background(), ""); err != nil {
|
||
t.Fatalf("run: %v (real=%d summary=%d targets=%v)", err, realCalls, summaryCalls, targets)
|
||
}
|
||
if realCalls != 4 || summaryCalls != 3 {
|
||
t.Fatalf("real=%d summary=%d targets=%v", realCalls, summaryCalls, targets)
|
||
}
|
||
if !(targets[0] > targets[1] && targets[1] > targets[2]) {
|
||
t.Fatalf("targets must shrink: %v", targets)
|
||
}
|
||
}
|
||
|
||
func atoiTest(s string) int {
|
||
n := 0
|
||
for _, r := range s {
|
||
if r < '0' || r > '9' {
|
||
break
|
||
}
|
||
n = n*10 + int(r-'0')
|
||
}
|
||
return n
|
||
}
|