feat: 修复审计落差 #1/#2/#3/#6 + 管理面板移动端适配
#1 工具布尔标记生效(agent-106, 安全红线): loop toolSchemas/runToolCalls 加 allow-set 安全门 restrictSchemas/applyToolSecurityGate——群聊/陌生人只暴露允许工具(默认关), 被拒not_allowed软提示+AuditEvent, owner全工具不变; 身份靠 ActorEnv(非提示词) #2 checkpoint 恢复(agent-109): run.go LoadLatest + resumableStopReason门控(只恢复in-flight工具循环,不出答案不续跑), 幂等不重放副作用, 新会话不误恢复; resume_test 4用例 #3 tool_result(agent-108, 协议层): Anthropic 原生tool_result块(tool_use_id/is_error)非user纯文本 + OpenAI tool消息补tool_call_id; Content加ToolCallID/IsError等; (loop接线层2处留待,见备注) #6 capacity_awaken推送(agent-107): onCapacity从只Debug改推WS hub(BroadcastCapacityAwaken, sessionId定向+payload对齐前端AwakenPayload) 管理面板移动端(agent-110): AdminApp 汉堡菜单+抽屉侧栏+标题truncate防换行, RBAC守卫不受影响 全仓 build/vet/test绿 + 前端build绿; 前端无涉差异已对齐 备注: #3 loop接线层2处(assistant工具轮append历史 + Message.ToolCallID/IsError)留给后续; #4代审/#5子代理触发/#7拆文件后置(主人说后三不急)
This commit is contained in:
Aska committed
1 parent
aedae2190b
commit
7e39153a69
24 files changed
+1286
-56
No files matched your search
@@ -0,0 +1,194 @@
|
||||
package loop
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"testing"
|
||||
|
||||
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
|
||||
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
|
||||
"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
|
||||
)
|
||||
|
||||
// fakeCheckpointStore is a seam for asserting the resume path without a real
|
||||
// database. It records every LoadLatest call and returns a scripted checkpoint or
|
||||
// error, so a test can prove the loop asks the store for the session's checkpoint
|
||||
// and resumes (or stays fresh) based on what comes back.
|
||||
type fakeCheckpointStore struct {
|
||||
latest *persist.Checkpoint
|
||||
latestErr error
|
||||
loadCalls []string
|
||||
}
|
||||
|
||||
func (f *fakeCheckpointStore) Save(context.Context, *persist.Checkpoint) error { return nil }
|
||||
|
||||
func (f *fakeCheckpointStore) LoadLatest(_ context.Context, agentID string) (*persist.Checkpoint, error) {
|
||||
f.loadCalls = append(f.loadCalls, agentID)
|
||||
return f.latest, f.latestErr
|
||||
}
|
||||
|
||||
// TestRun_ResumeFromCheckpoint_ContinuesNoRepeat simulates a long task that was
|
||||
// interrupted mid-tool-loop: the loop left a tool_calls checkpoint at step 2 with
|
||||
// the message window as it stood after folding two tool batches in, then the
|
||||
// process died before step 3 could run. A fresh Run on the same sessionID must
|
||||
// resume at step 3 — rebuilding the context from the checkpoint rather than
|
||||
// restarting — and must not re-execute the already-completed steps.
|
||||
func TestRun_ResumeFromCheckpoint_ContinuesNoRepeat(t *testing.T) {
|
||||
t.Parallel()
|
||||
ctx := context.Background()
|
||||
ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "final"}}}}
|
||||
a := newTestAgent(t, ad, nil)
|
||||
cp := newTestCheckpointStore(t)
|
||||
l := New(Options{Agent: a, Checkpoint: cp, MaxRounds: 8})
|
||||
|
||||
const sess = "sess-resume"
|
||||
interrupted := []model.Message{
|
||||
model.TextMessage(model.RoleUser, "do the task"),
|
||||
model.TextMessage(model.RoleTool, "tool aa result: va"),
|
||||
model.TextMessage(model.RoleTool, "tool bb result: vb"),
|
||||
}
|
||||
// Seed exactly what a crashed run leaves behind: a tool_calls checkpoint at
|
||||
// step 2 with the window computed after step 2's tool batch.
|
||||
l.saveCheckpoint(ctx, sess, 2, interrupted, agent.StopReasonToolCalls)
|
||||
|
||||
res, err := l.Run(ctx, sess, []model.Message{model.TextMessage(model.RoleUser, "do the task")})
|
||||
if err != nil {
|
||||
t.Fatalf("Run: %v", err)
|
||||
}
|
||||
|
||||
// Resume continues at step 3 (the interrupted step), not step 1.
|
||||
if res.Steps != 3 {
|
||||
t.Fatalf("Steps = %d, want 3 (resumed at step 3)", res.Steps)
|
||||
}
|
||||
// No completed step's tool call is re-executed.
|
||||
if res.ToolCallsExecuted != 0 {
|
||||
t.Fatalf("ToolCallsExecuted = %d, want 0 (completed steps not re-run)", res.ToolCallsExecuted)
|
||||
}
|
||||
if res.Text != "final" {
|
||||
t.Fatalf("Text = %q, want final", res.Text)
|
||||
}
|
||||
|
||||
// The resumed run makes exactly one model call (step 3), fed the restored
|
||||
// window — not a restart with a fresh conversation.
|
||||
if len(ad.reqs) != 1 {
|
||||
t.Fatalf("resumed run made %d model calls, want 1", len(ad.reqs))
|
||||
}
|
||||
got := ad.reqs[0].Messages
|
||||
if len(got) != len(interrupted)+1 {
|
||||
t.Fatalf("resumed request messages = %d, want %d (system + restored window)", len(got), len(interrupted)+1)
|
||||
}
|
||||
if got[0].Role != model.RoleSystem {
|
||||
t.Fatalf("leading message role = %q, want system (instruction)", got[0].Role)
|
||||
}
|
||||
for i, m := range interrupted {
|
||||
if got[i+1].Role != m.Role || messageText(got[i+1]) != messageText(m) {
|
||||
t.Fatalf("restored message %d mismatch: got %+v, want %+v", i, got[i+1], m)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// TestRun_NewSession_NotResumed proves a different sessionID never resurrects an
|
||||
// old checkpoint: the caller opens a brand-new session, LoadLatest misses, and the
|
||||
// loop runs fresh from step 1 with the caller's own messages.
|
||||
func TestRun_NewSession_NotResumed(t *testing.T) {
|
||||
t.Parallel()
|
||||
ctx := context.Background()
|
||||
call := `{"tool_calls":[{"name":"echo","args":{"q":"x"}}]}`
|
||||
ad := &mockAdapter{chat: []mgResp{
|
||||
{resp: model.Response{Content: call}},
|
||||
{resp: model.Response{Content: "done"}},
|
||||
}}
|
||||
td := dispatcherWith(t, &mockTool{name: "echo", result: "hi"})
|
||||
a := newTestAgent(t, ad, td)
|
||||
cp := newTestCheckpointStore(t)
|
||||
l := New(Options{Agent: a, Checkpoint: cp, MaxRounds: 8})
|
||||
|
||||
// sess-a holds an interrupted run's checkpoint.
|
||||
l.saveCheckpoint(ctx, "sess-a", 2, []model.Message{
|
||||
model.TextMessage(model.RoleUser, "a task"),
|
||||
model.TextMessage(model.RoleTool, "tool echo result: hi"),
|
||||
}, agent.StopReasonToolCalls)
|
||||
|
||||
res, err := l.Run(ctx, "sess-b", []model.Message{model.TextMessage(model.RoleUser, "go")})
|
||||
if err != nil {
|
||||
t.Fatalf("Run: %v", err)
|
||||
}
|
||||
if res.Steps != 2 {
|
||||
t.Fatalf("Steps = %d, want 2 (fresh two-step run)", res.Steps)
|
||||
}
|
||||
if res.ToolCallsExecuted != 1 {
|
||||
t.Fatalf("ToolCallsExecuted = %d, want 1", res.ToolCallsExecuted)
|
||||
}
|
||||
// The first model call must start from the caller's fresh message, never from
|
||||
// sess-a's restored window.
|
||||
first := ad.reqs[0].Messages
|
||||
if len(first) != 2 {
|
||||
t.Fatalf("fresh run first request messages = %d, want 2 (system + user)", len(first))
|
||||
}
|
||||
if messageText(first[1]) != "go" {
|
||||
t.Fatalf("fresh run first user message = %q, want 'go'", messageText(first[1]))
|
||||
}
|
||||
}
|
||||
|
||||
// TestRun_Resume_LoadLatestCalled asserts Run asks the store for the session's
|
||||
// latest checkpoint and resumes from it (LoadLatest → tool_calls checkpoint →
|
||||
// continue at the next step).
|
||||
func TestRun_Resume_LoadLatestCalled(t *testing.T) {
|
||||
t.Parallel()
|
||||
ctx := context.Background()
|
||||
ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "final"}}}}
|
||||
a := newTestAgent(t, ad, nil)
|
||||
|
||||
buf, err := json.Marshal(checkpointState{
|
||||
Step: 2,
|
||||
Messages: []model.Message{model.TextMessage(model.RoleUser, "keep going")},
|
||||
StopReason: agent.StopReasonToolCalls,
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatalf("marshal checkpoint: %v", err)
|
||||
}
|
||||
fake := &fakeCheckpointStore{latest: &persist.Checkpoint{ID: "sess:2", AgentID: "sess", Step: 2, State: buf}}
|
||||
l := New(Options{Agent: a, Checkpoint: fake, MaxRounds: 8})
|
||||
|
||||
res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "new")})
|
||||
if err != nil {
|
||||
t.Fatalf("Run: %v", err)
|
||||
}
|
||||
if len(fake.loadCalls) != 1 || fake.loadCalls[0] != "sess" {
|
||||
t.Fatalf("LoadLatest calls = %v, want exactly ['sess']", fake.loadCalls)
|
||||
}
|
||||
if res.Steps != 3 {
|
||||
t.Fatalf("Steps = %d, want 3 (resumed at step 3)", res.Steps)
|
||||
}
|
||||
got := ad.reqs[0].Messages
|
||||
if len(got) != 2 || messageText(got[1]) != "keep going" {
|
||||
t.Fatalf("resumed request = %+v, want [system, 'keep going']", got)
|
||||
}
|
||||
}
|
||||
|
||||
// TestRun_Resume_NotFound_Fresh asserts LoadLatest is still called on the fresh
|
||||
// path and that ErrNotFound yields a full restart from step 1 with the caller's
|
||||
// messages (no checkpoint to resurrect).
|
||||
func TestRun_Resume_NotFound_Fresh(t *testing.T) {
|
||||
t.Parallel()
|
||||
ctx := context.Background()
|
||||
ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Content: "hello"}}}}
|
||||
a := newTestAgent(t, ad, nil)
|
||||
fake := &fakeCheckpointStore{latestErr: persist.ErrNotFound}
|
||||
l := New(Options{Agent: a, Checkpoint: fake, MaxRounds: 8})
|
||||
|
||||
res, err := l.Run(ctx, "sess", []model.Message{model.TextMessage(model.RoleUser, "hi")})
|
||||
if err != nil {
|
||||
t.Fatalf("Run: %v", err)
|
||||
}
|
||||
if len(fake.loadCalls) != 1 || fake.loadCalls[0] != "sess" {
|
||||
t.Fatalf("LoadLatest calls = %v, want exactly ['sess']", fake.loadCalls)
|
||||
}
|
||||
if res.Steps != 1 {
|
||||
t.Fatalf("Steps = %d, want 1 (fresh run)", res.Steps)
|
||||
}
|
||||
got := ad.reqs[0].Messages
|
||||
if len(got) != 2 || messageText(got[1]) != "hi" {
|
||||
t.Fatalf("fresh request = %+v, want [system, 'hi']", got)
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user