mirror of
https://github.com/RuoJi6/ARTEX.git
synced 2026-10-27 17:03:49 +08:00
feat(agent): get_worker_trace 的 step_ids 超限改为降级返回而非报错
原先 step_ids 超过 5 个直接报错,逼 Agent 重新组织调用、白跑一轮。改为优雅 降级:先去重、剔除非法 id,超出 5 个则返回前 5 步的完整内容,并在结果里用 returned_step_ids / omitted_step_ids / notice 告知本次取了哪些、还有哪些没取, 并提示「若已足够则无需再取剩余」。Agent 立刻拿到数据,是否再取由它自行判断。 保留数量上限的原因不变:detail 是完整原文、不截断,上限用于兜住单个 tool result 的大小。get_task_worker_trace 经 delegateToTask 复用同一实现,一并生效; 两处工具描述同步更新。 验证: 新增 TestGetWorkerTraceStepIDsDegradeGracefully(含去重/非法 id/前5/ omitted/notice/在限内不加提示),连本地 PG 实跑通过;agent 全量无回归。
This commit is contained in:
Autumn-27 committed
1 parent
6a8e13fcbd
commit
412a993c25
4 files changed
+106
-9
No files matched your search
@@ -14,6 +14,12 @@
|
||||
|
||||
- 新建任务的「任务分类」由下拉选择改为可搜索输入框：输入即过滤已有分类，库里没有的名称回车（或点下拉里的「创建」）即时新建并选中，选中项以可移除的 tag 展示；仍限单个分类。
|
||||
|
||||
### Agent
|
||||
|
||||
#### 修改的功能
|
||||
|
||||
- `get_worker_trace` / `get_task_worker_trace` 的 `step_ids` 超过一次上限（5 个）时不再直接报错，改为返回前 5 步的完整内容，并在结果中通过 `returned_step_ids` / `omitted_step_ids` / `notice` 告知本次取了哪些、还有哪些未取，并提示「若已足够则无需再取」。传入的重复与非法 id 会先去重、剔除后再计数。
|
||||
|
||||
#### 修复的问题
|
||||
|
||||
- 任务列表改为按任务 id 降序（最新创建在最前），替换原先的按创建时间排序。同一时刻创建的多个任务时间戳相同、排序不稳定，配合每 10 秒轮询和内存 map 的随机遍历，会导致它们在列表里频繁互换位置。
|
||||
|
||||
@@ -272,3 +272,75 @@ func TestBlackboardToolsReadDirectSources(t *testing.T) {
|
||||
t.Fatalf("inherited trace search: %#v", hits)
|
||||
}
|
||||
}
|
||||
|
||||
// TestGetWorkerTraceStepIDsDegradeGracefully pins the over-cap behaviour: instead
|
||||
// of erroring, get_worker_trace returns the first 5 requested steps and tells the
|
||||
// model which ids it deferred, after de-duplicating and dropping invalid ids.
|
||||
func TestGetWorkerTraceStepIDsDegradeGracefully(t *testing.T) {
|
||||
d := testDB(t)
|
||||
defer d.Close()
|
||||
|
||||
task, err := d.CreateTask("trace-cap", "goal", nil, 0, 0)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
t.Cleanup(func() { _ = d.DeleteTask(task.ID) })
|
||||
store := d.Exploration(task.ExplorationID)
|
||||
intent, err := store.AddIntent(map[string]any{"summary": "cap work"}, 1, nil, "planner")
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
var stepIDs []int64
|
||||
for i := range 6 {
|
||||
sid, err := store.AppendActivity(db.Activity{
|
||||
NodeID: &intent, Worker: "w", Kind: "result", Tool: "HTTP",
|
||||
Summary: fmt.Sprintf("step %d", i), Detail: fmt.Sprintf("detail %d", i),
|
||||
})
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
stepIDs = append(stepIDs, sid)
|
||||
}
|
||||
|
||||
tools := NewToolSet(store, "worker")
|
||||
tools.SetAssetStore(d.Assets(), d.Companies())
|
||||
|
||||
// Request 7 ids: a duplicate of the first, an invalid 0, then all 6 real ids.
|
||||
// After dedup/cleanup that is 6 valid ids — one over the cap.
|
||||
requested := []int64{stepIDs[0], stepIDs[0], 0, stepIDs[1], stepIDs[2], stepIDs[3], stepIDs[4], stepIDs[5]}
|
||||
input, _ := json.Marshal(map[string]any{"intent_id": intent, "step_ids": requested})
|
||||
res := callReadJSON(t, tools.getWorkerTrace(), string(input)).(map[string]any)
|
||||
|
||||
returned, _ := res["returned_step_ids"].([]any)
|
||||
if len(returned) != 5 {
|
||||
t.Fatalf("returned_step_ids=%v, want the first 5", res["returned_step_ids"])
|
||||
}
|
||||
// First 5 distinct valid ids, in request order.
|
||||
wantReturned := []int64{stepIDs[0], stepIDs[1], stepIDs[2], stepIDs[3], stepIDs[4]}
|
||||
for i, raw := range returned {
|
||||
if int64(raw.(float64)) != wantReturned[i] {
|
||||
t.Fatalf("returned[%d]=%v, want %d", i, raw, wantReturned[i])
|
||||
}
|
||||
}
|
||||
omitted, _ := res["omitted_step_ids"].([]any)
|
||||
if len(omitted) != 1 || int64(omitted[0].(float64)) != stepIDs[5] {
|
||||
t.Fatalf("omitted_step_ids=%v, want [%d]", res["omitted_step_ids"], stepIDs[5])
|
||||
}
|
||||
if notice, _ := res["notice"].(string); notice == "" {
|
||||
t.Fatalf("notice missing — model would not know a step was deferred")
|
||||
}
|
||||
if steps, _ := res["steps"].([]any); len(steps) != 5 {
|
||||
t.Fatalf("steps=%d, want 5 detail rows", len(steps))
|
||||
}
|
||||
|
||||
// At or under the cap: no notice, no omitted list.
|
||||
okInput, _ := json.Marshal(map[string]any{"intent_id": intent, "step_ids": stepIDs[:3]})
|
||||
okRes := callReadJSON(t, tools.getWorkerTrace(), string(okInput)).(map[string]any)
|
||||
if _, hasNotice := okRes["notice"]; hasNotice {
|
||||
t.Fatalf("notice present for an in-cap request: %#v", okRes["notice"])
|
||||
}
|
||||
if _, hasOmitted := okRes["omitted_step_ids"]; hasOmitted {
|
||||
t.Fatalf("omitted_step_ids present for an in-cap request: %#v", okRes["omitted_step_ids"])
|
||||
}
|
||||
}
|
||||
+26
-7
@@ -1686,12 +1686,12 @@ func (t *ToolSet) getWorkerTrace() actool.CoreTool {
|
||||
"查看某条意图(work)的【执行过程】（区别于 get_worker_output 只给最终结论）。三种用法：\n"+
|
||||
"① 只传 intent_id → 返回该 work 每一步的摘要流（summary≤100字，含 step_id；只是动作轮廓，不含完整输出）；\n"+
|
||||
"② intent_id + q → 只返回命中关键字的步骤摘要（在摘要和完整输出里都搜；仍只给 summary，要看内容用③）；\n"+
|
||||
"③ intent_id + step_ids（最多5个）→ 返回这些步骤的完整内容(detail)。\n"+
|
||||
"③ intent_id + step_ids → 返回这些步骤的完整内容(detail)；一次最多取 5 个，超出只返回前 5 个并在 notice/omitted_step_ids 里告知未取的。\n"+
|
||||
"典型流程：先①/②定位可疑步骤的 step_id，再用③取其完整输出。不含思考(thinking)步骤。支持直接关联任务的历史 trace；其结果带 source_task_id/inherited=true 且只读。",
|
||||
obj(map[string]any{
|
||||
"intent_id": idp("意图 id（= work 句柄）"),
|
||||
"q": str("关键字：只返回摘要/完整输出命中它的步骤（可选；与 step_ids 互斥）"),
|
||||
"step_ids": map[string]any{"type": "array", "items": map[string]any{"type": "integer"}, "description": "要取完整内容的 step_id（来自①/②返回，一次最多 5 个）"},
|
||||
"step_ids": map[string]any{"type": "array", "items": map[string]any{"type": "integer"}, "description": "要取完整内容的 step_id（来自①/②返回；一次最多取 5 个，多传只返回前 5 个，其余在 omitted_step_ids 里列出）"},
|
||||
"limit": intp("摘要流/检索的返回上限（可选）"),
|
||||
}, "intent_id"),
|
||||
func(_ context.Context, in json.RawMessage) (actool.Result, error) {
|
||||
@@ -1713,16 +1713,26 @@ func (t *ToolSet) getWorkerTrace() actool.CoreTool {
|
||||
if intentNode == nil || intentNode.Kind != db.KindIntent {
|
||||
return actool.Errorf("intent_id 不属于本任务或其直接关联任务"), nil
|
||||
}
|
||||
// ③ detail drill-down by step ids (max 5), thinking excluded by the store.
|
||||
// ③ detail drill-down by step ids, thinking excluded by the store.
|
||||
if len(a.StepIDs) > 0 {
|
||||
// Dedup + drop invalid ids first so garbage/duplicates don't eat into
|
||||
// the per-call cap. detail is returned in full (untruncated), so the
|
||||
// cap bounds one tool result; over the cap we serve the first N and
|
||||
// tell the model exactly which ids were deferred, instead of erroring
|
||||
// and forcing it to re-plan the call.
|
||||
const maxStepIDs = 5
|
||||
var ids []int64
|
||||
seen := make(map[int64]bool)
|
||||
for _, raw := range a.StepIDs {
|
||||
if v := pid(raw); v > 0 {
|
||||
if v := pid(raw); v > 0 && !seen[v] {
|
||||
seen[v] = true
|
||||
ids = append(ids, v)
|
||||
}
|
||||
}
|
||||
if len(ids) > 5 {
|
||||
return actool.Errorf("step_ids 一次最多 5 个（先用不带 step_ids 的调用定位，再分批取）"), nil
|
||||
var omitted []int64
|
||||
if len(ids) > maxStepIDs {
|
||||
omitted = append(omitted, ids[maxStepIDs:]...)
|
||||
ids = ids[:maxStepIDs]
|
||||
}
|
||||
acts, err := t.ts.ActivityByIDsWithSources(ids)
|
||||
if err != nil {
|
||||
@@ -1743,7 +1753,16 @@ func (t *ToolSet) getWorkerTrace() actool.CoreTool {
|
||||
}
|
||||
steps = append(steps, step)
|
||||
}
|
||||
result := map[string]any{"intent_id": id, "steps": steps}
|
||||
result := map[string]any{"intent_id": id, "steps": steps, "returned_step_ids": ids}
|
||||
if len(omitted) > 0 {
|
||||
// returned_step_ids/omitted_step_ids let the model decide programmatically
|
||||
// whether another call is worth it; the notice states the same in prose.
|
||||
result["omitted_step_ids"] = omitted
|
||||
result["notice"] = fmt.Sprintf(
|
||||
"每次最多取 %d 个步骤的完整内容，本次已返回前 %d 个（%v），未取的 %d 个为 %v。"+
|
||||
"若这些内容已足够定位，则无需再取剩余步骤；确需继续时，用这些 step_id 再调一次。",
|
||||
maxStepIDs, len(ids), ids, len(omitted), omitted)
|
||||
}
|
||||
if intentNode.Inherited {
|
||||
inheritedMap(result, intentNode.SourceTaskID)
|
||||
}
|
||||
|
||||
@@ -345,11 +345,11 @@ func (s *Server) toolAddHint() actool.CoreTool {
|
||||
|
||||
func (s *Server) toolGetWorkerTrace() actool.CoreTool {
|
||||
return roTool("get_task_worker_trace",
|
||||
"看指定任务里某个 work(意图)的执行过程：get_task_worker_trace(task_id, intent_id) 看步骤摘要；再带 step_ids=[...] 取那几步完整内容(一次≤5)。",
|
||||
"看指定任务里某个 work(意图)的执行过程：get_task_worker_trace(task_id, intent_id) 看步骤摘要；再带 step_ids=[...] 取那几步完整内容(一次最多 5 个,多传只返回前 5 个)。",
|
||||
objSchema(map[string]any{
|
||||
"task_id": strParam("任务 id"),
|
||||
"intent_id": map[string]any{"type": "integer", "description": "意图 id(该任务里的 work)"},
|
||||
"step_ids": map[string]any{"type": "array", "items": map[string]any{"type": "integer"}, "description": "可选：要取完整内容的步骤 id(≤5)"},
|
||||
"step_ids": map[string]any{"type": "array", "items": map[string]any{"type": "integer"}, "description": "可选：要取完整内容的步骤 id(一次最多 5 个,多传只返回前 5 个,其余在 omitted_step_ids 里列出)"},
|
||||
}, "task_id", "intent_id"),
|
||||
func(ctx context.Context, in json.RawMessage) (actool.Result, error) {
|
||||
return s.delegateToTask(ctx, in, (*agent.ToolSet).GetWorkerTraceTool)
|
||||
|
||||
Reference in new issue
Block a user