vrr/qwen-code
2
0
Fork 0
mirror of https://github.com/QwenLM/qwen-code.git synced 2026-10-24 07:49:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 17

test(managed-agent): reuse a same-key cancel (FG11d)

Browse source
This commit is contained in:
Shaojin Wen 2026-10-21 13:52:05 +08:00
parent d625f663b9
commit acdab642b7
6 changed files with 56 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

7
.github/workflows/sdk-java.yml vendored
View file

@ -775,8 +775,9 @@ jobs:
# the fork to 10621 s and Verify to 188 minutes; the FG12d
# cancel-before-dispatch cell adds 18 s, moving the fork to 10639 s; the
# FG11d retry-reuse cell adds 11 s, moving the fork to 10650 s; the FG11d
# settled-outcome cell adds 9 s, moving the fork to 10659 s, wrapping
# fork + 600 s plus margin (11259 s of fork plus wrap);
# settled-outcome cell adds 9 s, moving the fork to 10659 s; the FG11d
# cancel-retry cell adds 12 s, moving the fork to 10671 s, wrapping
# fork + 600 s plus margin (11271 s of fork plus wrap);
# the Broker fault gates use 20,
# and admission plus O4 keep per-fork ceilings) plus ~13 minutes for the
# uncapped setup steps (Checkout.. cgroup delegation, the report
@ -881,7 +882,7 @@ jobs:
# fork leaves no per-test failure lines, so the main-CI failure
# analyzer can only file an undiagnosable per-commit issue (#13503).
# hosted-process-ci.test.js pins this ordering; the fork budget is
# 10659 s, so the step wraps to fork + 600 s plus margin (188 min).
# 10671 s, so the step wraps to fork + 600 s plus margin (188 min).
timeout-minutes: 188
env:
MAVEN_ARGS: '--settings ${{ runner.temp }}/setup-java-m2/settings.xml --toolchains ${{ runner.temp }}/setup-java-m2/toolchains.xml'

3
docs/design/2026-10-19-automation-lifecycle-fault-gates.md
View file

@ -65,6 +65,9 @@ asserts:
arriving after settlement is refused with `409 task_action_unavailable`
(`The task does not accept a cancellation now.`), and the run stays
`completed` with its occurrence `fired`;
- a cancel retried under the same key reuses its committed operation and changes
nothing — delivered: the second request answers the cancelled task, the run
stays `cancelled` and the occurrence settles once;
- a retry under the same key reuses the original operation and run — delivered:
the second `POST /runs` answers the run it already made, records no second
occurrence and starts no second Turn. That a parked cancel is never

6
docs/design/2026-10-19-automation-lifecycle-fault-gates.zh-CN.md
View file

@ -50,12 +50,14 @@ retire、retire 之后无 slot fire），FG11b 钉住账本周围的存储故障
- 被取消的 run 依据其已提交的证明进入 `cancelled`，occurrence 恰好结算一次；
- schedule、未来 slot、兄弟 run 与前台工作不受影响；
- 在任何尝试之前被取消的排队 input 结算且不调用模型。自然完成的 run 保留其结果
已交付：结算之后到达的取消被 `409 task_action_unavailable`（`The task does
not accept a cancellation now.`）拒绝，run 保持 `completed`、occurrence 保持
已交付：结算之后到达的取消被 `409 task_action_unavailable`（`The task does not
accept a cancellation now.`）拒绝，run 保持 `completed`、occurrence 保持
`fired`；
- `cancel-stops-only-the-target` 已交付：把 wake turn 扣在模型上时，公开 task
cancel 让 task 与 run 在一秒内结算为 `cancelled`，schedule 保持 `live`；反之在
收到 `202` 后立刻释放模型，turn 会先跑完——本格避开的正是这个顺序。
- 同一 key 重试取消复用其已提交的 operation 且不产生新动作——已交付：第二次请求
返回已取消的 task，run 保持 `cancelled`、occurrence 恰好结算一次；
- 同一 key 的重试复用原 operation 与 run——已交付：第二次 `POST /runs` 返回它
已经创建的 run，不产生第二个 occurrence，也不起第二个 Turn。被扣住的 cancel
绝不重投仍未覆盖。

7
packages/managed-agent-server/pom.xml
View file

@ -579,8 +579,11 @@
measured 5.3 s): healthy 10050 s, fork 10650 s,
Verify 188 minutes. The FG11d settled-outcome
cell adds 9 s (2x its measured 4.5 s): healthy
10059 s, fork 10659 s, Verify 188 minutes. -->
<forkedProcessTimeoutInSeconds>10659</forkedProcessTimeoutInSeconds>
10059 s, fork 10659 s, Verify 188 minutes. The
FG11d cancel-retry cell adds 12 s (2x its
measured 5.8 s): healthy 10071 s, fork 10671 s,
Verify 188 minutes. -->
<forkedProcessTimeoutInSeconds>10671</forkedProcessTimeoutInSeconds>
</configuration>
<executions>
<execution>

33
packages/managed-agent-server/src/test/java/com/alibaba/qwen/code/managedagent/HostedAutomationFaultGatesIT.java
View file

@ -1467,6 +1467,39 @@ class HostedAutomationFaultGatesIT {
request("DELETE", "/v1/agent-automations/" + id, null, "fg11d-settled-retire", 202);
}
// FG11d: retrying a cancel under the same key reuses its operation. The
// second request answers the already-cancelled task, records nothing new and
// starts no Turn.
@Test
@Timeout(600)
void fg11dCancelRetryReusesItsOperationOnMySql() throws Exception {
boot("fg11d");
String session = createSession("cancel-retry");
String id = request("POST", "/v1/agent-automations",
definition(session, "FG11A_HOLD", "0 0 1 1 *"), "fg11d-cancel-retry-base", 202).path("id").asText();
String run = request("POST", "/v1/agent-automations/" + id + "/runs", null, "fg11d-cancel-retry-run", 202)
.path("id").asText();
waiting().atMost(Duration.ofSeconds(60)).until(() -> modelPrompts.stream()
.anyMatch(prompt -> prompt.contains("manual:fg11d-cancel-retry-run")));
JsonNode task = request("GET", "/v1/agents/sessions/" + session + "/tasks", null, null, 200)
.path("data").path(0);
assertThat(task.path("kind").asText()).as("the newest task is the run's").isEqualTo("automation_run");
String taskId = task.path("id").asText();
request("POST", "/v1/agents/sessions/" + session + "/tasks/" + taskId + "/cancel", null,
"fg11d-cancel-retry", 202);
waiting().atMost(Duration.ofSeconds(120)).untilAsserted(() -> assertThat(runState(run))
.as("the run is cancelled").isEqualTo("cancelled"));
// The retry reuses the committed cancel instead of acting twice.
request("POST", "/v1/agents/sessions/" + session + "/tasks/" + taskId + "/cancel", null,
"fg11d-cancel-retry", 202);
assertThat(runState(run)).as("the retry changes nothing").isEqualTo("cancelled");
assertThat(jdbc.queryForMap("SELECT outcome FROM ma_schedule_occurrence WHERE tenant_id = ?"
+ " AND schedule_id = ? AND occurrence_key = ?", tenant, id, "manual:fg11d-cancel-retry-run"))
.as("the occurrence settles once").containsEntry("outcome", "fired");
heldModel.countDown();
request("DELETE", "/v1/agent-automations/" + id, null, "fg11d-cancel-retry-retire", 202);
}
private void boot(String letter) throws Exception {
assertThat(System.getProperty("mysql.url")).as("FG11 requires real MySQL").startsWith("jdbc:mysql:");
this.letter = letter;

12
scripts/tests/hosted-process-ci.test.js
View file

@ -766,13 +766,14 @@ describe('Hosted real-process gates', () => {
// 18 s (2x its measured 9.1 s), so the fork moves to 10639; the FG11d
// retry-reuse cell adds 11 s (2x its measured 5.3 s), moving the fork to
// 10650 and the step stays 188 minutes; the FG11d settled-outcome cell adds
// 9 s (2x its measured 4.5 s), moving the fork to 10659 and the step
// stays 188 minutes):
// 9 s (2x its measured 4.5 s), moving the fork to 10659; the FG11d
// cancel-retry cell adds 12 s (2x its measured 5.8 s), moving the fork to
// 10671 and the step stays 188 minutes):
// the relational check above tolerates
// anything up to step*60, so the value is pinned literally here —
// an unnoticed raise has no other red to land on.
if (profile === 'hosted-harness-mysql') {
expect(forkSeconds).toBe(10659);
expect(forkSeconds).toBe(10671);
}
},
);
@ -858,12 +859,13 @@ describe('Hosted real-process gates', () => {
// FG12d cancel-before-dispatch cell adds 18s (2x its measured 9.1s), so it
// becomes 10039s; the retry-reuse cell adds 11s (2x its measured 5.3s), so it
// becomes 10050s; the settled-outcome cell adds 9s (2x its measured 4.5s),
// so it becomes 10059s. The largest
// so it becomes 10059s; the cancel-retry cell adds 12s (2x its measured 5.8s),
// so it becomes 10071s. The largest
// ceiling (600s) plus
// that healthy sum still fits the fork. A larger ceiling lets failsafe
// kill the fork mid-run, and the classes left unrun never report, failing
// the gate undiagnosably (#13780).
const hostedForkHealthySeconds = 10059;
const hostedForkHealthySeconds = 10071;
const hostedForkSeconds = () =>
Number(
read('packages/managed-agent-server/pom.xml')