test: make multi-worker bob setup deterministic - #985
Conversation
📊 CI Metrics ReportSummary
By Role
Per-Test Breakdown
Trends✅ 2 test(s) improved (fewer LLM calls) Generated by HiClaw CI on 2026-07-13 09:58:10 UTC |
3a7a1e4 to
067fd41
Compare
|
@shiyiyue1102 Hi! This PR makes the multi-worker Just rebased onto current |
067fd41 to
78e9385
Compare
78e9385 to
43d1048
Compare
Pre-create the 3 workers (alice, bob, charlie) via hiclaw apply worker and the shared project via create-project.sh + immediate meta.json activation + MinIO sync, before sending the 4-phase task to Manager. Manager now focuses on phase coordination only — the worker + project setup steps that previously stalled on Manager's tool-guard approval prompt (jq | mv | cat on meta.json) are now in the test runner and never enter the LLM control loop. Mirrors the fix in agentscope-ai#985 for test-06-multi-worker.sh (bob creation flake). Both fixes isolate LLM-driven coordination from infrastructure setup so the test measures coordination quality, not LLM timing. Fixes the SHARD_B copaw/hermes test-14 failure seen on agentscope-ai#984 (run 28926397804, job 85816940745), where Manager hung for 30 minutes on a tool-guard approval prompt while trying to update meta.json status. Refs: agentscope-ai#984
|
CI update: run #1647 failed before any integration test started. All affected jobs exited during embedded install because the base workflow reads the new |
|
Fresh independent evidence after the repository secret recovered: draft #1041's base-branch run 29489698196 reached the full copaw/hermes SHARD_A suite and reproduced the setup race this PR removes. In unpatched The current red checks on #985 are still the older pre-test secret failure; this newer run does not include #985, but directly reproduces its target failure on the current base. No patch change is needed. |
1 participant
Summary
test-06-multi-workercreate Bob viahiclaw apply workerinstead of a Manager DM requestREPLYWhy
test-02-create-workeralready covers Manager-mediated worker creation. Repeating that LLM interaction intest-06made unrelated PR checks depend on Manager backlog and Tool Guard timing rather than multi-worker collaboration.Two refreshed #984 runs exposed both races:
29233390641: Bob was provisioned, but a Tool Guard approval was pending when the collaborative-task message arrived, so that message denied the tool instead of becoming a new task29236623909: all tests throughtest-05passed, buttest-06aborted because the Manager DM did not become quiet within 180 seconds, before Bob setup even startedDirect controller setup removes both dependencies while preserving strict checks for Bob's persisted state, runtime, Higress consumer, and MinIO files. The test still exercises the actual Manager collaboration flow after setup.
Verification
bash -n tests/test-06-multi-worker.shgit diff --checkmain(6f0c7da)I could not run the embedded integration test locally because this environment cannot access
/var/run/docker.sock(permission denied).