Orchestrata Ultimate › Plans › KIMI-TASK-flakes-and-unit
KIMI TASK — two masked flakes and one unversioned unit
Source: plans/KIMI-TASK-flakes-and-unit.md at 8a2cb42ad · last changed 2026-08-10 · first committed 2026-08-10 · 72 lines.
Prose pending.
Plan text
KIMI TASK — two masked flakes and one unversioned unit
Base: origin/master @ e1793c1f (pushed). Work in your own worktree with its
own env (RUL-021 — no symlinked .venv, no exported PYTHONPATH).
Three independent items. Do them in order; each is committable on its own.
Do not push without authorization in #agents.
1. tests/refactor/rename/test_lsp_transport.py — masked flake
Measured in the full suite on 2026-08-10 (32454 passed / 0 failed / 0 errors /
0 skipped, 1610 modules, 422s): this module failed inside its batch and
passed on the solo retry. It is logged in .flake_recoveries.ndjson.
A recovery is a masked red. The suite is green only because the runner ran it a second time on an idle machine.
- Reproduce it under load — the module fails in a BATCH, not solo. Run the batch it lands in, not the module by itself (that is the whole point).
- The transport is an LSP client over a subprocess pipe; look for a timeout or a read that assumes the server answers within a fixed window.
- Fix the CODE if the code is wrong, the TEST if the test assumes a speed it has no right to. Do not widen a timeout to make a race disappear — say what the race is.
- No
@pytest.mark.skip, no retry decorator, no mock (RUL-002).
2. Conversation-consumer flush flake
tests/knowledge_graph/test_conversation_consumer.py,
TestConversationQueueConsumerFlush::test_flush*. Eight recorded recoveries,
every one with the same signature: llen(processing_key) measured 0 where
the test expects 1.
Same class as #1 and already characterised — the assertion reads a Redis list length at a moment when the value it wants is in flight. Establish WHERE the window is (producer, consumer, or the test's own read) before changing anything, and fix the side that is actually wrong.
3. orchestrata-bot.service is unversioned
The unit runs live in ~/.config/systemd/user/orchestrata-bot.service and has
no copy under deploy/systemd/. Consequences:
deploy/systemd/install.shcannot install it, so a deployment to a fresh machine silently omits it.scripts/health_check.py's unit-drift check compares the two trees; a unit present in only one of them is invisible to it.
Do:
- Copy the live unit into
deploy/systemd/, matching the conventions of the four units already there (read them first —ExecStartPreKG-readiness gate,EnvironmentFilelayering,User=). - Register it in
install.sh. - Prove the drift gate now sees it: make the two copies differ, show the gate FAILS, restore, show it PASSES. Put that evidence in the commit message.
Do not hand-edit the live unit to match; the repo copy is the source of truth and the live one must be reproducible from it.
Standing rules
- Full
python scripts/gates.pybefore ANY commit (RUL-008). - Never run tests unobserved:
scripts/run_tests_streaming.py, backgrounded (RUL-009). Cap any run at 5 minutes. - No mocks / stubs / fakes / monkeypatch / skips (RUL-002).
- No AI attribution in commits — Mohamed is the only author (RUL-007).
- Report back in
#agentsaskimi_agent: what you changed, the commit sha, the branch, and the raw test numbers.