2.3 KiB
Row253 granted gate
Do not launch until Sage's CLEAR names this scope. No tests have run yet. Use this worktree, the frozen five blobs in CANDIDATE-SHAS.json and the plain node_modules link. No canonical-checkout node or queue writes during the slot.
Set TMPDIR with mktemp -d /mnt/storage/scratch/tmp/gXXX, export it and use
set -o pipefail. Record the grant and the exact date before launch. Use logs
under this packet's run1/, with the outer log named controller.log.
Neither helper nor wrapper argv may match the busy detector's patterns.
Sequence, stopping to inspect every result:
window.py --stage focused: isolated prior installer and acceptance RED examples, restoring bytes/status in finally, then both focused test files whole GREEN. Expected RED is recorded separately from gate failures.driver.py: 33 unique single-target changes, named base cases and killer runs. Each base must pass in under 30 seconds. bash -n is the only invalid classification. Runtime errors kill a change; unexpected timeouts stop for Sage/Mos. Every run records time, code, raw tee log and byte/status restore.window.py --stage whole: cli whole, webui whole and the ten shell files whole. No omitted tests, patterns or partial scripts in this phase.- Sum counts and seconds from raw logs by script. Sweep process cwd/ancestry for this worktree's node, tee, query and probe descendants. Report existing processes separately. Return the slot immediately, then prepare review.
Every driver invocation needs --grant <Sage-id> --out <record-directory>.
Before each actual test, driver.run checks Sage's busy guard and launches only
on CLEAR. It invokes a bash pipefail/tee pipeline through launch.py, which
loads the structured command after the guard. No command text is embedded in
wrapper argv. All execution is in a new process group and timed out groups
are killed and waited for. Query-owned groups are independently bounded by
product code; the orphan fixture also cleans its own recorded PID.
Two unexpected failures end the window. These drivers stop at the first unexpected base, focused or whole-phase failure for inspection, without automatically rerunning it. Any timeout is a stop. Send a stop to Sage/Mos and do not turn an ungranted retry into a new run.