Multi-agent coordination (2026-06-11)
⚠ 21:1xZ incident: a fleet-contention campaign (16/18 runs, ~70 min) and a
queued bench.py run were SIGKILLed mid-flight — likely a "cleanup duties"
sweep mistaking 32 active bench-fleet-tmp mounts for leaks. All rows were
lost (that lane flushed only at exit; now fixed with .partial.jsonl flushing
on branch mrmrs-751c3b). Do not kill processes that hold or wait on the
measurement lock (oak-bench-measurement.lock names the holder); active
wt-NNN mounts under oak-parallel-contention/runs/ belong to a
running campaign. Leaked-mount cleanup is safe only when no campaign process
is alive. The fable session's rerun chain (core-tuned + fleet) is queued
behind your mount_vs_clone run via the lock.
Two agent sessions are working this repo. We hit oak's false-divergence trap by racing one branch from two writers (see prompts/oak-fix-handoff.md P0). New protocol, per the owner:
- One workspace per agent, one branch per agent. This checkout
(
~/o/oak-benchmarks) belongs to the long-running session that started 2026-06-11 01:53. The fable session works in its own clone at~/o/oak-benchmarks-fableon branchmrmrs-751c3b(carries everything frombench-accuracy-pushas of ~20:55Z, 189 tests green). - Never push to another agent's branch. Fold work together by merging to
main (integrator clone:
oak switch <branch>+oak merge) or let the owner merge. - Measurement campaigns are running detached from THIS checkout (fleet contention, agent zero-shot/cheat-sheet, tuned-git core, mount-vs-clone; logs in /tmp/oakbench-fixtures/*.log). Please avoid editing scripts/ here until they finish — mid-campaign script edits change later reps. Code work should happen in your own branch/workspace and merge after.
- Disposable server repos in use: oak/bench-fleet-tmp (fleet lane), oak/bench-sync-tmp (sync + task-loop), oak/bench-large-mirror + github oakdotspace/bench-large-mirror (mount-vs-clone), github oakdotspace/bench-task-loop-tmp (task-loop git track).