ppilot command reference¶
ppilot is the scalable Run-production CLI. It exposes exactly two commands:
ppilot
├── run execute plan() / execute(item) with durable recovery
└── produce create independent pVisor Runs from a streaming planner
Dataset discovery, SQL, built-in analysis, find, import/export, and serving are
owned by pchronicle.
run¶
ppilot run plan.py --workers 8 --per-worker 2 --sink ./results
ppilot run plan.py --workers 8 --sink ./results --resume
ppilot run plan.py --check
The script defines plan() and execute(item). pPilot applies bounded
concurrency and backpressure, writes terminal results to the durable sink, and
uses stable task identity for resume and retry. --check validates the plan
and a sample execution without running the full workload.
produce¶
ppilot produce production.py --output ./runs --parallelism 8
ppilot produce production.py --output ./runs --parallelism 8 \
--cluster-network-limit 10mbps -- --dataset train
The planner's plan() may be a synchronous or asynchronous iterator. Each
item describes one Run:
def plan():
for index in range(100):
yield {
"id": f"task-{index:04d}",
"agent": "codex",
"command": ["codex", "exec", f"Solve task {index}"],
"cwd": "/work/eval",
}
Each emitted item gets its own pVisor workspace below --output. The planner
is streamed under the concurrency window, so large batches are not fully held
in memory. The command writes production-report.json; any failed Run makes
the command exit unsuccessfully after the report is durable.
--cluster-network-limit divides a conservative aggregate proxy rate across
the requested parallelism. It requires Gateway capture and does not cover
direct sockets that bypass the explicit proxy.
Runtime ownership¶
Both commands start an in-process, job-scoped Supervisor. pPilot owns planning, leases, retries, reconciliation, and collection. pVisor owns Run execution and the embedded Gateway. pChronicle owns trajectory Dataset operations.
The executable's --help is authoritative for flags and defaults.