EARLY
Before any session has run. Pre-weekend research and the gate's earliest non-blocking checkpoint.
System reference · not marketing copy
The race weekend runs through five machine phases. A single Conductor
decides which of ~20 registered agents fire in each one, in a fixed order, gated by
what's required and what already exists. Nothing here is invented — every name, order,
and rule below is read straight from scripts/conductor.py's live
default_registry() and f1analytics/conductor.py's dispatch loop.
GitHub Actions cron drifts 50–60+ minutes — too imprecise for session-timed dispatch — so the clock lives outside GitHub entirely.
conductor.py --plan-week reads the real FastF1 session schedule and
writes exact fire times to data/state/triggers_<season>_r<round>.json.
A cron every 15 minutes reads the latest trigger plan from the repo. Stateless — no KV, no Durable Object. Duplicate firings are safe by design.
Any trigger inside the lookback window POSTs a conductor-phase event
to GitHub with {season, round} as payload.
The hosted runner calls schedule.detect_phase(), then dispatches
every agent registered for that phase, in order.
requireda failure here halts the phase. optionalfailures are logged and the phase advances — retried on the next firing.
Before any session has run. Pre-weekend research and the gate's earliest non-blocking checkpoint.
Practice is running. The single hard stop in the whole week — nothing past this phase moves without a human.
Grid is set. A shorter blocking checkpoint before race day — 12 hours to decide, not 48.
The quiet phase — one last weather read ahead of lights out.
The long chain. Eleven optional agents run in sequence so each can see the last one's output, then one blocking gate closes the round.
An agent is never asked "did you run." Its declared output file existing IS the completion signal. A re-fire that finds the artifact already there just skips — no separate "already ran" bookkeeping to drift out of sync with reality.
data/state/conductor_state.json
key: "{season}:{round}:{phase}:{agent}"
FastF1 and Jolpica publish on their own clocks after a session flag — the first firing routinely finds a source not ready yet. An optional agent failing gets logged into the report and retried next firing; nothing else waits on it.
A det agent is a direct Python call — no subprocess. An llm agent carries a fixed three-part contract: a provider-neutral prompt file, a built input record the Conductor writes before calling it, and one pre-agreed output artifact the result lands at.