Add docs/runbooks/FLEET_DEVIN_LOCAL_RUN.md: how developers and coding agents
spin up platform-service + tracker-web + an agent-queue factory so a submitted
job is claimed and run autonomously by the Devin CLI against a target repo
(worked example: learning_ai_notes), pushing a branch and opening a real PR.
Covers: architecture + lifecycle, prerequisites incl. fresh-machine setup
(clone both repos, .env/Cosmos, pnpm -r build so host-run resolves @bytelyst/*
from dist/), all-localhost (no Docker) path as primary + Docker as the
Grafana/Prometheus option, local JWT minting, job submit, factory launch, observe,
PR-state reconcile, safety/cost, teardown, troubleshooting, and a copy-paste
quickstart.
Calls out two gotchas learned in practice: set AQ_FLEET_LEASE_RENEW_SEC < 90 so
the factory heartbeat beats the coordinator's 90s stale-factory reclaim window
(else a busy single-slot factory's in-flight lease is reclaimed mid-run and the
final report is fenced), and a WSL-on-Windows differences section (run inside
WSL, repos off /mnt/c, LF endings, gh/devin/node in WSL, localhost forwarding).
Generated with [Devin](https://cli.devin.ai/docs)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
e6611cae1a