菜单
Execute a flow in this repo by hand — locally for debugging, or through the API to exercise the real production worker. Use when asked to run, test, or trigger a flow outside its schedule, and to avoid corrupting production write paths while doing it.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Register, change, or debug Prefect deployments in this repo — schedules, triggers, parameters, work pools, job variables, secret injection, or dependency/python-version problems in flow runs. Use when editing prefect.yaml or when a flow run fails during pull or dependency install.
Read or debug the daily infra cost snapshot across Fly, Hetzner, Cloudflare, and Neon. Use when asked whether a teardown reduced spend, why a provider total moved, how a resource is attributed to a project, or when adding a cost connector. Also covers Fly volume and snapshot cleanup.
Diagnose the home-pool process worker on heavypad, where all flow runs actually execute. Use when a flow run is stuck in RUNNING, a run died without logs, the worker seems down, or you need to reach the worker box or restart it safely.
Query or inspect the self-hosted Prefect server at prefect-server.waow.tech, or the k3s cluster it runs on. Use when you need flow-run/deployment/block state, are told you have no credentials, are unsure which kubectl context is prod, or an /api/ call returns Unauthorized.
| name | adhoc-runs |
| description | Execute a flow in this repo by hand — locally for debugging, or through the API to exercise the real production worker. Use when asked to run, test, or trigger a flow outside its schedule, and to avoid corrupting production write paths while doing it. |
Two ways to run a flow, and they test different things. Pick deliberately.
Flow files under flows/ are ordinary modules and most have an
if __name__ == "__main__" entrypoint:
uv run python flows/diagnostics.py
uv run python flows/phi_atlas.py --dry-run
Good for iterating on logic. Two traps:
/tmp, not the production
analytics dir (/home/stoat/prefect-analytics on heavypad). Check where a flow
writes before running it, especially anything touching DuckDB or the spend log.Secret.load(...) still needs
PREFECT_API_URL/PREFECT_API_AUTH_STRING pointed at the live server —
see [[server-access]].Local execution does not exercise pull steps, job variables, or injected secret-block values. A flow that works locally can still fail as a deployment.
Runs on the real home-pool worker with real pull steps and job variables:
just prefect deployment run 'diagnostics/diagnostics' --watch
just prefect deployment run 'ingest/ingest' --watch
The form is flow-name/deployment-name. This is the only way to validate the
full path — and note it pulls pushed main, so an uncommitted local fix will
not be present ([[deployments]]).
diagnostics is the intended canary: it is retargeted to the home box
specifically to prove the worker pulls code and executes flows.
Some flows take a dry-run parameter, and some default to it — pds-records is an
operator flow whose default deployment is dry-run list mode. For anything that
writes PDS records, deletes remote resources, or publishes to R2, run dry first
and read the output before running for real.
Ask what the flow actually does at the boundary. Several flows here publish
externally — PDS records, R2 objects, Cloudflare Pages deploys, Discord alerts.
Re-running them is not free and is sometimes visible to other people. Snapshot
flows upsert at rkey=YYYY-MM-DD and so are idempotent within a day; that is a
property of those flows, not a general guarantee.
Cost flow specifics: [[infra-costs]]. Where execution happens: [[home-pool-worker]].