菜单
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.
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。
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.
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.
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.
基于 SOC 职业分类
| name | infra-costs |
| description | 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. |
The costs flow (cron 0 8 * * *, daily 08:00 UTC) collects provider billing,
writes an io.zzstoatzz.cost.snapshot PDS record at rkey=YYYY-MM-DD (upserted,
so re-runs are idempotent), and surfaces it at hub.waow.tech.
Use the pdsx MCP tools. list_records and get_record cover this with
structured output and no handle/DID/PDS resolution dance:
list_records(repo="zzstoatzz.io", collection="io.zzstoatzz.cost.snapshot", limit=30)
get_record(repo="zzstoatzz.io", collection="io.zzstoatzz.cost.snapshot", rkey="2026-07-24")
pdsx MCP does full record CRUD — reads and create_record/update_record/
delete_record. It is not read-only. Only blob upload, batch JSONL operations,
and permissioned spaces need the pdsx CLI.
Raw HTTP only as a fallback when the MCP isn't available. Records are public, so no auth is needed:
DID=$(curl -s "https://api.bsky.app/xrpc/com.atproto.identity.resolveHandle?handle=zzstoatzz.io" | jq -r .did)
curl -s "https://pds.zzstoatzz.io/xrpc/com.atproto.repo.listRecords?repo=$DID&collection=io.zzstoatzz.cost.snapshot&limit=30" \
| jq -r '.records | sort_by(.uri) | .[] | [(.uri|split("/")|last), (.value.total/100)] | @tsv'
Record fields are total, currency, byProvider, byProject, lineItems,
periodStart, periodEnd, generatedAt. There is no items field —
jq '.value.items' silently yields nothing and looks like zero cost.
Snapshots are a measurement, and the measurement has changed. Diff
lineItems by service, not byProvider/byProject totals, or a coverage
change will masquerade as a spend change.
Known discontinuity: on 2026-07-20 the Fly connector changed shape
(7a212e8). Before, one item per app (~26 items) with machines and volumes
bundled and snapshots not billed at all. After, a
app:compute / app:volumes / app:snapshots split (~55 items) with snapshot
storage newly counted — roughly $11/mo of previously invisible cost appeared.
Any before/after window spanning that date understates real savings.
Worked example: deleting the typeahead builder removed a $15/mo volume, yet the
typeahead project total moved only ~$2.40, because newly-counted snapshots
absorbed the difference.
There is no provider-side project tag. Attribution is by resource name
pattern, longest match wins, in packages/mps/src/mps/costs/projects.py
(_RESOURCE_PATTERNS). So relay.waow → relays while relay-api-staging →
plyr.fm. Adding a resource whose name doesn't match a pattern lands it in
misc or unattributed — add the pattern and a project_for assertion in
tests/test_costs.py.
A connector that raises is collected into failed_providers, and the flow then
refuses to publish rather than shipping a snapshot with a silent hole
(flows/costs.py). Never let a connector swallow an error and return partial
results — an omitted resource must never be counted as $0.
Hetzner specifics: Cloud API tokens are project-scoped (no account-wide
token), so the flow loads a hetzner-tokens Secret block of {label: token}
and merges all projects, deduping by server id. There are currently two project
tokens covering four Cloud servers — relay-eval, relay, zlay,
prefect-server — and that is the entire Hetzner footprint.
Hetzner Robot (the legacy dedicated/auction product) is deliberately out of
scope: there are no dedicated servers on this account. hetzner-edge on the
tailnet is not a separate machine — it is prefect-server under a different
tailnet hostname, same public IP. Don't build a Robot connector without first
confirming a dedicated server actually exists; one was written and reverted for
exactly this reason.
fly volumes destroy <id> is broken without -a — its app-lookup GraphQL
query fails with Field 'node' doesn't exist on type 'Queries'. This is
Fly-side, not a stale client. Always pass the app:
fly volumes destroy -y -a <app> <vol_id> [<vol_id>...]
fly volumes snapshots list hits the same broken query with no -a escape.
Use the Machines REST API instead:
TOK=$(fly auth token)
curl -s -H "Authorization: Bearer $TOK" \
"https://api.machines.dev/v1/apps/<app>/volumes/<vol_id>/snapshots" | jq .
curl -s -H "Authorization: Bearer $TOK" \
"https://api.machines.dev/v1/apps/<app>/volumes/<vol_id>" \
| jq '{size_gb,snapshot_retention,auto_backup_enabled,blocks,blocks_free}'
Destroyed volumes linger in a pending-destroy state and still appear via the
REST API after fly volumes list stops showing them. The connector ignores
them (3efa153), so they are not double-counted.
Every volume defaults to auto_backup_enabled: true,
snapshot_retention: 5 — a daily snapshot at ~00:02 UTC with a rolling 5-day
window. The count is a bounded steady state, not accumulation, and snapshots are
never read except on restore. Deleting them individually is pointless; they
regenerate. The only real lever is retention:
fly volumes update <vol_id> --snapshot-retention <days> -a <app>
Retention changes apply to future aging-out, so cost falls over the following days rather than immediately.
Judge retention by whether the data is authoritative or derived. typeahead-search
holds a derived index — typeahead-index rebuilds it every 3 days and
publishes to R2, explicitly "no ingress, no SLA" — so it is set to 2 days.
pds-zzstoatzz-io is authoritative PDS data and stays at 5. Do not lower
retention account-wide.
Running the flow by hand is [[adhoc-runs]]; the Secret blocks it needs are injected per [[deployments]].