━━ policy pack
chhhee10/cost-guard
failproofai policy pack — chhhee10/cost-guard
packs are published by their authors, not audited by failproof ai. read the source before you install.[ read more ↓ ]
━━ take part of it
with no flags you get the pack's own defaults and are shown the rest — what its author marked safe to switch on unattended, not everything it contains. these three skip that.
- one policy
- one category
- look first
━━ 17 policies · 4 categories
compute5
- block-gpu-instance-launch
Block launching 4+ GPU or A100/H100-class instances
- block-oversized-instance-launch
Block launching 24xlarge-and-up or bare-metal instances
- block-large-scale-up
Block scaling nodes or replicas to a large absolute count
- block-managed-cluster-create
Block creating an EKS/GKE/AKS/EMR cluster that bills while idle
- block-terraform-apply-without-plan
Block non-interactive terraform apply with no saved plan file
data4
- block-whole-bucket-recursive-copy
Refuse a recursive copy or sync whose source is a whole bucket
- block-cross-region-bucket-transfer
Refuse a bucket-to-bucket copy whose two ends are in different regions
- block-whole-database-export-or-restore
Refuse a full dump of a managed database, or a snapshot restore into a new instance
- block-unbounded-storage-retention
Refuse multi-year log retention and replication rules; flag versioning with no expiry rule
inference4
- cap-frontier-eval-row-count
Ask for a row cap before an eval loop runs a frontier model over a full dataset
- require-max-output-tokens-in-call-loops
Ask for an output-token cap when a loop calls a hosted model per item
- bound-inference-job-concurrency
Refuse unbounded fan-out on a job that makes one model call per item
- sample-before-embedding-or-finetuning
Ask for a sampled run before a fine-tune or an embedding pass over a full corpus
runaway4
- block-unreaped-background-spend
Refuse to detach a billable job with nothing to reap it
- block-unbounded-retry-loop
Refuse an unbounded loop that re-issues a billed call
- block-unbounded-fanout-to-paid-work
Refuse unlimited parallelism when each worker costs money
- require-deadline-on-billable-wait
Nudge for a deadline on a metered command that can hang
the hub reads this pack's manifest and verifies it against the release's own SHA256SUMS. it never downloads or runs failproofai-pack.mjs — what you see here is what policies add would install.
━━ review before you install
packs are published by their authors. failproof ai does not review, audit, or certify them, and listing here is not an endorsement. the hub checksums every pack against the SHA256SUMS in its own release — that proves the file is the one that release published, and nothing else. it is not a signature: whoever controls the release controls both files, so it says nothing about who wrote the code, and nothing about whether the code is safe.
read the source before you install. a policy runs inside your agent's hook path on every call it matches, and sees the tool input your agent is about to act on. treat a third-party pack exactly as you would any other dependency you are about to give that access to.
this pack's source: chhhee10/cost-guard ↗ · preview what it contains without downloading or running any of it with failproofai policies show chhhee10/cost-guard.