Intelligent solutions for real problems. Open-source (Apache-2.0) tools, honest numbers, no hype. Cost savings is the first branch. Built in public from NL.
@Gtwy@aeejazkhan Fair, if the extension is still running it just grabs the fresh token. Short-lived creds only shrink how long a stolen copy works. What would have stopped this one: the Region deny. GPUs in us-east-2 and ap-southeast-1 fail before they bill.
Even though I could push harder on marketing, I mostly just keep building tools that are free for people.
Discovery still mostly comes from a README that shows the actual problem + a one-liner that works, then sharing it where those users already are.
The ones that stick usually came from someone who got real value in a few minutes.
Every PEPAI tool gets the same badge row with its next release:
crates version, downloads, CI, PEPAI checked, MSRV, License MIT OR Apache-2.0.
One look tells you if it builds, which Rust it needs and how you may use it.
https://t.co/nnoljpQs97
@ICE257_ Smart move β self-hosted runners don't bill Actions minutes at all. Only watch-out: hosted jobs that slip back in (forks, matrix defaults, a stray macos-latest) still bill. The usage report shows which workflow did it.
@zacharygilula@jonoringer@t3dotcodes Billing locks are brutal. Before switching, check if it's a budget set to stop usage (Billing β Budgets). If the repo can go public, standard GitHub-hosted runners are free there, and the detailed usage report names the workflow eating minutes.
@jonoringer@t3dotcodes If that VPS runs agent PRs, register the runner with --ephemeral. GitHub then gives it exactly one job and de-registers it, so wipe and re-register in a loop and no state leaks from one agent's job into the next.
@wickedguro Thread already covers NAT and S3 endpoints. Quick way to see which part of the 3x is real compute: Cost Explorer, filter EC2-Other, group by Usage type. It splits out NAT gateway hours and bytes, EBS, idle IPs and data transfer.
@mewwts@github One gotcha moving Actions to Forgejo: uses: actions/checkout@... resolves against DEFAULT_ACTIONS_URL (https://t.co/9LgYaJ7aju by default), not GitHub. Their docs recommend fully qualified URLs, e.g. uses: https://t.co/gEsX0sMDcp
@kelleymuro True. Markup stays opaque until you break spend by day, provider and model. This offline CLI summarises tokens and cost from saved API responses β no live key: https://t.co/gUeT1sXd2X
@skormd Been there. Before the next surprise invoice, save a day of API responses and summarise tokens and cost per model offline β no live key needed. This CLI does that from files: https://t.co/gUeT1sXKSv
@DeRonin_ Solid dual-ship plan. Before you cut Datadog metrics, grab a DogStatsD or Prometheus sample and count unique name+tag combos β high-cardinality tags follow you into VictoriaMetrics too. Offline CLI: https://t.co/BrnQInGvmP
@xanderai 38x is hard to ignore. One check before the next stack: Datadog bills custom metrics per unique name+tag combo β unbounded tags (user_id, request_id) often dominate. Offline CLI on a DogStatsD or Prometheus sample: https://t.co/BrnQInFXxh
@syssignals Agreed on the silent NAT line. For a past window this offline CLI estimates NAT cost from VPC flow logs you already have β no cloud API call: https://t.co/hIk760PtCD
@AsumaCodes Solid recovery. Before the next minute cliff, pull Billing's detailed usage report β it names which workflow and runner SKU burned the free tier. This offline CLI ranks that CSV: https://t.co/9tBacggFys
@gianmauric Before moving everything: open Vercel Usage β Top Paths for Fast Data Transfer and Image Optimization over the same window as the bill. Often one route or image transforms drive spend more than the $20 Pro base.
@pfernan95dev Useful math. One more number before switching runners: Billing's detailed usage report breaks cost by workflow and runner SKU, so you see how much of that macOS line is one workflow vs many. This offline CLI ranks that CSV: https://t.co/9tBacggFys
@anant_hq 3x faster CI with that diff size points to a large generated/restructured workflow + dependency cache + cancel-in-progress concurrency.
Heavy matrix moved off pull_request, fast lint as required check, rest on merge_group.
Matches the numbers better than a full rewrite.