Stephen Way
@stephenway.bsky.social
Feedback loops should be instant, I think
Pinned
99% of your pipeline didn't change. Your CI should know that.
AI-generated changes should be cheap experiments, not pipeline landmines.
1) Right now, every AI-suggested refactor feels risky.
Not because the code is necessarily bad.
1) Right now, every AI-suggested refactor feels risky.
Not because the code is necessarily bad.
November 30, 2025 at 5:02 PM
AI-generated changes should be cheap experiments, not pipeline landmines.
1) Right now, every AI-suggested refactor feels risky.
Not because the code is necessarily bad.
1) Right now, every AI-suggested refactor feels risky.
Not because the code is necessarily bad.
Reproducible incidents: why epochs beat ad-hoc logs every time.
1) Most CI and infra incidents are technically “reproducible” only in theory.
1) Most CI and infra incidents are technically “reproducible” only in theory.
November 29, 2025 at 6:00 PM
Reproducible incidents: why epochs beat ad-hoc logs every time.
1) Most CI and infra incidents are technically “reproducible” only in theory.
1) Most CI and infra incidents are technically “reproducible” only in theory.
I’ve hit the founder stage where buying domains feels like collecting Beanie Babies.
Except Beanie Babies didn’t charge $79/year for .sh
Except Beanie Babies didn’t charge $79/year for .sh
November 29, 2025 at 5:03 PM
I’ve hit the founder stage where buying domains feels like collecting Beanie Babies.
Except Beanie Babies didn’t charge $79/year for .sh
Except Beanie Babies didn’t charge $79/year for .sh
Why restore_keys feel clever but still miss most reuse opportunities.
1) Restore_keys are a clever hack on top of a weak model.
You’re telling CI:
“If you can’t find an exact cache key, walk up this prefix ladder.”
It’s better than nothing.
1) Restore_keys are a clever hack on top of a weak model.
You’re telling CI:
“If you can’t find an exact cache key, walk up this prefix ladder.”
It’s better than nothing.
November 28, 2025 at 8:00 PM
Why restore_keys feel clever but still miss most reuse opportunities.
1) Restore_keys are a clever hack on top of a weak model.
You’re telling CI:
“If you can’t find an exact cache key, walk up this prefix ladder.”
It’s better than nothing.
1) Restore_keys are a clever hack on top of a weak model.
You’re telling CI:
“If you can’t find an exact cache key, walk up this prefix ladder.”
It’s better than nothing.
What would you do with a provable cache of every past computation? “Already computed” is a powerful primitive.
November 28, 2025 at 1:00 AM
What would you do with a provable cache of every past computation? “Already computed” is a powerful primitive.
AI as a stress‑test for your infra, not just your product. If your CI melts, that’s a signal about your architecture.
November 27, 2025 at 8:00 PM
AI as a stress‑test for your infra, not just your product. If your CI melts, that’s a signal about your architecture.
The hidden tax of waiting on CI when you’re supposed to be shipping.
November 27, 2025 at 5:00 PM
The hidden tax of waiting on CI when you’re supposed to be shipping.
The cost of not being able to rewind: incidents that linger for days because nobody can reconstruct the last known‑good world-state.
November 27, 2025 at 1:00 AM
The cost of not being able to rewind: incidents that linger for days because nobody can reconstruct the last known‑good world-state.
Why “more runners” stopped fixing your CI bottlenecks. Parallelizing guesses doesn’t turn them into guarantees.
November 26, 2025 at 8:00 PM
Why “more runners” stopped fixing your CI bottlenecks. Parallelizing guesses doesn’t turn them into guarantees.
AI doubled your code output, but your CI still acts like a monolith.
November 26, 2025 at 5:00 PM
AI doubled your code output, but your CI still acts like a monolith.
A day in the life of an epoch: from commit to deployment. One stable ID should trace the exact world-state your graph saw from first test to final rollout.
November 26, 2025 at 1:02 AM
A day in the life of an epoch: from commit to deployment. One stable ID should trace the exact world-state your graph saw from first test to final rollout.
Stop hitting “rerun job” and start rewinding to the exact state that passed.
1) “Rerun job” is a superstition button.
You’re not fixing anything.
1) “Rerun job” is a superstition button.
You’re not fixing anything.
November 25, 2025 at 5:03 PM
Stop hitting “rerun job” and start rewinding to the exact state that passed.
1) “Rerun job” is a superstition button.
You’re not fixing anything.
1) “Rerun job” is a superstition button.
You’re not fixing anything.
Your calendar as an infra smell detector. If whole afternoons disappear into “waiting on CI”, you don’t have a pipeline problem, you have a platform gap.
November 25, 2025 at 1:02 AM
Your calendar as an infra smell detector. If whole afternoons disappear into “waiting on CI”, you don’t have a pipeline problem, you have a platform gap.
If your CI had a time slider, what would you debug first? Imagine scrubbing across commits instead of grepping logs.
November 24, 2025 at 7:56 PM
If your CI had a time slider, what would you debug first? Imagine scrubbing across commits instead of grepping logs.
CI is slow because it guesses what changed instead of knowing.
1) Most CI pipelines don’t actually know which parts of the system a change touches.
They approximate it.
1) Most CI pipelines don’t actually know which parts of the system a change touches.
They approximate it.
November 24, 2025 at 5:04 PM
CI is slow because it guesses what changed instead of knowing.
1) Most CI pipelines don’t actually know which parts of the system a change touches.
They approximate it.
1) Most CI pipelines don’t actually know which parts of the system a change touches.
They approximate it.
Designing cache keys that carry proof, not hope. A good key says “I have computed this exact path before”, not “this folder looks familiar”.
November 24, 2025 at 3:04 AM
Designing cache keys that carry proof, not hope. A good key says “I have computed this exact path before”, not “this folder looks familiar”.
How many hours per week does your team lose to waiting on CI? If you don’t know the number, it’s probably bad.
November 23, 2025 at 8:56 PM
How many hours per week does your team lose to waiting on CI? If you don’t know the number, it’s probably bad.
99% of your pipeline didn't change. Your CI should know that.
November 21, 2025 at 5:02 PM
99% of your pipeline didn't change. Your CI should know that.
If the inputs haven’t changed, nothing else should have to either. Pipelines could remember past work. Kind of like our brains on good days.
November 19, 2025 at 4:59 PM
If the inputs haven’t changed, nothing else should have to either. Pipelines could remember past work. Kind of like our brains on good days.
What if secrets weren’t .env files or a Vault maze, but a project-scoped encrypted notepad that syncs across dev + CI with versioning and diffing?
Would you use that?
Would you use that?
November 18, 2025 at 7:58 PM
What if secrets weren’t .env files or a Vault maze, but a project-scoped encrypted notepad that syncs across dev + CI with versioning and diffing?
Would you use that?
Would you use that?
Looks like another VSCode AI clone
November 18, 2025 at 4:57 PM
Looks like another VSCode AI clone
You tweak two files. CI rebuilds everything. Why? This is why folks start dreaming about incremental systems. They just make sense.
November 17, 2025 at 7:59 PM
You tweak two files. CI rebuilds everything. Why? This is why folks start dreaming about incremental systems. They just make sense.
Ever feel like your CI is just doing busywork? Same commands, same files, over and over. There's a better way: cache what's already done.
www.dan-manges.com/blog/solvin...
www.dan-manges.com/blog/solvin...
November 14, 2025 at 7:58 PM
Ever feel like your CI is just doing busywork? Same commands, same files, over and over. There's a better way: cache what's already done.
www.dan-manges.com/blog/solvin...
www.dan-manges.com/blog/solvin...
At DoorDash, devs were losing 500+ hours a month to CI. That’s wild. If you've ever waited on a build only to learn nothing changed, you're not alone.
careersatdoordash.com/blog/how-do...
careersatdoordash.com/blog/how-do...
November 13, 2025 at 7:58 PM
At DoorDash, devs were losing 500+ hours a month to CI. That’s wild. If you've ever waited on a build only to learn nothing changed, you're not alone.
careersatdoordash.com/blog/how-do...
careersatdoordash.com/blog/how-do...