In our repos, Vercel CI is 2-6x faster than our previous solution. We’re excited for you to give it a try with one line of configuration in your turbo.json.
1{2 "tasks": {3 "lint": {4 "tags": ["ci"]5 }6 }7}
This is a big step forward in how CI will look and feel in the agentic future:
It’s faster, by executing only the tasks with changes
It’s easier to use, by automatically handling Remote caching and compute orchestration for you
And it’s easier to configure, by using the task graph that you’ve already written in your build system
We’re looking forward to hearing from you!
Share your experience
What did you find most valuable?
What issues did you encounter? Or gaps you found that we should address?
Native Vercel CI: cross-ref cache misses with identical task hashes
We are seeing pure test tasks rerun across Git refs in Native Vercel CI with Turborepo 2.11.10.
Reproduction:
Complete a successful native CI invocation with cache: true and deterministic inputs.
Push another ref and commit containing exactly the same Git tree.
Confirm that the actual TURBO_HASH values match the successful invocation.
The tasks still log “cache miss, executing” and run the tests again.
Same-ref reuse works. Team Remote Cache is enabled. Real declared outputs and explicit team scoping did not resolve this. A focused test using the installed Turbo binary against a local HTTP cache correctly restores across commits and refuses to cache a changed failing test.
Is native CI intentionally ref-scoped during the beta? Is there a supported way to reuse successful same-repository PR results after merge while preserving every check, provenance and trust boundaries, without adding credentials or upgrading the plan?
If this needs a backend fix, is there a tracking issue or a private channel where Vercel staff can receive the invocation identifiers? We are keeping repository names, commit hashes, invocation IDs, code and private logs out of this public post.
I recently ran into an issue with being overbilled by 20-30 minutes for some simple tasks that only actually took around 10-20 seconds to run - more details here, linking here for visibility:
Hey, Luke, sorry for the trouble. We had a bug that ended up overbilling some folks. We've applied credits to the affected customers so you'll see our mistake covered on your next invoice. Sorry again for the trouble!
Hi ,
This looks different from the cross-ref partitioning discussed above. We get no cache reuse at all, not even across commits on the same branch.
Setup
Vercel CI (beta) on a GitHub repo connected to one project on one team. Team Remote Caching is on (remoteCaching.enabled: true).
Turborepo 2.11.10, Vercel CLI 65.11.0, about 20 root tasks tagged ci, each with "outputs": [].
No remoteCache key in turbo.json, no --no-cache, and no TURBO_* overrides anywhere in the repo.
What we see
Every task log, on main and on PR branches, starts with:
• Running //#ci:
• Remote caching disabled
...
Cached: 0 cached, 1 total
Across our last 50 runs there were 147 cache miss, executing lines and zero hits.
On a single feature branch, eight consecutive successful commits each ran the same voice-agent test task in full (105 to 138 s). That branch never touches the task's inputs. Two of those commits produced an identical task hash (15ecd8fc0880444f), and both still missed. The same happened to 7 other tasks on that branch.
We added a probe that logs, inside each task, which cache variables are present (presence only, passed through with globalPassThroughEnv). All were unset: TURBO_TOKEN, TURBO_TEAM, TURBO_TEAMID, TURBO_API, VERCEL_ARTIFACTS_TOKEN, VERCEL_ARTIFACTS_OWNER, VERCEL_OIDC_TOKEN, and also VERCEL.
Example run IDs
Probe run: n7aHqSlFVjOHrrc4fEAMldcX
Same branch, same hash: ayb6SPWvraGTtj0NKAgctNjL and 48uSRFnpzMoeZn9bdkrruMmA
Main: XqV61oSKcnU5cEozbPfWKBim
Questions
Is "Remote caching disabled" expected inside native CI tasks, for example because the orchestrator checks the cache before it schedules a task? If so, why do identical hashes on the same branch still re-execute?
Should successive commits on one branch reuse each other's results? Should PR runs read results that main produced, as the docs suggest ("run a build on your default branch so that other branches get cache hits")?
Is there anything we need to enable or configure? We'd rather not set our own TURBO_TOKEN, since that would bypass the cache partitioning you described above.
Thanks!
Hi Vercel team,
I received a notice that my Hobby project has used 414% of the included CI CPU Time and will be paused on November 6, 2026.
I set up Vercel CI (beta) on November 1 for one repository, without realising its CPU time was capped at 3 hours on Hobby. In about five days it used 13h 48m. Much of that came from a smoke-test task that held 4 CPUs while it waited for each deployment to finish, and one run took 29 minutes.
Given that everything else was WELL within reason and the deploys were not by any means excessive for a few hours of a working session the usage there doesn't seem to make much sense compared to the rest of the tiers.
Thanks,
Steven
Hi Vercel team,
We noticed a sharp increase in our CI bill on November 12 and would appreciate your help investigating a possible CI cache-detection issue.
In the NephewAI/nephew repository, around 23:16 UTC on November 12, CI began launching hundreds of tasks per run, eventually roughly 1,200, rather than skipping them as cached. Inside the running tasks, Turbo frequently reported cache hits.
One concrete example:
Run: dz6seGH6YeAl7ycEW6dwGV2J
Task: /identity#check:size
Date: November 12, 2026 — all times UTC
23:21:22.045 — Vercel records the task starting
23:21:28.090 — bash -c turbo run /identity#check:size
23:21:29.159 — cache hit, replaying logs 6107d2557ee6442e
23:21:29.160 — Cached: 1 cached, 1 total
23:21:29.160 — Time: 715ms >>> FULL TURBO
23:21:31.349 — Vercel records the task completing as succeeded
That is about 9.3 seconds of recorded task runtime even though Turbo found a cached result. The approvals lint task in the same run showed similar behavior.
In another run, bhBLwYKKac34gDzxvKWy1gTd, the log search returned cache-hit messages for at least 935 launched tasks.
Our repository uses a cache proxy to support sharing artifacts through the team cache. We haven’t established whether that is related, whether scheduler behavior changed, or whether the artifacts became available between scheduling and execution. The logs demonstrate execution followed by cache hits, but we don’t have enough visibility to determine the underlying cause or reconcile the additional charges.
Could you investigate why these tasks were launched instead of skipped as cached, whether there was a relevant platform change around that time, and how much of the additional CI charge came from those executions?