Agents and control
The shared memory of fixes
When a deploy fails, CLI shows what fixed the same kind of failure in other projects (anonymized), and agents get it in the failed deploy itself.
Every time a failed deploy gets fixed on CLI, CLI remembers what fixed it. The next time anyone hits the same failure, that fix is suggested first, with how often it worked.
How to use it: open a failed deploy and look for Fixes that worked for others, for example "Set DATABASE_URL, worked 41 times". Your agent sees the same list.
Only the kind of failure and the kind of fix are shared, never your code, values or names. You can opt out in project settings.
Every failed deploy on CLI teaches it something. When a deploy fails, CLI takes a fingerprint of the failure. When the next deploy of that service works, CLI records what changed in between. The next time anyone hits the same failure, that fix is suggested, with how often it worked.
What you see
A failed deploy's page shows Fixes that worked for others, for example "Set DATABASE_URL, worked 41 times in 37 projects". Agents get the same list in the failed deploy itself (knownFixes in GET /v1/deploys/:id), so they can try the likeliest fix first. The full list:
GET /v1/deploys/:id/fixes
cli fixes dep_123Fixes your own project used before are marked as yours and shown straight away.
What's kept, and what isn't
- The failure: its error code and the line of the logs that says what went wrong, with names, paths, numbers, strings and addresses stripped out.
- The fix: variables added, removed or changed, by name only, never values; which settings changed (the new value only for port, health check path, CPU, memory and replicas); whether the code changed.
- Not kept: values, logs, project, service or repository names, or who did it.
A fix is only shown to other projects once it worked in at least two different ones. Changes that were tried and didn't fix it are counted too, so a fix that rarely works sinks.
Failures that aren't about the app (out of credit, rate limits, CLI's own problems) aren't recorded.
Opting out
A project admin can stop adding the project's failures and fixes:
PUT /v1/projects/:id/share-fixes {"share": false}It still sees suggestions from others. Operators can turn the whole thing off with CLI_FIX_MEMORY=off.