Follow-up to #1864, where the unclecode-litellm fork was adopted. The emergency that
motivated it has passed, and the fork now costs downstream users something it did not
cost back then.
The fork's premise no longer holds
When #1864 was filed, PyPI had quarantined the entire litellm package — every
version was uninstallable, so pinning a fork was the only way to keep pip install crawl4ai working.
That is no longer the case. On PyPI today:
litellm latest is 1.93.0
1.82.7, 1.82.8, 1.82.9 are gone — exactly the compromised releases
1.82.6 and everything from 1.83.0 onward are present and installable
The quarantine was lifted and only the bad releases were removed.
What the fork costs now
crawl4ai 0.9.2 requires unclecode-litellm==1.81.13, which installs a top-level
litellm package — the same import name shipped by the upstream distribution. An
environment that needs both ends up with two different versions written into one
site-packages/litellm/, and which files survive is decided by install order rather
than by either project.
Nothing errors at install time. It surfaces later as:
ImportError: cannot import name '_redact_string' from 'litellm._logging'
The same dependency pair passed on one CI run and failed on another two minutes later,
with nothing changed but resolution order. Because the two distributions have different
names, pip and uv see no conflict to report — they install both, and one silently
overwrites the other's files.
This is not fixable downstream. A [tool.uv] override-dependencies entry works for a
source install, but overrides are not recorded in wheel metadata, so anyone installing
a dependent package from PyPI resolves crawl4ai's requirements directly and hits the
collision again.
Suggested fix
Return to upstream litellm with a floor that excludes the compromised window:
1.83.0 is the first release after the three removed versions, so the floor cannot
resolve to a compromised build even if one were ever restored. This keeps pip install crawl4ai exactly as it is today — full package, LLM support included, no extras to
remember — while removing the shadowing.
If there is a reason to stay on the fork, the collision can also be resolved by having
it install as unclecode_litellm and importing it under an alias:
import unclecode_litellm as litellm
Same outcome for crawl4ai users, and the fork stops occupying a name it doesn't own.
A third option, if the fork is meant as a global drop-in replacement: publish the
crawler half as its own distribution (crawl4ai-core or similar), leaving crawl4ai
as the full LLM-enabled package most people install.
Happy to send a PR for whichever direction you prefer.
Follow-up to #1864, where the
unclecode-litellmfork was adopted. The emergency thatmotivated it has passed, and the fork now costs downstream users something it did not
cost back then.
The fork's premise no longer holds
When #1864 was filed, PyPI had quarantined the entire
litellmpackage — everyversion was uninstallable, so pinning a fork was the only way to keep
pip install crawl4aiworking.That is no longer the case. On PyPI today:
litellmlatest is 1.93.01.82.7,1.82.8,1.82.9are gone — exactly the compromised releases1.82.6and everything from1.83.0onward are present and installableThe quarantine was lifted and only the bad releases were removed.
What the fork costs now
crawl4ai0.9.2 requiresunclecode-litellm==1.81.13, which installs a top-levellitellmpackage — the same import name shipped by the upstream distribution. Anenvironment that needs both ends up with two different versions written into one
site-packages/litellm/, and which files survive is decided by install order ratherthan by either project.
Nothing errors at install time. It surfaces later as:
The same dependency pair passed on one CI run and failed on another two minutes later,
with nothing changed but resolution order. Because the two distributions have different
names, pip and uv see no conflict to report — they install both, and one silently
overwrites the other's files.
This is not fixable downstream. A
[tool.uv] override-dependenciesentry works for asource install, but overrides are not recorded in wheel metadata, so anyone installing
a dependent package from PyPI resolves crawl4ai's requirements directly and hits the
collision again.
Suggested fix
Return to upstream
litellmwith a floor that excludes the compromised window:litellm>=1.83.01.83.0is the first release after the three removed versions, so the floor cannotresolve to a compromised build even if one were ever restored. This keeps
pip install crawl4aiexactly as it is today — full package, LLM support included, no extras toremember — while removing the shadowing.
If there is a reason to stay on the fork, the collision can also be resolved by having
it install as
unclecode_litellmand importing it under an alias:Same outcome for crawl4ai users, and the fork stops occupying a name it doesn't own.
A third option, if the fork is meant as a global drop-in replacement: publish the
crawler half as its own distribution (
crawl4ai-coreor similar), leavingcrawl4aias the full LLM-enabled package most people install.
Happy to send a PR for whichever direction you prefer.