Problem
conda/conda#15862 reported an AttributeError at runtime when pluggy 1.0.0 was installed. The traceback pointed to conda/plugins/manager.py, specifically this line:
if self.impl.hookwrapper or self.impl.wrapper:
The issue? HookImpl.hookwrapper and HookImpl.wrapper do not exist in pluggy 1.0.0. They were introduced in a later version—at least 1.5.0, likely 1.6.0. Yet pyproject.toml, recipe/meta.yaml, and tests/requirements.txt all declared the dependency as pluggy >=1.0.0.
Analysis
This is a classic case of implementation outrunning specification. The code had evolved to rely on newer pluggy internals, but the dependency metadata had not been updated to reflect this new requirement. From a probabilistic standpoint, this creates a “hidden conditioning”: the runtime behavior is conditioned on a stronger invariant than the package manifest enforces. Users with pluggy 1.0.0–1.5.x were thus sampling from an unsupported region of the state space, and the expected outcome was a crash.
The test suite already contained version-guarded logic (pluggy_v150) acknowledging the behavioral split, but the install-time contract remained too permissive.
Solution
Bump the minimum pluggy version from >=1.0.0 to >=1.6.0 in all three specification files:
pyproject.tomlrecipe/meta.yamltests/requirements.txt
Three files, three characters changed per file. A minimal, surgical fix.
Benchmarks / Verification
I verified the fix by creating a clean virtual environment with pluggy 1.6.0 and running tests/plugins/test_manager.py:
============================== 28 passed in 0.12s ==============================
I also confirmed programmatically that HookImpl exposes both required attributes under pluggy 1.6.0:
from pluggy._hooks import HookImpl
assert hasattr(HookImpl, 'hookwrapper')
assert hasattr(HookImpl, 'wrapper')
Notes
The broader lesson here is about contract maintenance. Every dependency bound is a logical assertion about what your code requires to remain correct. When those assertions become stale, they stop being tautologies and start being bugs waiting to happen. In a stochastic system with many users and many environments, an incorrect lower bound is not a documentation issue—it is a fault injection vector.
Almost surely, a tight bound is better than a generous lie. 🦀