Skip to content

deps: raise the optional pymdp extra to inferactively-pymdp 1.0.5 - #28

Merged
docxology merged 1 commit into
mainfrom
chore/pymdp-1.0.5
Nov 12, 2026
Merged

docxology merged 1 commit into
mainfrom
chore/pymdp-1.0.5

Conversation

@docxology

Copy link
Copy Markdown
Collaborator

What changed

inferactively-pymdp 1.0.5 is the current PyPI release. Two live declarations named an older version:

File Line Before After
pyproject.toml 62 ([project.optional-dependencies].pymdp) inferactively-pymdp>=1.0.1 inferactively-pymdp>=1.0.5
uv.lock metadata specifier >=1.0.1 >=1.0.5
uv.lock inferactively-pymdp package block 1.0.4 + 1.0.4 hashes 1.0.5 + 1.0.5 hashes

The lock was regenerated with the repository tool (uv lock, uv 0.13.7 as [tool.uv].required-version), never hand-edited; a second uv lock is zero-diff. The locked hashes match PyPI: wheel sha256:5e423a331894c7d4dbda1d935d7b592556dfea229d99890aec02d5216e9f1d84, sdist sha256:c58461217fe16d9bf4c4df8e49411d36fff8ae067f4b42225f5c69c4832963a7.

Historical mentions are untouched: no CHANGELOG, dated evidence JSON, retirement inventory or frozen protocol/manuscript metadata in this repository names a pymdp version. The remaining pymdp hits are prose (optional-extra docs, glossary, references.bib) and the optional import guard in blockference/agent.py; none of them carries a version.

Gate (local front door: scripts/release_check.py)

uv sync --locked --extra dev, then:

  • fast arms (uv run python scripts/release_check.py): lint, types, test (83 passed, 1 skipped — the pinned-GNN skip), docs → passed
  • release gate (GNN_RUNTIME_PYTHON=<pinned GNN checkout>/.venv/bin/python uv run python scripts/release_check.py --arm all --timings ...): lint, types, test-full (97 passed, coverage 90.96% ≥ 90), docs, manuscript, gnn, smoke, package → passed, "release gate passed"

A --arm all run with only GNN_ROOT set goes red in manuscript (ModuleNotFoundError: No module named 'gnn' in scripts/gnn_importer_report.py), which is environmental and pre-existing: the identical command fails the same way at base b621126 in a detached worktree. manuscript and gnn arms read GNN_RUNTIME_PYTHON; setting it makes both green.

Behaviour of 1.0.5

Wheel contents 1.0.4 → 1.0.5: 39 Python modules, no module added or removed; changed pymdp/agent.py, pymdp/control.py, pymdp/envs/graph_worlds.py, pymdp/learning.py, pymdp/legacy/maths.py, pymdp/planning/mcts.py. python_requires stays >=3.10; the jax/jaxlib upper bound (<0.11.1) is removed and a modelfit extra dependency is added. pymdp.agent.Agent is retained, so the blockference.agent.BlockferenceAgent subclass still binds (uv sync --locked --extra dev --extra pymdp → tests/test_agent.py 2 passed, MRO parent pymdp.agent.Agent). The Agent change fixes flattened action-dependency handling and stops silently overriding sampling_mode; this package never constructs an Agent with B_action_dependencies, so no assertion had to change.

Hosted Actions are billing-locked org-wide; the local gate above is the evidence.

The optional `pymdp` extra named a >=1.0.1 floor and uv.lock resolved
inferactively-pymdp 1.0.4. Raise the floor to >=1.0.5 and regenerate the
lock with uv 0.13.7; the locked wheel/sdist hashes match PyPI for 1.0.5.

The adapter only subclasses pymdp.agent.Agent, which 1.0.5 keeps, so the
adapter surface is unchanged; 1.0.5 relaxes the jax/jaxlib upper bound and
fixes flattened action-dependency handling in Agent, neither of which this
package exercises.
@docxology
docxology merged commit 7f8469d into main Nov 12, 2026
0 of 4 checks passed
docxology added a commit that referenced this pull request Nov 12, 2026
…29)

The optional pymdp extra and its lock are recorded as done: the floor moved
from >=1.0.1 to >=1.0.5 and uv.lock resolves 1.0.5 (7f8469d, #28). The GNN
cutover item now cites its merge (b621126, #27), and the verification notes
record this pass: the docs arm and the gnn arm both passed at 7f8469d, with
gnn.importers reporting v5.1.0 @ 75c834bd85 and verdict ok.

One defect found while checking the tree is recorded as open: ignored.csv is a
cadCAD record dump committed by 574a5c1 and nothing references it; .gitignore
covers result.csv, results/ and output/ but not that name.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant