can-they-merge (ctm)

Can these Python dependency sets merge into one environment? And if not, into how few?

Researchers constantly stitch together frameworks, benchmarks, and baselines — each with its own requirements.txt, pyproject.toml, or conda environment.yaml. ctm answers two questions using real resolvers for ground truth:

1. ctm merge — do all these inputs co-resolve into a single environment?
2. ctm partition — when they don't, what is the minimum number of environments needed to cover a set of components/entrypoints?

ctm is a small, pure-Python tool (3.9+) that brings its own resolvers: it fetches pinned uv (PyPI) and micromamba (conda) binaries on first use, so you don't have to install or manage them.

Install

pip install can-they-merge        # or: uv tool install can-they-merge
ctm setup                         # fetch the bundled uv + micromamba resolvers

ctm setup is optional — the resolvers are also fetched automatically on first use. For air-gapped clusters, run ctm setup on a connected machine and copy the cache dir, or pre-stage the binaries and install from them:

ctm setup --from-dir /path/to/prestaged/bin
# or point at an existing location:
export CTM_BIN_DIR=/shared/tools/ctm-bin

Usage

Can they merge?

ctm merge requirements.txt pyproject.toml environment.yaml

NOT MERGEABLE ✗  (target: pypi)
  - numpy: <1.20 & >=1.24
  resolver said:
    × No solution found... Because you require numpy<1.20 and numpy>=1.24...

• Mix formats freely (auto-detected by extension).
• --target pypi (default, uses uv) or --target conda (uses micromamba).
• --json for a machine-readable report. Exit codes: 0 mergeable, 1 not mergeable, 3 resolver unavailable.

Minimum number of environments

Pass one dependency file per component:

ctm partition reqs-train.txt reqs-deploy.txt reqs-util.txt --exact

2 environment(s) needed [minimum]:
  env 1: reqs-deploy, reqs-util
  env 2: reqs-train

• --exact proves the count is the true minimum (branch-and-bound; ≤8 components). Without it, a sound greedy partition is returned and labeled as not-proven-minimal.

How it works

• Ground truth from real resolvers. ctm merge resolves the union with uv/micromamba; a pure-Python specifier pre-check supplies readable conflict explanations and catches obvious cases without a resolver call.
• Partitioning builds a conflict graph (no resolver calls), then a greedy-with-validation pass (every emitted environment is resolver-checked, so it's sound), with an optional exact branch-and-bound that memoizes every resolution to prove minimality.
• Over-approximate reachability (library: can_they_merge.reachability) statically maps entrypoints to the distributions they need, erring toward including a package rather than dropping it.

Configuration

Development

uv sync
uv run pytest -q                 # unit tests, no network
uv run pytest -q -m integration  # drives real uv against verified fixtures
uv run ruff check src tests

Design spec: docs/superpowers/specs/2026-06-11-can-it-merge-design.md.

License

MIT © Daniel Hou