[nicoloangileri]

Fixes #20.

This applies the change @Transurgeon suggested in #20, written so that it works with both the current cvxpy release and cvxpy master. That way it does not have to wait for cvxpy 1.10.

How we noticed

I was updating a cvxpy PR (cvxpy/cvxpy#3490), and its test_gcsopt check failed in tests/test_programs.py::TestConvexProgram::test_to_conic:

AssertionError: np.float64(-6.928203228560545) != -inf within 4 places

The PR doesn't touch anything that gcsopt uses, so I looked at the same check on cvxpy master. It has been red on every master commit since cvxpy/cvxpy#3525. The last green run was on e7e704c6 (2026-09-11) and the first red one on 773163eb (#3525, 2026-09-13). Since cvxpy's gcsopt.yml workflow runs this suite on every cvxpy PR, every cvxpy PR currently shows a red check. That is what led me to #20.

Change

Since the PR #22 changes, to_conic builds the chain with solver=cp.SCS. In cvxpy master that chain ends with ..., ConeFormat, SCS, so chain.reductions[:-1] keeps ConeFormat. The program then comes back in SCS's cone layout instead of the stuffed layout that _constrain_in_cone decodes.

The fix drops the solver-facing steps by type. The steps are Solver and, when it can be imported, ConeFormat. On cvxpy releases without ConeFormat, this is the same as removing the last reduction, as before.

Tests

pytest tests:

cvxpy	main	this PR
1.9.3 (PyPI)	21 passed	21 passed
master (1.10.0.dev, includes #3525)	2 failed (test_to_conic, test_to_conic_multiple_soc), 19 passed	21 passed

[nicoloangileri]

Hi @TobiaMarcucci, a gentle ping: cvxpy's test_gcsopt job still fails on every cvxpy PR because of the new ConeFormat step, and this PR keeps gcsopt working with both old and new cvxpy. Happy to adjust if you prefer a different approach. Thanks!

[TobiaMarcucci]

Sorry for the slow response. The fix looks good to me. Thanks very much!