[AI] Curate the public API instead of blanket-reexporting dependencies (#1286)

* Curate the public API instead of blanket-reexporting dependencies

Every package in the monorepo `@reexport`ed its dependencies wholesale, so
`using Optimization` put 263 names in scope and `using OptimizationOptimJL` put
298 — none of which either package owned. 34 came from SciMLOperators
(`AddVector`, `BlockDiagonalOperator`, `update_coefficients!`), and the surface
included `DynamicalSDEFunction`, `DDEProblem` and `EigenvalueProblem`. Of the
263, the entire repo referenced 65.

OptimizationBase now `using`s SciMLBase/ADTypes/SciMLLogging and states its
public API in explicit `export` lists: its own types plus the problem/solution
API and the nine `Auto*` types a user needs. Optimization re-exports that
curated set and nothing else. The solver packages stop passing OptimizationBase
through and expose only their own solvers, so `using OptimizationOptimJL` now
means `LBFGS`, `BFGS`, `IPNewton` and friends. Users load `Optimization`
alongside the solver package, which is what the docs already showed everywhere
but one page.

    Optimization         263 public, 0 owned  ->  39
    OptimizationBase     262 public, 6 owned  ->  38
    OptimizationOptimJL  298 public, 0 owned  ->  36

Dropping the reexports exposed five defects they had been masking:

- OptimizationMOI and OptimizationIpopt use `ADTypes.AutoSymbolics` and
  `ADTypes.AutoSparse` qualified without declaring ADTypes at all; they
  compiled only because the reexport chain happened to bind the name.
- OptimizationCMAEvolutionStrategy imported the SciMLBase *module* from
  OptimizationBase, and OptimizationLBFGSB took `ReturnCode` from there rather
  than from SciMLBase, which owns it.
- OptimizationPyCMA marked `all_qualified_accesses_via_owners` as broken; it
  passes once accesses go to their owners, so the marker is removed.

17 packages used `@SciMLMessage` without depending on SciMLLogging and now
declare it. `run_qa`'s reexport audit is satisfied without disabling it: the
nine glue packages allow exactly their backend module, the two umbrella
packages list their curated re-surface, and the remaining 18 need no
allow-list because they expose nothing they do not own.

Breaking: Optimization and OptimizationBase 5 -> 6, sublibraries by their
SemVer rule, with intra-monorepo compat bounds updated to match.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014Uj4Pu1LKHYqLSTLfPYDPo

* docs: use OptimizationBase in the PyCMA example

Matches getting_started.md and the LBFGSB/PRIMA/Sophia pages. Addresses
review feedback on the pycma.md hunk.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

* docs: use OptimizationBase as the entry package throughout

The docs were split between `using Optimization` and `using OptimizationBase`,
with getting_started.md, lbfgsb.md, prima.md and sophia.md already on the
latter. Settle on OptimizationBase everywhere, per review feedback on the
pycma.md hunk.

names(Optimization) is names(OptimizationBase) plus the symbol :Optimization,
and solve/init/solve!/reinit!/AutoForwardDiff/OptimizationProblem resolve to
identical bindings in both, so this is behaviour-preserving.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

* docs: stop relying on the SciMLBase binding leaked by @reexport

`@reexport using SciMLBase` exported names(SciMLBase), which always contains
the symbol SciMLBase itself, so the module binding landed in every namespace
that did `using Optimization`. Curating the exports removes it, and the docs
were depending on it in two places:

  - 21 `SciMLBase.OptimizationProblem`/`OptimizationFunction` calls, 18 of them
    in executed @example blocks, which would throw UndefVarError and fail the
    build (@example errors are not covered by the existing warnonly list).
    OptimizationBase exports both names, so drop the qualification.
  - 5 @docs entries under API/ naming SciMLBase types. No page sets
    CurrentModule, so Documenter resolves them in Main; make.jl now binds
    SciMLBase explicitly.

rosenbrock.md, linearandinteger.md and ode.md `using SciMLBase` for real and
are left as they are.

Verified on Julia 1.11: master binds SciMLBase and ADTypes for all four
entry-package x solver combinations tested, this branch binds neither.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

* docs: say which of Optimization/OptimizationBase to install

Every example now loads OptimizationBase, so README and the index install
instructions pointed at a package the docs never use. Both now lead with
OptimizationBase, and index.md gains a short section on the relationship.

`Optimization` re-exports OptimizationBase and adds nothing else — it pulls in
Logging/ConsoleProgressMonitor/TerminalLoggers/LoggingExtras but never
references them — so the new text says the two are interchangeable rather than
inventing a distinction between them.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

* OptimizationBase: drop the now-stale Reexport dependency

Removing every `@reexport` from OptimizationBase left Reexport in [deps] and
[compat] with no user, which fails Aqua's stale-dependency check. The
reexport check itself already passed; this is the leftover.

QA on OptimizationBase: 19 pass / 1 fail / 1 broken before, 20 pass / 1 broken
after. The remaining broken is `ei_broken = (:no_implicit_imports,)`, which is
already on master.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

* OptimizationLBFGSB: drop the now-stale OptimizationFunction import

`using SciMLBase: OptimizationStats, OptimizationFunction` predates this
branch, but on master `@reexport using OptimizationBase` re-exported every
name in names(OptimizationBase), OptimizationFunction among them — and
exporting a name counts as using it, so the import was not stale. Removing
the reexport orphans it, and ExplicitImports' no_stale_explicit_imports
errors. The name is used nowhere else in the package.

QA on OptimizationLBFGSB, Julia 1.11:
  master: 18 pass, 2 fail (JET + reexports), 1 broken
  before: 19 pass, 1 fail (JET), 1 error (stale import), 1 broken
  after:  19 pass, 1 fail (JET), 1 broken

The remaining JET failure is byte-identical to master's (same 9 toplevel
errors, from the LBFGSB struct colliding with the LBFGSB package binding)
and is already tracked as a follow-up.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A9dFnCQK2GVSNTCJP6ZMG2

---------

Co-authored-by: ChrisRackauckas-Claude <accounts@chrisrackauckas.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: 3 people

Project.toml

@@ -1,6 +1,6 @@
 name = "Optimization"
 uuid = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
-version = "5.6.6"
+version = "6.0.0"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
@@ -47,9 +47,9 @@ MLUtils = "0.4"
 ModelingToolkit = "11"
 Optim = "2"
 Optimisers = ">= 0.2.5"
-OptimizationBase = "5"
-OptimizationOptimJL = "0.4.10"
-OptimizationOptimisers = "0.3.16"
+OptimizationBase = "6"
+OptimizationOptimJL = "0.5"
+OptimizationOptimisers = "0.4"
 OrdinaryDiffEqTsit5 = "1, 2"
 Pkg = "1"
 Printf = "1.10"

README.md

@@ -22,11 +22,11 @@ unified interface.
 ## Installation
 
 Assuming that you already have Julia correctly installed, it suffices to import
-Optimization.jl in the standard way:
+OptimizationBase in the standard way:
 
 ```julia
 using Pkg
-Pkg.add("Optimization")
+Pkg.add("OptimizationBase")
 ```
 
 The packages relevant to the core functionality of Optimization.jl will be imported
@@ -71,7 +71,7 @@ the documentation, which contains the unreleased features.
 ## Examples
 
 ```julia
-using Optimization
+using OptimizationBase
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 x0 = zeros(2)
 p = [1.0, 100.0]

docs/Project.toml

@@ -165,40 +165,40 @@ ModelingToolkit = "10.23, 11"
 NLPModels = "0.21, 0.22"
 NLPModelsTest = "0.10"
 NLopt = "0.6, 1"
-Optimization = "5.5"
-OptimizationAuglag = "1.3, 2"
-OptimizationBBO = "0.4.6"
-OptimizationBase = "5"
-OptimizationCMAEvolutionStrategy = "0.3.6"
-OptimizationEvolutionary = "0.4.7"
-OptimizationGCMAES = "0.3.5"
-OptimizationIpopt = "1.1"
-OptimizationLBFGSB = "1.4"
-OptimizationMOI = "1.2"
-OptimizationMadNLP = "2, 1"
-OptimizationManopt = "1.2"
-OptimizationMetaheuristics = "0.3.5"
-OptimizationMultistartOptimization = "0.3.4"
-OptimizationNLPModels = "1.2"
-OptimizationNLopt = "0.3.9"
-OptimizationNOMAD = "0.3.5"
-OptimizationODE = "0.1.4"
-OptimizationOptimJL = "0.4.10"
-OptimizationOptimisers = "0.3.16"
-OptimizationPRIMA = "0.3.5"
-OptimizationPolyalgorithms = "0.3.5"
-OptimizationPyCMA = "1.3"
-OptimizationQuadDIRECT = "0.3.4"
-OptimizationSciPy = "0.4.6"
-OptimizationSophia = "1.3"
-OptimizationSpeedMapping = "0.2.3"
+Optimization = "6"
+OptimizationAuglag = "3"
+OptimizationBBO = "0.5"
+OptimizationBase = "6"
+OptimizationCMAEvolutionStrategy = "0.4"
+OptimizationEvolutionary = "0.5"
+OptimizationGCMAES = "0.4"
+OptimizationIpopt = "2"
+OptimizationLBFGSB = "2"
+OptimizationMOI = "2"
+OptimizationMadNLP = "3"
+OptimizationManopt = "2"
+OptimizationMetaheuristics = "0.4"
+OptimizationMultistartOptimization = "0.4"
+OptimizationNLPModels = "2"
+OptimizationNLopt = "0.4"
+OptimizationNOMAD = "0.4"
+OptimizationODE = "0.2"
+OptimizationOptimJL = "0.5"
+OptimizationOptimisers = "0.4"
+OptimizationPRIMA = "0.4"
+OptimizationPolyalgorithms = "0.4"
+OptimizationPyCMA = "2"
+OptimizationQuadDIRECT = "0.4"
+OptimizationSciPy = "0.5"
+OptimizationSophia = "2"
+OptimizationSpeedMapping = "0.3"
 OrdinaryDiffEq = "6, 7"
 Plots = "1"
 Random = "1"
 ReverseDiff = ">= 1.9.0"
 SciMLBase = "2.122.1, 3"
 SciMLSensitivity = "7"
-SimpleOptimization = "1"
+SimpleOptimization = "2"
 Symbolics = "7"
 Tracker = ">= 0.2"
 Zygote = ">= 0.5"

docs/make.jl

@@ -1,4 +1,7 @@
 using Documenter, Optimization
+# The `@docs` entries under API/ name SciMLBase types directly; nothing re-exports
+# the module binding any more, so bind it here rather than relying on that leak.
+using SciMLBase
 using OptimizationAuglag, OptimizationBBO, OptimizationBase
 using OptimizationCMAEvolutionStrategy, OptimizationGCMAES, OptimizationIpopt
 using OptimizationLBFGSB, OptimizationMadNLP, OptimizationManopt

docs/src/index.md

@@ -36,6 +36,20 @@ Optimization.jl is simply a bundle/interface over many of these dependencies.
 It may add some optional higher level behavior in the future but at this time
 the top level package does not add any extra behavior.
 
+### `Optimization` or `OptimizationBase`?
+
+`OptimizationBase` defines the interface — `OptimizationProblem`,
+`OptimizationFunction`, `solve` — and is what every solver package depends on,
+so it is what the examples throughout this documentation load:
+
+```julia
+Pkg.add("OptimizationBase")
+```
+
+`Optimization` re-exports that same interface and nothing else, so the two are
+interchangeable for everything shown here. Install whichever you prefer; you
+need only one of them, alongside a solver package.
+
 ## Contributing
 
   - Please refer to the

docs/src/optimization_packages/auglag.md

@@ -20,7 +20,7 @@ OptimizationAuglag.AugLag
 ## Example
 
 ```julia
-using Optimization, OptimizationAuglag, OptimizationOptimJL, ADTypes
+using OptimizationBase, OptimizationAuglag, OptimizationOptimJL, ADTypes
 
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 cons(res, x, p) = (res .= [x[1]^2 + x[2]^2])

docs/src/optimization_packages/blackboxoptim.md

@@ -77,12 +77,12 @@ OptimizationBBO.BBO_borg_moea
 The Rosenbrock function can be optimized using the `BBO_adaptive_de_rand_1_bin_radiuslimited()` as follows:
 
 ```@example BBO
-using Optimization, OptimizationBBO
+using OptimizationBase, OptimizationBBO
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 x0 = zeros(2)
 p = [1.0, 100.0]
 f = OptimizationFunction(rosenbrock)
-prob = SciMLBase.OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
+prob = OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
 sol = solve(prob, BBO_adaptive_de_rand_1_bin_radiuslimited(), maxiters = 100000,
     maxtime = 1000.0)
 ```

docs/src/optimization_packages/cmaevolutionstrategy.md

@@ -29,11 +29,11 @@ OptimizationCMAEvolutionStrategy.CMAEvolutionStrategyOpt
 The Rosenbrock function can be optimized using the `CMAEvolutionStrategyOpt()` as follows:
 
 ```@example CMAEvolutionStrategy
-using Optimization, OptimizationCMAEvolutionStrategy
+using OptimizationBase, OptimizationCMAEvolutionStrategy
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 x0 = zeros(2)
 p = [1.0, 100.0]
 f = OptimizationFunction(rosenbrock)
-prob = SciMLBase.OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
+prob = OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
 sol = solve(prob, CMAEvolutionStrategyOpt())
 ```

docs/src/optimization_packages/evolutionary.md

@@ -33,11 +33,11 @@ Algorithm-specific options are defined as `kwargs`. See the respective documenta
 The Rosenbrock function can be optimized using the `Evolutionary.CMAES()` as follows:
 
 ```@example Evolutionary
-using Optimization, OptimizationEvolutionary
+using OptimizationBase, OptimizationEvolutionary
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 x0 = zeros(2)
 p = [1.0, 100.0]
 f = OptimizationFunction(rosenbrock)
-prob = SciMLBase.OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
+prob = OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
 sol = solve(prob, Evolutionary.CMAES(μ = 40, λ = 100))
 ```

docs/src/optimization_packages/gcmaes.md

@@ -29,12 +29,12 @@ OptimizationGCMAES.GCMAESOpt
 The Rosenbrock function can be optimized using the `GCMAESOpt()` without utilizing the gradient information as follows:
 
 ```@example GCMAES
-using Optimization, OptimizationGCMAES
+using OptimizationBase, OptimizationGCMAES
 rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
 x0 = zeros(2)
 p = [1.0, 100.0]
 f = OptimizationFunction(rosenbrock)
-prob = SciMLBase.OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
+prob = OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
 sol = solve(prob, GCMAESOpt())
 ```
 
@@ -43,6 +43,6 @@ We can also utilize the gradient information of the optimization problem to aid
 ```@example GCMAES
 using ADTypes, ForwardDiff
 f = OptimizationFunction(rosenbrock, ADTypes.AutoForwardDiff())
-prob = SciMLBase.OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
+prob = OptimizationProblem(f, x0, p, lb = [-1.0, -1.0], ub = [1.0, 1.0])
 sol = solve(prob, GCMAESOpt())
 ```