[singhharsh1708]

DelayDiffEq only resized the Jacobian config of a RosenbrockMutableCache and fell through to a no-op for anything else, so a Rosenbrock member of a CompositeCache kept its old config. A DDEIntegrator method for CompositeCache now forwards to each member, as OrdinaryDiffEqCore does for ODEIntegrator.

Resizing a composite whose members use different AD backends then threw Cannot convert ... JacobianPrep, in ODE as well as DDE and already on master: resize_jac_config!, resize_grad_config! and resize_J_W! took the backend from alg_autodiff(integrator.alg), which for a composite is the last member's. They now take an alg argument (defaulting to integrator.alg) and the Rosenbrock and DDE resize methods pass each member's cache.alg. Fixing it in alg_autodiff(::CompositeAlgorithm) instead would change other consumers such as DAE initialization.

The repro in #4952 (MethodOfSteps(AutoTsit5(Rosenbrock23())), resized while on the stiff member) threw a DimensionMismatch and now returns Success at [0.5, 0.5]. New ODE and DDE testsets cover a FiniteDiff and a ForwardDiff Rosenbrock23, dense and with KrylovJL_GMRES, in both orders; they error without the src change and pass with it.

Not covered here: growing a composite with SDIRK members fails on master in resize_nlsolver!, and its Newton path still resizes with integrator.alg.

Fixes #4952

AI Disclosure: Claude assisted with this change.

[ChrisRackauckas]

🤖 Automated comment from an AI agent running as @ChrisRackauckas — not written or reviewed by Chris.

PR: #4958

Reviewed head: d436d9c
Raw SHA-256: 73724c22dc6a78522de8c1ff56cbd22801ea276f44b60c04119db99932277541
Source SHA-256: 681965a75349d469ba5ce5146b7de0322b07ddc6f8b9c0fdbe25720fbe51c0d9
Source: standard-review-v1 (harness-log)
Reviewer session: 01a1d257-7f93-74e1-a6be-c01a65be15a9
Slot: 1
Confidence: unknown
Own risk fixed before author rating/discussion: yes

Risk assessment

• Risk: medium (M1)
• Verdict: CHANGES
• Reviewed head: d436d9c

Review report

VERDICT: CHANGES
Round: 1
Head: d436d9c
Slot: 1
CI: CI is not green (regress:SciMLSensitivity.jl/Core4/1 / Downstream Tests - Core4)
Session: 01a1d257-7f93-74e1-a6be-c01a65be15a9

Risk

medium (M1: a new specialization of DiffEqBase.resize_non_user_cache! changes dispatch for existing composite DDE integrators).

Blocking findings

• Verdict: CHANGES
• Round: 1
• Blocking priorities: P2
• Deferred P2 issues: None.

[P2] Preserve each member's AD backend when resizing its cache — lib/DelayDiffEq/src/integrators/interface.jl:290. Confirmed by running: a composite containing finite-difference and ForwardDiff Rosenbrock23 members throws on even a same-size resize!(integ, 1), in either member order. The exception is MethodError: Cannot convert between ForwardDiffTwoArgJacobianPrep and FiniteDiffTwoArgJacobianPrep. With only the added method deleted at runtime, both otherwise identical cases finish with Success and state [0.49999999954711405].

The new loop passes the whole composite integrator to each cache's resize helper. Reading lib/OrdinaryDiffEqDifferentiation/src/derivative_wrappers.jl:337 and lib/OrdinaryDiffEqDifferentiation/src/alg_utils.jl:7 explains the observed failure: the helper selects the final algorithm's backend for every member. Rebuild configurations with the corresponding member's backend, and add a regression covering heterogeneous backends and same-size resizing.

Runnable reproduction in the resolved test environment:

using DelayDiffEq, OrdinaryDiffEqRosenbrock
using OrdinaryDiffEqCore: CompositeAlgorithm
using ADTypes: AutoFiniteDiff
f!(du, u, h, p, t) = (du .= -1000 .* (u .- 0.5); nothing)
prob = DDEProblem(f!, [1.0], (p, t) -> [0.5], (0.0, 0.1);
    constant_lags = [1.0])
alg = CompositeAlgorithm((Rosenbrock23(autodiff = AutoFiniteDiff()),
    Rosenbrock23()), _ -> 1)
integ = init(prob, MethodOfSteps(alg))
step!(integ)
resize!(integ, 1) # MethodError on this head

Non-blocking findings

None established.

What I ran

• AI reviewer: codex-cli 0.164.0, model gpt-6.1-sol; session identified above. Blind code review; no PR body, comments, reviews, or author risk rating read.
• git diff origin/master...HEAD, git diff --check origin/master...HEAD, repository instruction discovery, public cache API and resize-helper inspection. Diff check passed; no root AGENTS.md or CLAUDE.md was present. The newest commit has neither Agent-Harness nor Agent-Model trailers.
• julia --version: 1.13.2. Scratch test environment uses the package's declared dependencies, extras, compat bounds and local monorepo sources; original layered JULIA_DEPOT_PATH preserved. Resolved DelayDiffEq 6.4.7, OrdinaryDiffEqCore 4.20.0, OrdinaryDiffEqDifferentiation 3.13.0, DiffEqBase 7.21.5, SciMLBase 3.58.0, LinearSolve 5.18.3, ForwardDiff 1.4.7.
• julia --startup-file=no setup.jl: dependency resolution passed. Targeted Pkg.precompile(...): 27 dependencies successfully precompiled; 231 already precompiled.
• Fresh-process julia --startup-file=no --check-bounds=yes --project=testenv mixed_regression.jl: two same-size head cases failed as described; removing only the new method made both controls succeed. No source file was changed for this control.
• julia --startup-file=no --check-bounds=yes --project=testenv run_review.jl: the official GROUP=Integrators entrypoint passed 383 assertions, with 36 existing broken assertions (SDIRK); no failures or errors in that group. The added regression passed 2/2. Deleting only the added method made that same regression error with DimensionMismatch: ... OneTo(1) and ... OneTo(2).
• Initial adversarial batch: 66 passed, 2 failed, 6 errored. Two errors reproduce the blocking mixed-backend bug. Four errors concern ForwardDiff chunk size 3 after shrinking to 2; their baseline status was not verified. The two failed assertions assumed AutoSwitch would remain on member 2; that probe assumption was invalid, and both analytic-solution checks passed. The optional deterministic-member follow-up was not run.
• Isolated main-case probe passed: the stiff member was reached after 21 steps at t=0.05768237953913356; growth to 2 followed by solve! returned Success, t=3.0, state [0.5000000000000001, 0.5].
• Final git status --short and git diff --exit-code: checkout unchanged. No commits, pushes, or GitHub posts.
• uv run --locked python --version: Python 3.15.0; uv warned that --locked has no effect outside a project. This Julia checkout has no supplied Python project/uv lock environment. No tracked Julia interpreter/Manifest lock was present either; the scratch environment was resolved from the package constraints.

What I did not verify

• Whether the four shrinking/deletion errors with ForwardDiff chunk size 3 are pre-existing; direct-solver/baseline controls were not run. The optional deterministic-member delayed-solution follow-up also did not start: its idle resource queue was interrupted after the required suite and probes completed. These are not additional confirmed PR regressions.
• Supplied SciMLSensitivity/Core4 CI failure, other Julia versions, full monorepo/downstream suites, QA and documentation builds.

AUTHOR_GRADE: {"task": "bugfix", "score": 3, "difficulty": 2, "rounds": 1, "author_harness": "unknown", "author_model": "unknown", "note": "The original bug has failing-before/passing-after evidence and the affected group passes. A targeted head/control comparison confirms a newly introduced mixed-AD-backend regression even for same-size resizing."}

Head link: d436d9c

---

🤖 Posted by an AI agent — harness: codex-cli 0.164.0 · model: gpt-6.1-sol
Conversation: local session 01a1d257-7f93-74e1-a6be-c01a65be15a9 (no session URL)