feat: Add Magnetohydrodynamics (MHD) benchmark suite

.github/workflows/toolchain-image.yml

@@ -51,6 +51,7 @@ jobs:
           labels: ${{ steps.meta.outputs.labels }}
           build-args: |
             FLAP_VERSION=1.2.16
+            OPENCODE_VERSION=1.3.3
             CATCH2_REF=v3.13.0
             MDSPAN_REF=mdspan-0.6.0
             XTL_REF=0.8.2

.gitignore

@@ -8,3 +8,5 @@ __pycache__/
 .venv/
 venv/
 .DS_Store
+benchmarks/*/*/workspace/build/
+benchmarks/*/*/workspace/build-review/

.vscode/settings.json

@@ -0,0 +1,35 @@
+{
+    "files.exclude": {
+        "**/.git/objects/**": true,
+        "**/.git/refs/**": true,
+        "**/.pytest_cache/**": true,
+        "**/__pycache__": true,
+        "**/.ruff_cache/**": true,
+        "**/.venv/**": true,
+        "**/.opencode-data/**": true
+    },
+    "files.watcherExclude": {
+        "**/runs/**": true,
+        "**/.git/objects/**": true,
+        "**/.git/refs/**": true,
+        "**/.ruff_cache/**": true,
+        "**/.venv/**": true,
+        "**/.opencode-data/**": true,
+        "**/node_modules/**": true,
+        "**/__pycache__/**": true,
+        "**/build/**": true
+    },
+    "search.exclude": {
+        "**/runs/**": true,
+        "**/.git/**": true,
+        "**/.ruff_cache/**": true,
+        "**/.venv/**": true,
+        "**/.opencode-data/**": true,
+        "**/node_modules/**": true,
+        "**/__pycache__/**": true
+    },
+    "C_Cpp.files.exclude": {
+        "**/build/**": true,
+        "**/runs/**": true
+    }
+}

README.md

@@ -30,6 +30,7 @@ Configured in `agents_default.toml` (with per-agent overrides under `sample/*.to
 ## Available Benchmarks
 - Demo benchmark for Runge-Kutta 2 (RK2) midpoint method.
 - 3D wave equation solver with finite difference method.
+- Magnetohydrodynamics (MHD) solver.
 
 ## Quick Start
 

agents_default.toml

@@ -10,8 +10,10 @@ pass_env = [
 ]
 pre = [
   "mkdir -p \"$HOME/.local/share/opencode\"",
+  "mkdir -p \"$HOME/.config/opencode\"",
   "if [ -f /opencode-auth.json ]; then cp /opencode-auth.json \"$HOME/.local/share/opencode/auth.json\"; fi",
-  "if [ -f /opencode-config.json ]; then export OPENCODE_CONFIG=/opencode-config.json; elif [ -f /opencode-config.jsonc ]; then export OPENCODE_CONFIG=/opencode-config.jsonc; fi",
+  "if [ -f /opencode-config.json ]; then cp /opencode-config.json \"$HOME/.config/opencode/opencode.json\"; fi",
+  "if [ -f /opencode-config.jsonc ]; then cp /opencode-config.jsonc \"$HOME/.config/opencode/opencode.jsonc\"; fi",
 ]
 cmd = "stdbuf -oL -eL opencode run -m \"$BENCH_MODEL\" $BENCH_MODEL_OPTIONS_ARGS --thinking --dir / \"$(cat \"$BENCH_PROMPT_FILE\")\" -f \"$BENCH_SPEC_FILE\""
 

benchmarks/magnetohydrodynamics/README.md

@@ -6,13 +6,29 @@ This suite contains benchmark tasks for ideal magnetohydrodynamics solvers.
 
 - `shared/workspace/basic_equations.md`: suite-wide notation and flux conventions.
 - `shared/workspace/hlld.md`: HLLD algorithm notes for solver tasks.
-- `cpp-hlld/`: C++ HLLD approximate Riemann solver task.
+- `cpp-hlld-00/`: default C++ HLLD task with detailed solver guidance in spec.
+- `cpp-hlld-01/`: variant C++ HLLD task with reduced guidance but same test intent.
+- `cpp-full1d-00/`: easiest C++ full 1D ideal MHD variant (main+HLLD provided, solver scaffolded).
+- `cpp-full1d-01/`: reduced-guidance full 1D variant with only `evolve_ssp_rk3(...)` exposed.
+- `shared/eval/README.md`: hidden-eval contract for shared MHD scoring assets.
+- `shared/eval/mhd1d_shared.py`: shared helpers for CSV loading, score
+  windows, and comparison metadata.
+- `shared/eval/fixtures/mhd1d/`: hidden fixtures for full 1D variants.
 
 ## Notes
 
 - Shared workspace files are visible to the agent during benchmark runs.
 - Keep maintainer-only derivations, generators, and hidden fixtures outside the
   shared workspace.
+- `cpp-hlld-00` and `cpp-hlld-01` expose only
+  `hlld_flux_from_primitive(...)` in the public task API.
+- `cpp-hlld-00` and `cpp-hlld-01` keep public/hidden test intent aligned;
+  the main difference is prompt detail level.
+- `cpp-full1d-00` public tests compare solver CSV output against a golden file
+  with numeric tolerance (`1.0e-12`), and hidden tests use `nx=200` against a
+  hidden reference CSV with the same numeric policy.
+- Full 1D tasks score interior cells and emit CSV with lowercase magnetic-field
+  headers (`by`, `bz`).
 
 ## Reference credit
 

benchmarks/magnetohydrodynamics/cpp-full1d-00/eval/run.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+set -u -o pipefail
+
+cd /work
+export PYTHONPATH="/work:/eval_shared"
+
+status="passed"
+score="1.0"
+
+python3 -m pytest -q /eval/tests
+rc=$?
+if [ "$rc" -ne 0 ]; then
+  status="failed"
+  score="0.0"
+fi
+
+python3 - <<PY
+import json
+from pathlib import Path
+
+out = {
+  "status": "$status",
+  "score": float("$score"),
+}
+Path("result.json").write_text(json.dumps(out, indent=2) + "\n", encoding="utf-8")
+PY
+
+exit 0

benchmarks/magnetohydrodynamics/cpp-full1d-00/eval/tests/test_hidden.py

@@ -0,0 +1,55 @@
+import csv
+import os
+import subprocess
+from pathlib import Path
+
+from mhd1d_shared import TOLERANCE, assert_csv_rows_close
+
+SOLVER_TARGET = "cpp_full_solver1d"
+WORKSPACE_ROOT = Path("/work")
+REFERENCE_CSV_PATH = (
+    Path("/eval_shared") / "fixtures" / "mhd1d" / "brio_wu_reference.csv"
+)
+
+
+def _build_solver(build_dir: Path) -> Path:
+    subprocess.run(
+        ["cmake", "-S", ".", "-B", str(build_dir)], check=True, cwd=WORKSPACE_ROOT
+    )
+    subprocess.run(
+        ["cmake", "--build", str(build_dir), "--target", SOLVER_TARGET],
+        check=True,
+        cwd=WORKSPACE_ROOT,
+    )
+
+    binary_name = f"{SOLVER_TARGET}.exe" if os.name == "nt" else SOLVER_TARGET
+    solver_path = build_dir / "bin" / binary_name
+    assert solver_path.exists()
+    return solver_path
+
+
+def test_hidden_brio_wu_cli_matches_fixture(tmp_path: Path) -> None:
+    solver_path = _build_solver(tmp_path / "build")
+    output_csv_path = tmp_path / "brio_wu.csv"
+
+    completed = subprocess.run(
+        [str(solver_path), "200"],
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    output_csv_path.write_text(completed.stdout, encoding="utf-8")
+
+    output_rows = list(
+        csv.reader(output_csv_path.read_text(encoding="utf-8").splitlines())
+    )
+    reference_rows = list(
+        csv.reader(REFERENCE_CSV_PATH.read_text(encoding="utf-8").splitlines())
+    )
+
+    assert_csv_rows_close(
+        output_rows,
+        reference_rows,
+        tolerance=TOLERANCE,
+        expected_header=["x", "rho", "u", "v", "w", "p", "by", "bz"],
+    )

benchmarks/magnetohydrodynamics/cpp-full1d-00/spec.md

@@ -0,0 +1,74 @@
+# cpp-full1d-00
+
+Implement a 1D ideal-MHD solver CLI in C++.
+
+## Read first
+
+- `/work/basic_equations.md`
+- `/work/hlld.md`
+
+## Task
+
+The command-line entrypoint and the HLLD Riemann solver are already provided.
+The CLI accepts an integer `nx` argument for the number of grid points, performs the Brio-Wu Riemann problem, and writes the solution to stdout in CSV format.
+Your main task is to complete the solver implementation in `src/mhd1d.cpp`.
+
+The CLI output must match the provided golden CSV for `nx=100` within numeric tolerance (`1.0e-12`).
+
+## How to test
+
+Run the public checks from the workspace:
+
+```bash
+python3 -m pytest -q tests/test_public.py
+```
+
+## Local dev
+
+```bash
+pytest -q
+```
+
+To build manually:
+
+```bash
+cmake -S . -B build
+cmake --build build
+./build/bin/cpp_full_solver1d
+```
+
+## Numerical Algorithm
+
+- Riemann solver: HLLD
+- Primitive variables reconstruction: piecewise linear with MC2 slope limiter
+- Time integration: SSP-RK3
+- Boundary condition: symmetric (zero-gradient)
+
+## Files
+
+- `src/main.cpp`: complete CLI (already done)
+- `src/hlld.hpp`, `src/hlld.cpp`: complete HLLD implementation (already done)
+- `src/mhd1d.hpp`, `src/mhd1d.cpp`: solver scaffolding to complete
+
+## Functions to complete (in `src/mhd1d.cpp`)
+
+The easiest path is to implement these functions first:
+
+1. `primitive_to_conservative(...)`
+2. `conservative_to_primitive(...)`
+3. `compute_lr(...)`
+4. `compute_rhs(...)`
+5. `push_ssp_rk3(...)`
+6. `evolve_ssp_rk3(...)`
+
+Recommended implementation order:
+
+1. Reconstruction (`compute_lr`)
+2. Flux loop (`compute_flux_hlld` already calls provided HLLD)
+3. RHS assembly (`compute_rhs`)
+4. One RK3 step (`push_ssp_rk3`)
+5. Time loop (`evolve_ssp_rk3`)
+
+## Standards
+
+- C++17

benchmarks/magnetohydrodynamics/cpp-full1d-00/task.toml

@@ -0,0 +1,8 @@
+id = "cpp-full1d-00"
+suite = "magnetohydrodynamics"
+language = "cpp"
+time_limit_sec = 600
+eval_cmd = "/eval/run.sh"
+prompt = "Read /run/spec.md, /work/basic_equations.md, and /work/hlld.md, then solve the task in /work."
+use_shared_workspace = true
+use_shared_eval = true

benchmarks/magnetohydrodynamics/cpp-full1d-00/workspace/CMakeLists.txt

@@ -0,0 +1,61 @@
+cmake_minimum_required(VERSION 3.16)
+
+project(cpp_full_solver1d LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+include(FetchContent)
+
+find_package(Catch2 3 QUIET)
+
+if(NOT Catch2_FOUND)
+  FetchContent_Declare(
+    Catch2
+    GIT_REPOSITORY https://github.com/catchorg/Catch2.git
+    GIT_TAG v3.13.0
+    GIT_SHALLOW TRUE
+  )
+  FetchContent_MakeAvailable(Catch2)
+endif()
+
+add_library(mhd1d_solver
+  src/mhd1d.cpp
+  src/hlld.cpp
+)
+
+target_include_directories(mhd1d_solver PUBLIC
+  src
+  ../../../common/include
+)
+
+add_executable(cpp_full_solver1d
+  src/main.cpp
+)
+
+target_link_libraries(cpp_full_solver1d PRIVATE
+  mhd1d_solver
+)
+
+set_target_properties(cpp_full_solver1d PROPERTIES
+  RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin"
+)
+
+add_executable(cpp_full_solver1d_public_tests
+  tests/cpp/test_public.cpp
+)
+
+target_link_libraries(cpp_full_solver1d_public_tests PRIVATE
+  mhd1d_solver
+  Catch2::Catch2WithMain
+)
+
+target_include_directories(cpp_full_solver1d_public_tests PRIVATE
+  src
+  ../../../common/include
+)
+
+set_target_properties(cpp_full_solver1d_public_tests PROPERTIES
+  RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/tests"
+)

benchmarks/magnetohydrodynamics/cpp-full1d-00/workspace/README.md

@@ -0,0 +1,7 @@
+The public C++ workspace contains a Brio-Wu solver scaffold.
+
+- `src/main.cpp` is complete.
+- `src/hlld.cpp` is complete.
+- `src/mhd1d.cpp` contains TODO sections to implement.
+
+Shared workspace docs are already mounted for this benchmark.

benchmarks/magnetohydrodynamics/cpp-full1d-00/workspace/pyproject.toml

@@ -0,0 +1,7 @@
+[project]
+name = "magnetohydrodynamics-cpp-full1d-00"
+version = "0.0.0"
+requires-python = ">=3.10"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]