Decide the parameters in param.json

[su-zhongan]

I replaced the methane POSCAR on the dpgen official website with propane. At this point, how should 'sel' and 'rcut' be determined? Besides that, is there any need to modify the other parameters?

[dosubot[bot]]

Hi @su-zhongan! I'm Dosu and I'm helping the DeepModeling team.

When you replace methane with propane, the key parameter you must change is sel, because it depends on your molecule's composition and your type_map order. The sel list specifies the maximum number of neighbors of each type within the cutoff radius, and its order must match type_map [1].

Determining sel for propane (C₃H₈):

With type_map: ["C", "H"]:

• sel[0] = max number of C neighbors within rcut
• sel[1] = max number of H neighbors within rcut

The methane example uses sel: [16, 4], which means 16 H neighbors and 4 C neighbors — but that's for type_map: ["H", "C"]. If your type_map is ["C", "H"], then sel[0] counts C and sel[1] counts H. For propane, each atom has at most 2–3 C neighbors but up to 6–8 H neighbors within a 5.0 Å cutoff. So you'd want something like "sel": [6, 16] for type_map: ["C", "H"].

The easiest approach is to simply set "sel": "auto" — DeePMD-kit will automatically compute the appropriate values from your training data by scanning all frames, finding the per-type maximum neighbor count, and adding a safety margin (×1.1 by default) [1].

Determining rcut:

For a gas-phase organic molecule like propane, rcut: 6.0 Å is a common choice (the methane example uses 5.0 Å). Since propane is larger than methane, you may want to slightly increase it (e.g., 6.0 Å) to capture interactions between atoms further apart in the molecule. This is especially important if your simulation box contains multiple propane molecules.

Other parameters to check:

• stop_batch: The methane example uses only 20,000 steps which is very short. For propane you may want 50,000–100,000 with appropriate decay_steps to ensure training converges.
• model_devi_f_trust_lo/hi: Make sure these are set somewhat above your training force RMSE (typically 2–5× the RMSE). If they're too tight relative to the model quality, you'll get accurate_ratio = 0%.
• neuron sizes: The methane example's small descriptor network [10, 20, 40] may be sufficient for propane since it's still a simple hydrocarbon, but consider increasing if accuracy is poor.

A user recently reported exactly this issue (replacing methane with propane and getting accurate_ratio always 0%) — the root cause was that sel was inverted for the new system [2]. Fixing sel (ideally to "auto") and ensuring reasonable trust levels resolved it.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[dosubot[bot]]

@su-zhongan Here's everything you need:

1. Fix your POSCAR

Remove the selective dynamics lines. Your cleaned POSCAR should look like this:

La0.5Sr0.5FeO3 FeO2-terminated slab
   1.000000000000000
  5.5896837107350699  0.0000000000000000  0.0000030245792512
  0.0000000000000000  5.5704082918990396  0.0000000000000000
 -0.0000156431400586  0.0000000000000000 28.9098740415833007
   O    Fe   La   Sr
   24   8   4   4
Direct
  0.7283156044740990  0.2758054792891759  0.0790439136642109
  0.7283163668137300  0.2758054792891759  0.3514468584971760
  ... (all coordinates, NO "Selective dynamics", NO T/F flags)

Remove the Selective dynamics line and all T T T / F F F markers — dpgen strips them anyway [1], and fixed atoms during AIMD create inconsistent training labels [2].

2. run param.json

{
    "type_map":        ["O", "Fe", "La", "Sr"],
    "mass_map":        [16, 56, 139, 88],
    "init_data_prefix": "../",
    "init_data_sys":   ["init/LSFO_slab/02.md/sys-0024-0008-0004-0004/deepmd"],
    "init_batch_size": ["auto"],
    "sys_configs_prefix": "../",
    "sys_configs": [
        ["init/LSFO_slab/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00000[0-9]/POSCAR"],
        ["init/LSFO_slab/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00001*/POSCAR"],
        ["init/LSFO_slab/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00002*/POSCAR"]
    ],
    "sys_batch_size":  ["auto", "auto", "auto"],
    "numb_models": 4,
    "default_training_param": {
        "model": {
            "type_map": ["O", "Fe", "La", "Sr"],
            "descriptor": {
                "type":        "se_e2_a",
                "sel":         "auto",
                "rcut":        7.0,
                "rcut_smth":   0.5,
                "neuron":      [25, 50, 100],
                "resnet_dt":   true,
                "axis_neuron": 16,
                "seed":        12345
            },
            "fitting_net": {
                "neuron":     [240, 240, 240],
                "resnet_dt":  false,
                "seed":       12345
            }
        },
        "learning_rate": {
            "type":        "exp",
            "start_lr":    0.001,
            "stop_lr":     1e-8,
            "decay_steps": 5000
        },
        "loss": {
            "start_pref_e": 0.02,
            "limit_pref_e": 2,
            "start_pref_f": 1000,
            "limit_pref_f": 1,
            "start_pref_v": 0.0,
            "limit_pref_v": 0.0
        },
        "training": {
            "seed":       12345,
            "stop_batch": 400000,
            "disp_file":  "lcurve.out",
            "disp_freq":  1000,
            "save_freq":  1000,
            "save_ckpt":  "model.ckpt",
            "disp_training": true,
            "time_training": true
        }
    },
    "model_devi_dt":            0.002,
    "model_devi_skip":          0,
    "model_devi_f_trust_lo":    0.15,
    "model_devi_f_trust_hi":    0.35,
    "model_devi_clean_traj":    false,
    "model_devi_jobs": [
        {"sys_idx": [0,1,2], "temps": [600],      "trj_freq": 10, "nsteps": 1000, "ensemble": "nvt", "_idx": "00"},
        {"sys_idx": [0,1,2], "temps": [600],      "trj_freq": 10, "nsteps": 1000, "ensemble": "nvt", "_idx": "01"},
        {"sys_idx": [0,1,2], "temps": [600],      "trj_freq": 10, "nsteps": 2000, "ensemble": "nvt", "_idx": "02"},
        {"sys_idx": [0,1,2], "temps": [600, 800], "trj_freq": 10, "nsteps": 3000, "ensemble": "nvt", "_idx": "03"},
        {"sys_idx": [0,1,2], "temps": [600, 800], "trj_freq": 10, "nsteps": 5000, "ensemble": "nvt", "_idx": "04"},
        {"sys_idx": [0,1,2], "temps": [600, 800], "trj_freq": 10, "nsteps": 5000, "ensemble": "nvt", "_idx": "05"}
    ],
    "fp_style":        "vasp",
    "shuffle_poscar":  false,
    "fp_task_max":     30,
    "fp_task_min":     5,
    "fp_pp_path":      "./",
    "fp_pp_files":     ["POTCAR_O", "POTCAR_Fe", "POTCAR_La", "POTCAR_Sr"],
    "fp_incar":        "./INCAR_fp"
}

Key design choices:

• sel: "auto" — essential for a 4-element system [3]
• rcut: 7.0 — appropriate for oxide systems; 12.0 would work but significantly increases computation [4]
• ensemble: "nvt" — must use NVT for slabs, NPT would collapse the vacuum [5]
• f_trust_lo: 0.15 / f_trust_hi: 0.35 — start moderately; tighten after force RMSE improves. Calibrate to ~2–5× your training force RMSE [3]
• stop_batch: 400000 — complex 4-element oxide needs much more training than propane
• Larger networks ([25,50,100] descriptor, [240,240,240] fitting) — necessary for 4-element chemistry
• Progressive nsteps (1000 → 5000) — start short, increase as the model stabilizes
• Temperatures — 600K matches typical ammonia oxidation conditions; add 800K in later iterations for broader coverage

3. INCAR_fp (for fp labeling):

PREC    = Accurate
ENCUT   = 550
EDIFF   = 1E-5
NELM    = 200
IBRION  = -1
NSW     = 0
ISMEAR  = 0
SIGMA   = 0.05
KSPACING = 0.4
KGAMMA  = .FALSE.
LREAL   = Auto
LASPH   = .TRUE.
LMAXMIX = 6
LWAVE   = .FALSE.
LCHARG  = .FALSE.
ISYM    = 0

This is a single-point calculation (NSW=0, IBRION=-1) — dpgen's fp step only labels structures, no relaxation.

4. machine.json — you can reuse the same structure from your propane runs. Just update the remote_root paths. You need all three sections: train, model_devi, and fp.

Note: The init_data_sys and sys_configs paths above are examples — adjust them to match your actual directory structure after running init. Also, the fp_pp_files order must match type_map: O first, then Fe, La, Sr.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[su-zhongan]

How should I set up the param.json and machine.json files in run next, and why is there no 02.md file in the FeO2.POSCAR.01X01X01 file when I run the FeO2 surface init? @dosu

[dosubot[bot]]

@su-zhongan The missing 02.md directory is most likely caused by one of two things:

Most likely: You used dpgen init_surf instead of dpgen init_bulk

dpgen init_surf only implements stages 1 and 2 — it generates slabs and performs scale/perturbation, but it does NOT run MD or create 02.md [1]. This is different from dpgen init_bulk, which has all 4 stages:

Stage	What it does	Output
1	Relax initial structure	00.place_ele/
2	Scale & perturb	01.scale_pert/
3	Run short AIMD	02.md/
4	Collect MD data	Collects from 02.md/

Alternative: Your stages parameter doesn't include stage 3

If you used dpgen init_bulk but set "stages": [1, 2] in your param.json (as I suggested in my previous reply for the slab), then stage 3 was never executed and 02.md was never created [2].

How to fix this:

Change your param.json to include all 4 stages:

"stages": [1, 2, 3, 4]

Use dpgen init_bulk (not init_surf), since you already have your own POSCAR with from_poscar: true. Make sure you also have a valid md_incar (your INCAR.md file) in the specified path — without it, stage 3 will fail before creating 02.md.

Regarding the run param.json and machine.json:

The run configuration I provided in my previous reply is still applicable. The key points are:

• type_map must match your POSCAR element order: ["O", "Fe", "La", "Sr"]
• ensemble: "nvt" — mandatory for slabs to preserve vacuum [3]
• sel: "auto" for a 4-element system [4]
• stop_batch: 400000 for a complex oxide
• For machine.json, reuse your existing one from the propane runs — just update the remote_root paths

Once 02.md is successfully generated and data is collected (stage 4), the init_data_sys path in your run param.json should point to the deepmd data inside 02.md/. For example:

"init_data_sys": ["init/FeO2.POSCAR.01x01x01/02.md/sys-0024-0008-0004-0004/deepmd"]

Adjust the sys-XXXX directory name to match whatever dpgen creates based on your atom counts.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[su-zhongan]

So what should I do next to modify the INCAR, param.json, and machine.json files in my run? @dosu
PREC=A
ENCUT=400
ISYM=0
ALGO=Fast
EDIFF=1.000000e-06
LREAL=False
NPAR=1
KPAR=1
NELM=120
NELMIN=4
ISIF=2
ISMEAR=0
SIGMA=0.20000
IBRION=-1
NSW=0
LWAVE=F
LCHARG=F
PSTRESS=0
KSPACING=0.500000
KGAMMA=.FALSE.

{
"type_map": ["H","C"],
"mass_map": [1,12],
"init_data_prefix": "../",
"init_data_sys": ["init/CH4.POSCAR.01x01x01/02.md/sys-0004-0001/deepmd"],
"seed":12345,
"lmp_cmd": "lmp -var seed 12345",
"init_batch_size": [8],
"sys_configs_prefix": "../",
"sys_configs": [
["init/CH4.POSCAR.01x01x01/01.scale_pert/sys-0004-0001/scale-1.000/000000/POSCAR",
"init/CH4.POSCAR.01x01x01/01.scale_pert/sys-0004-0001/scale-1.000/000001/POSCAR",
"init/CH4.POSCAR.01x01x01/01.scale_pert/sys-0004-0001/scale-1.000/00000[2-9]/POSCAR"],
["init/CH4.POSCAR.01x01x01/01.scale_pert/sys-0004-0001/scale-1.000/00001*/POSCAR"],
["init/CH4.POSCAR.01x01x01/01.scale_pert/sys-0004-0001/scale-1.000/00002*/POSCAR"]
],
"sys_batch_size": [8,8],
"_comment": " that's all ",
"numb_models": 4,
"train_backend":"pytorch",
"default_training_param": {
"model": {
"type_map": ["H","C"],
"descriptor": {
"type": "se_e2_a",
"sel": [16,4],
"rcut_smth": 0.5,
"rcut": 5.0,
"neuron": [10,20,40],
"resnet_dt": true,
"axis_neuron": 12,
"seed": 12345
},
"fitting_net": {
"neuron": [120,120,120],
"resnet_dt": false,
"seed": 12345
}
},
"learning_rate": {
"type": "exp",
"start_lr": 0.001,
"stop_lr": 1e-8,
"decay_steps": 5000
},
"loss": {
"start_pref_e": 0.02,
"limit_pref_e": 2,
"start_pref_f": 1000,
"limit_pref_f": 1,
"start_pref_v": 0.0,
"limit_pref_v": 0.0
},
"training": {
"seed":12345,
"stop_batch": 20000,
"disp_file": "lcurve.out",
"disp_freq": 1000,
"numb_test": 4,
"save_freq": 1000,
"save_ckpt": "model.ckpt",
"disp_training": true,
"time_training": true,
"profiling": false,
"profiling_file": "timeline.json",
"_comment": "that's all"
}
},
"model_devi_dt": 0.002,
"model_devi_skip": 0,
"model_devi_f_trust_lo": 0.05,
"model_devi_f_trust_hi": 0.15,
"model_devi_clean_traj": false,
"model_devi_jobs": [
{"sys_idx": [0],"temps": [100],"press": [1.0],"trj_freq": 10,"nsteps": 1000,"ensemble": "nvt","_idx": "00"},
{"sys_idx": [1],"temps": [100],"press": [1.0],"trj_freq": 10,"nsteps": 3000,"ensemble": "nvt","_idx": "01"},
{"sys_idx": [2],"temps": [100],"press": [1.0],"trj_freq": 10,"nsteps": 5000,"ensemble": "nvt","_idx": "02"}
],
"fp_style": "vasp",
"shuffle_poscar": false,
"fp_task_max": 30,
"fp_task_min": 8,
"fp_pp_path": "./",
"fp_pp_files": ["POTCAR_H","POTCAR_C"],
"fp_incar": "./INCAR_methane"
}
{
"api_version": "1.0",
"deepmd_version": "3.1.2",
"train": [
{
"command": "dp",
"machine": {
"batch_type": "Shell",
"context_type": "local",
"local_root": "./",
"remote_root": "/home/xgduan/gdata/DPGEN_examlpe/00_gasous_CH4/run_node11/work",
"clean_asynchronously": false
},
"resources": {
"number_node": 1,
"cpu_per_node": 40,
"gpu_per_node": 1,
"group_size": 4,
"envs": {
"PATH": "/gdata01/opt/anaconda3-2025.12-1/envs/deepmd312/bin:$PATH"
}

        }
    }
],
"model_devi": [
    {
        "command": "lmp",
        "machine": {
            "batch_type": "Shell",
            "context_type": "local",
            "local_root": "./",
            "remote_root": "/home/xgduan/gdata/DPGEN_examlpe/00_gasous_CH4/run_node11/work",
            "clean_asynchronously": false
        },
        "resources": {
            "number_node": 1,
            "cpu_per_node": 40,
            "gpu_per_node": 1,
            "group_size": 1,
            "envs": {
                "PATH": "/gdata01/opt/anaconda3-2025.12-1/envs/deepmd312/bin:$PATH"}
        }
    }
],
"fp": [
    {
        "command": "mpirun -n 40 vasp.5.3.5.beef",
        "machine": {
            "batch_type": "Torque",
            "context_type": "SSHContext",
            "local_root": "./",
            "remote_root": "/home/xgduan/gdata/DPGEN_examlpe/00_gasous_CH4/run_node11/work",
            "clean_asynchronously": false,
            "remote_profile": {
             "hostname": "node00",
             "username": "xgduan",
             "key_filename": "/etc/ssh/node16-to-node00",
             "look_for_keys": false
  }
        },
        "resources": {
            "number_node": 1,
            "cpu_per_node":40,
            "gpu_per_node": 0,
            "queue_name": "normal40",
            "group_size": 4,
            "para_deg": 1,
            "envs": {
                "OMP_NUM_THREADS": "1",
                "MKL_NUM_THREADS": "1",
                "PATH": "/usr/local/bin:/gdata01/opt/anaconda3-2025.12-1/envs/deepmd312/bin:$PATH"}
         }
    }
]

}

[dosubot[bot]]

@su-zhongan Those are the methane example files — you need to make significant changes for your FeO₂-terminated La₁₋ₓSrₓFeO₃ slab. Here's exactly what to change in each file:

INCAR (fp_incar) — changes needed:

PREC    = Accurate
ENCUT   = 550          # ← increase from 400, O needs higher cutoff
ISYM    = 0
ALGO    = Fast
EDIFF   = 1.000000e-05 # ← slightly relaxed for single-point efficiency
LREAL   = Auto         # ← change from False, acceptable for ~40 atom slab
NPAR    = 4            # ← adjust to sqrt(num_cores)
NELM    = 200          # ← increase from 120, oxide convergence can be harder
NELMIN  = 4
ISIF    = 2
ISMEAR  = 0
SIGMA   = 0.05         # ← decrease from 0.20, this is an insulator
IBRION  = -1
NSW     = 0
LWAVE   = F
LCHARG  = F
PSTRESS = 0
KSPACING = 0.40        # ← decrease from 0.50 for better accuracy
KGAMMA  = .FALSE.
LASPH   = .TRUE.       # ← ADD: needed for La f-electrons
LMAXMIX = 6            # ← ADD: needed for f-electron SCF convergence

param.json — changes needed:

{
    "type_map": ["O", "Fe", "La", "Sr"],
    "mass_map": [16, 56, 139, 88],
    "init_data_prefix": "../",
    "init_data_sys": ["init/FeO2.POSCAR.01x01x01/02.md/sys-0024-0008-0004-0004/deepmd"],
    "init_batch_size": ["auto"],
    "sys_configs_prefix": "../",
    "sys_configs": [
        ["init/FeO2.POSCAR.01x01x01/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00000[0-9]/POSCAR"],
        ["init/FeO2.POSCAR.01x01x01/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00001*/POSCAR"],
        ["init/FeO2.POSCAR.01x01x01/01.scale_pert/sys-0024-0008-0004-0004/scale-1.000/00002*/POSCAR"]
    ],
    "sys_batch_size": ["auto", "auto", "auto"],
    "numb_models": 4,
    "train_backend": "pytorch",
    "default_training_param": {
        "model": {
            "type_map": ["O", "Fe", "La", "Sr"],
            "descriptor": {
                "type": "se_e2_a",
                "sel": "auto",
                "rcut_smth": 0.5,
                "rcut": 7.0,
                "neuron": [25, 50, 100],
                "resnet_dt": true,
                "axis_neuron": 16,
                "seed": 12345
            },
            "fitting_net": {
                "neuron": [240, 240, 240],
                "resnet_dt": false,
                "seed": 12345
            }
        },
        "learning_rate": {
            "type": "exp",
            "start_lr": 0.001,
            "stop_lr": 1e-8,
            "decay_steps": 5000
        },
        "loss": {
            "start_pref_e": 0.02,
            "limit_pref_e": 2,
            "start_pref_f": 1000,
            "limit_pref_f": 1,
            "start_pref_v": 0.0,
            "limit_pref_v": 0.0
        },
        "training": {
            "seed": 12345,
            "stop_batch": 400000,
            "disp_file": "lcurve.out",
            "disp_freq": 1000,
            "numb_test": 4,
            "save_freq": 5000,
            "save_ckpt": "model.ckpt",
            "disp_training": true,
            "time_training": true
        }
    },
    "model_devi_dt": 0.001,
    "model_devi_skip": 0,
    "model_devi_f_trust_lo": 0.15,
    "model_devi_f_trust_hi": 0.35,
    "model_devi_clean_traj": false,
    "model_devi_jobs": [
        {"sys_idx": [0,1,2], "temps": [600],      "trj_freq": 10, "nsteps": 1000, "ensemble": "nvt", "_idx": "00"},
        {"sys_idx": [0,1,2], "temps": [600],      "trj_freq": 10, "nsteps": 2000, "ensemble": "nvt", "_idx": "01"},
        {"sys_idx": [0,1,2], "temps": [600, 800], "trj_freq": 10, "nsteps": 3000, "ensemble": "nvt", "_idx": "02"},
        {"sys_idx": [0,1,2], "temps": [600, 800], "trj_freq": 10, "nsteps": 5000, "ensemble": "nvt", "_idx": "03"}
    ],
    "fp_style": "vasp",
    "shuffle_poscar": false,
    "fp_task_max": 30,
    "fp_task_min": 5,
    "fp_pp_path": "./",
    "fp_pp_files": ["POTCAR_O", "POTCAR_Fe", "POTCAR_La", "POTCAR_Sr"],
    "fp_incar": "./INCAR_LSFO"
}

Summary of key changes from the methane example:

Parameter	Methane (old)	LSFO slab (new)	Why
type_map	["H","C"]	["O","Fe","La","Sr"]	Match your elements
sel	[16,4]	"auto"	Essential for 4-element system [1]
rcut	5.0	7.0	Oxides need larger cutoff [2]
neuron (descriptor)	[10,20,40]	[25,50,100]	More complex chemistry
neuron (fitting)	[120,120,120]	[240,240,240]	4 elements need more capacity
stop_batch	20000	400000	Complex oxide needs much more training [3]
decay_steps	5000	5000	Keep the same
f_trust_lo/hi	0.05/0.15	0.15/0.35	Start relaxed, tighten later [1]
ensemble	nvt	nvt	Must use NVT for slabs — NPT would collapse vacuum [4]
temps	[100]	[600, 800]	Ammonia oxidation operating temperatures
model_devi_dt	0.002	0.001	Heavier atoms (La=139) need smaller timestep
ENCUT	400	550	Oxygen pseudopotential requires higher cutoff
SIGMA	0.20	0.05	LSFO is an insulator/semiconductor

machine.json — update the remote_root paths from the methane directory to your LSFO directory:

"remote_root": "/home/xgduan/gdata/DPGEN_examlpe/LSFO_surface/run_node11/work"

Everything else in machine.json (batch_type, SSH config, etc.) can stay the same. Also make sure fp_pp_files order matches type_map — O first, then Fe, La, Sr.

Important: Adjust the init_data_sys and sys_configs paths to match whatever directory names dpgen actually created during your init step.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[njzjz-bot]

The material-specific settings in Dosu's latest answer should not be treated as a validated LSFO input. Values such as ENCUT=550, rcut=7.0, SIGMA=0.05, 0.15/0.35 deviation thresholds, network sizes, and a 1 fs timestep cannot be derived reliably from the composition alone.

Several assumptions are especially unsafe:

• La₁₋ₓSrₓFeO₃ is a magnetic transition-metal oxide whose metallicity, Fe spin states, oxygen-vacancy chemistry, and possible DFT+U treatment depend on composition and the chosen reference method. Do not assume it is simply an insulator.
• ENCUT must be converged for the actual POTCAR set; k-points, smearing, spin/MAGMOM, U values, slab dipole treatment, and electronic convergence must also be tested and then kept consistent across every label.
• sel: "auto" is a reasonable starting mechanism, but rcut and network capacity require convergence against held-out errors and target properties.
• The MD timestep is controlled by the fastest modes and stability, not by La being heavy. An ammonia-oxidation model introduces H and N and may require a smaller timestep and training coverage for adsorbates, bonds breaking/forming, surfaces, defects, and gas-phase species.
• Isotropic NPT would collapse a slab's vacuum, but NVT is not the only possible slab ensemble; an explicit template can control only in-plane cell dimensions when physically appropriate.
• machine.json controls execution and resources, not model accuracy.

For an ammonia-oxidation project, the eventual type_map, mass_map, pseudopotentials, and datasets must include N and H as well as the substrate elements. Establish a converged, scientifically justified DFT protocol first; then generate balanced bulk, slab, defect, adsorbate, and reaction configurations, calibrate model-deviation thresholds against actual DFT errors, and validate energies, forces, stresses, and reaction energetics on independent structures. There is no reliable drop-in param.json that can replace those convergence tests.

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh