Deprecation of `metadata$model_params` in 0.9.0.9000 silently breaks brms

[anirban-mu]

Describe the bug

cmdstanr 0.9.0.9000 deprecated metadata$model_params (now NULL). brms' read_csv_as_stanfit() depends on this field to compute parameter dimensions for the internal stanfit object. With model_params = NULL, brms silently drops population-level coefficients — verified to lose 1–3 parameters in models with 3–8 predictors. No warnings or errors are raised.

CmdStan and cmdstanr are working correctly — all draws are present in the CSV output and in cmdstanr's draws objects. The issue is that the deprecation removed a field that brms relies on, without a compatibility shim or deprecation warning to downstream packages.

Per the NEWS.md entry for 0.9.0.9000, the deprecated field is:

model_params element of fit$metadata() list (variables element)

The replacement fields metadata$stan_variables and metadata$stan_variable_sizes are present and correct, but brms does not yet use them for dimension computation.

To Reproduce

library(brms)

set.seed(42)
n <- 200
dat <- data.frame(
  y  = rnorm(n),
  x1 = rnorm(n), x2 = rnorm(n), x3 = rnorm(n)
)

fit <- brm(y ~ x1 + x2 + x3,
           data = dat, family = gaussian(),
           chains = 1, iter = 500,
           backend = "cmdstanr")
nrow(fixef(fit))
# Returns 3, expected 4 (Intercept is missing)

The mechanism: brms greps csfit$metadata$model_params to determine vector dimensions. When NULL, grepl(pattern, NULL) returns logical(0), dimensions are never assigned, and rename_pars() silently drops parameters.

Expected behavior

fixef(fit) should return 4 rows (Intercept + x1 + x2 + x3), matching the backend = "rstan" result. Either:

1. model_params should remain populated alongside the new fields during the deprecation period, or
2. brms should be notified of the breaking change so it can migrate to stan_variable_sizes

Operating system

macOS (Apple Silicon, Darwin 25.3.0)

CmdStanR version number

cmdstanr 0.9.0.9000 (dev, SHA a8adbb1), CmdStan 2.38.0

Additional context

• Filed companion issue on brms: rename_pars() silently drops population-level coefficients with cmdstanr 0.9.0.9000 when metadata$model_params is NULL paul-buerkner/brms#1864
• brms version: 2.23.1
• R version: 4.5.3
• The bug affects any brms user on cmdstanr dev (0.9.0.9000) with 3+ population-level predictors
• Models with 1–2 non-intercept predictors have the latent bug (incorrect dims_oi) but no visible parameter loss
• The set of dropped parameters is deterministic for a given formula and scales with model size (1 lost at K=3, 3 lost at K=8)

[jgabry]

Thanks for reporting this! We can keep model_params in cmdstanr for now until there's a solution in brms. I just opened a PR: #1152