Browse interactively

mcmc browse opens an interactive browser over the project: the runs recorded in the store and the model files on disk. Typing mcmc on its own does the same thing when a terminal is attached; piped or redirected, bare mcmc still prints the command help, so scripts are unaffected.

mcmc browse

The picker

The box lists one scope at a time and filters as you type.

┌─  Runs 9   Models 2  ──────────────────────────────────────────────┐
│ ▸ gibbs                                                            │
├────────────────────────────────────────────────────────────────────┤
│ ❯ @2   model.jl      Gibbs 300x2              not converged · 5h ago│
│   @9   model.jl      Gibbs 300x2              not converged · 5h ago│
└ left/right scope · up/down pick · enter open · esc quit ───────────┘

Every word you type has to appear somewhere in the row, in any order: a run matches on its ref, id, model path, sampler, backend, and verdict, so gibbs failed finds failed Gibbs runs. Left and right switch between Runs and Models, up and down move, enter opens, and escape leaves.

Verdicts are colored: green for converged, red for a failed or unconverged run, yellow for a cancelled one.

What you can do with a run

Opening a run gives a menu that reuses the same code as the standalone commands, so what you see matches mcmc summary, mcmc diagnose, and friends:

• Summary and Diagnostics: the posterior table and the convergence table.
• Variables: every variable with its mean, sd, R-hat, ESS, and a sparkline of chain 1, so a stuck chain shows up as a flat line. Picking one draws its trace and density.
• Plots: any of the plot kinds that need only the draws, rendered in the terminal.
• Settings and artifacts: the same detail mcmc show prints.
• Report link: the report web-app deep link for the run.
• Run again: refits from the run’s frozen spec, so the settings are exactly the ones that produced it.
• Delete: removes the run from the ledger and deletes its directory, after a confirmation.

Escape backs out one level rather than quitting, so exploring is cheap.

What you can do with a model

The Models scope lists the files that actually declare a model, not every source file: a .jl file has to define a @model (or a build_model adapting one defined elsewhere), and a .stan file has to have a model or parameters block. Comments and docstrings do not count, so a sampler documented with a @model example stays out of the list. Each row shows the number of runs recorded from that file. From there you can run the model, read the file, or jump to just that model’s runs.

A file that declares a model but no entry function is listed as needs a build_model, because a fit calls that function with the data table. Opening it offers to add the missing line, showing it first:

build_model(data) = eight_schools(data.J, data.y, data.sigma)

The mapping from arguments to data columns is a guess from the model’s signature, so check it before running; mcmc run prints the same suggestion when a fit fails for want of an entry function.

Launching a run from the browser runs the real mcmc run, streaming progress the same way, and the new run appears in the list when it finishes. In a directory with no store yet, the first run creates one beside the model.

The same picker stands in for a missing argument elsewhere: mcmc run with no input offers these models instead of failing. Piped or redirected, it fails with the usual message, so scripts still get an error rather than a prompt.