Set up a cobaya run

Set up a cobaya run

Set up and execute a cobaya MCMC run named $0. The skill assembles a YAML from four independent decisions; treat them separately and ask the user when underspecified — don't pre-bin into preset "styles".

The four decisions:

1. Likelihood(s) — Planck / ACT / DES / a tSZ Cl^yy bandpower file / joint / etc.
2. Theory code — camb (CAMB), classy (CLASS), or classy_szfast.classy_sz.classy_sz (class_sz; itself a full Boltzmann code with CosmoPower emulators in fast mode, plus halo-model observables). Pick what supports the requested observables.
3. What's sampled vs fixed — any subset of cosmology, astro/HOD/profile, nuisances, foregrounds. Cosmology can be fully sampled, fully fixed in extra_args (e.g. with use_class_sz_no_cosmo_mode: 1 for class_sz), or partially constrained. Same for astro/nuisances.
4. Output — chain location, debug, timing, MPI, Rminus1_stop target, starting covmat.

Hints in $ARGUMENTS after $0 can shortcut any of these (e.g. planck_2018_lowl.TT,planck_2018_lowl.EE → likelihood set; --theory=class_sz → theory choice; --data-dir=… → bandpower workdir). Ask for anything left underspecified before writing.

Steps

1. Resolve the four decisions. From $ARGUMENTS, existing files in the workdir (e.g. an existing <run-name>.yaml to amend, a data/ dir, a reference-*/ baseline), and the user.

2. Sketch the YAML with the four standard blocks:

   • theory: { <code>: { stop_at_error: True, [extra_args: { ... }] } }
   • likelihood: { <name1>: null, <name2>: { ... } }
   • params: { ... fixed values, sampled priors, derived ... }
   • sampler: { mcmc: { Rminus1_stop: 0.05, max_tries: 1000, learn_proposal: True, [covmat: ...] } }
   • output: chains/<run-name> (or absolute path inside a workdir)
   • debug: True, timing: true

3. Install external packages if any are referenced (e.g. Planck likelihoods): cobaya-install -y -p packages/ <run-name>.yaml. Skip for runs that only use in-venv code (class_sz, classy_szfast, a standalone Likelihood module).

4. Validate with --test before launching: cobaya-run --test <run-name>.yaml. Reports the model assembly, initial point, per-component speed measurements; takes seconds.

5. Launch:

   • Single chain: cobaya-run <run-name>.yaml
   • MPI: mpirun -np N cobaya-run <run-name>.yaml (N parallel chains; bump for more cores)

6. Report back:

   • YAML path written, --test result
   • Chain location, max |R-1| reached, total samples, n_chains, acceptance rate from .progress
   • Any theory crashes (Traceback/error in stderr or <run-name>.log)
   • One-line suggested next step (tighten Rminus1_stop, switch to MPI, add a likelihood, etc.)

Cross-plugin handoffs

• For a tSZ Cl^yy power-spectrum bandpower likelihood (binned bandpowers + N×N covariance, fit via class_sz), the class-sz plugin's /class-sz:build-likelihood is the right entry point if loaded — it scaffolds a standalone likelihood module + a working YAML in one shot.
• For deeper class_sz / classy_szfast questions, defer to /class-sz:explain.

Workdir convention (when external module/data are involved)

<workdir>/
├── <likelihood-module>.py         # standalone Likelihood/Theory (optional)
├── <run-name>.yaml
├── data/                          # whatever the likelihood reads
└── chains/                        # cobaya output

cd <workdir> before running cobaya-run so any local module is importable.

Constraints

• Do not pip install cobaya or its deps — assume the venv is complete.
• Do not silently swallow theory errors; report them.
• Do not overwrite an existing chain — use resume: True or pick a new name.
• Quote ${VARS} and any path with spaces.
• Don't run from a directory that has a cobaya/ subdir (PEP 420 cwd collision — see /cobaya:explain Pitfalls).