run-simulation

name	run-simulation
description	Execute a geopolitical-forecasting simulation in a workspace previously scaffolded by `scaffold-simulation`. Use when the user wants to "run the simulation", "kick off the forecast", "resume a forecast", or "run geopol". Reads the workspace's .template-id sentinel and dispatches to the correct upstream run command, supporting resume of interrupted runs.
allowed-tools	Bash, Read, Glob

Run a simulation

Dispatch-by-sentinel: the same skill works across templates because it reads .template-id and looks up the run command in the registry.

Inputs to gather

1. Workspace path — defaults to the current working directory if it contains a .template-id file; otherwise ask.
2. Mode — fresh run (default) or resume. Resume reads as --resume-latest or --resume <run-id>; both are supported by the upstream templates.
3. Question / scenario — for fresh runs, the forecast question. If scenario.md exists in the workspace, default to the question recorded there.
4. Optional overrides — horizons, grounding tools, council-member models, etc. Forward verbatim to the upstream command. Don't try to validate them — let the template's CLI complain.

Steps

1. Locate sentinel. Read .template-id at the workspace root. If missing, stop and tell the user the workspace doesn't look scaffolded — they should run scaffold-simulation first.
2. Look up template. Find the matching entry in the plugin's templates/registry.json by id. If not found, stop — the registry may have been edited. List the known IDs.
3. Build the command. Use run_command for fresh, resume_command for resume. Append the user's question (quoted) for fresh runs. Append any overrides the user specified.
4. Execute in the workspace dir. Stream output. The upstream templates write checkpoints between stages; if the run fails partway, the user can re-run with resume mode and it'll skip completed stages.
5. On completion, locate the latest run dir under <workspace>/<reports_dir>/ (the registry says where) and report:
   • Run timestamp / id
   • Path to report.pdf (or template-specific equivalent)
   • Suggest next steps: bundle-run for a shareable artifact, grade-run later when horizons elapse.

Notes

• Don't second-guess the upstream CLI. Forward arguments verbatim.
• If resume is requested but no checkpoints exist, fall through to a fresh run and tell the user.
• Long-running. Consider run_in_background: true for the run command so the user isn't blocked, then poll for completion via the reports dir.