Skip to content

cli.md

Latest commit

 

History

History
121 lines (79 loc) · 5.66 KB

cli.md

File metadata and controls

121 lines (79 loc) · 5.66 KB

CLI reference

The lewron command is the console entry point. Every invocation is a subcommand:

lewron <command> [arguments] [flags]

Run lewron --help or lewron <command> --help for the built-in help.

All commands return exit code 0 on success and a nonzero code on failure; errors are rendered as a clean panel rather than a Python traceback. Set LEWRON_DEBUG=1 in the environment to also print the full traceback when debugging.

Model-selection flags

These flags are shared by run, stage, revise, and report. They are optional; both default to Sonnet.

Flag Default Applies to
--reasoning-model <id> claude-sonnet-4-6 auditing, physics reasoning, the report writer, bootstrap
--coding-model <id> claude-sonnet-4-6 code generation
--thinking-budget <int> unset extended-thinking token budget for reasoning calls

The selected models are recorded in the run's state.json, so a finished run documents which models produced it. See Python API → ModelConfig for the equivalent in code.

User commands

run — run the pipeline (and the Explorer)

lewron run --input <file-or-prompt> [--run-dir <dir>] [--allow-bypass] [model flags]
lewron run --run-dir <dir>            # resume an existing run

Runs the full toolbox pipeline and, when the build includes thermal analysis, flows straight into the Explorer. The build depth is inferred from your request — there is no build-mode flag.

Argument / flag Required Meaning
--input to start a run Path to an input file, or an inline prompt string. Omit it to resume.
--run-dir to resume Output directory. When starting a run it defaults to a directory named after the model. When resuming, it must already contain state.json.
--allow-bypass no Enables the Explorer's /bypass mode (skips all permission checks). Off by default; the Explorer's /edit auto-edit mode works without it.

To resume, pass --run-dir with no --input; LeWRON continues from the last completed stage, or re-enters the Explorer if the toolbox is already built.

revise — re-derive a stage with feedback

lewron revise <stage> "<feedback>" --run-dir <dir> [--run] [model flags]

Resets <stage> and all downstream stages, attaches your feedback, and re-derives. Stages that can be revised: 1a, 1b, 2a, 2b. (Bootstrap and the finite-T stage are not revised this way.)

Argument / flag Required Meaning
<stage> yes Stage to revise.
<feedback> yes Free-form text guiding the revision, e.g. "Use the paper's Higgs-basis convention."
--run-dir yes Existing run directory.
--run no Immediately resume the pipeline after queuing the feedback. Without it, the feedback is queued and LeWRON prints the lewron run command to apply it later.

For the tree-level stage, a revision is scoped automatically to the part of the derivation your feedback targets plus whatever depends on it — earlier and independent work is preserved. See The interactive workflow.

reset — reset a stage and everything downstream

lewron reset <stage> --run-dir <dir> [--yes]

Resets <stage> and all downstream stages back to not-run, discarding their generated outputs. Use it to redo a stage from scratch (as opposed to revise, which re-derives with guidance). For finer-grained control, the Python API also exposes reset_substep, which resets a single Stage 1b substep and the substeps after it.

Argument / flag Required Meaning
<stage> yes Stage to reset (1a, 1b, 2a, 2b, 3). Downstream stages reset automatically.
--run-dir yes Existing run directory.
--yes, -y no Skip the confirmation prompt.

report — rewrite the report from current artifacts

lewron report --run-dir <dir> [--stage <stage>] [model flags]

Regenerates notes/report.md from the artifacts already in the run, without rerunning any stage. This is the recovery path when a report write was interrupted but the artifacts are intact.

Argument / flag Required Meaning
--run-dir yes Existing run directory.
--stage no Which stage to report on. Defaults to the last completed stage.

install-skill — install the physics priors as a Claude Code skill

lewron install-skill [--uninstall]

Packages your physics priors as a global lewron-physics Claude Code skill under ~/.claude/skills/, so a plain Claude Code session can draw on them (run /lewron-physics, or let Claude load them on its own when you work on a phase transition). It is built from your effective experience library, so re-run it to refresh after editing your priors. This command is not tied to a run directory.

Argument / flag Required Meaning
--uninstall no Remove the installed skill instead of installing it.

Advanced commands

stage — run a single toolbox stage

lewron stage <stage> --run-dir <dir> [model flags]

Runs one toolbox stage in an existing run that already has all prior stages completed. Stage keys: 1a, 1b, 2a, 2b, 3. A normal analysis uses run instead; stage is for re-running an individual stage during development.

Notes

When a command pauses for you, pressing Enter accepts the default and typing quit or exit stops cleanly. The exact keyword contract is described in The interactive workflow.