cli

glrs is short for glorious. both names run the same binary, so use whichever you prefer.

project root

everything resolves against the project root: file tools, config discovery, rules. it is git rev-parse --show-toplevel, or the directory you started in when that is not a repository.

glrs works outside a repository. the first run still writes config, and the file tools still resolve against where you started.

commands

command	does
`glrs`	open a session
`glrs -p <prompt>`	one headless turn, then exit
`glrs doctor`	report what would run, without running it
`glrs update`	reinstall the latest release
`glrs <word>`	a subcommand an extension registered

--version and --help win wherever they appear. everything after -p is the prompt, taken verbatim, so it may contain what looks like a flag. a subcommand before -p wins; -p before a bare word wins.

flags

flag	effect
`-p, --print <prompt>`	one headless turn; reads piped stdin too
`--model <provider/id>`	model for this run, written to `GLRS_MODEL` before anything reads it
`--resume [session-id]`	resume that session, or open the picker when given no id
`--version`, `-v`	print `glrs <version>`
`--help`, `-h`	print the usage, the commands, the options, and an `Added by extensions:` block

--model must have a value, that value must not be another flag, and it must name a provider:

glrs --model           → --model needs a value.
glrs --model -p hi     → --model needs a model id, and "-p" is another flag.
glrs --model bare-id   → --model bare-id: --model needs provider/model-id, and "bare-id" names no provider. There is no default provider. See `glrs doctor`.

--resume <id> naming no session is Session not found: <id>; an empty store is No sessions to resume.

a long flag glrs does not know is lifted out before parsing, lowercased, and handed to the extension that registered it. --foo=x and --foo bar carry a value; --foo --bar carries "". an unclaimed flag is the notice (unknown flag: --foo). interactive sessions only.

subcommands

a first bare word glrs does not claim goes to the extensions: subcommands.

print mode

assistant text goes to stdout. the tool trail, retries, extension notes, [config] diagnostics and [provider] warnings go to stderr. piped stdin joins the prompt, fenced as <input>…</input>. nothing reaches the session store.

exit	when
`0`	the turn finished
`1`	the turn threw, SIGINT interrupted it (`[interrupted]`), the 100-step limit was reached (`[stopped at the step limit without finishing]`), an argv error, an unknown subcommand, or `Nothing to run: -p needs a prompt or piped input.`

doctor

glrs doctor reports what would run without running any of it: extensions are resolved, not loaded.

model: anthropic/claude-opus-5
provider: Anthropic
missing: ANTHROPIC_API_KEY
extensions: builtins (bundled)

with credentials present the third line is credentials: found. with no model the first is model: not configured, followed by the reason and the built-in provider ids. --json emits the same report as an object: diagnostics, model, provider, missing, note, extensions.

cli