Libreyolo Add CLI Command

Add or extend a LibreYOLO CLI command

The CLI is Typer with one custom layer: KeyValueCommand (libreyolo/cli/parsing.py) rewrites YOLO-style key=value tokens into --key value before Click parses, so every command accepts both grammars with full Typer type-checking, help, and completion intact. Understand that layer before touching anything; most CLI bugs live at its edges.

The map

• libreyolo/cli/__init__.py entrypoint(): registers everything. Special commands (version checks models formats cfg info metadata) come from commands/special.py; core modes (predict train val export ui monitor label doctor) are one module each under commands/; profile is a Typer sub-app (command group) - use that pattern for any new multi-subcommand tool.
• commands/command_utils.py, output.py, errors.py: shared option handling, machine-readable output, and error rendering. Reuse; don't reimplement --json printing or error exits per command.
• aliases.py: mode-aware translation of user-facing shorthand to internal config names (e.g. train mosaic to mosaic_prob, val conf to conf_thres). The single source of truth for name translation; never inline a rename in a command function.
• config.py: CLI-side config resolution.

Adding a flag to an existing command

1. Add it as a typed Typer parameter in the command function, named per the ecosystem convention (libreyolo-api-conventions; check the standard's documented name first).
2. If users know it by a different name than the internal config field, add the mapping to the right *_ALIASES dict in aliases.py.
3. Bool flags need a decision: negatable (--flag/--no-flag) or one-way (--json, --quiet)? key=false on a one-way flag must drop the flag, not emit a nonexistent --no-<flag>; the negatable set in rewrite_known_bool_flags encodes this and getting it wrong is a known bug class (see the comments in cli/__init__.py around the logging flags). Register new bool flags accordingly.
4. Mirror it in the Python API in the same PR (CLI and Python land together), and thread it through, verifying the value actually reaches the code that uses it; an accepted-and-ignored flag is a blocking review finding.
5. Defaults come from the loaded model/family where they are family-shaped (imgsz, thresholds): read them off the model, don't hardcode (REVIEW.md: "CLI defaults are family-derived").

Adding a new command

1. New module libreyolo/cli/commands/<name>.py exposing <name>_cmd (or a <name>_app Typer sub-app for command groups).
2. Register in entrypoint() with cls=KeyValueCommand (or app.add_typer for groups). Without KeyValueCommand the command silently loses key=value support, which users will file as a bug.
3. Support the output contract: --json (machine-readable to stdout, nothing else on stdout), --quiet, and correct exit codes via the shared error helpers. Agents and scripts consume these; they are part of why the CLI is self-describing.
4. Keep imports inside the command function or module lazy where heavy (torch etc.): libreyolo --help must stay fast, and the entrypoint's deferred imports exist for that reason.
5. Docs: the command shows up in libreyolo --help automatically; add it to the website docs command reference when user-facing.

Testing (the part that actually gates)

CLI tests are unit tests (tests/unit/cli/ for parsing/aliases/output/ errors, tests/unit/cli/commands/test_<cmd>.py per command) and run in the PR gate, so they must be hermetic: mock the model layer, use tmp_path, never download. Cover, at minimum:

• both grammars for the new surface (key=value and --key value),
• bool-flag behavior including key=false on one-way flags,
• alias translation if you touched aliases.py,
• --json output shape (assert keys, not prose),
• the failure path (bad input produces the intended error, not a traceback).

--help-json output comes from the declared Typer parameters, so it stays correct if you declared the flag properly and wrong if you hand-parsed argv; that is the test for whether you wired it right.

Traps

• Hand-parsing sys.argv in a command. Everything goes through Typer parameters + KeyValueCommand, or grammar support fractures.
• Renaming inside the command function instead of aliases.py.
• Printing progress to stdout (breaks --json consumers); human chatter goes to stderr / logging, respecting --quiet.
• A new command that imports torch at module top (slows every libreyolo invocation, including --help).
• Adding a track / task-prefixed verb without checking _strip_task_prefix in cli/__init__.py for how task-prefixed argv is normalized.

Related

• skills/libreyolo-api-conventions/: what to name it and whether it should exist on both surfaces (yes).
• skills/libreyolo-run-unit-tests/: keeping the new tests PR-gate-safe.
• tests/unit/cli/: the exemplars to copy.