nas: 4 new CLI commands + --group on model list (G4)

[probicheaux]

Description

Wraps the new public API routes from the agentic-surface-area onsite plan as CLI commands. Companion backend PRs (all live on staging):

• roboflow#11603 — G1, validator
• roboflow#11605 — G6, projection + ?group=
• roboflow#11610 — G2, public train cancel/stop + favorite
• roboflow#11612 — G3, training results
• roboflow-mcp#34 — sister G5 (MCP)

New CLI commands

roboflow train cancel <project>/<version> [--continue-if-no-refund]
roboflow train stop   <project>/<version>
roboflow train results <project>/<version>
roboflow model star   <model-id> [--unstar]

Plus roboflow model list -p <project> -g/--group <modelGroup> — the canonical "list NAS models per run" path. When --group is set, the list command hits the public /models endpoint with the enriched projection (hardware/latency/map5095/paretoOptimalFor/recommended) instead of walking versions via the SDK.

Adapter additions

In roboflow/adapters/rfapi.py:

• cancel_version_training, stop_version_training
• get_training_results
• list_project_models (with optional group=), get_model_by_url
• favorite_nas_model

How tested

Unit tests

+13 cases across test_train_handler.py and test_model_handler.py covering register, happy paths, the 409/CANNOT_CANCEL hint surfacing, the MODEL_NOT_NAS hint, the unstar flow, and the --group endpoint switch. 298/298 CLI tests pass locally; ruff check + ruff format clean.

CLI-COMMANDS.md got two new sections — Train, monitor, cancel, stop and NAS models — list, star, deploy.

E2E verified live on staging

Driven against api.roboflow.one on peter-robicheaux/beer-can-hackathon with API_URL=https://api.roboflow.one:

train results on a NAS parent (v410, 52 children):

$ roboflow --json train results peter-robicheaux/beer-can-hackathon/410 | jq '{jobType, modelGroup, modelCount, recommendedByHardware}'
{
  "jobType": "nas",
  "modelGroup": "pVYKOWUB6AUIVJMgPc7u-410-rfdetrNasGroup",
  "modelCount": 52,
  "recommendedByHardware": {"gpu": "C6KmrxA6h85kMmamGnsS"}
}

model list --group rendered a 53-row leaderboard:

$ roboflow model list -p peter-robicheaux/beer-can-hackathon -g pVYKOWUB6AUIVJMgPc7u-410-rfdetrNasGroup
URL                                                 TYPE        HARDWARE  LATENCY  MAP50  MAP5095  REC
peter-robicheaux/beer-can-hackathon-410-nas-gpu-…  rfdetr-nas                      100
peter-robicheaux/beer-can-hackathon-410-nas-gpu-b  rfdetr-nas  gpu       4.36     100    83.97
peter-robicheaux/beer-can-hackathon-410-nas-gpu-…  rfdetr-nas  gpu       1.09     100    62.57    ★
... 50 more rows

model star on a real NAS child:

$ roboflow --json -w peter-robicheaux model star 14CwSGmGetWh6rB0EnjL
{"success": true, "model": {"id": "14CwSGmGetWh6rB0EnjL", "favorites": {...}}}
$ roboflow --json -w peter-robicheaux model star 14CwSGmGetWh6rB0EnjL --unstar
{"success": true, ...}

train cancel on a finished version (text mode shows the actionable hint):

$ roboflow train cancel peter-robicheaux/beer-can-hackathon/318
Error: Cannot cancel non-running train job.
  Hint: Cancel only applies to in-flight runs. Check status with
        'roboflow train results <project>/<version>'.

E2E exposed one error-shape mismatch on the backend (flat {error: "Conflict", message: ...} lost the descriptive message through the CLI's parser); fixed in roboflow#11610 (nested {error: {message, code, type}}).

🤖 Generated with Claude Code

[digaobarbosa]

I don't thin we should talk about firestore here. It shows to the user.

[digaobarbosa]

[Medium] train cancel may report success even when the server did not cancel

• File: roboflow/cli/handlers/train.py:338-342 (refund semantics documented at L92-95).
• _cancel always emits status: "cancelled" and Training cancelled ... after any 2xx, but the documented default behavior is for the server to reply refund:false without cancelling unless --continue-if-no-refund is passed.
• Why it matters: agents/scripts will treat the CLI result as confirmation that an in-flight paid run was cancelled when the server may have only returned a refund-window check. Behavioral correctness issue on a destructive command.
• Fix: derive status from the payload — if cancelled:false or refund:false, output a distinct status: "not_cancelled" and surface a hint to rerun with --continue-if-no-refund. Add a unit test for that exact response.

[digaobarbosa]

Claude-only — [Medium] get_model_by_url is dead code

• File: roboflow/adapters/rfapi.py:154-161.
• No CLI command calls it. Either remove or comment as scaffolding for a follow-on PR.

[digaobarbosa]

Can delete this

[probicheaux]

✅ E2E verified against api.roboflow.one (backend roboflow#11646 deployed to staging via light-v2-api).

Branch installed via pip install -e . into a clean venv; commands run with API_URL=https://api.roboflow.one.

roboflow --json train results peter-robicheaux/beer-can-hackathon/410 (52-model NAS run):

{
  "trainingId": "peter-robicheaux/beer-can-hackathon/410",
  "versionId": "410",
  "status": "finished",
  "jobType": "nas",
  "modelGroup": "pVYKOWUB6AUIVJMgPc7u-410-rfdetrNasGroup",
  "modelCount": 52,
  "recommendedByHardware": { "gpu": "beer-can-hackathon-410-nas-gpu-ec8a0e" },
  "models": [{ "modelUrl": "beer-can-hackathon-410-nas-gpu-066866", "modelType": "rfdetr-nas", "metrics": {...} }]
}

trainingId is the slug form, no datasetId, recommendedByHardware value is a URL slug, every models[] row has modelUrl only.

roboflow --json -w peter-robicheaux model star beer-can-hackathon-410-nas-gpu-ec8a0e ✅

• Star: 200, response body is the public projection only — no id, dataset, owner, or projects[].
• Unstar (--unstar): 200, favorites[<uid>] = false. Round-trip clean.

roboflow --json -w peter-robicheaux model star bogus-not-a-real-slug ✅

• Output: {"error": {"type": "NotFound", "message": "Model not found in this workspace", "code": "MODEL_NOT_IN_WORKSPACE", "hint": "Verify the model URL and workspace. The slug is the same value 'roboflow train results' returns as models[].modelUrl."}}
• Exit code: 3 (not-found).