@tuned-tensor/cli 0.14.0 → 0.15.1

package/docs/local-runtime/architecture.md

@@ -1,8 +1,12 @@
 # Architecture
 
-This repository is the local CUDA runtime behind the unified `tt`
-workflow. The standalone `tt-local` product surface is deprecated.
-Its supported runtime boundary is text SFT of a registry-certified
+This document describes the local CUDA runtime bundled with the unified `tt`
+agent and CLI. The standalone `tt-local` product surface is deprecated. Local
+workflow commands need no TT token. Managed agent inference reuses a TT access
+token; BYO inference uses the selected provider's credentials. Cloud operations
+are explicit through `tt cloud` and use the TT account token.
+
+The local adapter runtime boundary is text SFT of a registry-certified
 Transformers causal LM into a LoRA adapter on CUDA. Certified checkpoints
 are `Qwen/Qwen3.5-2B` (pinned to Hugging Face snapshot
 `15852e8c16360a2fea060d615a32b45270f8a8fc`),
@@ -28,7 +32,7 @@ implementation details, not a public plugin framework.
 
 ## Boundaries
 
-The first version intentionally excludes:
+The local adapter runtime excludes:
 
 - arbitrary training or evaluation commands;
 - DPO, continued pretraining, and multimodal training;
@@ -40,7 +44,7 @@ The first version intentionally excludes:
 The next model or method should arrive with its own locked dependencies,
 resource defaults, data contract, and real CUDA acceptance test.
 
-TT Local currently ships two such methods:
+TT bundles two local training methods:
 
 - **Adapter** (`training/adapter`): certified Transformers checkpoints and
   PEFT LoRA SFT. Driven by `tt pipeline run`.
@@ -90,7 +94,7 @@ The bundled Python project owns four small operations:
 - `prefetch.py`: download and verify the pinned Hugging Face snapshot;
 - `evaluate.py`: generate predictions without receiving reference answers;
 - `train.py`: perform CUDA-only, model-specific LoRA SFT;
-- `serve.py`: load the base model and verified adapter for inference.
+- `serve.py`: prepare the verified base model and adapter for the vLLM serving runtime.
 
 Training uses assistant-only loss. Prompt tokens have label `-100`, and
 truncation removes older prompt tokens before it removes any answer token.
@@ -126,7 +130,7 @@ hide a valid artifact. Serving verifies the manifest again.
 The runner config controls only:
 
 - state, artifact, and Hugging Face cache paths;
-- CUDA or CPU inference for evaluation and serving;
+- CUDA or CPU evaluation, with local serving through the CUDA vLLM runtime;
 - deterministic generation and scoring limits.
 
 The training project, Python entrypoints, working directory, and child

package/docs/local-runtime/serving.md

@@ -2,9 +2,10 @@
 
 `tt serve` verifies TT artifacts and launches **vLLM 0.28.0**. vLLM owns model
 execution, batching, KV/prefix caching, streaming, and native tool-call parsing.
-There is no TT inference loop, HTTP server, or proxy between the harness and
-vLLM. The serving dependencies are locked separately from training, in an
-external cache environment; an npm upgrade does not mutate the training runtime.
+This local serving path connects the harness directly to vLLM; TT's managed
+agent proxy is a separate inference option. The serving dependencies are locked
+separately from training, in an external cache environment; an npm upgrade does
+not mutate the training runtime.
 
 ## Requirements and scope
 
@@ -77,7 +78,9 @@ tt serve local-<run-id> --config local-runner.json \
 
 Use the same launch options as the running server. Exporting config verifies the
 TT target but does not start a server. Adapter exports include both model IDs.
-For example, in a disposable coding-fixture directory:
+The exported `tt-local` provider ID names this local endpoint; it is not a
+separate CLI or package and uses no TT account token. For example, in a
+disposable coding-fixture directory:
 
 ```bash
 pi --provider tt-local --model 'base:Qwen/Qwen3.5-2B' \
@@ -107,7 +110,7 @@ OpenCode can connect directly using its existing custom provider support:
   "provider": {
     "tt-local": {
       "npm": "@ai-sdk/openai-compatible",
-      "name": "TT local",
+      "name": "TT local endpoint",
       "options": { "baseURL": "http://127.0.0.1:8000/v1" },
       "models": {
         "base:Qwen/Qwen3.5-2B": {

package/docs/local-runtime/spark.md

@@ -26,8 +26,10 @@ this GPU. Spark unified memory should show Nemotron and Muse Glimmer as
 trainable; consumer cards typically only fit Qwen LoRA. `tt status` and the
 shell agent reuse the cached report.
 
-TT Local requires Node 22+, `uv`, working CUDA PyTorch, and enough free space
-for the Hugging Face cache plus run artifacts.
+Local TT training requires Node 22.19+, `uv`, working CUDA PyTorch, and enough free space
+for the Hugging Face cache plus run artifacts. Local commands need no TT
+access token. Agent inference can use your TT token through the managed proxy,
+or your own provider credentials and model.
 
 ## Create a project
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tuned-tensor/cli",
-  "version": "0.14.0",
+  "version": "0.15.1",
   "description": "Local TT agent and CUDA fine-tuning CLI for Tuned Tensor workflows",
   "repository": {
     "type": "git",
@@ -13,12 +13,15 @@
   "scripts": {
     "build": "tsup",
     "dev": "tsup --watch",
-    "typecheck": "tsc --noEmit",
+    "typecheck": "tsc --noEmit && tsc --noEmit -p tsconfig.scripts.json",
     "test": "npm run test:cli && npm run test:local-runtime && npm run test:python",
     "test:cli": "vitest run",
+    "test:workflows": "vitest run src/__tests__/workflows",
+    "eval:agent": "tsx scripts/eval-agent.ts",
     "test:local-runtime": "node --import tsx --test \"test/local-runtime/**/*.test.ts\"",
     "test:python": "python3 -m unittest discover -s training/adapter/tests -p 'test_*.py' && uv run --frozen --project training/foundation python -m unittest discover -s training/foundation/tests -p 'test_*.py'",
-    "test:watch": "vitest"
+    "test:watch": "vitest",
+    "check": "npm run typecheck && npm test && npm run build"
   },
   "files": [
     "dist",

package/training/adapter/src/model_contract.py

@@ -110,7 +110,7 @@ def assert_certified_model_config(
             )
         return
     else:
-        raise ValueError(f"{label} is not a certified TT Local base-model architecture")
+        raise ValueError(f"{label} is not a certified TT base-model architecture")
 
     for name, expected in certified.items():
         actual = _field(text_config, name)

package/training/adapter/src/prefetch.py

@@ -147,7 +147,7 @@ def verify_snapshot(snapshot: Path, expected_model_id: str | None = None) -> tup
 
 
 def main() -> None:
-    parser = argparse.ArgumentParser(description="Prefetch a TT Local Hugging Face base model.")
+    parser = argparse.ArgumentParser(description="Prefetch a TT Hugging Face base model for local use.")
     parser.add_argument("--input", required=True)
     parser.add_argument("--output", required=True)
     args = parser.parse_args()

package/training/adapter/src/train.py

@@ -92,8 +92,8 @@ def assert_certified_request() -> None:
 def require_cuda() -> dict[str, Any]:
     if not torch.cuda.is_available():
         raise RuntimeError(
-            "TT Local bundled training requires an NVIDIA CUDA GPU. "
-            "Run this workflow on DGX Spark and verify it first with tt-local doctor."
+            "TT local training requires an NVIDIA CUDA GPU. "
+            "Run this workflow on a supported CUDA host and verify it first with tt doctor."
         )
     device_index = torch.cuda.current_device()
     properties = torch.cuda.get_device_properties(device_index)