autoforge-trainer 1.0.2__tar.gz → 1.2.0__tar.gz

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/AGENTS.md

@@ -69,6 +69,7 @@ the CLI below offers the same operations with `--json`.
 | PACK_EXISTS | already installed | use it, or reinstall with `--force` |
 | PACK_INVALID | pack fails validation; `data.problems` lists why | do not use it |
 | SHA_MISMATCH | source changed after indexing | do not install; tell the user |
+| PACK_NEEDS_NEWER_AUTOFORGE | pack sets `min_autoforge` above this version | tell the user: pip install -U autoforge-trainer |
 | INDEX_UNAVAILABLE | index unreachable and no cache | retry later or set AUTOFORGE_INDEX_URL |
 | PACK_SOURCE_ERROR | bad source string or git failure | check the source |
 | PLAN_INVALID | plan fails checks; `data.problems` lists why | fix the plan, validate again |

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/CHANGELOG.md

@@ -1,5 +1,34 @@
 # 變更紀錄
 
+## 1.2.0 — 2026-10
+
+- **網頁介面的 AI 助手**：在 `autoforge gui` 裡和內建 AI 代理對話。可以選 AI 大腦（Claude、OpenAI、本機 Ollama），用一句話說需求；AI 開始訓練、重新訓練或自動調參前會在頁面上請你同意，不同意時可以寫意見讓它修改；訓練完可以接著問。AI 用自己的身分操作，隱私規則和指令列相同。
+- AI 代理改成可以多輪對話（`AgentSession`），`autoforge train` 的行為不變。
+- 修正幾個還寫著舊套件名稱 `autoforge[...]` 的安裝提示。
+
+## 1.1.0 — 2026-10
+
+### 新方法包（P7 第二批）
+- `object-detection`：物件偵測，torchvision Faster R-CNN 微調，支援 CSV 或 COCO 標註，評估 mAP@0.5。
+- `image-segmentation`：影像分割，LR-ASPP / DeepLabV3，評估 mIoU，可匯出 ONNX。
+- `embedding-finetune`：文字向量微調（搜尋、RAG），對比學習，報告微調前後的 recall@1，可匯出 ONNX。
+- `speech-recognition`：Whisper 語音辨識微調，報告微調前後的 WER / CER。
+- `llm-preference`：LLM 偏好對齊（DPO + LoRA / QLoRA），可從 `llm-finetune` 的結果接著訓練。
+- `diffusion-lora`：Stable Diffusion 1.x/2.x 的 UNet LoRA，生成的圖片在網頁介面直接顯示。
+- `reinforcement-learning`：stable-baselines3（PPO / A2C / DQN / SAC），環境可用 Gymnasium 名稱或自己寫的 `make_env()`。
+- `recommendation`：推薦系統，隱式回饋 ALS 和 Item-KNN，和熱門排行的基準比較。
+- `timeseries-forecast` 0.2.0：新增 `global_lag_gbm`，多條序列共用一個模型。
+
+### 核心
+- 安裝方法包時檢查 `min_autoforge`，版本太舊會回 `PACK_NEEDS_NEWER_AUTOFORGE`。
+- 模型輸入新增 `audio` 類型（REST、MCP、網頁介面）。
+- 自動調參能判斷 WER、CER、MAPE、WAPE 是越低越好。
+- `ctx.check_updates()` 只在控制檔變動時才讀檔，每一步呼叫也沒有負擔。
+- 新的安裝選項 `[rl]`；`[deep]` 加上 diffusers、soundfile。
+
+### 已知限制
+- 新的 GPU 方法（偵測、分割、Whisper、DPO、Diffusion LoRA）只用極小模型在 CPU 上驗證過流程，尚未在真的 NVIDIA GPU 上跑過。
+
 ## 1.0.2 — 2026-10
 
 - **PyPI 套件名稱改為 `autoforge-trainer`**：`autoforge-ml` 在我們發佈前被別人註冊了。指令和 `import autoforge` 都不變。1.0.1 沒有上傳到 PyPI。

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: autoforge-trainer
-Version: 1.0.2
+Version: 1.2.0
 Summary: AI that trains AI: analyse your data and hardware, pick a method, train, tune and deploy models on your own machine — from a CLI, MCP, a web GUI or one sentence.
 Project-URL: Homepage, https://github.com/Honguan/autoforge
 Project-URL: Documentation, https://honguan.github.io/autoforge/
@@ -42,6 +42,8 @@ Requires-Dist: uvicorn>=0.30
 Provides-Extra: all
 Requires-Dist: accelerate>=0.30; extra == 'all'
 Requires-Dist: anthropic<2,>=1.14; extra == 'all'
+Requires-Dist: diffusers>=0.30; extra == 'all'
+Requires-Dist: gymnasium>=0.29; extra == 'all'
 Requires-Dist: nvidia-ml-py>=12.0; extra == 'all'
 Requires-Dist: onnx>=1.16; extra == 'all'
 Requires-Dist: onnxruntime>=1.18; extra == 'all'
@@ -49,6 +51,8 @@ Requires-Dist: onnxscript>=0.1; extra == 'all'
 Requires-Dist: peft>=0.12; extra == 'all'
 Requires-Dist: pillow>=10.0; extra == 'all'
 Requires-Dist: skl2onnx>=1.17; extra == 'all'
+Requires-Dist: soundfile>=0.12; extra == 'all'
+Requires-Dist: stable-baselines3>=2.3; extra == 'all'
 Requires-Dist: timm>=1.0; extra == 'all'
 Requires-Dist: torch>=2.3; extra == 'all'
 Requires-Dist: torchvision>=0.18; extra == 'all'
@@ -57,9 +61,11 @@ Provides-Extra: anthropic
 Requires-Dist: anthropic<2,>=1.14; extra == 'anthropic'
 Provides-Extra: deep
 Requires-Dist: accelerate>=0.30; extra == 'deep'
+Requires-Dist: diffusers>=0.30; extra == 'deep'
 Requires-Dist: nvidia-ml-py>=12.0; extra == 'deep'
 Requires-Dist: peft>=0.12; extra == 'deep'
 Requires-Dist: pillow>=10.0; extra == 'deep'
+Requires-Dist: soundfile>=0.12; extra == 'deep'
 Requires-Dist: timm>=1.0; extra == 'deep'
 Requires-Dist: torch>=2.3; extra == 'deep'
 Requires-Dist: torchvision>=0.18; extra == 'deep'
@@ -78,6 +84,9 @@ Requires-Dist: onnx>=1.16; extra == 'onnx'
 Requires-Dist: onnxruntime>=1.18; extra == 'onnx'
 Requires-Dist: onnxscript>=0.1; extra == 'onnx'
 Requires-Dist: skl2onnx>=1.17; extra == 'onnx'
+Provides-Extra: rl
+Requires-Dist: gymnasium>=0.29; extra == 'rl'
+Requires-Dist: stable-baselines3>=2.3; extra == 'rl'
 Description-Content-Type: text/markdown
 
 # AutoForge
@@ -89,7 +98,7 @@ Description-Content-Type: text/markdown
 - **方法包**：從 GitHub 安裝的「教學包」，AI 讀了就知道怎麼訓練
 - **隱私**：敏感資料只允許本地模型指揮
 
-> **1.0.2** · 文件網站：<https://honguan.github.io/autoforge/> · [變更紀錄](CHANGELOG.md)
+> **1.2.0** · 文件網站：<https://honguan.github.io/autoforge/> · [變更紀錄](CHANGELOG.md)
 
 ## 安裝與第一次使用
 
@@ -102,8 +111,8 @@ autoforge gui                     # 網頁介面
 | 安裝方式 | 包含 |
 |---|---|
 | `pip install autoforge-trainer` | 表格、文字（TF-IDF）、時間序列、異常偵測、自動調參、網頁介面、MCP、REST |
-| `pip install "autoforge-trainer[deep]"` | 再加上 Transformer、LLM 微調（LoRA/QLoRA）、圖片分類（NVIDIA GPU） |
-| `pip install "autoforge-trainer[all]"` | 再加上 Claude 代理、ONNX 匯出 |
+| `pip install "autoforge-trainer[deep]"` | 再加上 Transformer、Embedding、LLM 微調與 DPO、圖片分類、物件偵測、影像分割、語音辨識、圖像生成 LoRA（NVIDIA GPU） |
+| `pip install "autoforge-trainer[all]"` | 再加上強化學習、Claude 代理、ONNX 匯出 |
 | `docker run --rm -p 127.0.0.1:8800:8800 -v autoforge:/data ghcr.io/honguan/autoforge:cpu` | CPU 映像，預設啟動網頁介面 |
 | `docker run --rm --gpus all -p 127.0.0.1:8800:8800 -v autoforge:/data ghcr.io/honguan/autoforge:cuda` | NVIDIA CUDA 映像，所有方法包 |
 
@@ -119,8 +128,16 @@ PyPI 上的 `autoforge` 是另一個不相關的專案（`autoforge-ai`、`autof
 | `text-classification` | 文字分類（情緒、意圖、主題、垃圾訊息） | TF-IDF + 線性模型；Transformer 微調 | CPU / GPU |
 | `llm-finetune` | 大型語言模型微調 | LoRA / QLoRA（可匯出合併後的權重） | GPU（8GB+） |
 | `image-classification` | 圖片分類 | timm 預訓練模型微調（可匯出 ONNX） | GPU 建議，CPU 可跑小模型 |
-| `timeseries-forecast` | 時間序列預測（銷量、用電、流量） | 落後特徵 + 梯度提升，和「明天＝今天」的基準比較 | CPU |
+| `timeseries-forecast` | 時間序列預測（銷量、用電、流量），單一或多條序列 | 落後特徵 + 梯度提升；多序列共用一個模型；和「明天＝今天」的基準比較 | CPU |
 | `anomaly-detection` | 異常偵測（詐欺、故障、離群值） | Isolation Forest | CPU |
+| `object-detection` | 物件偵測（框出物品、瑕疵、人車） | torchvision Faster R-CNN 微調，mAP@0.5 | GPU 建議，CPU 可跑小模型 |
+| `image-segmentation` | 影像分割（逐像素標類別） | torchvision LR-ASPP / DeepLabV3，mIoU | GPU 建議，CPU 可跑小模型 |
+| `embedding-finetune` | 文字向量微調（搜尋、RAG 檢索、相似度） | 對比學習（同批負樣本），和微調前比較 recall@1 | CPU / GPU |
+| `speech-recognition` | 語音辨識 | Whisper 微調，WER / CER，和微調前比較 | GPU 建議 |
+| `llm-preference` | LLM 偏好對齊（選好的回答、避開壞的） | DPO + LoRA / QLoRA | GPU（8GB+） |
+| `diffusion-lora` | 圖像生成風格 / 主體 LoRA | Stable Diffusion 1.x/2.x UNet LoRA | GPU（8GB+） |
+| `reinforcement-learning` | 強化學習（遊戲、控制、自己的模擬器） | stable-baselines3：PPO / A2C / DQN / SAC | CPU（小網路）/ GPU |
+| `recommendation` | 推薦系統（商品、內容） | 隱式回饋 ALS、Item-KNN，和熱門排行比較 | CPU |
 
 每個方法包都附給 AI 讀的 `KNOWLEDGE.md`，以及在 CPU 上 30 秒內跑完的測試。自己做方法包：`autoforge pack new`（見 [docs/PACK_SPEC.md](docs/PACK_SPEC.md)）。
 
@@ -128,7 +145,7 @@ PyPI 上的 `autoforge` 是另一個不相關的專案（`autoforge-ai`、`autof
 
 ```bash
 git clone https://github.com/Honguan/autoforge && cd autoforge
-pip install -e ".[dev]"            # 加上 ,deep 安裝深度學習方法包的依賴
+pip install -e ".[dev]"            # 加上 ,deep 安裝深度學習方法包的依賴，,rl 安裝強化學習的依賴
 pytest -q
 ```
 
@@ -136,6 +153,7 @@ pytest -q
 
 ```bash
 autoforge gui      # 打開瀏覽器：上傳資料 → 選任務 → 選方法 → 訓練 → 調參 → 註冊 → 預測
+                   # 或在「AI 助手」頁面用一句話讓 AI 幫你做完（開始訓練前會先問你）
 ```
 
 ## 使用

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/README.md

@@ -7,7 +7,7 @@
 - **方法包**：從 GitHub 安裝的「教學包」，AI 讀了就知道怎麼訓練
 - **隱私**：敏感資料只允許本地模型指揮
 
-> **1.0.2** · 文件網站：<https://honguan.github.io/autoforge/> · [變更紀錄](CHANGELOG.md)
+> **1.2.0** · 文件網站：<https://honguan.github.io/autoforge/> · [變更紀錄](CHANGELOG.md)
 
 ## 安裝與第一次使用
 
@@ -20,8 +20,8 @@ autoforge gui                     # 網頁介面
 | 安裝方式 | 包含 |
 |---|---|
 | `pip install autoforge-trainer` | 表格、文字（TF-IDF）、時間序列、異常偵測、自動調參、網頁介面、MCP、REST |
-| `pip install "autoforge-trainer[deep]"` | 再加上 Transformer、LLM 微調（LoRA/QLoRA）、圖片分類（NVIDIA GPU） |
-| `pip install "autoforge-trainer[all]"` | 再加上 Claude 代理、ONNX 匯出 |
+| `pip install "autoforge-trainer[deep]"` | 再加上 Transformer、Embedding、LLM 微調與 DPO、圖片分類、物件偵測、影像分割、語音辨識、圖像生成 LoRA（NVIDIA GPU） |
+| `pip install "autoforge-trainer[all]"` | 再加上強化學習、Claude 代理、ONNX 匯出 |
 | `docker run --rm -p 127.0.0.1:8800:8800 -v autoforge:/data ghcr.io/honguan/autoforge:cpu` | CPU 映像，預設啟動網頁介面 |
 | `docker run --rm --gpus all -p 127.0.0.1:8800:8800 -v autoforge:/data ghcr.io/honguan/autoforge:cuda` | NVIDIA CUDA 映像，所有方法包 |
 
@@ -37,8 +37,16 @@ PyPI 上的 `autoforge` 是另一個不相關的專案（`autoforge-ai`、`autof
 | `text-classification` | 文字分類（情緒、意圖、主題、垃圾訊息） | TF-IDF + 線性模型；Transformer 微調 | CPU / GPU |
 | `llm-finetune` | 大型語言模型微調 | LoRA / QLoRA（可匯出合併後的權重） | GPU（8GB+） |
 | `image-classification` | 圖片分類 | timm 預訓練模型微調（可匯出 ONNX） | GPU 建議，CPU 可跑小模型 |
-| `timeseries-forecast` | 時間序列預測（銷量、用電、流量） | 落後特徵 + 梯度提升，和「明天＝今天」的基準比較 | CPU |
+| `timeseries-forecast` | 時間序列預測（銷量、用電、流量），單一或多條序列 | 落後特徵 + 梯度提升；多序列共用一個模型；和「明天＝今天」的基準比較 | CPU |
 | `anomaly-detection` | 異常偵測（詐欺、故障、離群值） | Isolation Forest | CPU |
+| `object-detection` | 物件偵測（框出物品、瑕疵、人車） | torchvision Faster R-CNN 微調，mAP@0.5 | GPU 建議，CPU 可跑小模型 |
+| `image-segmentation` | 影像分割（逐像素標類別） | torchvision LR-ASPP / DeepLabV3，mIoU | GPU 建議，CPU 可跑小模型 |
+| `embedding-finetune` | 文字向量微調（搜尋、RAG 檢索、相似度） | 對比學習（同批負樣本），和微調前比較 recall@1 | CPU / GPU |
+| `speech-recognition` | 語音辨識 | Whisper 微調，WER / CER，和微調前比較 | GPU 建議 |
+| `llm-preference` | LLM 偏好對齊（選好的回答、避開壞的） | DPO + LoRA / QLoRA | GPU（8GB+） |
+| `diffusion-lora` | 圖像生成風格 / 主體 LoRA | Stable Diffusion 1.x/2.x UNet LoRA | GPU（8GB+） |
+| `reinforcement-learning` | 強化學習（遊戲、控制、自己的模擬器） | stable-baselines3：PPO / A2C / DQN / SAC | CPU（小網路）/ GPU |
+| `recommendation` | 推薦系統（商品、內容） | 隱式回饋 ALS、Item-KNN，和熱門排行比較 | CPU |
 
 每個方法包都附給 AI 讀的 `KNOWLEDGE.md`，以及在 CPU 上 30 秒內跑完的測試。自己做方法包：`autoforge pack new`（見 [docs/PACK_SPEC.md](docs/PACK_SPEC.md)）。
 
@@ -46,7 +54,7 @@ PyPI 上的 `autoforge` 是另一個不相關的專案（`autoforge-ai`、`autof
 
 ```bash
 git clone https://github.com/Honguan/autoforge && cd autoforge
-pip install -e ".[dev]"            # 加上 ,deep 安裝深度學習方法包的依賴
+pip install -e ".[dev]"            # 加上 ,deep 安裝深度學習方法包的依賴，,rl 安裝強化學習的依賴
 pytest -q
 ```
 
@@ -54,6 +62,7 @@ pytest -q
 
 ```bash
 autoforge gui      # 打開瀏覽器：上傳資料 → 選任務 → 選方法 → 訓練 → 調參 → 註冊 → 預測
+                   # 或在「AI 助手」頁面用一句話讓 AI 幫你做完（開始訓練前會先問你）
 ```
 
 ## 使用

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/__init__.py

@@ -1,3 +1,3 @@
 """AutoForge: AI that trains AI."""
 
-__version__ = "1.0.2"
+__version__ = "1.2.0"

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/agents/loop.py

@@ -85,35 +85,84 @@ class Transcript:
             f.write(json.dumps({"t": time.time(), "kind": kind, **data}, ensure_ascii=False, default=str) + "\n")
 
 
-async def run_agent(
-    goal: str,
-    provider: Provider,
-    actor: str,
-    brain: BrainKind,
-    approve: ApproveFn | None = None,
-    on_event: EventFn | None = None,
-    max_steps: int = 40,
-    data_path: str | None = None,
-) -> AgentResult:
-    emit = on_event or (lambda e: None)
-    transcript = Transcript()
-    runs: list[str] = []
-    server = build_server(actor, log_level="WARNING")  # the agent reports tool errors itself
-
-    user_msg = goal
-    if data_path:
-        user_msg += f"\n\nData: {Path(data_path).expanduser().resolve()}"
-    transcript.write("start", goal=goal, actor=actor, brain=brain.value, data_path=data_path)
-
-    async with Client(server) as mcp:
-        listed = await mcp.list_tools()
+def describe_call(call: ToolCall) -> str:
+    """Plain-text summary of a gated call (what training would start), shown before asking the person."""
+    if call.name == "run_retry":
+        return f"retry run {call.arguments.get('run_id')}"
+    if call.name == "hpo_start":
+        a = call.arguments
+        head = (f"  hyperparameter search: {a.get('n_trials', 10)} trials, optimise {a.get('metric')}"
+                f" ({a.get('direction') or 'auto direction'})\n")
+        return head + describe_call(ToolCall(call.id, "run_start", {"plan": a.get("plan")}))
+    plan = call.arguments.get("plan") or {}
+    task = plan.get("task") or {}
+    method = (f"{plan.get('pack')} / {plan.get('recipe')}" if plan.get("pack")
+              else f"custom code {plan.get('custom_code_path')}")
+    lines = [
+        f"  task:            {task.get('task_type')} (target: {task.get('target')}, metric: {task.get('metric')})",
+        f"  data:            {plan.get('data_path')}",
+        f"  method:          {method}",
+        f"  device:          {plan.get('device', 'auto')}",
+        f"  hyperparameters: {json.dumps(plan.get('hyperparameters') or {}, ensure_ascii=False)}",
+    ]
+    return "\n".join(lines)
+
+
+
+class AgentSession:
+    """A conversation with the agent: enter once, `send()` each user message, exit at the end.
+
+    The MCP tools stay connected between messages and the provider keeps the whole history, so a
+    follow-up like "now tune it" builds on what the agent already did. `should_stop` is checked
+    between model turns (the GUI's stop button); a stopped turn ends with stop="stopped".
+    """
+
+    def __init__(self, provider: Provider, actor: str, brain: BrainKind, approve: ApproveFn | None = None,
+                 on_event: EventFn | None = None, max_steps: int = 40,
+                 should_stop: Callable[[], bool] | None = None, origin: str | None = None):
+        self.provider, self.actor, self.brain = provider, actor, brain
+        self.approve, self.max_steps = approve, max_steps
+        self.emit = on_event or (lambda e: None)
+        self.should_stop = should_stop or (lambda: False)
+        self.origin = origin
+        self.runs: list[str] = []
+        self.transcript: Transcript | None = None
+        self._client: Client | None = None
+
+    async def __aenter__(self) -> "AgentSession":
+        self.transcript = Transcript()
+        self.transcript.write("start", actor=self.actor, brain=self.brain.value, origin=self.origin)
+        server = build_server(self.actor, log_level="WARNING")  # the agent reports tool errors itself
+        self._client = Client(server)
+        await self._client.__aenter__()
+        listed = await self._client.list_tools()
         tools = [ToolSpec(t.name, t.description or "", t.input_schema) for t in listed.tools]
-        provider.start(SYSTEM_PROMPT.format(instructions=INSTRUCTIONS), tools)
-        provider.add_user(user_msg)
-
-        final_text, stop = "", "max_steps"
-        for step in range(1, max_steps + 1):
-            turn = await provider.complete()
+        self.provider.start(SYSTEM_PROMPT.format(instructions=INSTRUCTIONS), tools)
+        return self
+
+    async def __aexit__(self, *exc: Any) -> None:
+        if self._client is not None:
+            await self._client.__aexit__(*exc)
+            self._client = None
+        if self.transcript:
+            self.transcript.write("close", runs=self.runs)
+
+    async def send(self, text: str, data_path: str | None = None) -> AgentResult:
+        """One user message → model turns and tool calls until the model answers (or a limit is hit)."""
+        assert self._client is not None and self.transcript is not None, "use `async with AgentSession(...)`"
+        mcp, transcript, emit = self._client, self.transcript, self.emit
+        user_msg = text + (f"\n\nData: {Path(data_path).expanduser().resolve()}" if data_path else "")
+        transcript.write("user", text=text, data_path=data_path)
+        self.provider.add_user(user_msg)
+        runs: list[str] = []
+
+        final_text, stop, step = "", "max_steps", 0
+        for step in range(1, self.max_steps + 1):
+            if self.should_stop():
+                stop = "stopped"
+                emit(AgentEvent("info", {"message": "stopped by the user"}))
+                break
+            turn = await self.provider.complete()
             if turn.text:
                 final_text = turn.text
                 emit(AgentEvent("text", {"text": turn.text}))
@@ -136,8 +185,8 @@ async def run_agent(
                     content = json.dumps({"ok": False, "error": {"code": "BAD_TOOL_CALL", "message": call.parse_error}})
                     results.append(ToolResult(call.id, content, is_error=True))
                     continue
-                if call.name in GATED_TOOLS and approve is not None:
-                    ok, feedback = await approve(call)
+                if call.name in GATED_TOOLS and self.approve is not None:
+                    ok, feedback = await self.approve(call)
                     emit(AgentEvent("approval", {"name": call.name, "approved": ok, "feedback": feedback}))
                     transcript.write("approval", name=call.name, approved=ok, feedback=feedback)
                     if not ok:
@@ -148,21 +197,49 @@ async def run_agent(
                 res = await mcp.call_tool(call.name, call.arguments)
                 envelope = res.structured_content
                 if envelope is None:  # unstructured (e.g. argument validation error from the MCP layer)
-                    text = "\n".join(getattr(c, "text", "") for c in res.content)
-                    envelope = {"ok": False, "error": {"code": "BAD_TOOL_CALL", "message": text}}
+                    text_out = "\n".join(getattr(c, "text", "") for c in res.content)
+                    envelope = {"ok": False, "error": {"code": "BAD_TOOL_CALL", "message": text_out}}
                 if call.name in ("run_start", "run_retry") and envelope.get("ok"):
                     runs.append(envelope["data"]["run_id"])
                 if call.name == "hpo_start" and envelope.get("ok"):
                     runs.append(envelope["data"]["study_id"])
                 content = _truncate(json.dumps(envelope, ensure_ascii=False, default=str))
                 emit(AgentEvent("tool_result", {"name": call.name, "ok": envelope.get("ok"),
-                                                "error": (envelope.get("error") or {}).get("code")}))
+                                                "error": (envelope.get("error") or {}).get("code"),
+                                                "data": _brief(call.name, envelope)}))
                 transcript.write("tool_result", name=call.name, ok=envelope.get("ok"), content=content[:2000])
                 results.append(ToolResult(call.id, content, is_error=not envelope.get("ok", False)))
-            provider.add_tool_results(results)
-        else:
-            step = max_steps
+            self.provider.add_tool_results(results)
+
+        self.runs += runs
+        transcript.write("done", stop=stop, steps=step, runs=runs)
+        emit(AgentEvent("done", {"stop": stop, "steps": step, "runs": runs}))
+        return AgentResult(final_text=final_text, steps=step, stop=stop, runs=runs, transcript=str(transcript.path))
 
-    transcript.write("done", stop=stop, steps=step, runs=runs)
-    emit(AgentEvent("done", {"stop": stop, "steps": step, "runs": runs}))
-    return AgentResult(final_text=final_text, steps=step, stop=stop, runs=runs, transcript=str(transcript.path))
+
+def _brief(name: str, envelope: dict) -> dict[str, Any] | None:
+    """The few result fields a UI shows next to a tool call (ids and metrics), never whole payloads."""
+    data = envelope.get("data")
+    if not envelope.get("ok") or not isinstance(data, dict):
+        return None
+    out = {}
+    for k in ("run_id", "study_id", "status", "metrics", "model", "version", "name", "best_value", "improvement"):
+        v = data.get(k)
+        if isinstance(v, (int, float, bool, dict)) or (isinstance(v, str) and len(v) < 200):
+            out[k] = v
+    return out or None
+
+
+async def run_agent(
+    goal: str,
+    provider: Provider,
+    actor: str,
+    brain: BrainKind,
+    approve: ApproveFn | None = None,
+    on_event: EventFn | None = None,
+    max_steps: int = 40,
+    data_path: str | None = None,
+) -> AgentResult:
+    """One-shot agent (CLI `autoforge train`): a single goal, run until the model reports back."""
+    async with AgentSession(provider, actor, brain, approve, on_event, max_steps, origin="cli") as session:
+        return await session.send(goal, data_path)

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/agents/providers.py

@@ -136,7 +136,7 @@ class AnthropicProvider(Provider):
             try:
                 import anthropic
             except ImportError:
-                raise ProviderError("the anthropic package is not installed", hint='pip install "autoforge[anthropic]"')
+                raise ProviderError("the anthropic package is not installed", hint='pip install "autoforge-trainer[anthropic]"')
             client = anthropic.AsyncAnthropic()
         self.client = client
         self.model = model

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/cli/agent.py

@@ -8,38 +8,16 @@ from pathlib import Path
 
 import typer
 
-from autoforge.agents.loop import AgentEvent, run_agent
+from autoforge.agents.loop import AgentEvent, describe_call, run_agent
 from autoforge.agents.providers import ToolCall, make_provider, parse_brain
 from autoforge.cli.output import emit, errors_as_envelope, fail
 
 _SHORT = {"hardware_profile", "pack_list", "run_status", "run_wait", "run_metrics", "run_list"}
 
 
-def _describe_plan(call: ToolCall) -> str:
-    if call.name == "run_retry":
-        return f"retry run {call.arguments.get('run_id')}"
-    if call.name == "hpo_start":
-        a = call.arguments
-        head = (f"  hyperparameter search: {a.get('n_trials', 10)} trials, optimise {a.get('metric')}"
-                f" ({a.get('direction') or 'auto direction'})\n")
-        return head + _describe_plan(ToolCall(call.id, "run_start", {"plan": a.get("plan")}))
-    plan = call.arguments.get("plan") or {}
-    task = plan.get("task") or {}
-    method = (f"{plan.get('pack')} / {plan.get('recipe')}" if plan.get("pack")
-              else f"custom code {plan.get('custom_code_path')}")
-    lines = [
-        f"  task:            {task.get('task_type')} (target: {task.get('target')}, metric: {task.get('metric')})",
-        f"  data:            {plan.get('data_path')}",
-        f"  method:          {method}",
-        f"  device:          {plan.get('device', 'auto')}",
-        f"  hyperparameters: {json.dumps(plan.get('hyperparameters') or {}, ensure_ascii=False)}",
-    ]
-    return "\n".join(lines)
-
-
 async def _ask(call: ToolCall) -> tuple[bool, str]:
     typer.echo("\n" + typer.style("The agent wants to start training:", bold=True))
-    typer.echo(_describe_plan(call))
+    typer.echo(describe_call(call))
     answer = typer.prompt("Start? [y = yes / n = no / or type feedback for the agent]", default="n").strip()
     if answer.lower() in ("y", "yes"):
         return True, ""

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/engine/context.py

@@ -51,6 +51,7 @@ class TrainContext:
         self._last_check = 0.0
         self._stop = False
         self._adjust_seq = 0
+        self._control_mtime: tuple | None = None
 
     # ---- metrics / logging ----
 
@@ -80,8 +81,15 @@ class TrainContext:
 
     def check_updates(self) -> dict[str, Any]:
         """Hyperparameter changes sent with `run adjust` since the last call (already applied to
-        self.hyperparameters). Trainers that support live changes call this between steps."""
+        self.hyperparameters). Trainers that support live changes call this between steps; control.json
+        is only re-read when it changed, so calling this every step is cheap."""
         try:
+            st = self._control.stat()
+            # (mtime, size) plus a re-read every 2 s covers filesystems with coarse timestamps
+            key = (st.st_mtime_ns, st.st_size, int(time.monotonic() // 2))
+            if key == self._control_mtime:
+                return {}
+            self._control_mtime = key
             control = json.loads(self._control.read_text())
         except (OSError, ValueError):
             return {}

{autoforge_trainer-1.0.2 → autoforge_trainer-1.2.0}/autoforge/engine/hpo.py

@@ -19,6 +19,7 @@ from __future__ import annotations
 
 import json
 import math
+import re
 import secrets
 import sys
 import time
@@ -38,6 +39,7 @@ from autoforge.policy.privacy import brain_from_actor
 from autoforge.schemas.core import BrainKind, RunStatus, TrainingPlan
 
 MINIMIZE_HINTS = ("loss", "rmse", "mse", "mae", "error", "perplexity")
+MINIMIZE_WORDS = {"wer", "cer", "mape", "wape", "smape"}  # whole words only: "wer" must not match "lower"
 
 
 class StudyNotFound(AutoForgeError):
@@ -92,7 +94,9 @@ def studies_dir() -> Path:
 
 
 def infer_direction(metric: str) -> str:
-    return "minimize" if any(h in metric.lower() for h in MINIMIZE_HINTS) else "maximize"
+    name = metric.lower()
+    words = set(re.split(r"[^a-z0-9]+", name))
+    return "minimize" if any(h in name for h in MINIMIZE_HINTS) or words & MINIMIZE_WORDS else "maximize"
 
 
 def _validate_space(space: dict[str, Any]) -> list[str]: