labpilot-ai 0.1.2__tar.gz

labpilot_ai-0.1.2/PKG-INFO

@@ -0,0 +1,150 @@
+Metadata-Version: 2.4
+Name: labpilot-ai
+Version: 0.1.2
+Summary: Industrial AI copilot for labscript-suite experiments: voice commands, runmanager/BLACS/lyse control, optimization, Knowledge, Co-Sequence, and supervised logs.
+Author: LabPilot AI Developers
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: PyQt5>=5.15
+Requires-Dist: openai>=1.0
+Requires-Dist: PyYAML>=6.0
+Requires-Dist: numpy>=1.23
+Requires-Dist: pandas>=1.5
+Requires-Dist: h5py>=3.8
+Requires-Dist: matplotlib>=3.6
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: black; extra == "dev"
+Requires-Dist: ruff; extra == "dev"
+Provides-Extra: fit
+Requires-Dist: scipy>=1.9; extra == "fit"
+Provides-Extra: voice
+Requires-Dist: sounddevice; extra == "voice"
+Requires-Dist: faster-whisper; extra == "voice"
+Provides-Extra: opt
+Requires-Dist: scikit-optimize; extra == "opt"
+Requires-Dist: optuna; extra == "opt"
+Provides-Extra: docs
+Requires-Dist: pypdf; extra == "docs"
+
+# LabPilot AI
+
+LabPilot AI is an industrial-style local AI copilot for labscript-suite experiments. It does not modify labscript, runmanager, BLACS, or lyse internals. Instead, it wraps stable boundaries: registry whitelists, local safety validation, runmanager/BLACS/lyse adapters, HDF5-compatible data, SQLite/JSONL logs, and a PyQt5 desktop UI.
+
+The first release is designed to be safe by default: Mock LLM, Mock runmanager/BLACS, and Dry-run workflows can demonstrate the full loop before any real hardware action is enabled.
+
+## Main Features
+
+- Industrial PyQt5 UI: Command, Runmanager, BLACS Manual, Lyse, Optimizer, Co-Sequence, Experiment Log, Protocol, Knowledge, Directory, Diagnostics, Error Center, and Settings.
+- Voice input: manual recording, standby wake mode, Chinese/English transcription, CPU/GPU faster-whisper backends, scientific-term correction, and microphone signal display.
+- AI command schema: strict JSON actions using `type`, including `set_global`, `set_blacs_manual`, `engage`, `load_h5`, `run_single_lyse`, `run_multi_lyse`, `plot`, `fit`, `start_optimization`, and `generate_report`.
+- Safety layer: whitelist validation, type/range/array checks, high-risk confirmations, dry-run previews, diffs, rollback hooks, SQLite audit records, and Error Center reporting.
+- runmanager control: globals read/write, float/int/bool/array scan handling, shot preview, and engage through adapters.
+- BLACS manual control: mock/localhost bridge client, limited to registered manual channels.
+- lyse/HDF5 workflow: H5 folder loading, single/multi module selection, result table merging, JSONL/SQLite cache, plots, fits, and reports.
+- Optimizer: grid and Bayesian ask/tell, plus supervised auto loop: set parameters, engage, wait for new H5, run checked lyse modules, evaluate objective, and continue safely.
+- Knowledge base: local SQLite FTS index over sequence code, connection table, lyse modules, labscript source, manuals, papers, and registries. Remote LLMs only receive short retrieved snippets.
+- Protocol Designer: import text/Markdown/PDF text and image paths to generate experiment-design suggestions without executing them.
+- Co-Sequence: AI-assisted restricted diffs for only the active sequence file and active connection table, with validation, backup, review, color tags, and change logs.
+- Experiment Log: daily Markdown/LaTeX/DokuWiki logs from natural-language commands, run records, optimizer/analysis records, Co-Sequence changes, errors, and Knowledge snippets.
+- Directory: the primary path console for active sequence, connection table, globals, BLACS context, lyse folders, H5 output, Knowledge sources, manuals, papers, and log folders.
+- Package templates: PyPI installs include default configs, example plugins, `label.png`, and the manual templates.
+
+## First-Release Limits
+
+- LabPilot AI does not modify labscript-suite internals.
+- Co-Sequence never edits arbitrary files; it is restricted to the active sequence `.py` and active connection table `.py`.
+- Co-Sequence does not automatically run experiments after code edits.
+- lyse fitting and plotting parameters are stored in LabPilot analysis records, JSONL, SQLite, and reports. The first release does not write fitting results back into original H5 files.
+- Knowledge snippets are short context references only; LabPilot AI does not execute arbitrary source code from the Knowledge database.
+- Real runmanager/BLACS/lyse hardware workflows must be validated in the lab in stages: low-risk globals, one shot, small grid scan, then BLACS single-channel tests.
+
+## Install
+
+Use the labscript conda environment or another environment that can import the labscript suite:
+
+```powershell
+conda activate labscript
+cd E:\Labpilot\labpilot_ai
+pip install -e .
+```
+
+Optional extras:
+
+```powershell
+pip install -e ".[voice]"      # sounddevice + faster-whisper
+pip install -e ".[fit,opt]"    # scipy + scikit-optimize/optuna
+pip install -e ".[docs]"       # pypdf
+pip install -e ".[dev]"        # pytest/black/ruff
+```
+
+Launch:
+
+```powershell
+labpilot-ai
+```
+
+or:
+
+```powershell
+python -m labpilot_ai
+```
+
+## API Configuration
+
+DeepSeek/OpenAI-compatible settings can be entered in the UI or provided as environment variables:
+
+```powershell
+$env:DEEPSEEK_API_KEY="your key"
+$env:DEEPSEEK_BASE_URL="https://api.deepseek.com"
+$env:DEEPSEEK_MODEL="deepseek-v4-pro"
+```
+
+If no key is available, keep `Mock LLM` enabled to test parsing and safety flows offline.
+
+## Project Templates
+
+On first run, click `Init/repair project templates` in Settings to copy clean local templates:
+
+- `configs/global_registry.yaml`
+- `configs/blacs_manual_registry.yaml`
+- `configs/lyse_registry.yaml`
+- `configs/project_settings.yaml`
+- `plugins/single_modules/`
+- `plugins/multi_modules/`
+- `manual/`
+
+Existing local files are not overwritten. Registry saves create `.bak` backups.
+
+## Voice And CUDA Notes
+
+CPU mode is the most stable default:
+
+```yaml
+voice:
+  device: "cpu"
+  compute_type: "int8"
+  model_size: "small"
+  isolated_stt: true
+```
+
+GPU STT on RTX cards can use:
+
+```yaml
+voice:
+  device: "cuda"
+  gpu_compute_type: "float16"
+```
+
+If Windows reports `Library cublas64_12.dll is not found or cannot be loaded`, run Diagnostics. LabPilot AI includes CUDA DLL path discovery for CUDA Toolkit, `nvidia-cublas-cu12`, `nvidia-cudnn-cu12`, and CTranslate2 directories. See `manual/01_voice/cuda_troubleshooting.md`.
+
+## Release Checks
+
+```powershell
+cd E:\Labpilot\labpilot_ai
+python -m compileall -q labpilot_ai
+python -m pytest -q
+git diff --check
+```
+
+Before publishing, confirm the wheel/sdist does not contain API keys, experiment H5 data, `labpilot_outputs/`, pycache, build folders, or egg-info.

labpilot_ai-0.1.2/README.md

@@ -0,0 +1,121 @@
+# LabPilot AI
+
+LabPilot AI is an industrial-style local AI copilot for labscript-suite experiments. It does not modify labscript, runmanager, BLACS, or lyse internals. Instead, it wraps stable boundaries: registry whitelists, local safety validation, runmanager/BLACS/lyse adapters, HDF5-compatible data, SQLite/JSONL logs, and a PyQt5 desktop UI.
+
+The first release is designed to be safe by default: Mock LLM, Mock runmanager/BLACS, and Dry-run workflows can demonstrate the full loop before any real hardware action is enabled.
+
+## Main Features
+
+- Industrial PyQt5 UI: Command, Runmanager, BLACS Manual, Lyse, Optimizer, Co-Sequence, Experiment Log, Protocol, Knowledge, Directory, Diagnostics, Error Center, and Settings.
+- Voice input: manual recording, standby wake mode, Chinese/English transcription, CPU/GPU faster-whisper backends, scientific-term correction, and microphone signal display.
+- AI command schema: strict JSON actions using `type`, including `set_global`, `set_blacs_manual`, `engage`, `load_h5`, `run_single_lyse`, `run_multi_lyse`, `plot`, `fit`, `start_optimization`, and `generate_report`.
+- Safety layer: whitelist validation, type/range/array checks, high-risk confirmations, dry-run previews, diffs, rollback hooks, SQLite audit records, and Error Center reporting.
+- runmanager control: globals read/write, float/int/bool/array scan handling, shot preview, and engage through adapters.
+- BLACS manual control: mock/localhost bridge client, limited to registered manual channels.
+- lyse/HDF5 workflow: H5 folder loading, single/multi module selection, result table merging, JSONL/SQLite cache, plots, fits, and reports.
+- Optimizer: grid and Bayesian ask/tell, plus supervised auto loop: set parameters, engage, wait for new H5, run checked lyse modules, evaluate objective, and continue safely.
+- Knowledge base: local SQLite FTS index over sequence code, connection table, lyse modules, labscript source, manuals, papers, and registries. Remote LLMs only receive short retrieved snippets.
+- Protocol Designer: import text/Markdown/PDF text and image paths to generate experiment-design suggestions without executing them.
+- Co-Sequence: AI-assisted restricted diffs for only the active sequence file and active connection table, with validation, backup, review, color tags, and change logs.
+- Experiment Log: daily Markdown/LaTeX/DokuWiki logs from natural-language commands, run records, optimizer/analysis records, Co-Sequence changes, errors, and Knowledge snippets.
+- Directory: the primary path console for active sequence, connection table, globals, BLACS context, lyse folders, H5 output, Knowledge sources, manuals, papers, and log folders.
+- Package templates: PyPI installs include default configs, example plugins, `label.png`, and the manual templates.
+
+## First-Release Limits
+
+- LabPilot AI does not modify labscript-suite internals.
+- Co-Sequence never edits arbitrary files; it is restricted to the active sequence `.py` and active connection table `.py`.
+- Co-Sequence does not automatically run experiments after code edits.
+- lyse fitting and plotting parameters are stored in LabPilot analysis records, JSONL, SQLite, and reports. The first release does not write fitting results back into original H5 files.
+- Knowledge snippets are short context references only; LabPilot AI does not execute arbitrary source code from the Knowledge database.
+- Real runmanager/BLACS/lyse hardware workflows must be validated in the lab in stages: low-risk globals, one shot, small grid scan, then BLACS single-channel tests.
+
+## Install
+
+Use the labscript conda environment or another environment that can import the labscript suite:
+
+```powershell
+conda activate labscript
+cd E:\Labpilot\labpilot_ai
+pip install -e .
+```
+
+Optional extras:
+
+```powershell
+pip install -e ".[voice]"      # sounddevice + faster-whisper
+pip install -e ".[fit,opt]"    # scipy + scikit-optimize/optuna
+pip install -e ".[docs]"       # pypdf
+pip install -e ".[dev]"        # pytest/black/ruff
+```
+
+Launch:
+
+```powershell
+labpilot-ai
+```
+
+or:
+
+```powershell
+python -m labpilot_ai
+```
+
+## API Configuration
+
+DeepSeek/OpenAI-compatible settings can be entered in the UI or provided as environment variables:
+
+```powershell
+$env:DEEPSEEK_API_KEY="your key"
+$env:DEEPSEEK_BASE_URL="https://api.deepseek.com"
+$env:DEEPSEEK_MODEL="deepseek-v4-pro"
+```
+
+If no key is available, keep `Mock LLM` enabled to test parsing and safety flows offline.
+
+## Project Templates
+
+On first run, click `Init/repair project templates` in Settings to copy clean local templates:
+
+- `configs/global_registry.yaml`
+- `configs/blacs_manual_registry.yaml`
+- `configs/lyse_registry.yaml`
+- `configs/project_settings.yaml`
+- `plugins/single_modules/`
+- `plugins/multi_modules/`
+- `manual/`
+
+Existing local files are not overwritten. Registry saves create `.bak` backups.
+
+## Voice And CUDA Notes
+
+CPU mode is the most stable default:
+
+```yaml
+voice:
+  device: "cpu"
+  compute_type: "int8"
+  model_size: "small"
+  isolated_stt: true
+```
+
+GPU STT on RTX cards can use:
+
+```yaml
+voice:
+  device: "cuda"
+  gpu_compute_type: "float16"
+```
+
+If Windows reports `Library cublas64_12.dll is not found or cannot be loaded`, run Diagnostics. LabPilot AI includes CUDA DLL path discovery for CUDA Toolkit, `nvidia-cublas-cu12`, `nvidia-cudnn-cu12`, and CTranslate2 directories. See `manual/01_voice/cuda_troubleshooting.md`.
+
+## Release Checks
+
+```powershell
+cd E:\Labpilot\labpilot_ai
+python -m compileall -q labpilot_ai
+python -m pytest -q
+git diff --check
+```
+
+Before publishing, confirm the wheel/sdist does not contain API keys, experiment H5 data, `labpilot_outputs/`, pycache, build folders, or egg-info.

labpilot_ai-0.1.2/labpilot_ai/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.1.0"

labpilot_ai-0.1.2/labpilot_ai/__main__.py

@@ -0,0 +1,4 @@
+from .main import main
+
+if __name__ == "__main__":
+    main()

labpilot_ai-0.1.2/labpilot_ai/ai/__init__.py

@@ -0,0 +1 @@
+

labpilot_ai-0.1.2/labpilot_ai/ai/co_sequence_prompt.py

@@ -0,0 +1,86 @@
+import json
+from pathlib import Path
+
+
+def _read_limited(path, max_chars):
+    path = Path(path)
+    text = path.read_text(encoding="utf-8", errors="replace")
+    truncated = len(text) > max_chars
+    if truncated:
+        text = text[:max_chars] + "\n\n# [LabPilot truncated this file for prompt budget]\n"
+    return text, truncated
+
+
+def build_co_sequence_prompt(
+    instruction,
+    sequence_path,
+    connection_table_path,
+    *,
+    project_context="",
+    max_file_chars=60000,
+):
+    sequence_text, sequence_truncated = _read_limited(sequence_path, max_file_chars)
+    connection_text, connection_truncated = _read_limited(connection_table_path, max_file_chars)
+    return f"""
+You are LabPilot Co-Sequence, a code editing assistant for labscript experiments.
+You may propose changes ONLY to the selected sequence file and selected connection table file.
+Do not propose changes to any other file. Do not run hardware. Do not include Markdown fences.
+
+User instruction:
+{instruction}
+
+Project knowledge snippets:
+{project_context or "No local knowledge snippets were retrieved."}
+
+Selected editable files:
+1. sequence_path={sequence_path}
+truncated={sequence_truncated}
+```python
+{sequence_text}
+```
+
+2. connection_table_path={connection_table_path}
+truncated={connection_truncated}
+```python
+{connection_text}
+```
+
+Return strict JSON with this shape:
+{{
+  "summary": "short summary",
+  "risk_level": "low|medium|high",
+  "files": [
+    {{"path": "...", "unified_diff": "unified diff for that exact file"}}
+  ],
+  "consistency_checks": ["how sequence and connection table stay consistent"],
+  "color_tags": ["yellow: reason", "red: reason"],
+  "warnings": ["operator-facing warning"]
+}}
+
+Rules:
+- Every file path must be exactly one of the two selected paths.
+- Use unified diffs only. No full file replacement.
+- Keep sequence and connection table device/channel/global names consistent.
+- Prefer small, reviewable patches.
+- If the request is unsafe or underspecified, return an empty files list and put the issue in warnings.
+""".strip()
+
+
+def mock_co_sequence_plan(instruction, sequence_path, connection_table_path):
+    return {
+        "summary": "Mock Co-Sequence did not modify code. Enable an API key to generate real diffs.",
+        "risk_level": "low",
+        "files": [],
+        "consistency_checks": [
+            "Mock mode keeps sequence and connection table unchanged.",
+            f"Sequence target: {Path(sequence_path).name if sequence_path else ''}",
+            f"Connection table target: {Path(connection_table_path).name if connection_table_path else ''}",
+        ],
+        "color_tags": ["gray: mock"],
+        "warnings": ["No patch was generated in Mock LLM mode."],
+        "instruction": instruction,
+    }
+
+
+def compact_plan_text(plan):
+    return json.dumps(plan or {}, ensure_ascii=False, indent=2)

labpilot_ai-0.1.2/labpilot_ai/ai/command_schema.py

@@ -0,0 +1,30 @@
+ALLOWED_ACTIONS = {
+    "set_global",
+    "set_blacs_manual",
+    "engage",
+    "get_globals",
+    "load_h5",
+    "load_h5_folder",
+    "run_single_lyse",
+    "run_multi_lyse",
+    "plot",
+    "fit",
+    "start_optimization",
+    "stop_optimization",
+    "tell_optimization_result",
+    "evaluate_optimization_result",
+    "generate_protocol",
+    "generate_report",
+}
+
+
+def action_summary(action: dict) -> str:
+    typ = action.get("type", "")
+    if typ in {"set_global", "set_blacs_manual"}:
+        return f"{typ}: {action.get('name')} = {action.get('value')!r}"
+    if typ == "start_optimization":
+        names = ", ".join((action.get("parameters") or {}).keys())
+        return f"{typ}: {action.get('method')} {action.get('mode')} {action.get('objective')} over {names}"
+    if typ in {"plot", "fit"}:
+        return f"{typ}: {action.get('plot_type', action.get('model', ''))}"
+    return typ

labpilot_ai-0.1.2/labpilot_ai/ai/error_advisor.py

@@ -0,0 +1,39 @@
+def advise_error(error_text: str) -> str:
+    text = (error_text or "").strip()
+    low = text.lower()
+    hints = []
+
+    if any(token in low for token in ["unicode", "utf-8", "codec", "mojibake", "encoding", "decode error", "�"]):
+        hints.append("Check source/document encoding. Save Python and Markdown files as UTF-8, then rerun compileall before release.")
+    if any(token in text for token in ["乱码", "鏄", "涓", "鎵", "寮"]):
+        hints.append("This looks like mojibake. Reopen the affected file as UTF-8, replace corrupted strings, and avoid mixed terminal encodings.")
+    if "not in global whitelist" in low or "global" in low and "whitelist" in low:
+        hints.append("Add the variable to configs/global_registry.yaml with type, range, risk, aliases, and description.")
+    if "importerror" in low or "modulenotfounderror" in low or "no module named" in low:
+        hints.append("Install the missing optional dependency, or disable the related feature in Settings before continuing.")
+    if "blacs" in low and ("connection" in low or "refused" in low or "bridge" in low or "localhost" in low):
+        hints.append("Check the BLACS localhost bridge, the registered channel name, and Mock BLACS mode before programming hardware.")
+    if "runmanager" in low and ("connection" in low or "timeout" in low or "remote" in low):
+        hints.append("Check that runmanager is running, remote access is available, and the timeout is long enough.")
+    if "h5py has already been imported" in low or "h5_lock" in low:
+        hints.append("Restart Python and ensure labpilot_ai.bootstrap_labscript.install_h5_lock() runs before importing h5py.")
+    if "cublas" in low or "cudnn" in low or "cuda runtime" in low:
+        hints.append("For GPU STT, run Diagnostics and verify CUDA/cuBLAS/cuDNN DLL paths; switch to CPU/int8 if the DLL stack is unstable.")
+    if "openmp" in low or "libiomp5md" in low or "omp: error" in low:
+        hints.append("Keep STT in isolated_release mode or CPU/int8 mode to isolate OpenMP runtime conflicts.")
+    if "pypdf" in low or "pdf import requires" in low or "pdf indexing requires" in low:
+        hints.append("Install PDF support with pip install -e .[docs] or pip install pypdf.")
+    if "module" in low and ("not registered" in low or "no module" in low or "path does not exist" in low):
+        hints.append("Register the single/multi analysis module in configs/lyse_registry.yaml and verify the module path.")
+    if "objective" in low and ("missing" in low or "unsafe" in low or "variable" in low):
+        hints.append("Choose an objective from lyse results or use a safe expression over registered result fields.")
+    if "scipy" in low or "curve_fit" in low:
+        hints.append("Install scipy fitting dependencies with pip install -e .[fit], or choose a simpler fit model.")
+    if "h5" in low and ("timeout" in low or "folder" in low or "not found" in low):
+        hints.append("Verify the H5 output folder, shot naming rule, and whether the experiment actually produced a new H5 file.")
+    if "knowledge" in low or "fts" in low or "sqlite" in low:
+        hints.append("Rebuild the Knowledge index and confirm the configured source folders are readable.")
+
+    if not hints:
+        hints.append("Check the traceback, then reproduce in Dry run or Mock mode before touching hardware.")
+    return "\n".join(f"- {hint}" for hint in hints)

labpilot_ai-0.1.2/labpilot_ai/ai/json_parser.py

@@ -0,0 +1,14 @@
+import json
+
+
+def extract_json_object(text: str) -> dict:
+    text = (text or "").strip()
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        pass
+    start = text.find("{")
+    end = text.rfind("}")
+    if start >= 0 and end > start:
+        return json.loads(text[start:end+1])
+    raise ValueError(f"无法解析 JSON：{text}")

labpilot_ai-0.1.2/labpilot_ai/ai/llm_client.py

@@ -0,0 +1,127 @@
+import os
+import re
+
+from .co_sequence_prompt import build_co_sequence_prompt, mock_co_sequence_plan
+from .json_parser import extract_json_object
+from .prompt_builder import build_command_prompt
+
+
+class LLMClient:
+    def __init__(self, api_key=None, base_url=None, model=None, use_thinking=True, mock=False):
+        self.api_key = api_key or os.getenv("DEEPSEEK_API_KEY") or os.getenv("OPENAI_API_KEY")
+        self.base_url = base_url or os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com")
+        self.model = model or os.getenv("DEEPSEEK_MODEL", "deepseek-v4-flash")
+        self.use_thinking = use_thinking
+        self.mock = mock
+
+    def parse_command(self, user_text: str, global_registry: dict, blacs_registry=None, lyse_registry=None, project_context=None) -> dict:
+        if self.mock or not self.api_key:
+            result = self._mock_parse(user_text, global_registry)
+            if project_context:
+                result["project_context_used"] = True
+            return result
+
+        try:
+            from openai import OpenAI
+        except ImportError as exc:
+            raise RuntimeError("OpenAI-compatible SDK is not installed. Install the project dependencies or enable Mock LLM.") from exc
+
+        client = OpenAI(api_key=self.api_key, base_url=self.base_url, timeout=60.0)
+        kwargs = dict(
+            model=self.model,
+            messages=[
+                {"role": "system", "content": build_command_prompt(global_registry, blacs_registry, lyse_registry, project_context=project_context)},
+                {"role": "user", "content": user_text},
+            ],
+            stream=False,
+            temperature=0,
+            response_format={"type": "json_object"},
+        )
+        if self.use_thinking:
+            kwargs["reasoning_effort"] = "high"
+            kwargs["extra_body"] = {"thinking": {"type": "enabled"}}
+        response = client.chat.completions.create(**kwargs)
+        return extract_json_object(response.choices[0].message.content)
+
+    def _mock_parse(self, text: str, global_registry: dict) -> dict:
+        """Small local parser for testing without API. It only handles common cases."""
+        actions = []
+        low = (text or "").lower()
+        no_run = any(token in low for token in ["do not run", "don't run", "no run", "not run", "dry run"])
+        no_run = no_run or any(token in (text or "") for token in ["不要运行", "不运行", "别运行", "不跑"])
+
+        if any(token in low for token in ["tof", "time of flight"]) or any(token in (text or "") for token in ["飞行时间"]):
+            m = re.search(r"(?:from|从)\s*([\d.]+)\s*(?:ms)?\s*(?:to|到)\s*([\d.]+)\s*(?:ms)?.*?(\d+)\s*(?:points|点)", text or "", re.I)
+            if m and "duration_tof_ms" in global_registry:
+                actions.append(
+                    {
+                        "type": "set_global",
+                        "name": "duration_tof_ms",
+                        "value": {"linspace": [float(m.group(1)), float(m.group(2)), int(m.group(3))]},
+                    }
+                )
+            else:
+                m = re.search(r"(?:tof|time of flight|飞行时间).*?(?:to|set to|设为|设置为|改成|=)\s*([\d.]+)", text or "", re.I)
+                if m and "duration_tof_ms" in global_registry:
+                    actions.append({"type": "set_global", "name": "duration_tof_ms", "value": float(m.group(1))})
+
+        bool_map = {
+            "lyse_do_SG_F1_masked": ["sg mask", "遮罩"],
+            "lyse_update_centers_json": ["update centers", "centers_json", "centers json", "更新中心"],
+            "do_Rabi": ["rabi", "拉比"],
+            "do_pure": ["pure", "清除"],
+            "do_Ramsey": ["ramsey"],
+        }
+        for name, keys in bool_map.items():
+            if name in global_registry and any(key.lower() in low or key in (text or "") for key in keys):
+                val = not any(token in low for token in ["disable", "off", "turn off"]) and not any(
+                    token in (text or "") for token in ["关闭", "关掉", "不要打开"]
+                )
+                actions.append({"type": "set_global", "name": name, "value": val})
+
+        if not no_run and (any(token in low for token in ["run", "submit", "engage"]) or any(token in (text or "") for token in ["运行", "跑一次", "提交"])):
+            actions.append({"type": "engage"})
+        if any(token in low for token in ["read", "current", "globals"]) or any(token in (text or "") for token in ["读取", "查看", "当前参数"]):
+            actions.append({"type": "get_globals"})
+        return {"actions": actions, "comment": "Mock LLM parser result. For serious use, enable API."}
+
+    def propose_co_sequence_patch(
+        self,
+        instruction: str,
+        sequence_path,
+        connection_table_path,
+        *,
+        project_context=None,
+        max_file_chars=60000,
+    ) -> dict:
+        if self.mock or not self.api_key:
+            return mock_co_sequence_plan(instruction, sequence_path, connection_table_path)
+
+        try:
+            from openai import OpenAI
+        except ImportError as exc:
+            raise RuntimeError("OpenAI-compatible SDK is not installed. Install the project dependencies or enable Mock LLM.") from exc
+
+        client = OpenAI(api_key=self.api_key, base_url=self.base_url, timeout=90.0)
+        prompt = build_co_sequence_prompt(
+            instruction,
+            sequence_path,
+            connection_table_path,
+            project_context=project_context,
+            max_file_chars=max_file_chars,
+        )
+        kwargs = dict(
+            model=self.model,
+            messages=[
+                {"role": "system", "content": "You produce strict JSON patch plans for LabPilot Co-Sequence."},
+                {"role": "user", "content": prompt},
+            ],
+            stream=False,
+            temperature=0,
+            response_format={"type": "json_object"},
+        )
+        if self.use_thinking:
+            kwargs["reasoning_effort"] = "high"
+            kwargs["extra_body"] = {"thinking": {"type": "enabled"}}
+        response = client.chat.completions.create(**kwargs)
+        return extract_json_object(response.choices[0].message.content)