connectonion 0.6.2__py3-none-any.whl → 0.6.3__py3-none-any.whl

connectonion/cli/co_ai/prompts/connectonion/concepts/tools.md

@@ -0,0 +1,512 @@
+# ConnectOnion Tools
+
+> Built to “show, don’t tell,” with progressive disclosure from simple to advanced.
+
+---
+
+## Quick Start (60 seconds to first tool)
+
+**Three lines to a working tool. Then call it.**
+
+```python
+from connectonion import Agent
+
+def search(query: str) -> str:  # your first tool
+    return f"Found results for {query}"
+
+agent = Agent(
+    "helper", 
+    tools=[search],
+    max_iterations=5  # Simple search tasks
+)
+```
+
+**Run it**
+
+```pycon
+>>> agent("Find Python tutorials")
+'Found results for Python tutorials'
+```
+
+That’s it.
+
+---
+
+## Core Concepts (function tools)
+
+What you get:
+
+- Clear interfaces via Python type hints
+- Automatic tool schemas for the LLM and UIs
+- Docstrings become user-facing descriptions
+
+### Function tools with type hints
+
+Type hints are the interface. Keep signatures explicit.
+
+```python
+from connectonion import Agent
+from typing import List
+
+def top_k(query: str, k: int = 5) -> List[str]:
+    """Return the top-k result titles for a query."""
+    # ... your logic ...
+    return [f"{i+1}. {query} result" for i in range(k)]
+
+agent = Agent(
+    "helper", 
+    tools=[top_k],
+    max_iterations=8  # May need multiple searches
+)
+```
+
+**What the agent sees**
+
+```pycon
+>>> agent.list_tools()
+['top_k']
+```
+### Return types matter
+
+Use structured returns when the next step needs fields.
+
+```python
+from typing import TypedDict, List
+
+class SearchHit(TypedDict):
+    title: str
+    url: str
+    score: float
+
+def search_hits(query: str, k: int = 3) -> List[SearchHit]:
+    """Structured results for chaining and UI."""
+    return [
+        {"title": f"{query} {i}", "url": f"https://example.com/{i}", "score": 0.9 - i*0.1}
+        for i in range(k)
+    ]
+```
+
+**Real output**
+
+```pycon
+>>> agent("search_hits('vector db')")
+[
+  {"title": "vector db 0", "url": "https://example.com/0", "score": 0.9},
+  {"title": "vector db 1", "url": "https://example.com/1", "score": 0.8},
+  {"title": "vector db 2", "url": "https://example.com/2", "score": 0.7}
+]
+```
+
+### How Tool Schemas Are Generated
+
+ConnectOnion automatically converts Python functions into tool schemas for the LLM.
+
+**Your Python function:**
+
+```python
+def search(query: str, limit: int = 10) -> list[str]:
+#   ^^^^^^  ^^^^^  ^^^  ^^^^^  ^^^  ^^
+#     |       |     |     |     |   └── has default → not required
+#     |       |     |     |     └────── int → "integer"
+#     |       |     |     └──────────── parameter name
+#     |       |     └────────────────── str → "string"
+#     |       └──────────────────────── parameter name
+#     └──────────────────────────────── function name
+    """Search the web for information.
+#   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+#   └── first paragraph → description
+
+    Args:
+        query: The search query string.
+        limit: Maximum results to return.
+
+    Returns:
+        List of search result titles.
+    """
+#   ↑ Args and Returns sections are NOT sent to LLM (saves tokens)
+```
+
+**JSON schema sent to LLM:**
+
+```
+{
+  "name": "search",                    ← function.__name__
+  "description": "Search the web...",  ← first paragraph of docstring
+  "parameters": {
+    "type": "object",
+    "properties": {
+      "query": {"type": "string"},     ← query: str
+      "limit": {"type": "integer"}     ← limit: int
+    },
+    "required": ["query"]              ← no default value
+  }
+}
+```
+
+**Why this minimal approach?**
+- Docstrings serve many purposes (IDE hints, Sphinx docs, etc.)
+- We only extract what the LLM needs - no confusion, fewer tokens
+- Everything is optional - no docstring is fine too
+
+**No docstring? Also fine:**
+
+```python
+def add(a: int, b: int) -> int:
+    return a + b
+```
+
+Description defaults to `"Execute the add tool."`
+
+---
+
+## Stateful Tools (class-based tools)
+
+**✅ RECOMMENDED: Pass the class instance directly to ConnectOnion!**
+
+ConnectOnion automatically discovers all public methods with type hints when you pass a class instance. This is much cleaner than listing methods individually.
+
+Use class instances when tools need shared state, caching, or resources.
+
+### Browser automation with Playwright (navigate, screenshot, tabs)
+
+Requirements:
+
+- Install Playwright and browsers: `pip install playwright` then `playwright install`
+
+```python
+from typing import List, Optional
+from connectonion import Agent
+
+try:
+    from playwright.sync_api import sync_playwright, Page
+except ImportError:
+    raise SystemExit("Install Playwright: pip install playwright && playwright install")
+
+
+class Browser:
+    """Persistent browser session with navigation, screenshots, and tab control."""
+
+    def __init__(self):
+        self._p = None
+        self._browser = None
+        self._pages: dict[str, Page] = {}
+        self._active_tab: Optional[str] = None
+
+    def start(self, headless: bool = True) -> str:
+        """Start the browser session and open the first tab named 'main'."""
+        self._p = sync_playwright().start()
+        self._browser = self._p.chromium.launch(headless=headless)
+        self._pages["main"] = self._browser.new_page()
+        self._active_tab = "main"
+        return f"Browser started (headless={headless}) with tab 'main'"
+
+    def new_tab(self, name: str) -> str:
+        """Open a new tab with a friendly name, e.g., 'docs' or 'shop'."""
+        if not self._browser:
+            return "Error: Browser not started. Call start() first."
+        if name in self._pages:
+            return f"Tab '{name}' already exists"
+        self._pages[name] = self._browser.new_page()
+        self._active_tab = name
+        return f"Opened tab '{name}'"
+
+    def list_tabs(self) -> List[str]:
+        """List available tab names."""
+        return list(self._pages.keys())
+
+    def switch_tab(self, name: str) -> str:
+        """Switch the active tab by name."""
+        if name not in self._pages:
+            return f"Error: No tab named '{name}'"
+        self._active_tab = name
+        return f"Switched to tab '{name}'"
+
+    def goto(self, url: str, tab: Optional[str] = None) -> str:
+        """Navigate the active (or specified) tab to a URL and return the page title."""
+        if not self._pages:
+            return "Error: Browser not started. Call start() first."
+        target = tab or self._active_tab
+        page = self._pages[target]
+        page.goto(url)
+        return page.title()
+
+    def screenshot(self, path: Optional[str] = None, tab: Optional[str] = None) -> str:
+        """Save a PNG screenshot of the active (or specified) tab and return the filename."""
+        if not self._pages:
+            return "Error: Browser not started. Call start() first."
+        target = tab or self._active_tab
+        page = self._pages[target]
+        filename = path or f"{target}_screenshot.png"
+        page.screenshot(path=filename)
+        return filename
+
+    def close_tab(self, name: str) -> str:
+        """Close a tab by name."""
+        if name not in self._pages:
+            return f"Error: No tab named '{name}'"
+        self._pages[name].close()
+        del self._pages[name]
+        if self._active_tab == name:
+            self._active_tab = next(iter(self._pages), None)
+        return f"Closed tab '{name}'"
+
+    def close(self) -> None:
+        """Close all tabs and stop the browser."""
+        for page in list(self._pages.values()):
+            page.close()
+        self._pages.clear()
+        if self._browser:
+            self._browser.close()
+        if self._p:
+            self._p.stop()
+
+
+browser = Browser()
+
+# ✅ BEST PRACTICE: Pass the class instance directly!
+# ConnectOnion automatically extracts all public methods as tools
+agent = Agent(
+    "helper", 
+    tools=[browser],  # Just pass the instance - ConnectOnion does the rest!
+    max_iterations=15  # Browser automation often needs more steps
+)
+
+# With an LLM, the agent can use natural language prompts to decide which tools to call and in what order.
+# For example, you can give the agent a high-level instruction and it will use the available tools to accomplish the task:
+
+response = agent("Open https://example.com, take a screenshot, then open a new tab to https://playwright.dev and screenshot that too. List all open tabs at the end.")
+print(response)
+    browser.close()
+```
+
+### Todo list
+
+```python
+class TodoList:
+    """Simple todo list with add/list."""
+    def __init__(self):
+        self._items: list[str] = []
+
+    def add(self, text: str) -> None:
+        """Add a new todo item."""
+        self._items.append(text)
+
+    def list(self) -> list[str]:
+        """Return all todo items."""
+        return self._items
+
+todos = TodoList()
+
+# ✅ Two approaches - class instance is cleaner:
+# Option 1: Pass individual methods (more verbose)
+agent_verbose = Agent("helper", tools=[todos.add, todos.list], max_iterations=10)
+
+# Option 2: Pass class instance (recommended - cleaner!)  
+agent = Agent("helper", tools=[todos], max_iterations=10)
+```
+
+**Real session**
+
+```pycon
+>>> agent("add('buy milk')")
+>>> agent("add('book flights')")
+>>> agent("list()")
+['buy milk', 'book flights']
+```
+
+### Resource management
+
+Own the lifecycle. Close things you open. The `Browser` class above
+exposes a `close()` method and uses `try/finally` to guarantee cleanup.
+
+---
+
+## Advanced Patterns
+
+### Tool composition
+
+Small tools compose into bigger moves.
+
+```python
+def pick_top(hit_list: list[dict]) -> dict:
+    """Choose the highest score item."""
+    return max(hit_list, key=lambda h: h["score"])
+
+def search_then_pick(query: str) -> dict:
+    """Search then pick the best hit."""
+    hits = search_hits(query, k=5)
+    return pick_top(hits)
+```
+
+**Output**
+
+```pycon
+>>> agent("search_then_pick('weaviate vs pgvector')")
+{'title': 'weaviate vs pgvector 0', 'url': 'https://example.com/0', 'score': 0.9}
+```
+
+### Custom tool schemas
+
+Expose structured inputs with clear constraints.
+
+```python
+from dataclasses import dataclass
+from typing import Annotated, Literal
+
+Priority = Literal["low", "normal", "high"]
+
+@dataclass
+class Ticket:
+    title: str
+    description: str
+    priority: Priority
+    assignee: Annotated[str, "email"]
+
+def create_ticket(t: Ticket) -> dict:
+    """Create a ticket and return its metadata."""
+    return {"id": "T-1024", "title": t.title, "priority": t.priority, "assignee": t.assignee}
+
+agent = Agent(
+    "helper", 
+    tools=[create_ticket],
+    max_iterations=5  # Ticket creation is straightforward
+)
+```
+
+**Schema (example)**
+
+```json
+{
+  "name": "create_ticket",
+  "input": {
+    "type": "object",
+    "properties": {
+      "title": {"type": "string"},
+      "description": {"type": "string"},
+      "priority": {"enum": ["low", "normal", "high"]},
+      "assignee": {"type": "string", "format": "email"}
+    },
+    "required": ["title", "description", "priority", "assignee"]
+  },
+  "returns": {"type": "object"}
+}
+```
+
+---
+
+## Tool Development Guide
+
+### Design rules
+
+* One job per tool. Names are verbs: `search`, `summarize`, `create_ticket`.
+* Clean signatures. Prefer primitives and TypedDict/dataclass for structure.
+* First docstring line explains the value in plain language.
+* Deterministic by default. Document non-determinism.
+
+### Error handling
+
+* Validate inputs. Raise clear exceptions. Return helpful messages.
+* Timeouts for IO. Circuit-breakers for flaky deps.
+* Log stack traces during development.
+
+### Performance
+
+* Cache pure functions.
+* Batch remote calls.
+* Stream large outputs when useful.
+
+### Security
+
+* Treat inputs as untrusted.
+* Sanitize shell/SQL/HTML.
+* Scope credentials. Rotate secrets. Audit access.
+
+### Testing
+
+* Unit-test each tool. Avoid network in tests.
+* Golden tests for outputs.
+* Fuzz inputs for robustness.
+
+### Versioning and deprecation
+
+* Add `version="1.2.0"` metadata when behavior changes.
+* Keep old signatures working until removal.
+* Announce removals with dates.
+
+### Authoring checklist
+
+* [ ] Clear name and one-liner
+* [ ] Explicit types
+* [ ] Real output example
+* [ ] Errors handled
+* [ ] Tests written
+
+---
+
+## Appendix: Patterns at a glance
+
+* **Function tool** → simplest path to value.
+* **Class tool** → shared state, caching, external handles.
+* **Composition** → small tools, big outcomes.
+* **Custom schemas** → robust interfaces and UIs.
+
+---
+
+## FAQ: How tools are discovered and used
+
+- Tool schemas are built from function name, type hints, and first paragraph of docstring
+- Args/Returns sections in docstrings are NOT sent to LLM (saves tokens)
+- No docstring? That's fine - description defaults to "Execute the {name} tool."
+- **✅ For class-based tools: Pass the instance directly!** ConnectOnion auto-discovers all public methods with type hints
+- You can always call `agent.tools.tool_name.run(**kwargs)` to run tools without an LLM.
+
+## Class Instance vs Individual Methods
+
+**✅ RECOMMENDED (Clean & Automatic):**
+```python
+browser = BrowserAutomation()
+agent = Agent("browser_agent", tools=[browser])  # Auto-discovers all methods!
+```
+
+**❌ VERBOSE (Works but not recommended):**
+```python
+browser = BrowserAutomation()
+agent = Agent("browser_agent", tools=[
+    browser.start_browser,
+    browser.navigate, 
+    browser.take_screenshot,
+    browser.scrape_content,
+    # ... list every method manually
+])
+```
+
+ConnectOnion's class instance support makes your code much cleaner and more maintainable!
+
+---
+
+## Built-in Tools
+
+ConnectOnion provides ready-to-use tools. See [useful_tools/](useful_tools/) for full documentation.
+
+| Tool | Purpose |
+|------|---------|
+| [Shell](useful_tools/shell.md) | Execute shell commands |
+| [DiffWriter](useful_tools/diff_writer.md) | Edit files with diffs |
+| [TodoList](useful_tools/todo_list.md) | Track task progress |
+| [WebFetch](useful_tools/web_fetch.md) | Fetch web content |
+| [Gmail](useful_tools/gmail.md) | Gmail integration |
+| [Outlook](useful_tools/outlook.md) | Outlook integration |
+| [GoogleCalendar](useful_tools/google_calendar.md) | Google Calendar |
+| [MicrosoftCalendar](useful_tools/microsoft_calendar.md) | Microsoft Calendar |
+| [Memory](useful_tools/memory.md) | Persistent storage |
+| [Terminal](useful_tools/terminal.md) | Interactive terminal |
+| [SlashCommand](useful_tools/slash_command.md) | Custom commands |
+
+```python
+from connectonion import Agent, Shell, Gmail
+
+shell = Shell()
+gmail = Gmail()
+
+agent = Agent("assistant", tools=[shell, gmail])
+```

connectonion/cli/co_ai/prompts/connectonion/concepts/transcribe.md

@@ -0,0 +1,156 @@
+# Audio Transcription
+
+Convert audio files to text using Gemini's multimodal capabilities. Simple one-function interface for transcription.
+
+## Quick Start
+
+```python
+from connectonion import transcribe
+
+# Simple transcription (uses OpenOnion managed keys)
+text = transcribe("meeting.mp3")
+print(text)
+
+# With your own Gemini API key
+text = transcribe("meeting.mp3", model="gemini-3-flash-preview")
+```
+
+That's it! One function for audio-to-text.
+
+## With Context Hints
+
+Improve accuracy for domain-specific terms:
+
+```python
+# Technical meeting with specific names
+text = transcribe(
+    "standup.mp3",
+    prompt="Technical AI discussion. Names: Aaron, Lisa. Terms: ConnectOnion, OpenOnion"
+)
+
+# Medical transcription
+text = transcribe(
+    "consultation.mp3",
+    prompt="Medical consultation. Terms: hypertension, metformin, CBC"
+)
+```
+
+## With Timestamps
+
+```python
+text = transcribe("podcast.mp3", timestamps=True)
+# Output: "[00:00] Welcome to the show...\n[00:15] Today we're discussing..."
+```
+
+## Real Examples
+
+### Meeting Minutes
+
+```python
+def get_meeting_minutes(audio_path: str) -> str:
+    """Transcribe and summarize a meeting."""
+    from connectonion import transcribe, llm_do
+
+    # Step 1: Transcribe
+    transcript = transcribe(audio_path, prompt="Business meeting")
+
+    # Step 2: Summarize
+    summary = llm_do(
+        transcript,
+        system_prompt="Extract action items and key decisions as bullet points."
+    )
+    return summary
+```
+
+### Voice Notes Processing
+
+```python
+from pathlib import Path
+
+def process_voice_notes(folder: str) -> list[str]:
+    """Transcribe all voice notes in a folder."""
+    from connectonion import transcribe
+
+    results = []
+    for audio in Path(folder).glob("*.mp3"):
+        text = transcribe(str(audio))
+        results.append(f"# {audio.stem}\n{text}")
+    return results
+```
+
+### Use as Agent Tool
+
+```python
+from connectonion import Agent, transcribe
+
+def transcribe_audio(file_path: str) -> str:
+    """Transcribe an audio file to text."""
+    return transcribe(file_path)
+
+agent = Agent("assistant", tools=[transcribe_audio])
+result = agent.input("Transcribe the file meeting.mp3 and summarize it")
+```
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `audio` | str | required | Path to audio file |
+| `prompt` | str | None | Context hints for accuracy |
+| `model` | str | "co/gemini-3-flash-preview" | Model to use |
+| `timestamps` | bool | False | Include timestamps in output |
+
+## Supported Formats
+
+- WAV, MP3, AIFF, AAC, OGG, FLAC, M4A, WebM
+
+**Token cost:** 32 tokens per second of audio (1 minute = 1,920 tokens)
+
+## Models
+
+```python
+# OpenOnion managed keys (default - no API key needed)
+transcribe("audio.mp3", model="co/gemini-3-flash-preview")
+transcribe("audio.mp3", model="co/gemini-2.5-flash")
+
+# Your own Gemini API key (set GEMINI_API_KEY)
+transcribe("audio.mp3", model="gemini-3-flash-preview")
+transcribe("audio.mp3", model="gemini-2.5-flash")
+```
+
+## Error Handling
+
+```python
+from connectonion import transcribe
+
+try:
+    text = transcribe("nonexistent.mp3")
+except FileNotFoundError:
+    print("Audio file not found")
+except ValueError as e:
+    print(f"API error: {e}")
+```
+
+## Comparison with Agent
+
+| Feature | `transcribe()` | `Agent()` |
+|---------|----------------|-----------|
+| Purpose | Audio to text | Multi-step workflows |
+| Input | Audio files | Text prompts |
+| Output | Plain text | Agent responses |
+| Best for | Transcription | Complex tasks |
+
+```python
+# Use transcribe() for audio-to-text
+text = transcribe("meeting.mp3")
+
+# Use Agent for complex workflows with multiple tools
+agent = Agent("assistant", tools=[search, calculate])
+result = agent.input("Research and analyze...")
+```
+
+## Next Steps
+
+- Learn about [llm_do()](concepts/llm_do.md) for one-shot LLM calls
+- Explore [Agents](concepts/agent.md) for multi-step workflows
+- See [Tools](concepts/tools.md) for extending agents