feat: add Claude Agent SDK implementation

.claude/CLAUDE.md

@@ -4,19 +4,20 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## What This Is
 
-A monorepo containing two parallel implementations of **Casey**, an AI-powered IT helpdesk agent for Slack built with Bolt for Python. Both implementations are functionally identical from the Slack user's perspective but use different AI agent frameworks:
+A monorepo containing three parallel implementations of **Casey**, an AI-powered IT helpdesk agent for Slack built with Bolt for Python. All implementations are functionally identical from the Slack user's perspective but use different AI agent frameworks:
 
-- `pydantic-ai/` — Built with **Pydantic AI**
+- `claude-agent-sdk/` — Built with **Claude Agent SDK**
 - `openai-agents-sdk/` — Built with **OpenAI Agents SDK**
+- `pydantic-ai/` — Built with **Pydantic AI**
 
 All tool data (knowledge base, tickets, password resets, system status, permissions) is hardcoded for demo purposes.
 
 ## Commands
 
-All commands must be run from within the respective project directory (`pydantic-ai/` or `openai-agents-sdk/`).
+All commands must be run from within the respective project directory (`claude-agent-sdk/`, `openai-agents-sdk/`, or `pydantic-ai/`).
 
 ```sh
-# Run the app (requires .env with OPENAI_API_KEY; Slack tokens optional with CLI)
+# Run the app (requires .env with OPENAI_API_KEY or ANTHROPIC_API_KEY; Slack tokens optional with CLI)
 slack run          # via Slack CLI
 python3 app.py     # directly
 
@@ -32,13 +33,14 @@ pytest
 
 ```
 .github/              # Shared CI workflows and dependabot config
-pydantic-ai/         # Pydantic AI implementation
+claude-agent-sdk/    # Claude Agent SDK implementation
 openai-agents-sdk/   # OpenAI Agents SDK implementation
+pydantic-ai/         # Pydantic AI implementation
 ```
 
-CI runs ruff lint/format checks against both directories via a matrix strategy in `.github/workflows/ruff.yml`. Dependabot monitors `requirements.txt` in both directories independently.
+CI runs ruff lint/format checks against all directories via a matrix strategy in `.github/workflows/ruff.yml`. Dependabot monitors `requirements.txt` in all directories independently.
 
-## Architecture (shared across both implementations)
+## Architecture (shared across all implementations)
 
 Three-layer design: **app.py** → **listeners/** → **agent/**
 
@@ -59,15 +61,16 @@ Each sub-package has a `register(app)` function called from `listeners/__init__.
 
 ## Key Differences Between Implementations
 
-| Aspect | Pydantic AI | OpenAI Agents SDK |
-|--------|-------------|-------------------|
-| Agent file | `agent/casey.py` | `agent/support_agent.py` |
-| Agent definition | `Agent(deps_type=CaseyDeps)` | `Agent[CaseyDeps](model="gpt-4o-mini")` |
-| Model config | Passed at runtime via `run_sync(model=DEFAULT_MODEL)` | Set directly on agent constructor |
-| Tool definition | Plain async functions | `@function_tool` decorated functions |
-| Tool context param | `RunContext[CaseyDeps]` | `RunContextWrapper[CaseyDeps]` |
-| Execution | `casey_agent.run_sync(text, model=..., deps=..., message_history=...)` | `Runner.run_sync(casey_agent, input=..., context=...)` |
-| Result output | `result.output` | `result.final_output` |
-| Result messages | `result.all_messages()` | `result.to_input_list()` |
-| History type | `list[ModelMessage]` (framework-native) | `list` (generic, manually constructed) |
-| Feedback blocks | Native `FeedbackButtonsElement` | Native `FeedbackButtonsElement` |
+| Aspect | Claude Agent SDK | OpenAI Agents SDK | Pydantic AI |
+|--------|-----------------|-------------------|-------------|
+| Agent file | `agent/casey.py` | `agent/support_agent.py` | `agent/casey.py` |
+| App type | `AsyncApp` (fully async) | `App` (sync) | `App` (sync) |
+| Agent definition | `ClaudeSDKClient` with `ClaudeAgentOptions` | `Agent[CaseyDeps](model="gpt-4o-mini")` | `Agent(deps_type=CaseyDeps)` |
+| Model config | Managed by SDK (Claude models) | Set directly on agent constructor | Passed at runtime via `run_sync(model=DEFAULT_MODEL)` |
+| Tool definition | `@tool` decorated functions via MCP server | `@function_tool` decorated functions | Plain async functions |
+| Tool context param | `args` dict (no context param) | `RunContextWrapper[CaseyDeps]` | `RunContext[CaseyDeps]` |
+| Execution | `await run_casey_agent(text, session_id=...)` | `Runner.run_sync(casey_agent, input=..., context=...)` | `casey_agent.run_sync(text, model=..., deps=..., message_history=...)` |
+| Result output | `response_text` from collected `TextBlock.text` | `result.final_output` | `result.output` |
+| Conversation history | Session-based via `resume` (server-side) | `list` stored locally | `list[ModelMessage]` stored locally |
+| API key env var | `ANTHROPIC_API_KEY` | `OPENAI_API_KEY` | `OPENAI_API_KEY` |
+| Feedback blocks | Native `FeedbackButtonsElement` | Native `FeedbackButtonsElement` | Native `FeedbackButtonsElement` |

.github/dependabot.yml

@@ -5,6 +5,7 @@ updates:
       - "/"
       - "/openai-agents-sdk"
       - "/pydantic-ai"
+      - "/claude-agent-sdk"
     schedule:
       interval: "weekly"
     labels:

.github/workflows/ruff.yml

@@ -14,6 +14,7 @@ jobs:
         directory:
           - openai-agents-sdk
           - pydantic-ai
+          - claude-agent-sdk
     defaults:
       run:
         working-directory: ${{ matrix.directory }}

README.md

@@ -6,14 +6,15 @@ Built with [Bolt for Python](https://docs.slack.dev/tools/bolt-python/).
 
 ## Choose Your Framework
 
-This repo contains the same app built with two different AI agent frameworks. Pick the one that fits your stack:
+This repo contains the same app built with three different AI agent frameworks. Pick the one that fits your stack:
 
-| | [Pydantic AI](./pydantic-ai/) | [OpenAI Agents SDK](./openai-agents-sdk/) |
-|---|---|---|
-| **Framework** | [pydantic-ai](https://ai.pydantic.dev/) | [openai-agents](https://openai.github.io/openai-agents-python/) |
-| **Get started** | [View README](./pydantic-ai/README.md) | [View README](./openai-agents-sdk/README.md) |
+| App | Framework | Get started | Directory |
+|-----|-----------|-------------|-----------|
+| **Claude Agent SDK** | [claude-agent-sdk](https://docs.anthropic.com/en/docs/agents/claude-agent-sdk) | [View README](./claude-agent-sdk/README.md) | `claude-agent-sdk/` |
+| **OpenAI Agents SDK** | [openai-agents](https://openai.github.io/openai-agents-python/) | [View README](./openai-agents-sdk/README.md) | `openai-agents-sdk/` |
+| **Pydantic AI** | [pydantic-ai](https://ai.pydantic.dev/) | [View README](./pydantic-ai/README.md) | `pydantic-ai/` |
 
-Both implementations share the same Slack listener layer, the same five simulated IT tools, and the same user experience. The only difference is how the agent is defined and executed under the hood.
+All implementations share the same Slack listener layer, the same five simulated IT tools, and the same user experience. The only difference is how the agent is defined and executed under the hood.
 
 ## What Casey Can Do
 

claude-agent-sdk/.claude/CLAUDE.md

@@ -0,0 +1,19 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+See the root `../.claude/CLAUDE.md` for monorepo-wide architecture, commands, and a comparison of all implementations.
+
+## Claude Agent SDK Specifics
+
+**App (`app.py`)** uses `AsyncApp` from Bolt for Python. All listeners and Slack API calls are fully async (`await`).
+
+**Agent (`agent/casey.py`)** uses `ClaudeSDKClient` from the Claude Agent SDK. Tools are registered via `create_sdk_mcp_server()` and passed as `mcp_servers` in `ClaudeAgentOptions`. The `run_casey_agent()` function is async and returns `(response_text, session_id)`.
+
+**Tools (`agent/tools/`)** are defined with the `@tool` decorator from `claude_agent_sdk`. Each tool returns `{"content": [{"type": "text", "text": ...}]}`. Tools are registered into a single MCP server via `create_sdk_mcp_server()`.
+
+**Conversation history** is managed server-side by the Claude Agent SDK via sessions. The local `SessionStore` (`conversation/store.py`) only maps `(channel_id, thread_ts)` to session IDs. Sessions are resumed via `ClaudeAgentOptions(resume=session_id)`.
+
+**Feedback blocks** use the native `FeedbackButtonsElement` from `slack_sdk.models.blocks`. A single `feedback` action ID is registered.
+
+**Dependencies** (`agent/deps.py`) use `AsyncWebClient` from `slack_sdk.web.async_client` instead of the sync `WebClient`.

claude-agent-sdk/.env.sample

@@ -0,0 +1,9 @@
+# Optional, uncomment and set when running without the Slack CLI (python3 app.py).
+# SLACK_APP_TOKEN=YOUR_SLACK_APP_TOKEN
+# SLACK_BOT_TOKEN=YOUR_SLACK_BOT_TOKEN
+
+# Optional, uncomment and set when using a custom Slack instance.
+# SLACK_API_URL=YOUR_SLACK_API_URL
+
+# Required, set your Anthropic API key.
+ANTHROPIC_API_KEY=YOUR_ANTHROPIC_API_KEY

claude-agent-sdk/.gitignore

@@ -0,0 +1,43 @@
+# general things to ignore
+build/
+dist/
+docs/_sources/
+docs/.doctrees
+.eggs/
+*.egg-info/
+*.egg
+*.py[cod]
+__pycache__/
+*.so
+*~
+
+# virtualenv
+env*/
+venv/
+.venv*
+.env*
+!.env.sample
+
+# codecov / coverage
+.coverage
+cov_*
+coverage.xml
+
+# due to using tox and pytest
+.tox
+.cache
+.pytest_cache/
+.python-version
+pip
+.mypy_cache/
+
+# misc
+tmp.txt
+.DS_Store
+logs/
+*.db
+.pytype/
+.idea/
+
+# claude
+.claude/*.local.json

claude-agent-sdk/.slack/.gitignore

@@ -0,0 +1,2 @@
+apps.dev.json
+cache/

claude-agent-sdk/.slack/config.json

@@ -0,0 +1,6 @@
+{
+  "manifest": {
+    "source": "local"
+  },
+  "project_id": "c189134c-231e-4c96-b800-ea39c055aa77"
+}

claude-agent-sdk/.slack/hooks.json

@@ -0,0 +1,5 @@
+{
+  "hooks": {
+    "get-hooks": "python3 -m slack_cli_hooks.hooks.get_hooks"
+  }
+}