feat: add the Claude Agent SDK coding agent

.github/dependabot.yml

@@ -2,6 +2,7 @@ version: 2
 updates:
   - package-ecosystem: "npm"
     directories:
+      - "/claude-agent-sdk"
       - "/openai-agents-sdk"
     schedule:
       interval: "daily"

.github/workflows/lint.yml

@@ -14,6 +14,7 @@ jobs:
     strategy:
       matrix:
         directory:
+          - claude-agent-sdk
           - openai-agents-sdk
     defaults:
       run:

.github/workflows/tests.yml

@@ -14,6 +14,7 @@ jobs:
     strategy:
       matrix:
         directory:
+          - claude-agent-sdk
           - openai-agents-sdk
     defaults:
       run:

.gitignore

@@ -10,5 +10,3 @@
 # Slack CLI
 */.slack/apps.json
 */.slack/apps.dev.json
-*/.slack/cache/
-*/.slack/config.json

AGENTS.md

@@ -6,6 +6,7 @@ Guidance for coding agents working in this repository.
 
 A monorepo for a **Coding Agent** sample for Slack, built with Bolt for JavaScript. Each framework lives in its own directory as a standalone sample:
 
+- `claude-agent-sdk/` — built with the **Claude Agent SDK**
 - `openai-agents-sdk/` — built with the **OpenAI Agents SDK**
 
 Implementations share a structure and user experience, not code. Keep each one obvious on its own, written in its framework's idiom, rather than adding shared helpers.
@@ -37,6 +38,7 @@ ES modules with kebab-case filenames, JSDoc types checked by `tsc --checkJs`, Bi
 
 ```
 .github/             # Shared CI workflows and dependabot config
+claude-agent-sdk/    # Claude Agent SDK implementation
 openai-agents-sdk/   # OpenAI Agents SDK implementation
 ```
 
@@ -58,7 +60,7 @@ Implementations share structure, not code, so these parts are kept the same by c
 
 - **The seam between `listeners/` and `agent/`.** Listeners import only from `agent/index.js`, never from the agent framework, so `listeners/` is the same in every implementation. `agent/index.js` exports `answerRequest`, `findSession`, `runTurn`, `resumeTurn`, `stopTurn`, and `endSession`.
 - **Framework-free files.** `agent/turns.js`, `agent/context.js`, `agent/approvals.js`, `agent/sessions.js`, `agent/workspace.js`, `agent/outcome.js`, and `agent/prompts.js` don't import the agent framework. The framework lives in `agent/agent.js`, `agent/guardrails.js`, and `agent/tools/`. `turns.js` calls only `runRequest(prompt, deps)`, which returns the reply text, and `startRun(history, text, deps)` and `resumeRun(paused, approved, deps)`, which return the `RunOutcome` in `agent/outcome.js`. It calls them through its exported `runs` object, so tests can stand in for them. The paused run is a handle `turns.js` passes back without looking inside, so each framework's way of pausing fits behind it.
-- **What the framework side keeps true.** `turns.js` relies on these, so every implementation keeps them: `create_code_channel` sets `deps.codeChannelId`, and `archive_channel` sets `session.archived`. Every code channel tool skips its call once `session.stopped` is set. A paused run's history already records the pending command as not run, so `turns.js` can drop a paused run without resuming it. Beyond that, `turns.js` only checks whether `history` is empty, to know a turn is the first.
+- **What the framework side keeps true.** `turns.js` relies on these, so every implementation keeps them: `create_code_channel` sets `deps.codeChannelId`, and `archive_channel` sets `session.archived`. Every code channel tool skips its call once `session.stopped` is set. `turns.js` can drop a paused run without resuming it, since the session's history then shows its pending command as not run: the history already records it that way, or the next run answers it that way before anything else. Beyond that, `turns.js` only checks whether `history` is empty, to know a turn is the first.
 - **Prompts.** `agent/prompts.js` is the same word for word in every implementation.
 - **Tools.** Every implementation offers the same tools with the same names, arguments, and descriptions, written with its framework's tool helper. File tools stay inside the clone and can't write into `.git`. A tool that throws tells the model what went wrong.
 - **Command approvals.** `commandNeedsApproval(command)` is plain JavaScript and the same in every implementation, wrapped in the framework's own approval hook. Read-only commands and local git changes run without asking. Anything that runs the repository's code or reaches outside the clone asks first. Commands run without a shell, and `splitCommand` in `agent/workspace.js` splits a command the same way for the check and the run. On Windows, Node won't start `.cmd` and `.bat` programs like `npm` without a shell, so those don't run there.

README.md

@@ -7,9 +7,10 @@ A sample Slack agent that showcases [Slack Code](https://docs.slack.dev/ai/slack
 
 ## Choose your framework
 
-| App                   | Directory            | Get started                                  | Framework                                                   |
-| --------------------- | -------------------- | -------------------------------------------- | ----------------------------------------------------------- |
-| **OpenAI Agents SDK** | `openai-agents-sdk/` | [View README](./openai-agents-sdk/README.md) | [openai-agents](https://openai.github.io/openai-agents-js/) |
+| App                   | Directory            | Get started                                  | Framework                                                                  |
+| --------------------- | -------------------- | -------------------------------------------- | -------------------------------------------------------------------------- |
+| **Claude Agent SDK**  | `claude-agent-sdk/`  | [View README](./claude-agent-sdk/README.md)  | [claude-agent-sdk](https://platform.claude.com/docs/en/agent-sdk/overview) |
+| **OpenAI Agents SDK** | `openai-agents-sdk/` | [View README](./openai-agents-sdk/README.md) | [openai-agents](https://openai.github.io/openai-agents-js/)                |
 
 ## What it can do
 

claude-agent-sdk/.env.sample

@@ -0,0 +1,15 @@
+# Required, set your Anthropic API key.
+ANTHROPIC_API_KEY=YOUR_ANTHROPIC_API_KEY
+
+# Optional, uncomment and set when running without the Slack CLI (npm start).
+# SLACK_APP_TOKEN=YOUR_SLACK_APP_TOKEN
+# SLACK_BOT_TOKEN=YOUR_SLACK_BOT_TOKEN
+
+# Optional, uncomment and set for OAuth mode (node app-oauth.js).
+# SLACK_CLIENT_ID=YOUR_SLACK_CLIENT_ID
+# SLACK_CLIENT_SECRET=YOUR_SLACK_CLIENT_SECRET
+# SLACK_SIGNING_SECRET=YOUR_SLACK_SIGNING_SECRET
+# SLACK_STATE_SECRET=YOUR_SLACK_STATE_SECRET
+
+# Optional, uncomment and set when using a custom Slack instance.
+# SLACK_API_URL=YOUR_SLACK_API_URL

claude-agent-sdk/.gitignore

@@ -0,0 +1,18 @@
+# Claude Code
+.claude/*.local.json
+
+# Environment
+.env*
+!.env.sample
+
+# Logs
+logs/
+
+# macOS
+.DS_Store
+
+# Node.js
+node_modules/
+
+# OAuth
+data/

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": "452c1566-00e1-424d-bcca-3ee66ffe370b"
+}

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

@@ -0,0 +1,5 @@
+{
+  "hooks": {
+    "get-hooks": "npx -q --no-install -p @slack/cli-hooks slack-cli-get-hooks"
+  }
+}

claude-agent-sdk/AGENTS.md

@@ -0,0 +1,23 @@
+# AGENTS.md
+
+Guidance for coding agents working in this directory. Refer to the root `../AGENTS.md` for repository-wide structure and commands.
+
+## Claude Agent SDK specifics
+
+**Agents (`agent/agent.js`)** are two sets of `query()` options with pinned models: a routing `requestAgent` (`claude-haiku-4-5`) whose only tool is `create_code_channel`, and the coding `agent` (`claude-sonnet-5-5`) with every other tool. Tools are `tool()` definitions with zod arguments, one per file under `agent/tools/`. Each file exports a function of the run's deps (`AgentDeps` or `RequestDeps`, both in `agent/sessions.js`) that returns the tool, since the SDK calls a tool with only its arguments. Each run serves its tools with `createSdkMcpServer()`, so the model sees each as `mcp__coding-agent__<name>`. Optional arguments are `.optional()`, and the model leaves them out. Both agents set `tools: []`, `settingSources: []`, and `strictMcpConfig: true`, so they get none of Claude Code's built-in tools and none of the settings, `CLAUDE.md` files, or MCP servers on the machine. `verbatimPrompts: true` keeps an `@path` in a Slack message from reading a local file.
+
+**Runs** call the SDK's `query()` once per run from Bolt listeners, which runs the Claude Code CLI bundled with the SDK. Coding turns allow `MAX_TURNS` model calls, and a turn that uses them all (`error_max_turns`) ends with a message, keeping its history. A `PostToolUse` hook returns `continue: false` to end the run early once there's nothing left to do: after `create_code_channel` succeeds for the routing agent, and after `archive_channel` for the coding agent. A result with `is_error` throws, so the turn posts a warning. Every run also names its MCP server in `--mcp-config`: the SDK registers the server only once the CLI starts, and resuming a deferred call looks for its tool before that.
+
+**Turns** go through `agent/turns.js`, which doesn't import the SDK. A turn that throws posts a warning and moves the session back to `active`, keeping the history from the last finished turn. Model-written replies use `markdown_text`, since the model writes standard markdown.
+
+**Tool errors.** A tool that throws reaches the model as an error result with the error's message, which the SDK's MCP server does for every tool. A Slack API error a tool expects, a `WebAPIPlatformError`, is returned to the model as the tool's result.
+
+**Stopping** sets `session.stopped`, and so does archiving the channel from Slack. The `PreToolUse` hook from `gateToolCalls` (`agent/guardrails.js`) then denies every tool call and tells the model to wrap up, and the turn posts nothing when it finishes.
+
+**Conversation history** lives in the Claude Code session transcript. `session.history` is the list of each run's result message, and the next run resumes from the last one's `session_id`.
+
+**Command approvals.** The `PreToolUse` hook wraps `commandNeedsApproval()` from `agent/approvals.js` and returns `defer` for a `run_command` that needs approval, which ends the run with `deferred_tool_use` on its result. A paused run is a `PausedRun` (its history, ending with that result), parked in `pendingStore` with the ts of its Approve/Deny message, so a click only resumes the run its buttons were posted for. `resumeRun` resumes the session with a prompt that sends nothing: the CLI asks the hook about the waiting call again, and the hook answers with the decision. A newer message in the channel retires a pending approval, so a late click can't resume an older run over it. The next run then finds the waiting call on the last result, and the hook hands it to `answerUnanswered`, the `canUseTool` callback, which denies it as not run and interrupts that old turn, so no model call is spent on it before the new message.
+
+## Testing
+
+Tests use Node's test runner (`node --test`) and fake the Slack client with `fakeClient()` from `tests/helpers.js`, which mirrors the real client's methods. Listener tests stand in for the agent's runs with `t.mock.method(runs, 'startRun', ...)`, and `tests/agent/agent.test.js` stands in for the SDK with `t.mock.method(sdk, 'query', ...)`. `tests/invoke.js` calls a tool through its MCP server, the way a run does. `tests/agent/runs.test.js` runs the real agents and the bundled CLI against `tests/fake-model.js`, a local stand-in for the Anthropic Messages API that returns scripted tool calls and replies, passed to each run through the `env` option, so the SDK seams are tested without calling Anthropic.

claude-agent-sdk/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Slack Technologies, LLC
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.