[zimeg]

Summary

This pull request adds a Claude Agent SDK implementation of the coding agent in claude-agent-sdk/, ported from openai-agents-sdk/ with the Python Claude design from bolt-python-coding-agent#16.

• Matches the OpenAI Agents SDK version's experience: the same 16 tools with the same names, arguments, and descriptions, the same prompts, and the same approvals, stop, and archive behavior. listeners/, the framework-free agent/ files, and the listener, approvals, and workspace tests are copied unchanged, so only agent/agent.js, agent/guardrails.js, and agent/tools/ use the SDK.
• Tools are closure factories, one per file like the starter agent's, built for each run and served by an in-process MCP server. Claude Code's built-in tools are off, and so are the machine's settings, CLAUDE.md files, and MCP servers. verbatimPrompts keeps an @path inside a Slack message from reading a local file. Uses claude-sonnet-5-5 for coding and claude-haiku-4-5 for routing.
• Command approvals use the PreToolUse hook's defer decision. The run ends with the command waiting, and a click resumes the Claude Code session. The hook then allows or denies that same call.
• A dropped approval, from a newer message or the stop button, is answered as not run at the start of the next turn. canUseTool denies it with interrupt, so closing it costs no model call.
• A PostToolUse hook ends the run once a channel is created or archived. The stop button denies every tool call, a throwing tool reaches the model as an error result, and running out of turns ends with a message, keeping the session.
• Names the tool server with --mcp-config as well as the SDK option. The TypeScript SDK registers it too late for the CLI to resume a deferred call, which then fails as "tool not available". The Python SDK already passes the flag itself.
• Adds claude-agent-sdk/ to the lint and tests matrices, dependabot, the root README, and AGENTS.md.

tests/agent/runs.test.js drives the real SDK and bundled CLI against tests/fake-model.js, a local server that answers the Messages API from a script. The CLI's config goes to a temp directory, and the suite takes about 30s.

Testing

• Copy claude-agent-sdk/.env.sample to .env, set ANTHROPIC_API_KEY, and start the app with slack run from claude-agent-sdk/.
• Mention the agent in a thread with a question and confirm it answers in place, then with a task on a throwaway repo and confirm it links a new code channel.
• In the code channel, approve a test or build command and confirm it runs, then deny one and confirm the agent carries on without running it.
• Send a new message while an approval is pending and confirm the buttons show it was skipped and the agent answers the new message.
• Press stop mid-turn and confirm the agent says it stopped and posts nothing more from that turn.
• Say the work is done and confirm the summary is shared back to your original message and the channel is archived.

🤖 Generated with Claude Code

[zimeg]

🗣️ Wordings on the change before this might land?

[zimeg]

👾 note: This handles SDK specific implementations alongside the guardrails.js file. Others are similar across providers.

[zimeg]

🧪 note: Landing this now for coverage but will follow up to align between samples.

[zimeg]

📚 note: Leaving this without change for now but let's follow up to mirror pattern across other projects.

[zimeg]

🏁 Let's get this merged now and adjust ongoing before making this so public.