mwbrooks · Oct 23, 2026
@@ -12,6 +12,10 @@ A monorepo for a **Coding Agent** sample for Slack, built with Bolt for Python.
 
 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.
 
+## Slack plugin
+
+To install the Slack plugin for your coding agent, follow [Claude Code, Codex, Cursor, and Other Coding Agents](./README.md#claude-code-codex-cursor-and-other-coding-agents) in the README.
+
 ## Commands
 
 Run these from inside an implementation directory, such as `openai-agents-sdk/`:
@@ -29,6 +33,10 @@ ruff format --check .
 pytest
 ```
 
+## Code style
+
+No code comments. Make the code explain itself with clear names, tests, or types. Put the reason for a change in the PR description.
+
 ## Repository structure
 
 ```

@@ -0,0 +1 @@
+@AGENTS.md
@@ -2,6 +2,10 @@
 
 Guidance for coding agents working in this directory. Refer to the root `../AGENTS.md` for repository-wide structure and commands.
 
+## Slack plugin
+
+To install the Slack plugin for your coding agent, follow [Claude Code, Codex, Cursor, and Other Coding Agents](./README.md#claude-code-codex-cursor-and-other-coding-agents) in the README.
+
 ## Claude Agent SDK specifics
 
 **Agents (`agent/agent.py`)** are two sets of `ClaudeAgentOptions` with pinned models: a routing agent (`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 async `@tool` functions, one per file under `agent/tools/`, each with a JSON Schema for its arguments. They're served in-process by `create_sdk_mcp_server()`, so the model sees each as `mcp__coding-agent__<name>`. Both agents set `tools=[]`, `setting_sources=[]`, and `strict_mcp_config=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. `verbatim_prompts=True` keeps an `@path` in a Slack message from reading a local file.

@@ -0,0 +1 @@
+@AGENTS.md
@@ -2,6 +2,10 @@
 
 Guidance for coding agents working in this directory. Refer to the root `../AGENTS.md` for repository-wide structure and commands.
 
+## Slack plugin
+
+To install the Slack plugin for your coding agent, follow [Claude Code, Codex, Cursor, and Other Coding Agents](./README.md#claude-code-codex-cursor-and-other-coding-agents) in the README.
+
 ## OpenAI Agents SDK specifics
 
 **Agents (`agent/agent.py`)** are two `Agent`s with pinned models: a routing `request_agent` (`gpt-6-luna`) whose only tool is `create_code_channel`, and the coding `agent` (`gpt-6.1-sol`) with every other tool. Tools are async `@function_tool` functions, one per file under `agent/tools/`, that read `RunContextWrapper[AgentDeps]` (or `RequestDeps`, both in `agent/sessions.py`) for the Slack client and session. Both agents set `ModelSettings(truncation="auto")`, so a long session drops its oldest conversation instead of outgrowing the context window.

@@ -0,0 +1 @@
+@AGENTS.md
@@ -2,6 +2,10 @@
 
 Guidance for coding agents working in this directory. Refer to the root `../AGENTS.md` for repository-wide structure and commands.
 
+## Slack plugin
+
+To install the Slack plugin for your coding agent, follow [Claude Code, Codex, Cursor, and Other Coding Agents](./README.md#claude-code-codex-cursor-and-other-coding-agents) in the README.
+
 ## Pydantic AI specifics
 
 **Agents (`agent/agent.py`)** are two Pydantic AI `Agent`s: a routing `request_agent` whose only tool is `create_code_channel`, and the coding `agent` with every other tool. Neither sets a model. A `SelectModel` capability picks one when a run starts, by which API key is set: Anthropic (`claude-haiku-4-5` to route, `claude-sonnet-5-5` to code) when `ANTHROPIC_API_KEY` is set, else OpenAI (`gpt-6-luna` and `gpt-6.1-sol`). So no model client is created at import. Tools are async functions, one per file under `agent/tools/`, passed with `tools=[...]`. Each takes `RunContext[AgentDeps]` (or `RequestDeps`, both in `agent/sessions.py`) for the Slack client and session, and Pydantic AI reads its docstring for the description the model sees. The coding agent lists `AnthropicCompaction` and `OpenAICompaction`, so a long session compacts its oldest conversation instead of outgrowing the context window. Each only applies to its own provider's models.

@@ -0,0 +1 @@
+@AGENTS.md