upgrade

Ran tool The v0.21.0 release of mcp-go brings in a handful of features and fixes that we can immediately adopt to clean up and harden our Slack MCP client:

1. Upgrade to v0.21.0 – Bump your go.mod to require github.com/mark3labs/mcp-go v0.21.0. – Run go get github.com/mark3labs/mcp-go@v0.21.0 && go mod tidy.

2. Unified Transport Interface – The official client now exposes a transport‐agnostic interface that works over SSE, HTTP, or stdio. – We should refactor our internal client code to use the new transport.Interface instead of separate SSE vs. stdio logic. That will simplify switching modes and make our code more resilient to future additions (e.g. WebSockets).

3. omitempty on Tool Inputs – They fixed the issue where empty tool inputs came through as null by tagging maps with omitempty. – Audit our own tool‐input structs (in internal/llms/… or wherever you define ToolInput) and add json:"…,omitempty" to any optional fields so downstream clients (e.g. Claude Desktop) don’t silently skip tools.

   Example:

   type HelloToolInput struct {
     Name string `json:"name,omitempty"`
   }

4. DefaultArray Support – v0.21.0 adds DefaultArray so you can specify defaults for array parameters. – If any of your tools accept array arguments, declare them with the new helper so you don’t have to write manual nil checks.

5. Tool Annotation Metadata – They introduced a way to annotate tools with rich metadata (id, name, description, etc.). – We can decorate our registered tools with these annotations so that any UI (Cursor, Windsurf) can render nicer tool listings automatically.

6. Improved SSE Lifecycle – The new client fixes leaks, adds a cancelSSEStream helper, and renames done→closed. – Replace any direct SSE usage in our code with the new client’s stream API to get correct cleanup and timeout handling “for free.”

7. New Error Helper – There’s now a NewToolResultErrorFromErr constructor that wraps underlying errors cleanly. – Use this in your tool‐invocation code so you surface tool failures with proper context.

Next Steps

1. Update go.mod and imports.

2. Run go build / go test and fix any API‐breaking changes (there should be minimal; most of this is additive).

3. Replace custom transport logic with the unified transport interface.

4. Add omitempty tags and default‐array values where appropriate.

5. Verify that your Slack bot still passes the CI/CD pipeline (lint, tests, Docker, GoReleaser).

By pulling in v0.21.0 you’ll get a more robust, cleaner, and feature‐complete foundation—making it easier to maintain and less error‑prone when connecting to multiple MCP servers over different transports.

├── internal
│   ├── common
│   │   ├── errors
│   │   │   ├── app_errors.go
│   │   │   └── errors.go
│   │   ├── http
│   │   │   └── client.go
│   │   ├── logging
│   │   │   └── logger.go
│   │   └── types.go
│   ├── config
│   │   └── config.go
│   ├── handlers
│   │   └── llm_mcp_bridge.go
│   │   ├── handler.go
│   │   ├── gateway_handler.go
│   │   ├── registry.go
│   ├── llm
│   │   ├── langchain.go
│   │   ├── openai.go
│   │   ├── provider.go
│   │   └── registry.go
│   ├── mcp
│   │   ├── client.go
│   │   ├── mcp.go
│   │   └── server.go
│   └── slack
│       ├── client.go
│       └── llm_client.go

## Post-Restructuring Improvements

After restructuring the codebase to have a cleaner, more maintainable directory structure, there are still some code duplications that need to be addressed:

### Eliminate Duplication between Slack Client and LLM Implementation

The `internal/slack/client.go` file contains duplicate implementations of OpenAI API structures and direct API calls, when we've already consolidated this functionality in the `internal/llm` package. To fix this:

1. **Remove Duplicated OpenAI Structures**: 
   - Remove `openaiMessage`, `openaiRequest`, `openaiChoice`, and other OpenAI-specific structures from `slack/client.go`
   - Use the consolidated implementations from the `internal/llm` package instead

2. **Leverage the LLM Gateway Handler**:
   - Modify the Slack client to use the `LLMGatewayHandler` from `internal/handlers/gateway_handler.go` for all LLM operations
   - Replace direct API calls to OpenAI with calls through our abstraction layer
   - This centralizes LLM provider management and simplifies future changes

3. **Simplify LLM Interface in Slack Client**:
   - Refactor `callLLM` method in the Slack client to delegate to our gateway handler
   - This eliminates code duplication and ensures consistent handling across providers

### Implementation Plan

1. Create a lightweight LLM client in `internal/slack/llm_client.go` that wraps our gateway handler
2. Update `internal/slack/client.go` to use this client instead of direct API calls
3. Remove all OpenAI-specific code from the Slack client
4. Ensure all LLM operations go through our abstraction layer

This will complete the restructuring by ensuring we have a truly unified approach to LLM operations throughout the codebase.