slack-programming-bot

A Slack bot that receives coding tasks via @mentions, dispatches them to Claude Code (or a local LM Studio model), and streams progress + results back into the Slack thread. Supports multiple repos and multiple models (Anthropic cloud + local models on LM Studio).

How It Works

@bot fix the failing login tests in PMSS
@bot model=qwen3 investigate this bug
@bot add admin filter to the user list

The bot:

1. Extracts the model (if specified, otherwise uses default)
2. Sends an acknowledgment in the thread
3. Launches Claude Code against the correct repo (determined by the agent from repos.json)
4. Streams progress updates in-place
5. Posts a final summary with what was done

Requirements

• Python 3.11+
• uv
• Claude Code CLI logged in (uses your Claude plan, no API key needed)
• A Slack workspace with a bot app configured (see Slack Setup)

Quick Start

# Install dependencies
uv sync --all-extras

# Copy and fill in your Slack tokens
cp .env.example .env
# Edit .env with your SLACK_BOT_TOKEN and SLACK_APP_TOKEN

# Scan your repos to build the registry
uv run bot-cli scan-repos

# Start the bot
uv run bot-cli run

Configuration

Environment Variables (.env)

Variable	Required	Default	Description
SLACK_BOT_TOKEN	Yes	-	Bot User OAuth Token (xoxb-...)
SLACK_APP_TOKEN	Yes	-	App-Level Token (xapp-...) for Socket Mode
REPOS_BASE_DIR	No	/Users/Shared/github	Directory containing all your repos
OLLA_URL	Yes	-	Olla LLM router endpoint (e.g. http://127.0.0.1:40114/olla/proxy/v1). See /Users/Shared/github/llmhosting/.
DEFAULT_MODEL	No	sonnet	Default model when none specified
TASK_TIMEOUT_SECONDS	No	600	Max time per task in seconds

Model Registry (config/models.json)

Maps model aliases to providers. Add new models by adding a line:

{
  "sonnet": {"provider": "anthropic", "model_id": "claude-sonnet-4-20250514"},
  "opus": {"provider": "anthropic", "model_id": "claude-opus-4-6"},
  "haiku": {"provider": "anthropic", "model_id": "claude-haiku-4-5-20251001"},
  "qwen": {"provider": "local", "model_id": "qwen/qwen3.6-35b-a3b"}
}

• anthropic provider: uses Claude Code CLI directly (your logged-in plan)
• local provider: talks to a local LLM via the Olla router (OLLA_URL) and runs a built-in tool loop (Read, Write, Edit, Bash). model_id is whatever Olla routes — either an alias from Olla's model_aliases config, or a real backend model name.

Repo Registry (config/repos.json)

Generated by scan-repos, then hand-editable. Maps repo names to aliases, keywords, and frameworks:

{
  "predictmystepscore": {
    "path": "predictmystepscore",
    "aliases": ["predictmystepscore", "pmss"],
    "keywords": ["usmle", "score", "step"],
    "framework": "django"
  }
}

The agent uses this to figure out which repo you're referring to.

CLI Commands

# Scan REPOS_BASE_DIR for git repos and generate config/repos.json
uv run bot-cli scan-repos

# Start the bot
uv run bot-cli run

# Check if a model is configured
uv run bot-cli test-model sonnet

Usage in Slack

# Basic — uses default model
@bot fix the login tests

# Specify a model
@bot model=opus refactor the auth middleware

# The bot figures out the repo from context
@bot add stripe webhook handling to PMSS

Thread Conversations

The bot is thread-aware in two ways:

Follow-ups: After the bot completes a task, @mention it in the same thread to continue the conversation. The agent keeps its full context (files, git state, reasoning) across turns — it's the same session, not a fresh start.

@bot fix the failing login tests in PMSS
  → Bot: Fixed 3 tests. Created branch. Want me to push and create a PR?
@bot yes please push it
  → Bot: Pushed and created PR #42

Jumping into existing threads: You can discuss something with teammates in a thread, then drop the bot in mid-conversation. It fetches the full thread history and sees everything that was said.

@John: I think we should refactor the auth middleware to use JWT
@Eric: agreed, here's the file that needs changing: auth.py lines 40-60
@bot make the change suggested above for PMSS
  → Bot sees the full conversation and knows exactly what to do

Note: the bot only responds when @mentioned — it won't react to casual messages in threads.

What the bot posts

1. ACK — "Got it, working on this using sonnet..."
2. Progress — a single message updated in-place as the agent works
3. Summary — what was done, with cost if available

Using Local Models

Local models are served through Olla (config in /Users/Shared/github/llmhosting/), which routes requests across our Tailscale fleet of Ollama / LM Studio / vLLM / llama.cpp boxes. Set OLLA_URL in .env to point at the Olla instance, then add an entry in config/models.json:

"qwen": {"provider": "local", "model_id": "qwen/qwen3.6-35b-a3b"}

Then in Slack: @bot model=qwen investigate this bug.

The bot calls Olla over its OpenAI-compatible endpoint and runs its own tool-execution loop. We tried fronting LM Studio with the LiteLLM Anthropic-bridge so claude-agent-sdk could talk to it transparently, but the bridge mangled tool-call arguments — so non-Anthropic models go through a separate, simpler runner instead.

Project Structure

src/bot/
  config/settings.py   # Pydantic settings from .env
  slack/
    handler.py         # Slack event handling, model extraction
    feedback.py        # Thread updates (ACK, progress, summary)
  executor/
    worker.py          # Claude Agent SDK execution (anthropic provider)
    local_runner.py    # OpenAI SDK + tool loop (local provider, talks to Olla)
  app.py               # Slack Bolt + Socket Mode entrypoint
  cli.py               # CLI commands
config/
  models.json          # Model registry
  repos.json           # Repo registry (generated)

Development

# Install with dev dependencies
uv sync --all-extras

# Run tests
uv run pytest tests/ -v