IssueBridge

🏆 Built for the Slack Agent Builder Challenge (New Slack Agent track) — showcasing Slack's Real-Time Search API and the Model Context Protocol (MCP).

A maintainer agent that searches your GitHub issues and your Slack history — because half your project's answers never made it to GitHub.

Open-source maintainers already have AI triage for GitHub (Dosu and friends). But in active communities like vLLM's, half the project's knowledge lives in Slack: bugs reported informally in #help, workarounds shared in passing, decisions made in a thread that never reaches the tracker. No GitHub-side tool can see that — searching a live workspace in real time wasn't something an external agent could do at all, until Slack's Real-Time Search API. IssueBridge is the agent that closes that gap.

What it does

• New-issue triage. When an issue lands on the repo, IssueBridge checks both halves of the project's memory — the upstream vllm-project/vllm tracker and the community Slack — and posts a triage card in #maintainers with citations on both sides: "this matches closed issue #4821, and it was already answered in #help three days ago."
• One-click cited replies. Each card carries an LLM-drafted maintainer reply citing both sources. Post it as-is, or open it in an editable modal and adjust the wording first. A "Not a duplicate" button records the human override.
• The bridge runs both ways. When a reply is posted to GitHub, IssueBridge drops a note back into the cited Slack thread — the answer's original author learns their fix was reused. Structurally impossible for a GitHub-only tool.
• Ad hoc Q&A. DM the agent or @mention it in a channel; it searches both corpora and answers with a Sources: footer of clickable citations.

The two claimed technologies

Technology	Where it lives	What it does
Slack Real-Time Search API (assistant.search.context)	rts.py	Live workspace search, authorized by the action_token pulled from each message/mention event — the API's event-scoped permissioning model, not a static credential
Model Context Protocol (MCP)	github_mcp.py	GitHub access through the official github/github-mcp-server — issue search on upstream, add_issue_comment on the fork (plain REST available as a fallback)

Architecture

The flow: fork issue → poller (20s) → parallel search of upstream GitHub (MCP) + Slack history (RTS) → Claude Haiku ranks candidates and drafts a cited reply → Block Kit triage card in #maintainers → one click posts the cited comment back to GitHub — and a backlink to the cited Slack thread.

Built with Bolt for Python, Socket Mode, and the Claude Agent SDK (the conversational agent, with custom search tools registered over an in-process MCP server).

Try it

The demo runs in a Slack developer sandbox seeded with realistic community conversations modeled on real, closed vLLM issues. New issues are filed against a fork so the demo never touches the real project's tracker.

1. Ask the agent (DM or @mention):
   • Has anyone hit CUDA OOM errors with tensor parallelism?
   • How do I fix undefined symbol errors after pip install vllm?
   • What's the right max-model-len for AWQ quantization on a 24GB GPU?
2. Watch live triage: file an issue on the fork — a cited triage card appears in #maintainers within ~20 seconds.

Run it yourself

git clone https://github.com/ManasShouche/issuebridge && cd issuebridge
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.sample .env       # fill in ANTHROPIC_API_KEY, GITHUB_TOKEN, FORK_REPO
python check_setup.py     # pre-flight: tokens, channels, fork Issues enabled
python seed/seed.py       # seed the community history (once)
slack run                 # or: python3 app.py with SLACK_BOT_TOKEN/SLACK_APP_TOKEN set

Create the app from manifest.json (Slack CLI or api.slack.com/apps). The search:read.public scope powers Real-Time Search; reinstall after scope changes. Set GITHUB_MCP_URL to route GitHub calls through the MCP server.

Project structure

app.py                  Bolt Socket Mode entrypoint — starts the fork poller
poller.py               Watches the fork for new issues (20s)
dedupe.py               Parallel two-source retrieval + Haiku ranking → verdict JSON
rts.py                  Real-Time Search adapter (assistant.search.context)
github_mcp.py           GitHub via MCP server, REST fallback
cards.py                Block Kit triage cards — citations, stats footer, actions
triage.py               New issue → dedupe → card in #maintainers
agent/                  Claude Agent SDK core + tools (GitHub search, Slack search, emoji)
listeners/              Slack event/action handlers — DM, mentions, buttons, modal, App Home
seed/                   17 seeded community threads modeled on real closed vLLM issues

What's next

Multi-repo support · a weekly digest mode for maintainers who want a pull instead of a push · converting a #help thread into a GitHub issue with one click · a Marketplace-distributable version with per-workspace OAuth install.

License

MIT — see LICENSE.