🐙 Demo 2 — GitHub issue triage (agent run)

Goal: ask the agent to look at the open issues in a GitHub repo, fetch the repo's homepage to understand what the project is about, and produce a prioritized triage summary — without writing a single line of glue code.

Time to complete: ~10 minutes. Tools used: github, html_query. Surface: Chat (or Agent Probe on the Tools page).

🎯 Skip the typing — use the built-in template. Open Jobs → Job Templates and pick GitHub issue triage (id github-issue-triage). It seeds the prompt and lists GITHUB_TOKEN as a required secret so the UI can flag it. You'll still need to replace the OWNER/REPO placeholder.

💡 Why this scenario? It's the smallest demo that shows two new built-in tools cooperating. Swap the repo + URL and you have a generic "look at this project and tell me what to do" pattern.

Step 0 — (Optional) Set the GitHub token secret#

Anonymous GitHub calls work but you'll hit the 60 requests/hour rate limit quickly. Set a token (any read-only PAT will do) so the github tool authenticates automatically.

$body = '{ "value": "ghp_yourTokenHere", "description": "Read-only PAT for the github tool" }'
Invoke-RestMethod -Method Put `
  -Uri http://localhost:5010/api/secrets/GITHUB_TOKEN `
  -ContentType 'application/json' -Body $body

Verify (the list endpoint never returns plaintext):

Invoke-RestMethod http://localhost:5010/api/secrets
# → name: GITHUB_TOKEN, description: "Read-only PAT ...", updatedAt: ...

The github tool reads the secret at execute time and falls back to the GITHUB_TOKEN environment variable if no secret is set. See ACKNOWLEDGMENTS.md and the secrets section of 20-tools.md for details.

Step 1 — Open the chat#

Open the Web UI (http://localhost:5000) and click Chat in the sidebar. Make sure your active agent profile has both github and html_query enabled (they are by default for the system profile).

Step 2 — Send the prompt#

Paste this into the chat box and send:

Triage the open issues in elbruno/openclawnet-plan.
 
1. Use the github tool (action="list_issues", state="open", perPage=20) to fetch open issues.
2. Use the html_query tool to fetch https://github.com/elbruno/openclawnet-plan
   with selector "article" and limit 1, so you understand what the project is about.
3. Group the issues into three buckets and explain your reasoning briefly:
   - 🔥 Critical (block users / data loss / security)
   - 🟡 Important (visible bugs, missing features)
   - 🟢 Nice-to-have (polish, docs, refactors)
4. Pick the top 3 issues to work on next, with one-line justifications.
5. End with a one-sentence summary of the overall repo health.

You should see the agent stream three tool calls (you'll see them in the chat timeline):

github(action=list_issues, ...) returns a Markdown list of open issues.
html_query(url=https://github.com/elbruno/openclawnet-plan, selector=article) returns the rendered README content.
The agent then reasons over both pieces and writes the triage report.

Step 3 — (Alternative) Run from the Tools page in Agent Probe mode#

If you want to invoke just one tool at a time:

Open the Tools page in the sidebar.
Click the github card → Agent Probe tab → use the prompt:

"List the 10 most recent open issues in elbruno/openclawnet-plan"
Click the html_query card → Agent Probe tab → use the prompt:

"What does the H1 of https://elbruno.com say?"

Agent Probe runs the request through a real agent run with the single tool selected, so you see the prompt → tool call → response chain end-to-end.

Step 4 — Try the variations#

A few one-line tweaks that show how composable this is:

"Same triage but for pulls instead of issues" — the agent switches action to list_pulls.
"Now fetch the README of elbruno/elbruno.localembeddings with the github tool (action=get_file, path=README.md) and tell me what model it uses by default" — combines get_file with reasoning.
"Get the title (title selector) and meta description (meta[name=description], attribute=content) of three recent posts on https://elbruno.com" — multiple html_query calls in one run.

Where to go next#

Wire this same triage prompt into a scheduled job ("every Monday at 09:00") so you get a weekly digest. The Jobs UI takes the exact same prompt — see Demo 1.
Add the markdown_convert tool to the allowed list and ask the agent to also fetch and summarize the linked design doc of any issue that mentions one.
Read the full reference for the new tools in docs/manuals/20-tools.md.