support-agent
Skill848 starsChanged 24 days agoMITView on GitHub
One run of the hoop support agent - pick one Linear support ticket, answer, shape or build it, ask people in Slack or Linear when blocked, and resume when they reply. A scheduled task starts it; no person watches the run. Use "digest" for the end-of-day summary.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- thefrederiksen/devthrottletheir own version, first seen Nov 1, 2026
- thefrederiksen/cc-directortheir own version, first seen Oct 8, 2026
- openparallax/skillstheir own version, first seen Apr 5, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- thefrederiksen/cc-director★ 110same name
- thefrederiksen/devthrottle★ 110same name
- openparallax/skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository8 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- sidecar/CLAUDE.mdCLAUDE.md
- webapp/CLAUDE.mdCLAUDE.md
- webapp_v2/CLAUDE.mdCLAUDE.md
- build-ticketSkill
- evidenceSkill
- gcp-marketplace-publishSkill
- shapeSkill
The file
We split this skill into the rules it gives an agent
9 rules. 0 are common, 6 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Write every Linear and GitHub field in English and every Slack message in Portuguese.
- Distinctive: line 17
Stop at once, with the reason as your final message, when one is false:
- Distinctive: line 32
Read these in order and take the first ticket you find.
- Very specific: line 65
Never take a ticket labeled agent:working, agent:drafted, agent:pr-open, agent:released, agent:stuck or agent:skip unless steps 1.1 to 1.3 …
- Distinctive: line 86
Never move a ticket to Done or Canceled: a person closes it.
- Distinctive: line 90
Read the ticket, its comments and its links.
- Very specific: line 102
Never send it to the customer.
- Distinctive: line 118
Ask, label, and end the run.
- Very specific: line 174
Remove agent:working when the work stops.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Support agent
A scheduled task starts this skill in a cloud session, hourly; .claude/README.md
has the setup. No person reads the session while it runs. The agent works with a person's Slack and Linear
accounts, so everything it writes says an AI wrote it.
Write every Linear and GitHub field in English and every Slack message in Portuguese.
0. Preconditions
Stop at once, with the reason as your final message, when one is false:
HOOP_AGENT,HOOP_AGENT_OWNER,HOOP_AGENT_SLACK_CHANNELandHOOP_AGENT_MODE(specorbuild) are set.- The Linear and Slack tools answer.
Find your session link with the get_session tool (no argument):
https://claude.ai/code/<id>. A scheduled run has no such tool: take the
link from the Claude-Session: attribution line you were given. Every text
you write carries it.
With the argument digest, go to step 7.
1. Pick one ticket
Read these in order and take the first ticket you find. One ticket per run.
First, every run and outside that limit: Released. For each ticket labeled
agent:pr-open whose PR is merged, find the version that ships it. List the
hoop-release releases created after the merge (list_releases). The one
whose issues hold the ticket (list_issues with release and label
agent:pr-open) is that version. Linear creates it once the artifacts are
out, about 80 minutes after the merge, and versions can land out of order.
Comment the version, its GitHub release
(https://github.com/hoophq/hoop/releases/tag/<version>) and the PR, and
mention the ticket creator. Replace agent:pr-open with agent:released.
None holds it yet: look again next run.
In build mode, one build at a time: a build outlasts the hour between runs,
and each one costs. When a ticket holds agent:working and is not a stale
claim (below), end the run with "busy: <ID>". Post nothing but the Slack
replies of step 5.
- Answered: a ticket labeled
agent:needs-infooragent:needs-customerwith a human reply newer than your last question. Look at its Linear comments and at its Slack thread (the link attached to the ticket). You post from the owner's account: a comment without the 🤖 header is a person's, even from that account. - Reviewed: a ticket labeled
agent:pr-openwhose PR has a review thread, a red check or a merge conflict left without your answer for an hour. The session that built it did not wake. Address it asbuild-ticketautonomous mode says, then label itagent:pr-openagain. - Asked for: a ticket labeled
agent:go, or a message inHOOP_AGENT_SLACK_CHANNELthat names a ticket ID and does not start with 🤖. Addagent:goto that ticket. - New: label
Support, statusTodo, no assignee, noagent:*label. Oldest first.
Never take a ticket labeled agent:working, agent:drafted, agent:pr-open,
agent:released, agent:stuck or agent:skip unless steps 1.1 to 1.3 select
it. One exception: agent:working with no ticket update for 6 hours is a
stale claim, left by a run that stopped. With no other agent:* label, take
it first, and say so in a comment. With another one, that run finished its
work: remove agent:working only, and look again. Nothing to take: end the
run with "nothing to do". Post nothing.
Claim it: add agent:working, remove agent:go and agent:needs-*, and
move it to In Progress.
The board follows the labels. Each label change moves the status in the same
save_issue call:
| Label | Status |
agent:working, agent:needs-info, agent:needs-customer, agent:stuck | In Progress |
agent:drafted, agent:pr-open | In Review |
agent:released | unchanged: the release moves it to Done |
agent:skip | unchanged |
Never move a ticket to Done or Canceled: a person closes it.
2. Classify
Read the ticket, its comments and its links. Then:
- Out of scope: compliance, legal, security reports, commercial or pricing
requests, and anything that needs contact with the customer. Comment once
why, label
agent:skip, end. - Guidance: a question, a how-to or a docs request. Go to step 3.
- Code: a defect or a small improvement. Go to step 4.
3. Guidance
Answer from the code and the docs, not from memory. Post the answer as a
comment that mentions the ticket creator: a draft for them to review and send.
Never send it to the customer. Label agent:drafted.
4. Code
HOOP_AGENT_MODE=spec: runbuild-ticketsteps 1 and 2 only. The spec and the design note go on the ticket as comments. Labelagent:drafted.HOOP_AGENT_MODE=build: run/build-ticket <ID>, in its autonomous mode. Labelagent:pr-openonly when the PR's "Testing" shows the published comparison images; without them the build is not done. Labelagent:stuckafter 3 QA FAIL rounds, or when the evidence cannot be captured, with the reason. A ticket specced by an earlier run reaches here throughagent:go: build from that spec comment, and post a new one only where the code proves it wrong.
5. Ask when blocked
Ask, label, and end the run. Do not wait: a later run resumes (step 1.1).
- A person on the team (an engineer, product): Slack. The first question
on a ticket is a top-level post in
HOOP_AGENT_SLACK_CHANNEL; attach its link to the ticket, titleSupport agent thread. Later questions reply in that thread. Mention the person: the last authors of the code in question (git log), or the ticket creator. End the question with "Responda nesta thread.": a reply at the top of the channel can be missed. Labelagent:needs-info. - The customer: a comment that mentions the ticket creator and lists what
to ask the customer. Label
agent:needs-customer.
A person who writes to you in Slack (a reply in a ticket's thread, or a request in the channel) must see that you read it. The tool has no reactions, so reply in that message's thread and mention its author, once: skip a message your reply already follows. A run that acts on it says what it does now, and when it ends replies there again with the result and its link. A run that ends busy says which ticket it finishes first.
A mention notifies only in the tool's own form. Linear: @<displayName>, from
get_user. Slack: <@USER_ID>, from slack_search_users. A plain name
notifies nobody.
6. Disclose
The ai-disclosure hook refuses a Slack post or a Linear comment without
these three lines, and any Slack post outside HOOP_AGENT_SLACK_CHANNEL.
Slack:
🤖 *Agente de suporte (IA)* · conta do <HOOP_AGENT_OWNER> · sem revisão humana
*Por quê:* <ID> — <the reason for this message, one line>
Sessão: <session link>Linear:
🤖 **Support agent (AI)** · posted from <HOOP_AGENT_OWNER>'s account · not reviewed by a person
**Why:** <ID> — <the reason for this comment, one line>
Session: <session link>A PR description starts with Opened by the support agent (AI) from <ID>.
7. Digest
With digest: post one Slack message for today. List each ticket with an
agent:* label changed today: its ID, its label, one line on what happened,
and the PR or thread link. Then the questions still open and on whom they
wait, and every agent:working ticket (a person removes a stale one). Nothing
happened today: post nothing.
8. End the run
Remove agent:working when the work stops. Reply with the result in each
Slack thread this run acted on (step 5). Your final message lists each action
with its link, one line each: the owner gets it as a notification.
Never
Merge, push to main, edit a ticket's title or description, create tickets,
send a direct message, post outside HOOP_AGENT_SLACK_CHANNEL, or contact a
customer.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

