Dispatch

dispatch

Read AGENTS.md, then read .agent-teamflow from the repo root, then follow the workflow below. Treat the user's remaining request text as $ARGUMENTS.

---

Split a brain dump into parallel issues for multiple team members, write a workflow file, and push.

Run this workflow in an isolated worker if the current agent runtime supports one; otherwise run it in the main session. Keep intermediate docs/CLI output concise and report only the final result (workflow file path, issue URLs, commit/push status).

What this command does

You are distributing today's work across multiple team members. The user gives you a brain dump. You decide who does what, write a workflow file, file issues, and push.

Setup

Read .agent-teamflow from the repo root. Extract issueTracker, project, owners, and paths.workflowDir (default docs/workflow if missing) — refer to it as `` below.

The owners map keys are the assignee aliases (e.g. "alice", "bob"). Values are their integration branch names — but for dispatch, you primarily need the keys as assignee identifiers.

Inputs

$ARGUMENTS — the brain dump, either inline text or a file reference.

• If $ARGUMENTS starts with /, ~/, or ./, or ends in .md/.txt, Read it and use the file contents as the brain dump.

Tasks may include explicit assignment hints (e.g. "alice: fix X", "bob does Y"). Honor them. Match case-insensitively against owners map keys. If no hints, distribute by area of expertise inferred from recent file activity, or alternate round-robin as fallback.

Hard constraints

• No cross-assignee blockers. Tasks within one assignee may be sequential; tasks across assignees must be independent. If the brain dump implies a cross-assignee dependency, restructure so the dependent piece lives entirely with one assignee, or flag it in the workflow file under ## Dependencies (manual coordination) and do NOT create the blocked issue yet.
• No emojis anywhere — workflow file, issue titles, bodies, commit messages.
• Read before write. If the project has per-area context docs (check AGENTS.md or a docs/ directory), read the relevant ones before writing issue bodies. Issue bodies must reference actual files, not vague "update the X page" phrasing.

Execution

1. Parse the brain dump

• Extract individual tasks.
• Detect explicit assignee hints. Record them.
• For tasks without hints, decide based on recent file activity in per-owner areas, or round-robin.
• Verify the no-cross-assignee-blocker constraint. Restructure if violated.

2. Read relevant context docs

If the project has per-area context docs, read the ones relevant to each task before writing issue bodies.

3. Write the workflow file

Path: /workflow-.md (use local time: date +%Y%m%d-%H%M).

Create / if it doesn't exist.

Structure:

# Workflow 

## Brain dump (source)

## Split

### 

1. **** — 
   - Touches:  / 
   - Acceptance: 
   - Issue: 

2. ...

### 

1. ...

## Dependencies (manual coordination)

## Notes

4. Create issues

For each task, create sequentially:

# GitLab
glab issue create --title "" --description "" --assignee 

# GitHub
gh issue create --title "" --body "" --assignee 

Issue body must contain:

• Context: which area
• What to do: concrete description referencing actual files
• Acceptance criteria: bullet list
• Source: relative path to the workflow file

Capture the returned issue URL for each task.

5. Patch the workflow file with issue URLs

After all issues are created, edit the workflow file and fill in the Issue: lines with the returned URLs.

6. Commit and push

git add /workflow-.md
git commit -m "chore(workflow): dispatch  tasks"
git push origin HEAD

Do not touch integration branches, staging, or main.

7. Report

• Workflow file path
• Per-assignee bullet list of issues with URLs
• Commit SHA + branch pushed
• Any assumptions or flagged dependencies

Error handling

• If an issue creation fails: keep going for the rest, note the failure in the final report. Do NOT roll back.
• If the workflow file already exists at the exact timestamp: append -2, -3, etc.
• If git push fails: report the error and the local commit SHA; do not retry blindly.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: lkim0402
• Source: lkim0402/agent-teamflow
• License: MIT

Install and usage instructions live in the source repository linked above.