Task Dispatch

Task Dispatch

You manage cross-agent communication between the claw and agents in other repos. Tasks are dispatched via BRAIN_TASKS.md files that live in target repo roots. Repo agents read tasks, do work, and reply -- creating a bidirectional conversation channel.

The user's input is: $ARGUMENTS

Input Detection

1. Dispatch: "send a task to prototype-creator", "next time I open rhoai-prototypes I need to...", "add a task to archie" --> Dispatch mode

2. Check replies: "check agent replies", "any replies from my repos", "what did agents do" --> Reply check mode

3. Scan activity: "what have I been working on", "scan my repos", "project activity", "repo status" --> Scan mode

4. List tasks: "what tasks are pending in prototype-creator", "list tasks for archie" --> List mode

5. Empty: Ask what they want to do.

Configuration

Projects directory: ~/Desktop/

This is where Evan's project repos live (prototype-creator, rhoai-prototypes, archie, automated-usability-testing, my-cursor-claw, etc.).

Dispatch Mode

1. Confirm the repo name and task description with the user before dispatching.
2. Run:

python3 .cursor/skills/task-dispatch/scripts/dispatch_task.py \
  <repo_name> "<task description>" \
  --repos-dir ~/Desktop/ \
  --context "<optional context>"

3. After dispatching, check if the target repo has the Brain Tasks instruction block in its .cursor/rules/ or CLAUDE.md. If not, offer to add it:

"This is the first task dispatched to [repo]. Want me to add the instruction block so the agent there knows how to read and reply to tasks?"

If yes, read scripts/claw_tasks_instruction.md and add it to the target repo's .cursor/rules/ as a new rule file or append to existing CLAUDE.md.

Reply Check Mode

Run:

python3 .cursor/skills/task-dispatch/scripts/dispatch_task.py \
  --check-replies --repos-dir ~/Desktop/

Present unread agent replies grouped by repo. For each reply:

• Show the task and full conversation thread
• Ask how to respond (reply back, mark complete, or note for later)

To reply:

python3 .cursor/skills/task-dispatch/scripts/dispatch_task.py \
  <repo_name> --reply <N> --reply-message "response" --repos-dir ~/Desktop/

To complete:

python3 .cursor/skills/task-dispatch/scripts/dispatch_task.py \
  <repo_name> --complete <N> --repos-dir ~/Desktop/

Scan Mode

Run:

python3 .cursor/skills/task-dispatch/scripts/scan_repos.py \
  --repos-dir ~/Desktop/ --days 7

This scans four sources across all repos:

1. Git commits -- recent commit messages and dates
2. Claude Code sessions -- first user prompt from each session
3. Cursor agent transcripts -- first user prompt from each session
4. BRAIN_TASKS.md -- pending tasks and unread agent replies

Present the output as a structured summary. When repo activity overlaps with calendar/Slack context, call it out explicitly.

List Mode

python3 .cursor/skills/task-dispatch/scripts/dispatch_task.py \
  <repo_name> --list --repos-dir ~/Desktop/

The Conversation Format

BRAIN_TASKS.md uses checkbox tasks with blockquote conversation threads:

- [ ] Add batch evaluation support
> **Claw:** Add batch evaluation support
> *Dispatched on 2026-06-23 14:30*
> Context: running one persona at a time is slow
>
> **Agent:** Implemented batch mode in evaluate.py. Added --batch flag.
> One question: should failed evals halt the batch or continue?
>
> **Claw:** Continue with a warning. Log the failure in the report.
>
> **Agent:** Done. Failed evaluations show a warning icon.
>

• Claw: lines are from this workspace (the dispatcher)
• Agent: lines are from the repo agent (the worker)
• Unread = last speaker is Agent without subsequent Claw response

What NOT to Do

• Don't dispatch without confirming the repo name and task with the user
• Don't modify files in other repos beyond BRAIN_TASKS.md and the instruction block
• Don't assume a repo exists -- the script will error if it's not found
• Don't dispatch to this repo (my-cursor-claw) -- that's circular