The relay packet and the routine-or-judgment table (free sample)

Here are the last ten handoffs between my agents [paste]. Rewrite each as a relay packet.
Mark any field you had to guess, and tell me which missing field caused the most rework.

Read the packets that touch auth, payments or data deletion yourself.

# The relay packet and the routine-or-judgment table (free sample)

A free sample from Cross-Harness Relay by PaperChaseWebb. Works with Grok Bot, Claude Code, Codex, Hermes or any mix of agents.

Every time work moves from one agent to another, the next one starts cold. Whatever was settled in the last chat is gone unless it travels with the work. These two pieces fix most of that.

## The relay packet

Every handoff carries the same packet, as a file or a message.

```
RELAY PACKET
From seat / harness: [planner / harness A]
To seat / harness: [builder / harness B]
Task: [one line]
Approved by owner: [yes, date | not needed because routine]
Inputs: [plan section, ticket id, branch]
Context the next seat lacks: [decisions made in chat, constraints, things already tried]
Done means: [the exact check that must pass]
Out of bounds: [files, systems, actions]
Open questions: [list, or "none"]
Proof so far: [test command and result, links to logs]
```

"Context the next seat lacks" does the most work. "Done means" stops two agents from passing the same task back and forth forever.

## Routine or judgment

Write the route down. An agent left to decide where work goes will decide differently every time.

| Routine (can go to the cheaper or fallback agent) | Judgment (stays with the main agent and you) |
|---|---|
| Scaffolding from an approved plan | Architecture and data model choices |
| Writing tests for existing behavior | Anything touching auth, payments, secrets or permissions |
| Mechanical refactors | Deleting data or changing migrations |
| Simple bug fixes with a failing test already written | Bugs with no clear cause |
| Docs and comments from code | Public copy, pricing, legal wording |