Skip to main content
AI & Automation

Hermes To OpenClaw Packet Contract

Defines the executable work packet that Hermes should hand to OpenClaw so the hands receive bounded intent instead of vague chat

April 8, 2026

Hermes To OpenClaw Packet Contract

If OpenClaw is the hands, Hermes must stop talking to it like a vague assistant.

The packet contract is the missing interface between planning and execution.

Purpose

Hermes shapes intent into an executable packet.

OpenClaw accepts, executes, reviews, verifies, or escalates that packet.

The packet exists so agents can work with bounded autonomy instead of inventing scope on the fly.

Packet Rule

No serious engineering work should begin without a packet or a clearly equivalent written task card.

If the packet is missing critical fields, OpenClaw must either:

  • downgrade the work to A0 discovery
  • request the missing fields
  • escalate to huddle or governance when the missing field is actually a trust-boundary question

Required Packet Fields

Every packet should contain:

  • packet_id
  • packet_type
  • source
  • objective
  • autonomy_class
  • scope_in
  • scope_out
  • write_policy
  • acceptance_criteria
  • review_path
  • verification_path
  • dependencies
  • blockers
  • linked_refs
  • delivery_target
  • receipt_required

Field Definitions

packet_id

Stable execution identifier for the work packet.

Format recommendation:

HP-YYYYMMDD-###

packet_type

Allowed types:

  • investigate
  • implement
  • review
  • verify
  • deploy
  • draft-governance
  • repair

source

Where the packet came from:

  • Hermes
  • Board
  • Huddle
  • Council
  • Operator
  • Sprint Lead

objective

One sentence describing the user-visible outcome.

autonomy_class

Must be one of:

  • A0
  • A1
  • A2
  • A3
  • A4

scope_in

Explicitly allowed boundary.

Examples:

  • one module
  • one service
  • one docs surface
  • one sprint card

scope_out

Explicitly forbidden expansion.

Examples:

  • no auth changes
  • no schema changes
  • no canonical vault edits
  • no cross-node rollout

write_policy

State where OpenClaw may write:

  • workspace-only
  • draft-surface-only
  • repo-and-nearby-docs
  • release-surface-with-approval

acceptance_criteria

The conditions that make the packet done.

These must be testable.

review_path

Who performs skeptical review.

Examples:

  • reviewer required
  • reviewer plus architect
  • huddle before merge

verification_path

What evidence is required before the packet may be treated as accepted.

Examples:

  • unit tests pass
  • health check receipt captured
  • docs updated
  • operator signoff required

dependencies

Upstream work or required systems.

blockers

Open conditions that stop implementation.

linked_refs

Relevant proposal, huddle, decision, blocker, or decree references.

OpenClaw must not invent these.

delivery_target

Where work runs or lands.

Examples:

  • local workspace
  • argobox-lite
  • DragonCanyon
  • ArgoMesh capability: build-worker

receipt_required

Defines whether OpenClaw must emit a formal receipt after execution.

Default: true

Execution Rules For OpenClaw

When a packet arrives, OpenClaw should do this in order:

  1. validate the packet fields
  2. map the packet to the correct role lane
  3. check blockers, remands, and write policy
  4. execute only inside scope_in
  5. stop when scope_out would be crossed
  6. emit a receipt with evidence
  7. hand off to review or verification as required

Refusal And Escalation Rules

OpenClaw must not silently proceed when:

  • autonomy_class is missing
  • acceptance criteria are missing
  • the write surface is ambiguous
  • a remand is open on a linked reference
  • the packet asks OpenClaw to settle doctrine, governance, auth, schema, or trust-boundary questions by itself

In those cases, the correct action is escalation, not improvisation.

Execution Receipt

Every non-trivial packet should end with a receipt containing:

  • packet_id
  • status
  • owner_role
  • files_changed
  • tests_run
  • docs_updated
  • assumptions
  • risks
  • follow_ups
  • handoff_target

Minimal Packet Example

packet_id: HP-20260408-001
packet_type: implement
source: Hermes
objective: Replace the remote OpenClaw startup prompt with the hands contract.
autonomy_class: A2
scope_in:
  - /home/node/.openclaw/workspace/AGENTS.md
  - /home/node/.openclaw/workspace/OPENCLAW-HANDS-CHARTER.md
  - /home/node/.openclaw/workspace/OPENCLAW-HERMES-PACKET-CONTRACT.md
scope_out:
  - no openclaw.json seat renaming
  - no canonical vault write-policy expansion
write_policy: workspace-only
acceptance_criteria:
  - remote AGENTS.md points to the hands charter and packet contract
  - remote runtime prompt no longer treats OpenClaw as the whole organism
review_path: reviewer required
verification_path:
  - remote files present
  - role map visible in operator guide
dependencies:
  - SSH access to argobox-lite
blockers: []
linked_refs:
  - OPENCLAW-HANDS-CHARTER
  - OPENCLAW-REMOTE-ROSTER
delivery_target: argobox-lite
receipt_required: true

The Practical Result

Hermes becomes responsible for shaping work clearly.

OpenClaw becomes responsible for executing that work faithfully.

That is how the hands of the collective stay fast without becoming reckless.

openclawhermesargoboxpacketexecutionautonomy