ogp-agent-comms（OGP Agent 通讯）

Prerequisites

The OGP daemon must be installed. If you see errors like 'ogp: command not found', install it first:

npm install -g @dp-pcs/ogp
ogp-install-skills
ogp setup
ogp agent-comms interview
ogp config show

Full documentation: https://github.com/dp-pcs/ogp

OGP Agent-Comms Configuration

This skill is an interactive wizard for configuring how your agent handles incoming agent-comms messages from federated peers.

When to Use

Use this skill when:

• User wants to set up agent-to-agent communication policies
• User wants to configure what topics their agent can discuss with peers
• User wants different policies for different peers
• User says things like "configure agent comms", "set up agent communication", "how should my agent respond to X"

Overview

Agent-comms policies control HOW your agent responds to incoming messages (separate from scope grants which control WHETHER messages are allowed).

Two layers:

1. Scope grants (doorman) - Controls which intents/topics are ALLOWED
2. Response policies (this skill) - Controls HOW your agent RESPONDS
3. Human delivery preferences - Controls WHERE and WHEN your agent surfaces federated requests back to the human

Important default: when a federation request is approved, OGP auto-enables the general topic for that peer. Everything else still needs explicit policy if the user wants something more restrictive or more open.

Multi-Agent Routing (v0.2.28+): When using notifyTargets in your OGP config, federation messages can be routed to specific agents based on the message context. Each agent can have its own agent-comms policies.

Human delivery behavior (v0.4.2+): Use humanDeliveryTarget plus inboundFederationPolicy.mode to tell the agent whether it should forward everything, summarize, act autonomously, or wait for approval before replying.

OGP now evaluates delegated-authority rules—global defaults, per-peer overrides, message-class-specific rules (agent-work, human-relay, approval-request, status-update), and topic overrides—before finalizing delivery behavior. This determines whether the agent may reply directly or must escalate to the configured human session. On OpenClaw, /hooks/agent is the preferred delivery path; it tries to request the actual human session key only when hooks.allowRequestSessionKey=true and the requested key matches configured prefixes. When the flag is false, the hook runs in the default session (agent:main:main), so Telegram sender identity continues to appear as injected/system/cli, which is the accepted v0.4.2 limitation documented in the README and changelog. Lifecycle events such as federation approvals now surface through these same hook-driven turns, so humans see the lifecycle update without needing to poll another channel.

On OpenClaw, human-facing federated work should be handled through the platform's hook-driven agent turn (/hooks/agent). Direct sessions.send injection is a lower-level fallback or sync mechanism, not the primary behavioral path.

Multi-Framework note: Policies are framework-local. Use ogp --for openclaw ... for OpenClaw state and ogp --for hermes ... for Hermes state.

Interactive Flow

When invoked, guide the user through this flow:

Step 1: Check Prerequisites

# Verify OGP is running and has peers
ogp config show
ogp --for openclaw status
ogp --for openclaw federation list --status approved

If no approved peers, inform the user they need to federate first.

Step 2: Re-run the Delegated-Authority Interview First

When the user is configuring overall human-delivery or autonomy behavior, start with the canonical interview command instead of approximating it manually:

ogp --for openclaw agent-comms interview

Use the Hermes variant when the active framework is Hermes:

ogp --for hermes agent-comms interview

This command updates:

• humanDeliveryTarget when the framework supports it
• inboundFederationPolicy.mode
• delegated-authority defaults for human surfacing, relay handling, approval-required topics, and trusted-peer autonomy

After the interview, continue with per-peer or per-topic policy commands only if the user wants something more specific.

Step 3: Show Current Configuration

# Show current policies
ogp --for openclaw agent-comms policies

Step 4: Ask What to Configure

Present options:

1. Global defaults - Apply to all peers (current and future)
2. Specific peer(s) - Configure individual peers
3. View current policies - Just show what's configured

Step 5: For Specific Peers - Multi-Select

If configuring specific peers, show the list and allow multi-select:

Use the peer list from:

ogp --for openclaw federation list --status approved

Prefer the peer ID shown in the list. Some commands also accept a display-name match, but the peer ID is the safest input.

Example interaction:

Select peers to configure:
  [x] Stanislav (302a300506032b65)
  [ ] Leonardo (5f8b2c...)
  [x] Alice (9d4e1f...)

Selected: Stanislav, Alice

Step 6: Configure Topics

Ask which topics the agent should engage on:

Which topics should your agent discuss with these peers?
  [x] memory-management
  [x] testing
  [x] general
  [ ] calendar (add custom)
  [ ] personal (add custom)

Add custom topic: _______

Step 6: Configure Response Level

For each topic (or all topics), set the response level:

Step 7: Save Configuration

# Save policies for selected peers
ogp --for openclaw agent-comms configure <peer-id> \
  --topics "memory-management,testing,general" \
  --level full

# Or set global defaults
ogp --for openclaw agent-comms configure --global \
  --topics "general,testing" \
  --level summary

Step 8: Confirm and Show Result

# Show the updated configuration
ogp --for openclaw agent-comms policies <peer-id>

CLI Commands

View All Policies

ogp --for openclaw agent-comms policies

Shows global defaults and per-peer overrides.

View Peer Policy

ogp --for openclaw agent-comms policies 302a300506032b65

Shows effective policy for a specific peer (global + overrides).

Configure Global Defaults

ogp --for openclaw agent-comms configure --global \
  --topics "general,testing" \
  --level summary \
  --notes "Default: be helpful but don't overshare"

Configure Specific Peer

ogp --for openclaw agent-comms configure <peer-id> \
  --topics "memory-management,testing,general" \
  --level full \
  --notes "Stan is a trusted collaborator"

Configure Multiple Peers at Once

ogp --for openclaw agent-comms configure 302a300506032b65,5f8b2c,9d4e1f \
  --topics "testing" \
  --level full

Note: Prefer peer IDs. Partial peer IDs or display-name matches may work, but peer IDs are canonical.

Add Topic to Existing Policy

ogp --for openclaw agent-comms add-topic <peer-id> calendar --level escalate

Remove Topic

ogp --for openclaw agent-comms remove-topic <peer-id> personal

Reset to Global Defaults

ogp --for openclaw agent-comms reset <peer-id>

Policy Inheritance

1. Global defaults apply to all peers
2. Per-peer policies override globals for that peer
3. Topic-level settings are the most specific

Example:

Global: { "general": "summary", "testing": "full" }
Stan (302a300506032b65):   { "memory-management": "full" }

Effective for Stan:
  - general: summary (from global)
  - testing: full (from global)
  - memory-management: full (from Stan-specific)

Response Policy Schema

Stored in the active framework's peers.json under each peer:

{
  "id": "302a300506032b65",
  "alias": "Stanislav",
  "responsePolicy": {
    "memory-management": {
      "level": "full",
      "notes": "Stan is working on similar architecture"
    },
    "testing": {
      "level": "full"
    },
    "calendar": {
      "level": "escalate",
      "notes": "Ask me before sharing schedule"
    }
  }
}

Note: The alias field (formerly petname) is stored with the peer, but agent-comms commands should still prefer the peer ID shown by ogp federation list.

Global defaults live in the active framework config, for example ~/.ogp/config.json or ~/.ogp-hermes/config.json:

{
  "humanDeliveryTarget": "telegram:123456789",
  "inboundFederationPolicy": {
    "mode": "summarize"
  },
  "agentComms": {
    "globalPolicy": {
      "general": { "level": "summary" },
      "testing": { "level": "full" }
    },
    "defaultLevel": "summary",
    "activityLog": true
  },
  "notifyTargets": {
    "main": "telegram:123456789",
    "scribe": "telegram:987654321"
  }
}

How Your Agent Uses These Policies

When an agent-comms message arrives:

1. Doorman checks if the intent/topic is allowed (scope grants)
2. Your agent receives the message via the local platform backend
3. Your agent looks up the response policy:

- Check peer-specific policy for this topic - Fall back to global policy for this topic - Fall back to defaultLevel

1. Your agent responds according to the level:

- full: Engage openly - summary: Brief, high-level response - escalate: "Let me check with my human and get back to you" - deny: "I'm not able to discuss that topic"

OpenClaw delivery path: OGP should hand human-facing federated work to OpenClaw through /hooks/agent, letting the agent run a real turn and deliver via the configured human channel. A compact sessions.send note may still be used to keep session state synchronized, but that is secondary.

Multi-Agent Routing: When notifyTargets is configured, messages are routed to the appropriate agent who then applies their own policies.

Human Delivery: When humanDeliveryTarget is configured, the agent should treat "tell my human X" as a delivery obligation to that configured channel, not merely as something to mention in whatever session happens to be active.

Activity Logging

When enabled, all agent-comms interactions are logged:

# View activity log
ogp --for openclaw agent-comms activity

# View for specific peer
ogp --for openclaw agent-comms activity <peer-id>

# View last N entries
ogp --for openclaw agent-comms activity --last 20

Log format:

2026-03-23 11:52:14 [IN]  Stanislav → testing: Hello from Stan!
2026-03-23 11:52:15 [OUT] → Stanislav: Hi Stan! Test received successfully.
2026-03-23 11:55:22 [IN]  Leonardo → calendar: What's David's availability?
2026-03-23 11:55:23 [OUT] → Leonardo: [ESCALATED] Checking with David...

Example Configurations

Trusted Collaborator (Full Access)

ogp --for openclaw agent-comms configure 302a300506032b65 \
  --topics "memory-management,testing,general,code-review" \
  --level full \
  --notes "Trusted peer, full collaboration"

Business Contact (Limited)

ogp --for openclaw agent-comms configure 5f8b2c... \
  --topics "general,status-updates" \
  --level summary \
  --notes "Professional contact, keep it high-level"

New Federation (Cautious)

ogp --for openclaw agent-comms configure --global \
  --topics "general,testing" \
  --level escalate \
  --notes "Default: check with human for new peers"

Human Delivery Modes

These now have a first-class CLI interview via ogp agent-comms interview, and they remain part of the same operational policy:

• forward — forward all inbound federated requests/replies to the human
• summarize — summarize and escalate only important/actionable items
• autonomous — act independently unless blocked or explicitly told to relay something
• approval-required — do not act or reply until the human approves

Make the user's intent explicit during setup and documentation. Examples:

• "Tell me everything that comes in from peers."
• "Summarize most things, but escalate if you need me."
• "Act on your own unless you are blocked."
• "Never reply to a peer without clearing it with me first."

Multi-Agent Setup

With notifyTargets configured in ~/.ogp/config.json:

{
  "notifyTargets": {
    "main": "telegram:123456789",
    "scribe": "telegram:987654321"
  }
}

Each agent can have independent policies. The main agent might have full access for operational topics, while the scribe agent handles content-related discussions.

Troubleshooting

Agent not following policies

1. Verify the policy is saved:

   ogp --for openclaw agent-comms policies <peer-id>

1. Check if the topic is in scope grants (doorman):

   ogp --for openclaw federation scopes <peer-id>

1. Restart daemon to reload config:

   ogp --for openclaw stop && ogp --for openclaw start --background

Policy not taking effect for new peer

New peers inherit global defaults. Configure them specifically:

ogp --for openclaw agent-comms configure 302a300506032b65 --topics "..." --level "..."

Multi-Agent Routing Issues

If notifications aren't reaching the right agent:

1. Check notifyTargets and humanDeliveryTarget in the active framework config
2. Verify the target format: telegram:chat_id or a raw session key like agent:main:telegram:direct:<chat-id>
3. For OpenClaw, verify hooks are enabled and the hook token is present
4. For OpenClaw Gateway RPC debugging, use wss://localhost:18789 against a TLS-enabled local gateway, not ws://
5. Remember that a successful direct session injection is not the same thing as the agent correctly handling a human-delivery obligation