# PALO plugin implementation excerpts

These excerpts document the connected draft dated 27 October 2026. They are abbreviated for explanation, not ready-to-deploy configuration. No credential, tenant identifier or client identifier is included.

## Skill routing and evidence boundary

Exact excerpts from palo-agentic-action-design/SKILL.md, selected from the longer file:

```yaml
name: palo-agentic-action-design
description: "Use when designing a future agent action: mandates, delegation limits, Action Claims, Effect Contracts, approval requirements or recovery plans. For an actual timeout, incident or disputed past result, start with PALO Evidence Review."
```

```text
Keep authorization, execution, observed effect and business outcome separate.

Never mint an executable claim, signature, token or approval.
```

These instructions guide drafting behavior. They do not intercept other tools, enforce access control or establish deterministic prevention of sensitive output. The local synthetic no-echo failure illustrates that boundary.

## MCP configuration

Selected fields from mcp.json. The omitted client identity, issuer endpoints and resource must be configured and verified for the actual environment. Scope prefixes are abbreviated. This excerpt parses as JSON but is not a valid complete connection configuration.

```json
{
  "mcpServers": {
    "palo-reader": {
      "type": "streamable-http",
      "url": "https://guide-api.paloframework.org/mcp-guide",
      "extensions": {
        "com.openai": {
          "auth": {
            "type": "oauth",
            "client": {
              "mode": "provided",
              "tokenEndpointAuthMethod": "client_secret_post"
            },
            "baseScopes": [
              ".../palo:guide",
              ".../palo:knowledge:read"
            ]
          }
        }
      }
    }
  }
}
```

## Intended connected workflow

1. A user selects a PALO documentation workflow in the host.
2. The host requests the Reader with the minimum necessary arguments under the intended delegated identity.
3. The Reader checks authentication, the admitted client and the allowed scope before returning guidance or canonical source material.
4. The skill uses the returned sources to prepare a draft, retaining citations and unresolved evidence.
5. An accountable reviewer evaluates the draft. Target business actions require separate authorization and execution controls.

The connected path is not qualified yet. Successful authenticated discovery and actual calls under the intended ChatGPT identity must be demonstrated before describing this sequence as operational.