Protocol Model

IGRE uses a small command taxonomy. The taxonomy is intentionally strict because each verb carries different safety semantics.

Verbs

Verb	Reads	Mutates	Purpose
lfetch	local true systems	ledger only	Record local grounding.
fetch	remote true systems	ledger only	Record remote grounding.
intend create-*	ledger fetch artifacts	ledger only	Build a proposed intervention.
intend list/status	ledger intents	nothing	Inspect proposed or attempted interventions.
laction	ledger and local systems	local true systems	Execute a local mutation.
action	ledger and remote systems	remote true systems	Execute a remote mutation.

Intent Anatomy

An intent is action-ready. It may contain one unit or many units. It should include:

• selected tool and action target
• source fetch lineage
• execution plan
• data paths
• success and failure checks
• invariants and blind spots
• rollback classification
• risk profile
• per-unit payloads when the workflow is batch-shaped

Execution Plans

An execution plan makes multi-step behavior explicit:

{
  "steps": [
    {
      "name": "create_remote_record",
      "stage": "ACTION",
      "required": true,
      "mutates": ["remote.records"]
    },
    {
      "name": "verify_remote_record",
      "stage": "FETCH",
      "required": true,
      "reads": ["remote.record.<created_id>"]
    },
    {
      "name": "update_local_cache",
      "stage": "LACTION",
      "required": false,
      "mutates": ["local.cache.records"]
    }
  ]
}

Remote verification reads are fetches even if the action runner performs them internally.

Action Validation

Before mutation, laction and action must validate:

• the intent exists and still lints
• source fetches exist and are fresh
• SDK/catalog versions still match
• the requested command target matches the intent target
• the command action name matches the selected tool
• the intent’s plan allows the side effects about to happen

Any validation error blocks mutation.