escalation-hitl-patterns

name	escalation-hitl-patterns
description	Escalation triggers, HITL deployment patterns (co-pilot, shadow, AI-first-with-escalation), the human approval artifact contract, and handoff packet requirements. Owned by Hermes; consumed by Hestia.
user-invocable	false
argument-hint	<evaluate-trigger|build-packet|record-approval>
allowed-tools	["Read","Write","Grep"]

Escalation & HITL Patterns

Humans are collaborators, not fallbacks. Escalation is a terminal state and a success of the covenant — the moment the house correctly recognizes a matter beyond its authority.

Purpose

Define when the workflow crosses to humans, how it crosses (packet requirements), and the only mechanism by which human authority flows back in (the approval artifact).

When to use

• Hermes on every trigger evaluation, packet build, and approval record.
• Hestia when deciding dispatch and terminal states.
• /escalate and /support-ticket commands.

The three HITL deployment patterns

Pattern	What it means	When
Co-pilot	AI drafts; human agent reviews and sends	walk phase; high-risk intents
Shadow	AI runs the full graph offline on live/past tickets; humans compare	crawl phase; /support-shadow
AI-first with escalation	AI resolves low/medium complexity; escalates on triggers	run phase; the default production mode

Promotion between patterns follows deployment-roadmap criteria — never promote on enthusiasm.

Escalation triggers (canonical list)

1. Explicit human request (Article I — immediate, no persuasion).
2. Regulatory flag (jurisdictional right-to-human, sector mandates).
3. Monetary/irreversible action without an approval artifact (Article V).
4. Low confidence: thin retrieval coverage, low self-assessment, or two failed Themis cycles.
5. Sustained negative sentiment (2+ turns) despite de-escalation.
6. SLA breach risk: P1 at 45 minutes unresolved.
7. Pipeline failure: Themis or Eunomia unable to run (fail closed).

Precision matters as much as recall: false escalations train humans to ignore true ones (escalation-correctness rubric).

The handoff packet

A customer must NEVER have to repeat themselves after escalation. The packet (schema in hermes.md) always carries: the portable-context token, a history digest, every attempted action with executed: true|false flags, the KB passages already consulted (cited), the named trigger, and the squad's recommendation. Eunomia clears the packet before it crosses the boundary.

The approval artifact contract (Article V)

The ONLY mechanism by which monetary/irreversible authority enters the squad:

# hearth/approvals/APPROVAL-<ticket-id>-<seq>.yaml
approval_id: APPROVAL-<ticket-id>-<seq>
ticket_id: ...
action: <refund | credit | plan-change | cancellation | deletion>
scope: <exact object>
amount_limit: <number | null>
issued_by: <named human>
issued_at: <ISO-8601>
expires_at: <ISO-8601, default +24h>
status: approved | denied

• Issuer: a human, via /hydra:approve (orchestrated) or explicit in-chat confirmation that Hermes records verbatim (standalone).
• Consumers: the pre-tool-privilege.ps1 hook and Hermes's PreToolUse gate. Both require match on ticket_id + action + scope, status approved, and expires_at in the future.
• Deny-by-default: absent, expired, mismatched, or denied = the action does not execute, in every mode, degraded or not.

Failure modes

• Bot wall: containment metrics suppressing trigger 1. Forbidden; Themis scores it as false deflection.
• Context-poor handoff: a packet missing attempted-actions flags forces the human to re-derive state — escalation-correctness fails on packet-completeness.
• Approval drift: reusing an approval for a second action or a larger amount. The artifact binds to exact action + scope + limit.