Human-in-the-loop purchase-completion agent (standalone plugin + checkout skill)

[juniperbevensee]

Summary

Experimental human-in-the-loop purchasing capability, shipped as a standalone, opt-in HSM artifact — zero hermes-agent image footprint, zero fork divergence.

• Plugin infra/templates/plugins/human_escalation/ — registers escalate_to_human + check_pending_escalation tools and a pre_gateway_dispatch resume hook. The escalation bus over chat: relay an SMS/2FA code, confirm a charge, hand off a VNC/QR/link — all as chat messages, replies resolved on the next turn.
• Skill infra/templates/skills/checkout/SKILL.md — reusable, resume-aware checkout playbook with an always-confirm-the-charge gate. Moshtix is instance feat: surfaces P0 — restart, disconnect, edit, pairing visibility #1.
• Docs — image-vs-hsm-boundary.md updated with the default-bundle-vs-optional-artifact distinction, this worked example, and a skill-packaging caveat (file skill vs register_skill).

Design (Approach A′ — no-patch async)

Uses only existing extension points (tool registration, the pre_gateway_dispatch hook contract, the send_message tool, session env). The escalation posts to chat and returns (turn ends); the user's reply is a normal inbound message that the hook rewrites to nudge the resume-aware skill, which continues from the live (persistent) Camofox browser tab. No gateway/agent source patched.

Tests

24 co-located unit/integration tests pass (store / escalate / hook / full resume cycle). Standalone-load proven via a loader replica with no plugins parent package on sys.path.

Test Plan — manual E2E on a real agent (gates merge-to-default; spends real money at the end)

• Install on personal hermes via HSM (plugins.enabled += human_escalation; install the checkout skill). Confirm tools register at boot.
• Smoke: ask it to escalate_to_human for a code → it posts + ends turn → reply → it resumes via check_pending_escalation. (Validates agent reliably re-enters on a bare reply.)
• Moshtix dry-run to the confirmation gate → decline → assert nothing charged.
• Escalate, wait 3–5 min, then reply → assert resume against the still-live Camofox tab. (Tab persistence already validated standalone on the mini.)
• One real small purchase with affirmative confirm → assert confirmation number; capture Moshtix selectors into the skill.
• Verify Signal target routing (platform:chat_id) and the MEDIA: attachment path for a QR/VNC link_handoff.

Residual real-runtime risks (validate in E2E, not blockers)

Signal send_message target routing for colon-containing chat_ids; MEDIA: attachment dir allowance; agent honoring "END YOUR TURN" + calling check_pending_escalation; restart-mid-wait (record persists, browser tab won't); hook fires pre-auth (auth still gates after).

🤖 Generated with Claude Code