A2A Operationalization Index — Spine Footers, JSONL Prompts, Model Variants, Permissions

<aside> 🛰️

Single canonical operationalization surface for the 18-page spine. Each row binds one canonical page to its upstream/downstream graph, its next executable step, the runtime where that step executes, the A2A-optimized JSONL prompt to drive that step, and the explicit model variant policy.

Envelope: envelope_version: 0.3.0 (per ‣).

Visibility: workspace can_view (set on this page) — propagate manually to any surface that should be readable mesh-wide. Attestation / secret pages stay restricted.

Anti-fabrication: entries marked SPEC are not yet deployed; entries marked LIVE have been confirmed in the source page. Per Greg × /maru R2, do not assert deployment without operator confirmation.

</aside>

0. Conventions

DVLO plane — one of Relational, Tool-Risk, Execution, or Cross-plane.
Execution location — one of Cloudflare Worker (gc-mcp-server), Cloudflare Worker (gc-diamond-node), Cloudflare Worker (notion-bridge), Cloudflare Worker (retraining-worker), Cloudflare Worker (W-OPS), Local podman (Yennefer repo only), Notion API (via notion-bridge), On-chain (Base 8453), Postman runner, or Operator console.
JSONL prompt — Agent-to-Agent message tape, one JSON object per line, each line a complete message. First line is always {"role":"system",...} with model, dvlo_plane, envelope_version. Subsequent lines alternate user / assistant / tool with intent_class, capsule_event_type where relevant.
Model variant — primary + fallback. Mistral variants are scoped to the Yennefer repo unless explicitly marked mesh-wide. vibe-cli rename is Yennefer-repo only — Skills Activation Manifest, kimi.task.submitted capsule keys, gc-mcp-server transport labels, and standalone Kimi Claw skill row do not propagate the rename.

1. Canonical Model Variant Matrix (mesh-wide)

Role	Primary	Fallback	Reasoning depth	Scope
Greg (architect review)	Claude Opus 4.1	GPT-5	extended	mesh-wide
/maru (constraint reframe)	Claude 3.5 Sonnet	Claude Opus 4.1	standard → extended on dissonance ≥ 0.6	mesh-wide
Pareto (UI conductor)	Claude 3.5 Sonnet	GPT-5	standard	mesh-wide
Kiro (implementation)	Claude 3.5 Sonnet	GPT-5-codex	standard	mesh-wide
Codex (terminal ops)	GPT-5-codex	Claude 3.5 Sonnet	standard	mesh-wide
Foundry (eval / harness)	Claude Opus 4.1	GPT-5	extended	mesh-wide
Yennefer daemon (thermo proof)	DeepSeek-R1-0528	Claude Opus 4.1	extended	mesh-wide (proof verification)
Retraining classifier promotion	Mistral Codestral 2501	GPT-5-mini	standard	mesh-wide
vibe-cli (replaces Kimi)	Mistral Large 2.1 (codestral)	—	standard	Yennefer repo ONLY — does not propagate
W-OPS rollup author	GPT-5-mini	Claude 3.5 Haiku	standard	mesh-wide
gc-synthesis	Claude Opus 4.1	Gemini 2.5 Pro	extended	mesh-wide
diamondNode attest verifier	—	—	n/a (Ed25519 only, no LLM)	mesh-wide
Self-destruct dual-signature	Claude Opus 4.1 + Greg charter	—	extended	guardian-tier only
HAVIS deterministic ensemble	HAVIS/Pareto(e) + Prune(π) + Codex(1) + Kiro + Claude(i) + Gemini CLI(0) + Copilot + Foundry	—	per-agent	mesh-wide (8-agent S24 ensemble)

2. Spine Surfaces — Upstream / Downstream / Next Step / Execution / JSONL / Variant

2.1 — ‣

Upstream: §6.9 rename slip closure (‣) · DVLO — Dyadic Verifiable Lane Orchestration: A Three-Plane Agent Architecture (Genesis Conductor × LidLift × Lane Harness × Edins Dyadic/PEAN × Instinct/DWE) three planes
Downstream: every plugin tool that writes (genesis-conductor-plugin — One-Click MCP Server, Skill & Plugin for diamondNode); §6.17 invariants inherited by gc.intent.submit, gc.intent.execute, gc.memory.store, gc.settlement.settle
Status: LIVE (charter ratified)
Next step: Wire the four invariants (anti-fabrication, brand separation, reversibility, Trace-as-Consent) into gc-mcp-server middleware as a single enforceCharterV617() guard fired before every write tool.
Execution location: Cloudflare Worker (gc-mcp-server) — src/middleware/charter.ts
A2A JSONL prompt: