Consolidated. This note's findings now live in ../MIN-STUDIES.md §12. The original is kept for its measurements; do not re-derive from it.

Open decisions for Young — ALL CLOSED

✅ D1–D5 are all CLOSED, and so are the four earlier rulings.

Re-verified individually against live state on 2026-10-10. Each was checked against the filesystem, not carried forward from this document.

Full records: young-rulings-2026-10-10-four.md · young-rulings-2026-10-10-second.md

D1–D5 — status

Nothing in D1–D5 remains open.

The six still genuinely open — all named, all owned

⚠️ Item 1 carries a caution from the vault's own rule: "a deletion committed here propagates there." The vault's prod remote points at the production host, which LAB rules put off-limits for Min. So .solos's destination needs a deliberate choice, not a copy of the vault's remote list.

The four earlier rulings (for the record)

The second set

D1 — Where does the registry live?

Options

• cortex/registry/agents.yaml — a declarative file. The registry is data.
• A Python module (cortex/registry/registry.py) — the registry is code.
• cortex/personas/AGENT_IDENTITIES absorbs the role and cortex/registry is retired.

Recommendation: a declarative file. Identity, ports, lanes, skills and tools are configuration, not logic. A file can be read by anything — Python, a shell script, the Hermes config generator, a test — without importing the app. V2's cortex/registry was Python and nothing imported it.

Open question: cortex/registry/registry.py already exists (orphaned). Keep, absorb, or delete it?

D2 — Where do tasks live?

Options

• Move the whole task store into cortex/tasks/ (V3 owns everything).
• Keep a database as the store; cortex owns the contract (vocabulary + wire mapping).

Recommendation: the contract, not the store. V2's core/queue.py is the one component that works correctly — atomic claim with BEGIN IMMEDIATE, a sound state machine, separation-of-duties enforcement. Relocating it is a large refactor of working code to satisfy a diagram. Cortex should own the states and the mapping; the storage engine is an implementation detail.

D3 — How is a skill assigned to an agent?

Options

• By plane — brain skills to brain agents, heart to heart. Simple.
• By agent — an explicit list per agent. Precise, verbose.
• By capability — a skill declares capabilities; an agent declares capabilities; they match. Flexible, needs both sides maintained.

Recommendation: plane by default, with a per-agent override. Plane covers most cases and cannot drift (it is one field). The override handles the genuine exceptions — only atlas gets fleet-reconcile.

D4 — SDK migration: replace or run both?

Options

• Full replacement — retire the hand-rolled server in one move.
• Run both, prove with the Inspector, then migrate per agent.

Recommendation: both, then migrate. The Inspector should prove the SDK version is more compliant before the fleet depends on it. That converts the migration from a belief into a measurement — and it is exactly the discipline whose absence produced V2's non-spec states.

Sol's finding makes this materially safer than I knew

Verified 2026-10-10 in the SDK's own source: a2a.compat.v0_3.JsonRpc03Adapter accepts our exact current method names.

src/a2a/compat/v0_3/types.py:1559   method: Literal['message/send'] = 'message/send'
src/a2a/compat/v0_3/types.py:1212   method: Literal['tasks/get'] = 'tasks/get'
src/a2a/compat/v0_3/types.py:1125   method: Literal['tasks/cancel'] = 'tasks/cancel'
src/a2a/compat/v0_3/types.py:979    method: Literal['tasks/resubscribe'] = 'tasks/resubscribe'

So the migration is a compatibility flag, not a rewrite. Nothing that speaks to Heart must change on day one — which removes the main risk of running both.

This strengthens the recommendation: adopt the SDK with v0.3 compat on, prove conformance with the Inspector, then flip to v1.0 method names and re-prove. Two proofs instead of one leap.

D5 — Phase 0 first?

Phase 0 is the registry, and it is independent of the SDK work.

Recommendation: yes, immediately. It is small, it is load-bearing, and it fixes the one defect V3 must not reproduce — V2's misaddressed A2A peers, where three ports formed a rotation and returned success from the wrong agent.

Also worth a ruling