B0 (decision, Jack): ratify depth-1 redesign gates G-1/G-8/G-9

[aarontrowbridge]

Decision needed (Jack) — ratify the depth-1 redesign's three fork-facing gates

This is the gate that unblocks all fork-side work (B1–B8) in the Amico agent/skill architecture depth-1 redesign. Design spec is in Aaron's vault (spec-20260708-112732-amico-agent-architecture-depth1-redesign); happy to walk through it live. No fork code changes until this is ratified — B1–B8 are drafted but intentionally unfiled.

Context (one paragraph)

The Amico agent stack was built for Claude Code (root session = orchestrator). Amicode is an opencode fork. Source audit finding: opencode does not hard-block subagent nesting — it's a per-agent task permission default, overridable in config (CC blocks it hard; the fork is soft). So the 1117-line orchestrator could be ported, but shouldn't — control flow (iteration loop, DAG walk, escalation) belongs in a score + extension, not an LLM, for reproducibility/cost/provenance. Agents flatten to depth-1 leaves; the tool surface standardizes on a CLI that both runtimes call.

The three decisions

• G-1 — orchestration home: opencode score (data-driven flow) vs extension TS. Recommendation: score manifests for flow + amicode_* tool validation for enforcement; TS only for genuinely new stage-kinds. Open sub-question: does the experiment-iteration loop fit the stage model or need a TS driver?
• G-8 — depth policy: flat-by-default, with a governed task-grant escape hatch in the opencode ring only. Recommendation: built-in general/explore stay ungranted; nothing granted now; the skill/agent lint hard-fails a task grant lacking a justifying frontmatter annotation.
• G-9 — tool-transport boundary: CLI-first spine (amico verbs, the amico-run lineage) + optional amico mcp-serve facade + opencode-plugins reserved for interview/session-bound UX + remote HTTP for the cloud ring. Interacts with the open-core release — this puts the entitlement gate (gate.ts, exit-64) in the OSS shell. A runnable draft scaffold exists (Aaron's machine, ~/harmoniqs/amico-cli-draft, bun test green) generalizing the real amico-run; the amicode_* migration is the load-bearing sub-decision.

What ratification unblocks

B1 generalize amico-run → amico CLI · B2/B3 migrate bookkeeping tools to verbs · B4 harness-reframe prototype · B5 MCP facade · B6 one-source agent/skill compiler · B7 retrieval tools · B8 full orchestrator dissolution.

Acceptance criteria

• G-1 decided (score / TS / hybrid)
• G-8 depth policy ratified
• G-9 tool-transport boundary + amicode_* triage ratified (coordinated with the open-core split)

Blocked by

None — this is the decision gate for Track B.

[aarontrowbridge]

Gate decisions (recorded 2026-07-09). Aaron delegated G-1/G-8/G-9 to me tonight to unblock Track B. Jack — these are my calls per the spec's recommendations; your async refinement or veto is genuinely welcome before we build beyond the foundational CLI slice. Nothing is being merged into the fork tonight.

• G-1 (orchestration home): opencode score for flow (stages / gates / loopbacks) + amicode_* tool validation for enforcement; extension TS only for what the stage model can't express (the experiment-iteration loop likely needs a thin TS driver). Score-first.
• G-8 (depth policy): flat-by-default. Governed task-grant escape hatch in the opencode ring only; built-in general/explore stay ungranted; nothing granted now; the skill/agent lint hard-fails a task grant lacking a justifying annotation.
• G-9 (tool-transport): CLI-first spine (amico verbs, the amico-run lineage) + optional amico mcp-serve facade + opencode-plugins reserved for interview/session-bound UX + remote HTTP for the cloud ring. amicode_* triage per spec §7.3. The entitlement gate lives in the CLI, so it ships in the OSS shell.

Proceeding tonight: B1 (generalize amico-run → the amico CLI verb router) as a PR only. The fork's CI-token blocker means nothing merges without a human regardless. Holding B2–B8 — especially the B4 harness reframe and the B8 orchestrator-dissolution epic — for review on top of a working B1 and for Jack's input.

— posted on Aaron's behalf (delegated), via Claude Code

[Rchari1]

Superseded by #368 Autoresearch Studio — the loop is the product (design-of-record `spec-20260813-162300`, decomposed into WS1–WS7 #371–#375).

The `B*` depth-1 redesign (`spec-20260708-112732`) gated by B0 #107 was never ratified; the harness axis has moved to opencode open + plan-first, telaio by subscription (#398) with:

• WS1 WS3 — QEC pack landing: the second domain proves the interface #371 pack interface — one manifest per domain, score as field of pack (`1922950 WS1 — The pack interface` + follow-up `84a0784`)
• WS4 WS4 — Telaio research rail: campaign interview → skills injection → campaign execution #372 / WS5 WS5 — Campaign machinery: ledger entity, warrant bounds, fleet dispatch, scheduled runs #373 telaio research rail + campaign machinery (warrant-bounded, amico-CLI boundary) — subsumes the orchestrator-dissolution / harness-reframe intent
• Modes: default to plan, retire the pulse-designer agent shell (#389) #395 (`053f952`) plan-first default (`default_agent:plan` at `opencode_config.ts:434`) + pulse-designer retirement

This `B*` slice has had no activity since 2026-07-09 and is now tracked under the WS epics. Closing as superseded — please reopen against #368/WS* if any slice still applies.