Ship Helix 11-wave upgrade as primarily-vendored absorption (167 brand…
Decision
Ship Helix 11-wave upgrade as primarily-vendored absorption (167 brands, 4 reference impls, 4 new skills, 6 skill extensions) with deliberate scaffold-only scope on agent-router transport adapters (W2.2) and helix: frontmatter retroactive migration (W3.1) — over a single-session-impl-everything alternative.
Rationale
Premortem identified 4 real follow-up risks the next session must own: (1) vendor working tree (~170MB at /root/aj-workspace/vendor/) has NO sync cron — will go stale; manual git pull documented but not automated, violates feedback_proactive_cleanup_hooks.md unless evolution-backlog gets a cleanup-cron entry. (2) agent-router SCAFFOLD must not be confused with impl — SKILL.md flags it (“scaffold not production”), but a future session reading “agent-router exists” may falsely assume routing works; mitigation is the explicit “Roadmap” section in SKILL.md. (3) helix: frontmatter spec is a doc, not a runtime — autoloader still does regex; without a Phase 3 autoloader patch (separate session), the typed metadata is dead weight. (4) Bias detected: framing-effect on “skipped MEMORY.md per anti-bloat” — that decision was correct (skills auto-discovered) BUT the LACK of an evolution-backlog entry for the 4 follow-ups is a real gap. Counter-bias: write the follow-ups to evolution-backlog NOW so they’re not forgotten. Calibrated confidence is 0.78 not 0.95 because: SMOKE only validated slop-lint (1 of 11 deliverables), the other 10 are file-presence-validated not behavior-validated; agent-router and collab-primitives are scaffolds; helix: spec is unproven by autoloader integration. Decision stands as the right tradeoff for “march toward god grade” — getting 11 waves to scaffold-or-shipped beats 3 waves fully impl’d because Helix benefits from the catalog/spec layer immediately, the impl layers can compound across future sessions.
Alternatives Rejected
Outcome
Pending