• ad-creative (2.5.0 → 2.6.0): added Mode 4 — the Creative Strategy Loop, the upstream workflow layer for deciding which ads are worth making (built so the same capability ships in any agent running the library, incl. Magister; workflow shapes informed by public creative-strategist practice, expressed originally). New references/creative-roadmap.md: three-signal synthesis (account performance via the platform CLIs; customer/brand language via the Grounded Inputs corpus + customer-research; external organic via scraping, social listening, and competitor-profiling) on a monthly-deep-dive + weekly-refresh cadence with a receipts rule (no insight enters the icebox without its source); concepts as segment × motivation × angle × format hypotheses ranked on a six-tier evidence scale; account-state branching (exploration state: wide net-new mix, per-metric win redefinition, iterate only on hits, four root causes to check; scaling state: winner-variation-heavy mix with visually-distinct variations, sub-angle probes, a standing exploration allocation); the roadmap artifact (icebox → quarterly themes → monthly slate with T1 iteration / T2 remix / T3 production tiers and a hard capacity check); per-concept briefs handed to Modes 1–3; and the monthly creative retro (winners/losers/metric-wins/learnings/kills, judged at concept level, every learning landing as an icebox update, re-rank, or kill). New references/hook-system.md: hooks as three components (visual action / spoken line / caption) with the no-duplication rule; the segment → motivation → format → hook generation pipeline output as a hook matrix; an eight-move opening menu; the diagnostic funnel mapping thumbstop/hold/CTR/CVR to which component to fix; the on-ramp rule (every hook test is an on-ramp test); fidelity laddering tied to production tiers; and inherited grounding rules plus organic language mining. SKILL.md adds the Mode 4 section and 'creative strategy,' 'creative roadmap,' 'creative retro,' 'hook writing' triggers. New eval (id 9) covers exploration-state diagnosis, evidence-ranked slates, and the receipts rule.
• customer-research (2.0.0 → 2.0.1): added the zero-review persona fallback to Persona Generation — a four-step proxy ladder (own differentiator → direct competitors' reviews → marketplace comparables → adjacent brands sharing the audience) with provisional-persona tagging and replacement as first-party evidence arrives.
• marketing-loops (1.1.0 → 1.2.0): added the monthly-creative-retro loop to the Paid section of references/loop-catalog.md — closes ad-creative's Mode 4 loop on a schedule: pulls last month's performance, drafts the retro artifact and next month's capacity-checked slate, flags the account-state call for human confirmation, one-retro-per-month idempotency, and never launches or pauses ads. Closes #434.