Session 38 — Untangle + Cleanup of v4 Page Deployment Systems

0 · Strategic review notes — risks & open items before cascade

Review log — Codex 5.5 + Gemini Pro, 2026-05-10

HIGH findings — addressed inline

• H1 (Codex) — Pipeline §3 step 5 said --channel blog / --channel landing “NEVER --channel article”, contradicting AGENTS.md & CLAUDE.md which mandate --channel article for disease-adjacent topics. Fixed — pipeline now reads copy_preflight.py --channel article with the canonical disease-adjacent topic list.
• H2 (Codex) — Theme deployment ordering contradicted itself: §3 said serialised UK→US→DE→FR “never overlap stores”; §9.B said the 4 stores “can run parallel.” Fixed — both sections now state: serialised within a single store dispatch (no overlapping PUTs to the same store), separate per-market dispatches MAY run in parallel because they target different repos.

MEDIUM findings — addressed where cheap

• M1 (Codex) — GRAPHITI_API_KEY status: kept READY (fallback OK) since flat-file fallback returns canonical answers; flagged in §0 risks and §9.C with explicit “fix when convenient.” Not blocking Mission 25.
• M2 (Codex) — sync_canonical_messaging.py needed connections-registry consultation. Fixed in §9.D step 2.
• M3 (Codex) — Wave-3-review BLOCKER wording was ambiguous. Fixed in §9.C — row title now reads “Resolution of 5 wave-3-review items” and explicit “Items are not yet resolved” note.
• M4 (Codex) — mkt-creative-brief “verify reads same source as mkt-ad-copy” — deferred to Mission 25 dispatch as a hard pre-flight check rather than a passive note.
• M (Gemini, verified) — Performance / cost metrics absent. Deferred — not material to go/no-go decision; banked for the next cross-session digest.

LOW findings — ignored unless flagged

• L1 (Codex) — apps.np.org vs apps.nutritionalproducts.org: both valid (alias). No change.
• L2 (Codex) — GPT-Image-2 / FLUX 2 Pro service-contract references: covered by viz-image-pipeline SKILL.md. No change.
• L1/L2 (Gemini) — Wording nits on §9.A discount-codes risk and §9.C BLOCKER phrasing. Phrasing on §9.C resolved as part of M3 fix above.

Risks & open items, surfaced explicitly

• Enforcement tool fallback mode — check_copy_against_canonical.py currently runs flat-file fallback because GRAPHITI_API_KEY env-var resolution needs a fix. Flat files return canonical answers correctly, so this is non-blocking; flagged so it isn’t forgotten.
• 5 unresolved Calum decisions at wave-3-review (B1, B3, B4, B6, B7). Recommend resolving before Mission 25 cascade so new commissions don’t collide with pending decisions on existing copy.
• Wave 3 backlog — 10 items at projects/briefs/canonical-messaging-multi-track-2026-05-09/wave-3-backlog.md. Not blockers; documented so they aren’t lost between sessions.
• Theme-repo state across 4 stores — should all be at canonical state post-Wave 2 Phase 2; verify before Mission 25 dispatch (any half-deployed fixes get clobbered by new commissions).
• Mirror staleness risk — if the Drive canonical-messaging doc has been edited since the last Pinecone ingest, run tools/sync_canonical_messaging.py before any new copy work.

1 · Where we started

The original mandate was narrow: backfill tactics docs for ten v4 landing pages that had been built without a pre-publish strategy doc. Those ten pages were a snapshot of how things had been done up to that point — commission a page, ship it, audit later.

While walking through the ten pages, two customer-facing copy bugs surfaced:

• The footer disclaimer on multiple LPs read “FOCUS PRO/ELITE suitable for adults 18+”. The canonical age rule on the actual product labels is 10+ with band-specific guidance (e.g. ORIGINAL from age 10, PRO from 14, ELITE from 16). The 18+ wording isn’t just inflated — it contradicts the label and would shrink the addressable audience.
• A FAQ entry titled “Is it safe for students under 18?” framed PRO/ELITE as adults-only, again contradicting the label and the authorised wording.

Those two bugs were the visible symptom. Behind them sat the deeper system problem: copy was drifting across pages because there was no canonical source of truth, no enforced gate before publish, and no automated check between “what the label / authorised-claim says” and “what got typed into a Liquid file at 2am.” The whole rest of the session traces back to that observation.

2 · The arc — what got built (chronological)

What started as a tactics backfill turned into a cleanup of the entire pipeline that produces v4 pages. In rough order, here’s what landed:

3 · The new SOP for how a v4 page is made

Every customer-facing page now flows through this pipeline, in order. Publisher tooling refuses to POST when an upstream gate is missing — that’s intentional, not a bug. The headline change: we used to publish first and audit later. Now we plan first, gate continuously, and only PUT when every upstream layer has signed off.

4 · What was actually deployed (the receipts)

Wave 2 Phase 2 — per-market summary

Other deploys

• D3 Phase 2b — 10 FLUX product-in-scene renders deployed across 4 markets · commit adebcad
• B2 Strategy A discount migration — 4 stores migrated to canonical welcome codes · commits 5774860 6e00b3f 47f0242 + 4 theme-repo commits
• Wave 3 Track A — UK anti-coffee sweep, planpath hygiene · commits 93304ea f72e02b
• Track A2/A2b — Pinecone + Graphiti ingest of canonical messaging · commits 3a4c712 8d6022b
• Track A6 — canonical-messaging publish gate built · commit c8cf3a8
• Wave 3 dashboards — review + summary published to apps.np.org · commits b9f69d95 3c76355e

5 · Operations completed (OPS01 / OPS02 / OPS03)

• OPS01 — Shopify token-mint canonical path. Caught and banked: tokens now mint to F:/Agentic-OS/.env, never F:/Claude Root/.env. Two copies of shopify_get_token.py existed; running the wrong one wrote tokens nobody read. Fix: always cd F:/Agentic-OS first. Memory pointer at feedback_shopify_token_mint_canonical_path.md.
• OPS02 — Wave 2 page deployments. 4 stores. 30 theme files. 88 v4 pages audited and gate-consulted. 5-log ritual per market.
• OPS03 — Wave 3 cleanup + canonical-messaging hybrid build. Pinecone + Graphiti operational and queryable. Enforcement tool live (check_copy_against_canonical.py) with 3/3 smoke tests passing.

6 · Where we are right now

• 4 markets shipped Wave 2 Phase 2 — UK / US / DE / FR all live.
• Pinecone — 1,016 records populated, queryable, Tier 1 contract.
• Graphiti — 869 triplets populated, queryable, Tier 1 contract.
• Enforcement tool — live, 3/3 smoke tests pass, flat-file fallback active until env-var fix.
• 5 items pending Calum review at wave-3-review dashboard (B1, B3, B4, B6, B7).
• 3 items resolved autonomously — B2 Ritalin substance / B5 anti-coffee ambiguous=0 / B8 US Dopamine H1.
• Wave 3 backlog — 10 items logged for next session at projects/briefs/canonical-messaging-multi-track-2026-05-09/wave-3-backlog.md.
• 18+ disclaimer bug — FIXED through Wave 2 Phase 2 sweep across all 4 markets.
• Adults-only FAQ bug — FIXED through canonical-messaging gate consultation per page.

7 · What this unblocks

• New page commission — A new page brief flows through the SOP and lands clean. The agent doesn’t have to invent a process; the process exists.
• Copy drift detection — Pinecone + Graphiti gate catches drift before it ships. Every commit that touches customer-facing copy can be checked against canonical wording in seconds.
• Compliance reviews — Auditable trail (.qc-log/canonical-overrides.jsonl, deferred-to-Calum log, per-phase audit reports). If ASA / EFSA / ANSM / FDA ever asks “how did this claim get on the page?” we can answer.
• Cross-market consistency — The DE-canonical hero pattern (referenced in B4) can now be ported to other markets through the same gate, without each port becoming an ad-hoc rewrite.
• Mission 25 (campaign architecture → creative briefs → LP commissioning) inherits the SOP automatically. The next campaign doesn’t need a new playbook; it uses this one.

8 · Wave 3 review — ALL 8 RESOLVED CLOSED

Cascade-level deploys shipped today (parallel to B-item review)

• Bucket 1 focus-hub suppression — chat widget + Forms popup hidden + Related Articles fix — UK / US / DE / FR all live (32 PUTs, all 200, zero customer-visible regressions)
• Track B brain-fog quiz hero redesign — 8 surrealist concepts → 16 WebPs → 4-market live deploy with responsive overlay/stack pattern — SHIPPED · total cost £0.31
• DE Dopamine pain-led port — dual-file fix needed THREE locations on DE (section + body_html + page.title). All 3 updated. Live H1 now reads "Der Dopamin-Leitfaden: Warum Sie sich antriebslos fühlen — und was wirklich hilft". DONE
• Canonical-messaging Drive watcher (Track A5) — lazy per-day diff with sentinel; 79 zombie records purged from Pinecone — LIVE
• Pre-PUT assertion suite codified MANDATORY in AGENTS.md — caught a real Liquid 422 pre-render on DE leg — LIVE
• Creative-review dashboard fix — 5 broken bottom tabs (Article Guides / Read Articles / Collections / Exam City / Quizzes) restored. Root cause: .dockerignore excluded screenshots from Coolify build context.
• Pages registry hygiene sweep — 34 rows updated (Last QC + QC Result + Notes), selective Sheet writes only.
• Iron sweep + ingredient gate (Layer 6) — 6 phases complete. Phase 1 Pinecone: quote:other:top-sant corrected (first pass over-removed zinc/iodine/choline alongside iron; re-fixed to preserve valid minerals, drop iron only). Phase 2 themes: 13 files PUT across 4 markets. Phase 3 Shopify pages: iron counts UK 7→0, US 12→0, DE 12→0, FR 13→0. Phase 5 Layer 6 gate: 280-line gate + 31-ingredient lexicon × 3 langs + 29/29 pytest tests, wired with BZ_INGREDIENT_GATE_ENABLED=advisory flag. DONE
• Ingredient rollback + canonical audit GSD (post-iron-sweep) — 4 Phase-1 audits all clean (Pinecone 0/90 collateral, body_html 0/4, themes 0/13). Phase 2 rollback: no PUTs needed. Phase 3 scientific references rebuild: 12 snippet PUTs across UK/US/DE/FR (PRO/ELITE/Bundle) with EFSA-verbatim wording + DE/FR translations. Phase 4 body copy sweep: 7 surface PUTs with WARN-OVERRIDE + desktop/mobile QC PASS. Canonical SoT data contract at reference/data-contracts/product-ingredients.md. Citation Library v1.0: 28 ingredients × 138 citations, manual-only sync. DONE
• B7 Path B — DE cognitive-enhancers collection — Live at brainzyme.de/collections/cognitive-enhancers-smart-pills in EUR with 3 FOCUS products. Phase 9 re-finding: original "DE→UK URL misroute" premise NOT corroborated by live Google Ads state — collection now stands as defensive infrastructure. DONE
• B8 US/UK Dopamine Four-Layer hero — Template reassign on both markets (page_id 120747917504 US / 697036571005 UK). New H1 / subtitle / H2 live. All 18 US section markers + 9 UK markers verified. DONE

9 · Mission 25 integration — go/no-go before cascade

Calum is about to dispatch the Mission 25 cascade (campaign architecture → creative briefs → LP commissioning → asset renders → ad setup). This section answers: does anything Session 38 built conflict with Mission 25, and what does Mission 25 need to know to inherit cleanly?

A · Overlap zones — what touches what

B · Things Mission 25 must know (inheritance brief)

• Tactics-gate is mandatory pre-publish — tools/lib/tactics_gate.py. Every LP needs projects/briefs/{slug}/tactics.md per _tactics-template.md.
• Canonical messaging gate is mandatory pre-customer-facing-edit — per AGENTS.md §“Before customer-facing copy.” Run tools/check_copy_against_canonical.py per chunk before any PUT.
• Discount codes per market — WELCOME10 (UK/US), WILLKOMMEN10 (DE), BIENVENUE10 (FR). Don’t mint new topic-prefixed codes.
• Image methodology — GPT-Image-2 watercolour ligne-claire for atmospheric scenes + ONE FLUX 2 Pro photorealistic product-in-scene shot per LP in the nutrition section. viz-image-pipeline §1.5.
• 5-log ritual after every deploy — DEPLOYMENT_LOG / pointer / session bank / pages_registry / Component Registry. Non-negotiable.
• Defer-to-Calum criteria — substance changes / RED-compliance items / code disables >100 lifetime uses / net-new Shopify mints / cross-domain routing / Codex REJECT verdicts.
• Theme writes serialised within a single store dispatch — one PUT at a time per store, never overlap PUTs to the same store. Separate per-market dispatches MAY run in parallel (UK / US / DE / FR are different repos), but each market’s own writes are internally sequential.
• Codex 5.5 plan-mode review on every plan — HIGH inline (don’t halt), MEDIUM if cheap, LOW ignored unless flagged.
• Gemini Pro for imagery only — never for text plans; hallucinates strategic content.

C · Pre-cascade checks — UPDATED 2026-05-10 18:50 UTC

D · Recommended sequencing GREEN LIGHT

Session 25 working context: F:/Agentic-OS/context/sessions/session-25-competitor-creative-intel.md · Campaign Planner v2 brief: F:/Agentic-OS/projects/briefs/campaign-planner-v2/PHASE-A-BRIEF.md · Phase B sign-off dashboard: /phase-b-signoff/