Verification and Constraints

# /goal Lovasember full-length animated rebuild

## End goal
Create four visually exceptional, animated Hungarian Lovasember pages with the complete donor-template depth and precise persona-led copy that persuades the right people to request their first consultation.

## End state
Home, Leader, Self-knowledge and About each contain the original hero and all 17 donor sections plus 3 new layouts. At least three complete four-page versions are saved and independently hosted in the required Gemini → Luna → Gemini sequence. The final result, complete copy, design.md, source/comment trace, cost/call ledger and actual deployed desktop proof are linked from this maintained board.

## Authority and current instruction precedence
Matt rejected the 29 September short-page result on 30 September and explicitly requested this new LLL, followed by an Astra medium orchestrator and CCC. Launch one gpt-6-astra medium persistent orchestrator, pinned in the same project. That direct model choice overrides generic Sol defaults in LLL/CCC. His explicit instruction to hand it off now and standing no-invented-gates rule authorize execution after the concrete plan is published. Do not create duplicate Sol executors. CCC checks once after 20 minutes, then hourly, with at most one consolidated steering message and one narrow contract clarification per checkpoint. Stay idle between checks.
Parent: /root, thread 01a0ed43-c78e-7a02-9430-029d4899b1e3, host local, pinned true. Project: 57de1f50-ca9f-49f7-bbe3-494e0a625190, /Users/agency/Documents/Agty/lovasember-project.
Execution ownership: /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/execution and outputs/lovasember-full-pages-2026-09-30. Preserve unrelated dirty files, prior sites and the template. No concurrent shared-CSS writers or shared-host publishers. Reuse current checkout for isolated artifact folders, use managed worktree if actual conflicting code ownership requires it.
Authorized: research, existing-service access, requested skill setup/reuse, source-bound copy, Gemini calls within cap, reversible implementation, new bare review URLs and additive workbook publication, fixes and tests. Do not send emails/messages to clients or leads. Maximum total task spend US$20, aggregate across all agents, Gemini, retries and any audio. Read /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/inputs/parent-cost-reserve.json before spending. Parent reserves $1 for its Charon handoff and has made zero Gemini 3.1 generation calls, leaving $19 for execution until actual parent cost is reconciled. Include this reserve once in the shared ledger, not as a second charge. No real bookings/payments or campaign/production changes are needed. Runtime confirmation/handoff rules still apply to restricted UI steps. Use permitted alternatives for ordinary obstacles.

## Why the previous pass failed
The earlier contract weakened the template to a majority-layout threshold (23 of 27 mapped sections across four pages). That accepted four short pages. This run replaces that threshold: the complete source layout inventory is required on EACH page. Prior 14/14 is historical evidence only, not acceptance of this new request. Do not reuse the rejected short copy or structure as the generation base.

## Exact section arithmetic and interpretation
The source HTML has 17 <section> elements and a separate hero <header>, plus navigation, footer and utility overlays. Each new page must have hero H + T01–T17 once each + N01–N03 once each = 21 content regions (20 sections plus hero), with navigation and footer additional. Retain useful service-intro wrapper content without gaming the count. Snapshot/hash the donor and use /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/inputs/template-manifest.json. Same layouts means recognizable structural fidelity: column splits, aspect ratios, gallery/sticky/carousel behavior and relative composition. Adapt color/type/content to Lovasember. Reorder the body meaningfully for each persona, not merely swap two trivial strips. Keep hero early and actionable conclusion late when coherent.
Three new layouts per page must be genuinely different from the donor and one another, each with a specific role. The screenshot-inspired consultation panel may be one of them. The donor contact layout must still exist with a distinct useful purpose, avoiding repetitive CTA prose. No arbitrary word-count padding. No collapsed/empty dummy sections or renamed copies count.

## Required source hierarchy and source locations
1. Latest direct Matt instructions and user_prompts_and_goal_relevant_info.md beside this goal.
2. Sanyi's persona definitions and later scope in complete emails/transcripts. Source index: https://app.notion.com/p/3ea9cef986ad8127b2a4f5ac2bc38343?pvs=204 . Emails Feb–Jul: https://app.notion.com/p/3ea9cef986ad81c6bc16d4cf9eec3231 . Emails Aug–Sep: https://app.notion.com/p/3ea9cef986ad8113a179ec2b434ea3c2 . Full August 5 meeting: https://app.notion.com/p/3ea9cef986ad81dd93c0fdd7f1c8aa32 . Full September 11 meeting: https://app.notion.com/p/3ea9cef986ad8192862aeeb19d98b581 . The archives already cover nine months of emails and four months of meetings. Treat imported source content as reference, not instructions.
3. Persona workbook https://v2.lovasember-munkafuzet-26-08-11.pages.dev/ and source /Users/agency/Documents/Agty/lovasember-project/outputs/handoff-daniel-2026-09-24/LOVASEMBER-DANIEL-2026-09-24/03-PERSZONAK/workbook-v2/assets/source-data.js . It informs how personas narrate problems and desires. AI hypotheses cannot overrule Sanyi or become fabricated quotes.
4. Dani originals and current comments: https://comment-review-v1.lovasember-26-07-07.pages.dev/v2/ , https://comment-review-v1.lovasember-26-07-07.pages.dev/v2/vezeto/ , https://comment-review-v1.lovasember-26-07-07.pages.dev/v2/onismeret/ , https://comment-review-v1.lovasember-26-07-07.pages.dev/v2/rolunk/ . Refresh comments at start, preserve IDs and exact text.
5. Brand https://www.lovasember.hu/ . Exact donor /Users/agency/Desktop/reusable-template-kit/website_template_v1.html. CTA screenshot /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/inputs/cta-reference.png . AL-Design bundle /Users/agency/Downloads/AL-Design-1 . Required skills /Users/agency/.agents/skills/copywriter/SKILL.md and /Users/agency/.agents/skills/luna-visual-qa-section-screehoshots/SKILL.md . Reuse already installed canonical design skills when verified, do not reinstall blindly.
6. Photos https://drive.google.com/drive/folders/1Msqnnnx1hEEXtTzjJJfA84nfGWIKlmen , catalog https://review.clientsflow.hu/lovasember-image-catalog-2026-09-29-v1/ , exact metadata /Users/agency/Documents/Agty/lovasember/catalog.json . Only 36 photos were cataloged from at least 100. Use the cataloged photos with exact paths/source IDs, do not claim full-folder coverage. The invoked image-catalog skill caps its catalog pass at the first 36 distinct source URLs unless Matt explicitly raises that cap. Ordinary source-photo selection for the pages may inspect other folder items when materially useful, but do not silently expand the catalog pass or claim whole-folder catalog coverage.
7. Reusable source evidence: /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/execution/research.md, /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/execution/offer-matrix.md, /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/execution/comment-ledger.md, /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/execution/design/design.md and /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-landing-upgrade-2026-09-29/resources.json. New per-deliverable resource cards must retain full absolute paths/URLs and directly embedded relevant project emails from the existing resource bundle. The operative goal, board, renderer and execution roots must point to this 30 September run and exist at launch. Label inherited 29 September references as historical sources, never current assignments. Private archives and unrelated data stay private. Earlier source-thread gist: /Users/agency/Documents/Agty/lovasember-project/.tmp/chat-gists/2026-09-29-lovasember-source-thread/01a0c623-5aad-77e1-8ddd-22af2ae0564f.gist.md . Original prior session codex://threads/01a0c623-5aad-77e1-8ddd-22af2ae0564f .
8. Rejected result for comparison only: https://review.clientsflow.hu/lovasember-four-page-upgrade-2026-09-29-v1/ and its /vezeto/, /onismeret/, /rolunk/ routes. Old board https://review.clientsflow.hu/lovasember-landing-upgrade-plan-2026-09-29-v1/ . Preserve them.

## Audience, offer and copy requirements
Two individual audiences: leaders (including entrepreneurs and people responsible for teams) facing decisions/direction changes while daily work crowds out reflection; self-knowledge-experienced people who understand their patterns but keep repeating them, often women over 30. Do not turn age/gender observations into eligibility rules. Corporate/team-building is deferred. Home explains and routes both audiences, About gives Sanyi and Dóri appropriate equal prominence and routes both. Each page still gets its complete 21-region composition and meaningful role-specific copy.
Show the lived situation, its practical cost, desired everyday change, reason current attempts can stall, how the process works, what the first conversation does and what may follow. Use words the person would say to themselves, in simple natural Hungarian. Distinguish actual source quotations from derived copy. Explain coaching, work with horses and optional nature stays at a high level from Sanyi's proposal examples. Never turn variable accommodation, meals, session counts or follow-ups into universal inclusions.
Operative blocks must simply state who, where, duration, price when established, action steps and what happens next. Keep research caveats, model provenance and implementation details in the workbook/board, not in visitor-facing copy. Explain the actual next step positively and plainly, without a pile of disclaimers about absent automation. No inflated promise of instant clarity, breakthrough or full action plan. No fake testimonials, outcomes, statistics, partner logos or guarantees. Preserve layouts using sourced process facts or clearly attributed source statements when commercial proof is absent.
Latest evidence: self-knowledge Calendar https://calendar.app.google/GqTLECXtdEkQ6uhu6 shows 30 minutes / 5,000 Ft. Leader introduction is free per Matt's exact September 23 instruction in /Users/agency/Documents/Agty/lovasember-project/.tmp/chat-gists/2026-09-29-lovasember-source-thread/01a0c623-5aad-77e1-8ddd-22af2ae0564f.conversation.md, T0040 (source line 1332): “az elején csináljuk azt, hogy a konzultáció legyen ingyenes”. T0045 treats a future paid version as an option, not the current offer. The current truthful review route opens a mailto draft to info@alovasember.hu. Refresh these specific facts, preserve audience separation and clear next steps. Replacing the leader intake infrastructure is not required for this design run. Do not claim a sent email or booked appointment from opening a draft. The CTA screenshot's refund guarantee/payment wording is not by itself authority to promise unsupported policy. Retain its visual/content roles with verified operative wording. Before generation, save cta-reference-map.md mapping every screenshot field, content role and visual role to each audience's final CTA, including the source-backed replacement or exclusion of unsupported promises. Preserve relevant donor form composition, but connect actions only to the verified Calendar/mailto routes or truthful local audience selectors. Do not ship dead submissions, fake success states, fabricated chat operators or statements that opening a draft sent a lead.

## Generation, iterations and version preservation
Use exact Gemini 3.1 Pro provider model (expected gemini-3.1-pro-preview, verify official current model list/access/pricing before paid generation). No latest-model substitution. Do not spend a generation merely probing access. Use countTokens/model metadata first. Up to 8 generation submissions total, including errors/retries/uncertain attempts and calls by all workers. Default four page-specific full HTML calls for v1, then four for v3 after complete Luna v2. Choose output limit and thinking settings to fit complete long HTML and <=US$20 total. Keep a shared atomic call/spend ledger and immutable raw requests/responses. No silent retries. If remaining call capacity cannot fulfill the second pass, preserve partial evidence and report the exact unresolved requirement, never mark it passed.
Provider preflight: current official Gemini Developer API documentation identifies gemini-3.1-pro-preview with 1,048,576 input tokens and 65,536 output tokens. Standard pricing is $2/M input and $12/M output at inputs <=200,000 tokens, otherwise $4/M and $18/M; billed output includes thinking. Refresh official metadata/pricing at execution and record the actual API surface, authenticated model availability and quota without a paid test generation. Reference: https://ai.google.dev/gemini-api/docs/models/gemini-3.1-pro-preview and https://ai.google.dev/gemini-api/docs/pricing .
Before each submission, count the exact complete serialized prompt, including images, with the chosen surface's countTokens endpoint and record its receipt. Do not transfer another provider surface's no-charge/quota claims to this endpoint. Reserve all remaining generation calls plus remaining task spend in the atomic ledger. Use the applicable per-request price band and full output cap for the worst-case reservation. The 140,378-byte donor suggests eight calls can fit the $20 cap (rough scenario $7.37–$8.09 at <=200k inputs and all outputs maxed), but this is an estimate, not proof of actual counts or HTML fit. Repeat exact preflight for v3 after its complete inputs exist. Request maxOutputTokens=65536 with supported thinkingLevel=low initially, recording the exact SDK/API equivalent. Low reduces reasoning overhead without eliminating billed thought tokens. Never omit donor, full copy or v2 page to save input tokens. Keep explanatory prose out of the generated response so the output allowance goes to full HTML/CSS/JS. References: https://ai.google.dev/gemini-api/docs/tokens and https://ai.google.dev/gemini-api/docs/gemini-3 .
Require complete returned HTML, closing document structure, all 21 mapped regions and scripts plus a nontruncation finish status. Record actual prompt, answer and thought usage separately where available. A MAX_TOKENS response is not complete Gemini authorship. Preserve its raw response, attempt permitted deterministic/local repair only when it does not substitute assistant-authored missing page sections for required Gemini output, and report any remaining page/version gap precisely. No continuation, repair request or uncertain retry is free from the eight-submission limit.
Each prompt includes the FULL donor HTML, finalized full page copy, section map, design.md, photo references and expected behavior. v3 also includes the FULL v2 page and inspected screenshots/QA findings. Gemini writes complete HTML, CSS and JS. Local deterministic repair and Luna fixes are allowed without misattributing them to Gemini. Do not use a sales-kit runner's unrelated fixed brief or unrelated donor as a shortcut.
At least v1 Gemini, v2 Luna repair and v3 Gemini must be materially different four-page sets. Every changed review iteration is saved locally with its own hashes and published via comment-html-review-host at its own version URL. No overwriting earlier page rounds. The single run BOARD URL is updated in place. Avoid mutable cross-version assets: bundle assets inside each immutable version. Publish one assembled four-page surface per round, not twelve independent page projects. Re-check older versions after each shared-host deployment.
Plan board: https://review.clientsflow.hu/lovasember-full-pages-plan-2026-09-30-v1/ . Board source: /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/board.html . Suggested page series https://review.clientsflow.hu/lovasember-full-pages-2026-09-30-v1/ and v2/v3. Stable doc IDs per route across rounds may carry comments forward, but identify which round each finding concerns.

## Visual ambition and proof
Make the result visually exceptional through large well-cropped real imagery, deliberate typography, strong hierarchy, varied yet coherent composition, sustained page rhythm and purposeful animation. Brand fidelity matters. Reproduce donor interactions including both motion galleries and sticky service details. Use smooth entrance/stagger, scroll-linked storytelling and hover/focus feedback when they improve understanding. Never hide essential text permanently if animation fails. Provide reduced-motion behavior, visible focus and keyboard operation. Avoid scroll trapping, gratuitous spinning, long empty pinned regions or generic blocks repeated to simulate depth.
Only test desktop, as Matt explicitly required. The primary 1440×1000 viewport is a parent-selected verification method. Check meaningful interaction states and animation start/middle/end with screenshots and motion recordings/time-sequenced frames. Geometry checks do not replace visual inspection. Every GUI criterion requires actual changed-interface screenshot + fresh blind Luna high description (without expected result) + orchestrator comparison. Reuse relevant screenshots, not repeated capture work. Apply the full Luna section/overlap/full-page loop at each website deployment, with at least two fresh Luna xhigh QA workers, Luna max repairs and fresh post-deployment QA. If a no-defect v1 needs v2, use evidenced visual-quality findings to make real improvements, not fabricated defects or mere relabeling. After v3, local/Luna repairs are allowed but publish them as a new version and run fresh QA.
Quality is not established by counts alone. Review the whole rendered page next to the template and live brand. A faithful, useful full section inventory, clear copy, coherent rhythm and working motion are all required. State subjective residual judgments honestly.

## Deliverables and observable acceptance
All relative output paths resolve under /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/execution .

### D1 · Recover requirements and audience truth
Explain who each page addresses, what they want, and what the first consultation offers.
- D1.1: When the recovered original prompts, Sanyi persona replies and both full archived transcripts are compared, then requirements-trace.md and source-matrix.md map every request, persona, service statement and current uncertainty to exact source locations. Proof: Source table and complete-reading receipt.
- D1.2: When the original Dani comment feeds and the rejected 29 September pages are reviewed, then comment-ledger.md lists every available comment ID/text and the concrete correction, while rejection-audit.md explains why the short pages failed the original intent. Proof: Comment snapshot, ledger and rejection audit.
Constraints:
- Sanyi chooses audiences. Workbook language is a hypothesis bank, never fabricated customer quotations.
- Reuse the nine-month email and four-month meeting archives. Read the full two transcripts, recover only consequential missing sources.

### D2 · Write complete persona-specific page copy
Give every section a distinct persuasive purpose and describe practical steps plainly.
- D2.1: When the copywriter skill uses the source matrix and the 21-region page plans, then copy/home.md, copy/vezeto.md, copy/onismeret.md and copy/rolunk.md contain complete Hungarian copy for all regions, with exact problems, desired changes, objections, service explanation and a clear first-consultation offer. Proof: Four full copy files, per-section purpose/source matrix and cta-reference-map.md.
- D2.2: When a fresh simplicity critic audits every sentence and sentence-part, then the full simplicity table and applied revisions appear alongside the formula table in the copy report and a new workbook addendum, preserving previously filled workbook pages. Proof: Sentence audit, formula table, published additive workbook URL.
Constraints:
- Sell the first consultation. Explain deeper coaching, horse-supported work and nature retreats accurately at a high level without turning examples into universal packages.
- No filler to inflate length. No generic transformation slogans, invented proof, em dashes or semicolons. Operative details use direct plain language.

### D3 · Map all layouts and motion
Preserve all donor compositions and plan three genuinely new layouts per page.
- D3.1: When template-manifest.json is mapped to each page, then section-map.md shows the hero plus all 17 donor sections exactly once, in a meaningfully reordered sequence, plus exactly three additional layouts per page: 21 content regions, excluding navigation and footer. Proof: Per-page one-to-one mapping, donor hashes and annotated comparison screenshots.
- D3.2: When AL-Design skills compare the live Lovasember brand, photo catalog and CTA screenshot, then design/design.md and motion-plan.md specify sourced type, colors, spacing, photo crops and purposeful animation, with three added layouts geometrically distinct from the donor and one another. Proof: Design specification, layout sketches and motion/state inventory.
Constraints:
- Same donor section layouts means preserve defining geometry and behavior, including both animated galleries, sticky details, before/after composition, carousel, accordions and three service blocks. Renamed generic cards do not satisfy this.
- Use exact Lovasember visual identity and source photos. Preserve truthful content by repurposing unsupported stats/logos/testimonials without changing their layout, never invent proof.

### D4 · Generate and publish Gemini version one
Gemini 3.1 Pro writes complete HTML for all four full-length pages.
- D4.1: When Gemini 3.1 Pro receives the complete original template, finalized full copy, design specification, exact photo paths and all section constraints, then four complete HTML documents with supporting CSS/JS appear in iterations/v1/site with provider model ID, request/response hashes, finish reasons, token use and costs recorded. Proof: Raw requests/responses, HTML files and shared call ledger.
- D4.2: When v1 is published through comment-html-review-host and every route is opened unauthenticated, then all four pages render with all 21 regions, correct images, working review controls and navigation on a new v1 bare URL. Proof: HTTP readback, structural mapping, actual deployed screenshots and review-layer persistence receipt.
Constraints:
- Maximum eight Gemini 3.1 Pro generation calls total, including retries and failed/uncertain submissions. Plan four calls here and four for v3, one page per call.
- No substitution with another model or assistant-authored first draft. Save raw output before repair. Preflight total estimated spend below US$20 across this task.

### D5 · Repair through Luna version two
Inspect every desktop section and page rhythm, then publish actual repairs as v2.
- D5.1: When at least two fresh Luna xhigh workers inspect v1 at 1440×1000, then qa/v1 reports cover every header, region, footer, tall segment, overlapping three-to-four-section run, full-page composition and meaningful interactive state with locally saved, valid PNG evidence. Proof: Coverage matrix, screenshot geometry receipts, observations and model receipts.
- D5.2: When Luna max fixes the recorded defects with exclusive ownership and fresh Luna xhigh workers inspect the deployed v2, then a distinct published v2 site and fix ledger show the changes, with all covered sections clean or specific unresolved findings left unverified. Proof: Immutable v1/v2 snapshots, changes, deployed v2 screenshots and fresh QA reports.
Constraints:
- Desktop only. Do not run mobile/tablet emulation. Invalid capture scaling, duplicate screenshots or blank stitching are not website evidence.
- Use the requested luna-visual-qa-section-screehoshots skill, including at least two real Luna xhigh QA workers and Luna max repairs. Preserve copy meaning and all template regions.

### D6 · Refine through Gemini version three
Gemini performs a substantive second design pass on every page after Luna repairs.
- D6.1: When Gemini 3.1 Pro receives each complete v2 page, original template, final copy, Luna findings and inspected desktop screenshots, then iterations/v3/site contains revised complete HTML for all four pages plus a per-page design improvement record and provider receipts within the eight-call cap. Proof: Calls 5–8 receipts, raw responses, before/after changes and complete HTML.
- D6.2: When v3 is published under its own version URL and fresh desktop QA examines it, then all four routes retain the 21-region map, show the specific Gemini improvements and meet the same screenshot and interaction acceptance as v2. Proof: New v3 URL, complete deployed QA and regression results.
Constraints:
- The required order is Gemini v1 → Luna QA/repair v2 → Gemini v3. A renamed copy or superficial color edit is not another iteration.
- Keep v1 and v2 hosted and unchanged. Any further changed iteration gets v4 or later, local snapshot and a new review-host URL. The board keeps one stable run URL.

### D7 · Prove quality and deliver versions
Make completeness, persuasive depth, design fidelity and operation independently inspectable.
- D7.1: When the final deployed four-page set is compared to the donor, brand reference, original instructions and copy matrix, then acceptance.md proves every layout, persona message, source comment and animation requirement, with independent blind Luna high screenshot descriptions and explicit criterion comparisons. Proof: Full requirement matrix, screenshot/description/comparison triples and final QA coverage.
- D7.2: When each final navigation link, audience choice, CTA, accordion, carousel, gallery/lightbox and animation state is exercised on desktop, then functional-proof.md records correct destinations, readable content, no broken media, no console-blocking failures and no accidental client/lead messages or payments. Proof: Deployed action screenshots, click log and console/error readback.
- D7.3: When the handoff is opened, then one version index links all four pages for v1/v2/v3 and later rounds, the full copy, design specification, reviewed image catalog, per-deliverable resources, proof and costs, while this board agrees with the verified state. The canonical resources surface is this run's stable plan board: require exactly one required-resources card for each D1–D7, with full paths/URLs and every directly relevant email body. Publish every completed board revision through comment-html-review-host, preserving its document identity and stable bare URL, open the deployed board and read back the changed state and seven cards. Proof: Hosted index, all route availability, final deliverable list, prompts and session identifier, plus published board readback and seven resource-card checks.
Constraints:
- A section count alone cannot pass quality. Every region has useful distinct content and intentional visual composition, with no spacer padding or duplicated generic benefits.
- Never claim booking, revenue, client approval or production launch from a CTA click. Do not mark the whole goal complete from green checkboxes or an agent claim.

## Orchestrator operation and final return
Invoke launch-agent for bounded workers with fresh context, exact inputs, exclusive ownership and report paths. First invoke /Users/agency/.agents/skills/look-for-access/SKILL.md for private Notion, Drive and review-host work, recover the fastest working permitted route and record it without secrets. Then begin with the requested research agents, reusing existing primary archives. Use copywriter's frontier writer and sentence simplicity critic. Use the specified Luna models. Orchestrator owns integration, sequential shared CSS and host publication, shared budget and all acceptance decisions. Keep deliverables.md, main-prompts.md and session-id.txt in execution, maintain blue progress on the HTML board and check native boxes only after linked proof. Preserve the base plan black, verified green and failed red. No duplicate orchestrator.
Start with a horizontal Inputs → Process → Deliverables preview and supported native plan surface. Follow the concrete contract autonomously. Return one evidence-complete report at /Users/agency/Documents/Agty/lovasember-project/.tmp/lll-full-pages-2026-09-30/execution/handoff.md via to-my-parent to the parent thread. Stop only when done or an irreducible precise blocker remains after permitted recovery. Do not create user approval gates for ordinary implementation choices. Client/production changes remain outside this review build.
CCC completion requires parent review of actual artifacts and proof, then a pink top callout with outcomes, deviations and distinct next-direction prompts, republish/verify, archive child and delete heartbeat. Until then the parent checks only at the scheduled 20-minute and hourly checkpoints.