Project home · Build and content workflow
The local candidate provides 16 full stack guides across content websites, web applications, services, command-line tools, automation and native mobile. Human and agent paths use the same task-based recommendation policy and complete records. The reviewed 23 October base was pushed as 9aa317f. The task/rewrite-source follow-up is local and held for Almir’s review; production delivery is tracked separately from local verification.
The optional context router validates explicit requirements and returns resources, candidates, exclusions and questions. It calls no model and stores no application data. Known-stack work preserves the choice; new-stack comparisons and migration planning keep constraints and unknowns explicit. The visiting agent interprets the application conversation and reasons about the result.
Initial HTML, matching page Markdown, JSON exports, canonical sitemap, crawler policy and factual contact resources support ordinary browsing and direct retrieval. Production uses Cloudflare Pages with Functions; local responses have noindex headers and exclude analytics. Aggregate pageviews load only on the canonical production origin. The preview uses the same routing and format handlers.
Core official guidance and all 394 cited source addresses were checked on 13 October 2026. That is the dated scope of the original URL check, not a blanket current freshness claim; per-record research dates and scope remain separate in the freshness snapshot. Recipes, full dependency combinations, application capacity and migrations require qualification in the actual project; this reference does not certify compatibility or measured cross-stack rankings.
Build and local rendered/direct checks cover desktop and phone views, contact representations, routing, malformed input, true missing pages and resource parity. These are website checks, not execution evidence for every application recipe. The real demonstration is included in the released site; the dated status below records its release and verification boundary.
The current 16-profile catalogue has first-party generator records: official framework/language initializers, IDE templates or documented CLI routes, conditional incremental generators, generated defaults to retain and baseline checks. Plain Mix project choices are included with the Phoenix profile; supervision is conditional. Profiles without a full application generator explicitly distinguish module/package initialization from the application code still needed. Each record owns its own review and evidence date; 18 October is the date of the historical catalogue verification below, not a shared review date for every current record.
The shared workflow requires generation before hand-authoring a new skeleton, preserves existing repositories, records deviations and separates review from execution. Agent entry guidance links the rules. The normal compiler exposes matching HTML, complete Markdown, structured records, the full bundle and /generators.json, and requires coverage for every curated profile. Generator review dates remain distinct from imported research freshness.
The original 18 October review did not execute commands or native templates. Later disposable starter trials are scoped in their generator and profile records; they do not qualify real applications or the complete catalogue. Stack Menu can provide explicit instructions and a generation receipt; enforcement in a consuming agent or build runner requires that consumer to honor and verify the guidance.
Local verification on 18 October: normal build passed; all 14 profiles passed direct checks for source/record/catalogue agreement, command coverage in HTML, complete and negotiated Markdown and the full bundle, MIME, Vary: Accept, HEAD, local noindex and genuine missing-page status. The catalogue includes 22 starting points, 34 conditional follow-up entries and 85 unique cited sources. Desktop browser inspection of the Elixir generator section passed; phone rendering was not inspected in this pass. An independent read-only integration review found no actionable defects. These checks verify reference delivery, not application scaffolding execution.
Almir selected Oghogho’s design for implementation, based on optimumBA/design at 9c62989. The 21 October release adapted the homepage, library, comparison, application-type filtering, migration planning and all 14 guides then in the catalogue. Shared navigation, fonts, mark, favicon and share artwork follow that direction. The current local source has 16 guides. Redesign record owns the original review and deliberate adaptations. The production release is recorded below.
Complete agent records, Markdown, generators and routing remain intact. At release, recommendations filtered application type; the visiting agent still reasons about requirements. During the 20 October review, the illustrated conversation was visibly labeled and the video area linked to agent instructions while recording was pending. The real recording was added and released on 21 October below.
The rewrite picker now makes its list immediately visible before moving keyboard focus into it; opacity and position still animate. This repairs selection with the keyboard. Native browser checks covered arrows, Home/End, type-ahead, Enter, Escape and Tab, including focus restoration. Mouse selection passed in Safari. The normal production build passed. The repair was local during review and was included in the 21 October release.
UI architecture preferences now filter choices only for an explicit web component. API and unspecified-component requests retain the same candidates regardless of that preference. Direct router checks covered those cases and confirmed that an explicit web/client request still filters normally. Routing documentation reflects this rule. The repair was local during review and included in the 21 October release.
Human and agent migration paths now share a documented-role assessment. A pairing such as Next.js to Go CLI requires a named component boundary and retained interfaces, or an explicit interface redesign with changed/removed behavior. Human briefs remain unavailable until those details are supplied; interface redesigns cannot claim compatible replacement. Pair changes and shared-link reloads require fresh scope details. Details stay in page memory and are omitted from URLs. Agent routing preserves both selected profiles and asks about role, component and hard-constraint conflicts before replacement planning; an explicit component documented by the source narrows the role assessment.
Verification at the 20 October review covered all 196 profile pairs, real routing requests and constraint cases, ordinary same-role transitions, required fields, pair resets, copy/download content and shared-link reloads. Desktop and 320/390-pixel phone views passed, including both themes. The normal build and Pages Functions bundle passed; 1,087 internal resource references resolved, and all 28 complete per-profile Markdown/JSON records matched the pre-redesign baseline. An independent source review found no further actionable defects. These checks verify the website’s behavior and guidance, not execution or compatibility of application migrations. At that review point, production was unchanged and the video was still pending; the 21 October release below supersedes that status.
Almir supplied a real ChatGPT screen recording for a family meal planner and shared shopping list. His exact seven source ranges and 4×/16× speeds produce a 16.5-second silent H.264 MP4; 1:26–1:27 and everything from 1:34 onward are omitted. The 2304×1612 capture proportions are preserved. The original and reproducible edit recipe remain with the maintainer; only the 2.1 MB web export and an actual-frame poster enter the website source.
The centered hero and agent panel now use the real recording, replacing both the pending panel and simulated conversation. Native controls, inline/fullscreen playback, speed disclosure and a text alternative are included. Playback is visitor-initiated; opening the page does not preload the recording. The local adapter serves correct media MIME and byte-range responses; published media bypasses Functions through the normal static asset route.
The normal build passed. Browser review checked desktop and 390/320-pixel phone layouts, light/dark presentation, actual video playback/pause, seeking, fullscreen, player coordination, copy feedback and the text summary. Direct checks covered video GET/HEAD/ranges/invalid ranges, media MIME, all 24 page titles/descriptions, 1,093 internal references, the 23-page canonical sitemap, robots, real 404, negotiated Markdown, security headers and local noindex. The existing 1200×630 social card was regenerated and visually inspected; Open Graph and large-card metadata agree across the built pages. Local analytics remains production-origin gated. An independent source review found no further actionable defects.
This release superseded the earlier local-only state. Signed commit 2ce8993dfff5e16b4764ccc07426f5df5a7fb0ec and Cloudflare Pages deployment e031b7ef-7198-4372-a304-b1b477392769 identify the release. At approximately 20:42 EAT, the canonical site served the redesign and real demonstration; 23 sitemap pages, 37 direct resource/media cases, and production browser checks passed. Earlier Google/Bing indexing and reporting observations retain their original dated limits.
The catalogue now recommends one editorial starting point for each of 13 defined jobs, with the reason, tradeoff and conditional alternatives. Astro fills the content-site role; Expo fills a shared iOS/Android project role. TanStack Start 1.0, Hugo, plain HTML, Flutter, Hono and other ecosystem choices are scoped alternatives rather than additional equal-looking guides. All 16 full guides remain available for deliberate stack choices and rewrites; no healthy existing profile was reclassified as obsolete or source-only.
Local repairs preserve backend responsibilities in frontend-only rewrites, visible keyboard scope focus, feasible realtime advice and record-driven catalogue views. Next.js, Rails, Laravel and Django document their API compositions. Generator guidance corrects FastAPI initialization, Bubble Tea v2 and explicit Next.js scaffold choices. Agent marks now have pinned upstream copies and complete attribution. Current tool corrections and per-record review dates remain separate from the imported 13 October source-availability receipt.
Scoped disposable qualification: Astro's official basics starter installed, built and served HTTP 200 using create-astro 5.3.1 / Astro 7.4.2 / Node 24.21.0. Expo's pinned SDK 58 template generated and installed, Doctor passed 20/20 and an iOS JavaScript export succeeded. Its generated lint and TypeScript checks failed; no native build, device, signing, store or migration qualification is claimed. The complete recipe records retain versions, commands and limits.
Earlier candidate checks covered task defaults, hard-constraint conflicts, chosen-stack preservation, format negotiation, media ranges and local noindex. Browser review covered desktop and phone layouts, dark presentation and guide/comparison navigation. The final card presentation has one primary guide button and shorter secondary links; alternatives fill the available width with aligned footers. These are candidate checks, not production acceptance.
The history reconstruction built each intermediate revision and checked its stored message, signature and assigned changes. Final direct checks resolved 1,192 initial HTML href/src references, 26 distinct titles and descriptions, 25 canonical sitemap pages and all 32 per-profile Markdown/JSON resources. The 16-profile catalogue has 13 task defaults, and all 256 documented role/responsibility pair assessments ran without an inconsistent clarification result. These are website and documentation checks, not application setup or migration-compatibility qualification.
This reviewed base was subsequently pushed as 9aa317fd6d85c4108343bb4a4062b77eed58cce9 on 23 October; GitHub signature verification passed for all 17 commits. Deployment was queued at that readback. This is source-publication evidence, not a later live-delivery verification. The follow-up below remains local.
Almir authorized these local changes and their signed commits. Pushes remain on hold until his approval.
The reviewed base 9aa317f was pushed on 23 October. Its recorded deployment was queued at readback; this is source-publication evidence, not a later live-delivery check.
Expo alternatives are arrays of complete strings in the curated record. The generated lint/TypeScript starter failures remain a separate, unresolved qualification limit.
A requested component absent from the source now produces an explicit clarification instead of treating target coverage as sufficient source evidence.
Mobile component labels no longer inject SwiftUI or Compose resources into a chosen Expo build. Client architecture and a shared backend remain separate questions.
Selecting a chooser task updates its result in place instead of triggering reveal scrolling.
Chooser option and clear-selection controls grow with wrapped text, preserving readable labels and focus indicators on narrow screens.
Inline text links have underlines and visible hover/focus states in the shared interactive styles.
Generated page Markdown separates form labels and select options instead of concatenating their text.
The Rust CLI alternative makes asynchronous-runtime complexity conditional on the workload; ordinary synchronous CLIs do not require an async runtime.
The chooser separates one primary recommendation, conditional alternatives and all 16 reference guides. Cards use readable primary actions and shorter secondary links rather than treating every guide as a recommendation.
A short existing-services/tools reminder is nonblocking. Stack Menu remains about choosing stacks for software work and planning rewrites, with explicit choices preserved.
Nine neutral task choices replace technology-shaped jobs. Human and optional agent routing share defaults, explicit preferences, legacy task aliases and applicable alternatives. An undecided mobile scope retains native candidates and asks a consequential question; explicit platform choices narrow the recommendation. Matching Python/Elixir/React defaults do not emit unsupported-preference warnings.
Older ?kind= links resolve to a task and, for iOS/Android, the original platform preference. Changing or clearing the choice removes the stale parameter.
Create React App, AngularJS, Vue 2, WordPress and WooCommerce have HTML, Markdown and JSON source references plus offline discovery. They are source-only entries beside the 16 full-guide targets; supported WordPress/WooCommerce are not called obsolete. Scoped briefs preserve kept/replaced/retired responsibilities, owners, interfaces and checks. Unknown sources receive research guidance. Upgrade assessments need no unrelated target. Downloads read the current mode at click time. The AngularJS UpgradeModule citation, linked WordPress inventory and Rewrite navigation are corrected.
The browser permits identical source/target profile IDs and produces maintenance/upgrade guidance about actual versions and contracts. The router already supported same-profile questions; this change closes the human-flow gap.
The rebuilt candidate passed the 27 selection cases, 336 documented-role pairs and 57 HTTP requests in the final review, with 1,371 internal references and no missing targets. The independent repair harness passed 63 assertion groups, 16 HTTP reads, all six ordinary/scoped download-mode combinations and 23 Python/JavaScript alternative-list comparisons. These are direct checks, not an application regression suite. Sol’s repair verification passed. Opus’s repair source review accepted the nine repairs, found a minor error-message regression that was corrected, and its final narrow source check passed. Opus did not execute the browser or runtime checks.
Actual browser checks confirmed undecided/iOS mobile selection, applicable alternatives, matching-preference warnings, the linked WordPress inventory and Rewrite navigation. Safari saved a Rails-to-Django Markdown brief after changing staged transition to target architecture; its file content matched the new mode. A fresh Vue 2 assessment download retained its target-independent upgrade guidance. The new mobile question, selector and card were visually inspected at 390 pixels. Earlier 320-pixel and wide-view checks retain their stated scope; this repair pass did not repeat dark-theme or media checks. Maintainer-held evidence is in stack.menu-review/2026-10-23/review-repairs/, including exact saved downloads, executable checks, reviewer receipts and screenshots.
All 15 local revisions built successfully and passed direct chosen-stack, task and migration routing checks. Their stored messages and valid signatures were read back; the final site sources match the reviewed candidate.
The complete candidate is signed and committed locally. No changes have been pushed. Review and browser checks do not establish production delivery, application setup or migration execution.
Python-urllib readers still receive Cloudflare HTTP 403 / error 1010; the cause remains unresolved. Maintainer-held surface and access records own the operational diagnostics.