Rolino fixed changes

Repository secret scans now exclude two exact non-secret test labels without weakening scans for other values or files.

Production dependency installs now use patched URL parsing and MySQL driver versions required transitively by the application, authentication, and database toolchain, so the high-severity package audit passes.

Remote MCP now creates credential rate-limit buckets only after authentication from the verified workspace grant, so fake bearer values cannot create unbounded database rows. Old request-limit rows and unused anonymous OAuth client registrations now expire through bounded maintenance, while active connections remain valid. Codex CLI 0.152.1 is the verified minimum and 0.145 is clearly unsupported.

Concurrent platform feature grants now resolve as one idempotent grant instead of sometimes failing on their unique workspace boundary. The command palette also ignores malformed keyboard events before reading shortcut keys, and shared interface state now updates without cascading effect renders.

OAuth consent now removes only stale consent records that have no matching active workspace grant, while keeping valid connections. OAuth and remote MCP requests now use shared bounded rate limits, and database tests prove cross-user isolation, cross-workspace isolation, scope enforcement, and immediate revocation. Codex URL-only setup can now register a safe native public OAuth client through the rate-limited standard endpoint without a host-specific client record.

Removing the obsolete host-specific MCP OAuth client now also revokes its active workspace grants, clears pending consent selections, and keeps deployment readiness tied to the latest database migration.

CLI and remote MCP browser authorization now show the complete opt-in capability list, so users can grant draft and scheduling access instead of receiving an unavoidable read-only connection. Hosted Codex setup now uses native Codex configuration and OAuth discovery, keeps protocol initialization public, and challenges protected tool discovery at `tools/list`. Agent setup names the two permissions required to prepare and schedule social posts.

Create Post now uses the project's last explicit destination choice from browser or agent drafts, preserves existing project choices during migration and saved empty choices afterward, falls back to its only connected account on first use, and removes disconnected destinations during duplication and automatic checks.

YouTube connection now replaces a stale local callback address with Rolino's configured public address in production, so onboarding cannot send users to `localhost`.

Weekly Blog plan checks now save a complete proposal set atomically and show health only when the completed job matches the exact active plan revision and week.

Weekly Blog plan checks now use each opportunity once, fail visibly when a selected proposal cannot be completed, and reject duplicate topics, expired evidence, changed page evidence, and incompatible work types before acceptance.

Blog performance recommendations now require exact-page evidence for page-specific signals; only the query-spread detector can use an observed query-to-page relationship.

Blog plan review dialogs now stay open after an action error and show the recovery message beside the unchanged proposal.

Blog replacement generation now bounds managed-article context, validates new ideas against those articles, and ignores replacement-job states from older item versions.

Blog replacement ideas now roll back cleanly when the editorial slot changes during activation, reject incomplete evidence snapshots, prevent conflicting recovery checks while generation runs, and show accurate creation dates in idea history.

Blog evidence setup now reports failed source analysis instead of remaining in a preparing state, refreshes while either recommended source is pending, limits the search-analysis action to Search Console, and keeps Google property selection usable on mobile screens.

Remote MCP now negotiates both the current Codex `2025-06-18` protocol and the newer `2026-07-28` protocol instead of rejecting Codex during startup.

Signed-in CLI and Codex OAuth requests now load the registered client, open the workspace consent screen, and complete authorization instead of failing or showing an invalid-request message.

Direct Codex UI connections can now fetch and validate ChatGPT-hosted OAuth client metadata on Node 24 instead of failing with `invalid_client` before authorization.

Codex Streamable HTTP setup now writes only the hosted MCP URL and OAuth mode, so Codex can use current OAuth discovery and Client ID Metadata Documents. It also repairs a stale Rolino configuration marker instead of stopping setup.

CLI OAuth coverage now verifies that a cold authorization request seeds the registered client and reaches sign-in instead of returning a server error.

The landing hero preview no longer overlaps the star divider or the **For creators, teams, and their agents** heading at tablet widths.

Home now shows one explicit UTC cross-brand schedule, reports the complete draft-review total, keeps the full weekly plan without a silent record limit, exposes semantic section headings, and uses larger mobile actions.

Home now separates dated weekly plan progress from cadence targets, removes empty seven-day grids, gates per-brand planning behind connection recovery, capitalizes greeting names, and gives the first blocker clearer priority.

Onboarding now keeps the analyzed website stable during review, handles concurrent completion requests safely, removes expired discovery records during regular maintenance, and keeps the color review within the shared theme rules.

Website onboarding now prepares its first review from one bounded homepage read and one small evidence-backed extraction focused on description, audience, voice, and supported positioning. It keeps a slow but valid homepage allowance, returns safe metadata and branding when extraction is late, and prepares deeper context without blocking Review.

Homepage branding now recognizes Firecrawl's top-level official logo result as well as nested logo and favicon candidates, without treating a general social preview image as a logo.

Brand discovery now avoids weak audience, voice, and color guesses; Brand Studio also keeps Voice and Audience separate and never exposes internal voice tokens.

The final onboarding screen now confirms that setup is saved instead of claiming that background research and ideas are already ready.

Onboarding now keeps its supporting visual at tablet-landscape widths and centers the form when the visual is hidden, without leaving an empty half-card gap.