Give each project a clear boundary

Friday clarified who owns what across Steve’s projects: repositories, production services, error evidence and component behaviour. The coordinated deployment established independent delivery paths. Later investigations produced reviewable diagnostics, while a measured grid experiment supplied an inconvenient but useful result: using the library did not save builder tokens in that pilot.

• GrowthScout, db3.ai and Framework · complete the repository split and production cutover. Scout’s folders and deployment responsibilities were separated from the company apps and framework examples. Migrations and schema moved beside Scout’s server registries without changing their contents or history. The coordinated release verified Scout, db3.ai website/Cloud and framework documentation on identified revisions, with public health and browser checks. db3.ai now owns its database, storage and deployment configuration; its runbook covers host moves and rollback. The framework gate passed 969 tests, and db3.ai passed 64 tests plus its build. AI execution and customer database provisioning remain disabled pending dedicated setup; a Workers-native deployment would require adaptation.
• GrowthScout and db3 · retain why an agent finished before enforcing its save guard. The article investigation confirmed that the saved-article guard correctly rejects a run that finishes without saving. The missing evidence was terminal provider status and incomplete reason, lost before validation failed. Draft changes store a bounded completion summary in existing attempt records and add the stable article_not_saved classification, without extra article text or altered retries. SDK/queue regressions use simulated provider responses. Final framework checks passed 227 AI tests, docs and package consumers; Scout passed its 71-test local gate. Both PRs remain unmerged, and the Article Runs screen does not yet display the summary. The historical cause remains unknown.
• GrowthScout · diagnose keyword-review claims without changing contention behaviour. A separate diagnostics PR records bounded claim stages and database error fields, excluding SQL and article text. Real SQL tests show that failed inserts roll back the conversation and claim, preserve the article and make no provider calls; concurrent opens and refreshes admit one review. The final local gate passed 3,508 tests with one skip, builds, types, conventions and all 69 migrations. The production contention remains unexplained. Hosted CI could not start because of an account-level execution block, and the PR remains unmerged and undeployed.
• DOM Studio · measure the grid workflow before claiming efficiency. Two builders implemented the same 200-invoice tracker, one with a native Vue table and one with DOM Studio 0.2.5. Both ultimately met all 18 acceptance checks. Builder usage was 2,046,878 tokens for the custom grid and 5,241,164 for DOM Studio: 156.1% more in this pair. Most input was cached, and these are token totals rather than financial costs. Different browser interfaces, permission friction and an extra repair invocation confound the comparison; supervision and independent review usage were excluded. Findings and evidence remain local. A repeat with identical tooling and a compact component contract is proposed, not a demonstrated improvement.
• DOM Studio · extend the flush-layout contract to month calendars. DomMonthCalendar gained layout="flush", application-supplied gutter classes and an optional bottom-border control. The surrounding page owns the outer borders; edge events align to its padding while internal dividers remain. A documented example, generated types and production build passed, alongside desktop/mobile checks for continuous mode, week numbers and no horizontal overflow. This remained local and unpublished, separate from Thursday’s grid package release.
• db3.ai · repair site checks and explain the planned Studio journey. A three-file commit updated CI database networking, the database-app test setup and browser-cursor screenshot behaviour, then was pushed to main. The resulting remote site-check run failed at npm test, so remote validation remains incomplete. Separately, local homepage copy explains desktop or browser access through a db3 account, collaborative design and Codex-assisted building, with Cloud resources underneath. It is labelled in development; the build and two site tests passed, but that later copy was not deployed. The earlier verified Cloud cutover does not establish these planned capabilities as available.
• GrowthScout and Framework · record frontend responsibilities and simplify the launcher. Guidance now assigns UI and interaction wiring to Vue components, substantial reactive behaviour to composables, and Vue-independent logic to plain TypeScript functions. Extraction follows responsibility rather than creating oversized composables. These documentation edits were included in the coordinated all-current-work release; they do not constitute a component refactor. A later local launcher change makes npm run dev start Scout and updates the guides. Fourteen launcher tests, conventions and whitespace checks passed; that later change remains uncommitted.
• Personal site · refresh the public profile and remove an unused webhook. The GitHub profile now centres on db3.ai, Test Engineer and the planned Codex integration, with a corrected website link and framework repository pinned first. Development instructions moved into their own guide. The unused Outrank webhook server, processing code, test script, commands and environment settings were removed. Builds passed, commits were pushed, and the final profile quote’s deployment workflow succeeded. The quote and profile changes were verified publicly; the integration and proposed funded pilot remain future work, without an approval or delivery claim.
• GrowthScout · design planning for one editor and multiple viewers. The growth-planning draft now centres on presenting and editing a plan in the same view. A logged-in owner or agency representative records decisions and confirms the plan; separate viewers have read-only access. Solo owners can complete the workflow without a meeting, and multi-editor collaboration is outside the scope. A separate logging review recommends automatically propagating deliberate request, job and authorised identity context, with bounded recent activity rather than arbitrary memory dumps. Both are design directions, not newly implemented planning or context-propagation systems.

The diagnostic work kept the distinction I care about: preserving better evidence is a completed improvement, while explaining a historical failure requires that evidence to exist. The grid pilot makes the same point from another angle. Keep the result, keep its limitations, and make the next experiment easier to compare.