Dev tasks

Build plan for one header-level search box across every internal source we own (docs, tasks, CRM, prompts, voice notes, images, spaces, people): an instant Postgres full-text keyword layer plus a Gemini-powered pgvector semantic layer that finds things by meaning, merged into one grouped result list. Replaces the indocs-only command palette and folds the indocs search into the same engine. Internal sources only — Drive/Gmail/Slack explicitly deferred. Five phases; per-row access control flagged as the real hard part.

10-phase agile build plan covering the whole ~102-page indocs corpus: unified Supabase registry, Google-Drive-inspired docs browser (sidebar + grid/list + global search + right-click menus + star/pin), request-time markdown rendering, in-browser Lite Editor with GitHub Contents API round-trip, Google-Docs-style per-doc ACLs, generalized /docs skill replacing /devtask, lightweight cross-format embeds, and worktree-first feature-flagged rollout with explicit verification + rollback per sprint.

Build plan for a dashboard that shows what each team member accomplished, with merged session time, completion notes, result links, and logged hours comparison — one page to answer “what did everyone do?”

Build plan for an autonomous daily agent that reads archetype briefs from Obsidian, searches the web for matching candidates, scores fit with a local model, and populates the Relations CRM pipeline via PostgREST. Runs on Alpuca via Hermes scheduled tasks.

Build plan for a human-operated Telycam Vision+ SE control page in the Control intranet, backed by a local LAN controller and a command-history ledger that records both manual UI actions and machine/API commands.

A focused cut on top of the existing realtime stack: a richer onboarding interview, dual-artifact profile pipeline (user-visible profile + host-only dossier), a dossier-driven host prompt with reactive triggers, and a self-enroll Android lobby. The minimum needed to run the first real test of the personalization hypothesis — that the AI can produce demo-worthy nudges drawn from each guest's interview.

How our team works with Claude Code: two lanes (worktree vs direct-to-main), five skills (/brainstorm, /feature, /journal, /ship, /wrap), and a build journal that keeps multi-week projects coherent across sessions. Lead with the quick-reference table at the top.

An on-demand intranet page where any authorized user can generate a structured work summary for any time range. Two view modes (By Person, By Surface Area), AI explanation strips that describe what each cluster of activity accomplished, org-level summary bullets, traceability links, and a self-improving prompt loop driven by feedback reactions.

Build plan for an automated testing & monitoring system that verifies every Sponic surface nightly, auto-extends coverage as new dev work lands, and surfaces activity / agents / findings in a dedicated intranet section. Includes full frontend spec and a 5-phase sequenced build plan (~3.5–4 weeks).

Build plan for a speaking AI host that listens via mics in a LiveKit room, replies with a Live2D-style Simli avatar, plugs into the existing Sponic translation app for multilingual support, and reuses the same agent code over SIP for outbound phone interviews. 8 phases, ~$15–25 per dinner, ~$1.20 per 30-min call.

Parked build plan for the iOS Sponic app — a SwiftUI iPhone client that joins the AI dinner-host LiveKit room, mirroring the Android app feature-for-feature. Blocked on Apple Developer Program enrollment; covers prerequisites (D-U-N-S, account approval), bundle-ID decisions, 5 dev phases (~3.5–4 days), and TestFlight distribution.

How Sponic manages documentation so both humans and AI agents can search the same source of truth. Markdown-in-git, MkDocs Material for the human layer, Postgres + pgvector for agent search. Includes deep SSG comparison.

Where every credential lives (BW devops-sponic), the unlock recipe, image-gen wrapper rule, per-machine session-recording hook patch, and a paste-ready prompt that bootstraps a fresh Claude Code session.

End-to-end map of how Sponic Gardens email works: Cloudflare Email Routing for inbound, Resend SMTP for outbound, Gmail in front of both. Current rule list, verified destinations, DNS records, gotchas, and how to add a new alias.

Configure Gmail "Send mail as" with Resend SMTP so you can compose and reply from [email protected] (or any alias) inside your normal inbox. Replies route back via Cloudflare Email Routing.

Move every generated image (and metadata) into a queryable catalog: Cloudflare R2 bucket + Supabase public.images table + a generateImage() wrapper. Self-contained prompt for a fresh Claude session.