I spent Saturday mostly in two worlds. The first was listing-ops, where the vision pipeline hit its wall on RTX 6000 context. I squeezed image sizes, capped OCR tokens, bumped per-slot batch from four to two, and routed text tasks to a dedicated endpoint. It was dependency hell in the best way: every fix exposed the next, but by the last commit dots.ocr was grounded, the text tasks were on their own route, and the whole thing actually fits now. I also fixed owner balance to net payouts properly and gave David that "view as" toggle so he can check his own numbers.
The second world was bigger and more interesting: tax-planning-ops v3. I had a design spec that mapped every surface from the Desk down to Season, then spent the day building it. Daylight and Paper token system, dark mode that actually works, a primitive SVG chart kit, routing scaffold. Then the swarm ported in waves: Book, Brief, Playbook, Case File, Explained, Staff, Return, Plan, Entities, Lab, Intake, Memo, Present, and finally the Desk itself as the firm command deck. Old v1 paths redirect, login lands on /. It's a complete frontend rewrite sitting on top of the same engine. Twelve commits for mockups, design, and routing, then another fourteen to actually render every surface.
Accounting-ops was a quieter day but still meaningful. I shipped the in-app Guide help agent on /help, which reuses the chat system and strips the thinking tags so end users see clean answers. Fixed the Cloudflare 403 on chat by fixing the User-Agent. Surfaced real LLM errors instead of infinite spinners. And for the dashboard, I built a GL daily-balance endpoint that gives the cash ribbon real historical data, plus live burn/cash/runway mini-cards in the command palette. Released 0.6.0.
I had a design spec that mapped every surface from the Desk down to Season, then spent the day building it.
Center-deep-research got its most significant upgrade yet: a full consensus layer. Sources are now graded on credibility, subagent drafts feed into the gate, conflict detection is wider, and the synthesis is credibility-weighted. Fan-out budget went from four to six sites per subagent. The GUI got a Space dashboard, live AgentTimeline with subagent counts, and a polish pass that made everything feel alive. Fixed the Traefik router collision and the healthcheck that was failing on 406.
Contact-ops quietly got GDPR export and erasure engines, public self-serve signup, and a self-contained cell deploy topology that makes standing up a new cell just a config exercise. Listing-ops got a tenant kill-switch and tenancy hardening across the board. Ops-center shipped the floating Guide bubble for end users and the full admin federation contracts surface with pricing policies.
One hundred and one commits across seven repos. Two main arcs: a vision pipeline that finally fits, and a tax platform that got a face worth looking at.
Real product captures — click any to enlarge.