Implementation recipes
These are not feature bullets. They are the guides that ship inside the product, on the Implement tab of each module. Each one carries a goal, prerequisites, the full file inventory, ordered steps with real code, verification at every stage, and a definition of done. Below is what each one covers.
Showing 8 of 173 recipes.
An in-product assistant that answers from the signed-in organization's projects and tasks, streams its reply as it forms, and still holds the conversation after a hard refresh. Its tools are pinned to the caller's organization so the model can never read another tenant, plans can gate access, and the provider key never reaches the browser.
tools that cannot read another tenant
Assistant replies grow in every open tab as the model writes, at a fraction of the write cost a naive token-by-token stream would spend. A refresh mid-answer resumes where the text left off, awkwardly split network chunks never drop characters, and a straggling chunk cannot reopen a reply that already finished.
streaming backpressure and late chunks
Documents nobody has touched in a month surface in a weekly report. A new agent card appears with its own toggle and Run now button, scheduling, retries, and run history already handled, and runs produce a readable report even before any model key is configured.
per-organization opt-in for scheduled runs
Agents propose, people decide. Every run's output lands in a review queue where a member approves or rejects each suggestion, only approved ones ever write product data, and each decision records who made it and when. Rejections and malformed output leave your records untouched.
rejected output leaving no trace
AI spend stops being a surprise. Every run's tokens count against the plan's daily allowance, an organization past its limit fails fast before the provider is ever called, and organizations nearing the cap finish the day on a cheaper model. Spend and current tier show on the usage and agents screens.
refusing the call before spending money
Model replies arrive as typed, validated data instead of prose you parse by hand. A reply with the wrong shape gets exactly one repair attempt, and a second miss marks the run failed with a readable reason — nothing malformed is ever stored or acted on.
one repair pass, then fail cleanly
Agents react to what happens in the product, not just the clock. A burst of fifty writes starts one run after a short settling window, repeat events inside that window coalesce instead of queueing, a cooldown stops back-to-back runs, and the existing scheduled sweeps keep working untouched.
debounce windows and duplicate triggers
New tables join the ⌘K palette as first-class results under their own heading, worked through end to end on an invoices table. Isolation lives inside the index itself, so one tenant's records never surface in another organization's palette, and the palette component needs no changes at all.
per-tenant isolation inside a shared index