Stop the person-facing screens saying things that are not true

[guidovizoso]

Four defects found in a UI/UX pass over the app, plus the regressions that review then found in the fixes for them. The thread running through all of it: a screen should not assert something it has not established.

What changes

A routine's firing no longer reads as a message the person wrote. A scheduled routine arrives as role: "user" carrying a model-facing frame — three sentences of scaffolding — which the transcript drew in an ordinary bubble, and the channel titler used to name the conversation. The frame is now declared once in shared/routine-firing.ts and read back on both sides, so the server still persists exactly what it persisted before. The transcript draws a firing as a scheduled event, and the scroll machinery stops treating it as the person having just sent something.

A tool line stops printing its plugin's name twice. mcp__notes__search_notes read "Search notes notes". The server is dropped when the label already names it as the thing acted upon — matched word by word, singular-aware, and excluding the label's leading verb, so mcp__posts__post_message keeps its "posts" attribution rather than losing it to a coincidence with "Post".

Both agent screens say what is going on. Empty states where there was a heading over nothing; loading states that hold their own height; error states styled as the empty state with destructive text and border. isError is gated on the query never having returned, so a failed background refetch shows the roster it still has rather than blanking it — TanStack Query keeps data across a failed refetch, so error-with-data is an ordinary state, not a contradiction.

The home screen's shared Bots scroll instead of squeezing. A fifth public Bot used to compress all five; they are now a carousel with controls on the heading's baseline.

Also here: PageShell gained a scroller, because _authed/_app is deliberately a non-scrolling viewport and Skills was clipping 246px below the fold with no way to reach it. And the sidebar toggle stands down while the sidebar is already on screen, never while it is the way back.

Review

The branch went through a full review-fix loop. Round 2 found that three of round 1's own fixes were regressions, each corroborated by more than one reviewer and each reproduced before being accepted:

• readFiring had started returning null for a frame wrapping a blank instruction — but null means "not a firing" to every caller, so that drew the scaffolding as the person's own words. The exact defect the branch exists to remove.
• The new error arm outranked cached data, so a transient refetch failure blanked a working roster and, on /, printed "there's no one to send this to yet" beside a still-enabled composer.
• singular() matched the label's verb rather than its object, dropping the server for mcp__posts__post_message and mcp__lists__list_files — attribution main had kept.

A later round found a fourth: the empty-slice case, where one section rendering real cards proves the sibling's emptiness is a fact rather than a failure.

Every state was then checked in a browser against the real app rather than only in tests — the destructive states measure oklch(0.577 0.245 27.325) for both text and border, matching the --destructive token exactly, at the same 180px the other arms occupy.

Notes for the reviewer

• The branch is behind main; happy to rebase.
• Repo-wide bun test has pre-existing failures in agent-handoff, agent-computer and agent-bot/agent-langgraph (the last two are missing-module errors, reproduced on main). None are in files this branch touches. bun test app is 370 pass / 0 fail.
• One product gap was found and deliberately not fixed: on /, a query that succeeds and returns zero agents leaves the composer disabled with no explanation. It previously said "couldn't be loaded", which was false; removing that lie exposed the silence underneath. New copy is a product decision.