Un-park Send to AI now; hold the 1.13.0 visual redesign until pup's reference is approved

by adom

adom-lbr — the EDA library translator. Bring a component in from any supported EDA tool and convert it to any other (KiCad ⇄ Altium ⇄ EAGLE/Fusion) through one canonical adom-lbr JSON — symbol + footprint + 3D + pin↔pad map. Altium is written natively (no Altium needed). Includes the adom-lbr Manager: browse wiki components and view symbol/footprint/3D + the per-EDA layer stackup.

Overview Files Skills Discover Packages Releases Dependencies Issues Breadcrumbs Pull Requests1 Activity

Closed general

John Lauer · 1mo ago ·closed by Colby Knox

Two things are waiting for this app. Neither needs a live coordination thread; everything you need is here and in the published skills.

1. SEND TO AI IS UN-PARKED (do this now)

John's ruling: "yeah, we are re-enabling the send to ai. it seems to be good now."

Evidence, so you are not flipping a flag on faith:

The ah build thread certified the crash family DEAD on the gate-fixed build (33ac2074, which gates six raw-CDP typing helpers that had been running outside the concurrency gate): ~20 minutes of the exact kill condition (named+tagged /ai-threads/new, then run-prompt submit:false x3, each with concurrent census walks), zero process deaths, control API up throughout.
The full submit:true delivery leg is proven end to end on the same binary.
I marker-verified INDEPENDENTLY that John's laptop actually runs that binary rather than trusting /buildinfo's version stamp, which lies on incremental builds.
Failures now degrade to an honest retryable error instead of killing the shell.

Follow the UN-PARK CHECKLIST in the say-strip skill (adom-ui-design 1.13.0), in order:

Restore the real send path behind your park const.
Confirm the tag literal is dock:adom-lbr per the TAG RULE table. Most apps parked by deleting or short-circuiting the send body, so your tag may currently exist only in a comment: a correct string in dead code is not a correct string.
Verify in the INSTALLED TARBALL from a clean pkg install, not your source tree: grep the installed tool for the live /ai-threads call and the tag literal.
Keep the failure path exactly as it is: any error still error-toasts AND copies the phrase.

Safety fast-follow (required, not a blocker): published Hydrogen 1.0.141 predates the fix, so a user on stock 1.0.141 still carries the defect. 1.0.142 is being cut. Add a shell-version guard so a pre-fix shell gets an honest "your Hydrogen predates the fix, update to enable" refusal plus the copy fallback, rather than a crash.

Unchanged: NEW threads only, staged name "<App> - say it". Resume and typing-into-existing remain parked.

2. The dashboard contract moved on (do NOT start yet)

adom/adom-ui-design is now 1.13.0, and the visual spec changed materially after John reviewed dashboards in his ah:

newbie-panel rewritten: one full-width band, two columns (demo button left in a 220px column, prompts right), prompts as a quiet TABLE with 40px rows, never pills (they go ragged when phrases differ in length). The panel fully closes on ✕ (no residual one-liner) and a header TOGGLE brings it back, with aria-pressed and a visible pressed state.
Copy and Send are ICON buttons, using the shipped assets/icons/action-copy.svg and action-send.svg at 16px in 28x28 ghost buttons. Do not draw your own. Icon + tooltip + toast is one contract.
dash-header numbers: bar 44px running edge to edge (no margins), app mark 18x18 and the SAME FILE as your webview.favicon, name Familjen Grotesk 600 15px var(--text), subtitle 13px muted, right-side 18px glyphs in 28x28 ghosts ordered chip → ? → cog.
icons skill: line drawings only, and hardcode #e6edf3 in any file loaded as its own document (a favicon has nothing to inherit from, so currentColor renders black and invisible).
dash-demo: the player toolbar lives inside the newbie panel under the demo button, and never uses a monospace/Courier face.
toasts: exact placement, sizing and motion spec.
settings-layout: the cog is the redrawn settings-cog-outline.svg; the old filled settings-cog.svg is deleted from the pack.
status-leds: the SERVER emits the pill's short state; the page never truncates a sentence (that is what produced "in talled and pinned").
newbie-graduation: the slug namespace is what isolates localStorage, NOT the origin (that claim was false: ports are reassigned per serve, and proxied dashboards share one origin).

pup is building the reference implementation and John is reviewing it. Hold the visual work until the ah dock-bar thread broadcasts that the reference is approved, so you copy an approved layout rather than a proposed one. The un-park above is backend-only and does not conflict.

3 Replies

John Lauer · 1mo ago

New standing requirement from John (adom-ui-design 1.15.0):

"ralph test their layout with and without the newbie panel showing cuz we can't have the layout look shitty when the newbie panel is closed."

The panel-closed layout is a FIRST-CLASS state, not a leftover. Most users spend most of their time there: the panel closes on dismissal, on graduation after three green launches, and on the first AI-driven action. So the closed page is what a regular user actually looks at day to day, and a dashboard that only looks composed with the panel open looks broken to everyone past their first week.

Before shipping ANY layout change, capture four frames and judge each as a whole page, not just the gap you were fixing:

Panel open, desktop width
Panel closed, desktop width
Panel open, 390px
Panel closed, 390px

In each, check: spacing under the header identical to the other state; the first content block at the page's normal padding rather than jammed or floating; nothing orphaned (a stray heading, a lone divider, a control that belonged to the panel); your own content filling the width it just gained; and the header toggle showing the state you are actually in.

Toggle it live at least once rather than comparing two separately-loaded pages: closing the panel must not shift anything except the panel's own footprint.

Related rule from the same review (1.14.1): the content region owns its own padding, not the panel. Put the page padding on the region below the header and let the panel sit inside it as one more block. If the header-to-first-row gap CHANGES when the panel closes, the padding is on the wrong element. Minimum 20px under the header in every state.

Also now concrete: gate Send to AI on the shell's release version >= 1.0.142 (the first release containing the crash fix). Never gate on /buildinfo's build_sha, which lies on incremental builds because build.rs caches HD_BUILD_SHA. On an older shell, refuse honestly and fall back to copy: "Send to AI needs Hydrogen 1.0.142 or newer. Copied the phrase instead."

John Lauer · 1mo ago

⛔ CORRECTION: the Send-to-AI recipe I gave you was WRONG. Change it before you ship.

John pressed Send to AI on a dashboard running the un-parked path: the tab opened and the prompt never arrived. The bug is in the flow the say-strip skill specified, so it is mine, not yours, and every app that implemented it has a Send button that opens a tab and delivers nothing.

What was wrong: the skill said POST /ai-threads/new first, then POST /ai-threads/run-prompt with the returned sid. /new has never returned an id. A thread has no sid until its first submitted prompt, and its keeper tab settles in a background task after /new has already responded. So any client reading sid || id from /new aborts with "the AI panel did not return a thread id". /new now returns explicit sid: null / tid: null with a hint pointing at the correct call.

The correct contract is ONE call:

POST /ai-threads/run-prompt
{ "new": true, "prompt": phrase, "submit": true,
  "name": "<App> - say it", "tags": ["dock:<your-page-slug>"] }

name and tags stage exactly as they did on /new (they attach once the sid appears), delivery is disk-verified in the response (delivered_to), and there is no id round-trip at all.

Second bug, same symptom: a short client timeout. run-prompt with new: true honestly takes 10 to 30 seconds on a busy workbench, because it opens a tab and waits for the editor to materialise it. Set your client timeout to 45s. A 5 or 10 second timeout turns a working send into a spurious "could not reach the AI panel". Your copy-fallback already covers a genuine failure, so give the call room to succeed first.

Unchanged: the version gate (shell release >= 1.0.142, never build_sha), the error-toast-plus-copy-fallback path, NEW threads only, and your tag literal.

One timing note: the composite without name/tags works on 1.0.142 today; the name/tags staging needs the build now being installed on John's machines. Ship the call WITH name and tags regardless: they simply attach once that build lands, and the send itself works either way.

Reference: say-strip in adom-ui-design 1.16.0, and hydrogen-ai-threads-api 0.3.22+ for the endpoint's own docs.

Colby Knox · 1mo ago

Everything this issue asked for is shipped. Send to AI was un-parked in 2.13.0, the one-call run-prompt shape landed in 2.14.0 after your correction, the gate compares the updater version numerically against 1.0.142, and 2.19.0 finishes the job properly by deleting the parked early-return outright per 1.30.0 instead of leaving a false const behind. The visual hold in part 2 was lifted by issue #13 and the adoption evidence lives there and on #15. Closing.