[Birfy]

A real mouse click-through of a running FlyBot (web, DeepSeek V4 Flash on Ark) found interactive UI defects. This fixes every one of them, including the stacking defect that made a chat message paint over the Tasks and Settings dialogs — the owner's screenshots showed it plainly after my first pass wrongly reported it unreproducible.

Fixes #1347

What changed

The Details panel wrapped its own controls. It opened at 320px, so a settings row could not hold a label and its control on one line: "Name" came out as "Nam" + "e", "Test connection" as "Test connecti" + "on", Main Bot's "Change" as "Ch/an/ge". DETAILS_WIDTH is 380 now; a row's label is one line with an ellipsis; .bt-soft / .bt-mini in a row are flex:none;white-space:nowrap so a control never breaks its own words; and .bt-panel declares container:bt-panel/inline-size so @container bt-panel (max-width:300px) stacks the row at the panel's 240px floor, where nothing can be clipped to fit a width that is not there. A width already saved under the old 320 policy is kept (savedWidth floors at MIN_DETAILS), so the change does not silently widen a panel someone sized.

A sidebar preview could show the untrusted fence. Not only the complete wrap unwrapUntrusted handles: a line whose opener arrived already cut showed from="User"> B… in the row. core/src/shared/preview.js visibleText unwraps complete wraps and then removes leftover marker fragments until neither changes the text, and folds to one line. It runs where each row is built — previewOf (Bot rows), the group row in sidebar.js, and preview in host/channel.js — so no surface has to remember it.

A Bot chat with nothing said yet was a blank column. The shell draws headerBlank with no title for a session that has said nothing, so the pane under the composer read as white space. BotChatStrip now draws on the same blank fact #998 uses for a group: the Bot's name, its role, "No messages yet. Say something to {name}.", and Details. Only for a local Bot — a friend's chat keeps its own empty hint (#173) — and it disappears with the shell's own header, so there are never two. The first version of this drew the strip for a friend's Bot too; the dock test caught it and the condition is !bot.remote.

The copy button was named for a different action. "Share conversation" only copied Markdown and said nothing. It is now "Copy conversation", and a successful copy announces "Conversation copied as Markdown"; a refused clipboard write is reported instead of swallowed.

"Standard mode" looked clickable and was not. DSH draws that label as a span with no handler and a hover title. It now reads as a status: no fill, no radius, cursor:help, description still on hover.

A provider refusal arrived as a cut-off JSON body. A real Inbox showed 404: {"code":"UnsupportedModel","message":"The requested model does not support …, — truncated before anything actionable. core/src/shared/failure-copy.js failureSentence keys on the provider's stable code and returns a sentence naming the feature and where to change the model, in the chat's dead-turn lines and in the Inbox; an unmapped code is returned unchanged rather than replaced with a guess.

The stacking bug: reproduced and cured on your 0.2.7 tree

You re-verified on instance A, whose checkout is 0.2.7 (e109361). The fix existed only on this branch, so that instance necessarily still showed the defect — which is exactly what the screenshots show. To remove all doubt I ran the A/B on your exact 0.2.7 tree, porting only the stacking CSS onto it:

forced-promotion scenario, red pixels inside the dialog	pristine 0.2.7	0.2.7 + this CSS
Tasks	615	0
Settings → Connectors	59	0

The screenshot scenario itself measured with a long right-aligned pill: it overlaps the Tasks dialog by 12090 px² and Settings by 1587 px² (your ~1864 is the same shape) — overlap in the layout, and zero pixels painted inside the dialog.

Cause. The chat column had no stacking context of its own (position:static, z-index:auto), so nothing in it was boxed in. Anything the browser promotes to its own composited layer — an entrance animation's transform (bt-row-in), a filter, a will-change, a sticky box — was ordered against the root stacking context instead of against the column, and a dialog's layer is just a z-index box at the root. So a promoted bubble sorted above the dialog. A DOM walk (elementsFromPoint) reports the dialog as topmost in that state, which is why my first pass missed it; a headless screenshot forces a repaint and hides it.

Fix, three parts.

1. Load-bearing: the chat column is one stacking context — position:relative; z-index:0; isolation:isolate — sitting below .bt-panel (40), .bt-pane (45) and the dialogs (120/121).
2. Tiers: dialogs 120/121, popovers 130, palette 140 (they used to share z-index:60 with the popover layer), and every overlay layer is its own stacking context so scrim and dialog move as one.
3. The scrim actually separates the surfaces. It was rgba(0,0,0,.32), which left the thread legible behind a dialog — that is what "the pill at full opacity behind the Inbox" and "assistant text readable across columns" describe, and no z-index can fix it. Now rgba(12,12,14,.62) light, rgba(0,0,0,.7) dark; still one layer, still dismissed by a click.

Bug 2 was not touched: your re-verify and my frame-by-frame recording agree it unmounts cleanly.

Investigated, not changed, with the reason

• Trajectory toolbar left painted over Chat. The owner's fresh re-verify found it gone, and my frame-by-frame recording on DSH 0.2.0-rc.2 and 0.2.1-rc.3 agrees: the toolbar is removed within a frame of the click across Chat → Trajectory → Chat, an 80 ms and a 250 ms switch, after scrolling the table, from Inbox and from Settings, and through a sidebar chat switch. The 61px band under the tabs is ordinary scroll padding, identical before and after a Trajectory visit. No code changed for it.
• Chat bubbles above modals — fixed and A/B-verified on the 0.2.7 tree; see the section above. My earlier elementsFromPoint walk was the wrong instrument: it answers DOM order, and this is a paint-order defect.
• rate_limited flood in Feedback. This is 交付群 invite join stuck on Asked to join… — host never seats, no sidebar row (regression after #1082) #1161's retry storm, already fixed: seatOpenClaims reads one open link per minute and pauses for the server's retry on rate_limited / locked_out (room-membership.js). Pinned by core/test/room-ticket-seat.test.js ("one open link is read once per minute, not on every inbox tick"). No code change; the entries in the owner's Feedback list are from before that fix.

Verification at the tip

Tip SHA: 7f8c7da

Ran at the tip, TZ=UTC, DSH_HARNESS=core/dev/harness:

• npm run verify — every stage green: core lint + arch-docs (40 changed files), plugin lint, build-client.mjs --check, ds-bot-patches check (10 patches apply, core/ equals ds-bot/ + patches), core 1874/1874, social 525/525, tree 125/125, timeline 125/125, gate scripts 529/529, interop 37/37, live-e2e lint.
• One full-suite run failed remote-room.test.js on ENOTEMPTY: directory not empty, rmdir in its after hook — a teardown race, not an assertion. The file passes 69/69 twice on its own and the whole suite passes 1874/1874 on the re-run.
• Red without the change: on origin/main with only the new test files copied in, 10 of the 11 ui-clickthrough tests fail plus the new dock test and narrow-pane. All of them are defect tests — each names a property the base violates. The one that passes on base is the preview-sanitizer test, because the helper it exercises is new in this PR.
• Browser check against a live bench on DSH 0.2.1-rc.3: panel 380px, "Test connection" 112×28 on one getClientRects() line with scrollWidth === clientWidth (nothing clipped), "Change" 67×28 the same; the blank Helper chat renders "Helper · Researcher · No messages yet. Say something to Helper. · Details".
• The stacking fix was verified by pixels on a live bench of this branch (Tasks / Settings / Inbox all 0 inside the dialog) and A/B'd on the owner's 0.2.7 tree (615 → 0 and 59 → 0). Popovers still float: account menu and in-panel model menu both topmost at 130.

Not run, and why: the smoke bench (no index.js or src/host/ behaviour change — channel.js is one preview string), Desktop, real-model steward.

Leftovers

• The stacking fix is structural (one stacking context per surface plus explicit tiers), so it covers a promoted bubble from any future rule; the numbers above are the forced-promotion case, the worst shape the report describes.
• The UnsupportedModel mapping covers the codes the provider actually sends; a new provider code falls through unchanged by design.

Docs updated in this PR: core-client, core-host, plugin-runtime.

Owner: Dev
Reviewer: Reviewer

Checkpoint

{
  "pr": 1346,
  "claim": "Dev",
  "issue": 1347,
  "summary": "Interactive UI defects from a real click-through, now including the stacking one. The Details panel opens at 380 with labels and buttons that never break mid-word. A shared preview helper strips untrusted fences from sidebar rows. A blank Bot chat names itself. The copy button is named for copying and confirms it. The mode word reads as a status. Provider refusals read as sentences that say what to change. The chat column is now one stacking context below the panel and dialogs, so a message can no longer paint over Tasks, Settings or an Inbox scrim.",
  "files": [
    "core/src/client/styles.js",
    "core/src/client/remote-bot-settings.js",
    "core/src/client/pane-width.js",
    "core/src/client/bot-settings.js",
    "core/src/shared/preview.js",
    "core/src/shared/failure-copy.js",
    "core/src/client/flybot-slots.js",
    "core/src/client/details-drawer.js",
    "core/test/ui-clickthrough.test.js"
  ],
  "checks": { "ran": "npm run verify", "result": "green", "at": "7f8c7daeebe21e7eea8274f99f973629e93fff48" },
  "next": "none — ready for review"
}

  

[Birfy]

Written-by: Dev

Superseded by #1362, which already merged the same #1347 fixes onto main (regenerated patches, non-frozen branch). Leaving this cursor/ PR closed rather than fighting the freeze seal.

[Birfy]

Superseded by #1362 (merged 2026-10-27T06:54:30Z → main 907a1959); this head was left dirty and needs no rebase.