[Bug] Turns whose final assistant message has no text block render no branch or copy actions at all

[gorban]

Summary

There are turns from which a session cannot be forked at all, and the UI gives no hint why:

1. If a turn's final assistant message contains no non-empty text block (for example a length stop that produced only a reasoning block), closing is null, and TurnTailNodeView returns before rendering MessageIconActions — so neither the branch nor the copy action exists in the DOM.
2. If a turn ended with reason.kind === "error", branchUnavailable is true and the branch action renders but is inert.

In the attached session that is turns 20–26 of 26: the last seven turns have no reachable fork point. The remaining turns are not the latest turn, so their actions are revealed only on hover.

Environment

• dsh 0.1.6-alpha.2, commit ddefc45fbc7f8e46dd73185e68295696d1297887, web profile
• session session-2496881f-8d36-4ce6-93d9-1954a31930c7

Reproduction

1. Drive a turn to a length stop that returns a reasoning block only — e.g. the five one-token turns in the attached log (turns 22–26), or any truncated reasoning-only reply.
2. Look at that turn's footer: there is no branch icon and no copy icon. The whole action row is missing.
3. Separately, end a turn with a runtime error (turns 20–21 in the log). The branch icon renders but aria-disabled is set and clicking does nothing.

Current behavior

packages/client/ui-chat/src/client/conversation-nodes/turn-tail.ts:

const closing = finalized.findLast(hasText) ?? null
branchUnavailable: closing === null || latestTranscriptSeq !== closing.finalNode.seq

hasText requires block.kind === 'text' && block.text.trim() !== ''. A reasoning-only assistant message fails it, so closing is null — even though a finalized assistant node does exist (finalized includes it).

packages/client/ui-chat/src/client/chat/TurnTailNodeView.tsx:

if (closing === null) return tail === null ? null : <div …>{tail}</div>

The early return skips the entire <MessageIconActions … onBranch={…} /> block, so the user loses copy and branch on exactly the turns that are most likely to be unusable — a truncated turn is the moment you most want to branch away from it.

For error turns, latestTranscriptSeq is advanced past the closing node by the error event itself:

|| (event.type === 'turn/end' && event.data.reason.kind === 'error')

so branchUnavailable is true and the control is dead.

Expected behavior

• Keep the action row rendered whenever a finalized assistant node exists, independent of whether it produced text; branchUnavailable can stay for the enabled/disabled decision without removing the row.
• On a turn that ended in an error, the fork target should be the last transcript seq, not "nothing" — that turn is precisely where a user wants to retry from.
• Some affordance should explain where forking is possible (the current rule — "last message of a completed turn" — is only discoverable from a hover tooltip).

Relation to existing reports

• client: both session-fork gestures swallow their failure, so a failed fork is indistinguishable from a dead control #4200 already covers the swallowed rejection in the two fork gestures (forkAt/forkSession), and I am not re-reporting it. This report is the separate defect that on some turns the control is never rendered or is unconditionally disabled, so the swallow is never even reached.
• fork 之后，fork 前那条消息仍在排队，但界面上不显示 #6499 / session/fork inherits the next turn's inbox insert → phantom queued input in the child #7118 / Forking a running session replays its pending prompt in the child #7175 concern fork state correctness after a fork succeeds; DSH | dsh-fork-relink | fork 后子 agent 记录自动跟随新会话 #6150 is a plugin. None covers the missing affordance.

Evidence bundle

https://gist.github.com/gorban/4725c87390cb391047a2f14b56b98d28 — EVIDENCE.md section 4; the session log shows the reasoning-only assistant messages for turns 22–26 and the two reason.kind: "error" turn ends for turns 20–21.