fix(tui): a Host-started Turn in /side makes the copied parent transcript visible again

[liuxiaocs7]

Summary

A TUI side conversation (/side) is a Session copy that carries the parent's transcript. The driver hides it by recording #hiddenTranscriptThroughTurnId and filtering every transcript it hands the runner through visibleTranscriptMessages(). One path skipped that filter: #reattachStartedTurn built the attached Turn for a Host-started Turn from the raw opened.messages. When a Host-started Turn arrived on the side Session (e.g. reconnecting while a Turn was running there), pi-tui-runner replaced the visible transcript with those unfiltered messages — so the parent conversation the user opened /side to get away from reappeared in full.

Every sibling already applies the filter (switchSession, and the transcript-replacement subscriptions), so this was a gap rather than a deliberate exception. The fix routes this path through the same filter:

messages: visibleTranscriptMessages(opened.messages, this.#hiddenTranscriptThroughTurnId),

Present since #3759, which introduced both sides of this path.

Fixes #3881

Verification

• npm --workspace maka-agent run build then node --test dist/__tests__/runtime-host-session-driver.test.js — all 52 tests pass.
• Added a regression test (hides the copied parent transcript when a Host starts the side successor turn) that drives a Host-started successor Turn on a side Session and asserts the copied parent transcript stays hidden. Confirmed it fails without the one-line fix (the parent transcript leaks back in) and passes with it.
• npx biome lint and npx biome format clean on both changed files; tsc -p packages/cli/tsconfig.json --noEmit passes.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code authored the one-line fix and the regression test; the affected commit carries a Generated-by: Claude Code trailer.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

Approving.

What this PR does: a TUI /side conversation is a Session copy that carries the parent's transcript, hidden by recording #hiddenTranscriptThroughTurnId and filtering every transcript handed to the runner through visibleTranscriptMessages(). #reattachStartedTurn was the one path that built the attached Turn from raw opened.messages, so a Host-started Turn on the side Session — e.g. reconnecting while a Turn was running there — put the parent conversation back on screen in full. This routes that path through the same filter.

Verified against head:

• Every sibling really does apply the filter, so this was a gap and not a deliberate exception: runtime-host-session-driver.ts:251 (transcript-replacement subscription), :563 (switchSession), :1243 (the other subscription). :1155 is the one that did not, and now does.
• Ordering is right. #hiddenTranscriptThroughTurnId is assigned only in switchSession (:555) and cleared only in startNewSession (:703); #adoptLoadedConfiguration does not touch it, so the value the reattach reads is the one the switch established for that same Session. The generation and channel guards run before the adopt, so a stale reattach bails out before reaching this line.
• visibleTranscriptMessages(messages, undefined) is a no-op, so ordinary (non-side) Sessions are unaffected.

The regression test drives a Host-started successor Turn on a side Session and asserts the copied parent transcript stays hidden, which is the actual reported path rather than a proxy for it.

AI use: Claude Code assisted with tracing the call sites and the field's lifecycle; the verification and conclusion are my own.

简体中文

Approve。

这个 PR 在做什么：TUI 的 /side 会话是带着父会话 transcript 的 Session 副本，靠 #hiddenTranscriptThroughTurnId 加 visibleTranscriptMessages() 过滤来隐藏。#reattachStartedTurn 是唯一漏掉过滤的路径，于是当 Host 在 side 会话上启动 Turn（例如那里正在跑 Turn 时重连），用户本来用 /side 想避开的父会话又整段冒出来。这个 PR 把该路径接回同一个过滤器。

已核对：另外三处调用点（:251 / :563 / :1243）确实都做了过滤，所以这是遗漏而非有意例外；该字段只在 switchSession:555 赋值、startNewSession:703 清空，#adoptLoadedConfiguration 不碰它，因此 reattach 读到的正是同一 Session 在 switch 时确立的值，且 generation/channel 守卫在 adopt 之前就会拦掉过期的 reattach；对非 side 会话该调用是 no-op，无影响。回归测试跑的是真实上报路径，不是替身。