[fallleave001]

Closes #1638.

A side chat copies its parent's whole history, so searching it as-is would return the parent's words twice. Only the entries after its boundary message (pi-web:side-boundary) are its own, and this searches exactly those: the inherited history stays the parent row's job, and a side chat without a boundary contributes nothing rather than a duplicate of its parent. The route's relation?.kind !== "side" filter is gone, and the title fallback uses the side chat's own name — its firstMessage is the parent's question.

Picking a hit opens the parent and then the side chat's tab, as you described: a side chat is never selected as the main chat. When another chat holds the panel, the tab is queued in the parent's saved panel state, so the existing side-tab restore applies it as the parent arrives; when the parent is already open, the tab opens directly. The row shows the side chat's own title with "Side chat of {parent}" under it, and an untitled side chat is numbered among its parent's side chats like the Agents switcher numbers it.

The queued tab survives a search that read a catalogue newer than the page's: a side chat missing from sessionCatalog is listed before the parent is selected, because a restore keeps only the sessions the page can see, and the next save would have written the shortened state back over the queue. sideChatHit() counts the hit among its parent's side chats even when the caller's list does not list it yet, so the row in the results and the tab its pick opens carry the same number. Re-picking a side chat replaces its saved tab rather than adding a second one, so a renamed side chat keeps its current label instead of the first one saved.

Three invariants the review tightened:

• A side chat with no boundary also gives no title hit: its name alone is not searchable content.
• A side chat's placeholder name ("Side chat") is never searched — it is app-generated text, and matching it would change what a query means.
• A side chat whose parent is no longer listed is skipped: the sidebar folds it into a parent row that no longer exists, so a hit on it could not be opened.

Lines before the boundary are dropped before the length check and the parse, so the inherited prefix costs reads only: a line that never mentions the boundary type is neither parsed nor able to claim a truncation.

Tests:

• lib/session-search.test.mjs: inherited text matches the parent only; a term split across the boundary matches neither; the boundary boilerplate never matches; a side chat without a boundary contributes nothing, not even its name; an unnamed side chat's placeholder name is not searched; a side chat whose parent is gone is not returned; the title fallback uses its own name. Each of the last four fails when its guard is removed.
• lib/session-search-hit.test.mjs: the parent, title and numbered label of a side hit; null for an ordinary session and a subagent; a parent missing from the list still yields the parent id.
• components/AppShell.search-side-hit.test.mjs: picks the parent and queues the tab; opens the tab when the parent already holds the panel; refreshes the list rather than selecting the side chat when the list lags; a side chat the page's list has not seen yet is listed before its tab is queued; an ordinary session is not handled here.
• components/side-tab-state.test.mjs: picking a side chat again replaces its saved tab instead of adding a second one.

Cost, measured on the 428-session library (13 side chats, 38 MB of inherited text, median of 5 runs, searchSessionContents over the whole catalogue):

query	without side chats	with
t490	1288 ms	1342 ms
session search	2087 ms	2164 ms

Both stay well inside the 3 s budget and return the same 30 rows. Verified on the real library: picking a side chat's hit moves the URL to the parent's ?session= and opens the side chat's tab. Limits, ordering and the 200-character cap are unchanged.

[agegr]

Approved with changes I pushed to this branch (76a662a6, 0bdbcf4f, plus merges of main, no conflicts). Thanks, this does what #1638 asked for, the pick included.

Search. Checked against the live API with sessions built by createSideSession(): a side chat matches on its own text only, inherited text matches the parent only, a query with one inherited and one own term returns nothing, and the boundary text and the placeholder name never match. A user message that merely contains pi-web:side-boundary in the inherited part is not taken for the boundary (I added a test for that). Limits, ordering and the 200-character cap are untouched.

Picking a hit, in headless Chrome:

case	result
another session open	URL moves to the parent's ?session=, the side chat's tab opens selected
parent already open	URL unchanged, the tab opens
tab already open	focused, not duplicated
390px phone	parent opens, drawer closes, the panel shows the side tab, as the switcher's tab button does
same hit twice, then reload	one tab, restored after the reload
row	own title over "Side chat of {parent}"; untitled ones are "Side chat 2" / "Side chat 3", as in the switcher

What I changed:

1. A re-picked hit moved its tab to the end. With tabs Cache sizing, Side chat 3, Side chat 2, picking the Cache sizing hit from another session gave Side chat 3, Side chat 2, Cache sizing, again at every pick. The "replace on re-pick" filter in appendSideTab() caused it and was not needed: the restore already keeps one tab per side chat, and a tab's label is read from the list, not from the saved entry. appendSideTab() is back to main's.
2. Picking the hit of the side chat already shown in the main chat brought a hidden right panel back (on a file tab): the handler opened a tab that the one-session-one-chat effect closed at once. Such a hit is now picked like any other.
3. One helper for both callers. The queued-tab block was a copy of /side's late-landing path; handleSelectSideChat and handleSideSession now share openSideChatTab().
4. Less list-lag handling. I kept the one line that matters (the hit is listed before its tab is saved, since a restore keeps only listed sessions) and dropped the listed branch and both extra list reloads: the running poll closes that window within about 2.5 s, and missing it only means clicking again. A forced-lag run still opens the tab.
5. Docs: docs/agents/side-sessions.md and ADR 0007 still said search skips side chats; AGENTS.md gets file-map lines for lib/session-search.ts and lib/session-search-hit.ts.

On a merge with current main: tsc, lint and the full npm test are clean (3542 tests); node e2e/run.mjs passed (51 checks) before the last merge of main and was not re-run after it.

Two things left as they are, both the same as the switcher's tab button: the tab opens at its end, not scrolled to the matched message (only the main chat takes a search target), and when the main chat shows a sibling side chat the tab opens beside it without moving the main chat to the parent.