(sidebar): exclude subagents from project-level archive-all - #149

Problem

buildProjectsFromCache (session-cache.js) groups every session_cache row by projectPath — subagent rows included. They are indexed from <folder>/<parent>/subagents/agent-*.jsonl and carry the same folder and projectPath as their parent, with parentSessionId set. So the project.sessions array the renderer receives is a flat list of parents and children, and every consumer has to drop the children itself. processProjectSessions does (allSessions.filter(s => !s.parentSessionId)); the DB side does too (getTotalCounts, WHERE parentSessionId IS NULL).

The project header's Archive all sessions button did not:

• the confirmation prompt counted subagents — "Archive all 87 sessions…" on a project with a handful of real ones;
• every subagent transcript was archived, so it vanished from under its parent;
• stopSession was called on any subagent id that happened to be in activePtyIds.

The status bar's N sessions total (renderDefaultStatus) had the same defect, and therefore disagreed with the stats panel's own total for the same data.

Fix

Filter on parentSessionId in both places. Two one-line changes; no behavioural change for top-level sessions.

Tests

test/dom-project-archive-all.test.js — 2 jsdom tests. Verified failing before the fix:

actual: 'Archive all 4 sessions in dev/myproj?'  expected: /Archive all 1 session in /
actual: [ 's-top-1', 's-sub-1', 's-sub-2', 's-sub-orphan' ]  expected: [ 's-top-1' ]

The second test puts all four ids in activePtyIds, so it also pins the stopSession leak.

Full suite: 673 tests, 666 pass, 0 fail, 7 pre-existing skips. ESLint: 0 errors.

renderDefaultStatus is not covered — there is no jsdom harness for app.js. That one line is verified by reading only.

Out of scope

The slug-group archive button has the same class of bug and is handled in #134.

One related defect is documented but not fixed here: when a search matches only subagent transcripts in a project, refreshSidebar narrows sessions to those matches, processProjectSessions filters them all out, and the project drops out of the results entirely. Surfacing it needs a UI decision, not a filter. Written up in .ai/contexts/subagent-observability.md.

Knock-on fixed in the same PR: the project vanished instead

Review caught a regression the two-line filter introduced. Once every top-level session is archived, the surviving subagent rows keep project.sessions non-empty while filtered (top-level only) is empty — so the guard at sidebar.js:442 returned null and the render loop continued before reaching buildSessionsList and its orphan-subagent bucket. Header and orphans both gone: the whole project disappeared from the default view although its data was still there.

Not data loss — "Show Archived" brings it back. But the old behaviour hid this by accident: archiving the children too left project.sessions genuinely empty, so the disappearance was correct. The filter alone turned a correct disappearance into an incorrect one, and systematically, from the first surviving subagent.

The guard now keeps a project that still has indexed subagents and no active filter:

const keepForOrphanSubagents = !anyFilterActive && subagentIndex.size > 0;

subagentIndex is already built two lines above and is the exact object buildSessionsList consults for its orphan bucket — no second mechanism. When filtered is empty, allTopLevelIds is empty by construction, so every indexed subagent is an orphan: size > 0 is the precise "there is something to show", not an approximation. !anyFilterActive carries weight — without it, a starred/running/today filter or an active search would resurface every project owning a subagent, since a subagent satisfies none of those filters.

Both halves are mutation-proven:

without !keepForOrphanSubagents → ✖ project header must survive when unarchived subagents remain
                                    actual: null, expected: true
without !anyFilterActive        → ✖ an active filter with no matching top-level session must still hide the project
                                    actual: {}, expected: null

Guard cases verified unbroken: empty project, project whose top-level sessions are filtered out, empty project under an active filter, _projectMatchedOnly, project with surviving top-level sessions.

Full suite after rework: 685 tests, 678 pass, 0 fail, 7 pre-existing skips. ESLint: 0 errors.