[Bug] Undeletable "(no messages)" rows: the saved-session catalog skips the draft lifecycle filter

[artgas1]

Summary

The agents view lists rows titled (no messages) that cannot be removed. They are not stale files and not orphaned leases: the daemon is holding a real, healthy, message-less session, while the saved-session catalog draws a second row for the same file that no lifecycle filter ever sees.

activeLifecycleForSession classifies a message-less resident session as draft, and AgentsViewMode.reconcileCatalogs drops it with shouldShowAgentsViewSession. The same method then passes this.savedSessions into reconcileUnifiedSessions unfiltered, and a saved-only record is never classified, so the file returns as an inactive row with no activeSessionId. Ctrl+X cannot resolve it, because the daemon still owns the file, and the draft has no visible row to stop.

Steps to reproduce

Three daemon commands from a clean state, no TUI needed (0.8.1, isolated session dir):

await client.request({ type: "create", config: { cwd, sessionDir } }); // a session nobody prompts
await client.request({ type: "detach", activeSessionId });             // the client leaves
await client.request({ type: "list" });
await client.request({ type: "list_saved_sessions", cwd, sessionDir, scope: "all" });
await client.request({ type: "delete_saved_session", sessionPath });

Observed:

list:   {"id":"15fe65b6f984","activeSessionId":"15fe65b6f984","lifecycle":"draft","activity":"working",
         "messageCount":0,"attachedClients":0,"file":"/tmp/ghost-repro/sessions/01a0583f-....jsonl"}
saved:  {"id":"01a0583f-....","path":"/tmp/ghost-repro/sessions/01a0583f-....jsonl",
         "messageCount":0,"first":"(no messages)"}
delete: "Cannot delete the currently active session"

The file is 696 bytes: session, model_change, thinking_level_change, service_tier_change, session_state. No messages.

Through the UI: start a session, never send a message, leave it. The row appears under Inactive as (no messages), and Ctrl+X twice answers Session became active; stop it before deleting forever. On my machine three of them accumulated, one of them 12 hours old, each still holding a workerState: "ready" worker with a live workerPid and attachedClients: 0.

Expected behavior

Either the row does not exist, which is what the daemon already decides for the resident half of the same session, or the row is deletable. Today it is neither.

Prime Agent version

0.8.1 (npm). Code paths re-read on main at c718bf3; the mechanism is unchanged there.

Environment

macOS 15 (arm64), background daemon plus agents view.

Why the file is there at all

hasUserContent() counts model_change, thinking_level_change, and service_tier_change as user intent through CONTENT_ENTRY_TYPES, and every session writes those three at startup. So isEmptyDraftContent is false for every abandoned session, the detach-time discard never fires, the file stays, and the worker stays resident with it. One ghost per abandoned session.

The row is not the only symptom

prime-agent --continue picks a session with findMostRecentSessionForCwd, which sorts *.jsonl by mtime and filters on the header cwd only. An abandoned draft is the newest file in that directory, so it wins. Checked against the real function with two fixture files, a conversation written at 10:00 and a draft at 11:00:

prime-agent --continue would open: /tmp/continue-check/sessions/ghost.jsonl
is the ghost: true

So the ghost is not only an undeletable row: it also captures -c until the user opens something else in that directory. The view patch below does not change this, and it should not - the honest fix is either to stop persisting an abandoned draft or to make -c skip message-less files. The second one is not free: findMostRecentSessionForCwd deliberately reads one header line per file and would need message counts, so it is a decision for whoever owns that path.

Patch

Branch: artgas1/prime-agent@main...artgas1:prime-agent:fix/agents-view-hidden-draft-rows (bf2085d). PR #1919 was auto-closed by the contribution gate, as CONTRIBUTING describes; the branch remains available for review or cherry-pick. The same change as a mailbox patch: append .patch to that compare URL. Checked by applying it with git am to a fresh worktree of c718bf3, where agents-view-state, agents-view-mode, and the #502 regression file pass (105 tests), and biome and tsgo are clean.

Catalog rows without a daemon summary get the same live-only rule as daemon rows, reusing the existing classifier so on-disk and resident answers cannot drift apart:

export function shouldShowAgentsViewSavedSession(saved: AgentConnectionSavedSessionInfo): boolean {
	return inactiveLifecycleForSession(saved) === "live";
}

applied to the saved-only branch of reconcileUnifiedSessions; daemon records are still enriched by their saved twin. This is the same direction as #1852: one answer to "does this session have messages", not two.

Validation, since CI does not run for unvouched authors: test/agents-view-state.test.ts 68 passed, two new tests, both fail without the guard (expected [ Array(4) ] to deeply equal [ 'live' ]). Also green: agents-view-mode, agents-view-inactive-reply, agents-view-missing-cwd, session-view-search, daemon-session-list, session-manager/session-state, test/suite/regressions/502-unified-session-view.test.ts. npm run check clean.

Blast radius

The saved catalog has one consumer in the TUI, listDaemonSavedSessions at agents-view-mode.ts:2202, so the guard changes one surface. Explicit resume is unaffected: --resume <id|path> goes through resolveSessionPath, which reads the directory itself, so a hidden draft can still be opened on purpose. And the project already treats a message-less session as unresumable one layer up, in formatResumeHint: "Omit ephemeral and unflushed empty sessions because neither can be resumed."

What the patch does not do, and one open choice

• The draft file and its resident worker still survive, now without a row. A sweep for message-less drafts with no attached client is a separate change, and it is the half of fix(coding-agent): delete empty draft sessions on shutdown and sweep ghosts at startup #1079 that [Bug] Ghost sessions: empty '(no messages)' rows undeletable from agents view, orphaned leases never reclaimed, partial fix #1079 has a regression #1531 already isolated as safe.
• The opposite fix is possible: keep the row and let delete_saved_session discard an idle, message-less draft instead of refusing. I did not take it because activeLifecycleForSession states that a conversation-less session must not surface a row, so the visible row looks like the defect. If you prefer the delete-side behavior, say so and I will prepare that shape instead.
• The guard covers drafts only. An earlier shape of it also hid archived and crash files, which replaying the view over a live catalog proved wrong: see the comment below. Hiding archived rows belongs to Complete queued, interrupted, and archived session lifecycle recovery #1382, not to a view filter.

Related

• [Bug] Ghost sessions: empty '(no messages)' rows undeletable from agents view, orphaned leases never reclaimed, partial fix #1079 has a regression #1531 reports the same symptom on 0.7.2/0.7.3 and attributes it to orphaned leases. On 0.8.1 the lease theory does not hold for my case: the sessions were held by live, ready workers, and the repro above needs no dead process at all.
• Complete queued, interrupted, and archived session lifecycle recovery #1382 covers archived session revival.
• refactor(coding-agent): read messageCount instead of a shadow flag #1852 removed the shadow flag in favor of messageCount; this is the same idea one layer over.

[artgas1]

Second pass: I replayed the real view pipeline over a live catalog instead of trusting unit tests, and it caught a defect in my own patch.

Method

Snapshot of a running 0.8.1 daemon: 39 session summaries and 15 saved files, with one ghost created on purpose (create then detach, no prompt) and removed again afterwards. The snapshot is then fed through exactly what reconcileCatalogs does - shouldShowAgentsViewSession, reconcileUnifiedSessions, buildAgentsViewRows - once from main and once from the branch, and the rendered rows are diffed.

What it caught

The first shape of the guard (inactiveLifecycleForSession(saved) === "live") removed three rows, and only one of them was the ghost:

- (no messages)                       0 messages   <- the ghost, correct
- "Ты продолжаешь ветку портации..."  3587 messages
- "Сигнал: DEV-1729, живой трафик"     214 messages

Both extra rows are real sessions carrying an archived marker on disk. One of them has active -> archived -> active -> archived in its session_state history, so the marker is not a hide flag at all: the session was deactivated and later revived, which is the case #1382 is about. Hiding those rows would have removed the only way back to them.

Correction

The guard now excludes drafts only:

export function shouldShowAgentsViewSavedSession(saved: AgentConnectionSavedSessionInfo): boolean {
	return inactiveLifecycleForSession(saved) !== "draft";
}

Re-run over the same live snapshot: exactly one row disappears, the ghost. No row is added, and the order of the remaining 18 rows is identical. Tests now pin archived and crash rows as visible so this cannot regress; 68 pass in agents-view-state, 174 across the eight agents-view, daemon-session-list, session-state, and #502 regression files, npm run check clean.

Branch updated: bf2085d. artgas1/prime-agent@main...artgas1:prime-agent:fix/agents-view-hidden-draft-rows

The wider point for whoever picks this up: today the on-disk archived marker does not keep a session out of the agents view, and at least in my catalog it coexists with active revivals, so any future "hide archived" behavior should be decided in #1382 rather than in a catalog filter.

[artgas1]

#1920 landed while this was open (created 14:34Z, three minutes before this report, so independent work rather than a reply). It changes the severity here, and it is worth being precise about which half it covers.

What it fixes for this bug, reading the merged diff rather than running it: an empty, unnamed, idle session's worker is now passivated when its last client disconnects, so the daemon stops owning the file. delete_saved_session therefore stops answering Cannot delete the currently active session, and the row in this report becomes deletable. The same PR also stops pinning an empty session at "working".

What it does not change: the row still exists. reconcileCatalogs still passes the saved-session catalog through unclassified, so a message-less draft file is still drawn as an inactive (no messages) row, while the daemon's own rule says a conversation-less session must not surface one. After #1920 that is an inconsistency rather than a dead end - the user can now delete the row, they just should not have been shown it.

Branch rebased onto 9f5edc1: 5e8f60bd0. agents-view-state and daemon-session-list pass together (95 tests, including the new empty-session cases from #1920), npm run check clean.

Two smaller notes that survive #1920: the draft file itself still stays on disk after eviction (passivation keeps it), and prime-agent --continue still picks it, since findMostRecentSessionForCwd sorts by mtime with no message-count filter.

[artgas1]

Rebased after the agents-view roster/direct-transport stack landed, and replayed the saved-only contract on its merged base (0749e0667) instead of assuming the overlapping refactor covered it. The branch is now rebased through latest main (3da8c5a1c); the only intervening commit is #1954 and it does not touch the four source/test files in this patch.

The row is still present. On a pristine 0749e0667 worktree, reconciling [draft, archived, used] saved sessions with no daemon summaries returns all three; on the fixed branch it returns [archived, used]. The live catalog is filtered first, but the saved-only branch of reconcileUnifiedSessions is still accepted without lifecycle classification.

Branch updated to dab1797fb: artgas1/prime-agent@main...artgas1:prime-agent:fix/agents-view-hidden-draft-rows

The roster-stack rebase had one conflict only: keep the new rosterAgentIdForSummary import and add inactiveLifecycleForSession. The later move from 0749e0667 to 3da8c5a1c was clean. I also corrected the patch prose for #1920: empty-session eviction now makes the stale row deletable, but it keeps the saved draft file and does not fix the catalog inconsistency.

Validation:

• 137 related agents-view / daemon regression tests passed on the roster-refactor base; agents-view-state passes 69/69 again on 3da8c5a1c.
• npm run check passes on 3da8c5a1c (953 files, no fixes; type/static/installer/browser checks clean).
• Sabotage on the roster-refactor base: removing only the saved-draft guard makes both new regressions fail and renders the draft as an inactive row; restoring it makes both pass.
• A regenerated mailbox patch applies cleanly to a detached pristine 3da8c5a1c worktree with git apply --check.

No maintainer action is needed to resolve rebase drift; the branch is current and remains available for cherry-pick.