swarm tool fails with 'Not in a swarm. Use a git repository to enable swarm features.' after /clear (member never re-registered)

[sorairo-fullstack]

Summary

After /clear, every swarm tool action that needs membership (spawn, list, start, …) fails with:

Not in a swarm. Use a git repository to enable swarm features.

The message points at git, but the repo is a valid git worktree (git rev-parse --is-inside-work-tree → true) and features.swarm = true. The actual cause is that the session created by /clear is never registered as a SwarmMember.

Version: v0.75.0 (5ae2385), Windows x86_64.

Root cause

In crates/jcode-app-core/src/server/client_session.rs, handle_clear_session:

1. Removes the old session's member entry (intentional, per the comment "Do not migrate the old session's swarm membership"):

   let swarm_id_for_update = {
       let mut members = swarm_members.write().await;
       members.remove(client_session_id).and_then(|member| member.swarm_id)
   };

2. Then calls update_member_status(&new_id, "ready", …) — but update_member_status_with_report_tldr (swarm.rs) only does members.get_mut(session_id) and silently no-ops in the else branch when the member does not exist. It never inserts.
3. The only function that creates a member, ensure_client_swarm_member, is called exclusively from handle_subscribe. /clear never re-runs it, and the connection stays subscribed, so no new subscribe happens.
4. The clear_done log then reports swarm_id_updated=true, which is misleading: it only means the old member had a swarm id to remove, not that the new session was registered.

Result: the new session has no SwarmMember entry, so every swarm_id_for_session(id) → members.get(id).and_then(|m| m.swarm_id) returns None, and all comm handlers answer "Not in a swarm."

Evidence (from ~/.jcode/logs/jcode-2026-08-10.log)

17:43:37 EVENT event=SESSION_LIFECYCLE … phase=clear_done old_session_id=session_guppy_… new_session_id=session_skunk_… swarm_id_updated=true
17:45:00 EVENT event=TOOL_LIFECYCLE … error="Not in a swarm." (swarm spawn)
17:45:08 EVENT event=TOOL_LIFECYCLE … error="Not in a swarm. Use a git repository to enable swarm features." (swarm list)

Every phase=swarm_member_registered event of the day belongs to sessions that went through handle_subscribe; none names the /clear-born session. Its state file ~/.jcode/state/swarm/session_session_skunk_….json does not exist.

Checked 5ae238574..origin/master on client_session.rs: no later commit fixes this. #816's fix (reset swarm plan state on clear, 5496ebb) is an ancestor of v0.75.0 and doesn't cover member re-registration.

Repro

1. Open a jcode TUI session in a git repo, use any swarm action (works).
2. /clear.
3. swarm list → "Not in a swarm. Use a git repository to enable swarm features."

Workaround: quitting the client and jcode resume on the same session re-registers the member via handle_subscribe.

Suggested fix

In handle_clear_session, after swapping in the fresh agent and before update_member_status, re-register the replacement session, e.g. call ensure_client_swarm_member(&new_id, client_connection_id, …, swarm_enabled, …) (the old member's swarm_enabled can be captured before the remove). Alternatively, make the update_member_status no-op branch log a warning so the next silent drop of this kind is visible.

Also worth a tweak: the "Use a git repository to enable swarm features" wording sends users debugging git/config when membership is the actual gate.

[sorairo-fullstack]

Additional downstream symptom, same root cause: the swarm panel hotkey (Alt+N by default) permanently reports "Swarm view closed" after /clear.

inline_swarm_gallery_active() (crates/jcode-tui/src/tui/app/tui_state.rs) requires !inline_swarm_members().is_empty(), and inline_swarm_members() scopes to the session's spawn subtree. Since every swarm spawn fails with "Not in a swarm." in the /clear-born session, the member list can never become non-empty, so cycle_swarm_panel_view() always falls into the Chat branch and the panel can never open.

[1jehuang]

Thanks for the detailed root-cause analysis. I reproduced the lifecycle gap and opened #876 with a fix that registers the /clear replacement as a fresh swarm root while still removing the old swarm and plan participation. The focused clear-session regression test and affected-crate checks pass.

--- — Jcode agent (automated triage), on behalf of @1jehuang