feat: shells tracked in kanban

[boozedog]

(Note: please feel free to reject if this isn't consistent with your vision for how the kanban board is supposed to work!)

As mentioned in #176 I was thinking it'd be nice if shells that are started with an associated agent appear in the kanban board under the relevant status column to make it easier to see what they are doing. Useful for doing work on the "main" worktree. (Which maybe I shouldn't be doing, LOL!)

Screenshot of feature:

[marcus]

Code Review: Shells tracked in kanban

Good PR — the architectural direction makes sense and the implementation is clean. The kanbanData struct as a computed snapshot is a solid pattern. A few things to address:

🚨 Must-fix: Accidental result symlink

The diff includes:

diff --git a/result b/result
new file mode 120000
--- /dev/null
+++ b/result
@@ -0,0 +1 @@
+/nix/store/sw4mkz9fjpirdvfk95m0qf7y6rhf320n-sidecar-v0.72.0-dev+dirty

This is a Nix build artifact (nix build output symlink). It must be removed from this PR — it references a path on your local machine and will be broken for everyone else. Add result to .gitignore if it isn't already, and drop it from this branch.

Architecture ✅

The kanbanData struct approach is right: compute column assignments once per render/update cycle rather than calling buildKanbanData() multiple times. However, I see buildKanbanData() is currently called multiple times per operation (e.g., moveKanbanColumn builds it, moveKanbanRow builds it again, syncKanbanToList builds it again). This is fine for correctness but redundant work. Worth a follow-up to thread kd through the call stack where operations are batched, though not a blocker.

Shell-to-column Mapping

The status mapping in kanbanShellStatus:

case AgentStatusError:
    return StatusPaused, true

Routing error shells to StatusPaused is consistent with how worktrees handle StatusError (also mapped to StatusPaused since both need user intervention). Good consistency.

The ordering within columns — worktrees first, then agent-shells — is a reasonable choice, but it means a shell in the Active column will always appear below all worktrees regardless of recency. Is this intentional? If agent-shells should be interleaved by last-activity time, that's a separate feature, but worth calling out as a known limitation.

syncKanbanToList Fix

The old code did p.selectedShellIdx = p.kanbanRow which assumed kanban row directly mapped to shell slice index. The new code correctly searches by TmuxName:

for i, s := range p.shells {
    if s.TmuxName == shell.TmuxName {

This is the right fix, and the test updates (TmuxName: "shell-1") confirm the intent. ✓

syncListToKanban

The new implementation correctly handles agent-shells that have been routed to status columns:

for colIdx, status := range kanbanColumnOrder {
    wtCount := len(kd.worktrees[status])
    for i, s := range kd.shells[status] {
        if s.TmuxName == shell.TmuxName {
            p.kanbanCol = colIdx + 1
            p.kanbanRow = wtCount + i

The wtCount + i offset is correct since worktrees come first in each column. ✓

renderKanbanShellCardLine Improvements

The new status text logic is much more informative — showing claude · waiting, claude · done, claude · error etc. is a clear UX improvement over just shell · running. The orphaned shell case (◌ icon, · offline) is a nice touch.

One nit: the orphaned check runs before the agent-chosen check, which means an orphaned shell with an agent will show ◌ · <agent> · offline but reach the first branch. Trace through:

if shell.IsOrphaned {
    statusIcon = "◌"
} else if shell.ChosenAgent != AgentNone && shell.ChosenAgent != "" {
    if shell.Agent != nil { ... }
} else if shell.Agent != nil {
    statusIcon = "●"
}

For line 1 (status text), the orphaned branch correctly shows · offline using ChosenAgent. The logic is correct, just requires careful reading. A comment on the case-1 branch would help the next reader.

view_list.go

The agent status label update (running → active/waiting/done/error) is consistent with the kanban card changes. Good to keep the list view and kanban in sync. ✓

Minor

The comment removal churn (removing // Highlight selected column header, // Build content for panel, // Apply styling, etc.) is fine — most were redundant. But a few (like // td-693fc7 references) are tracking IDs that provide useful context; keep those.

The renamed variable i → j in the empty-cells loop is correct (avoiding shadowing the outer loop variable). ✓

Summary

1. Remove the result symlink — blocker, must fix before merge
2. Multiple buildKanbanData() calls per operation — minor, follow-up
3. Worktrees-before-shells ordering within columns — intentional? If so, document it

Everything else looks solid. The kanban integration is architecturally clean and the shell routing logic handles the edge cases (orphaned, no agent, agent with various statuses) correctly.

[marcus]

Hey @boozedog! This is a great first PR — you identified the gap, filed the issue, then went ahead and implemented it. That's exactly the kind of contribution we love. 🙌

Review notes:

✅ The struct is a clean abstraction — pre-computing column assignments before rendering is the right move.

✅ kanbanShellStatus() is a nice, focused function with clear routing logic. Mapping AgentStatusError → StatusPaused makes sense (both need user intervention).

✅ The status label improvements in renderShellEntryForSession (showing "waiting"/"done"/"error" instead of always "running") are a welcome bonus.

One thing to address:

There's a result symlink at the repo root that points to a Nix store path — that's a build artifact from nix build and shouldn't be committed. Mind removing that from the branch?

(A result entry in .gitignore would prevent this in future Nix builds — though I see @deephack1982's Nix flake PR is also pending, so there may be a nice opportunity to coordinate there.)

Once that's cleaned up, this is looking solid. ✦