Skip to content

agent.spawn / spawn_batch with no --pane or --workspace still target the focused workspace, not the caller's (follow-up to #242) #246

Closed

Description

@mkh63d

Follow-up to #242 (fixed in 2.13.0). That fix made an explicit --pane authoritative, but the case with no --pane and no --workspace still resolves to the focused workspace instead of the caller's. Seen on 2.13.

Symptom

An orchestrator is started in workspace A. While it runs, the user switches to workspace B. Some of the time the agents show up in B (where the user is looking) instead of A (where the orchestrator lives).

What I checked (static reading of v2.13.1, not yet reproduced live)

  • The CLI attaches caller: $WMUX_SURFACE_ID to every sendV2 call (src/cli/wmux.ts, sendV2).
  • layout.grid, pane.list, pane.split, surface.create … are callerScoped in v2-bridge.ts, so they follow the caller's workspace (0.42.0: V2 commands resolve to the wrong window once a second window exists, and unknown flags silently execute with defaults #143).
  • agent.spawn / agent.spawn_batch are not in that table. They go through resolveSpawnTarget (src/main/agent-spawn-target.ts), whose params type is { paneId?, workspaceId? } — caller is never read.
  • With neither given, resolveSpawnWorkspace falls back to lookups.activeWorkspaceId(), i.e. the first window's active workspace, and paneLoads(workspaceId) then picks a pane in it.

So a bare wmux agent spawn --cmd … (or spawn-batch without --workspace) issued from a pane in a background workspace, or in a second window, files the agent under whatever workspace is focused.

Why the bundled plugin can still hit it

wmux-orchestrator's spawn-agents.sh passes --pane from layout grid, so after #242 that path is pane-authoritative. But:

  1. Any orchestration recipe that spawns without --pane (hand-written wmux agent spawn, spawn-batch) is unprotected.
  2. layout grid is only as good as the caller resolution. If $WMUX_SURFACE_ID no longer resolves (stale id after a session restore mints new pane ids, or claude --resume in a different pane), resolveCallerTarget falls back to the focused window/workspace and the grid itself is created where the user is. I have not verified that this happens in practice — it is a suspect, not a finding.

Suggested fix

Give resolveSpawnWorkspace a middle step: explicit --workspace → the caller's workspace (via resolveCallerTarget / __wmux_locateSurface, looking at every window) → active workspace. Same for the pane-load lookup in spawn_batch, which today reads the first window only. A caller that does not resolve should probably fall through to the current behaviour rather than error, matching resolveCallerTarget.

Test: two workspaces, caller surface in the non-active one, agent.spawn with no pane/workspace → agent lands in the caller's workspace.

Workaround

Pass both explicitly: --workspace <id from wmux current-workspace> and --pane <id>.

Activity

  1. mkh63d commented on Sep 19, 2026

    @mkh63d
    ContributorAuthor

    Reproduced live on 2.13.0, plus evidence from a real orchestration run. The static reading in the report holds.

    Controlled repro

    Caller pane in a background workspace, a different workspace focused:

    caller workspace (wmux current-workspace) : ws-c2ca9a68  "Personal-1"
    active workspace (isActive)               : ws-86a07e37  "Work - Product Owner-1"
    
    $ wmux agent spawn --cmd hostname --label t
    → agent record: workspaceId = ws-86a07e37, paneId = pane-59b706e9
    

    The agent is filed under, and its pane taken from, the focused workspace, not the caller's. wmux current-workspace answers Personal-1 from the same shell in the same breath, so caller resolution is live and working — agent.spawn just never asks.

    Control: the same spawn with an explicit --workspace ws-c2ca9a68 lands correctly in Personal-1 (pane-596c6570). So the flag is honoured; only the default is wrong.

    It happens in practice

    agent list from a six-agent spec run earlier the same day, all spawned from one session whose pane lives in Personal-1, all with & $wmux agent spawn --cmd $cmd --label $labels[$n] — no --pane, no --workspace:

    label spawned workspaceId
    spec-01-style-library 10:37:13 ws-c2ca9a68 Personal-1
    spec-02-visual-dna 10:38:09 ws-e6d34f5b Work - Code
    spec-03-swipe-discovery 10:38:09 ws-e6d34f5b Work - Code
    spec-04-moodboard-export 10:38:10 ws-e6d34f5b Work - Code
    spec-05-art-direction 10:38:10 ws-e6d34f5b Work - Code
    spec-06-project-persistence 10:38:11 ws-e6d34f5b Work - Code
    spec-01-…-retry 11:08:56 ws-c2ca9a68 Personal-1

    Agent 1 went to the caller's workspace, agents 2-6 went wherever the user had clicked in the intervening 56 seconds, and the retry half an hour later went back. 02/04/06 → one pane, 03/05 → another: distributeAgents round-robin over the active workspace's panes, which is the no---pane signature.

    Correction to my report

    The bundled wmux-orchestrator's spawn-agents.sh does pass --pane (from wmux layout grid, which is caller-scoped), so on 2.13.0 that path is already correct thanks to #242. The breakage is in hand-rolled spawn loops, which is what an agent writes when it drives the CLI directly. That makes the wrong default a footgun for exactly the caller this API is for.

    I also could not reproduce the second suspect in my report (stale WMUX_SURFACE_ID misplacing layout grid) — in every check here caller resolution was correct. Treat that part as unsubstantiated.

    Separate finding

    While testing the --workspace workaround I found it silently fails whenever the --cmd value contains double quotes, which also corrupts the command itself. Different root cause (PowerShell 5.1 native argument passing + the CLI's pairwise flag parser), filed separately as a follow-up to #154.

  2. amirlehmam commented on Sep 29, 2026

    @amirlehmam
    Owner

    Fixed in v2.14.0. With neither --pane nor --workspace, agent spawn / spawn-batch now resolve the caller's workspace ($WMUX_SURFACE_ID → __wmux_locateSurface, asked of every window) before falling back to the focused one; an unresolvable caller falls through to the old behaviour silently. Thanks for the live repro and the agent list table — that made the round-robin signature obvious.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions