[liuedcson]

Background

Two follow-ups to the process-registry work in #349.

A bug in that change. session_shutdown fires for five reasons and only one of them ends the
session:

reason: "quit" | "reload" | "new" | "resume" | "fork"

new, resume and fork replace the runtime while the old conversation stays on disk — and, in a
multi-session host, on screen — and reload keeps the same session object. The cleanup released on
all five, so opening a new conversation, forking one, resuming another or reloading settings each
stopped that session's background commands: the dev server the user was watching died. new also
reaches the handler from a session that is still open elsewhere.

process-registry.ts promises the opposite — bash "must survive a session switch and a fork" —
and the comment beside the call claimed a switch or fork never reaches this handler. That comment is
what made the bug look correct, so it is replaced with the reason list.

Nothing to read them with. A session can leave background commands and background subagent lanes
running, and the transcript never shows either. The lane state had a permanent widget; commands had
no surface at all.

实现

1. Release on quit only (features/step.ts)

if (event.reason === "quit") releaseSessionProcesses(ctx);

quit covers a host that is exiting and a session that is deleted (WebSession.delete →
close() → dispose()). A host killed outright is covered by the process.exit sweep.

2. /ps (features/step-ps.ts)

Lists the session's background commands with pid, state, elapsed time and log path; stops one
(/ps kill <pid>) or all (/ps kill all). Folds after 16 rows, like the terminal ps output it is
modelled on.

• kill all narrows to command pids. releaseOwnerGracefully is owner-wide and kind-agnostic by
  design, so passing the whole owner would also stop the session's subagent trees — not what the
  command promises.
• It reports before it awaits: releasing a tree takes the SIGTERM grace window, and three silent
  seconds read as ignored.
• A repeat kill all during that window counts the commands already stopping instead of answering
  "Stopped 0".
• notify is captured before the await, because a fork during the wait invalidates ctx and a
  stale context throws on read.

3. /subagents (features/step-subagents-command.ts)

Lists each running lane with its tasks and the step each child is on — the active tool, or the
thinking/reply text while it reasons, so a child that is reasoning does not read as idle.

Reads SUBAGENT_LANE_CHANNEL rather than the lane map: it observes lanes without extending their
lifetime. Settled lanes are evicted from its map (it never shrank otherwise). skipped is not
counted as failed — it is a chain step that never ran, a normal ending. Elapsed time stops at
updatedAt, so a settled task no longer reads as still running. A lane whose roster is not yet
populated says starting instead of printing a bare heading.

4. Both are terminal commands

Web and Desktop draw the same state in their own interface, and the notify a transcript command
depends on is a no-op in an RPC host — a command there would look like nothing happened.
registerCommand runs once per process and cannot see the host's mode, so the check is at
invocation.

5. Session status line (features/step.ts)

publishBackgroundCommandStatus reports N background commands in the footer, subscribed per
session and released on shutdown. The transcript command answers "which one"; the status line
answers "is anything running" without being asked.

6. No lane widget

Removed the permanent aboveEditor widget for background lanes. A panel that sits above the editor
for as long as a lane runs competes with the conversation, and its rows cannot show which step a
child is on. publishLaneSnapshot is called from the lifecycle wrapper before the host hook, so
snapshot publishing is unchanged — verified by deleting the call, which fails five tests in
step-subagent-events.test.ts. The blocking subagent call keeps its inline list: that is the
turn's own output and ends with it.

Progress notifications to the model stay out of the transcript, now stated as the deliberate rule it
is rather than justified by the deleted widget. Completion and needs-input still display.

验证

New tests (19):

• test/step-ps.test.ts (8) — owner scoping, folding, kill <pid> / kill all correctness, refusal
  in a non-terminal host, usage errors.
• test/step-subagents.test.ts (7) — lane/task/activity rendering, thinking fallback, settled lanes
  drop out, folding, refusal in a non-terminal host.
• test/session-process-release.test.ts (4) — session_shutdown releases on quit and on nothing
  else. Reverting the guard makes this test fail, and that was checked.

Also: test/subagent-background-tui.test.ts now asserts no widget is ever set; re-adding one makes
it fail, which was checked. The assertion waits for both lanes to settle and then drains a repaint
interval, rather than sampling for a fixed 50 ms — a late publish escaped the short sample.

pnpm run check passes, including all 16 repository guards.

边界

• Real Windows hosts are not verified. The parent-death supervisor's taskkill path is exercised by
  inference only.
• /ps and /subagents do not page; they list up to 16 rows and fold the rest.