fix(gateway): route plain background live output to its spawn-time UI owner (#61719 residual)

[ayushnangia]

Summary

Live agent.terminal.output chunks from a plain background=true process (no notify_on_complete, no watch_patterns) — a delegated child's, in particular — are dropped by the desktop router: the process has no positive UI owner. This completes the remaining scope of #61719 after its core landed.

Root cause

Owner-routing for post-turn completions landed in merged 54d0948d3 ("route post-turn completions by owner"), and drain_notifications now requires positive ownership for routed events. But the sweeper's flagged problem on #61719 survived the merge:

• tools/terminal_tool.py captured routing metadata only inside the background and (notify_on_complete or watch_patterns) gate — while live output is emitted for every background process (process_registry.on_output).
• _owner_sid_for_process() (tui_gateway/server.py, _wire_agent_terminal_output) matched only by session_key equality against live TUI sessions. A delegated child's ProcessSession.session_key is the subagent's internal key, which never matches a window — so its chunks were emitted with sid="", and a session-less event from a registry reader thread is dropped by write_json (documented at the _live_transports registry).

Changes

• tools/process_registry.py: ProcessSession.origin_ui_session_id — spawn-time UI owner, persisted/restored in the crash checkpoint alongside the existing watcher metadata.
• tools/terminal_tool.py: capture HERMES_UI_SESSION_ID for every background spawn, outside the notify/watch gate — same session-env pattern the codebase already uses for UI routing (tools/desktop_ui.py:39, tools/delegate_tool.py:3191). Empty outside TUI/desktop contexts; best-effort with fallback.
• tui_gateway/server.py: _owner_sid_for_process() prefers the recorded origin while that window is still live, then falls back to the legacy session_key match — so existing routing behavior is byte-identical for processes without a recorded origin or after the owning window closes.

Validation

case	before	after
delegated child, plain background=true live output	sid="" → dropped	routed to spawn-time owner window
plain background, no recorded origin (non-desktop spawn)	session_key match	unchanged
recorded origin but window closed, unknown key	—	falls back to "" (no leak to other windows)

• 2 new tests red on unfixed main, green with fix: test_background_spawn_captures_ui_origin_without_notify (capture without notify/watch) and test_agent_terminal_output_routes_by_spawn_time_ui_owner (child-shape routing + fallback + stale-origin cases)
• scripts/run_tests.sh tests/tools/test_notify_on_complete.py tests/test_tui_gateway_server.py tests/tools/test_process_registry.py — 636 passed, 0 failed

Scope notes

• Completion/watch notifications are untouched — they already carry watcher metadata and flow through the owner-routing from 54d0948d3.
• A checkpoint-recovered process keeps its recorded origin; if that window is gone after restart the sid simply no longer matches a live session and routing falls back — no stale delivery.

Salvages the remaining scope of #61719 by @soria-clawd-bot (sweeper review problem: UI origin captured only when notify/watch enabled). Refs #54785, 54d0948d3.

[alt-glitch]

This was generated by AI during triage.

Related: #62201 has the broader owner-routing implementation across notifications and delegation. This PR isolates the remaining plain-background live-output path.

[ayushnangia]

Correct framing, thanks — mapping the relationship precisely for whoever picks between them:

• fix(tui): preserve exact background notification owners #62201 (last commit Jul 10) is the broad implementation, but it predates merged 54d0948d3, which already landed the completion/notification owner-routing core on main — a large share of its 12-file diff now overlaps what's merged.
• This PR is the residual only, rebuilt against current main: spawn-time UI-owner capture for every background spawn + _owner_sid_for_process preference, 144 added lines, no overlap with the merged routing.

If maintainers prefer salvaging #62201's broader scope instead, the two red-on-main tests here (plain-background capture without notify/watch; delegated-child-shape routing with live-window guard) transfer directly as its regression matrix.

[teknium1]

Thanks for isolating the remaining live-output routing path. The underlying routing gap is present on current main: tui_gateway/server.py:8850-8858 can only map a process by session_key.

Problems

• The proposed assignment is too late for a reliable fix. tools/process_registry.py:798-807 starts the local reader before spawn_local() returns, and that reader calls _emit_output() at tools/process_registry.py:964-974. The PR assigns origin_ui_session_id only after the spawn call returns, so immediate process output can still be emitted without a UI owner.
• The new capture test stubs spawn_local() with a SimpleNamespace, so it cannot cover that startup ordering.

Suggested changes

• Capture the UI sid before spawning and pass it into spawn_local() / spawn_via_env() so ProcessSession has the owner before any reader or poller starts.
• Add an immediate-output regression case using the real registry lifecycle.

Automated hermes-sweeper review.

[ayushnangia]

Both points addressed in 574911b61 (rebased onto current main):

• Startup race: spawn_local() / spawn_via_env() now accept origin_ui_session_id and set it on ProcessSession at construction — before the reader thread (started inside spawn_local()) or the sandbox log poller exists. terminal_tool captures HERMES_UI_SESSION_ID before the spawn call and passes it in; the post-spawn assignment is gone.
• Real-lifecycle coverage: test_immediate_output_carries_ui_owner_through_real_registry spawns echo through the real registry with an on_output recorder that captures the session's owner at emission time — the first chunk must already carry it. The stubbed capture test now asserts the kwarg is handed to the spawn (fails on the post-spawn variant).

scripts/run_tests.sh across notify_on_complete, tui_gateway_server, process_registry, terminal_task_cwd, watch_patterns: 587 passed, 0 failed. One pre-existing exact-kwargs assertion in test_terminal_task_cwd was extended with the new parameter.