fix(tui): preserve exact background notification owners

[kevinmanase]

Summary

Background completion and watch notifications now stay bound to the exact live UI session that started them, so output from one chat cannot surface in another chat that shares the same durable conversation identity. The ownership model distinguishes concurrent runtime sessions, preserves the originating stream owner through completion, and keeps terminal/watch attribution scoped to that owner.

Validation

• scripts/run_tests.sh tests/test_tui_gateway_server.py tests/tools/test_notify_on_complete.py tests/tools/test_process_registry.py tests/tools/test_terminal_task_cwd.py tests/tools/test_watch_patterns.py tests/tools/test_zombie_process_cleanup.py tests/tui_gateway/test_delegation_session_lifecycle.py — 549 passed, 0 failed.
• Codex CLI review against main — no findings.
• Greptile CLI could not dispatch because NousResearch/hermes-agent is not connected to Greptile.

Related

Related: #61719

---

[alt-glitch]

This was generated by AI during triage.

Related: #61719 and this branch share the core UI-session ownership patch; #62201 adds exact live-owner/background-stream coverage. The branch is now substantially behind current main, which also includes merged ownership routing in #65419, so it needs a rebase/split to show its remaining delta before a maintainer chooses consolidation. Not a duplicate.

[teknium1]

Thanks for covering the exact live-UI-owner path, including delegated child execution and checkpoint/event propagation.

Current main still has the reported defect: tui_gateway/server.py:8667-8687 has multiple session pollers dequeueing a process-global queue, tools/process_registry.py:1189-1202 applies the positive ownership callback only to async-delegation events, and tui_gateway/server.py:8834-8842 routes terminal streams to the first matching durable session key. The proposed ownership propagation and fail-closed routing directly address those paths, with regression coverage for completion, watch, stream, checkpoint, and executor-boundary cases.

This overlaps open PR #61719, as noted in the existing maintainer comment, so the two implementations should be compared and one selected rather than merged together.

Automated hermes-sweeper review.

[kevinmanase]

This overlaps open PR #61719, as noted in the existing maintainer comment, so the two implementations should be compared and one selected rather than merged together.

Thanks. I compared the histories directly. e3f398f90 in #62201 is patch-equivalent to #61719's 54dbd8bd5, so these are not competing implementations. #62201 then adds 96129c9c7 and ecad36e27, which model a unique live owner and preserve exact background stream ownership. Those follow-ups address #61719's open review gap for plain background live output.

I think #62201 is the branch to keep and #61719 can be treated as superseded.

[kevinmanase]

@teknium1 rebased on current main and resolved. The branch was ~7,500 commits behind, so this needed a rework rather than a straight rebase.

On the #61719 overlap you raised. Checking against current main, #65419 already landed the ownership routing, and main also solved the delegated-child ContextVar problem a different (better) way, with contextvars.copy_context() instead of the bind_ui_session_id helper I had. So I dropped that whole commit rather than re-litigate it, and kept only what main is still missing. That also settles the #61719 comparison: the part the two branches shared is upstream now, so there is nothing left to choose between.

The defect is still live on main. Two concrete gaps:

• ProcessSession has no origin_ui_session_id field at all, so a background process never records which UI tab commissioned it. tools/terminal_tool.py never sets one either.
• _owner_sid_for_process in tui_gateway/server.py returns the first session whose session_key matches. session_key is the durable conversation key and several live tabs can share one, so live agent.terminal.output goes to whichever tab comes first in dict iteration order.

Net effect: a user can see output from a command they never ran, while the tab that actually ran it shows nothing.

The fix. Track the commissioning tab on the process itself, carry it on completion and watch_match notifications, persist it through the checkpoint so it survives a gateway restart, and prefer it when routing. Two deliberate fallbacks so nothing regresses:

• A process with no recorded origin (older checkpoints, CLI sessions) keeps the previous key matching.
• If the origin tab is gone, routing falls through to a live tab on the same conversation. The key-match loop now skips finalized tabs, which would otherwise swallow the output.

Verification.

• tests/test_tui_gateway_server.py 573 passed, tests/tui_gateway/ 455 passed, and the process_registry / notify_on_complete / watch_patterns / zombie_cleanup suites pass.
• The 9 new tests all fail without the fix, so they cover the bug rather than passing vacuously.
• Ruff clean.

One thing to note if you run the suites in bulk: tests/tools/ tests/tui_gateway/ in a single pytest process has a large number of order-dependent failures on main already. I measured it both ways, 113 failed on clean main and 114 on this branch out of ~6,900. Run per-directory they are green. The residual tests/tools/ failures are confined to test_voice_mode.py, test_wake_word.py and test_web_tools_config.py, and I confirmed those fail identically with my changes reverted (7 failed / 96 passed either way). They look like missing optional deps in my environment, not anything from this PR.

Ready for review.

[alt-glitch]

This was generated by AI during triage.

Related: #61719 and #73351 address the same cross-session background-output family through overlapping but distinct ownership paths. This PR uses the exact originating live UI session; a maintainer should choose the preferred implementation.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

fix(tui): preserve exact background notification owners

1. Fallback destination is iteration-order dependent — tui_gateway/server.py _owner_sid_for_process: when the origin tab is finalized/gone and several live tabs share the session_key, the fallback picks whichever live tab comes first in _sessions iteration order. The tests pin two-tab cases, but with 3+ live tabs on one conversation the destination is effectively arbitrary. Recording the most-recently-active tab (a last-activity timestamp on the session dict) would make the fallback deterministic.
2. Broad except on the session-context read — tools/terminal_tool.py: get_session_env("HERMES_UI_SESSION_ID", "") is wrapped in a bare except Exception: pass. A stale or recycled HERMES_UI_SESSION_ID (e.g. inherited through delegation from a tab that has since closed) silently routes output to a different live tab via the fallback. Consider validating the sid against known live sessions before pinning it, or at least logging when the inherited value does not resolve.
3. Forward compatibility of the new event key — tools/process_registry.py completion/watch events now carry origin_ui_session_id; verify all consumers of the completion queue (gateway watchers, CLI, dashboard) tolerate the extra key rather than strict-schema parsing it away.