Conversation
alt-glitch
added
type/bug
… parent's gateway routing _INHERIT_PARENT_ROUTING_SQL copied session_key/chat_id/chat_type/thread_id/user_id/ display_name/origin_json/transport_profile onto ANY child whose parent row ended on 'compression'. A delegate or branch fork of such a parent is not that conversation's continuation, so it must not take over the route: two live rows holding one routing key lets peer recovery repoint gateway traffic into a subagent's transcript (#116322). Exclude children whose `_delegate_from` / `_branched_from` marker names the queried parent id. Value-match rather than presence: a compression continuation inherits model_config verbatim (marker included), so a presence check would misclassify it — the same idiom _NON_CONTINUATION_CHILD_FILTER_SQL already uses. Cherry-picked from #109699 (Lei-k) minus its unrelated tools/delegate_tool_dispatch.py nested-dispatch hunk; test relocated to tests/hermes_state/ where the suite now lives.
…its its route Pins the reason the exclusion matches the marker's VALUE against the parent id instead of its presence: a continuation copies model_config verbatim, so a `_branched_from` marker rides along. A presence check (#116377's variant) would leave the continuation unroutable — exactly the shape this test fails on against that branch. Salvages #109699 (@Lei-k). Supersedes #116377 (@strzhao). Co-authored-by: strzhao <strzhao@users.noreply.github.com>
|
Closing after revalidation against upstream main 9573f44; no successor PR is needed. The DB routing-inheritance part is now covered by merged #116740 (which closes #116322). For the remaining nested-dispatch part, my earlier direct _dispatch_background regression bypassed the authoritative model-facing depth gate. Testing the actual AIAgent._execute_tool_calls entrypoint and registry fallback gives 6 passing cases: depth 0 selects background=True, while depths 1 and 2 select background=False. The coordinator independently reran all six against that exact main SHA. This validates dispatch selection with delegate_task stubbed, not a live-provider E2E claim. Related lineage reviewed: #96424, #103486, #115616, and #52756; #109744/#109730 also clarify the existing model-facing contract. The private-function test failing alone is therefore not evidence of a reachable model-path defect, and adding another policy gate here would duplicate existing behavior. Thank you for the upstream fixes. This closes the superseded/redundant proposal without claiming that this PR itself was merged. AI-assisted revalidation. |
2 participants
What does this PR do?
Keeps a delegated worker from taking ownership of its coordinator's conversation route in two related cases:
delegate_task(background=true)inside a delegated-child context runs inline, so the requesting worker receives its own results instead of detaching them onto the coordinator's chat/history route. Normal route-owning sessions retain background delegation.Related issue and provenance
Refs #92859 (duplicate live routing keys; this does not claim to fix every cause).
Selectively adapted from Lei-k#22, source commit
0fd2bd67daa9420bf89a14d5393b84114a33e293, merged in the fork ase54e3bae9ac1e33cff1c25bab4d681c532d8be2c. Source authorship is preserved.The fork owner reports a week of successful operation. That is user-reported fork experience, not an upstream end-to-end validation claim.
Scope and deduplication
Based on upstream main
b6b53c69a6ed49cb099cf1bfe76b5e6edd718e5a, not fork main. Only four files change. No schema migration, config, dependency, workflow, deployment or runtime changes.This is deliberately not a wholesale cherry-pick:
switch_sessionrebinding and has separate API history-delivery rules. Open fix(gateway): keep delegate completions on parent session #92620/fix(gateway): never bind a routing key to a delegate subagent session (#92859) #93208 address completion/bind-time ownership; this PR is limited to the earlier insert-time inheritance edge and nested-dispatch seam.The separate gateway connection wording change from fork #16 is not bundled, in accordance with the contributing guide's one-logical-change rule. Its reset recovery/status dedup portions overlap existing upstream submissions and are not reintroduced here.
Verification
Canonical runner, isolated local environment, bounded workers:
SessionDBfor routing; nested dispatch uses real dispatch/aggregation with the model-calling worker stubbed.tests/state/ tests/hermes_state/): 527 passed, 9 skipped, 64 files.git diff --check: passed.tests/test_hermes_state.pyexceeded the per-file timeout. A longer rerun was interrupted; no full-suite pass is claimed. No live provider, gateway transport or production E2E was exercised.5c7bb4897bdb540562a47d5776a2460be72d9a60: PASS, no blocking findings, with real temporary-DB and dispatch-context probes. Two non-blocking notes remain for maintainer review: fallback-note line wrapping, and the branch-marker test being a defensive SQL invariant rather than an observed gateway branch failure. Nested inline work consumes the worker's timeout budget, consistent with merged fix(delegate): nested orchestrators get their workers' results back — delegate_task exempt from the 420 s tool deadline; summary budget uses current prompt, not the session sum #103486. This Draft is for maintainer/user review, not approval to merge.Checklist