huangruiteng · Oct 28, 2026 · Oct 27, 2026
@@ -172,6 +172,35 @@ without treating quoted content as authority. Directory metadata tests do not
 qualify model interpretation, receiver adoption or the original-channel return;
 observe each separately in the live journey.
 
+#### Topic owner and misroute correction / 主题归属与误派纠正
+
+Use the bounded public [routing cases](../../../../examples/evaluations/chat-routing.public.json)
+with the existing [intake evaluator](../../../../examples/evaluations/chat-intake.py).
+Two Agents share a generalist role; the first Agent's id resembles the owner's
+shorthand, but relevant current work and recent delivery identify the other as
+the RFC researcher. A delivered earlier misroute is evidence of where the
+message went, not ownership. Require the topic owner's exact Goal/Agent pair
+and preserve the original artifact, purpose, privacy boundary and return audience.
+Then explicitly name the other pair for a separate verification request: the
+topic-owner inference must not override that instruction. A standalone
+explanation requires no handoff; equally relevant evidence requires one focused
+clarification. Missing delivery permission cannot be repaired by substituting
+the tempting role match. Qualified new work needs no pre-existing matching Todo.
+
+Supply a delivered correct-topic handoff with no read/adoption record. Do not
+resend it. If the old recipient still has uncorrected intent, send only the
+bounded ownership correction through the existing permitted path. Once both
+corrections are delivered, answer from their receipts with zero further effects.
+Neither receipt proves cancellation, receiver adoption or the RFC revision.
+Preserve unknown effects and do not interrupt unrelated work or change authority.
+
+Offline tests check prompt/managed-skill delivery, parser/scorer oracles and
+negative outputs; they do not qualify model reasoning. Release-only live runs
+use `--cases examples/evaluations/chat-routing.public.json` and the existing
+explicit paid-call opt-in below. Review the complete rationale and correction
+brief, including no invented cancellation/adoption. Score actual discovery,
+receiver assessment and original-channel return separately in the live journey.
+
 #### Repair and merge / 修复并合并
 
 GQ03/GQ07 include “修复并合并这个 PR。” / “Fix and merge this PR,” with

@@ -30,6 +30,8 @@ def score(case, response):
         errors.append(f"expected_{case['expected']}_got_{observed}")
     if response.get("proposals") or response.get("protected_action"):
         errors.append("unrequested_action")
+    if response.get("context_handoff") and response.get("goal_draft"):
+        errors.append("competing_intents")
     if case["expected"] in {"draft", "handoff"} and response.get("gate"):
         errors.append("redundant_gate")
     if case.get("target"):

@@ -31,7 +31,7 @@
       "dict_any_count": 0
     },
     "loopx/chat_agent.py": {
-      "lines": 2005
+      "lines": 2007
     },
     "loopx/chat_runtime.py": {
       "any_count": 37,

@@ -380,11 +380,17 @@ The page does not expose host addresses or read host stores; `view=agent_route`
 remains the exact-identity, current-observation path. A registration or binding
 does not prove presence, execution readiness or delivery authority.
 
-Repeated role names alone do not require an owner question. The conversation's
-explicit identity corrections, established assignment, project and declared
-responsibilities guide recipient selection; binding metadata can support that
-context but cannot override an explicit identity or grant. Preserve both Goal
-and Agent identity. Forwarded requests retain their original speaker/addressee
+Repeated role names alone do not require an owner question. Resolve the subject,
+project, purpose and destination before matching a role alias or Agent id.
+Compare plausible recipients by conversation, current work and responsibilities;
+when the user refers to recent research, inspect relevant recent deliveries.
+A previous send proves delivery, not subject ownership: recheck a challenged
+route rather than preserving an earlier mistake. Explicit identity corrections
+take precedence; binding metadata cannot override them or a grant. Preserve both
+Goal and Agent identity. Correcting delivery does not retract an old misroute:
+reuse verified receipts and send only a still-needed bounded ownership correction
+through the existing path. Never claim cancellation or adoption from delivery.
+Forwarded requests retain their original speaker/addressee
 in the existing collaboration brief rather than redefining every “you” as the
 receiver. Clarify only competing interpretations that still change the action.
 

@@ -3,7 +3,7 @@ name: loopx-manager
 description: Inspect authorized LoopX Goals, Todos and deliveries to explain progress, identify owner decisions and delegate intent to the right worker.
 ---
 
-<!-- loopx-managed-manager-skill:v4 -->
+<!-- loopx-managed-manager-skill:v5 -->
 
 # LoopX manager
 
@@ -48,9 +48,15 @@ qualified responsible Agent; keep new constraints and corrections attached to
 that work. A matching Todo or historical owner is not required: use the authorized
 directory's responsibilities and context to choose a qualified recipient, and
 describe that as a new selection rather than proven historical ownership.
-Resolve shorthand from the known conversation and project, stating a material
-assumption. Ask only when competing interpretations would change the action;
-do not require a link or Agent id you or an authorized recipient can resolve.
+Resolve shorthand against the requested subject, project, purpose and destination.
+A role alias or matching Agent id is a discovery lead, not a selected recipient.
+Compare plausible recipients using the permitted directory and relevant current
+work; when the user means recent research, inspect relevant recent deliveries
+as well. Directory order, a generic profile, a delivery target or an available
+host binding cannot establish topic ownership. An explicit Goal/Agent pair or
+owner correction takes precedence. State a material assumption; clarify only
+competing interpretations that still change the action. Do not require a link
+or Agent id you or an authorized recipient can resolve.
 Do not turn every sentence into a fresh assignment. Missing facts
 require a relevant permitted read or bounded verification by a qualified peer,
 not an assumption of completion or a demand for owner intervention. Restricted
@@ -67,6 +73,23 @@ path, finish that response before separate repair and report the repair as
 unfinished. Do not promise an unsupported background continuation. Distinguish
 submission, delivery, receiver decision and actual completion throughout.
 
+A previous handoff proves where a request was sent, not who owns its subject.
+When feedback challenges the recipient or corrects the subject, verify that
+route before preserving an alleged established owner. Do not turn an earlier
+routing mistake into an ownership commitment. Keep the complete original
+request, accepted corrections, purpose, destination and original return audience
+in the existing collaboration brief.
+
+A corrected handoff does not retract a delivered misroute. Check the original
+and correction receipts. Reuse a correction already delivered to the right
+recipient; absent read/adoption records do not justify another send. If the old
+recipient still has uncorrected intent, carry a bounded ownership correction
+through that existing collaboration path, referring to the original request.
+Do not repeat an already delivered ownership correction, claim cancellation,
+erase effects/history or interrupt unrelated work. Corrected delivery, adoption
+and the actual artifact remain separately verified phases. This repair grants
+no new delivery audience, task edit, execution or cancellation authority.
+
 ## Answer shape
 
 Follow the task's depth. A short factual question needs a direct answer and its

@@ -455,9 +455,11 @@ def _agent_item_text(message: dict[str, Any]) -> str:
     "If the user asks to hand off before separate self-repair, submit the handoff first. Reuse an existing delivered request after checking its receipt; never resend merely because reading or adoption is unrecorded. When submission occurs only in the final response envelope, finish that response before separate repair and report the remaining phase honestly. "
     "When actual work remains, reuse qualified existing work and its responsible Agent before creating or delegating another request; preserve new corrections without treating them as duplicate intent. "
     "An exact matching Todo or previously assigned owner is not a prerequisite for requested work. Use the authorized directory's responsibilities and context to select a qualified recipient; distinguish that selection from proof of historical ownership. "
-    "Resolve ordinary shorthand from the known conversation and project context, disclosing a material assumption; ask only when competing interpretations would change the action. Do not ask the user to supply a link or Agent id you can resolve or have an authorized qualified recipient verify. "
+    "Resolve ordinary shorthand against the requested subject, project, purpose and destination. A role alias or matching Agent id is a discovery lead, not a selected recipient. Compare plausible recipients using the permitted directory and relevant current work; when the user means recent research, inspect relevant recent deliveries as well. Disclose a material assumption; ask only when competing interpretations would change the action. Do not ask the user to supply a link or Agent id you can resolve or have an authorized qualified recipient verify. "
+    "A previous handoff proves where a request was sent, not who owns its subject. When feedback challenges the recipient or corrects the subject, verify that route before preserving an alleged established owner. "
     "An explicit Goal/Agent identity or owner correction takes precedence over a similar role name. Preserve the full goal_id/agent_id pair. "
     "When role names repeat, compare the relevant conversation, established assignment, project and declared responsibilities before deciding that the recipient is ambiguous. Registered host bindings can support that context, but a missing binding does not prove an Agent is offline and an available route cannot override an explicit identity or grant. "
+    "A corrected handoff does not retract a delivered misroute. Check the original and correction receipts; reuse delivered corrections, and carry any still-needed bounded ownership correction to the original recipient through the existing collaboration path. Preserve actual effects and history; do not claim cancellation or interrupt unrelated work. "
     "For forwarded or quoted requests, preserve the original speaker and addressee: 'you' does not automatically refer to the receiving Agent. Explain the resolved referent in the existing collaboration brief when it matters; do not invent one. "
     "Carry the original project, purpose, destination and owner corrections in that brief. A qualified recipient's role, available store or execution route cannot replace them; when the intended destination is unresolved, verify the relevant context or ask one focused question instead of substituting another destination. "
     "Delegate only work or verification that remains necessary and needs that recipient's context or execution grant. "

@@ -1008,7 +1008,8 @@ def manager_session_model_allocation(
 # 21: migrate the former transcript-in-objective representation to persisted
 #     native history. Merely removing it from new prompts leaves old upstream
 #     conversations overfull and unable to compact before their next request.
-MANAGER_CONTEXT_VERSION = 21
+# 22: resolve role shorthand by subject evidence; a past handoff is not ownership.
+MANAGER_CONTEXT_VERSION = 22
 
 # An installed manager workspace keeps the marker it was written with. The
 # writer refreshes that workspace skill while the file still carries any
@@ -1018,8 +1019,9 @@ def manager_session_model_allocation(
 MANAGED_SKILL_MARKER_V2 = "<!-- loopx-managed-manager-skill:v2 -->"
 MANAGED_SKILL_MARKER_V3 = "<!-- loopx-managed-manager-skill:v3 -->"
 MANAGED_SKILL_MARKER_V4 = "<!-- loopx-managed-manager-skill:v4 -->"
-MANAGED_SKILL_MARKERS = (MANAGED_SKILL_MARKER_V1, MANAGED_SKILL_MARKER_V2, MANAGED_SKILL_MARKER_V3, MANAGED_SKILL_MARKER_V4)
-MANAGED_SKILL_CURRENT_MARKER = MANAGED_SKILL_MARKER_V4
+MANAGED_SKILL_MARKER_V5 = "<!-- loopx-managed-manager-skill:v5 -->"
+MANAGED_SKILL_MARKERS = (MANAGED_SKILL_MARKER_V1, MANAGED_SKILL_MARKER_V2, MANAGED_SKILL_MARKER_V3, MANAGED_SKILL_MARKER_V4, MANAGED_SKILL_MARKER_V5)
+MANAGED_SKILL_CURRENT_MARKER = MANAGED_SKILL_MARKER_V5
 
 
 def manager_skill_text() -> str:

@@ -272,8 +272,9 @@ def test_review_evidence_retains_brief_without_provider_payload():
 
 @pytest.mark.parametrize("profile", ["restricted", "trusted_owner"])
 @pytest.mark.parametrize("custom_owner_instructions", [False, True])
+@pytest.mark.parametrize("previous_marker", ["v3", "v4"])
 def test_followthrough_refresh_reaches_existing_workspaces_without_replacing_owner_text(
-    tmp_path, profile, custom_owner_instructions
+    tmp_path, profile, custom_owner_instructions, previous_marker
 ):
     from loopx.chat_agent import _turn_prompt
     from loopx.chat_manager import manager_workspace, MANAGED_SKILL_CURRENT_MARKER
@@ -283,7 +284,7 @@ def test_followthrough_refresh_reaches_existing_workspaces_without_replacing_own
     skill = workspace / ".agents/skills/loopx-manager/SKILL.md"
     previous = "Custom owner instructions" if custom_owner_instructions else "# LoopX managed manager instructions\nOlder guidance"
     instructions.write_text(previous)
-    skill.write_text("<!-- loopx-managed-manager-skill:v3 -->\nOlder managed skill")
+    skill.write_text(f"<!-- loopx-managed-manager-skill:{previous_marker} -->\nOlder managed skill")
     manager_workspace(tmp_path, runtime_profile=profile)
     current = instructions.read_text()
     if custom_owner_instructions:
@@ -293,9 +294,79 @@ def test_followthrough_refresh_reaches_existing_workspaces_without_replacing_own
     assert MANAGED_SKILL_CURRENT_MARKER in skill.read_text()
     assert "standalone explanation or opinion request" in skill.read_text()
     assert "finish that response before separate repair" in skill.read_text()
+    assert "A previous handoff proves where a request was sent, not who owns its subject" in skill.read_text()
+    assert "A corrected handoff does not retract a delivered misroute" in skill.read_text()
     turn = _turn_prompt("A follow-up about existing work.", runtime_profile=profile)
     assert "do not end at agreement or design advice" in turn
     assert "never resend merely because reading or adoption is unrecorded" in turn
+    assert "A role alias or matching Agent id is a discovery lead" in turn
+    assert "no new delivery audience" in skill.read_text()
+
+
+def test_manager_refresh_preserves_an_unmanaged_owner_skill(tmp_path):
+    from loopx.chat_manager import manager_workspace
+
+    workspace = manager_workspace(tmp_path)
+    skill = workspace / ".agents/skills/loopx-manager/SKILL.md"
+    skill.write_text("# Owner-managed skill\nKeep this local instruction.")
+    manager_workspace(tmp_path)
+    assert skill.read_text() == "# Owner-managed skill\nKeep this local instruction."
+
+
+ROUTING_SUITE = json.loads((ROOT / "examples/evaluations/chat-routing.public.json").read_text())
+
+
+@pytest.mark.parametrize("case", ROUTING_SUITE["cases"], ids=lambda case: case["id"])
+def test_routing_oracles_reject_alias_misroutes_duplicate_sends_and_added_authority(case):
+    """Frozen decisions exercise the production parser and release scorer, not inference."""
+    response = {"message": "Evidence: " + " ".join(case.get("required_evidence_refs", []))}
+    if case["expected"] == "handoff":
+        response["context_handoff"] = {**case["target"], "brief": {
+            "schema_version": "collaboration_brief_v0", "purpose": "Assess the scoped request",
+            "context": " ".join(case["required_handoff_refs"]), "constraints": ["No new authority."],
+            "inputs": [], "acceptance": ["Return the actual disposition."],
+            "return_requirement": "Return a scoped result.",
+        }}
+    wire = evaluation.CHAT_REVIEW_OPEN_TAG + json.dumps(response) + evaluation.CHAT_REVIEW_CLOSE_TAG
+    parsed = evaluation.parse_agent_response(wire)
+    assert evaluation.score(case, parsed)["passed"]
+    context = ROUTING_SUITE["contexts"][case["context_ref"]]
+    for recipient in context["agent_directory"]:
+        wrong = {**parsed, "context_handoff": {
+            "goal_id": recipient["goal_id"], "agent_id": recipient["agent_id"],
+            "brief": (response.get("context_handoff") or {}).get("brief"),
+        }}
+        if case.get("target") != {k: recipient[k] for k in ("goal_id", "agent_id")}:
+            assert not evaluation.score(case, wrong)["passed"]
+    for effect in ({"proposals": [{"kind": "todo"}]}, {"goal_draft": {"objective": "Duplicate work"}},
+                   {"protected_action": {"operation": "deploy", "target": "unrequested"}}):
+        assert not evaluation.score(case, {**parsed, **effect})["passed"]
+    if case["expected"] == "handoff":
+        assert not evaluation.score(case, {**parsed, "gate": {"kind": "confirmation"}})["passed"]
+        for omitted in case["required_handoff_refs"]:
+            brief = {**response["context_handoff"]["brief"], "context": " ".join(
+                ref for ref in case["required_handoff_refs"] if ref != omitted)}
+            assert not evaluation.score(case, {**parsed, "context_handoff": {
+                **case["target"], "brief": brief}})["passed"]
+
+
+def test_routing_fixture_separates_topic_identity_delivery_and_permission():
+    contexts = ROUTING_SUITE["contexts"]
+    topic = contexts["topic"]
+    assert topic["agent_directory"][0]["agent_id"] == "generalist-2"
+    assert topic["agent_directory"][1]["agent_id"] == "topic-investigator"
+    assert len({row["profile_role"] for row in topic["agent_directory"]}) == 1
+    assert topic["handoff_receipts"][0]["agent_id"] == "generalist-2"
+    assert topic["recent_deliveries"][0]["agent_id"] == "topic-investigator"
+    for receipt in contexts["settled"]["handoff_receipts"]:
+        assert receipt["delivery"]["status"] == "delivered"
+        assert receipt["read"]["status"] == receipt["decision"]["status"] == "not_recorded"
+    assert not contexts["new-work"]["goals"][0]["current_todos"]
+    assert not contexts["ambiguous"]["recent_deliveries"]
+    assert len({row["scope_summary"] for row in contexts["ambiguous"]["agent_directory"]}) == 1
+    denied = contexts["ungranted"]
+    assert denied["agent_directory"][1]["context_delivery"] == "not_granted"
+    assert denied["context_delegation"]["targets"] == [topic["context_delegation"]["targets"][0]]
 
 
 def test_followthrough_oracles_distinguish_context_not_question_wording():

@@ -1145,8 +1145,9 @@ def start(**kwargs):
     }
 
 
+@pytest.mark.parametrize("previous_context_version", [None, 21])
 def test_legacy_manager_migrates_without_project_and_refreshes_each_turn(
-    monkeypatch, tmp_path
+    monkeypatch, tmp_path, previous_context_version
 ):
     store = ChatSessionStore(tmp_path / "runtime" / "chat")
     runtime = ChatRuntimeController(
@@ -1176,6 +1177,8 @@ def start(**kwargs):
         upstream_mode="chat",
         codex_home=str(runtime.codex_home),
     )
+    if previous_context_version is not None:
+        store.update_session(legacy["session_id"], manager_context_version=previous_context_version)
     session, resumed = runtime.open_session(
         goal_id="different-project",
         agent_id="codex",