fix(nous-portal): give auxiliary calls the Portal sticky routing key

The Nous Portal profile publishes a top-level `session_id` that the Portal
uses as its sticky routing key, pinning a conversation to one upstream
endpoint so explicit Anthropic `cache_control` breakpoints stay warm
(those caches are instance-local, so a reroute cold-writes instead of reads).

`NousProfile.build_extra_body` sourced that key only from the explicit
`session_id` argument. Auxiliary call sites — compression, title generation,
vision, web_extract, session_search, MoA slots — funnel through
`agent.auxiliary_client`, which has no session handle and never passes one.
They carried the `conversation=` tag but NO sticky key at all, so each one
routed independently of the conversation it belonged to.

Resolve the key the way `nous_portal_tags` already resolves that tag:
ambient lineage-ROOT contextvar first, explicit argument as fallback. Fixes
every aux call site at once with no per-call-site plumbing.

Also publish the root in the `_compress_context` forwarder when nothing is
ambient. Out-of-turn entry points (`/compact`, the gateway `/compress`
command and its hygiene sweep, partial head compression) call it outside
`run_conversation`'s scope, so the summarizer's aux call ran untagged and
unkeyed; the caller's value is restored in `finally` so a compaction never
leaks its tag.

Scope note: this does NOT keep the compaction turn's own prompt cache warm.
Compaction replaces the history with a summary and rebuilds the system
prompt, so that request is a cold write on any endpoint — what a stable key
buys is the turns AFTER compaction reading the cache it wrote. Under the
default `compression.in_place: true` (#38763) the session id does not rotate
at compaction, so for the main loop the ambient root and the explicit
argument already agree; the ambient resolution additionally holds the key
stable for installs that opt back into rotating compaction and across
delegate-subagent trees.

Salvaged from #71747. Subsumes the aux-call half of #70883 (@webtecnica),
which threaded `session_id` through `call_llm` to close the same gap.

Tests: 4 cases in tests/agent/test_portal_tags.py; both fix halves verified
to fail against the pre-fix behavior via sabotage runs.

plugins/model-providers/nous/__init__.py

@@ -2,7 +2,7 @@
 
 from typing import Any
 
-from agent.portal_tags import nous_portal_tags
+from agent.portal_tags import get_conversation_context, nous_portal_tags
 from providers import register_provider
 from providers.base import ProviderProfile
 
@@ -14,15 +14,35 @@ class NousProfile(ProviderProfile):
         self, *, session_id: str | None = None, **context
     ) -> dict[str, Any]:
         body: dict[str, Any] = {"tags": nous_portal_tags(session_id=session_id)}
-        if session_id:
-            # Top-level session_id → provider sticky routing key. Pins every
-            # turn of a session to the same upstream endpoint so explicit
-            # Anthropic cache_control breakpoints stay warm instead of
-            # cold-writing a fresh cache on each reroute (Anthropic/Vertex/
-            # Bedrock caches are instance-local). Mirrors the OpenRouter
-            # profile; without it the portal falls back to hashing the opening
-            # messages, which breaks pinning whenever those shift.
-            body["session_id"] = session_id
+        # Top-level session_id → provider sticky routing key. Pins every
+        # turn of a session to the same upstream endpoint so explicit
+        # Anthropic cache_control breakpoints stay warm instead of
+        # cold-writing a fresh cache on each reroute (Anthropic/Vertex/
+        # Bedrock caches are instance-local). Mirrors the OpenRouter
+        # profile; without it the portal falls back to hashing the opening
+        # messages, which breaks pinning whenever those shift.
+        #
+        # Resolve it exactly like ``nous_portal_tags`` resolves the
+        # ``conversation=`` tag: ambient context first (the lineage ROOT id
+        # published by the agent loop), explicit argument as fallback.
+        #
+        # The gap this closes is the auxiliary call sites — compression,
+        # title generation, vision, web_extract, session_search, MoA slots.
+        # They funnel through ``agent.auxiliary_client`` which has no session
+        # handle, so they never pass ``session_id``: they carried the
+        # ``conversation=`` tag but NO sticky key at all, and each one routed
+        # independently of the conversation it belongs to. Reading the same
+        # ambient contextvar the tag already uses fixes that with zero
+        # per-call-site plumbing.
+        #
+        # For the main loop the two agree anyway under the default
+        # ``compression.in_place: true`` (#38763), where compaction keeps the
+        # session id; the ambient root additionally keeps the key stable for
+        # installs that opt back into rotating compaction, and across
+        # delegate-subagent trees.
+        sticky_key = get_conversation_context() or session_id
+        if sticky_key:
+            body["session_id"] = sticky_key
         provider_preferences = context.get("provider_preferences")
         if provider_preferences:
             body["provider"] = provider_preferences

run_agent.py

@@ -6462,13 +6462,43 @@ class AIAgent:
         ``force=False``.
         """
         from agent.conversation_compression import compress_context
-        return compress_context(
-            self, messages, system_message,
-            approx_tokens=approx_tokens, task_id=task_id, focus_topic=focus_topic,
-            force=force,
-            defer_context_engine_notification=defer_context_engine_notification,
-            commit_fence=commit_fence,
+        from agent.portal_tags import (
+            get_conversation_context,
+            reset_conversation_context,
+            set_conversation_context,
         )
+        # Out-of-turn compaction entry points — ``/compact`` (cli.py), the
+        # gateway ``/compress`` command and its hygiene sweep (both of which
+        # build a throwaway agent), and partial head compression — call this
+        # forwarder directly, outside ``run_conversation``'s ambient scope.
+        # With nothing ambient the summarizer's auxiliary call carries no
+        # conversation tag and no Portal sticky key, so it routes independently
+        # of the conversation it belongs to. Publish the root here as a
+        # fallback; in-turn callers already have it set to the same value, so
+        # this is a no-op for them.
+        #
+        # Note this does NOT keep the compaction turn's own prompt cache warm:
+        # compaction replaces the history with a summary and rebuilds the
+        # system prompt, so that request is a cold write on any endpoint. What
+        # it buys is the turns AFTER compaction reading the cache it wrote.
+        token = None
+        if get_conversation_context() is None:
+            root = self._conversation_root_id()
+            if root:
+                token = set_conversation_context(root)
+        try:
+            return compress_context(
+                self, messages, system_message,
+                approx_tokens=approx_tokens, task_id=task_id, focus_topic=focus_topic,
+                force=force,
+                defer_context_engine_notification=defer_context_engine_notification,
+                commit_fence=commit_fence,
+            )
+        finally:
+            # Restore whatever the caller had, so a compaction never leaks its
+            # tag into the surrounding scope.
+            if token is not None:
+                reset_conversation_context(token)
 
     def _set_tool_guardrail_halt(self, decision: ToolGuardrailDecision) -> None:
         """Record the first guardrail decision that should stop this turn."""

tests/agent/test_portal_tags.py

@@ -217,3 +217,112 @@ def test_nous_provider_profile_uses_helper():
     assert profile is not None
     body = profile.build_extra_body()
     assert body["tags"] == nous_portal_tags()
+
+
+def test_nous_sticky_key_matches_conversation_tag():
+    """Sticky routing key must resolve like the ``conversation=`` tag does.
+
+    The load-bearing case is the auxiliary call sites (compression, titles,
+    vision, MoA slots): they pass no ``session_id`` at all, so before this
+    resolution they carried the conversation tag but NO Portal sticky key and
+    routed independently of their conversation.
+
+    The explicit-argument case matters for installs that opt out of the
+    default ``compression.in_place: true`` (#38763) and therefore still rotate
+    ``agent.session_id`` at compaction, and for delegate-subagent trees that
+    should tag as the parent conversation.
+    """
+    from agent.portal_tags import (
+        conversation_tag,
+        reset_conversation_context,
+        set_conversation_context,
+    )
+    from providers import get_provider_profile
+
+    profile = get_provider_profile("nous")
+    token = set_conversation_context("root-conversation")
+    try:
+        # Rotated segment id passed explicitly — root still wins, both places.
+        body = profile.build_extra_body(session_id="segment-after-compaction")
+        assert body["session_id"] == "root-conversation"
+        assert conversation_tag("root-conversation") in body["tags"]
+
+        # Auxiliary call sites pass no session_id but inherit the context.
+        aux = profile.build_extra_body()
+        assert aux["session_id"] == "root-conversation"
+    finally:
+        reset_conversation_context(token)
+
+
+def test_nous_sticky_key_falls_back_to_explicit_session_id():
+    """Outside any agent turn the explicit session_id remains the sticky key."""
+    from providers import get_provider_profile
+
+    profile = get_provider_profile("nous")
+    body = profile.build_extra_body(session_id="explicit-only")
+    assert body["session_id"] == "explicit-only"
+    assert profile.build_extra_body().get("session_id") is None
+
+
+def test_compress_context_publishes_root_when_called_out_of_turn(monkeypatch):
+    """Out-of-turn compaction must still carry the conversation tag.
+
+    ``/compact``, the gateway ``/compress`` command and its hygiene sweep call
+    ``_compress_context`` directly, outside ``run_conversation``'s ambient
+    scope. That call ships the full uncompressed history — the largest prompt
+    of the session — so losing the sticky key there reroutes it to a cold
+    endpoint.
+    """
+    import agent.conversation_compression as cc
+    from agent.portal_tags import get_conversation_context
+    from run_agent import AIAgent
+
+    seen = {}
+
+    def _fake_compress(agent, messages, system_message, **kwargs):
+        seen["conversation"] = get_conversation_context()
+        return ([], "")
+
+    monkeypatch.setattr(cc, "compress_context", _fake_compress)
+
+    class _Agent:
+        def _conversation_root_id(self):
+            return "root-abc"
+
+    AIAgent._compress_context(_Agent(), [], "sys")
+
+    assert seen["conversation"] == "root-abc"
+    # The scope is local: nothing leaks into the caller's context.
+    assert get_conversation_context() is None
+
+
+def test_compress_context_preserves_ambient_context(monkeypatch):
+    """In-turn compaction inherits the turn's root and restores it untouched."""
+    import agent.conversation_compression as cc
+    from agent.portal_tags import (
+        get_conversation_context,
+        reset_conversation_context,
+        set_conversation_context,
+    )
+    from run_agent import AIAgent
+
+    seen = {}
+
+    def _fake_compress(agent, messages, system_message, **kwargs):
+        seen["conversation"] = get_conversation_context()
+        return ([], "")
+
+    monkeypatch.setattr(cc, "compress_context", _fake_compress)
+
+    class _Agent:
+        def _conversation_root_id(self):
+            # A rotated segment id must never win over the ambient root.
+            return "segment-after-compaction"
+
+    token = set_conversation_context("outer-root")
+    try:
+        AIAgent._compress_context(_Agent(), [], "sys")
+        assert seen["conversation"] == "outer-root"
+        assert get_conversation_context() == "outer-root"
+    finally:
+        reset_conversation_context(token)