fix(gateway): resume follows the compression tip so post-compression replies render

Auto-compression ends the live session and forks a continuation child
(linked via parent_session_id). A long-lived parent keeps its own flushed
message rows, so resolve_resume_session_id()'s empty-head walk never
redirected it — resuming the parent id reloaded the pre-compression
transcript and dropped every turn generated after compression, including
the assistant's response. On the desktop this is the recurring "I sent a
message, came back, and the reply isn't there" report on large sessions:
the chat's routed id is the pre-rotation id, and both the gateway
session.resume RPC and the REST /messages read anchored on it.

Fix the resolver at the chokepoint: resolve_resume_session_id() now
follows the compression-continuation chain forward via get_compression_tip()
before its existing empty-head descendant walk. get_compression_tip() only
follows children whose parent ended with end_reason='compression' (created
after the parent was ended), so delegation/branch children never hijack a
resume. This fixes every resume caller at once (REST /messages, CLI
--resume, gateway /resume).

session.resume in tui_gateway was the one resume path that never called the
resolver — it used the raw target id directly. Route it through
resolve_resume_session_id() too (non-lazy only; lazy watch windows must
stay on their exact child branch). Resolving up front also re-anchors the
live-session fast path so a still-live rotated session is reused by its new
key instead of rebuilding a duplicate agent on the stale parent.

Tests:
- resolve_resume_session_id follows the tip even when the parent retains
  messages, and is not confused by a delegation child.
- session.resume binds the agent to the continuation tip and returns the
  post-compression reply.

tests/hermes_state/test_resolve_resume_session_id.py

@@ -83,6 +83,46 @@ def test_walks_from_middle_of_chain(db):
     assert db.resolve_resume_session_id("c") == "d"
 
 
+def test_follows_compression_tip_when_parent_retains_messages(db):
+    # The bug behind the desktop "I came back and the reply isn't there" report
+    # on large sessions: auto-compression ends the live session and forks a
+    # continuation child, but a long parent keeps its own flushed message rows.
+    # The empty-head walk below never redirects a non-empty head, so resuming
+    # the parent id reloaded the pre-compression transcript and the response
+    # generated *after* compression (which lives in the continuation) was
+    # missing. resolve_resume_session_id must follow the compression-tip chain
+    # forward even when the parent still has messages.
+    base = int(time.time()) - 10_000
+    db.create_session("root", source="cli")
+    db.append_message("root", role="user", content="pre-compression turn")
+    db.end_session("root", "compression")
+    db.create_session("cont", source="cli", parent_session_id="root")
+    db.append_message("cont", role="assistant", content="post-compression reply")
+    # Force deterministic ordering so the continuation's started_at is clearly
+    # at/after the parent's ended_at (the get_compression_tip discriminator).
+    db._conn.execute("UPDATE sessions SET started_at = ?, ended_at = ? WHERE id = 'root'", (base, base + 50))
+    db._conn.execute("UPDATE sessions SET started_at = ? WHERE id = 'cont'", (base + 100,))
+    db._conn.commit()
+
+    assert db.resolve_resume_session_id("root") == "cont"
+
+
+def test_compression_tip_not_confused_with_delegation_child(db):
+    # A delegation/branch child is created while the parent is still live (the
+    # parent is NOT ended with end_reason='compression'), so resuming the
+    # parent must stay on the parent, not get hijacked into the subagent branch.
+    base = int(time.time()) - 10_000
+    db.create_session("conv", source="cli")
+    db.append_message("conv", role="user", content="parent turn")
+    db.create_session("subagent", source="cli", parent_session_id="conv")
+    db.append_message("subagent", role="assistant", content="delegated work")
+    db._conn.execute("UPDATE sessions SET started_at = ? WHERE id = 'conv'", (base,))
+    db._conn.execute("UPDATE sessions SET started_at = ? WHERE id = 'subagent'", (base + 100,))
+    db._conn.commit()
+
+    assert db.resolve_resume_session_id("conv") == "conv"
+
+
 def test_prefers_most_recent_child_when_fork_exists(db):
     # If a session was somehow forked (two children), pick the latest one.
     # In practice, compression only produces single-chain shape, but the helper

tests/test_tui_gateway_server.py

@@ -954,6 +954,65 @@ def test_session_resume_uses_parent_lineage_for_display(monkeypatch):
     assert captured["history_calls"] == [("tip", False), ("tip", True)]
 
 
+def test_session_resume_follows_compression_tip(monkeypatch, tmp_path):
+    """Resuming a rotated-out parent id must load the continuation's messages.
+
+    Regression for the desktop "I came back and the reply isn't there" report:
+    auto-compression ends the live session and forks a continuation child, so a
+    resume on the parent id (the desktop's routed id when the chat was opened
+    before it rotated) used to reload the pre-compression transcript and drop
+    the response generated after compression. session.resume must follow the
+    compression tip via resolve_resume_session_id.
+    """
+    from hermes_state import SessionDB
+
+    db = SessionDB(db_path=tmp_path / "state.db")
+    base = int(time.time()) - 10_000
+    db.create_session("parent_root", source="tui")
+    db.append_message("parent_root", role="user", content="pre-compression turn")
+    db.end_session("parent_root", "compression")
+    db.create_session("cont_tip", source="tui", parent_session_id="parent_root")
+    db.append_message("cont_tip", role="assistant", content="post-compression reply")
+    db._conn.execute(
+        "UPDATE sessions SET started_at = ?, ended_at = ? WHERE id = 'parent_root'",
+        (base, base + 50),
+    )
+    db._conn.execute("UPDATE sessions SET started_at = ? WHERE id = 'cont_tip'", (base + 100,))
+    db._conn.commit()
+
+    captured = {}
+
+    def fake_make_agent(sid, key, session_id=None, session_db=None, **kwargs):
+        captured["agent_session_id"] = session_id
+        return types.SimpleNamespace(model="test", provider="test")
+
+    monkeypatch.setattr(server, "_get_db", lambda: db)
+    monkeypatch.setattr(server, "_enable_gateway_prompts", lambda: None)
+    monkeypatch.setattr(server, "_set_session_context", lambda target: [])
+    monkeypatch.setattr(server, "_clear_session_context", lambda tokens: None)
+    monkeypatch.setattr(server, "_make_agent", fake_make_agent)
+    monkeypatch.setattr(
+        server, "_session_info", lambda agent, *a: {"model": "test", "tools": {}, "skills": {}}
+    )
+    monkeypatch.setattr(
+        server, "_init_session", lambda sid, key, agent, history, cols=80, **_kwargs: None
+    )
+
+    try:
+        resp = server.handle_request(
+            {"id": "1", "method": "session.resume", "params": {"session_id": "parent_root"}}
+        )
+    finally:
+        db.close()
+
+    # The agent must bind to the continuation tip, and the returned transcript
+    # must include the post-compression reply (which lives only in the tip).
+    assert resp["result"]["session_key"] == "cont_tip"
+    assert captured["agent_session_id"] == "cont_tip"
+    texts = [m.get("text") for m in resp["result"]["messages"]]
+    assert "post-compression reply" in texts
+
+
 def test_session_resume_passes_stored_runtime_to_agent(monkeypatch):
     captured = {}