Skip to content

The agent that grows with you

fix(gateway): reuse a session's context breakdown until its transcript moves

Tekniuma7e69984ce97
diff --git a/gateway/session_authority.py b/gateway/session_authority.py
index a3b7011b029..d85b652650b 100644
--- a/gateway/session_authority.py
+++ b/gateway/session_authority.py
@@ -58,6 +58,9 @@ class LiveSession:
# (a contended SQLite writer must not stall loop-side lock takers); readers holding that lock
# still see the claim as started, so no snapshot shows a terminal row without its completion.
settling: dict = field(default_factory=dict)
+ # ``(revision key, payload)`` of the last ``session.context_breakdown``: Desktop re-polls it on
+ # every idle mount and reconnect, and a rebuild walks the whole stored transcript (#140667).
+ context_breakdown: tuple | None = None
# A viewer attached while a guarded deferred pick awaits consent: the next boundary asks again.
reask_pick: bool = False
# A historic messaging row mounted for reading after its lane moved on (``/new``): ``route`` is
diff --git a/gateway/session_tui_reads.py b/gateway/session_tui_reads.py
index 6ec610044d4..a5b7b1399bc 100644
--- a/gateway/session_tui_reads.py
+++ b/gateway/session_tui_reads.py
@@ -163,8 +163,25 @@ async def context_breakdown(connection, ref, params) -> dict:
from agent.context_breakdown import compute_session_context_breakdown
from agent.context_file_sources import context_file_sources_for_agent
with owner_scope(authority), _session_workspace(authority.runner, live.source):
+ # A revision probe is two indexed reads; the rebuild walks the whole transcript and the
+ # system prompt. Reuse the payload while neither the transcript nor the agent moved.
+ key = (id(agent), _agent_context_marks(agent),
+ authority.db.conversation_revision(stored, include_ancestors=True))
+ cached = live.context_breakdown
+ if cached is not None and cached[0] == key:
+ return dict(cached[1])
conversation = authority.db.get_messages_as_conversation(stored, include_ancestors=True)
payload = compute_session_context_breakdown(agent, conversation)
payload['context_files'] = context_file_sources_for_agent(agent)
- return payload
+ live.context_breakdown = (key, payload)
+ return dict(payload)
return await asyncio.to_thread(build)
+
+
+def _agent_context_marks(agent) -> tuple:
+ """What the breakdown reads off the agent besides the transcript: model, toolset size, window,
+ measured prompt tokens, compressions and the usage anchors."""
+ comp = getattr(agent, 'context_compressor', None)
+ return (getattr(agent, 'model', None), getattr(agent, 'provider', None), len(getattr(agent, 'tools', None) or ()),
+ *(getattr(comp, name, None) for name in ('context_length', 'last_prompt_tokens', 'compression_count')),
+ repr(getattr(agent, '_turn_base_usage_anchor', None)), repr(getattr(agent, '_usage_anchor', None)))
diff --git a/hermes_state_messages.py b/hermes_state_messages.py
index 299e8b59d46..172dfb0d14f 100644
--- a/hermes_state_messages.py
+++ b/hermes_state_messages.py
@@ -1532,6 +1532,16 @@ class SessionMessagesMixin(SessionDetachedResultsMixin):
repair_alternation=repair_alternation, include_row_ids=include_row_ids,
include_summary_markers=repair_alternation)
+ def conversation_revision(self, session_id: str, include_ancestors: bool = False) -> tuple:
+ """A cheap fingerprint of what ``get_messages_as_conversation`` (live rows) would load: the
+ lineage ids plus the active rows' count and highest id. Any append, rewind, undo or compaction
+ moves it, so a payload derived from the transcript can be reused while it holds."""
+ ids = self._resume_lineage_ids(session_id) if include_ancestors else [session_id]
+ row = self._read_one(
+ f"SELECT COUNT(*), COALESCE(MAX(id), 0) FROM messages WHERE session_id IN ({_placeholders(ids)})"
+ " AND active = 1", tuple(ids))
+ return (tuple(ids), int(row[0]), int(row[1]))
+
def _dedupe_replayed_user(self, messages, msg, exact_user_clones) -> tuple[bool, Any]:
"""Ancestor-lineage dedupe of one decoded user *msg* -> ``(skip, exact_clone_key)``. Rotation
column-clones the concurrent tail into the child, so copies need not be adjacent: the exact
diff --git a/tests/gateway/test_tui_session_reads.py b/tests/gateway/test_tui_session_reads.py
index c63b92a8b66..663213b8a6c 100644
--- a/tests/gateway/test_tui_session_reads.py
+++ b/tests/gateway/test_tui_session_reads.py
@@ -137,3 +137,35 @@ async def test_context_breakdown_verb_serves_the_owner_agents_categories(owner,
monkeypatch.setattr(context_file_sources, 'context_file_sources_for_agent', lambda a: [])
live = await owner.call('session.context_breakdown', session_id=sid)
assert live['categories'][0]['tokens'] == 3000 and live['model'] == 'fake-model', live
+
+
+async def test_context_breakdown_is_reused_until_the_transcript_moves(owner, monkeypatch):
+ """#140667: Desktop re-polls ``session.context_breakdown`` on every idle mount and reconnect,
+ and each poll rebuilt the prompt over the whole stored transcript (seconds on a 9,501-row
+ session). The owner reuses the payload while the transcript revision is unchanged and
+ recomputes once a row lands."""
+ sid = (await owner.call('session.create', request_id='r', source='cli', cwd=str(owner.workdir)))['session_id']
+ from agent import context_breakdown, context_file_sources
+ agent = SimpleNamespace(model='fake-model', provider='custom')
+ owner.authority.runner._cached_agent_for = lambda route: agent
+ computed = []
+
+ def compute(a, messages):
+ computed.append(len(messages))
+ return {'categories': [{'id': 'conversation', 'label': 'Conversation', 'tokens': len(messages),
+ 'color': '#fff'}], 'context_max': 128000, 'context_used': 1,
+ 'context_percent': 0, 'estimated_total': len(messages), 'context_estimated': True,
+ 'context_source': 'local_estimate', 'model': 'fake-model'}
+ monkeypatch.setattr(context_breakdown, 'compute_session_context_breakdown', compute)
+ monkeypatch.setattr(context_file_sources, 'context_file_sources_for_agent', lambda a: [])
+ first = await owner.call('session.context_breakdown', session_id=sid)
+ again = await owner.call('session.context_breakdown', session_id=sid)
+ assert computed == [0] and again == first
+ from gateway.session_tui_reads import _stored_id
+ stored = _stored_id(owner.authority, SimpleNamespace(session_id=sid))
+ owner.authority.db.append_message(stored, 'user', 'hello')
+ moved = await owner.call('session.context_breakdown', session_id=sid)
+ assert computed == [0, 1] and moved['estimated_total'] == 1
+ await owner.call('session.context_breakdown', session_id=sid)
+ assert computed == [0, 1]