fix(agent): never persist empty-response recovery scaffolding

Ephemeral empty-response/prefill recovery scaffolding (the synthetic
assistant "(empty)" turn, the user nudge, the terminal "(empty)"
sentinel, and the thinking-only prefill placeholder) exists only to
drive the next API retry; the in-memory loop pops it before appending
the real response. The append-only flush did not mirror that, so a
mid-turn persist could commit scaffolding to the SQLite session store
(and JSON log), and a resumed session would replay synthetic
"(empty)"/nudge turns as genuine context — re-poisoning the empty-retry
boundary forever.

Filter ephemeral scaffolding at both durable-write sites
(_flush_messages_to_session_db + _save_session_log), by flag not
position, so buried scaffolding (an answered nudge leaves the synthetic
pair mid-list) is skipped too. Covers all three flags including
_thinking_prefill.

Adapted onto current main's identity-tracking flush.

Cherry-picked from #41281 by petrichor-op.
This commit is contained in:
petrichor-op 2026-07-01 00:47:07 -07:00 committed by Teknium
parent 8db6ed7bd9
commit f2a528fb59
3 changed files with 120 additions and 0 deletions

View file

@ -213,6 +213,28 @@ from agent.tool_dispatch_helpers import (
from utils import atomic_json_write, base_url_host_matches, base_url_hostname, env_float, is_truthy_value, model_forces_max_completion_tokens
# Internal flags that mark a message as ephemeral empty-response/prefill
# recovery scaffolding: the synthetic assistant "(empty)" turn and user nudge
# injected after an empty response, the terminal "(empty)" sentinel, and the
# thinking-only prefill placeholder. These exist only to drive the next API
# retry; the in-memory loop pops them before appending the real response.
# Persistence must mirror that, otherwise an append-only flush can commit them
# to the session store and a resumed session replays synthetic "(empty)"/nudge
# turns as if they were genuine context.
_EPHEMERAL_SCAFFOLDING_FLAGS = (
"_empty_recovery_synthetic",
"_empty_terminal_sentinel",
"_thinking_prefill",
)
def _is_ephemeral_scaffolding(msg: Any) -> bool:
"""Return True when ``msg`` is internal recovery scaffolding that must never
be persisted to the durable transcript (SQLite session store or JSON log)."""
return isinstance(msg, dict) and any(
msg.get(flag) for flag in _EPHEMERAL_SCAFFOLDING_FLAGS
)
_MAX_TOOL_WORKERS = 8
@ -1706,6 +1728,17 @@ class AIAgent:
for msg in messages:
if not isinstance(msg, dict):
continue
# Never write ephemeral recovery scaffolding to the session
# store. The flush is append-only (it only advances
# _last_flushed_db_idx via identity tracking), so a synthetic
# message committed by a mid-turn persist cannot be un-written
# when the end-of-turn drop removes it from the in-memory list —
# the resumed transcript would then replay synthetic
# "(empty)"/nudge/thinking-prefill turns as if they were genuine
# context. Skip regardless of position: an answered nudge leaves
# the synthetic pair buried mid-list, not just at the tail.
if _is_ephemeral_scaffolding(msg):
continue
msg_id = id(msg)
if msg_id in flushed_ids:
continue
@ -2430,6 +2463,10 @@ class AIAgent:
try:
cleaned = []
for msg in messages:
# Mirror the SQLite flush: ephemeral recovery scaffolding is
# internal retry state, never durable transcript content.
if _is_ephemeral_scaffolding(msg):
continue
if msg.get("role") == "assistant" and msg.get("content"):
msg = dict(msg)
msg["content"] = self._clean_session_content(msg["content"])

View file

@ -46,6 +46,7 @@ ACP_REGISTRY_MANIFEST = REPO_ROOT / "acp_registry" / "agent.json"
# Auto-extracted from noreply emails + manual overrides
AUTHOR_MAP = {
"jnibarger01@gmail.com": "jnibarger01", # PR #35130 salvage (ReDoS-bound threat-pattern filler + FTS5 query cap + V4A Move-File approval/traversal targets)
"290868363+petrichor-op@users.noreply.github.com": "petrichor-op", # PR #41281 salvage (never persist ephemeral empty-response recovery scaffolding to the SQLite session store / JSON log; filter by flag not position)
"syahidfrd@gmail.com": "syahidfrd", # PR #17059 salvage (tag unverified senders in Slack thread context to mitigate indirect prompt injection)
"22971845+H2KFORGIVEN@users.noreply.github.com": "H2KFORGIVEN", # PR #22523 salvage (turn-pair preservation: never orphan the last user ask at head_end during compaction)
"5823452+sgabel@users.noreply.github.com": "sgabel", # PR #13139 salvage (redact secrets in user-facing approval prompts)

View file

@ -3,6 +3,28 @@
from run_agent import AIAgent
class _CapturingSessionDB:
"""Minimal SessionDB stand-in that records every appended message."""
def __init__(self):
self.rows = []
def append_message(self, session_id, role, content=None, **kwargs):
self.rows.append({"role": role, "content": content})
return len(self.rows)
def _agent_with_capturing_db():
agent = AIAgent.__new__(AIAgent)
agent._persist_user_message_idx = None
agent._persist_user_message_override = None
agent._session_db = _CapturingSessionDB()
agent._session_db_created = True
agent._last_flushed_db_idx = 0
agent.session_id = "sess-test"
return agent
def _agent_with_stubbed_persistence():
agent = AIAgent.__new__(AIAgent)
agent._persist_user_message_idx = None
@ -92,3 +114,63 @@ def test_persist_session_strips_marked_terminal_empty_sentinel():
assert messages == [{"role": "user", "content": "continue"}]
assert agent.flushed_session_db_messages[-1] == messages
assert all(not msg.get("_empty_terminal_sentinel") for msg in messages)
def test_flush_never_writes_buried_empty_recovery_scaffolding():
"""When an empty-after-tools nudge is followed by a tool-calling response,
the synthetic ``(empty)`` + nudge pair stays buried in the live message
list (only the trailing copies are ever dropped). The append-only flush
must skip it regardless of position, otherwise the synthetic turns land in
the session store and pollute every resumed transcript.
"""
agent = _agent_with_capturing_db()
messages = [
{"role": "user", "content": "run the task"},
{
"role": "assistant",
"content": "",
"tool_calls": [{"id": "call_1", "type": "function",
"function": {"name": "x", "arguments": "{}"}}],
},
{"role": "tool", "content": "{}", "tool_call_id": "call_1"},
# Synthetic recovery scaffolding, now buried because the model answered
# the nudge with another tool call rather than terminating.
{"role": "assistant", "content": "(empty)", "_empty_recovery_synthetic": True},
{
"role": "user",
"content": "You just executed tool calls but returned an empty response.",
"_empty_recovery_synthetic": True,
},
{
"role": "assistant",
"content": "",
"tool_calls": [{"id": "call_2", "type": "function",
"function": {"name": "x", "arguments": "{}"}}],
},
{"role": "tool", "content": "{}", "tool_call_id": "call_2"},
{"role": "assistant", "content": "All done."},
]
agent._flush_messages_to_session_db(messages, conversation_history=[])
persisted = agent._session_db.rows
assert all(row["content"] != "(empty)" for row in persisted)
assert all("empty response" not in (row["content"] or "") for row in persisted)
# Only the genuine turns reach the store, in order.
assert [r["role"] for r in persisted] == [
"user", "assistant", "tool", "assistant", "tool", "assistant",
]
assert persisted[-1]["content"] == "All done."
def test_flush_skips_thinking_prefill_scaffolding():
agent = _agent_with_capturing_db()
messages = [
{"role": "user", "content": "hi"},
{"role": "assistant", "content": "", "_thinking_prefill": True},
{"role": "assistant", "content": "Hello!"},
]
agent._flush_messages_to_session_db(messages, conversation_history=[])
assert [r["content"] for r in agent._session_db.rows] == ["hi", "Hello!"]