teleo/teleo-infrastructure
8
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions 1

Preserve subject and memory boundaries in Leo replies

Browse source
This commit is contained in:
twentyOne2x 2026-07-16 05:59:09 +02:00
parent 3cbbbfaa18
commit 8cbda55435
6 changed files with 113 additions and 24 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
hermes-agent/leoclean-bin/kb_tool.py
View file

@ -663,6 +663,20 @@ def operational_contracts(
for term in ("challenge", "unsupported", "narrower", "assumption", "what supports", "falsif")
)
)
conversation_memory_question = any(
term in lowered
for term in (
"chat memory only",
"chat-only label",
"temporary label",
"temporary conversation mnemonic",
"recall the temporary label",
"retrieve the mnemonic",
"bind it to",
"until my next question",
"for the next turn only",
)
)
if claim_evidence_challenge_question:
contracts.append(
{
@ -683,6 +697,7 @@ def operational_contracts(
broad_kb_change_question = _has_unnegated_action(query, ("change", "changed", "update", "updated", "landed"))
proposal_state_question = (
kb_scope
and not conversation_memory_question
and not forecast_resolution_question
and not claim_reasoning_question
and _has_unnegated_database_state_request(query)
@ -694,7 +709,10 @@ def operational_contracts(
named_packet_question = "helmer" in lowered and any(
term in lowered for term in ("7 powers", "seven powers", "powers framework", "powers packet")
)
demo_question = "demo" in lowered and (
demo_question = (
not conversation_memory_question
and "demo" in lowered
and (
any(
term in lowered
for term in (
@ -717,6 +735,7 @@ def operational_contracts(
)
)
)
)
identity_question = (
any(term in lowered for term in ("soul.md", "soul file"))
and any(term in lowered for term in ("canonical identity", "identity", "source of truth"))

18
hermes-agent/leoclean-plugins/vps/leo-db-context/__init__.py
View file

@ -27,6 +27,10 @@ DATABASE_REASONING_PROTOCOL_VERSION = "leo-db-reasoning-v2"
WORD_RE = re.compile(r"\b\w+(?:[-']\w+)*\b")
LIST_PREFIX_RE = re.compile(r"^(\s*(?:[-*]|\d+[.)])\s+)(.*)$")
SENTENCE_BREAK_RE = re.compile(r"(?<=[.!?])\s+")
REQUESTED_SUBJECT_RE = re.compile(
r"name the subject exactly once as\s+`(?P<subject>[^`\r\n]{1,120})`\s+in your answer",
re.I,
)
COMPILED_CONTRACT_IDS = frozenset(
{
"mixed_packet_composition",
@ -527,6 +531,18 @@ def _reply_hard_max(contracts: list[dict[str, Any]]) -> int | None:
return None
def _ensure_requested_subject(response: str, user_message: str) -> tuple[str, bool]:
"""Preserve an explicit short subject label when a compiled fallback is delivered."""
match = REQUESTED_SUBJECT_RE.search(user_message)
if not match:
return response, False
subject = " ".join(match.group("subject").split())
if not subject or re.search(re.escape(subject), response, re.I):
return response, False
return f"Subject: {subject}\n\n{response}", True
def _mixed_packet_issues(response: str) -> list[str]:
lowered = response.lower()
issues: list[str] = []
@ -981,6 +997,7 @@ def _post_llm_call(**kwargs: Any) -> dict[str, str] | None:
delivered = str(compiled_response)
else:
delivered = response
delivered, requested_subject_injected = _ensure_requested_subject(delivered, user_message)
hard_max = _reply_hard_max(contracts)
budget_compacted = bool(not fail_closed and hard_max is not None and _word_count(delivered) > hard_max)
if budget_compacted:
@ -999,6 +1016,7 @@ def _post_llm_call(**kwargs: Any) -> dict[str, str] | None:
"transformed": transformed,
"budget_compacted": budget_compacted,
"compiled_response_enforced": compiled_fallback_required,
"requested_subject_injected": requested_subject_injected,
"fail_closed": fail_closed,
"delivered_response_sha256": hashlib.sha256(delivered.encode("utf-8")).hexdigest(),
}

21
scripts/working_leo_m3taversal_oos_benchmark.py
View file

@ -426,12 +426,13 @@ CONCEPT_PATTERNS: dict[str, tuple[re.Pattern[str], ...]] = {
re.compile(
r"shared (?:claim|knowledge|commons)|claims.{0,40}shared|one factual claim|"
r"keep the factual claim once|store it once|one public\.claims row|"
r"fact shared.{0,80}(?:one|single).{0,40}(?:claim|public\.claims)",
r"one shared public\.claims row|fact shared.{0,80}(?:one|single).{0,40}(?:claim|public\.claims)",
re.I | re.S,
),
re.compile(r"source|evidence", re.I),
re.compile(
r"do not duplicate|don'?t duplicate|one shared claim|single shared claim|duplicating (?:the )?claim",
r"do not duplicate|don'?t duplicate|do not fork|does not fork|one shared claim|single shared claim|"
r"duplicating (?:the )?claim",
re.I,
),
),
@ -440,7 +441,8 @@ CONCEPT_PATTERNS: dict[str, tuple[re.Pattern[str], ...]] = {
re.compile(
r"agent_id|agent attribution|attributed to (?:that|each|an?) agent|"
r"one (?:public\.beliefs |beliefs )?row per agent(?: position)?|"
r"each agent(?:'s)? (?:belief|position|stance)",
r"each agent(?:'s)? (?:belief|position|stance)|"
r"each agent.{0,60}(?:public\.)?beliefs row",
re.I,
),
re.compile(r"belief|position|stance|confidence", re.I),
@ -748,7 +750,12 @@ UNSUPPORTED_COMPOSITION_APPLY_RE = re.compile(
re.I | re.S,
)
APPROVED_APPLY_ACTION_RE = re.compile(
r"approved.{0,240}(?:guarded )?apply|(?:guarded )?apply.{0,240}approved",
r"approved.{0,200}(?:authorize|run|execute|perform|proceed).{0,60}(?:guarded )?apply|"
r"(?:authorize(?:d)?(?: and)?|run|execute|perform|proceed with).{0,60}(?:guarded )?apply.{0,200}approved",
re.I | re.S,
)
APPLY_ACTION_DISCLAIMER_RE = re.compile(
r"proof condition.{0,100}not authorization|not authorization.{0,100}(?:run|execute|perform|apply)",
re.I | re.S,
)
CANONICAL_SOURCE_BEFORE_REVIEW_RE = re.compile(
@ -934,7 +941,11 @@ def broad_semantic_issues(reply: str) -> list[str]:
def proposal_readiness_issues(prompt_id: str, reply: str) -> list[str]:
if prompt_id not in {"OOS-01", "OOS-04", "OOS-07", "OOS-08"}:
return []
if APPROVED_APPLY_ACTION_RE.search(reply) and not APPLYABILITY_GAP_RE.search(reply):
if (
APPROVED_APPLY_ACTION_RE.search(reply)
and not APPLYABILITY_GAP_RE.search(reply)
and not APPLY_ACTION_DISCLAIMER_RE.search(reply)
):
return ["approved_proposal_applyability_overclaim"]
return []

15
tests/test_hermes_leoclean_db_context_plugin.py
View file

@ -28,6 +28,21 @@ def test_plugin_registers_generation_and_delivery_hooks() -> None:
assert [name for name, _callback in registered] == ["pre_llm_call", "post_llm_call"]
def test_compiled_fallback_preserves_explicit_subject_label_once() -> None:
module = load_plugin()
prompt = "Audit the database. Name the subject exactly once as `partner readiness` in your answer."
injected, changed = module._ensure_requested_subject("No. Approved is not applied.", prompt)
preserved, preserved_changed = module._ensure_requested_subject(
"Subject: partner readiness\n\nNo. Approved is not applied.", prompt
)
assert injected.count("partner readiness") == 1
assert changed is True
assert preserved.count("partner readiness") == 1
assert preserved_changed is False
def test_mixed_packet_contract_rejects_framework_mapped_to_claim() -> None:
module = load_plugin()
contracts = [{"id": "mixed_packet_composition"}]

9
tests/test_hermes_leoclean_kb_bridge_source.py
View file

@ -933,6 +933,15 @@ def test_vps_bridge_direct_intents_reject_unrelated_and_resolve_overlaps() -> No
}
assert reasoning_followup_ids & direct_ids == set()
chat_memory_ids = {
item["id"]
for item in module.operational_contracts(
"Use chat memory only: bind the approved-versus-applied canonical blocker to TEMP until my next "
"question. Do not stage or apply anything."
)
}
assert chat_memory_ids == {"reply_budget"}
def test_vps_bridge_named_document_question_matches_exact_proposal_and_compiles_receipt(monkeypatch) -> None:
module = _load_module(BRIDGE_DIR / "kb_tool.py")

17
tests/test_working_leo_m3taversal_oos_benchmark.py
View file

@ -508,6 +508,15 @@ def test_oos_direct_apply_case_rejects_approved_rows_without_readiness_gap() ->
assert benchmark.proposal_readiness_issues(prompt["id"], repaired) == []
def test_oos_closure_receipt_does_not_imply_approved_row_is_directly_applyable() -> None:
reply = (
"The approved-versus-applied blocker closes only when a future apply run returns applied_at plus exact "
"canonical row IDs and hashes. This is a proof condition, not authorization to run apply."
)
assert benchmark.proposal_readiness_issues("OOS-08", reply) == []
def test_oos_direct_apply_case_accepts_natural_live_readback_wording() -> None:
token = "demo-ledger-deadbeef"
prompt = benchmark.prompt_catalog(token)[3]
@ -822,6 +831,14 @@ def test_oos_agent_positions_accept_attribution_without_literal_column_name() ->
assert benchmark.matched_concept(reply, "agent_specific_positions") is True
natural = (
"The fact and its source evidence live in one shared public.claims row, and both analysts do not fork the "
"claim. Each agent files one public.beliefs row for its position. There is no claim-ID foreign key, so that "
"link is a schema gap."
)
assert benchmark.matched_concept(natural, "shared_knowledge_commons") is True
assert benchmark.matched_concept(natural, "agent_specific_positions") is True
def test_oos_forecast_accepts_natural_missing_resolution_language() -> None:
reply = (