mirror of
https://github.com/pewdiepie-archdaemon/odysseus.git
synced 2026-10-06 23:12:22 +02:00
fix: close Wave 1.1 completion-gate audit findings
- Headless consumers (task scheduler, background follow-up) now treat a completion-gate final_response as the authoritative answer instead of collecting deltas only. A gated replacement no longer leaves scheduled output empty, which used to trigger an extra, ungated grace-summary model call. - The scheduler closes the agent stream with contextlib.aclosing, so the approval-pause break unwinds the gate's journal and teacher-takeover context in its own task. Chained runs no longer inherit a stale parent_run_id, and later finalization no longer raises ContextVar reset errors. - On provider error, the completion gate applies the live answer's statement filter to persisted round_texts. Diagnostics and the failure note survive; claims rejected by the gate cannot reappear on reload.
This commit is contained in:
Alexandre Teixeira
@@ -92,13 +92,8 @@ def _current_run_claims(statement: str, *, execution_required: bool) -> list[tup
|
||||
return claims
|
||||
|
||||
|
||||
def completion_answer(text: str, ledger: EvidenceLedger, decision: CompletionDecision) -> tuple[str, str]:
|
||||
"""Keep explanatory prose; remove unsupported assertions and attach facts.
|
||||
|
||||
Exit status proves neither test counts nor coverage. A bad assertion is
|
||||
removed at statement boundaries instead of erasing an entire explanation.
|
||||
The execution outcome remains separate from a discarded model assertion.
|
||||
"""
|
||||
def _supported_prose(text: str, ledger: EvidenceLedger, decision: CompletionDecision) -> tuple[str, str]:
|
||||
"""Remove unsupported assertions at statement boundaries; add no notice."""
|
||||
incomplete = decision.reason if not decision.can_complete and decision.status != CompletionStatus.AWAITING_USER else ''
|
||||
execution_required = _execution_obligation(ledger.requirements)
|
||||
kept = []
|
||||
@@ -128,6 +123,19 @@ def completion_answer(text: str, ledger: EvidenceLedger, decision: CompletionDec
|
||||
else:
|
||||
kept.append(statement)
|
||||
prose = ''.join(kept).strip() if removed else text
|
||||
return prose, removed
|
||||
|
||||
|
||||
def completion_answer(text: str, ledger: EvidenceLedger, decision: CompletionDecision) -> tuple[str, str]:
|
||||
"""Keep explanatory prose; remove unsupported assertions and attach facts.
|
||||
|
||||
Exit status proves neither test counts nor coverage. A bad assertion is
|
||||
removed at statement boundaries instead of erasing an entire explanation.
|
||||
The execution outcome remains separate from a discarded model assertion.
|
||||
"""
|
||||
incomplete = decision.reason if not decision.can_complete and decision.status != CompletionStatus.AWAITING_USER else ''
|
||||
execution_required = _execution_obligation(ledger.requirements)
|
||||
prose, removed = _supported_prose(text, ledger, decision)
|
||||
if incomplete or (removed and execution_required and decision.status in {CompletionStatus.UNVERIFIED, CompletionStatus.AWAITING_USER}):
|
||||
reason = incomplete or removed
|
||||
missing = (' Missing artifacts: ' + ', '.join(decision.missing_artifacts) + '.'
|
||||
@@ -336,6 +344,13 @@ def with_completion_gate(func):
|
||||
if not provider_error:
|
||||
metadata['round_texts'] = [safe_answer]
|
||||
metadata['completion_gate_reason'] = reason or unsafe_draft or 'receipt_summary'
|
||||
if provider_error and isinstance(metadata.get('round_texts'), list):
|
||||
# Failed rounds stay as per-round diagnostics, but they are
|
||||
# rendered again on reload. Apply the same statement filter
|
||||
# as the live answer so a rejected claim cannot reappear.
|
||||
metadata['round_texts'] = [
|
||||
_supported_prose(text, ledger, presentation_decision)[0] if isinstance(text, str) else text
|
||||
for text in metadata['round_texts']]
|
||||
if isinstance(metadata.get('thinking'), str):
|
||||
_, unsafe_thinking = completion_answer(metadata['thinking'], ledger,
|
||||
replace(presentation_decision, can_complete=True))
|
||||
|
||||
@@ -42,6 +42,7 @@ async def _drain_agent(sess, messages):
|
||||
saves, so the frontend rebuilds them as standard agent-thread tool cards."""
|
||||
from src.agent_loop import stream_agent_loop
|
||||
full = ""
|
||||
final_replaced = False
|
||||
tool_events = []
|
||||
round_num = 1
|
||||
async for chunk in stream_agent_loop(
|
||||
@@ -68,7 +69,17 @@ async def _drain_agent(sess, messages):
|
||||
if isinstance(delta, str):
|
||||
if d.get("thinking"):
|
||||
continue
|
||||
if final_replaced:
|
||||
# A later answer supersedes the replacement, as the
|
||||
# completion gate treats it.
|
||||
full = ""
|
||||
final_replaced = False
|
||||
full += delta
|
||||
elif d.get("type") == "final_response":
|
||||
# The completion gate may present its sanitized answer as one
|
||||
# replacement instead of deltas.
|
||||
full = str(d.get("content") or "")
|
||||
final_replaced = True
|
||||
elif d.get("type") == "agent_step":
|
||||
round_num = d.get("round", round_num)
|
||||
elif d.get("type") == "tool_output":
|
||||
|
||||
+70
-56
@@ -1976,6 +1976,7 @@ class TaskScheduler:
|
||||
except Exception:
|
||||
pass
|
||||
full_text = ""
|
||||
final_text_replaced = False
|
||||
tool_results = []
|
||||
approval_pause = None
|
||||
|
||||
@@ -1997,62 +1998,75 @@ class TaskScheduler:
|
||||
)[1:]
|
||||
except Exception:
|
||||
_task_fallbacks = []
|
||||
async for event_str in stream_agent_loop(
|
||||
endpoint_url=endpoint_url,
|
||||
model=model,
|
||||
messages=messages,
|
||||
max_rounds=_task_max_rounds,
|
||||
session_id=session_id,
|
||||
owner=task.owner,
|
||||
headers=headers,
|
||||
disabled_tools=disabled_tools,
|
||||
relevant_tools=relevant_tools,
|
||||
fallbacks=_task_fallbacks,
|
||||
workload="background",
|
||||
):
|
||||
if event_str.startswith("data: ") and not event_str.startswith("data: [DONE]"):
|
||||
try:
|
||||
data = json.loads(event_str[6:])
|
||||
# Capture text from all event types, not just delta
|
||||
if "delta" in data:
|
||||
if data.get("thinking"):
|
||||
continue
|
||||
full_text += data["delta"]
|
||||
elif data.get("type") == "tool_output":
|
||||
# Tool results — capture summary so we have SOMETHING even
|
||||
# if the model never produces a final text response
|
||||
tool_summary = data.get("stdout") or data.get("output") or data.get("result") or ""
|
||||
if isinstance(tool_summary, str) and tool_summary.strip():
|
||||
tool_results.append(f"[{data.get('tool', '?')}] {tool_summary[:500]}")
|
||||
approval = data.get("ask_user")
|
||||
if (
|
||||
isinstance(approval, dict)
|
||||
and approval.get("kind") == "tool_approval"
|
||||
):
|
||||
approval_pause = {
|
||||
"tool": data.get("tool") or "tool",
|
||||
"approval_id": approval.get("approval_id"),
|
||||
}
|
||||
# Scheduled tasks have no interactive surface that
|
||||
# can safely resume a one-use grant. Retire the
|
||||
# record immediately instead of leaving it pending
|
||||
# and report an explicit manual-action boundary.
|
||||
try:
|
||||
from src.tool_approvals import tool_approval_store
|
||||
tool_approval_store.consume(
|
||||
approval_pause["approval_id"],
|
||||
decision="deny",
|
||||
owner=task.owner,
|
||||
session_id=session_id,
|
||||
)
|
||||
except Exception:
|
||||
logger.debug(
|
||||
"Could not retire scheduled-task approval",
|
||||
exc_info=True,
|
||||
)
|
||||
break
|
||||
except (json.JSONDecodeError, KeyError):
|
||||
pass
|
||||
# Close the stream in this task on every exit, including the
|
||||
# approval-pause break, so the agent run's context state unwinds here.
|
||||
async with contextlib.aclosing(stream_agent_loop(
|
||||
endpoint_url=endpoint_url,
|
||||
model=model,
|
||||
messages=messages,
|
||||
max_rounds=_task_max_rounds,
|
||||
session_id=session_id,
|
||||
owner=task.owner,
|
||||
headers=headers,
|
||||
disabled_tools=disabled_tools,
|
||||
relevant_tools=relevant_tools,
|
||||
fallbacks=_task_fallbacks,
|
||||
workload="background",
|
||||
)) as agent_stream:
|
||||
async for event_str in agent_stream:
|
||||
if event_str.startswith("data: ") and not event_str.startswith("data: [DONE]"):
|
||||
try:
|
||||
data = json.loads(event_str[6:])
|
||||
# Capture text from all event types, not just delta
|
||||
if "delta" in data:
|
||||
if data.get("thinking"):
|
||||
continue
|
||||
if final_text_replaced:
|
||||
# A later answer supersedes the replacement,
|
||||
# as the completion gate treats it.
|
||||
full_text = ""
|
||||
final_text_replaced = False
|
||||
full_text += data["delta"]
|
||||
elif data.get("type") == "final_response":
|
||||
# The completion gate may present its sanitized
|
||||
# answer as one replacement instead of deltas.
|
||||
full_text = str(data.get("content") or "")
|
||||
final_text_replaced = True
|
||||
elif data.get("type") == "tool_output":
|
||||
# Tool results — capture summary so we have SOMETHING even
|
||||
# if the model never produces a final text response
|
||||
tool_summary = data.get("stdout") or data.get("output") or data.get("result") or ""
|
||||
if isinstance(tool_summary, str) and tool_summary.strip():
|
||||
tool_results.append(f"[{data.get('tool', '?')}] {tool_summary[:500]}")
|
||||
approval = data.get("ask_user")
|
||||
if (
|
||||
isinstance(approval, dict)
|
||||
and approval.get("kind") == "tool_approval"
|
||||
):
|
||||
approval_pause = {
|
||||
"tool": data.get("tool") or "tool",
|
||||
"approval_id": approval.get("approval_id"),
|
||||
}
|
||||
# Scheduled tasks have no interactive surface that
|
||||
# can safely resume a one-use grant. Retire the
|
||||
# record immediately instead of leaving it pending
|
||||
# and report an explicit manual-action boundary.
|
||||
try:
|
||||
from src.tool_approvals import tool_approval_store
|
||||
tool_approval_store.consume(
|
||||
approval_pause["approval_id"],
|
||||
decision="deny",
|
||||
owner=task.owner,
|
||||
session_id=session_id,
|
||||
)
|
||||
except Exception:
|
||||
logger.debug(
|
||||
"Could not retire scheduled-task approval",
|
||||
exc_info=True,
|
||||
)
|
||||
break
|
||||
except (json.JSONDecodeError, KeyError):
|
||||
pass
|
||||
|
||||
if approval_pause is not None:
|
||||
return (
|
||||
|
||||
Reference in New Issue
Block a user