erabo/odysseus
1
0
Fork 0
mirror of https://github.com/pewdiepie-archdaemon/odysseus.git synced 2026-08-05 02:45:28 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

fix: apply fake-success guard to all models, not just qwen finetune

Browse source 
kimi-k3 (the default assistant model) repeatedly confabulated confident
sent/done confirmations for destructive/send-style requests where zero
tools actually ran. The existing guard already did the right thing but
was gated behind _ody_qwen_finetune_model, so it never fired for the
model actually in use.
This commit is contained in:
romanreiki 2026-07-30 12:37:41 +00:00
parent a1aa0c082a
commit fc912df640
1 changed files with 11 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
src/agent_loop.py
View file

@ -5156,12 +5156,17 @@ async def stream_agent_loop(
full_response = strip_tool_blocks(full_response).strip()
if _ody_qwen_finetune_model:
full_response = _normalize_ody_qwen_text_artifacts(full_response)
if (
not tool_events
and _looks_like_destructive_request(_last_user)
and _looks_like_success_claim(full_response)
):
full_response = "I couldn't make that change because no matching tool action completed."
# Fake-success guard: applies to ALL models, not just the qwen finetune.
# If the user asked for a send/delete/reply-style action and zero tools
# actually ran this turn, don't let the model's prose claim it happened —
# kimi-k3 and others have repeatedly confabulated confident "sent"/"done"
# confirmations for actions that never executed (see Odysseus memory log).
if (
not tool_events
and _looks_like_destructive_request(_last_user)
and _looks_like_success_claim(full_response)
):
full_response = "I couldn't make that change because no matching tool action completed."
_response_before_tool_summary = full_response
if tool_events:
for _ev in reversed(tool_events):