Guardrail misclassifying refusals as tool failures
Cluster 113895. Data as of
Nothing open on GitHub. Fix merged in #114848 on 18 Sep 2026. Check it covers your case.
The short answer
- Known?
- Yes. 3 related items grouped here, 1 issue already closed.
- Fix event
- A fix was merged to main in PR #114848. Check it covers your case. No release data is tracked.
- Tracker
- The maintainers' tracker marks this
resolved, deciding PR #114848. That is separate from the fix event. - Still open
- Nothing open on GitHub.
These items address a bug where the guardrail system incorrectly counts its own refusals (such as read-dedup or harness refusals) as tool failures, causing unwarranted escalation to repeated_exact_failure_block. Fixes aim to distinguish genuine tool failures from guardrail-initiated refusals.
Open issues0
Open PRs0
Fixes on record
-
#114848 fix(guardrails): file-tool loop refusals no longer count as failures or escalate to repeated_exact_failure_block (#113895, salvage #113897)
files:
agent/display.py,agent/tool_guardrails.py,agent/tool_result_classification.py,tests/agent/test_tool_guardrails.py,tools/file_tools.py
1 PR closed without merging
From the close comments: merged via another change 1
-
#113897 fix(guardrails): a harness refusal is not a tool failure · merged via another change →
9814fb9,#114848,#113895,#113897
Closed issues · 1
- Closed#113895Guardrail counts its own read-dedup refusal as a tool failure, escalating to repeated_exact_failure_block over calls that never failed
How this cluster was grouped
-
Link evidence (
structural links throughout): every member is connected by structural evidence: shared changed files, or a closing/superseded reference. -
Pairwise link density
0.67: the share of item pairs that carry a link, counting a structural link in full and a relevance-only link at 0.36.1.00means every pair is linked.
Pick up a Lab task
Check whether a fix covers the open issues here, with your own agent. Bring back evidence for maintainers.
Open the LabAt a glance
- Merged PRs
- #114848
- Closed issues
- 1
- Attempts
- 1 PR closed without merging
- Fix link checked
- 18 Sep 2026