claude-code - 💡(How to fix) Fix [Bug] Differential refusal behavior between subprocess and main turn in same session [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#52990Fetched 2026-04-25 06:15:21
View on GitHub
Comments
0
Participants
1
Timeline
3
Reactions
0
Participants
Timeline (top)
labeled ×3

Root Cause

The two paths ran in the same logical session moments apart. Subprocess performed the active work; main turn was a passive report to the user. Filing because the differential seems worth a data point — the subprocess path presumably shares classifier layers with the main turn, yet one fired and the other did not, on material strictly less sensitive than the work that preceded it.

Code Example

[]
RAW_BUFFERClick to expand / collapse

Bug Description Observed differential model behavior on 2026-04-24 at 17:44:25Z during a routine setup task. Same session, same account context:

  • Path A — claude --chrome -p subprocess: completed a multi-step browser navigation flow (navigate, form fill, button click) without refusal.
  • Path B — main chat turn, 7-word status update to the user ("the account is set up and I'm logged in"): refused.

The two paths ran in the same logical session moments apart. Subprocess performed the active work; main turn was a passive report to the user. Filing because the differential seems worth a data point — the subprocess path presumably shares classifier layers with the main turn, yet one fired and the other did not, on material strictly less sensitive than the work that preceded it.

Client: Claude Code 2.1.119. No expectation of individual response — logging in case it helps calibrate aggregate behavior.

Environment Info

  • Platform: linux
  • Terminal: xterm-256color
  • Version: 2.1.119
  • Feedback ID: 70159513-97c6-49d4-9b63-1a7e30b7d1f4

Errors

[]

extent analysis

TL;DR

Investigate the classifier layers shared between the subprocess and main turn to identify potential discrepancies in handling sensitive data.

Guidance

  • Review the code for differences in how the subprocess and main turn handle user input and sensitive data to identify potential causes for the differential behavior.
  • Verify that the classifier layers used in both paths are identical and properly configured to handle the specific use case.
  • Check for any recent changes or updates to the classifier layers that may have introduced the discrepancy.
  • Test the behavior with different input scenarios to see if the issue is reproducible and to gather more data points.

Example

No code snippet can be provided without more context, but a potential area of investigation could be the input processing and sanitization steps for both the subprocess and main turn.

Notes

The issue may be related to the specific configuration or version of the classifier layers used, and further investigation is needed to determine the root cause. The fact that the subprocess and main turn share classifier layers but exhibit different behavior suggests a potential discrepancy in how they handle sensitive data.

Recommendation

Apply workaround: Investigate and refine the classifier layers to ensure consistent handling of sensitive data, as the current behavior may indicate a calibration issue.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING