Preflight Checklist
What's Wrong?
Describe the bug
Governs sessions over current user corrections, causing repeating classifier/warning loops
Sessions repeatedly act on stale persisted state instead of the customer's
current corrections: outdated credentials, superseded commands, and recorded
methods reloaded from access settings, auto-memory, and configuration across
updates. The product's own safety/permission classifiers then fire on the
stale action, producing a warning and dispute cycle that consumes the
session. After I correct the method, a later session falls back to the stale
version and the identical loop restarts. Related, already-tracked instances
of stale state governing behavior: #6499, #41259, #29002, #60742, #13785.
This issue is for the composite loop: stale state -> classifier fire ->
dispute -> correction -> stale reinjection -> repeat, which none of those
issues covers end to end.
What Should Happen?
Expected behavior
Corrections are durable. Persisted credentials, commands, and methods are
invalidated or flagged stale on update and on user correction. Classifier
events triggered by the product's own reloaded state do not consume the
customer's usage allowance.
Error Messages/Logs
Steps to Reproduce
Steps to reproduce
- Establish a working method/credential in a project; let sessions record it
(memory/settings).
- Change the method or credential; correct the assistant in-session.
- Update the app and/or open fresh sessions over following days.
- Observe the assistant fall back to the stale recorded method, trip
permission/safety warnings on it, and re-litigate what was already
corrected.
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
No response
Claude Code Version
1.26832.0
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Other
Additional Information
https://gist.github.com/53f85e6d48eaec4a661209fcf3f49ae6.git
Preflight Checklist
What's Wrong?
Describe the bug
Governs sessions over current user corrections, causing repeating classifier/warning loops
Sessions repeatedly act on stale persisted state instead of the customer's
current corrections: outdated credentials, superseded commands, and recorded
methods reloaded from access settings, auto-memory, and configuration across
updates. The product's own safety/permission classifiers then fire on the
stale action, producing a warning and dispute cycle that consumes the
session. After I correct the method, a later session falls back to the stale
version and the identical loop restarts. Related, already-tracked instances
of stale state governing behavior: #6499, #41259, #29002, #60742, #13785.
This issue is for the composite loop: stale state -> classifier fire ->
dispute -> correction -> stale reinjection -> repeat, which none of those
issues covers end to end.
What Should Happen?
Expected behavior
Corrections are durable. Persisted credentials, commands, and methods are
invalidated or flagged stale on update and on user correction. Classifier
events triggered by the product's own reloaded state do not consume the
customer's usage allowance.
Error Messages/Logs
Steps to Reproduce
Steps to reproduce
(memory/settings).
permission/safety warnings on it, and re-litigate what was already
corrected.
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
No response
Claude Code Version
1.26832.0
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Other
Additional Information
https://gist.github.com/53f85e6d48eaec4a661209fcf3f49ae6.git