[Bug] Auto mode classifier returns no verdict for project-local write operations
Bug Description
Claude Code 2.1.278 on macOS, using a Claude Pro login.
/status reports:
Auto mode server: Enabled
Normal project-local Write operations repeatedly fail with:
The server-side auto mode classifier gave no verdict (it skipped this action), so auto mode cannot determine the safety of Write.
The issue reproduces across fresh Claude Code sessions and across different
ordinary project-local targets, including both documentation and source files:
claude-platform/docs/general_platform_block_d_v1.mdclaude-platform/lib/repo_intel.pyclaude-platform/docs/repo_intelligence_recipes.md
Read-only operations continue to work normally.
claude auto-mode config shows that local file operations within project scope
are allowed, the current repository is trusted, and there is no apparent policy
rule that explains these writes being blocked.
The issue also persists after starting fresh sessions.
After identifying the problem, I did not use Bash, heredocs, Python, or another
write mechanism to bypass the failed classifier decision.
Expected behavior
The server-side auto mode classifier returns an allow or deny verdict for
ordinary project-local Write operations.
Actual behavior
The classifier repeatedly produces no verdict, causing Auto mode to fail closed
and preventing ordinary file writes.
This is not shown as an ordinary classifier denial with a policy reason.
Environment
- Platform: macOS / Darwin
- Terminal: Apple Terminal
- Claude Code: 2.1.278
- Login: Claude Pro
- Auto mode server: Enabled
- Feedback ID:
2d4845ff-cdde-4235-b282-7a120ea0a7d4
Showing cached comments. Read the full discussion on GitHub ↗
3 Comments
Claude Code 2.1.220
full plugin
→ project-local Write PASS
Claude Code 2.1.278
same full plugin
→ Write NO VERDICT
2.1.278
same plugin, only hooks/hooks.json description shortened
→ Write PASS
Additional isolation:
Claude Code 2.1.278, server-side Auto mode.
Same repo/plugin:
permission/bypass behavior and prior classifier failures:
first ordinary Edit can return NO VERDICT
The error explicitly says the server-side classifier skipped the action
and retrying will not change it.
This remains after fixing a separate local oversized-plugin-metadata defect.
Update: I now have a deterministic reproducer and a minimized boundary.
Environment:
The failure requires all three conditions together:
claude -ppath does not reproduce).Result:
"The server-side auto mode classifier gave no verdict..."
"Classifier unavailable"
One-variable controls pass:
claude -ppath -> PASSA manifest-only plugin is sufficient to reproduce, so hooks, skills and plugin
metadata size are not required for the failure.
The client treats files inside the loaded plugin root as sensitive; those
mutations reach the Auto classifier instead of the normal in-cwd auto-accept
path. For the failing interactive topology the classifier is unavailable /
skips the action and the client correctly fails closed.
I implemented a local structural workaround:
when self-developing the plugin, load an immutable plugin snapshot from outside
the project cwd while continuing to edit the original source tree.
Using the same interactive reproducer, same Auto mode, same cwd, same prompt
and same source mutation target, changing only the loaded-plugin topology:
The local workaround avoids the failing topology but does not fix the underlying
Claude Code/server-side classifier behavior. The original reproducer still fails.
I can provide a sanitized stdlib-only minimal reproducer if useful.