claude auto-mode critique returns "No critique was generated" for a large autoMode block; works with a small one

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

claude auto-mode critique always fails on my configuration, printing:

Analyzing your auto mode rules…

No critique was generated. Please try again.

Retrying does not help. The configuration itself is valid and working: auto mode runs normally, and claude auto-mode config prints the effective config correctly.

The CLI distinguishes three outcomes here, and it matters which one this is:

| Message | Meaning |
|---|---|
| Failed to analyze rules: <error> | the model query itself failed |
| No custom auto mode rules found. | there was nothing to review |
| No critique was generated. Please try again. | what I get |

I confirmed the first is reachable and genuinely distinct: pointing CLAUDE_CONFIG_DIR at a directory with no credentials produces Failed to analyze rules: Could not resolve authentication method….

So in my case the model call is completing and returning empty text, rather than erroring or finding nothing to critique.

The failure is size-dependent. With my real autoMode block (15 custom environment entries plus one custom soft_deny rule) it fails every time. Replacing autoMode with a single small rule, leaving the rest of settings.json untouched, makes it succeed immediately and return a long, detailed critique. Swapping back makes it fail again.

What Should Happen?

claude auto-mode critique should either return a critique, or report an error that says what actually went wrong.

If the underlying cause is a truncated or empty model response, that case is worth distinguishing from a generic "please try again". The message as written implies a transient failure, whereas in my case it is deterministic and will never succeed at my configuration size.

For comparison, the /auto-mode-setup path already inspects stop_reason and reports truncation explicitly. The critique path appears to fall through to this generic message whenever the extracted text is empty.

Error Messages/Logs

$ claude auto-mode critique
Analyzing your auto mode rules…

No critique was generated. Please try again.


For contrast, the genuine query-failure path (isolated config dir with no credentials), showing these are different code paths:

$ CLAUDE_CONFIG_DIR=/tmp/cfgtest claude auto-mode critique
Analyzing your auto mode rules…

Failed to analyze rules: Could not resolve authentication method. Expected one of apiKey, authToken, credentials, config, or profile to be set. Or for one of the "X-Api-Key" or "Authorization" headers to be explicitly omitted


No debug output is available for this command - see Additional Information.

Steps to Reproduce

1. Populate autoMode in ~/.claude/settings.json with a substantial block. Mine is 15 custom environment entries plus one custom soft_deny rule, roughly 13,000 characters of custom rule text. The effective config printed by claude auto-mode config (defaults plus custom) is roughly 75,000 characters.
2. Run claude auto-mode critique.
3. Observe No critique was generated. Please try again. Retry - same result.
4. Now replace the whole autoMode block with a single small rule, leaving everything else in settings.json unchanged:

{
  "autoMode": {
    "soft_deny": [
      "$defaults",
      "Never delete the build output directory without being asked."
    ]
  }
}

1. Run claude auto-mode critique again. It succeeds and returns a long, detailed critique.
2. Restore the original block. It fails again immediately.

Variables ruled out:

• Not nesting - fails both from a shell inside a running Claude Code session and from a standalone terminal.
• Not the model - fails with the default and with --model claude-sonnet-5.
• Not the custom rules - it also failed when environment was the only populated section and there were no custom allow/soft_deny/hard_deny rules at all.
• Not transient - about six attempts over several hours.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.233 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

The critique prompt embeds the entire classifier system prompt together with the user's custom rules. The practical effect is that the more environment context you add - which the documentation actively encourages, saying "the more specific context you give, the better the classifier can distinguish routine internal operations from exfiltration attempts" - the more likely critique is to fail.

The feature therefore stops working at exactly the configuration size where having your rules reviewed is most valuable, and it fails in a way that reads as transient rather than permanent.

Configuration contents are omitted deliberately, as they describe private infrastructure. I can supply sizes, structure, or a synthetic block of equivalent size on request.