Repository navigation
Conversation
Raises the default max tokens so thinking tokens don't truncate the verdict.
Also report a clear error when the check runs out of output tokens.
zentron
marked this pull request as ready for review
October 11, 2026 12:26
APErebus
approved these changes
Oct 11, 2026
zentron
deleted the
roberterez/sf-3843-update-the-claude-code-steps-injection-check-to-claude-haiku
branch
October 13, 2026 01:23
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This change moves the Claude Code step's prompt injection check to Claude Haiku 5.5. The default model goes from
claude-haiku-4-5toclaude-haiku-5-5.Why max tokens goes from 1024 to 2048
Haiku 4.5 had no thinking. The output was only the small JSON verdict, so 1024 was plenty.
Haiku 5.5 has thinking on by default at
mediumeffort. Thinking tokens count towardmax_tokens. They use up the budget before the verdict is written.If the budget runs out, the response has no verdict. The check then throws and the step fails unless the fail-open option is set.
A higher cap costs nothing extra. You pay only for the tokens used, at $0.50 per million output tokens. The cap only stops a long thought from cutting off the verdict.
2048 is double the old value. It still limits runaway thinking. It is an estimate until we measure real usage.
Token usage logging
The check now logs its input and output tokens at verbose level. This lets us compare real output tokens against the 2048 cap.
If the model runs out of output tokens, the step now fails with a clear message. The old message was "returned no verdict content".
Not in this change
The step still sends no effort setting. Haiku 5.5 uses its default effort of
medium.This change doesn't need a server change. The matching UI text and docs updates are in separate PRs:
Linear: SF-3843
🤖 Generated with Claude Code