Quality seems to have degraded after upgrading from beta to v1 (auto pr)

[gobenpark]

After upgrading from beta to v1, there appears to be mismatches with inline comments where comments are not being placed on the correct lines but are slightly off. The only change made was adding more detailed content to the system-prompt in claude_args.
Additionally, the following tool configurations have been added:

--allowedTools "mcp__github_inline_comment__create_inline_comment,mcp__github__create_pending_pull_request_review,mcp__github__add_comment_to_pending_review,mcp__github__submit_pending_pull_request_review,mcp__github__get_pull_request_diff,mcp__github__get_pull_request_files,mcp__github__get_pull_request"
--disallowedTools "mcp__github__create_and_submit_pull_request_review"
--model bedrock-claude-sonnet-4-20250514-v1

[0xBunzy]

I have Claude Code Max plan. Does claude-code-action@v1 support OAuth tokens from Max plans (CLAUDE_CODE_OAUTH_TOKEN), or does it require a separate paid API key from console.anthropic.com?

[gobenpark]

@0xBunzy I believe both are supported.

[gobenpark]

Adding mcp__github__get_file_contents to allowTools improved performance a bit. I think it's because the contents get added to the context. When this MCP doesn't work properly, inline comments aren't placed on the correct lines. In beta, comments seem to be written on the appropriate lines - are there any differences in the prompts between beta and the current version?

[chrisxaustin]

The quality is pretty terrible now, while also consuming far more tokens.

For fairly minor PRs claude is using ~30k tokens and giving me a lot of hallucinations.