[Bug] `run-gemini-cli@v0` fetches wrong PR even when correct repo and PR number provided

[eshwarvijay]

TL;DR

Description

The google-github-actions/run-gemini-cli@v0 action is failing to fetch pull request data for the correct repository and PR number, instead trying fallback PRs (1–6) and returning 404 errors.

Reproduction Steps

1. Use Gemini workflow on pull_request event, triggered for PR # in amber-ml/amber-ml.
2. Workflow log shows:
   Then Gemini tries PRs 1 through 6 and finally says unable to proceed with the code review.

Expected behavior

Gemini should fetch and review the exact PR it was triggered for—no arbitrary probing of PRs 1–6.

Observed behavior

Gemini disregards correct PULL_REQUEST_NUMBER and REPOSITORY, instead attempting PRs 1–6, all of which fail with 404.

Action YAML

jobs:
gemini-code-review:
 runs-on: ...
 env:
   PULL_REQUEST_NUMBER: ${{ github.event.number }}
   REPOSITORY: ${{ github.repository }}
 steps:
   - uses: google-github-actions/run-gemini-cli@v0
     with:
       gemini_cli_version: ${{ vars.GEMINI_CLI_VERSION }}
       gemini_api_key: ${{ secrets.GEMINI_API }}
       use_vertex_ai: false
       use_gemini_code_assist: false
       settings: |
         { ... }
       prompt: |
         Review this PR...

Log output

Warning: Unexpected input(s) 'google_api_key', 'gemini_debug'
Error executing tool get_pull_request: ... pulls/<actual-number>: 404 Not Found
... retries PRs 1–6 with the same error

Additional information

No response

[gemini-cli]

🤖 Hi @eshwarvijay, I've received your request, and I'm working on it now! You can track my progress in the logs for more details.

[michael-airspace]

Neat idea, but gemini is pulling random old PRs to review which don't match current PR.

[hassib-ashouri]

How is this working on this repo; or is not working here too??

[tleonhardt]

We have seen this problem as well. It appears to work correctly when a PR is first created. However, sometimes when a manual re-review is requested by posting a @gemini-cli /review comment in the PR, sometimes it works, but most of the time it goes out and reviews some old PR that has already been merged.

Here is a specific example where we saw this recently in the cmd2 open-source Python library repo:

• Open PR: Documentation review python-cmd2/cmd2#1511
• Incorrect 4-year old PR that was reviewed when manual re-review was requested: Updated async_alert() to account for self.prompt not matching Readline's current prompt python-cmd2/cmd2#1145 (review)