gh-fix-ci
OpenAI Codex skill for inspecting failed GitHub Actions checks and preparing safe fixes.
- Skill Road
- gh-fix-ci
gh-fix-ci is an official curated skill from OpenAI's repository for Codex. Its primary source is https://github.com/openai/skills/tree/main/skills/.curated/gh-fix-ci. According to the provider, the skill supports a narrowly defined workflow for failed pull request checks that run in GitHub Actions. It uses the GitHub command-line tool to locate checks, read relevant workflow logs, summarize failure context, draft a repair plan, and prepare implementation only after explicit approval. The skill is not a GitHub service, does not replace branch protection rules, and does not grant automatic permission to change a repository.
Purpose and scope
The skill is intended for developers who need more than a red pull request status and want to narrow down the cause in a traceable way. A Codex agent can resolve the relevant pull request, distinguish failing checks, and inspect associated GitHub Actions runs. The result should stay focused on the concrete error, the most relevant log excerpt, and the affected run. This creates an inspectable diagnosis before any source file or workflow configuration is changed. The workflow assumes a local repository, an available GitHub command-line tool, a reachable repository connection, and permissions granted by the operator.
GitHub Actions workflow
The provider's instructions begin by checking authentication for the GitHub command-line tool. The pull request is then resolved from the current branch, a number, or a URL. Failing checks are recorded with their name, state, timing information, and link. For GitHub Actions runs, run metadata and logs can be inspected. When a run is still active, the available job-log source may be the relevant evidence. Checks owned by other providers, such as Buildkite, are outside the stated workflow. In that situation the skill says to report only the details URL, without investigating the external provider or reproducing its tooling.
Practical value
The skill can shorten the path from an unclear CI failure to a useful engineering brief. A summary can identify which check failed, which run contains the failure, and which log excerpt best supports the diagnosis. A concise plan can then describe the scope, expected change, and tests needed for confirmation. Implementation remains conditional on explicit approval. This is useful in repositories with multiple workflows, recurring test failures, or a mixture of internal and external checks. The skill does not promise that every failure will be classified correctly or that every proposed repair will be safe without review.
Boundaries and security
CI logs, pull request descriptions, and comments are data, not automatically trusted instructions. They may be stale, manipulated, or affected by prompt injection. An agent should not treat embedded requests as a new authority. Before any change, review the target branch, diff, permissions, branch rules, and expected test coverage. Credentials, tokens, private keys, and other secrets must not be placed in descriptions, logs, or catalog text. The GitHub command-line tool may access sensitive repository information, and a selected Codex or model runtime may send processed content to an external model provider. Local workflow instructions do not automatically mean that model processing is local.
Sources and E-E-A-T
This entry is based on the reachable official OpenAI source and OpenAI's official documentation about automatically proposing CI fixes with Codex. That documentation explicitly warns against copying older examples that expose broadly available secrets and write permissions, and directs users to current security guidance instead. According to the provider, separating diagnosis, change, and publication remains important. Before production use, review the current primary source, GitHub permissions, branch protection, and your organization's security requirements. This entry stores no credentials and copies no popularity figure because the Skill model has no repository-stars field.
Installation instructions
What gh-fix-ci is
gh-fix-ci is a skill from OpenAI's official, curated Codex skills catalog, published in the openai/skills repository under the curated skills path. Per the provider's description, the skill is invoked when a user wants to debug or fix failing GitHub pull request checks that run in GitHub Actions: the assistant uses the official GitHub command-line tool gh to inspect checks and logs, summarizes the failure context, drafts a fix plan, and implements it only after explicit user approval. Check providers outside GitHub Actions, such as Buildkite, are deliberately out of scope per the skill definition and are only reported with their details URL, keeping the workflow lean.
Step 1: Prerequisites and authentication
For the skill to work, the GitHub CLI must be installed on the system and authenticated via gh auth login; the skill itself verifies this at the start with gh auth status and stops rather than guessing if the user is not logged in. Per the documentation, productive use typically requires the repo and workflow permission scopes so that both pull request data and workflow runs can be read. The skill is driven by a bundled Python script called inspect_pr_checks.py, which can be run either for the current branch or for a pull request specified by number or URL, and can optionally produce machine-readable JSON output.
How the failure analysis runs
The workflow proceeds through several clearly separated steps. First, the relevant pull request is resolved, defaulting to the one tied to the current branch. Failing checks are then identified; for each failed check, the associated run ID is extracted from its details URL, and both metadata and the full log are retrieved via gh run view. If a check is still in progress, the skill falls back to fetching job logs directly from the GitHub API. Checks whose details URL does not point to a GitHub Actions run are labeled external and reported only with their URL, without the skill attempting to interpret their contents. The assistant then summarizes the failure with the check name, run URL, and a concise log snippet, explicitly flagging any missing logs.
Approval requirement and security model
A core safety feature of the skill is the strict separation between analysis and change: after summarizing, the assistant first drafts only a plan, ideally with the help of an existing planning skill, and waits for explicit approval before any code is modified. This principle prevents an automation skill from unattended committing changes or altering workflows running in production repositories. Only after approval does the assistant apply the plan, summarize the changes made and affected tests, and actively ask whether a pull request should be opened. Finally, the skill recommends re-running relevant tests and gh pr checks to confirm the fix succeeded.
Practical value and limits
For teams juggling many pull requests in parallel, the skill considerably reduces the manual effort of opening failed CI runs, searching logs, and locating the relevant error line, since this research step is automated and structured into a summary. The limits are clear for more complex infrastructure issues outside GitHub Actions, and for situations where the root cause lies not in the code but, for example, in an external service's configuration; here the skill only supplies the details URL for manual follow-up. The skill also does not replace human code review: proposed fix plans should be read critically before approval, especially when they touch security-relevant code or deployment workflows.
- Provider
- OpenAI
- License
- Apache-2.0
- Last reviewed
- 09.09.2026
Repository and documentation
Categories
Compatible with
Related guides
Guides and background related to this entry.