Skip to content

AUTHENTICATION.md

Latest commit

 

History

History
95 lines (71 loc) · 3.53 KB

AUTHENTICATION.md

File metadata and controls

95 lines (71 loc) · 3.53 KB
Raw

Authentication and agent selection

All reusable AI workflows call one shared run-ai-agent adapter. Set GH_ACTION_AI_AGENT to select the implementation for every inheriting repository:

Value Action Credential
claude (default) anthropics/claude-code-action GH_ACTION_AI_API_KEY
codex openai/codex-action OPENAI_API_KEY

The credentials stay separate. The adapter validates and passes only the secret for the selected agent. Changing GH_ACTION_AI_AGENT is therefore the only switch required by callers that use secrets: inherit.

Required configuration

Configure these as GitHub organization or repository variables and secrets:

Name Kind Required when Purpose
GH_ACTION_AI_AGENT Variable Optional claude or codex; defaults to claude
GH_ACTION_AI_API_KEY Secret Claude selected Anthropic API key
OPENAI_API_KEY Secret Codex selected OpenAI API key

An explicit-secret caller must forward both credentials so changing the selector does not require another workflow edit:

jobs:
  run:
    uses: dryvist/ai-workflows/.github/workflows/<name>.yml@main
    secrets:
      GH_ACTION_AI_API_KEY: ${{ secrets.GH_ACTION_AI_API_KEY }}
      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Callers that use secrets: inherit need no change after both organization secrets are available to the repository.

Optional Claude configuration

Existing Claude variables remain supported:

Name Purpose
GH_ACTION_AI_BASE_URL Anthropic-compatible endpoint; empty uses the upstream default
GH_ACTION_AI_MODEL Global Claude model override
GH_ACTION_AI_MODEL_CODE Code-task override
GH_ACTION_AI_MODEL_ISSUES Issue-task override
GH_ACTION_AI_MODEL_PLAN Planning-task override
GH_ACTION_AI_MODEL_REVIEW Review-task override

No model identifier is hard-coded. If no applicable model variable is set, the upstream action chooses its default.

Optional Codex configuration

Name Maps to
GH_ACTION_AI_CODEX_RESPONSES_API_ENDPOINT responses-api-endpoint
GH_ACTION_AI_CODEX_MODEL model
GH_ACTION_AI_CODEX_EFFORT effort
GH_ACTION_AI_CODEX_VERSION codex-version

Leave these variables empty to use the Codex Action defaults. This avoids coupling inherited workflows to model or CLI identifiers that change over time.

Security boundary

The adapter runs Codex with drop-sudo and an explicit read-only or workspace permission profile. AI jobs receive read-only GitHub permissions. Workflows that publish comments, labels, commits, or pull requests perform that operation in a fresh deterministic publisher job with only the minimum GitHub permission and, where required, the existing GitHub App identity.

Do not give a model an App token or a write-capable GITHUB_TOKEN. Treat issue bodies, pull-request descriptions, comments, and repository files as untrusted prompt input.

Verify both agents

Run the same dogfood workflow once per selector value:

gh variable set GH_ACTION_AI_AGENT --org dryvist -b claude
gh workflow run dogfood-ci.yml --repo dryvist/ai-workflows
gh run watch --repo dryvist/ai-workflows

gh variable set GH_ACTION_AI_AGENT --org dryvist -b codex
gh workflow run dogfood-ci.yml --repo dryvist/ai-workflows
gh run watch --repo dryvist/ai-workflows

A successful run for each value proves that selection, credential forwarding, prompt rendering, and the provider-specific adapter path are wired correctly.