Configuration
Runtime images
Choose trusted repository tools without weakening provider credential isolation.
Repository tools and inference are separate
The agent container holds repository tools and scoped GitHub access. A separate trusted inference sidecar receives the selected provider credential and exposes authenticated inference only on localhost.
Agent image configuration can never replace the website-specified trusted sidecar image that protects provider root credentials.
Choose an image source
The CLI first checks its Docker daemon, pulls missing default or registry images, loads selected images into kind, and uses imagePullPolicy Never. Local images must already exist. Every reference must include a tag or digest.
- Default: the CLI pulls the website-specified role image from GHCR when it is missing, then loads it into kind.
- Local: build or pull a tagged image into the runner machine's Docker daemon before installation.
- Registry: use a tagged or digest-pinned reference and run docker login separately when authentication is required.
Custom image contract
Custom images must include Codex, Node.js, GitHub CLI, and /opt/orchestrator/claude-runner.mjs. Use only images you trust because ordinary agent containers receive scoped GitHub access and a per-job inference-proxy token.
Frequently asked questions
Can Issue Planner use a custom image?
No. Issue Planner uses the fixed trusted planning image.
When does a changed image take effect?
The connected runner prepares it once and later jobs use it. An already-running workload keeps its starting image.
Should registry credentials go in the image reference?
No. Run docker login separately and never embed credentials in an image name or repository setting.
Do agent containers receive provider API keys?
No. Root provider credentials stay in the trusted inference sidecar.