Sandbox Image Topology

Fullsend agents run inside sandboxed containers. Two images exist in a parent-child hierarchy; which image an agent uses depends on whether it needs a compiled-language toolchain.

ghcr.io/nvidia/openshell-community/sandboxes/base   (upstream)
  +-- fullsend-sandbox                                (base sandbox)
        +-- fullsend-code                             (extends base with Go)

Image	Agents	Run frequency	Key additions over parent
fullsend-sandbox	triage, prioritize, retro	High (most agent runs)	Claude Code, pi (pinned; opt-in runtime, #6464), Codex (pinned; opt-in runtime, #6920), ripgrep + fd (pi grep/find backends), jq, gitleaks, acli, pre-commit, gitlint, tirith
fullsend-code	code, fix, review	Lower (code/fix are the least-run agents; review runs per-PR)	Go toolchain, scan-secrets, gopls, lychee

Harness definitions that map agents to images live in the fullsend-ai/agents repo (the image: field in each harness YAML). The GitLab scaffold (internal/scaffold/fullsend-repo-gitlab/) uses the same execution model: a single generic agent template (fullsend-agent.yml) calls fullsend run "${STAGE}", parameterized by the $STAGE pipeline variable set by API-triggered pipelines dispatched by the cron poller. fullsend run resolves the harness and creates the sandbox container from the harness image: field. Image Containerfiles live in images/sandbox/ and images/code/. The CI build pipeline is .github/workflows/sandbox-images.yml.

When reviewing CI changes: If a PR modifies image pulling, caching, or pre-warming logic in action.yml, consider which agent types are affected. Changes that only benefit fullsend-code have a smaller blast radius (fewer agent runs) than changes to fullsend-sandbox. A cache or pull optimization may not be worth the complexity if it only helps the least-frequently-run agents.