feat(agents): support cloud execution

[AbysmalBiscuit]

Enable cloud agents to take an issue through implementation, verification, and a draft PR using devkit and Lev's workflow preferences. Start with Claude Code cloud, keeping the shared instructions and scripts reusable by Codex later.

Local user config, helper scripts, and installed skills do not automatically reach a cloud VM. Cloud setup must supply them, and context recovery must preserve ongoing work instead of restarting the workflow.

Scope

• Keep the shared skill under .agents/skills/cloud and Python scripts under .agents/scripts. Symlink .claude/skills to ../.agents/skills; keep harness-specific lifecycle configuration separate.
• Provision devkit binaries and command links, register the devkit and Superpowers plugins, and generate devkit.local.toml with command/write enforcement, commit tasks, and the PR body template. Bundle the patch-commit helper in the repository rather than reference a personal home directory.
• Before implementing, find the issue's existing spec, plan, and progress. Execute a ready plan; use judgment about brainstorming when requirements or design remain unresolved. Resolve routine choices autonomously and ask when human input is needed.
• Use full workflow instructions at startup and a short recovery brief after resume or compaction. Recover the active task, reconcile running work, and continue from the first unfinished action while preserving pending questions and approvals.
• Push completed work and open a draft PR. Include issue references and actual model/harness attribution, and leave ## TL;DR (human written) empty.

Acceptance criteria

• A fresh Claude cloud session has working devkit commands, plugins, enforcement, and generated config without local-machine dependencies.
• Cached setup, session resume, and compaction preserve the intended behavior without repeating completed work or reinstalling tools on every hook.
• Local sessions do not generate cloud preferences; generated files stay out of commits.
• Tests beside the scripts use *.test.py names and cover config generation, paths containing spaces, hook routing, and patch commits preserving unrelated staging.
• An actual cloud run completes a representative issue and opens a correctly formatted draft PR. Document setup, verification, and the remaining Codex adapter work.

Starting point

A local prototype exists in .agents/scripts, .agents/skills/cloud, and .claude/settings.json. Its Python tests pass locally, including a patch commit through devkit. It has not been validated in Claude's cloud environment, and the prototype files are not yet committed.

References

• Claude cloud environments
• Claude lifecycle hooks