Review and change agents

On this page

Review the agent's changes before you commit or merge them. Use Git, your IDE, and the project's checks. You can ask another agent to review the work.

Reviewing an agent's work before keeping it

  1. Keep a recovery point outside the capsule. Commit your current work, then push it to a remote or keep an independent backup. A local commit alone is not protected from processes that can write .git. A linked Git worktree shares repository data and is not an independent backup. Keep the backup or remote outside the agent's write permissions.

  2. Inspect every modified and untracked file. Use your IDE's Git diff tool or run these checks in the terminal:

    git status --short
    git diff
    git diff --staged
    

    Watch for unintended edits to lockfiles, CI definitions, .devcapsule/, or skipped tests, plus stray scratch files left in the tree.

  3. Run the project's checks yourself. Execute your unit tests, linter, and type checker in the capsule terminal rather than relying solely on the agent's summary—especially for newly added tests, which can mirror a flawed assumption in the implementation.

  4. Discard or refine cleanly. Revert unwanted hunks with git restore -p (and remove untracked scratch files after previewing git clean -nd), or ask a second agent to review the diff before you commit and push.

Running multiple agents or switching on a checkout

Claude Code (claude), OpenAI Codex (codex), and the Antigravity CLI (antigravity) install into separate /opt/... prefixes and keep independent persistent state directories (~/.claude, ~/.codex, ~/.gemini). You can keep more than one enabled on the same checkout—for instance, implementing a feature with one agent and asking another to critique the branch.

Manage your checkout's agents from a host terminal in the project directory (or request add, omit, and restore from the running session's web console). The launcher records accepted requests immediately. The running environment changes only after you stop the capsule and launch it again:

# Try an agent for a single launch without saving it to the checkout
devcapsule project run --with codex-agent

# Add an agent to this checkout's saved selection (shared manifest unchanged)
devcapsule project config add claude-code-agent

# Remove a checkout-added agent, or omit an agent offered by the project
devcapsule project config omit antigravity-agent

# Re-enable a project-offered agent that you previously omitted
devcapsule project config restore antigravity-agent

(Capability names and unambiguous component aliases—claude-code, codex, antigravity-cli—both work on add, omit, and restore. Required capabilities declared with require in .devcapsule/devcapsule.toml cannot be omitted.)

Use restore only for project offers. To bring back a checkout addition that you removed, use add again. If an agent has an explicit secret binding, remove it with config unbind before omitting the agent; see Sign in and defaults.

Your sign-in and settings stay intact when you switch

Running devcapsule project config omit <agent> removes that agent from the next environment selection. It does not delete its saved state or old images. Its persistent directory under your configuration home keeps settings and session files for later use. Omitting an agent is not a way to revoke its provider credentials or hide files from another agent.

Read the source on GitHub Back to top ↑