cli

Trusted Compute workspaces

This is an opt-in Agent Host profile for authorized operator work. Add "access_mode": "trusted" to a registered execution profile, retain its exact account/model/effort, and set network to enabled. Omission or restricted retains the original file-task behavior. The installed runner must include this contract before a trusted task can be admitted.

Files and outputs

Include the project in read_roots and write_roots. Add each approved shared directory to both arrays. Roots must exist on the execution host and be canonical; no missing mount is replaced with a new empty directory. Descendants inherit access. Normal repository links are accepted; the native command sandbox still restricts resolved targets to approved paths. Codex receives explicit permission for .git, .codex, .agents and AGENTS.md entries inside approved roots, because a parent write grant otherwise leaves native metadata protection in place. Linked worktrees may need their external Git metadata directory added to the declared roots too.

outputs.deliverables remains a list of output locations, each covered by a write root. It no longer has to equal every write root for trusted tasks. Output manifests are checked against deliverables, including real paths and regular files, before hashes are recorded. The project and deliverables must remain disjoint from outputs.runtime in the Compute definition.

Trusted write roots may include the RUDI runtime tree, which also contains accounts, job control files and receipts. This explicitly gives up protection against a malicious job modifying that tree. Hashes are checked before execution; they are not an immutable custody guarantee against a trusted operator with runtime access. Native account selection, version pins, job deadlines, process cleanup and failure receipts remain in effect. Do not use trusted tasks to run untrusted instructions or treat MCP-side effects as filesystem-sandboxed.

Tools and native configuration

The host’s ~/.rudi/bins/rudi-router must exist and be executable. Trusted tasks add that one server; the router discovers its installed tools dynamically. They do not maintain a per-stack tool list or copy secrets into profiles. The native process gets the installed RUDI and user executable directories in PATH, but does not inherit ambient API keys or change the selected native account.

Codex loads configuration from the selected account home, retains native MCP, app and plugin support, enables web search, and pins the OpenAI provider. An optional rudi-trusted.config.toml in that home is loaded as a native profile. For shared host integrations, an operator can link that file to the host’s existing Codex config. This references configuration without copying credentials, replacing the account home, or adding config.toml to an otherwise restricted account. Restricted tasks never select this named profile. Existing files must be inspected and preserved; do not replace a different profile silently.

Trusted Codex tasks discover enabled MCP servers through the selected native configuration and set their default tool approval mode, plus the native app fallback, to approve. Otherwise unattended approval_policy=never can deny even a read-only router call that requests a prompt. Explicit per-tool, per-app and managed restrictions remain in effect. Discovery is bounded and fails closed on errors or server names that cannot safely form native override keys; configuration contents and secrets are never copied into arguments or errors. This does not authorize external actions beyond the approved task.

Claude exposes its default native tool set and configured MCPs, adds the RUDI router, and uses noninteractive native tool permissions inside the existing OS filesystem sandbox. Trusted mode additionally grants the execution host’s ~/.rudi read/write access for router runtimes, secret mediation, caches and outputs, plus read access to ~/.local, /opt/homebrew and /usr/local for installed executables. These runtime grants apply even if a declared shared runtime root is a mount of another computer. Native account/temp/system runtime grants also remain. Hooks remain disabled. Native user/project configuration is available; only connections actually configured and authenticated for that identity can be used. Do not share .claude.json across identities: it contains account state. Configure additional MCPs with the native per-account management surface. Cloud connections and desktop-only plugins depend on native CLI and account support; file access alone does not provision them.

Native MCP servers can have service authority beyond command filesystem roots. Each tool’s own authorization and the task’s instructions continue to govern external actions. Availability does not authorize sending messages, publishing, deploying, deleting records or changing permissions.

Host selection and rollout

Choose the execution host separately from the native identity. Do not silently move jobs between computers. Shared names resolve as follows on the verified two-Mac setup (substitute the actual host user):

Workspace RUDI Mac Admin Mac
Clients records /Volumes/Clients ~/Clients
Shared RUDI /Volumes/RUDI ~/RUDI
RUDI Runtime /Volumes/RUDI Runtime ~/.rudi
Personal ~/Personal Requires a separately verified location

Personal-account preferences belong in private host configuration. Always bind and verify the selected host-local catalog entry; never change global login or fall back to another account. A Personal task requires the chosen execution host to have the intended files; Admin cannot read the workstation’s Personal files while the workstation is down. The workstation’s current native versions differ from the qualified Admin worker, so local scheduling also requires worker qualification before activation.

Before adoption: publish/review the source, update the clean accepted Admin checkout, build a new immutable bundle, verify provider identities, exercise project and child writes plus Git metadata, enumerate MCPs and router tools, and verify native supported connections. Then deliberately revise each selected task with its broader roots and trusted mode. Preserve old bundles and task revisions for rollback; do not rewrite active pinned profile files or switch worker services implicitly. No existing task gains access from this source change alone.