Recipes for the connected developer

Use this recipe when you alternate between Claude Code and Codex. The device connection is shared. Each agent needs its own provider authentication and observed activation.

Prepare the first repository

• Check macOS or Linux, Node.js 22 or later, and npm.
• Obtain the approved setup distribution and personal enrollment command.
• Choose the assigned working repository for both initial inspections.

Run the supplied setup command with both agents selected. Complete native sign-in in each agent. For Codex, open /hooks. Read and trust the Betterloop hooks.

npx @betterloop/setup setup --harness claude-code --harness codex

Open a fresh session in each agent. Use the same approved working repository for the initial comparison. Inspect each connection independently.

betterloop doctor --harness claude-code --project /path/to/repository
betterloop doctor --harness codex --project /path/to/repository

Decide whether both are usable

Review native authentication, signed policy, activation, and required controls in each result. A working Claude Code session cannot establish Codex hook trust.

If one agent lacks a mandatory control, withhold acceptance for that agent. Review compatibility with the administrator. Retain the other agent’s independent result.

For Harbor Search, this means recording two search-api inspections. Shared enrollment establishes the device relationship; the separate records establish what each agent reports.

Use this recipe for a resource approved for developer self-service. Approval makes that version available for selection. It does not install the resource everywhere.

Team-required resources remain controlled by reviewed rules. Personal selections can tighten restrictions. Shared resource identities and permission limits still apply to those personal selections.

Inspect the prepared selection

betterloop library browser
betterloop library list
betterloop library show org.review

The read-only catalog session lasts ten minutes. Read the actual instructions, supporting files, dependencies, and exact version before copying the prepared local command.

For hooks, inspect the executable script and supported lifecycle event. For an MCP connection, inspect account authorization requirements before installation.

Apply it to the chosen agents

betterloop library install org.review --version 1.0.0 --harness claude-code,codex
betterloop library status

Read the result for each agent. Complete any native sign-in or trust step it requires. Open fresh sessions if activation changed. Run doctor afterward.

The selection applies across repositories on this device for those agents. Library status describes local selection state. Later execution requires separate activity evidence.

Move to another approved version

betterloop library update org.review --version 1.1.0 --harness claude-code

Selections remain pinned until an explicit update. Approved dependencies also use exact versions. Optional resources require Mimir 0.26.0 when Mimir is selected.

An organization approval change becomes available after reviewed team baseline adoption and publication. Run sync to receive that release.

If Claude Code succeeds and Codex fails, keep the working selection. Repair the Codex requirement. Repeat the resource operation only for Codex.

Use this recipe with Claude Code or Codex native optimization. Drafts remain private and scoped to repository, runtime, and enrollment.

A draft can qualify through two distinct tasks. A correction with an existing explicit verification receipt can also qualify. Duplicate observations and candidates are deduplicated.

betterloop optimize --harness codex --action list
betterloop optimize --harness codex --candidate UUID --action review

Read the draft exactly as presented. Inspect its evidence and intended applicability. Decide whether it belongs only locally, should be dismissed, or merits shared review.

betterloop optimize --harness codex --candidate UUID --action keep
betterloop optimize --harness codex --candidate UUID --action dismiss

Keeping or dismissing a candidate suppresses further notices for that candidate. A submitted revision remains immutable. A changed draft receives a new identity.

Review the separate submission before sharing

betterloop contribute --harness codex --candidate UUID

Read the displayed public content separately from the private draft. Type submit only when you intend to upload those displayed fields.

The MCP bridge cannot make that submission automatically. Public fields exclude script bytes, private instructions, source, local paths, commands, and transcripts.

Assessment alone does not publish a lesson. A different human must approve the exact eligible revision. Signing and staged delivery follow that review.

Use this recipe only for a learning entry actually delivered in the current session. Exposure is neutral. It does not establish that the task succeeded.

Obtain the entry identifier and matching session identifier. Select a local check that evaluates the intended outcome of the work.

betterloop verify --harness codex --entry ENTRY_ID --session SESSION_ID -- npm test

Run the verification command. Read its outcome. A passing receipt can support a later success report tied to that existing exposure.

An unrelated passing command cannot verify the task. Finishing a model turn cannot verify it either. A failed check alone does not establish lesson causality.

Describe the check that passed and its actual scope. These receipts are client evidence. They do not supply independent device attestation.

For Harbor search-api, the selected check should exercise the changed search behavior. A generic command that exits successfully would not substantiate that result.