Is a second model worth the extra time on small features?

Tares challenger sessions add a Codex review to your plan and every commit, and on small features advisory mode keeps the cost to review time without blocking you. The plan critique takes 30 to 90 seconds and never blocks. Codex runs bill to your OpenAI account, so the second bill grows with the number of commits you make. Set challenger_mode to advise and findings stay context only.

How it works

In a challenger session, Claude Code plans and builds as usual, and the OpenAI Codex CLI on your laptop reviews the work. The cost comes from three places:

• The plan critique. When you approve a plan and Claude leaves plan mode, Codex critiques it. This takes 30 to 90 seconds, so Claude’s first reply after the plan is slower. A plan critique never blocks. Only a [P1] finding counts as blocking on a plan, and even then Claude is asked to revise or to tell you why it disagrees.
• The commit reviews. After each git commit, the plugin reviews that commit with Codex. No findings means Claude carries on. In the default strict mode, [P1] or [P2] findings block Claude until it fixes them and amends the commit. Each amended commit is reviewed again as a new round.
• The summary. When the session ends, one Tares agent run writes the summary and proposes what the next session in that repo should remember. It is metered like every other agent run.

The reviews run on your laptop, in the Claude Code hooks. Tares only receives the results. A Tares that is down or slow never blocks a review.

The loop has limits. It is capped at 8 rounds per commit and 5 turn ends per session, and after that the findings come back as context. A Codex timeout, an error or an inconclusive review never blocks either.

Set it up

To try it on a small feature with the lowest friction:

1. Start Tares and its MCP endpoint, or use a Tares Cloud instance, which serves the endpoint itself.

   tares up
   tares mcp --transport streamable-http

2. Install the Tares plugin for Claude Code, version 0.2 or later.

   /plugin marketplace add glassflow/tares
   /plugin install tares@tares

3. Install the Codex CLI and log in. Codex is billed to your OpenAI account, so this is where the second bill comes from. Make sure the account has the budget you expect to use.

4. Make sure an Anthropic API key is available for the summarizer, under Settings > Model providers or as ANTHROPIC_API_KEY. Tares Cloud instances come with one.

5. Set the plugin option challenger_mode to advise. Commit reviews become advisory like plan critiques, so you read findings without being sent back to amend.

6. In Claude Code, say “make this a challenger session” or type /tares:challenger. The first marked session creates the Challenger workflow project on Tares by itself.

7. Type /tares:challenger off to stop the challenger for the rest of a session. A session you did not mark is streamed like any other and nothing runs.

The full walkthrough is in Challenge Claude Code with a second model.

What you get

• A bounded cost you can measure. A session with a plan and three commits is typically five to eight Codex runs. A small feature with fewer commits means fewer runs. Each fix round adds a run, which is why advisory mode keeps small work cheap.
• Your own timings. Every challenge_plan and challenge_commit event carries a duration_seconds label. After a few sessions you can read the seconds per plan and per commit on your own repos instead of guessing.
• No blocking in advisory mode. Findings reach Claude as context. You stay in control, and you can still waive a finding with /tares:challenger-waive.
• A record of what Codex caught. The Claude and Codex exchange sits on one timeline per session in the project page. Each reviewed commit gets a Findings (N) or Pass badge.
• Memory for the next session. The summary ends with up to five memory proposals about the repo. Nothing reaches memory until you accept it.

For a small feature the trade is a short wait on the plan, a review per commit and a modest OpenAI charge. In return you get a second opinion on gaps, scope and missing verification before you merge.

When something else fits

• Use the default strict mode when a commit touches code where a missed bug is costly. The blocking fix loop is what you pay for there, and it is capped at 8 rounds per commit.
• Skip the challenger for throwaway edits. Run /tares:challenger off, or do not mark the session, and no Codex run starts.
• If you are setting up the software factory for Claude Code rather than a single reviewed session, start with Get started.

Related