groupwork
A second agent on the work, and a result you can cite
A second opinion that has already been told your conclusion is not a second opinion, it is agreement with extra steps. groupwork runs four named patterns - attack an idea, judge material independently, rule on finished work against its constraints, or put one hard question to a panel whose members answer alone - and none of them is told your view, which is enforced in code rather than promised.
Install groupwork
Add the DBHQ marketplace, once
/plugin marketplace add dbhq-uk/marketplaceInstall the skill
/plugin install groupwork@dbhq
Both commands work in Claude Code and in Codex. The marketplace is a public repository you can read first: dbhq-uk/marketplace (external site)
What it is good at
Withholding your view in code, not in prose
No pattern will carry your conclusion. The red team pattern is started in an empty directory, and on Codex a permission profile lets it read only that directory and the system's runtime files. Give it your draft conclusion and any run of six of its words found in the brief refuses the run before it costs anything; declare you have none and the citation says the check did not run.
Writing the record at the moment of the run
Every run keeps the exact brief it was sent, and its citation carries that brief's hash, what the run withheld and whether the leak check ran. The withheld line is copied from the pattern definition rather than typed afterwards from memory, and runs that fail, time out or are stopped are logged too. Nobody writes the independence claim, so nobody can write one that is not true.
Putting one hard question to a panel
panel sends the same brief to at least two members at once, on different providers where more than one is ready, and each answers alone in its own session without sight of the others. The result says when every answer came from one model family, because agreement then counts for less. One optional critique round shows each member every first answer, unlabelled.
Treating the counterpart as a provider, not the product
Codex and opencode sit behind one three-function contract, with nothing vendor-specific outside a providers directory, and opencode brings other model families through the same adapter. Swapping which agent gives the second opinion does not change the patterns.
Saying which counterparts are actually usable
It reports which are installed and which are signed in, which are not the same thing, and asks opencode which models it can actually reach. A run that could never succeed - a model the installed CLI is too old for, or verify with no constraints to rule on - is refused before it costs anything.
What it does not do
It does not delegate work
There is no go-and-build-this mode. Another tool already does that well, and it is the mode where a sandbox mistake costs real money.
It does not write findings into your repository
It records the run and hands back a citation. Where a finding belongs, and whether it belongs anywhere, stays your call.
It does not score or benchmark the counterpart
It reports what the other agent said. It does not rank models, grade answers or claim one counterpart is better than another.
GitHub Copilot cannot be the counterpart yet
An adapter for Copilot is in the source, written from its documentation, but it has never been run against the real CLI and could not complete a run. So it is not registered: the provider check does not list it and a run cannot ask for it. It comes back after one real run has shown it works.
Read-only is not the same guarantee on every provider
Codex enforces it in the kernel, and where Codex cannot start the sandbox that keeps the red team pattern out of your repository, the run fails before any model call rather than going ahead with the repository readable. opencode has no sandbox: there, read-only is a set of deny rules opencode enforces itself, which is weaker, and the citation names the provider.
Read it before you install it
Every line is public and MIT licensed. The README is the long version of this page, including the parts that are still rough.