Skip to content
Agentmart
Submit a tool

cross-check

Have a different agent harness review an iteration as adversary, re-implementer, estimand auditor, or gotcha scanner before concluding it.

Skill by HudoGriz

IndexedMIT
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/hudogriz-cross-check

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Have a different agent harness review an iteration as adversary, re-implementer, estimand auditor, or gotcha scanner before concluding it.

Use cases: agent-building

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/cross-check

What's inside

Skill name
cross-check
Description
Have a different agent harness review an iteration as adversary, re-implementer, estimand auditor, or gotcha scanner before concluding it.
License
MIT
Instructions
40 lines in SKILL.md · view on GitHub

Files (5)

  • SKILL.md
  • adversary.md · roles/
  • estimand-auditor.md · roles/
  • gotcha-scanner.md · roles/
  • reimplementer.md · roles/

SKILL.md (first lines)

# Cross-check with a foreign harness

```bash
arh ask --role adversary -n N
arh ask --role reimplementer -n N --harness codex
arh ask --role estimand-auditor -n N
arh ask --role gotcha-scanner -n N
```

A same-family review can reproduce the producer's reasoning errors. `arh ask`
therefore requires a distinct configured model family for a review that satisfies
the results gate. `--same-family` records a weaker check.

Configure harnesses in `.arh/config/harnesses.md`.

A provider refusal is not a review and spends no round. Exit 75 is a usage or rate
limit, authentication or a timeout: retry after a bounded wait; the reset time
printed is an upper bound, not a schedule. Exit 77 means
the provider refused the content; `arh ask` then tries each harness in
`verifier_fallback`, which must be another model family than the producer's.

Every run writes a `CROSSCHECK_<role>_<harness>_<timestamp>.md` record and
execution metadata. The verdict is **evidence, not a ruling**. Evaluate each
finding, record dispositions and reasons, and preserve disagreements rather than
silently deleting inconvenient review output.

The reviewer is asked to number its findings (`F1:`, `F2:` ...). Answer each in
`REVIEW_RESPONSE.md` by its id, with a decision word (accepted, rejected, fixed,
deferred, disputed) and the reason; with several reviews, put each review's
answers under a heading that names its `CROSSCHECK_...` file. `arh gate results`
lists every finding and fails while one is unanswered (`review_response` in
`project.md`).

An `UNSOUND` verdict is answered as a whole, in the same place:
`Verdict: rebutted, because <reason>` when the iteration stands despite it, or
`Verdict: accepted` when it does not. An accepted one leaves the iteration
unconcluded, and the correction is a new iteration. A second, milder review does
not answer the first.

A rebuttal is then judged by a reviewer, not by you:

```bash
arh ask --rebuttal -n N            # the reviewer who said UNSOUND answers WITHDRAWN or MAINTAINED
```

The judgment is written to `REBUTTAL_<harness>_<timestamp>.md`, bound to the
review and to your rebuttal's text; editing the rebuttal afterwards voids it.
The gate passes only on `WITHDRAWN` (`unsound_review = require-adjudication`,
the default). A `MAINTAINED` verdict stands: leave the iteration unconcluded and
correct it in a new one. One judgment per review (`ask_max_adjudications`).

From HudoGriz/autoresearch-hpc, MIT.

Signals

Stars
0
Forks
1
Open issues
0
Last commit (this folder)
Oct 15, 2026 (14 days ago)
Last release
Sep 24, 2026 (1 month ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from HudoGriz

Cite this page

cross-check skill. Agent Mart. https://agentmart.dev/skills/hudogriz-cross-check (last updated 2026-10-24)