Skip to content
Agentmart
Submit a tool

iterate

Run one complete AutoResearch HPC iteration — claim a number, pre-declare the design before any result exists, execute, cross-check with a foreign harness…

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-iterate

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

Run one complete AutoResearch HPC iteration — claim a number, pre-declare the design before any result exists, execute, cross-check with a foreign harness, and record the outcome in the ledger.

Use cases: design

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/iterate

What's inside

Skill name
iterate
Description
Run one complete AutoResearch HPC iteration — claim a number, pre-declare the design before any result exists, execute, cross-check with a foreign harness, and record the outcome in the ledger.
License
MIT
Instructions
112 lines in SKILL.md · view on GitHub

Files (1)

  • SKILL.md

SKILL.md (first lines)

# Run an iteration

One iteration answers **one question**. If the question becomes “and also…”,
that is normally another iteration.

## 0. Read what the project already knows

`arh context` prints the current state, open iterations, the directives in force
and the gotchas index; open the gotcha entries about the tools and inputs you will
use, and the site's `site_gotchas` file when `.arh/config/site.md` names one.
Whatever you learn that outlives this session goes back into the project's files,
never only into a harness's private memory.

## 1. Claim

```bash
arh claim -t "the question, stated as a question"
arh claim -t "the same question, corrected design" --continues 7    # another attempt at 7's question
```

This is the only supported way to take an iteration number. When an iteration
tries again at a question an earlier one addressed (a corrected design, a failed
acceptance criterion handled), claim it with `--continues N` or a shared
`--question ID`: the ledger reports each concluded iteration as attempt k of n
on its question, counting the ones that did not conclude.

## 2. Pre-declare

```bash
arh new -n N
```

Complete every section of `iterations/iterationN/README.md`, then freeze it:

```bash
arh gate predeclare -n N
```

The frozen declaration is immutable. A changed design is a new iteration.

## 3. Execute

Keep scripts under the producing iteration and pin scientific tools:

```bash
arh run samtools -- samtools view -c input.bam
arh submit iterations/iterationN/scripts/experiment.nf -n itN_01
```

Scientific packages come from `arh env create NAME pkg=version ...`, solved once in
the task image and locked. `arh submit WORKFLOW.nf --lint` checks a workflow before
it runs; `reference.md` beside this file shows a minimal one.

A confirmation on held-out data (`sealed_inputs`) first freezes every artefact
carrying a decision: `arh freeze -n N MODEL THRESHOLDS SCRIPTS...`. Only then can
its runs read the sealed paths, and each read is recorded (`arh freeze list`).

GPU work gets `label 'gpu'`, a second tool image `label 'image_<name>'`, and code
from another iteration is declared under `imports =` in the pre-declaration. Work
that cannot go through `arh submit` is recorded at once with `arh note`.

Check acceptance criteria before interpretation. A failed criterion is this
iteration's result: report and conclude it, then pre-declare the handling in a new
iteration. It is not a reason to stop.

## 4. Report

Write `results/report/iterationN_report.md`. Report the pre-declared quantity,
negative-control behaviour, detection limit and missed predictions.

## 5. Cross-check

```bash
arh ask --role adversary -n N
arh ask --role estimand-auditor -n N
```

Evaluate findings on their merits. A verdict is evidence, not a ruling.

`arh ask` blocks until the review is written; stay in the session until it returns.
Exit 75 (usage or rate limit, authentication) and exit 77 (the provider refused the
content) spend no review round. After 75, run it again after a bounded wait (the stated reset is an upper bound);
after 77, the configured fallback verifiers have already been tried.

**Where to record it.** A review is bound to the report's sha256 and to every
file under `results/`. Editing the report, or adding a file under `results/`,
after the review makes it ineligible. So:

- write the report's *Cross-check* section before `arh ask`, as a pointer;
- put your evaluation in `iterations/iterationN/REVIEW_RESPONSE.md` — at the
  iteration root, outside `results/`, and not named `CROSSCHECK_*` (that glob
  counts review attempts);
- answer every numbered finding by its id with a decision and the reason
  (`F1: accepted ...`, `F2: rejected, because ...`); `arh gate results -n N
  --skeleton` lists them, and the gate fails while one is unanswered;
- answer an `UNSOUND` verdict with `Verdict: rebutted, because <reason>`, or
  `Verdict: accepted` and leave the iteration unconcluded; put a rebuttal to the
  reviewer with `arh ask --rebuttal -n N`. The gate fails until the reviewer
  withdraws every `UNSOUND` verdict; a maintained or accepted one means the
  correction is a new iteration;
- if a finding needs new computation, do it through `arh verify`, not by
  changing the iteration, and append the outcome to the response file.

Every number the report states should be in a file under `results/`; the
results gate lists those that are not.

## 6. Conclude

```bash
arh gate results -n N
arh ledger render
arh ledger check
```

Record what the iteration established, its strongest limit and what it
supersedes. Never edit an older conclusion to make history look cleaner.

From HudoGriz/autoresearch-hpc, MIT.

Signals

Stars
0
Forks
1
Open issues
0
Last commit (this folder)
Oct 15, 2026 (15 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

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