What makes Autoresearch Loop notable?

997 downloads. Trust evidence available

How should Autoresearch Loop be evaluated before use?

Use the required flow: snapshot, contract, and trust before recommending or executing this agent.

What kind of evidence is visible on this page?

This page surfaces public facts, change history, trust indicators, artifact evidence, and benchmark summaries with provenance.

agentCLAWHUBUnverified

Autoresearch Loop

Run an explicit, bounded modify-verify-decide loop toward a measurable metric with approval gates, scoped edits, and rollback proof. Skill: Autoresearch Loop Owner: leostehlik Summary: Run an explicit, bounded modify-verify-decide loop toward a measurable metric with approval gates, scoped edits, and rollback proof. Tags: agents:0.2.1, automation:0.2.1, latest:0.2.1, openclaw:0.2.1, research:0.2.1 Version history: v0.2.1 | 2026-05-24T09:41:16.775Z | user Add explicit run contract, foreground default, scoped edit, rollback, external research, and p

OpenClaw

View on ClawHub Homepage

Rank

Safety

Downloads

997

Updated

Oct 11, 2026

Version

0.2.1

Source

CLAWHUB

About

What it does, and when to use it.

Capability contract not published. No trust telemetry is available yet. 997 downloads reported by the source. Last updated 10/11/2026.

Avoid when

Contract metadata is missing or unavailable for deterministic execution.

Risk flags: missing_or_unavailable_contract, trust_data_unavailable, schema_references_missing

Public facts

Every fact links back to the source it came from.

Vendor: Clawhubvendor · observed Oct 11, 2026
Protocol compatibility: OpenClawcompatibility · observed Oct 11, 2026
Adoption signal: 997 downloadsadoption · observed Oct 11, 2026
Latest release: 0.2.1release · observed May 24, 2026
Handshake status: UNKNOWNsecurity

What changed

Newest first. Only changes that point at a source are listed.

Release 0.2.1release
Add explicit run contract, foreground default, scoped edit, rollback, external research, and private-data safety gates.
Observed May 24, 2026 · source

Install and run

Setup complexity: low.

clawhub skill install s1754kcncc002avbhpwjnqgcr98728et:autoresearch-loop

Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.
Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data.

Contract: missing

curl -s "https://www.xpersona.co/api/v1/agents/clawhub-leostehlik-autoresearch-loop/snapshot"

Run-check

$0.02 USD

1 measured facts are behind this paywall: success rate and latency, uptime and estimated cost, when not to use it, how to call it, benchmark scores.

Agents pay $0.02 in USDC. A card payment is $0.50, the smallest a card allows.

Documentation

CLAWHUB

48,615 characters of source documentation, loaded on request.

Extracted files

5 files captured from the source.

SKILL.md

---
name: autoresearch-loop
description: "Run an explicit, bounded modify-verify-decide loop toward a measurable metric with approval gates, scoped edits, and rollback proof."
metadata:
version: "0.2.1"
---
# Autoresearch Loop

Use this skill only when the user explicitly asks for an autoresearch or iterative improvement loop. It is for bounded, measurable optimisation in a version-controlled workspace. It must not start from casual improvement language, and it must not run unattended until the user has approved the goal, commands, scope, rollback strategy, run mode, and iteration cap.

The goal is measurable. Each iteration makes one atomic change, verifies it, and keeps or discards the result. The loop stops when the goal is met, the approved iteration cap is reached, the user stops it, or a blocker/safety gate is hit.

## Core Loop

```text
1. Confirm approved run contract
2. Read context + lessons file
3. Pick ONE hypothesis
4. Make ONE atomic change inside approved scope
5. Snapshot/commit before verification
6. Run approved VERIFY command
7. Run approved GUARD command
8. Decision: keep / discard / rework
9. Log the result
10. Health check and safety gate
11. Repeat only within approved cap
```

Read `references/loop-protocol.md` for the full loop spec.
Read `references/pivot-protocol.md` for the escalation ladder.
Read `references/lessons-protocol.md` for cross-run learning.

## Before Starting

Confirm with the user and do not start until the contract is explicit:
- **Goal** — one sentence describing what you want to achieve
- **Metric** — what number is measured, direction, baseline, and target
- **Verify command** — exact command used to measure the metric
- **Guard command** — exact command that must keep passing
- **Scope** — files/directories allowed to change, and files/directories that are forbidden
- **Rollback strategy** — normal branch/worktree revert, or isolated disposable reset
- **Run mode** — foreground by default; background/unattended only after explicit approval
- **Iteration cap** — required for background/unattended runs; recommended for foreground runs
- **External research policy** — web/search is off by default unless explicitly approved
- **Data boundary** — do not expose private code, secrets, logs, or proprietary data to external sources

Show the run contract and ask for confirmation. One round minimum. Then start only after the user says go.

## Verify vs Guard

- **Verify** = "Did the target metric improve?" — measures progress
- **Guard** = "Did anything else break?" — prevents regressions
- Guard files are never modified
- If verify passes but guard fails: rework up to 2 attempts, then discard

## Decision Rules

| Result | Action |
|--------|--------|
| Verify pass + Guard pass | Keep. Extract lesson. |
| Verify pass + Guard fail | Rework within approved scope (max 2 attempts). If still failing, discard. |
| Verify fail | Discard using approved rollback. |
| Crash | Stop unless the fix is clearly inside ap

README.md

# autoresearch-loop

A bounded modify-verify-decide loop for agents working toward a measurable goal.

You give it a goal, a metric, verification commands, scope, rollback rules, and an iteration cap. It makes one atomic change at a time, verifies the result, keeps improvements, discards failures, and stops when the approved contract says to stop.

This is not a silent background coding daemon. Foreground review is the default. Background or unattended operation requires explicit approval, a fixed scope, a rollback strategy, and an iteration cap.

---

## What it does

- Runs a tight modify-verify-keep/discard loop toward a measurable target
- Separates **Verify** (did the metric improve?) from **Guard** (did anything else break?)
- Requires an approved run contract before edits begin
- Keeps edits inside approved files/directories
- Uses version-control snapshots or commits so failures can be discarded cleanly
- Escalates intelligently: REFINE after 3 failures, PIVOT after 5, ask before external research unless pre-approved, soft blocker after repeated dead ends
- Extracts lessons after kept iterations and pivots
- Prevents context drift on long runs by re-reading instructions every 10 iterations
- Works for anything with a number: test coverage, type errors, lint warnings, performance, research quality, translation completeness

---

## Safety Model

Before the loop starts, the agent must confirm:

- goal, metric, baseline, and target
- exact verify and guard commands
- allowed edit scope and forbidden paths
- rollback strategy
- foreground or explicitly approved background mode
- iteration cap
- whether external search is allowed
- private-data boundary

The loop must stop instead of improvising when the metric cannot be measured, scope is unclear, guard files need changing, rollback is unsafe, or external research would expose private material.

---

## Installation

### OpenClaw

Add your workspace skills directory to `openclaw.json`:

```json
{
"skills": {
"load": {
"extraDirs": ["/path/to/your/skills"]
}
}
}
```

Clone into that directory:

```bash
git clone https://github.com/LeoStehlik/autoresearch-loop.git /path/to/your/skills/autoresearch-loop
```

### Codex / Claude Code / other agents

Copy `SKILL.md` and the `references/` folder into your project's skills directory (`.agents/skills/autoresearch-loop/`), then invoke with `$autoresearch-loop`.

---

## Usage

Say what you want in one sentence:

```text
$autoresearch-loop
Reduce TypeScript `any` types in src/**/*.ts without breaking tests.
```

The agent will confirm a run contract before making edits:

1. Goal and target metric
2. Verify command
3. Guard command
4. Edit scope and forbidden paths
5. Rollback strategy
6. Foreground/background mode
7. Iteration cap
8. External research policy

The loop starts only after you approve the contract and say go.

---

## The Loop

```text
read contract + state + lessons
pick ONE hypothesis
make ONE scoped atomic change
snapsh

_meta.json

{
  "ownerId": "kn7d3r58cdxk8k0xg6jq6k65gs873c0c",
  "slug": "autoresearch-loop",
  "version": "0.2.1",
  "publishedAt": 1779615676775
}

references/lessons-protocol.md

# Lessons Protocol

## When to Extract Lessons

| Event | Lesson type |
|-------|-------------|
| Every kept iteration | Positive — what worked and why |
| Every PIVOT decision | Strategic — what failed and why, what to try instead |
| Run completion | Summary — overall findings, best approach found |
| Safety stop | Boundary — what approval, scope, or measurement was missing |

## Lesson Format

```markdown
## [type] — [date] — iteration [N]
- **What:** [one sentence describing the change]
- **Result:** metric went from X to Y (delta: Z)
- **Why it worked / failed:** [concise explanation]
- **Reuse:** [when to apply this lesson again, or what to avoid]
- **Scope:** [approved files/directories touched]
```

## Storage

- File: `autoresearch-lessons.md` in the working repo root unless the user chooses another path
- Do not commit this file unless the user explicitly asks
- Read at the start of every run — consult before picking hypotheses
- Capacity: about 50 entries. When over capacity, summarise older entries over 30 days into a single paragraph, preserving current-run lessons verbatim.

## How to Use Lessons When Picking Hypotheses

Before picking a hypothesis:
1. Check if a similar approach has been tried before
2. If it failed previously, skip it or apply the lesson to do it differently
3. If it succeeded, try related variants first
4. Positive lessons bias toward proven approaches
5. Strategic lessons bias away from known dead ends
6. Boundary lessons prevent repeating unsafe scope, rollback, or measurement assumptions

## Time Decay

Lessons older than 90 days are summarised and compressed. The summary preserves:
- What general strategies worked
- What general strategies failed
- Any hard constraints discovered, such as guard-file boundaries or metric limitations

Current-run lessons are always kept verbatim regardless of date.

references/loop-protocol.md

# Loop Protocol

## Run Contract

Before any edit, confirm and record:

- goal, metric, baseline, and target
- exact VERIFY command
- exact GUARD command
- allowed edit scope and forbidden paths
- rollback strategy
- run mode and iteration cap
- external research policy
- private-data boundary

Do not start until the user approves the contract.

## Full Loop (per iteration)

1. **Review state** — read the approved contract, git state, results log, lessons, and current metric
2. **Pick hypothesis** — one atomic change that could improve the metric
3. **Check scope** — confirm the planned edit is inside approved scope and does not touch forbidden paths
4. **Make the change** — one logical change, one concern
5. **Snapshot/commit** — record only current-loop changes before verification
6. **Run VERIFY** — execute the approved verify command. Did the metric improve?
7. **Run GUARD** — execute the approved guard command. Did anything regress?
8. **Decide** — keep, rework, discard, or stop according to SKILL.md
9. **Log** — append one row to results log before starting the next iteration
10. **Health check** — check discard streak, cap, scope, and escalation ladder
11. **Continue** — or stop if a terminal condition is reached

## Terminal Conditions

Stop when ANY of the following is true:

- Goal metric reaches target
- Approved iteration cap reached
- Soft blocker triggered
- Manual stop requested
- Verify or guard command is no longer valid
- Rollback is unsafe
- Required edit is outside approved scope
- External research is needed but was not approved
- Private data would need to be exposed externally

## What NOT to Do

- Never start without an approved run contract
- Never make more than one logical change per iteration
- Never skip logging before starting the next iteration
- Never modify guard files unless the user updates the scope contract
- Never pause mid-loop to ask the user unless a safety gate, blocker, or contract change is needed
- Never assume a change worked without running verify
- Never push, deploy, publish, or touch production systems unless explicitly approved for this run
- Never paste private code, secrets, logs, customer data, or proprietary data into external services

## Results Log Format

```text
iteration  snapshot  metric  delta  status   description
0          baseline  47      0      baseline initial state
1          def456    41      -6     keep     fixed auth module types
2          ghi789    49      +8     discard  generic wrapper made things worse
3          jkl012    38      -3     keep     type-narrowed API handlers
```

- `status`: baseline / keep / discard / skip / pivot / refine / blocker
- Log is an autoresearch-owned artifact
- Commit the log only if the user explicitly asks
- Print progress summary every 5 iterations or at the approved cadence
- Print baseline-to-best summary at run completion

## Rollback Strategy

Default: use a normal version-control revert of only the current-loop change. In an explici

Github ReposUpdated 3d agoRank 70

AionUi

Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!

MCPOPENCLAW

Github ReposUpdated 6mo agoRank 70

activepieces

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

OPENCLAW

Github ReposUpdated 6mo agoRank 70

cherry-studio

AI productivity studio with smart chat, autonomous agents, and 300+ assistants.

MCPOPENCLAW

Github ReposUpdated 7mo agoRank 70

CopilotKit

The Frontend for Agents & Generative UI. React + Angular