Resume

What Is Resume?

The Resume skill for Claude Code (from the autoresearch-agent toolkit) is a powerful utility designed to pick up paused or context-limited experiments exactly where they left off. Often, iterative research or code optimization experiments are stopped due to shifting priorities, context limitations, or the need for further analysis. The Resume skill automates the process of reconstructing an experiment’s full context—including history, configuration, and results—so you can seamlessly continue your workflow without manual reconfiguration or data loss. This skill is especially useful in environments where experiments are managed across different branches or stored as structured artifacts.

Resume is invoked via the /ar:resume command and operates by checking out the relevant experiment branch, loading all associated configurations and histories, and providing a concise summary of the current experiment state before allowing further iteration. The skill is open source and available at its GitHub repository.

Why Use Resume?

Modern software development, especially in research-driven or experimental contexts, often involves running multiple experiments in parallel or over long periods. This can result in:

• Interrupted workflows: Experiments are paused due to changing priorities or resource constraints.
• Context loss: Important details about the experiment’s state, results, and rationale may be forgotten or misplaced.
• Manual overhead: Resuming work often requires manually piecing together configuration files, git branches, and experimental history.

The Resume skill addresses these challenges by automating the restoration of experiment context. Its core benefits include:

• Time savings: No need to manually reconstruct the environment or search for experiment details.
• Accuracy: Ensures continuity by retrieving the exact state and history at the point of pause.
• Insightful summaries: Provides a synthesized overview of what has been done, what worked, and what failed, supporting better decision-making for next steps.
• Consistency: Maintains experiment tracking standards, reducing human error and improving reproducibility.

By leveraging Resume, teams can maintain momentum in experimentation, reduce cognitive overhead, and make data-driven decisions with confidence.

How to Get Started

Installation and Setup

First, ensure you have the autoresearch-agent and its skills installed per the instructions in the official repository. The Resume skill does not require separate installation if you have the agent set up.

Basic Usage

To invoke Resume, use the /ar:resume command:

/ar:resume

This command will list all available experiments, allowing you to select which one to resume. If you know the specific experiment you want to continue, specify it directly:

/ar:resume engineering/api-speed

Under the Hood

Upon execution, Resume:

• Lists available experiments if none is specified, using:

  bash

  python {skill_path}/scripts/setup_experiment.py --list

• Checks out the experiment-specific git branch:

  bash

  git checkout autoresearch/{domain}/{name}

• Loads all relevant files for full context:

  bash

  cat .autoresearch/{domain}/{name}/config.cfg
  cat .autoresearch/{domain}/{name}/program.md
  cat .autoresearch/{domain}/{name}/results.tsv
  git log --oneline -20

• Summarizes experiment status and recent history for the user.

Key Features

1. Intelligent Experiment Selection:
Resume lists all tracked experiments, showing their current status (active, paused, or done) based on the age and state of results.tsv. This makes it easy to identify which experiments can be resumed.

2. Full Context Loading:
On resumption, the skill checks out the relevant git branch and loads all associated artifacts, including configuration, strategy, results, and recent commit logs. This ensures no information is lost, and the context is fully reconstructed.

3. Comprehensive Status Summaries:
Before proceeding, Resume summarizes the experiment’s target, metrics, outcomes (e.g., best result, total iterations, kept/discarded/crashed runs), and recent trends or patterns. For example:

Resuming: engineering/api-speed
  Target: src/api/search.py
  Metric: p50_ms (lower is better)
  Experiments: 23 total — 8 kept, 12 discarded, 3 crashed
  Best: 185ms (-42% from baseline of 320ms)
  Last experiment: "added response caching" → KEEP (185ms)

4. Seamless Iteration:
With context restored, you can immediately continue experimenting, iterating, or analyzing, without manual setup.

Best Practices

• Consistent Experiment Naming: Use clear and descriptive names for experiments to make selection and resumption straightforward.
• Frequent Pausing and Summarizing: Pause experiments at logical milestones and use Resume to generate summaries; this aids collaboration and future analysis.
• Regular Commits: Commit changes regularly in experiment branches to ensure all progress is tracked and resumable.
• Review Summaries: Always review the summary Resume provides before proceeding, to avoid redundant or conflicting iterations.

Example workflow for best results:

## Start a new experiment
/ar:start engineering/api-speed

## Work iteratively, pausing as needed
/ar:pause engineering/api-speed

## Later, resume from where you left off
/ar:resume engineering/api-speed

Important Notes

• Git Integration Required: Resume relies on git branches for experiment tracking. Ensure your experiments are managed under version control.
• Single-Experiment Context: Only one experiment context can be active at a time; switching experiments requires a new Resume invocation.
• Results and Configurations: All experiment metadata (config, results, strategy) must reside in the expected .autoresearch/{domain}/{name}/ directory structure for correct operation.
• Data Integrity: Always verify that your results and configuration files have not been corrupted or manually altered outside the Resume workflow.
• Skill Updates: Check the official repository for the latest updates and bug fixes.

By following these guidelines, the Resume skill can significantly streamline experiment management and iteration, ensuring efficient and reproducible research workflows.