Files

102 lines
2.7 KiB
Markdown

---
name: checkpoint-research
description: Run multi-step research tasks with checkpoint tracking. Saves progress to an MD file after each step so work survives timeouts and can be resumed. Use for competitive analyses, multi-source investigations, or any research with 3+ sequential steps.
version: 1.0.0
author: case
metadata:
openclaw:
emoji: "📋"
category: "research"
tags:
- research
- checkpoint
- resumable
- analysis
---
# Checkpoint Research
Run multi-step research with automatic progress tracking. If you time out, the next run picks up where you left off.
## When to Use
- Competitive analyses (multiple competitors to research)
- Multi-source investigations (several URLs/topics to cover)
- Any research task with 3+ sequential items that might timeout
## How It Works
1. Create a progress file at the specified path
2. Before each step, read the progress file to check what's done
3. Skip completed steps (marked `[x]`)
4. After completing each step, write findings to the file and mark `[x]`
5. If timed out mid-run, next invocation reads file and continues
## Usage
When given a research task, structure it as follows:
### 1. Initialize Progress File
```markdown
# [Research Title] — Progress
## Status: IN PROGRESS
## Steps
- [ ] 1. [Item name]
- [ ] 2. [Item name]
- [ ] 3. [Item name]
## Final Sections
- [ ] Summary / Verdict
- [ ] Recommendations
## Findings
*(Written below as each step completes)*
```
Save to: `data/investigations/[topic]-progress.md`
### 2. Research Loop
```
For each unchecked item:
1. Read progress file
2. Find first unchecked `- [ ]` item
3. Research it (web_search, web_fetch, etc.)
4. Append findings under ## Findings with a ### heading
5. Mark the item `[x]` in the checklist
6. Write updated file to disk immediately
7. Continue to next item
```
### 3. Completion
When all items are `[x]`:
1. Write final sections (summary, recommendations)
2. Change `## Status: IN PROGRESS` to `## Status: COMPLETE`
3. Deliver the full report
## Template for Spawning
When spawning a sub-agent with checkpoint research:
```
Read the progress file at [PATH] first.
If any items are marked [x], skip them — continue from the first unchecked item.
After researching each item, IMMEDIATELY write your findings to the file and mark it [x].
When all items are complete, write the final sections and mark status as COMPLETE.
```
## Cost Impact
~2-3K extra tokens per full run vs non-checkpoint approach.
Saves 20-30K+ tokens on any timeout (avoids full re-run).
Net positive whenever timeout risk exists (tasks with 5+ web fetches).
## Progress File Location
Default: `/home/wdjones/.openclaw/workspace/data/investigations/[topic]-progress.md`