Skip to content
Back to skills

Checkpoint

ASecurity

Record progress to subtask: Phase completion (from autoworker:code) or test results (from autoworker:test). Auto-detects upstream type from conversation context. Ends by calling autoworker:dispatch.

  • 17 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added May 26, 2026
toolsgotesting

Security analysis

A100/100

Scanned May 27, 2026

npx -y skills add phj128/autoworker --skill checkpoint --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Checkpoint?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Checkpoint
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/phj128-checkpoint/badge)](https://www.skillsdirectory.com/skills/phj128-checkpoint)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: checkpoint
description: |
  Record progress to subtask: Phase completion (from autoworker:code) or test results (from autoworker:test).
  Auto-detects upstream type from conversation context. Ends by calling autoworker:dispatch.
---

# autoworker:checkpoint — Record Progress (Phase Check-off / Test Results)

Called after every autoworker:code or autoworker:test completion. Performs different record-keeping based on upstream type, then always calls autoworker:dispatch.

## Execution Flow

### 1. Locate Subtask

```
Glob `subtask_*.md` (exclude subtask_template.md) →
  0 found → stop, prompt to create subtask
  1 found → use directly (backward compatible)
  multiple → grep `status:` to filter:
    - Files without status field treated as active (backward compatible)
    - Exactly 1 active → use it
    - 0 active → list all files + status, prompt user to choose
    - >1 active → report anomaly
→ Read → locate "## Plan", "## Verification Plan", and "## Test Results" sections
```

### 2. Determine Upstream Type

Prioritize conversation context; when context is insufficient, infer from file state:

- **Upstream is autoworker:code**: Conversation contains Phase implementation content; or file has unchecked Phase with corresponding code files already modified
- **Upstream is autoworker:test**: Conversation contains test execution output; or file has all Phases checked but incomplete test layers

### 3a. Upstream is autoworker:code → Phase Check-off

1. Extract the just-completed Phase number and Step list from conversation context
2. Edit subtask's "Plan" section: check off all Steps for that Phase `[x]`
3. Append completion record in "Progress Log" section:

```markdown
**Phase X complete**
<brief change description>
```

### 3b. Upstream is autoworker:test → Test Results Write

1. Extract test level and per-item results from conversation context
2. Edit subtask's "Test Results" section, format:

```markdown
### L<N>
- `<command>`: <output summary> PASS/FAIL
```

3. Check off passed items `[x]` in the "Verification Plan" section

### 4. Judgment Criteria

**Hard standard for "pass"**:
- Function completes expected task and returns **meaningful results**
- Returning empty array/empty string without error ≠ pass
- "No exception thrown" ≠ pass

### 5. Output Summary

```
Checkpoint recorded:
- Type: Phase completion / Test record
- Content: Phase X checked off / L<N> results written
→ Invoking autoworker:dispatch
```

### 6. Chain: Immediately Invoke autoworker:dispatch

**After outputting the summary, immediately invoke `autoworker:dispatch`. Do not wait for user instructions, do nothing else.**

## Important Notes

- **Do not return to user mid-way**: After recording, go directly to autoworker:dispatch — don't report or ask
- **Do not skip verification plan items**: The verification plan is a carefully considered product — don't ad-hoc decide "this one doesn't need testing"
- **Chaining is mandatory**: Must invoke autoworker:dispatch after completion, cannot skip or manually substitute

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.