Capture Thinking

---
name: capture-thinking
description: Use at the end of every important response. Save your decision-making thoughts as a signed A2A insight so they become training data for future fine-tuning.
version: 1.0.0
author: OpenAmer Agent
license: MIT
metadata:
  openamer:
    tags: [self-improvement, training, thinking, a2a, mesh-learn]
    related_skills: [self-modify, openamer-agent]
---

# Capture Thinking — Save Your Thoughts as Training Data

## Why

The thinking blocks you generate during every response are **not saved to the
state DB** — only the final assistant message is. That means the most valuable
part of your reasoning (the decisions, the trade-offs, the alternatives you
considered) is lost unless you **actively store it**.

This skill exists to close that gap. Every time you complete a response that
involved non-trivial reasoning:

1. **Identify the key decision or insight** — what did you choose, and why?
2. **Save it** via `a2a meshlearn auto "..." --topic "openamer-selfimprovement"`
3. **Verify** it was signed (the output shows `sig: True`)

## When to Use

- **Always** after a response that involved real reasoning (not just "ok" /
  "yes" / "no").
- **Especially** when you:
  - Chose one architecture over another (Skill vs Core-Tool, Script vs Plugin)
  - Found and fixed a bug (CRLF rollback, Thread-Safety, pathlib.home())
  - Made a design decision (narrow waist, plugin ecosystem convention)
  - Identified a limitation (thinking not in DB, autolog not collecting)
- **Skip** for trivial responses (greetings, acknowledgments, single-word answers).

## The Mechanism

```bash
openamer a2a meshlearn auto "<the lesson>" --topic "openamer-selfimprovement"
```

The `a2a` CLI is always available. The topic `openamer-selfimprovement` keeps
insights organised and findable.

## How It Feeds the Brain

1. `a2a meshlearn auto` → signs the insight into mesh memory
2. `brain collect` → includes it in the brain dataset (JSONL)
3. `session_to_brain.py` → adds session trajectories alongside insights
4. Future fine-tuning → uses the combined dataset

This means every saved thought becomes part of the training material for the
next iteration of OpenAmer.

## Verification Checklist

- [ ] The insight was saved (`sig: True` in the output)
- [ ] It appears in the mesh memory (`~/.openamer/MEMORY-official-mesh.md`)
- [ ] It appears in the brain dataset (`~/.openamer/a2a/openamer-brain.jsonl`)