72 lines
2.6 KiB
Markdown
72 lines
2.6 KiB
Markdown
# Executive Summary: Self-Calling Iteration Framework in xml-pipeline
|
|
|
|
**Date**: December 29, 2025
|
|
**Component**: Core reasoning mechanism for LLM personalities
|
|
|
|
### Principle
|
|
|
|
In xml-pipeline, **multi-step reasoning, planning, tool use, and iteration are implemented through open, auditable self-calls** — not hidden loops or state machines.
|
|
|
|
The LLM personality participates in the conversation exactly like any other listener: it can send messages to its own root tag.
|
|
|
|
### How It Works
|
|
|
|
1. **Conversation thread = memory**
|
|
Every message (user, personality, tool) is appended to history keyed by `convo_id`.
|
|
|
|
2. **Self-reflection = self-message**
|
|
To think step-by-step, the LLM emits its own root tag with the same `convo_id`:
|
|
```xml
|
|
<ask-grok convo_id="123">
|
|
First, let's outline the steps...
|
|
</ask-grok>
|
|
```
|
|
|
|
3. **MessageBus routes it back** to the same personality instance.
|
|
|
|
4. **Personality appends its own message** to history and calls the LLM again.
|
|
|
|
5. **LLM sees full context** — including its previous thoughts — and continues.
|
|
|
|
6. **Iteration emerges naturally** from repeated self-calls until final response.
|
|
|
|
### Key Properties
|
|
|
|
- **No hidden state** — the thread history is the only memory
|
|
- **No special controller** — MessageBus and `convo_id` do all coordination
|
|
- **Fully auditable** — every thought, plan, and intermediate step is a logged message
|
|
- **Tool use fits identically** — tool calls are messages to other root tags
|
|
- **Termination is natural** — LLM decides when to emit final response tag (e.g., `<grok-response>`)
|
|
|
|
### Structured Planning Support
|
|
|
|
Personalities are encouraged to use visible structures:
|
|
|
|
```xml
|
|
<todo-until condition="all primes under 100 listed">
|
|
<step done="true">List 2</step>
|
|
<step done="true">List 3</step>
|
|
<step done="false">List 5</step>
|
|
</todo-until>
|
|
```
|
|
|
|
This enables:
|
|
- Self-coordination (LLM reads its own plan)
|
|
- GUI progress rendering
|
|
- Explicit termination conditions
|
|
|
|
### Owner Control
|
|
|
|
- `iteration-capacity` config (`high`/`medium`/`low`/`none`) tunes how strongly the personality is prompted to iterate
|
|
- All behavior governed by owner-provided prompts and response templates
|
|
- No personality can loop indefinitely without owner consent
|
|
|
|
### Result
|
|
|
|
The framework turns **conversation into computation**.
|
|
|
|
Loops, conditionals, and planning become **visible message patterns** rather than hidden code.
|
|
|
|
The organism reasons by talking to itself in the open — producing a complete, transparent trace of thought at every step.
|
|
|
|
This is the core iteration mechanism for all LLM personalities in xml-pipeline.
|