Explanation
Understanding-oriented. Read these when you want to know why the template is shaped the way it is — grouped by domain (same order as the other sections).
Agent core & runtime
| Page | Question it answers |
|---|---|
| Architecture | How do the A2A handler, LangGraph runtime, and LiteLLM gateway fit together? |
| Model output | Native reasoning, and the thin guard that strips provider-leaked <think> from answers |
| Mid-turn steering | How can I redirect the agent mid-turn without stopping and losing its work? |
| LiteLLM gateway | Why route every call through a gateway instead of the provider SDK? |
Knowledge & memory
| Page | Question it answers |
|---|---|
| Memory & knowledge store | How does the agent remember across sessions — and why extract, not dump? |
A2A, fleet & delegates
| Page | Question it answers |
|---|---|
| A2A protocol | What does A2A actually require, and where do naive implementations go wrong? |
| Cost & trace propagation | Why do we emit cost-v1 and parse a2a.trace, and why that specific shape? |
Operate & deploy
| Page | Question it answers |
|---|---|
| Security & trust model | What's the trust posture — auth, redaction, origin checks, sandboxing? |
| Tuning & cost | Which cost/perf levers (compaction, aux-model routing, execute_code, prefix caching, failover) exist, and when to flip them? |
Architecture decisions
The ADRs record every significant design decision in MADR format (numbered, never deleted — superseded instead).