Agent Loop & Error Recovery

The core agentic loop: receive message, call LLM, execute tools, iterate, respond.

Agent Loop Flow

Persist user message — stored with importance score
Auto-route model — classify query complexity (if not overridden)
Build system prompt — base prompt + matched skills + known facts
Retrieve context — tri-hybrid memory retrieval
Iterate (up to max_iterations):
- Collect pinned old memories + recent messages (deduplicated)
- Build OpenAI-format message list
- Call LLM with error-classified recovery
- If tool calls → execute each, persist results, continue loop
- If no tool calls OR final iteration → return text response
Max iterations reached → return timeout message

The call_llm_with_recovery method classifies errors and responds accordingly:

Error Type	Strategy
Auth / Billing	Return immediately to user — no retry
Rate Limit	Wait `retry_after_secs` (capped at 60s), retry once
Timeout / Network / Server Error	Wait 2s, retry once; on failure, fall back to previous model
Not Found (bad model)	Immediately switch to fallback model
Unknown	Propagate as error

Last Known Good

After every successful LLM call, the current config is saved as config.toml.lastgood. This enables automatic recovery from bad config changes.

During the loop, each tool call receives: