FAQ

FAQ

Frequently asked questions about GolemCore Bot.

---

What is GolemCore Bot?

An open-source autonomous agent framework (Java/Spring Boot) built around:

• ToolLoop (single-turn internal LLM↔tools loop)
• typed outcomes (TurnOutcome, FailureEvent, RoutingOutcome)
• transport-only routing with OutgoingResponse as the single source of truth

---

How does it avoid “tool-call mismatch” errors when models switch?

Provider/model switches can introduce incompatible tool-call IDs/names. The bot builds a request-time conversation view that can:

• normalize/mask tool-call IDs and names
• flatten tool interactions on model switch

Raw history is preserved.

---

What is OutgoingResponse?

OutgoingResponse is the canonical transport payload prepared in the pipeline and consumed by ResponseRoutingSystem. Routing is transport-only (no domain decisions; no raw-history writes).

---

What happens if something fails mid-turn?

Failures are captured as typed events (FailureEvent) and a TurnOutcome is finalized. FeedbackGuaranteeSystem ensures the user receives a minimal response even if upstream systems fail.

---

Why do some runs stop with “deadline exceeded”?

Each turn is protected by guardrails:

• wall-clock deadline
• max LLM calls
• max tool executions

Defaults are configured to support long-running autonomous work. See Configuration.

---

How does model routing work?

The bot routes per request using a tier (balanced/smart/coding/deep), influenced by:

• user tier selection
• active skill preference (model_tier)
• dynamic upgrades

See Model Routing.