FAQ
Frequently asked questions about Dsh-Go. The full version is available in docs/FAQ.md.
Q1. What exactly is Dsh-Go?
A pure-Go, in-process DeepSeek Harness Agent reference implementation — it translates the official DSH’s Turn/Step dual loop, Event Sourcing, Goal planning, tool governance, and other core capabilities seam by seam into Go code.
Q2. How does it relate to Eino / LangChain?
Different species. Eino/LangChain is a “framework to use” (general LLM orchestration); Dsh-Go is a “reference implementation to read” (a transcription of DSH). It neither intends to nor is capable of replacing the former.
Q3. What advantages does it have over them?
Three asymmetric hard strengths: ① word-for-word alignment with the official semantics (events/error codes/Goal four states); ② Event Sourcing + fold projection (a structural difference, 3437× incremental); ③ prefix-cache friendly (DeepSeek cost reduction).
Q4. Can’t mature frameworks do compliance auditing?
They can. “Being auditable” is not a unique capability; the difference lies in capability ownership and strength of evidence: mature frameworks rely on external hooks (callbacks/manual persistence), while Dsh-Go auditing is built-in (append-only events are the source of truth, intermediate states can be replayed).
Q5. What enterprise scenarios is it suitable for?
Primary battlefield = audit replay + DeepSeek cost reduction + controlled tool execution, compounded, and anchored on planning-based Agent platforms built on DeepSeek.
Q6. Does this project have real value?
The technical value is real (word-for-word reproduction + benchmark data), but it ≠ external production-adoption value (zero ecosystem, no endorsement). It is verification + reference value — suitable for personal use, teaching, or a portfolio, not as a production replacement.
Q7. Recommended learning path to get started?
Start with examples/tutorial, then read through pkg/session / pkg/goal, and finally dive into pkg/agent / pkg/tools / pkg/llm.
Q8. Useful for replicating / re-implementing an Agent?
It is a ready-made transcription template: every seam is split into independent small modules + comments + tests — effectively a “disassembly manual of an Agent core implementation”.
Q9. What is Event Sourcing?
Q10. What is fold projection?
See Tutorial: fold projection.
Q11. Does it support models other than DeepSeek?
Currently only provider_deepseek is built in; however, the LLMAdapter seam is extensible — implementing 3 methods is enough to connect a model (an OpenAI-compatible provider approach is described in docs/FAQ.md Q11).
The full version (including comparison tables, code, and criteria) is available in docs/FAQ.md.