OMS carries Solana reference material with it, so the model starts a task with facts that are current to the pinned versions instead of whatever it remembers from training. There are two parts: a program reference that works offline, and a pinned documentation set the model queries through tools.

Built-in program reference

The program reference works offline. It holds 50 programs (21 native and SPL programs and 29 applications) with 1,559 instructions between them. For 18 programs it holds facts only, meaning names, tags and account order, because the source license is not on the allowlist or is unclear. A compact index of programs (program id, tag style and instruction names) is part of every system prompt and stays under 17,500 bytes. For detail, the program_reference tool returns discriminator bytes, accounts in order with signer and writable flags, and argument types, with no network needed. You can read the same data yourself with /program <name|id> [instruction]. Applications were chosen from a DefiLlama ranking snapshot dated 2026-11-13. The data is generated from pinned sources, and a --check run fails if the checked-in file differs from a fresh build. The system prompt also carries a reviewed static reference pack, built from the MIT-licensed Solana Foundation solana-dev 2.7.0 skill and the Helius helius 1.1.5 and svm 1.2.0 skills.

Documentation set

OMS draws on a pinned set of Solana documentation. As built on 2026-11-13, it has 46 sources, 1,777 documents and 8,148 chunks. One more source, idl-mpl-core, is skipped because its license is not on the allowlist. Every source is pinned to a version and a full commit. Examples are Anchor 1.2.2, Agave 4.4.1, Kit 8.6.0, Surfpool 1.7.0, platform-tools 1.58, Pinocchio 0.12.0, LiteSVM 0.18.2 and Mollusk 0.17.1. Only sources with a license on this list are built: MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, CC-BY-4.0 and CC0-1.0. If any source has a missing field, a missing license file or a license off the list, the build aborts.

Lookup tools

The model has nine lookup tools for this set: search_docs, read_doc, list_sources, decode_error, idl_lookup, detect_topics, list_topics, get_topic and api_lookup. Every result is wrapped in an <untrusted-knowledge> frame, with variants of the closing tag neutralized, and capped at 16,000 bytes. Each call is a separate request with a 20 second timeout, and redirects count as errors. The access token never enters the sandbox, the prompt or a tool result.