RSO Context Alpha 0.11.0

Download 0.11.0 · Quick-start · Installation · Manual · Changes · Verification

RSO means Recursive Semantic Octree, from Paul Griswold's research paper. Research origin explains the graphics-inspired proposal and which parts this application implements.

The research idea borrows subdivision from 3D graphics: split a request into smaller tasks, check each part, assemble the answer, then check it backward against the original request. A failed check would direct further work to the part that needs repair. This application implements a local evidence-ledger subset of that proposal; it doesn't implement or validate the full research architecture.

RSO Context is a local evidence ledger for agents working with project files. It gives different agents a shared way to retrieve current requirements, decisions, source references, and recorded approvals from the same bounded project folder.

For example, a project file might say, "Decision: export the preview at 24 fps." RSO can return that sentence with its file path, line range, and content hash. If the file changes, the next ingest records a new version. If someone explicitly approves a claim, their named validation can travel with the evidence.

The files remain the source of truth. RSO doesn't learn from chat, call a model, use embeddings, or change model weights. It uses Python and local SQLite full-text search. Its command-line interface returns JSON; the optional browser viewer runs on localhost. An optional local stdio MCP adapter (rso-context mcp) uses the same CLI functions and ledger. It pins mcp==2.2.0 in an isolated runtime and is not required for ordinary CLI use. See INSTALL.md.

Start here

Download RSOContextAlpha-0.11.0.zip from the 0.11.0 release and extract it. The release page includes a SHA-256 checksum. GitHub access is required while the repository is private. A release ZIP is tied to a version; Code > Download ZIP downloads the selected branch's current source instead.

New to GitHub or command-line tools? Follow the beginner quick-start. It covers downloading the ZIP, installing Python and Git, setting up RSO, and trying a complete example. It also includes a request you can paste into a local coding agent to have it perform setup.

1. Read INSTALL.md for Windows, Linux, or macOS setup. An agent asked to install RSO should perform the installation and checks itself.

2. Point it at one project folder, then ask a real question:

   rso-context use --agent <current-agent-name> --path <bounded-project-folder>
   rso-context query "What export settings are required?" --agent <current-agent-name> --path <bounded-project-folder>
   

3. Read MANUAL.md for the workflow, examples, command reference, and troubleshooting. SKILL.md is the agent adapter, with references/protocol.md describing the packet contract.

Replace angle-bracket placeholders before running commands. use registers the folder, ingests its files, and returns a resume packet. query retrieves evidence for the actual task.

MCP access and searchable context are separate. Reaching a folder through MCP does not register or ingest it. Run rso_use once for a project that was never ingested; after that, query it directly. Re-ingest after file changes. Copies keep their own project identity; RSO does not copy or update their source files.

What you can do

Need	Workflow
Start with current project decisions	use, then query with the actual task
Hand work to another agent	Share the project files and database; use distinct agent names
Check where an answer came from	Inspect source spans and hashes; use explain for a saved packet
Pick up an edited requirement	Run use again; unchanged files are skipped
Separate suggestions from approval	Use propose; reserve validation for a named, authorized decision
Integrate through MCP	Run rso-context mcp --setup --client all, then call rso_query with the project folder; no startup call (rso_use only for a never-ingested folder)
Inspect the ledger locally	Run the optional read-only localhost viewer with admin

What's new in 0.11.0

• Smaller packets: the compact budget is 4 KB and the query limit 4, with per-file dedup, merged adjacent spans, capped claims and trimmed metadata, so packets keep inline evidence instead of filling up with metadata.
• No startup ritual: agents call rso_query directly. rso_use on a known project is a cheap freshness check with a small targeted refresh.
• Better recall per byte: stemmed, prefix-indexed search with AND-first query planning (schema revision 8, compatible with running 0.10 servers).
• Faster ingest: files are hashed outside the write lock, unchanged files cause no writes, vendored and generated files are skipped, and a query does at most one write transaction.
• New prune-orphans command (dry run by default) removes chunks of superseded file versions.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗