Ru documentation¶
Current operating documentation and accepted architecture decisions live here. Isolated historical verification evidence records old snapshots without making them current operating instructions. Completed implementation plans and design drafts belong in Git history.
Open plans may sit under docs/plans/ with a status: pending or
status: in-progress marker. Generated plans and specs under
docs/superpowers/ must be removed before verification/integration.
test_completed_implementation_archives_are_not_kept_in_the_checkout checks
these locations so obsolete drafts cannot accumulate in the checkout.
Guides¶
Task-oriented pages for people using Ru; start here before the reference material below.
- Install Ru
- Connect a provider and run your first task
- Permissions and approvals
- Using the terminal UI
- Teach Ru your project
- Subagents
- Troubleshooting
Current system¶
- Windows distributions and updates
- Architecture and ownership
- Component implementation: Host, Harness, Broker, GUI, TUI, Providers
- Domain language
- Context management and checkpoints
- Database history, backup and maintenance
- Current runtime behavior
- Long runs and their ceilings
- Agent performance diagnostics and offline benchmark
- The stdio RPC bridge
- TypeScript SDK for the bridge
- TUI known limitations and debugging
- Provider catalogs and model-specific reasoning
- Operations and verification
- Agent tools and truthful execution
- Browser DOM and DevTools automation through MCP
- Skill setup, invocation and continuity
- Workspace project memory
- Coding-agent capability and acceptance map
- Custom Terminal/SWE/Everyday coding benchmark
- Logging and diagnostics
- Python dependency ownership
- Python runtime boundary
- Shared workspace coordination
- Driving Ru without a provider
- ASP.NET Core public route contract
- Canonical route data
- Contributing
- Documentation site: how these pages are published
- Agent guidance
Decisions¶
- A provider route table for Models.dev providers
- Native continuation under the same model and a window-sized handoff
- No Continue gate; a deny settles the call, not the run
- Advisor: the main model consults a second pinned model on any provider (proposed)
- Self-evolution runs: Ru proposes changes to its own source (proposed)
- Detached background subagents and manual PR integration (proposed)
- Advisory verification of declared checks
- Tool descriptor table
- Broker-owned language-server tools
- Explicit pinned GitHub skill package installation
-
Explicit MCP Streamable HTTP behind the existing Broker authority
- Bounded programmatic coding tools
-
Pinned coding tool profiles (legacy compatibility superseded by 0029)
- MCP servers through the Broker behind one
mcptool - Bounded session-run handoff across model changes
- Full access is literally full
- Always allow parallel sessions in one workspace
- Product packages and explicit updates
- Architecture decision records
- ASP.NET Core Public Host
- Typed Worker protocol v2 (superseded by 0012)
- ADR 0003 reserved
- Canonical Ru data location
- Local conversational-model retirement
- Structured logging and activity
- Capability profiles
- Shared workspace coordination
- Broker execution authority (superseded in part by 0014)
- Product logic outside WinUI code-behind
- Model autonomy and workspace trust boundary
- In-process model runner
- SQLite business persistence
- In-process execution (terminal sandbox portion superseded)
- Retire terminal OS sandbox
- Standard workspace Python files
- Workspace terminal policy
- Approval patterns and durable folder grants
- Computer use through the Broker
- Provider-hosted web search
- Named custom providers
Git history is the recovery source for removed dated plans, specifications, reviews and verification records.
Agent permissions use the canonical, session-scoped plan / standard /
full_access profiles defined by the Protocol catalog and described in
tool policy; CONTEXT.md defines the
domain terms. This index does not restate the matrix. Under
full_access everything else runs automatically, and structured protected
paths are allowed like any other target. Workspace trust is explicit once per
canonical identity: the
GUI confirms it and CLI run/chat use --trust-workspace. Full Access
requires the GUI gate or CLI --confirm-full-access.
The canonical profiles do not synthesize web_approval_required; that legacy
reason remains available to an explicit ask rule. Current approval reasons also
include external_directory_required, terminal_external_cwd_requires_approval,
terminal_timeout_requires_approval, rule_asks, and
terminal_protected_path_requires_approval (never reusable as allow_always).
Legacy command-classification reasons are retained for recorded activity.
allow_always remembers a bounded pattern as a durable project rule: a
terminal command prefix (resolved executable plus at most two leading non-flag
argv tokens, with the exact launch directory), or an approved external folder
as a path pattern covering its subtree (ADR 0018).
Flag-first invocations and shell-mode scripts stay exact, and rule_asks /
terminal_protected_path_requires_approval never become a reusable rule.
A folder grant can stay session-scoped (the Broker grant) or be remembered for
the project; workspace trust follows the canonical identity. Terminal children in both standard and full_access run as the
current Windows user without OS filesystem or network confinement. Standard authorizes workspace launch directories under
ADR 0017; the structured protections hold in plan and standard. Hard timeout,
output cap, cancellation, process-tree cleanup, credential storage and
diagnostic redaction remain enforced.
- Subagents: coding/review task browser, explicit model requests, main-agent controls and guarded integration.