Grounded Claims make OpenWiki's docs degrade gracefully instead of going silently stale

OpenWiki is a CLI that generates and maintains a Markdown wiki for a codebase or personal knowledge base, using an agent to read sources and keep documentation current. The output stays in your repository as plain Markdown files you own — no proprietary platform, no lock-in.

The most substantive piece is Grounded Claims. Rather than tracking only when a Markdown file was last touched, OpenWiki tracks the specific factual propositions behind each page — behavior, architecture, data flow, security boundaries — each pointing to exact source evidence like a file and line range, with the version observed when the claim was established. Before any update run, it checks every persisted evidence version. A stale or unresolved claim forces work on its owning page even if the planner would otherwise skip it. Page completion is a durability boundary: claims are reconciled, evidence is rechecked, and the result is proven before that page's job is marked done. This means the wiki degrades gracefully as code changes rather than silently going stale.

The resumable page-job architecture follows a defined sequence: begin, submit plan, next page, submit page, finish. A durable ordered queue means an interrupted run can resume on the same checkout. Ephemeral CI runners start fresh after failure unless their workspace is preserved — an honest limitation the README states plainly.

Coding-agent integrations let Codex, Claude Code, or OpenCode drive the documentation lifecycle using their own authenticated model sessions and native repository tools. OpenWiki owns the durable queue, claims validation, and finalization; the coding agent owns research and authoring. The same five MCP operations are exposed to the host agent. This means you can reuse an existing Copilot or Claude subscription without provisioning separate inference credentials.

Thirteen model providers are supported, from OpenAI and Anthropic to Bedrock, Gemini Enterprise on Vertex AI, GitHub Copilot, OpenRouter, and any OpenAI-compatible endpoint including local servers like Ollama. Each provider section covers credential handling, known edge cases, and CI configuration in enough detail to be actionable.

Mermaid diagrams are validated after every run. A diagram that fails validation degrades to a plain text fence with an explanatory comment, and the next update repairs it — preventing broken renders from accumulating silently.

Personal brain mode ingests from eight connectors: Notion, Gmail, Slack, X, Hacker News, web search, local git repos, and custom MCP servers. The LangSmith connector is separate and works differently: it enriches a code wiki with runtime trace data pulled via the LangSmith SDK, so documentation reflects actual observed behavior rather than only static source analysis.

Telemetry is on by default but narrow: a single event per run recording the command, outcome, mode, provider, and connector names. File contents, model output, error messages, and repository names are explicitly excluded.