wiki-agent

A query-driven, targeted ingest tool that pulls context from one specific AI agent raw conversation history into your current session. It is built for cross-referencing: while working in one agent you can retrieve how you solved a problem in another. Unlike bulk history ingest, it finds only the sessions about a specific topic, extracts the relevant portions, distills them into an Obsidian wiki, and returns a synthesized answer you can act on immediately.

It is invoked through commands that select the target agent and an optional query: /wiki-claude, /wiki-codex, /wiki-hermes, /wiki-openclaw, /wiki-copilot, and /wiki-pi. With no query it falls back to recent-sessions mode, ingesting the last five unprocessed sessions from that agent. Before starting it resolves configuration (an inline name override, walking up the working directory for a .env file, then a home config) to find the Obsidian vault path, reads the vault manifest to know what is already ingested, and warms context from hot.md.

The workflow locates each agent history root using documented default paths and config overrides, builds a session inventory from the cheapest available index rather than opening every file, and scores sessions against the query using name/title match, working-directory or project match, and recency before selecting the top three to five. It then extracts only the relevant blob per session using agent-specific strategies, for example grepping Claude JSONL and pulling a surrounding window, filtering Codex rollout event and response items, reading short Hermes memory files fully, or scanning OpenClaw MEMORY.md sections. The outcome is a focused knowledge transfer between agents rather than a full archive.