"] \ [--org-id "<org-id>"] ``` For the literal ref `latest`, replace `--host auto` with the explicit current host. NEVER pass `--conversation-id`. Omitted, Claude uses the session id and Codex/Cursor use the same host-prefixed id as automatic capture. That keeps one conversation per session and makes re-imports incremental. A fresh id would split the session's memory. If nothing new lands, automatic capture already did its job; do not work around that with a new id. Very large transcripts are AUTO-CHUNKED (default threshold ~3.5MB): the script sends disjoint slices sequentially under one conversation_id and waits for each slice's extraction (the session gist folding forward) before the next — payloads beyond ~8MB fail server-side as one shot, so never disable chunking for huge sessions. This is slow but unattended; just let the command run. 3. Report back the returned `conversation_id`, `path` (should be `"agentic"`), `messages_received`, and `scope`, plus the platform from the script's `source platform :` line (the response does not echo it). Tell the user: - **where its memory landed** — their personal memory, with the transcript in their Sessions view; not the repo's brain; - extraction runs in the background (allow several minutes for large sessions) — as it completes, `list_sessions` lists the session, `read_memory(<conversation_id>)` opens it, and `search_memory(kind="episode", session_id=<conversation_id>)` finds its gist and episodes. Don't promise facts or directives from it: an import creates none; - re-importing the same session later is **incremental** (only new records are processed; the gist folds forward — nothing duplicates). 4. If the script prints an auth error, no setup is needed — it opens the browser ONCE for approval and then mints a personal access key (`mhk_…`) that later runs and the capture hooks reuse without a browser. `/memhub:login` does the same thing deliberately if you would rather provision it up front. This is the plugin's own credential, not the `/mcp` connector's, so being connected in `/mcp` will not satisfy it. If it can't find the session id, ask the user for the transcript path. Never use curl or raw HTTP; never pass transcript content as tool arguments.

--- description: Use when the user wants a past Claude Code, Codex or Cursor session saved into their memory because automatic capture missed it — a session from before the plugin was installed, one whose upload failed, or one from a host without capture hooks (e.g. "import this session into memhub", "save session abc123 to my memory", "upload yesterday's conversation"). Saves it to the user's own memory, not a team brain, at any size. argument-hint: " [title...]" allowed-tools: 'Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/capture.py" import *), Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/capture.py" list *), mcp__plugin_memhub_memhub__list_orgs, mcp__plugin_memhub-staging_memhub__list_orgs' --- **Plugin root:** Resolve this skill's plugin root once: it is the ancestor of this file containing the `scripts/` directory and a plugin manifest (`.claude-plugin/plugin.json` or `plugin.json`). A trusted host variable such as `CLAUDE_PLUGIN_ROOT` or `CURSOR_PLUGIN_ROOT` may already point there; use it only when it resolves to that same ancestor. Substitute the resulting absolute path as `` below; do not infer it from the workspace cwd. Commands show `python3`; on native Windows use `py -3`. Import a past Claude Code, Codex, or Cursor session into the user's own MemHub memory on demand — for a session automatic capture missed (it predates the install, its upload failed, or its host ran without capture hooks). The unified capture script locates the host session, normalizes it when needed, and ships it to the `import_conversation` MCP tool — **do NOT call the MCP tool yourself and do NOT read or paste transcript content**; sessions can exceed a million tokens and the script handles any size without putting the transcript in model context. This is a terminal operation. Arguments: `$ARGUMENTS` - First token = a session id or native session path (required). Paths and bare ids are host-detected. A bare id is accepted only when exactly one of Claude, Codex, or Cursor owns it; cross-host collisions are refused. - Remaining text = an optional conversation title. - If invoked without arguments (e.g. the user said "import this session"), ask which session they mean. For "this/the current session", determine the current host and run `capture.py list --host --limit 20`; select the current session only when its id or cwd is unambiguous, otherwise ask. The literal ref `latest` always requires an explicit current host. Do exactly this: 1. **It goes to the user's personal memory — never to a brain.** A session is the author's: the transcript lands in their Sessions view, and its **gist** and task **episodes** (never facts, never directives) land in their personal memory. The repo's room is not an option, and neither is a brain the user names: the import takes no brain. If the user asked to "put it in a brain", say so plainly and offer what does reach a brain — saving a document with `/memhub:save-artifact`, or `/memhub:handoff-session` to pass the work to a teammate. For an account in several orgs, `--org-id` picks whose personal memory (`list_orgs` names them); omitted, the default org. 2. Run the import in the terminal — one command, substitute the real values: ```bash python3 "/scripts/capture.py" import \ --session "" --host auto [--title ""] \ [--org-id "<org-id>"] ``` For the literal ref `latest`, replace `--host auto` with the explicit current host. NEVER pass `--conversation-id`. Omitted, Claude uses the session id and Codex/Cursor use the same host-prefixed id as automatic capture. That keeps one conversation per session and makes re-imports incremental. A fresh id would split the session's memory. If nothing new lands, automatic capture already did its job; do not work around that with a new id. Very large transcripts are AUTO-CHUNKED (default threshold ~3.5MB): the script sends disjoint slices sequentially under one conversation_id and waits for each slice's extraction (the session gist folding forward) before the next — payloads beyond ~8MB fail server-side as one shot, so never disable chunking for huge sessions. This is slow but unattended; just let the command run. 3. Report back the returned `conversation_id`, `path` (should be `"agentic"`), `messages_received`, and `scope`, plus the platform from the script's `source platform :` line (the response does not echo it). Tell the user: - **where its memory landed** — their personal memory, with the transcript in their Sessions view; not the repo's brain; - extraction runs in the background (allow several minutes for large sessions) — as it completes, `list_sessions` lists the session, `read_memory(<conversation_id>)` opens it, and `search_memory(kind="episode", session_id=<conversation_id>)` finds its gist and episodes. Don't promise facts or directives from it: an import creates none; - re-importing the same session later is **incremental** (only new records are processed; the gist folds forward — nothing duplicates). 4. If the script prints an auth error, no setup is needed — it opens the browser ONCE for approval and then mints a personal access key (`mhk_…`) that later runs and the capture hooks reuse without a browser. `/memhub:login` does the same thing deliberately if you would rather provision it up front. This is the plugin's own credential, not the `/mcp` connector's, so being connected in `/mcp` will not satisfy it. If it can't find the session id, ask the user for the transcript path. Never use curl or raw HTTP; never pass transcript content as tool arguments.