# Resume > Find a past conversation and resume it in the current worktree, including sessions that ran in another agent or another git worktree. Use when the user wants to continue, pick up, or reopen earlier work. - Skill: `erwint/resume` (Agent Skill) - Install (CLI): `npx skillmds@latest add erwint/resume` - Raw SKILL.md: https://api.skillmd.com/api/skills/erwint/resume/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: erwint (https://skillmd.com/u/erwint) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/erwint/resume --- # Resume Help the user resume a previous conversation — including one that ran in a different git worktree, or in a different agent — without making them switch directories or hunt for a session id. Everything below is `remaimber`, a CLI on PATH. If this agent also exposes the remaimber MCP tools (`find_context`, `get_segments`, `get_session`), prefer them for locating the passage: same data, structured results. Hosts namespace MCP tools by server, so they appear as `mcp__remaimber__find_context`, `mcp__remaimber__get_segments` and so on. The bare names used below are the short form; if a lookup for the bare name finds nothing, the tool is probably there under the namespaced one. ## How sessions are identified remaimber captures a durable identity for every session at start: `repo_id` (stable across all worktrees of a repo) and `subpath` (the monorepo sub-project). This is why a session started in a now-deleted temporary worktree is still findable. Each session also records the agent that produced it. That decides how it opens: | agent | native resume | |--------|-------------------------------| | codex | `codex resume ` | | claude | `claude --resume ` | | pi | `pi --session ` | `remaimber resume ` prints the right one — use it rather than assembling the command yourself, since Claude Code sessions also need their transcript linked under the current directory first, which that command does. ## Resume the part, not the whole thing Sessions here run to thousands of messages. When the user names a topic — "resume X, but only the part where we did Y" — do **not** load the whole conversation. Find the passage instead. ``` remaimber resume --match 'the part where we set up a mail relay on the nas' # any conversation remaimber resume b2bd8168 --match 'smtp relay' # within one remaimber resume b2bd8168 --segments 4 --print # explicit part remaimber resume b2bd8168 --since 2026-08-06T11:18 --until 2026-08-06T11:45 # by time ``` The MCP equivalents are `find_context` (topic in plain words, searches every conversation) and `get_segments` (`session_id` plus `match`, locates the passage inside one conversation; without `match` it lists every segment so you can choose by summary). Both return summaries first and messages only on `include_messages: true` — read the summaries, decide which passage is right, *then* pay for the text. Both also return alternatives: a one-word topic is often ambiguous, so if the top passage looks wrong, check the runners-up before widening the search. ## Steps 1. Identify the target: - Topic described, session unknown → `remaimber resume --match ''`. - Session known, part wanted → `remaimber resume --match ''`. - Nothing specific → run `remaimber resume` with no arguments to list this repo's sessions across all worktrees and agents, newest first. 2. Present the top candidates: session id (first 8 chars), agent, subpath, branch, time span, and summary. If one clearly matches, pick it; otherwise ask. 3. Decide with the user how to resume: - **Partial (usual case):** load the chosen passage's messages, summarize what was done and what's unfinished, and continue here. No restart, and it works even when the session came from another agent. The branch a session ran on is a fact about the past: check where it stands now (`remaimber resume ` prints this) before switching. It may be checked out in another worktree, where `git checkout` refuses outright, or already merged here, where switching steps back from the work you have. - **Native full resume:** run `remaimber resume ` and hand the user the command it prints. Check the branch first. Note this always resumes the *whole* session — partial resume is a way of reading part of it as context, not a smaller transcript. 4. Respect the liveness warning: if `remaimber resume` reports the session looks **live** elsewhere, do NOT resume it — warn that it would corrupt the transcript, and suggest closing that session first. ## Notes - Always prefer the session's own `cwd`/identity for any path you show — never reverse-engineer it from the project key (that encoding is lossy). - If a session has no segments it was never summarized; `get_segments` will say so. Fall back to `get_session`, or suggest `remaimber summarize `.