agentsclimarketplace

Recap

Skill noluyorAbi/claude-code-recap/skills/recap

Show recent Claude Code sessions across all projects, so the user can re-enter work after a reboot or context switch. Lists per session the absolute project path, a short summary, where the session stopped, last activity, turn count, git branch, model, and a ready-to-paste resume command; can search sessions by what was said in them, flag projects with uncommitted or unpushed work, and re-open every session at once in new terminal tabs. Use when the user asks "what was I working on", "which projects did I touch recently", "where did I leave off", "list my recent sessions", "how do I get back into that session", "open all my sessions again", "which session did I fix X in", "where do I still have uncommitted work", or invokes /recap. Reads only local files; the default run is instant and offline.From its SKILL.md

Install
npx -y skills add noluyorAbi/claude-code-recap --skill recap

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

What its file declares

Copied from the file, not written here

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

9.9 KB, as published. Nobody here has run it

recap: re-entry radar for Claude Code sessions

One command, full overview of recent sessions across every project on this machine. Built for the post-reboot moment: which projects, what was each session about, and the exact command to jump back in.

How to run it

Personal or project install (the skill directory is on disk next to this file):

python3 ~/.claude/skills/recap/recap.py [flags]

Plugin install (Claude Code exports the plugin root):

python3 "$CLAUDE_PLUGIN_ROOT/skills/recap/recap.py" [flags]

Nothing to install beyond Python 3 (stdlib only, no third-party packages).

FlagMeaning
(none)last 15 sessions, newest first, instant, no network
--since 7donly sessions active in the window (30m, 24h, 7d, 2w)
--project foofilter by substring of the absolute project path; separators are interchangeable (my_app, my-app, work/my_app)
--grep PATTERNonly sessions whose conversation matches PATTERN (regex, case-insensitive); the matching excerpt replaces the ↳ line
--dirtyannotate each row with its project's uncommitted and unpushed counts (read-only git calls)
--limit Nmax rows (default 15)
--jsonmachine-readable output with full session ids and resume commands
--no-tailhide the ↳ line that says where each session stopped
--smartreal one-sentence summaries via ONE claude -p haiku call (network, ~10s)
--pickchoose a row interactively, then cd + claude -r into it directly
--openopen EVERY listed session in its own new terminal tab and resume it there
--claude-flags "..."extra flags for each resumed claude (e.g. "--chrome --dangerously-skip-permissions"); also shown in the printed resume lines and --json
--terminal NAMEwhich terminal --open drives: iTerm2, Terminal, tmux, wezterm, kitty, gnome-terminal, konsole (default: auto-detected)
--yes / -yskip the --open confirmation prompt (required when stdin is not a TTY)
--dry-runwith --open: print the tabs and commands, open nothing
--colorforce ANSI colors when output is piped (e.g. inside Claude Code)
--plainno ANSI at all (also honors NO_COLOR / FORCE_COLOR env)

Usage examples

# after a reboot: what was I doing?
python3 ~/.claude/skills/recap/recap.py

# everything in one repo this week
python3 ~/.claude/skills/recap/recap.py --since 7d --project my-repo

# which session did I solve this in?
python3 ~/.claude/skills/recap/recap.py --grep "connection pool" --limit 5

# where did I leave work uncommitted?
python3 ~/.claude/skills/recap/recap.py --dirty --since 7d

# jump straight back into a session
python3 ~/.claude/skills/recap/recap.py --pick

# re-open yesterday's whole working set in terminal tabs, browser enabled, no permission prompts
python3 ~/.claude/skills/recap/recap.py --since 24h --limit 12 \
  --open --yes --claude-flags "--chrome --dangerously-skip-permissions"

# preview first, open nothing
python3 ~/.claude/skills/recap/recap.py --open --dry-run

# feed into other tooling
python3 ~/.claude/skills/recap/recap.py --json --limit 50

The --open flag

Restores a whole working set at once. It opens one new tab per listed session and types the resume command into it. Rules:

  • The session recap itself runs in is skipped automatically (matched on CLAUDE_CODE_SESSION_ID), so it never re-opens itself.
  • Sessions whose project directory no longer exists are skipped and reported, never opened.
  • Scope comes from the normal filters, so --limit / --since / --project decide exactly which tabs appear. Check with --dry-run before committing.
  • Without --yes it asks for confirmation; with piped stdin (Claude Code's Bash tool) it refuses to open unless --yes is passed.
  • --claude-flags is passed through verbatim to every tab. --dangerously-skip-permissions disables all permission checks in each opened session; the script prints a warning before doing it.
  • The terminal is auto-detected: tmux when running inside it (even on macOS), else iTerm2/Terminal on macOS via osascript, else wezterm, kitty, gnome-terminal or konsole. --terminal NAME forces one. Where none can be driven, --open says so and opens nothing; --open --dry-run needs no terminal and still prints the commands.
  • Every opened tab outlives claude: quitting a resumed session leaves a shell in that project instead of closing the restored window.

When Claude runs this skill

Run the script with Bash, ALWAYS with --color (tool output is piped, auto-detection would strip the colors; Claude Code renders ANSI). Add --since/--project if the user narrowed the question, then relay the table. Quote the resume command (cd <path> && claude -r <id>) for any session the user wants to re-enter. Only add --smart if the user explicitly asks for better summaries; it makes a network call.

Reach for --grep when the user is looking for a session by what happened in it ("where did I fix the connection pool"), not by when it ran. Reach for --dirty when the question is about unfinished work ("what did I leave half-done", "where do I still have uncommitted changes"); it runs read-only git commands, so it is safe but not instant.

Use --open only when the user explicitly asks to re-open or resume the sessions. Because the Bash tool pipes stdin, --open needs --yes; treat the user's request as the confirmation, and set the scope with --limit/--since so no unwanted session gets a tab. Pass --claude-flags only with flags the user named; state the permission-bypass warning in your reply whenever --dangerously-skip-permissions is among them.

Data sources and guarantees

  • ~/.claude/history.jsonl is the fast index (prompt text, timestamp, project path, session id). CLAUDE_CONFIG_DIR is honored when set.
  • ~/.claude/projects/<encoded>/<session>.jsonl transcripts are parsed only for the displayed rows (title, branch, model, turn count). Project paths always come from the cwd/project fields, never decoded from folder names (that encoding is lossy). Because claude -r only finds a session from the directory it started in, recap resolves the path to the first recorded cwd whose encoding matches the transcript's folder, so a mid-session cd never produces an unresumable command; a session with no cwd at all borrows one from a sibling transcript.
  • Summary preference: Claude Code's own ai-title line, else the first real user prompt, else the first slash command the session ran, else (no prompt). A transcript's user role also holds what Claude Code injects into it, and none of it counts as a prompt: command blocks, local command output, the body of an invoked skill, hook text, task notifications, ! bash runs and interrupt markers.
  • The ↳ line is where the session stopped: Claude's last reply when the run finished, the unanswered prompt when it was interrupted, whichever came last. --json always carries it as lastMessage / lastSpeaker, so --no-tail only quiets the table.
  • --grep searches the conversation. A transcript is rejected by a raw byte scan before it is parsed, using a word every match must contain, derived from the pattern only when that is provably safe (never across an alternation, a quantifier or whitespace).
  • --dirty runs git status --porcelain and one rev-list per project directory, cached and only for the rows actually shown. Not-a-repository, no upstream and no git installed all yield the same non-answer: no counts on the row.
  • --smart privacy: the only path that leaves the machine. It shells out to the local claude CLI once and sends, for the listed sessions only, each session's 8-character id prefix, its title (first 150 characters), and its first user prompt (first 300 characters). No file bodies, no transcript contents, no other sessions. If claude is not on PATH, --smart is skipped with a warning and the offline summaries are used.
  • Turn count groups assistant streaming chunks by message id and is labeled approximate.
  • Broken or partial JSONL lines are skipped, never fatal. Times shown in the local timezone.
  • Display: day-grouped timeline (Today green, Yesterday amber), 14-day activity sparkline in the header, per-project colored dot (stable hash), four-level gray hierarchy with one accent color, per-session resume line indented with spaces only (safe to copy), and a small footer with a GitHub-star CTA and author link (clickable OSC 8 hyperlinks on supporting terminals). Falls back to plain text when piped or when NO_COLOR is set.
  • Read-only: the tool never writes or deletes anything. It sees only sessions still on disk (Claude Code prunes old ones). --open, --pick and --dirty are the only side-effecting paths: --open drives a terminal, --pick execs claude -r in the chosen directory, --dirty runs read-only git commands in the project directories. None of them writes or touches session data.

Optional shell alias

alias recap='python3 ~/.claude/skills/recap/recap.py'

What ships with it: 1 file

29.8 KB alongside SKILL.md, 1 of them executable

Gives 0 of the 12 instructions most pr commit review skills give

Counted across 1,055 of the 1,911 authors here whose files we hold, read 2026-09-06

  • Use conventional commit message formatin 150 of 1055, across 145 files
  • Announce skill usage at startin 78 of 1055
  • Use imperative mood for commit descriptionsin 54 of 1055, across 51 files
  • Add directory to gitignore if not ignoredin 52 of 1055, across 41 files
  • Use imperative mood for commit subjectin 52 of 1055
  • Run tests to verify clean baselinein 42 of 1055, across 32 files
  • Push branch to originin 40 of 1055, across 38 files
  • Verify worktree directory is ignored before creationin 39 of 1055, across 32 files
  • Delete branches after mergingin 38 of 1055, across 30 files
  • Create worktree with new branchin 37 of 1055, across 32 files
  • Wrap body text at 72 charactersin 36 of 1055, across 34 files
  • Auto-detect and run project setupin 35 of 1055, across 27 files

Said here and by no other author read

  • Always use the color flag when piping output
  • Relay the session table to the user
  • Quote the resume command for requested sessions
  • Use the smart flag only if explicitly requested
  • Use the grep flag to search session conversations
  • Use the dirty flag to check for uncommitted work

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once.

Keep looking

Skills are one crate of 613,373. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.