Find Session

Lists recent Claude Code sessions (or searches them) using the Get-ClaudeSessions.ps1 helper in this skill directory. The helper reads each session transcript under ~/.claude/projects/ and reports, per session: its session ID, original launch directory (where it must be resumed from), last-active time, git branch, and first prompt.

With -Match, it also scans every message in the transcript, so a topic raised in the middle of a long session is found — not just one named in the opening prompt.

Instructions

1. Run the helper. It lives next to this file. Pass $ARGUMENTS as the -Match filter when the user gave a search term; omit it to list the most recent sessions.

   # List the 15 most recently active sessions
   & "$env:USERPROFILE\.claude\skills\find-session\Get-ClaudeSessions.ps1" -Limit 15
   
   # Search everything — id, dir, branch, first prompt, and full conversation
   & "$env:USERPROFILE\.claude\skills\find-session\Get-ClaudeSessions.ps1" -Match "$ARGUMENTS" -Limit 15

   Always pass -ExcludeSession <current session id> when searching. The live conversation contains the search term by definition — the user just typed it — so without this it ranks first and buries the real answer. Take the id from the scratchpad path in your system prompt (the UUID segment).

   & "$env:USERPROFILE\.claude\skills\find-session\Get-ClaudeSessions.ps1" `
       -Match "$ARGUMENTS" -ExcludeSession '<this-session-uuid>' -Limit 15

   A full-corpus body scan takes ~1-3s. Add -HeadOnly to restrict the match to id / dir / branch / first prompt if you only want to identify a session by where it ran.

2. Display the returned objects as a clear list, best match first. The launch directory is the single most important field — it's where the user must cd before resuming. For each session show:

   <ShortId> — "<Summary>"
     Launch dir  : C:\work\12345_ShortTitle   ← cd here to resume
     Branch      : feature/12345_ShortTitle
     Last active : 2026-06-16 14:02  (0d 2h ago)
     Mentioned   : 2026-06-16 13:41  (16 times, last by assistant)
     Context     : "...<Snippet>..."
     Resume      : cd "C:\work\12345_ShortTitle"; claude --resume <full SessionId>

   Report LastMention when it's present — for "when did we last discuss X" that is the answer, and it can differ a lot from LastActive (a session touched today for unrelated work may have discussed the term days ago). Quote the Snippet so the user can confirm it's the thread they meant.

   The helper emits a ready-to-paste ResumeCmd. You can also format the raw objects directly, e.g. ... | Format-Table ShortId, MatchedIn, MatchHits, LastMention, Branch, Summary -AutoSize -Wrap.

3. Rank and qualify by MatchedIn. Results are already ordered so real mentions come first, but say which kind each is:

   MatchedIn	Meaning
   Head	id / launch dir / branch / first prompt
   Body	said in conversation by the user or Claude — the strongest signal
   Head+Body	both
   Tool	only inside tool traffic (a grep result, an API response, a file read) — usually incidental, e.g. a /daily run that swept the transcripts. Mention these as "possibly incidental", don't lead with them.

   MatchHits counts spoken occurrences; ToolHits counts tool traffic.

4. No matches / no sessions. If the helper returns nothing with a filter, say so and offer to list recent sessions instead (re-run with no -Match). If ~/.claude/projects doesn't exist, tell the user no sessions are tracked yet.

Notes

• Last-active comes from the transcript file's modified time — the real signal for "which session did I touch most recently."
• Subagent sidechain transcripts (agent-*.jsonl) and empty/aborted sessions are excluded automatically.
• The per-directory sessions-index.json files are unreliable placeholders in this environment, so the helper deliberately reads the transcripts directly.
• Don't hand-roll this with rg / Select-String over the .jsonl files. The helper exists because raw grep on transcripts is wrong in three ways, all observed in practice:
  • False positives — a bare number matches metadata. Searching 12345 "hit" a session that never discussed it: the digits sit inside the message UUID c1234567-.... Token counts and request ids collide the same way. The helper only matches inside message content.
  • Injected context — every session carries CLAUDE.md/MEMORY.md inside <system-reminder> blocks, so grepping any term those files mention matches nearly every session. The helper strips those blocks before counting.
  • Missed files — Select-String -List silently skipped a transcript holding 13 occurrences (very long single-line JSON). The helper's per-line IndexOf scan does not.
• Only lines whose raw text contains the needle get parsed as JSON, which keeps the scan fast (~1-3s over ~175 MB / ~95 transcripts, incl. a 46 MB one).