Agent-Skills.md

Agent Skills: Session Finder

Index and search Claude Code sessions using semantic embeddings (Gemini). Find past sessions by topic, relaunch the best match. Triggers on "find session", "which session did I", "relaunch the session where", "session about X".

UncategorizedID: glebis/claude-skills/session-finder

Repository

glebis/claude-skills
glebisLicense: MIT
37956

Install this agent skill to your local

pnpm dlx add-skill https://github.com/glebis/claude-skills/tree/HEAD/session-finder

Skill Files

Browse the full folder contents for session-finder.

Download Skill

Loading file tree…

session-finder/SKILL.md

Skill Metadata

Name
session-finder
Description
Index and search Claude Code sessions using semantic embeddings (Gemini). Find past sessions by topic, relaunch the best match. Triggers on "find session", "which session did I", "relaunch the session where", "session about X".

Session Finder

Semantic search across Claude Code sessions using Gemini embeddings.

Commands

Index sessions

python3 ~/.claude/skills/session-finder/scripts/session_finder.py index [--max-age-days 90]

Search

python3 ~/.claude/skills/session-finder/scripts/session_finder.py search "query" [--top 5]

Open best match directly

python3 ~/.claude/skills/session-finder/scripts/session_finder.py open "query"

Stats

python3 ~/.claude/skills/session-finder/scripts/session_finder.py stats

How it works

  1. Document extraction — deterministic, no LLM. Each session JSONL is parsed into a structured document:
    • away_summary events (pre-existing Claude recaps) if available
    • First user message (task description)
    • Follow-up user messages (condensed)
    • First assistant response
    • Tools used
    • Project name
  2. Embedding — documents are embedded with gemini-embedding-exp-03-07 via llm CLI
  3. Storage — SQLite at ~/.claude/session-finder.db
  4. Search — query is embedded, cosine similarity ranks all sessions, top match is the default to open

Workflow

When user asks to find a session:

  1. Run search with their query
  2. Present results with confidence scores
  3. Offer to resume the top match via claude --resume <id>