History Find — AI skill for Claude Code

history-find

**English** · [中文](README.zh-TW.md)

A [Claude Code](https://claude.com/claude-code) skill that finds a past Claude Code or Codex conversation from a vague memory — when you remember *what you did* but not the words you used.

The problem

You remember designing something months ago — a batch import, a refactor plan, a pricing model — but you can't remember the exact wording you used, so keyword search finds nothing. Meanwhile the work is spread across a parent session, two forks, and a subagent transcript.

What it does

Instead of trusting one search tool, it works in layers:

fuzzy memory
    -> metadata narrowing        (dates, projects, titles, paths)
    -> content search            (actual prompts and replies)
    -> model reranking           (task, artifacts, actions, sequence)
    -> chain reconstruction      (parent + forks + subagents as one piece of work)
    -> 1-5 evidence-backed candidates

It returns a ranked shortlist with a confidence label, the evidence behind each match, and the exact session paths — plus an explicit list of what it *did not* check, so a miss is never reported as "doesn't exist".

What makes it different

• Keeps evidence layers separate. Metadata proving a session exists is not the same as message content proving what was discussed. A model-generated title is a clue, not proof.
• Never reports a false zero. An empty full-text table means "this index is empty", not "your history is empty" — the skill says so and keeps looking.
• Reconstructs work chains. One task often spans a parent session and its forks. It returns the parent where your decisions live, not just the child that happened to match a keyword.
• Read-only by default. It does not resume, rename, export, or delete anything, and it does not send transcript content to external services.

Requirements

Works with whatever you have; it degrades instead of failing:

• [claude-log](https://github.com/badlogic/lemmy/tree/m