recall

name	recall
description	YOU MUST use this skill for anything involving past Claude Code conversations in ~/.claude/projects/: when the conversation starts with 'this session is being continued from a previous conversation that ran out of context', or to recover context from a previous conversation, or to discover/catalog which conversations exist (by date, project, message count), or to search and read past conversation logs. One entry point for cataloging, searching, and rendering Claude Code JSONL transcripts.
metadata	{"category":"productivity","requires":{"bins":["uv","rg"]}}

recall — Query Claude Code Conversations

One entry point for working with past Claude Code conversations stored as JSONL in ~/.claude/projects/. This file is a router: find your task below, then read the matching reference for full command syntax.

Routing table

Task	Reference	Tool
Find which conversations exist (filter by date, project, message count); catalog/stats	references/discovery.md	scripts/query_conversations.py
Read or recover a specific conversation; line-range search within one	references/rendering.md	scripts/VCC.py
Trivial in-session "did I already mention X?" lookup, zero-dependency	references/fast-path.md	rg

Typical pipeline: discover with the catalog, then read with VCC. Examples are in references/discovery.md.

/recall — Recover context from a previous conversation

Trigger: The conversation opens with a context-continuation summary ("This session is being continued from a previous conversation..."), or the user says /recall.

Action:

1. Find JSONL filename — The continuation message ends with read the full transcript at: <path>.jsonl. Find that JSONL filename. If no JSONL path is present (e.g. the user invoked /recall manually without a continuation header), use the /searchchat method to locate the JSONL before proceeding to steps 2–3.

2. Follow VCC rendering workflow — Follow the full workflow in references/rendering.md (compile → read .min.txt → --grep → jump to .txt) with the JSONL file found in step 1. The summary is lossy — only the original JSONL is authoritative.

3. Verify against current state — After recovering the conversation, cross-check with reality:

   • REALLY read files — Read all files referenced in the conversation with your Read tool. Pre-loaded content from system-reminders or prior conversation turns does NOT count — it may be truncated, stale, or lossy. You must issue a fresh Read call for each file, even if you believe you already have its content. 99.97% of the time, the user has externally modified nearly every file between sessions. If you skip reading and miss their external changes, your failure rate is effectively 100%.
   • Key details — Identify specific values, paths, configs, and logic mentioned in the conversation. Compare against the actual files to catch drift or errors.
   • Understand the journey — Trace the user's intent, decision sequence, and direction changes. Understand not just what was done, but how and why they got there.

/readchat — Read a specific conversation log

Trigger: The user wants to read a specific JSONL conversation log.

Action: Follow references/rendering.md with the user-specified file.

/searchchat — Search across conversation logs

Trigger: The user wants to search across conversation logs in ~/.claude/projects/.

Action: For corpus-wide filtered discovery use references/discovery.md; for line-range content search within conversations use VCC --grep per references/rendering.md; for a quick zero-dependency check use references/fast-path.md.