AMindToThink/finding-old-chats

Finding Old Chats

Overview

Claude Code stores every session as a JSONL file under ~/.claude/projects/<dashed-cwd>/<uuid>.jsonl. When the user asks for "chats from <date>", they mean their wall-clock day, not UTC — and the timestamps inside the JSONL are UTC. A naive timestamp.startswith("YYYY-MM-DD") filter silently drops sessions the user clearly considers part of that day.

This skill bundles a script (scan_chats.py) that does it right: convert local day → UTC window, then test each chat by interval-overlap, not substring match.

The trap this skill exists to prevent

Real failure on 2026-05-07: user asked to review May 1 chats. The agent filtered with startswith("2026-05-01"), missed two sessions whose first timestamp was 2026-05-02T00:32Z and 2026-05-02T00:43Z — both are May 1 evening in Central US time, but every timestamp inside them starts with 2026-05-02.

The mirror trap: a session whose first timestamp is 2026-05-01T03:55Z is April 30 evening local, not a May 1 chat — but it would match the naive UTC filter.

Either way: do not use UTC-substring filtering. Use the script.

Where transcripts live

~/.claude/projects/<dashed-cwd>/<session-uuid>.jsonl

Directory name is the absolute working-directory path with / → -, leading dash. Example:

• cwd: /home/cs29824/matthew/icl-diversity
• dir: ~/.claude/projects/-home-cs29824-matthew-icl-diversity/

Each JSONL line is {"type": "user"|"assistant"|"summary"|..., "timestamp": "2026-05-01T04:25:36.489Z", "message": {...}}. Inside message.content, content can be a string or a list of {type: text|tool_use|tool_result, ...} parts.

Quick reference

# Default: use system timezone
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01

# System TZ is UTC but you think in local time — pass an IANA name
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01 --tz America/Chicago

# Don't know the user's TZ — pad the window
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01 --buffer-hours 12

# Different project
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01 --project /path/to/project

# Show first user message + tail of each chat
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01 --tz America/Chicago --full-msgs

# Just the chats whose last assistant turn looks like an unanswered question
~/.claude/skills/finding-old-chats/scan_chats.py --date 2026-05-01 --tz America/Chicago --unfinished-only

The script prints the UTC window it actually scans, so the caller can sanity-check the timezone.

Picking the right time-window arguments

digraph tz_decision {
    "User says \"chats from <date>\"" [shape=box];
    "Do you know user's IANA TZ?" [shape=diamond];
    "Pass --tz <name>" [shape=box];
    "Is system TZ UTC?" [shape=diamond];
    "Default (system tz)" [shape=box];
    "Pass --buffer-hours 12" [shape=box];

    "User says \"chats from <date>\"" -> "Do you know user's IANA TZ?";
    "Do you know user's IANA TZ?" -> "Pass --tz <name>" [label="yes"];
    "Do you know user's IANA TZ?" -> "Is system TZ UTC?" [label="no"];
    "Is system TZ UTC?" -> "Pass --buffer-hours 12" [label="yes"];
    "Is system TZ UTC?" -> "Default (system tz)" [label="no"];
}

--buffer-hours 12 casts a wider net (36-hour UTC window for one local day) and accepts some over-reporting in exchange for never missing a chat. Skim the printed UTC bounds and ignore obvious outliers.

Detecting unfinished sessions

A chat is heuristically "probably unfinished" when:

• The last assistant message ends with ? in its final 500 chars or ends with ] (cut off mid-tool_use), AND
• No real user message follows. A "real" user message is not a [tool_result...] echo, a <system-reminder>, a <command-name> slash-command marker, a <bash-input>/<bash-stdout> echo, or a <local-command-caveat> echo.

scan_chats.py --unfinished-only flags these. Expect false positives:

• Rhetorical-question closers ("…or are we done here?", "Is this surprising?")
• Long sessions where the trailing ? is hundreds of chars before genuine resolution

Always read the actual last messages before reporting "unfinished" to the user. The flag is a triage shortcut, not a verdict.

Common mistakes

• ❌ timestamp.startswith("YYYY-MM-DD") — misses sessions that started after local-midnight-but-before-UTC-midnight.
• ❌ find -newermt on the JSONL files — uses file mtime, which moves whenever the IDE re-opens or re-writes them. Use timestamps inside the file.
• ❌ Treating system-reminders as the "first user message". The user's first real message is often several entries in. Use is_real_user_msg filtering.
• ❌ Reading 10 MB JSONLs with the Read tool — single-shot read can be near the file-size limit and is wasteful. The script streams line-by-line.
• ❌ Trusting the JSONL's last-message timestamp as "when the session ended". A session that started May 2 and was last touched May 7 may have been pending the entire week — first/last bracket the activity, the gap matters.
• ❌ Trying to compute "did this session end on action X" from the chat alone. Cross-check git log and current repo state for the work that was supposedly done.

Implementation

See scan_chats.py in this directory. ~200 lines, stdlib only (zoneinfo from Python 3.9+).