| name | t3ex-find-thread |
| description | Find a past T3 Code, Codex, or Claude Code conversation by what was said in it, and return its id and resume command. Use when the user is looking for an older thread, remembers a topic but not the id, or asks which session covered something. |
| disable-model-invocation | true |
| argument-hint | <topic or phrase> [repo] [recent] |
Find an Older Thread
Search every local transcript store at once — T3 Code, Codex, and Claude Code —
and return the thread the user is thinking of, with the command that resumes it.
The usual next step is t3ex-fork-thread, which continues a T3 Code thread
under a different model. This skill is how you get the id to hand it.
Run the search
find_thread.py sits next to this file:
python3 find_thread.py "<pattern>"
Python 3.10+ and ripgrep. The pattern is a case-insensitive regex, so
indicator|spinner works.
| Flag | Effect |
|---|
--source t3code|codex|claude | Limit to one store, repeatable |
--repo TEXT | Only threads whose working directory contains TEXT |
--since N | Only threads active in the last N days |
--limit N | Results to show (default 10) |
--scan-limit N | Transcript files to open per store, newest first (default 40) |
--include-deleted | Include T3 threads deleted in the UI |
--paths | Print the backing file for each hit |
--json | Machine-readable output |
Results are newest first — when someone asks for "that older thread", they
almost always mean the most recent one that matches.
Reading the output
3. [t3code] Start Localhost Dev Server
when 2026-07-25 08:33 · 16 matches · codex/gpt-5.6-sol
repo /Users/madda/dev/active/Maestro3
said This site can't be reached localhost:3001: Connection refused…
also as codex 019f97a1-e620-7a92-abc6-047811367bfb
resume /t3ex-fork-thread ce419dc7-4b8f-4d7c-b438-cdeed11ba120
also as means the same conversation exists in a second store. A T3 Code thread
run on Codex writes a rollout file too; the script folds them into one result and
reports the T3 Code id, because that is the one t3ex-fork-thread takes.
Narrowing
A bare word matches far more than you expect. Every session's system prompt
carries the full list of installed skills, so a term like beads or chrome
hits thousands of transcripts that never discussed it.
Start narrow and widen only if nothing lands:
- An exact phrase the user remembers, or a distinctive identifier — an error
string, a file name, a port number.
- Add
--repo when they name a project. This is the single most effective
filter.
- Add
--since when they say "yesterday" or "last week".
- Only then loosen to
term1|term2.
If the script reports transcripts went unread, it means the pattern was too
broad — narrow it rather than raising --scan-limit, which just reads more
irrelevant files.
Confirm before answering
Match counts include assistant replies and tool output, not just what the user
typed, so a high count is not proof the thread is about the topic. Before naming
a winner, look at the actual text:
python3 find_thread.py "<pattern>" --paths --limit 3
rg -n -i "<pattern>" "<the file path>" | head -20
Quote the line that proves the match. If two threads are close, name the best
one and mention the runner-up in a sentence — do not make the user guess.
If nothing matches, say which stores were searched and what pattern was used.
Do not silently widen to a vaguer term and present those results as the answer.
Where the transcripts live
| Store | Location | Id |
|---|
| T3 Code | ~/.t3/userdata/state.sqlite | thread UUID |
| Codex | ~/.codex/sessions/YYYY/MM/DD/rollout-<ts>-<uuid>.jsonl | trailing UUID |
| Codex archive | ~/.codex/archived_sessions/ | trailing UUID |
| Claude Code | ~/.claude/projects/<path-slug>/<uuid>.jsonl | filename |
~/.codex-profiles/*/sessions and ~/.claude-profiles/*/projects are searched
too, but on this machine they are nearly empty — the real history is under
~/.codex and ~/.claude. Claude project folders encode the repo path with
dashes: /Users/madda/dev/active/Maestro3 becomes
-Users-madda-dev-active-Maestro3.
Escape hatch — record shapes
For anything the script does not surface, read the transcripts directly.
Codex rollouts. First line is session_meta, carrying session_id, cwd,
and originator — t3code_desktop means the session was driven by T3 Code
rather than the CLI. Conversation lines are response_item with
payload.role and payload.content[].text.
Claude transcripts. ai-title holds a generated title, last-prompt the
most recent prompt, and user/assistant the conversation. user records also
carry cwd, sessionId, gitBranch, and timestamp.
T3 Code. See the schema table in t3ex-fork-thread.
Three traps:
- The first user message is never the real prompt. Both harnesses inject
# AGENTS.md instructions, <recommended_plugins>, <environment_context>
and similar ahead of it. Skip those wrappers.
- Claude subagent turns are inline. Filter
isSidechain: true out or a
subagent's prompt reads as the user's.
~/.codex/sessions/index/catalog.jsonl is partial — about 123 of 4,900
rollouts. Useful for metadata, useless as a search index.
Related
t3ex-fork-thread continues a T3 Code thread once you have its id.
