[Bug]: Imported Claude Code threads drop the user's renamed title (custom-title / customTitle) #12276

Version: T3 Code 0.0.43-nightly.20260917.1866, macOS.

Summary

Claude Code stores a user-set thread title in its transcript as a custom-title record. The importer never reads it — it reads only aiTitle — so any Claude session the user has explicitly renamed imports under a title derived from the first user message instead of the name the user chose.

Steps

1. In Claude Code, rename a session (e.g. to Refactor auth module).
2. Import that project's sessions into T3 Code through onboarding.
3. Look at the imported thread's title.

Actual: the thread is titled from the first user record. When the session's first message is a slash command, the title becomes the raw wrapper — e.g. <command-name>/effort</command-name>.

Expected: Refactor auth module — the title the user set in Claude Code.

Cause

Claude Code writes two distinct title records into ~/.claude/projects/<project>/<session>.jsonl:

{"type":"ai-title","aiTitle":"...","sessionId":"..."}         // auto-generated
{"type":"custom-title","customTitle":"...","sessionId":"..."} // the user's rename

AgentSessionScanner.ts handles only the first:

• TranscriptRecord declares aiTitle but not customTitle (L118), so customTitle is dropped at decode.
• The only title assignment from a record is if (record.aiTitle?.trim()) title = record.aiTitle.trim(); (L407).
• Everything else falls through to the first-user-message derivation at L501.

Worth noting: shouldRetainDecodedRecord (L514-526) already keeps the record — a custom-title record carries sessionId, so it satisfies record.sessionId !== undefined. The record survives the read and is then discarded purely for lack of a schema field.

In a sample of local transcripts, 53 files carried customTitle against 8 carrying aiTitle — so for renamed sessions this is the common case, not an edge one.

Suggested fix

Add the field to TranscriptRecord and prefer it over the generated title:

customTitle: Schema.optional(Schema.String),

if (record.aiTitle?.trim()) title = record.aiTitle.trim();
if (record.customTitle?.trim()) title = record.customTitle.trim();

Ordering matters: Claude Code writes both kinds, and an explicit user rename should win over the generated one. Both appear repeatedly through a transcript, so last-write-wins on the existing linear scan yields the current title without extra bookkeeping.

Related

#10513 is the Codex-side of the same first-user-message derivation. This one is narrower: for Claude sources a user-authored title already exists inside the transcript and is being dropped.