Date imported agent sessions by their last message instead of file mtime, and allow importing after onboarding

[SriCharanK2002]

Follow-up to #21755. The 30-day import window stays as it is; this is about two things around it.

1. Date imported sessions by their last message, not the file's modified time

Session import treats a transcript's file mtime as its last activity (lastActiveAtMs = transcript.mtimeMs in AgentSessionScanner.ts). That one value decides three things:

• whether the transcript is inside the 30-day window,
• the newest-first order used when the per-import budget runs out ([Bug]: Shared import record limit skips valid conversations #17484),
• the imported thread's updatedAt / settledAt, which is how it's sorted and dated in the sidebar.

Agent CLIs keep writing to transcripts after the conversation ends, so mtime often drifts from real activity. On one machine (Claude Code 2.1.303, codex-cli 0.162.1):

• 47 of 75 top-level Claude Code transcripts have an mtime 2 to 75 days (median 10) after their last timestamped message. The records written afterwards carry no timestamp: cost-state, last-prompt, ai-title, mode, permission-mode, and similar bookkeeping.
• 33 of them share one mtime date (2026-10-10), although their last messages range from 2026-09-28 to 2026-10-07. After import, 31 Claude threads all show 2026-10-10 as their last update, and they sort together in the sidebar out of real order.
• One session last active 75 days earlier was imported as "recent", because a bookkeeping write had refreshed its mtime.
• Codex shows the same pattern from a different cause. Opening an old conversation in the Codex app rewrites its rollout (Session import ignores compressed Codex rollouts (.jsonl.zst), so their conversations are never discovered #21755), so mtime reflects when the file was last opened rather than when the user last worked in it.

A concrete approach

• Compute lastActivityAt from the transcript itself: the newest timestamp on a conversational record. For Claude that's user / assistant records that aren't sidechain, meta or compact summaries. For Codex it's event_msg (user_message) and response_item message records. These are the same records parseAgentSessionRecords already keeps.
• Keep it cheap during the scan. Read the file backwards from the end in 8 KiB chunks, the same chunk size readCwd uses, until the first complete conversational record with a timestamp, under a small per-file byte cap. Bookkeeping records sit at the tail and are small, so this usually takes one or two reads. If the cap runs out, fall back to mtime as today.
• Use lastActivityAt everywhere mtime is used as activity: window eligibility, budget ordering, candidate lastActiveAt in the onboarding picker, and the imported thread's updatedAt / settledAt. Keep mtime and file identity for what they're good at, which is detecting a changed file between scan and read (sameTranscriptIdentity).
• Regression tests:
  • a fixture whose mtime is recent but whose last message is 60 days old (excluded, or dated correctly);
  • one whose last message is recent but whose mtime is old (included);
  • a Claude transcript ending in untimestamped cost-state / last-prompt records (dated by the last message).

2. Import history after onboarding

The import step is only reachable in the welcome wizard. After onboarding, the only way back on desktop is location.hash = '#/welcome' in DevTools (see #10961). Re-running the import is already safe to repeat: thread IDs are deterministic, completed transcripts are skipped by file identity, and existing threads are left alone. On this machine, a second run recovered 42 Claude Code sessions the first run had dropped, without touching the 147 existing threads.

A per-project Import agent history action (in the project menu and the command palette) that calls agentSessions.import for that project would cover:

• sessions started outside T3 after onboarding ([Bug]: Session import is unreachable after onboarding, so externally started sessions can never be imported #10961),
• sessions cut off by the per-run budget ([Bug]: Shared import record limit skips valid conversations #17484),
• older conversations brought back into the 30-day window by reopening them in Codex.

It would also help if the result said why threads were skipped: "N had no user message, N hit the import limit (run again), N couldn't be read". Today it's just "N threads could not be imported".

Prepared by Claude Code (claude-opus-5-5) via t3 triage.