[ScriptedAlchemy]

Why

Cache reads are 70% of list-price spend across 30 days of real sessions, because every call re-sends the whole thread. Once a thread's prompt cache has expired, the next request is uncached anyway, so a fresh session with less history costs little extra and makes every later call cheaper.

What changed

• A new freshSessionOnColdReturn server setting is off by default. When it is on, a person's message to a thread that is over 100k tokens and has been idle for an hour skips resume. The thread continues in a new native session through the existing context handoff.
• Web and mobile settings get a toggle, and the portable handoffs guide gets a section.
• The replay that priced this change and its alternatives moved to a private repo, next to the private chats it reads. This branch no longer carries it.

Scope

• The change covers every provider, because it reuses the existing resume-fallback path.
• Commands, agent messages, system messages, and threads with background work never trigger a reset.
• Before an elective reset, the reconstructed handoff must fit the target model and the resumed session must still have no background work. Otherwise T3 keeps the usable existing session.
• Idle time includes accepted root-turn activity in the current native conversation, so a long turn's old token report cannot make a recently completed turn appear idle.
• The change does not use a decision model to choose what survives a reset. Real models are still being evaluated on that question in the same private repo, and none has earned a place yet.

Tradeoffs

• Resetting during active work saves more on Codex (10.5% at over 150k tokens) but interrupts the work, so this change starts with the safest moment.
• Claude's server-side clear_tool_uses lost money or roughly broke even at every setting replayed.
• No provider has a supported way to edit one old tool result.

Blast radius

The setting is off by default, so nothing changes until someone turns it on. When it is on, the agent loses older context in that session. The agent can read omitted history back through the thread-reading tool, but nobody has measured the effect on answer quality.

Verification

• All 80 ProviderTurnStartService.test.ts tests pass, including three added regression tests. They cover a recent completion after old telemetry, an oversized replacement handoff, and background work that appears during handoff preparation. All three failed before the fixes.
• vp run check:changed passes on the latest commit: formatting, lint, both Knip passes, GraphQL, typecheck across 36 packages, seven test files, and React compiler checks.
• A 30-day replay at API rates, which charges for measured re-reading, shows Claude saving 18.7% ($56 of $302) and Codex saving 5.0% ($149 of $2,969). The replay with no resets matches T3's priced cost within 0.02%, and moving it out of this repo reproduced its output byte for byte. Claude's 95% confidence interval on 52 sessions reaches just below zero.
• This change has had no live run. Nobody has used the toggles in a real client, so there are no screenshots yet.

GPT-6 Astra Pro in ChatGPT did the follow-up review and fixes.

Claude Opus 5.5 in Claude Code, through T3 Code.

🤖 Generated with Claude Code

[changeset-bot]

⚠️ No Changeset found

Latest commit: 2aa714b

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR