Resin

Part of Genshin mods. In the game, resin is the capacity that refills on a clock and is spent on what matters. A session has the same: its prompt cache, its context window and its usage limits. Resin shows all three at a glance and makes the usual answers to them one press each.

The row

Shown once the session has sent a request, and gone again after a /clear or a resume:

• cache — minutes until the prompt cache expires, counted from when the last request or warm was sent rather than from its reply, since that is when the cache was read, and how much context a cold cache would re-send at full price.
• context — tokens used against the window, as the engine reports them.
• 5h and 7d — the limit windows' percentages, when the account reports them.
• cost — what the session would have cost on API billing, the engine's own total.

A figure takes the warning colour past its threshold — the cache under ten minutes, the context or a limit at four-fifths — and a figure the engine has not measured yet is left out rather than shown as zero. The countdown redraws once a minute.

How to use it

Each is a button, or with the band focused a key: w, c and h.

• Warm renews the cache: a one-word side question through a fork of the session, which re-sends the conversation's prefix and adds no row to the transcript.
• Compact runs the engine's own compaction.
• Handoff is the whole relay in one press. A fork writes a handoff of the session — the goal, what is done, what is next, the files and decisions that matter, the open questions — then the mod clears the conversation and sends the handoff as the first prompt of the fresh one. The buttons give way to a note while it is written, and the clear waits for the handoff, so a fork that fails clears nothing and a toast says why. A clear or submit that fails after the fork is toasted the same way, and the buttons come back either way.
• Five minutes before the cache expires, a toast says so, once per expiry and never while a turn is running, since the turn renews the cache itself.
• /resin off hides the row, the toasts and the usage reserve's section.

Usage reserve

A plan's usage windows are spent by the session and its agents, and a compute run that is still owed needs some of the window left to start. So when the five-hour or the weekly window passes nine-tenths used (USAGE_RESERVE_PERCENTAGE in constants.ts), the mod puts a usage reserve into the session and keeps it there until that window resets.

flowchart TD
  M["Engine measurement<br/>reads the five-hour and weekly windows"] --> P{"A window past the line<br/>with a reset time?"}
  P -->|yes, five-hour before weekly| Q{"Reserve already on?"}
  Q -->|no| ON["Reserve starts"]
  ON --> T["Toast names the window and its reset, once"]
  ON --> S["System section, held still while the reserve lasts"]
  Q -->|yes| HOLD["Reserve holds, no second toast"]
  P -->|no, reserve on| OFF["Reserve lifts<br/>the section stops being sent"]
  P -->|no, reserve off| NONE["Nothing shown"]

• What starts it — each measurement the engine pushes reads the limit windows. The first window at or past the line, five-hour before weekly, starts the reserve, and a toast names it with its reset time in local time, once. Nothing is shown when the reserve ends.
• What the session is told — one system section naming the window and its reset time, then an instruction: start no new agent or workflow that writes code or designs, have each running one commit what builds and end on a handoff spec, commit and push the work in hand, write every open item down so it can be resumed cold, and clean up idle servers, shells and monitors. The text holds still while the reserve lasts, so it does not break the prompt cache. It is sent only while the resin mod is on.
• What lifts it — the window's reset drops its reading under the line, and the next measurement clears the reserve, so the section simply stops being sent. No button or command does this.

The steps a session follows under a reserve are the repository's own, in the throughput skill's "The usage reserve" section.

Key files

Notes

• The cache's lifetime is the engine's choice, an hour on a subscription and five minutes on some API plans. The engine does not report it, so it is the mod's one constant, CACHE_LIFETIME_MS.
• A window the engine reports with no reset time does not start a reserve, since both the reserve's section and its lift depend on that time.
• The reserve belongs to the account's usage window rather than to the conversation, so a /clear or a resume keeps it until the next measurement reads the window again.
• The published cache mod also compacts by itself past a threshold. That is left to the engine's own auto-compact, which already reads the window, so the mod adds no second policy beside it.