bandicot
Bandicot Docs
Chapter IIModels · Available now

Decisions

Decisions is a small, fast model that makes quick calls while agents work: screen picks, risk checks, compaction. Choose Jev or a model on your Mac.

Available now. A current user-facing path exists within the stated scope.

What Decisions is

Decisions is a small, fast model that Bandicot asks quick questions while an agent works, so the agent's own model can stay on the real task.

The agent's own model does the thinking and writing. Decisions answers the short questions around it: which button to press, whether a tool call is clearly safe, which part of an old conversation to keep. Each question has one answer, so a small model is enough and it replies in a fraction of a second.

You do not have to set anything up to use Bandicot. Decisions is a choice you can make in AI Gateway to cut cost and speed up agents. Older notes and posts call it System One.

Open the Decisions tab

Open Home (double-tap ⌘), then Account › AI Gateway › Decisions, which has three cards: Compaction, Decisions model and What Decisions does while agents work.

The cards read top to bottom. Compaction says when an agent makes its older history smaller. Decisions model is where you pick which model answers. The last card lists every job with its own switch, and ends with two folds: Compare models on these jobs and How screen picks work, and the benchmark.

The same model is named on the Brain tab of Memory, and on the Decisions tab of Efficiency.

Choose the Decisions model

In the Decisions model card, press the name of the model you want: Jev runs on TypeSafe, and a Laya model runs on your Mac.

Next to the title, a status line says what answers now: local · <model>, TypeSafe · <model>, or no Jev model. A hosted model such as Jev stays greyed out until TypeSafe is connected, and its button then says to connect TypeSafe first. A model you installed on this Mac needs no connection.

Jev is made by TypeSafe, answers from there, and is billed to your TypeSafe account per million input tokens. Laya runs on your Mac, costs nothing per ask, works offline and keeps the question on the machine. Some models carry a sees images mark: they can also be given a picture of the app window.

The pick applies to every job below it. On the Rack, the Decisions rows follow this pick unless you put a different model on one of them.

  1. Open AI Gateway › Decisions.
  2. To use Jev, paste your key into Paste a TypeSafe key and press Connect. The status line then says TypeSafe.
  3. To use a model on this Mac, install a Laya model first: press ⇧⌘M and choose a Decisions model, or press Download & use for decisions where the picker offers it. See Local models.
  4. In Decisions model, press the model's name. A note confirms Decisions model set to <model>.

If it goes wrong

The Jev button is greyed out and its hint says to connect TypeSafe.
Paste your TypeSafe key and press Connect, or choose a model installed on this Mac.
The list says there are no Jev models yet.
Connect TypeSafe, or install a decision model from Models.
Connect stays disabled.
The key field is empty. Paste the key first.
A job asks for a Decisions model and says none is turned on.
Pick one here. The Models tab shows None turned on until you do.

What Decisions does while agents work

The last card lists every job Decisions does, each with a switch, and a switch takes effect on the very next question, even in the middle of a turn.

Each row says what the job does, when it runs, and what happens when you turn it off. An off job is not skipped: the agent does it the plain way. A row that is off also counts how many times it was held back.

The right of each row shows how many questions it asked and the average time, and the header counts them all with the total paid.

The jobs, in the order you meet them
JobWhat it doesWhen it runsIf you turn it off
Screen picksWhat it doesChooses the button or field an agent presses in an app.When it runsEach time an agent working an app names its next step.If you turn it offThe agent reads the list of controls and picks itself.
Risk checkWhat it doesLets clearly safe tool calls run without asking you.When it runsBefore a tool call that would ask your permission.If you turn it offYou are asked as usual.
Agent asks directlyWhat it doesA tool agents call for a quick yes, no or pick.When it runsOnly when an agent decides to call it.If you turn it offAgents no longer have the tool.
CompactionWhat it doesKeeps the parts of older history that matter, word for word.When it runsWhen an agent reaches Compact at, or you type /compact.If you turn it offThe agent's own model writes a summary instead.
Context rankingWhat it doesScores older history ahead of time so compaction is instant.When it runsBetween turns, while the agent waits for you.If you turn it offCompaction scores everything when it runs, a little slower.
Help for the agentWhat it doesPoints the agent to a tool or skill it has not loaded.When it runsDuring a turn, once per step at most.If you turn it offThe agent finds tools on its own.
Install adviceWhat it doesSuggests a plugin or extension from your marketplace when nothing installed fits.When it runsAt an agent's first request, and when it starts a new goal.If you turn it offAgents use what is installed, and can still ask you.
Pattern watchWhat it doesSpots what you prefer and saves it to memory.When it runsAfter two idle minutes, at most every half hour.If you turn it offNothing is saved by it.
Worth rememberingWhat it doesPicks what from a finished turn is worth keeping.When it runsAfter a turn ends, in the background.If you turn it offEverything is offered, unranked.
Work profileWhat it doesLearns how you work from each turn.When it runsAfter each turn, in the background.If you turn it offThe profile stops learning.
Next-step adviceWhat it doesSuggests the next step for automations.When it runsWhen an automation finishes a step.If you turn it offNo advice.
Memory tidyWhat it doesMerges and cleans your notes.When it runsOn the charger, idle five minutes, at most hourly.If you turn it offNotes stay as they are.

Turn every job on or off at once

The Every job switch at the top of the card turns all jobs on or off together, and its line shows how many are on, how many questions were asked and what was paid.

The line reads like <n>/<total> on · <n> asks · $<amount> paid. The amount is only what was sent to a hosted model; a model on your Mac adds nothing.

Compaction is shared with the Compaction card above it. One setting, two places: flipping either one moves the other.

  1. Open AI Gateway › Decisions.
  2. Find the What Decisions does while agents work card.
  3. Press the Every job switch to turn all twelve jobs off, or on again.
  4. To keep most of them, leave it on and turn off the single rows you do not want.

Screen picks

When an agent works an app, Screen picks chooses which button or field to press, and it presses only when it is sure.

An agent working in an app states its goal. Bandicot lists every button and field it can press and asks Decisions twice, once with the list forward and once reversed, which control the goal needs. It presses only when both answers agree and the pick clears seventy percent with a lead of fifteen points. Otherwise the agent gets the list back and chooses itself. A wrong press is worse than a slow one, so Decisions is built to hand back rather than guess.

While Screen picks is on, two small switches appear under its row. Jev when connected makes screen picks use Jev whenever TypeSafe is connected, even if your pick above is a model on this Mac; the line beside it names the model that answers now. Send a screenshot also sends the app window to a model marked sees images, on Mac apps. If that ask fails, the same question is asked again as text, so a pick never breaks. A text-only model says it reads text only.

To see each pick, open an agent and press ⌘8, then the Decisions tab: every screen pick shows a bar with the seventy percent line, and under the line it was handed back. Asks that you switched off show there as held back. The fold How screen picks work, and the benchmark explains this in the app and compares the models Bandicot tested.

Compaction: what stays when history is shortened

In the Compaction card, Decisions chooses what stays keeps the parts of older history that matter word for word, and Compact at sets the size at which an agent compacts.

When an agent's context fills up, older history is made smaller so the agent can keep working. It happens at the point you set, or when you type /compact. With the switch on, Decisions keeps the parts that matter word for word and drops the rest: fast, and nothing is rewritten. With it off, the agent's own model writes a summary instead. Either way, compaction itself always happens; off only changes who shortens it.

Compact at offers Automatic and fixed sizes, each shown only when it is below the open agent's window. A bar shows where compaction falls in that window, as a share of its size. The setting applies to the agent you have open. Usage and cost shows the context meter, and an agent's earlier turns stay searchable after compaction, as Memory explains.

Fixed choices for `Compact at`
ChoiceTokens
AutomaticBandicot picks the point for the model's window.
Fixed sizes32,000, 64,000, 100,000, 128,000, 200,000, 300,000 and 500,000

Compare models on your own asks

The fold Compare models on these jobs prices the real questions Decisions asked in the last thirty days on each model, so you see what a switch would cost.

Job to compare picks one job or Every job. The line above the table counts your asks and their input and output tokens. Each row is a model, with where it runs (This Mac, TypeSafe, OpenCode, Anthropic or OpenAI), its price per million tokens in and out, what your asks would cost, the cost per 1,000 asks, and a score from Bandicot's own screen test. The Terminal-Bench 4.0 column shows the score the lab published for models that have one. A model on your Mac reads free. The row of the model in use is marked in use.

Under the table, one sentence compares the model in use with the dearest one. Prices are each model's list price, and where a job wrote no output tokens a general-purpose model is charged a typical amount per ask. Models that were not tested show not run.

What Decisions costs

A model on your Mac costs nothing per ask, and Jev bills your TypeSafe account per million input tokens; nothing else is added by Bandicot.

Decisions work counts against the automatic AI allowance in Usage and cost, priced at TypeSafe's rate, so a hosted pick cannot spend past the cap you set. The Every job line shows what was paid, and the Compare models on these jobs fold shows what each alternative would have cost.

To spend nothing on Decisions, install a Laya model and pick it. To spend less, turn off the jobs you do not use: an off job asks nothing.

A hosted pick means the question text leaves your Mac. A model on your Mac keeps it on the machine. Background checks that Bandicot runs to compare its choices run only for a model on your Mac, so a hosted model is never billed for them, and the receipts are on the Brain tab of Memory.

If it goes wrong

The account allowance is used up. Raise it in AI Gateway → Usage.
Raise the allowance under AI Gateway › Usage, or switch Decisions to a model on your Mac.
The screen-pick line says TypeSafe not connected.
Jev when connected is on but there is no key. Connect TypeSafe, or turn that switch off.

Jev and Laya

Jev and Laya are the two families of Decisions models: Jev runs on TypeSafe and Laya runs on your Mac.

Both answer with a choice and how sure they are, not with a paragraph, which is why they are quick. Pick one for everything, or put a different model on a single row of the Rack. The agent's own model, whether Claude, ChatGPT or a local one, is a separate choice in Connect a model.

Jev and Laya
JevLaya
Runs onTypeSafeYour Mac
NeedsA TypeSafe keyA download from the model library
CostPer million input tokensNothing per ask
OfflineNoYes
Sees imagesNewer releases do, and sees images shows on the modelNo

Questions

What is Decisions?

A small, fast model that answers quick questions while an agent works, such as which button to press or whether a tool call is safe. It is the current name of what older posts call System One.

Do I need Decisions to use Bandicot?

No. Turn the jobs off and the agent does each one itself, only a little slower. Pick a model to speed agents up and spend less on the big model.

Does Decisions send my data anywhere?

Only if you pick a hosted model such as Jev, which sends the short question to TypeSafe. A model on your Mac keeps it on the machine.

How do I stop it spending money?

Pick a Laya model that runs on your Mac, or turn off Every job. The automatic AI allowance in Usage also caps what a hosted pick can spend.

Can an agent press the wrong button because of Decisions?

It is built to hand the choice back instead of guessing: it presses only when two asks agree and the pick is clear, and otherwise the agent decides itself.