Adaptive

Adaptive makes FlupCode watch how your sessions go and help from what it sees: it points the agent to the skills that fit a task, warns you when the agent repeats itself, and proposes new skills from finished sessions. It never widens a permission, it installs nothing without your approval, and turning it off puts FlupCode back exactly as it was.

This page is how to turn it on and set it up. How it works inside is in the reference.

Turn it on

Open Settings → Adaptive and pick a Level:

Level	What it does
Off	Nothing runs. Your choices are kept for when you turn it back on.
Observe	Watches your sessions and notes what it would do, without changing anything.
Assist	Suggests skills and warns about loops while you work.
Custom	Your own mix of the four capabilities below.

Observe is the safe place to start: nothing the agent sees changes.

The four capabilities

Each one has its own switch under the level, so you can turn on only what you want.

Capability	Choices	What you get
Context	Off, Observing, Acting*	Works out which earlier parts of the conversation the agent still needs.
Skill suggestion	Off, Suggesting	Points the agent to the skills that fit the task.
Loop warnings	Off, Warning	Warns you when the agent repeats the same step. It never pauses the turn.
Learning	Off, Proposing	Proposes new skills from finished sessions. Nothing is installed without your approval.

* Acting changes what the agent sees, and it has not passed its offline evaluation yet. Leave Context on Observing unless you are testing it.

Freeze learning stops new proposals without turning Learning off: you can still review the pending ones, and the skills you already approved stay in use.

Choose a predictive model

Out of the box every decision is made by built-in rules, and nothing leaves your machine. A predictive model can answer some of those decisions instead. This step is optional.

Under Predictive model, Which model answers each decision has one selector per decision:

Decision	Jev	Small model
Which skills fit	Yes	Yes
Whether the task is finished	Yes	Yes
Why a step failed	Yes	Yes
Which context to keep	Yes	No
Whether a session is worth learning from	Yes	No
Whether a run moves to its fallback model	Yes	No

Each selector offers None (built-in rules) and the models that can answer that decision. A row tells you what its model is still waiting for: a project, the decision, permission to send data, or the key. Until all of them are there, the built-in rules keep deciding.

There are two models to choose from.

Jev

Jev is a hosted predictive model (api.typesafe.ai). It answers every decision, and it needs an API key from its provider.

1. In Predictive model, find Model key in Jev's section and paste the key. If a key is already saved, choose Change first.
2. Choose Save key, then confirm. The key is stored encrypted on this machine and is never shown again. It is read on every request, so there is nothing to restart.
3. Under Data shared with the predictive model, in Sharing with Jev, choose Add project and enter the path of each project Jev may be asked about.
4. In the same section, turn on each decision Jev may answer. Each one asks you to confirm, and covers Jev only.
5. Back in Which model answers each decision, pick Jev for those decisions.

To set the key from the environment instead, start FlupCode with FLUPCODE_TYPESAFE_API_KEY set. The environment wins over a saved key, and the panel then shows Key set by the environment and can no longer change it.

Small model

The small model is the small_model of your OpenCode config, asked through the engine with the provider you already connected. It needs no key of its own, and it appears in the selectors only when small_model is set.

It still sends your data to that provider, so it needs the same consent as Jev: add the project and turn on the decisions under its own sharing section, then pick it in the selectors.

One thing to know: each answer is a full round trip through the engine and takes seconds, while a decision waits 400 ms by default. With that default the small model almost always answers too late and the built-in rules decide. To use it for a decision, raise that decision's deadline in the config file (the panel shows it read-only):

// ~/.config/opencode/opencode.json
{
  "flupcode": {
    "adaptive": {
      "decisions": { "completion": { "timeoutMs": 60000 } },
    },
  },
}

What is sent, and to whom

Consent is given per provider, per project and per decision. Allowing the small model never lets anything reach Jev, and the reverse. What a model receives is redacted first, and a model that is not allowed for a project is simply not asked there.

The same settings, written by hand:

// ~/.config/opencode/opencode.json
{
  "flupcode": {
    "adaptive": {
      "models": { "skillRelevance": "jev" },
      "egress": {
        "providers": {
          "jev": { "enabled": true, "projects": ["/work/app"], "kinds": { "skillRelevance": true } },
        },
      },
    },
  },
}

A key is never written to this file: it lives in the encrypted vault or in the environment.

Check that it works

• The summary of Predictive model names each decision and its model, with the reason when one is not ready, for example Whether the task is finished: Jev (key missing). A decision that is ready shows only its model.
• Is it worth asking? shows, per decision, whether the model is being asked: Measuring its value while it warms up, then Asked. Paused: it does not help here means its answers did not pay for themselves, and the built-in rules took over again.
• Data & budget shows the tokens spent this month and the predictive model's cost over its recent decisions.

Budget and history

Under Data & budget you can set a Monthly budget (tokens) for the predictive model. When it is spent, the built-in rules decide until the next month. The same section can remove decision history older than a number of days.

Turn it off

Set the level to Off: every capability stops at once, nothing is deleted, and the skills you already learned still load.

To force it off from outside the app, start FlupCode with FLUPCODE_ADAPTIVE_DISABLED=1. The level then stays Off and the panel says the environment set it.