Menu Documentation

Background model

One machine-wide model for titles, summaries, compaction, recall and sub-agents.

Your conversation is not the only thing Ogcode asks a model to do. Titles get generated, long turns get summarised, unfamiliar commands get assessed, and recall questions get answered. These are all real LLM calls, and they do not have to run on the model you are talking to. The background model is a machine-wide choice of which model they use.

What counts as background work

Everything that runs behind the scenes rather than in your transcript:

TaskWhat it does
Session titlesNames a new session from its first message.
Turn-memory summariesWrites the structured note kept after a turn and indexed for recall.
Context compactionSummarises a long turn so the agent can continue from where things stand.
Command risk assessmentDecides whether an unfamiliar shell command is routine enough to run in Auto mode.
Intent tool matchingPicks which lazily loaded tools a plan step needs.
Auto-indexDescribes files for the project index refreshed after a turn.
Deep searchRuns the research pipeline behind a deep search query.
Memory recallAnswers a recall question from past turn summaries.
Task sub-agentsThe read-only task children, and background tasks started to keep working on their own.

Smaller jobs such as the note text transform follow the same setting when they are not handed a model of their own.

Where to set it

Open Settings → General and find the Background tasks group; the Background model row is its first setting. The choice is machine-wide: it is stored on the server and applies to every session, agent and project there, so you set it once rather than per session. On a remote deployment, set it through that server’s Settings page; the same value is reachable at GET/POST /api/models/background if you script it.

Pick any model from any configured provider. Changing it affects the next background call; there is nothing to restart.

Inherit from the active session

The default option, Inherit from active session, leaves the setting empty on purpose. With nothing chosen, each background call uses the model and provider of the session doing the work: exactly what Ogcode did before the setting existed.

Choosing a model pins it instead. The setting is honoured only while that model is actually being served; if you later remove it from the catalogue or disable its provider, background calls quietly fall back to inheriting the session’s model rather than failing, and saving a model that is not being served is refused outright. Switching back to Inherit from active session clears the pinned model and provider.

What it does not change

Your conversation always runs on the session’s model: the one picked in the composer, or the one passed with --model to ogcode run. The background setting never redirects a turn, so a small, fast background model is a way to keep titles and summaries cheap without touching the model you actually work with.

The reverse direction has a limit worth knowing: the background model only replaces a model that would otherwise be inherited from the session. When a job names a model of its own — a manual index build with an explicit model, or a note text transform given one — that explicit choice wins and the background setting is not consulted.

A sensible split

If you have a fast, inexpensive model available, set it here and keep the stronger one for sessions: titles and summaries are short, mechanical jobs where the smaller model is usually indistinguishable. If you have a single provider or are unsure, leave it on Inherit from active session; that is a perfectly good setup.

Next

From side project to enterprise codebase.

One Go binary, any model, your keys. Ogcode builds the context each task needs, so you get the job done without burning through your wallet.

Or install from a terminal: curl -fsSL https://ogcode.in/install.sh | sh