Prompt & context

Compacting the context100×30 · 0:37

/compact replaces the conversation so far with a summary of it, and hands the tokens back — the gauge on the status bar moves as it lands.

The system prompt

The built-in prompt is deliberately short. Local models have small context windows and degrade quickly when instructions crowd out the actual task: every token of prompt is a token the model does not spend on your code.

You can replace it. Open /settings, select system prompt, and press enter to edit it inline — empty means the built-in one. Per-tool prompt overrides go in a [tool_prompts] table in the config.

Whichever prompt is in force, koda appends the facts a model cannot know on its own: the workspace path, the project type, and the current date, time and UTC offset. That last one matters more than it sounds — it is why koda dates a changelog entry from your clock instead of guessing a year from its training data.

Your instructions, and any AGENTS.md / CLAUDE.md / .koda.md in the workspace root, are appended either way.

Fitting a small context

A local model with an 8k window spends most of it on tool results it can no longer use: the first read of a file that has since been edited twice, the listing from six steps back, the 400-line build log whose only payload was ok.

Dropping the oldest messages to make room frees exactly the wrong thing. The first message in a session is the description of the job.

So when a request does not fit, koda decides what the model still needs, in escalating passes, stopping as soon as it fits:

1. Supersede. A read_file answered again later, or a file edited since it was read, is stale. Its body becomes one line saying what replaced it.

2. Squeeze. What survives is truncated on a ladder. The step in flight keeps everything. Recent steps keep their detail. Older ones keep their head and tail. Anything naming what you just asked about keeps twice as much as anything that does not.

3. Drop. Only now are whole exchanges removed, oldest first — and never the opening request.

The opening request is the description of the job, so it is the one message no pass will remove. All three run on a copy at send time — the conversation on disk is untouched, and under budget none of them run at all.

This runs on a copy at send time. Your history, the session file and /compact are untouched, so nothing is lost for good. Raise context_tokens and the next request carries the full detail again. Under budget it does nothing at all.

Curation and compaction are different tools

	What it does	What it changes
Curation	Trims the copy about to be sent.	Nothing. Runs every request, automatically.
/compact	Writes a real summary and keeps that.	The conversation itself.

/compact is the tool for a session that has genuinely moved on. Curation is what keeps a long turn working in the meantime. auto_compact_at = 0.85 runs compaction automatically once context passes that fraction of the budget; 0 disables it.

The memory block scales too

Memory in the prompt is sized to the window: one remembered fact per roughly 1k of context, between four and twenty. A model with 8k does not spend its context on notes instead of the task, and a model with 200k gets the full set.

Reasoning effort

Thinking models can be told how hard to think. /reason cycles off → low → medium → high, or pass a level directly (/reason high), or set reasoning_effort in the config.

koda sends it as reasoning_effort on the request. Servers that do not support the field ignore it, and off omits it entirely. Reasoning deltas are displayed dimmed — expandable with ctrl+t — and never sent back in the next request.