Managing context

The model can only see so much at once — its context window. Hive manages that budget for you automatically, shows you where it stands at all times, and gives you explicit controls when you want them.

The context pill

The chat header shows a live pill — ⬡ 42% — with the share of the model's window the next reply will use (system prompt + the messages that fit). It turns amber at 80%. Click it to open the Context pane in the right rail.

The Context pane (⬡)

The full breakdown, updated live while replies stream:

• Budget bar — tokens planned vs. the model's window.
• What's in the prompt — system prompt, kept history, loaded skills (with their token cost), attached vaults.
• What fell out — how many older messages overflowed and were folded into an automatic summary.
• One-click unloading — remove a skill or unmount a vault right from the pane to reclaim budget.

/summarize and /compact

Long conversations eventually outgrow any window. Hive auto-summarizes overflow — older messages are collapsed into a running summary that rides in the system prompt — but two / commands give you the explicit version (type / in the composer; both are tagged Context):

/summarize

Posts a model-written summary of the conversation so far as a new message. The transcript is kept — nothing is removed. Use it for a recap, or to hand a newcomer the gist.

/compact

Collapses the conversation into a single summary checkpoint. The existing messages are removed and the summary becomes the new head of the chat, freeing context immediately. This is the explicit version of the automatic overflow handling.

/compact removes messages

Hive confirm-gates it. The earlier messages are gone from the active conversation afterward — reach for /summarize if you want to keep the full history.

Customizing the instructions

Both commands follow an instruction you can rewrite in Settings → Models → Context commands:

• Two text fields — one per command — with the built-in default shown as the placeholder. Blank = default.
• Want /compact to preserve code snippets verbatim, or /summarize to output only decisions and action items? Write it there.
• The /summarize instruction also guides the automatic overflow summarization, so long-chat memory follows the same rules you set.

Changes apply immediately (persisted in settings.json); no restart.

Sizing the window itself

Hive infers the context window from the model name (Claude, GPT-4o, …). For Ollama / LM Studio / custom endpoints it can't, so set it yourself:

• Settings → Models → Add runtime → "Context window in tokens", or
• context_window = 32768 on the runtime in hive.config.toml.

The planner budgets against your number — the pill and pane reflect it.

What else occupies the window

Everything the pane counts, and where to manage it:

Occupant	Managed in
Skills	Skills pane (✦) / Context pane
Vaults	Vaults pane / Context pane
Attachments & @file	The composer — you choose what to include
Conversation history	Automatic windowing + /compact

See Your first chat for the rest of the composer's / menu.