Slash commands and compaction · Current design

Slash commands dispatch; compaction produces system messages

The entry for context compaction is /compact. The composer executes optional-argument commands immediately; compaction start, success and failure each produce a system message. Measurement, nothing-to-compact behavior and ring lifecycle belong in the occupancy design.

Backend compaction pipeline available Optional-argument commands execute on selection or Enter System messages for started / finished / failed See occupancy design for ring suggestions and panel actions

Summary

/compact asks the model to summarize old conversation turns, records the summary in the DAG, and supplies “summary + recent turns” afterward.

Optional arguments ([…]) execute immediately on selection or Enter. Menu closing animations no longer intercept Enter. Without a session, the UI asks the user to send a message first instead of claiming success.

What compaction does

Models have finite context windows. As conversation grows, older turns must be reduced. OpenProgram uses a rolling summary: only one summary is active at a time. Further compaction combines the old summary and additional turns into a new summary instead of stacking summaries.

The next model request actually contains:

[System prompt] [Current summary] [Recent uncompacted turns] [New user message]

Three trigger paths

Trigger	Condition	Status
Manual `/compact`	Execute in the composer; optionally keep the latest N tokens	Emits `action: compact`. Without a session, show an alert instead of running an empty operation.
Automatic pre-turn compaction	Turn budget ≥ 80% of the window. Compact first, then fall back to snip	Dispatcher handles this in `loop_runner.py`. System messages use the same wording.
Context-window overflow recovery	Model returns 413 / prompt_too_long; at most once per turn	`reactive.py`; uses the same compaction events.

The ring color reflects occupancy, and the panel offers one-click compaction; see Occupancy and compaction.

Compaction steps

Read

Read rendered history from SessionDB: the existing summary and turns remaining on the branch. Recompaction does not reread original text already covered by a summary.

Threshold

Stop if fewer than four messages remain; the manual path emits a system explanation. If rendered history is already summary plus tail with no valid cut, emit nothing-to-compact without invoking the summarization LLM.

Summarize

Make a separate LLM call. By default, retain a recent tail of 10% of the window, clamped to 8k–40k tokens and at least five nonempty messages. This can be overridden by /compact 2000.

Persist

Write a summary_* node (role=llm, name=context/summary); covers_ids marks the replaced range. Session metadata records the summary for subsequent compaction. Old turns are not deleted and HEAD does not move.

Graph

The DAG displays a capsule. The old summary becomes superseded and stops participating in rendering. Covered original conversation text remains in storage.

Ring and /context panel

The lower-right ring shows current occupancy divided by window size; opening it reveals token categories. Occupancy uses the compacted rendered view, matching the compaction report. Colors: green below 50%, yellow at 50–80%, red at ≥80%.

The occupancy design defines running-compaction state and the Compact now panel interaction; this page does not create separate behavior.

Slash commands

When input starts with / and contains no space, the menu expands upward from the composer. After a space appears, it closes over 380ms. Enter checks whether the first word is a registered command: execute if registered, otherwise send ordinary chat.

The built-in registry is in slash-commands.ts. It also fetches /api/commands to include skills, plugins and user commands. Built-ins have dispatch precedence.

Command	Selection / Enter
`/compact`	Compact immediately. The `[keep_recent_tokens]` argument is optional; execution does not require it.
`/rewind`, `/branch`	Run immediately (arguments optional).
`/context`	Open the ring panel.
`/clear`, `/new`	New session.
`/task`, `/skill`, `/merge`	Fill the prefix and wait for required arguments.
`/help`	Set input to `/`, making the menu appear again.
`/sandbox`, `/doctor`, `/memory`	Run their respective actions; with a session, doctor sends its report as a user message.

Implemented slash-command fixes

Optional arguments (only […]) execute immediately on selection / Enter./task and similar required-argument commands still fill a prefix first.
When the menu is closing, it does not intercept Enter; execution proceeds to runCommand.
Without a session, retain input and alert “Start a conversation first” rather than clearing it as if successful.
compaction_started / finished / failed each have a system message. Nothing-to-compact is a fourth state defined in the occupancy design.

Minor unfinished items

If the unified registry also names an item compact, the menu shows duplicate names. Dispatch still selects the built-in.
Jump to latest has z-index 7 while the composer has 5; the button may overlap the menu position.
/help is not a separate page.
/doctor sends its diagnostic report as chat when a session exists.

Key files

Responsibility	Path
Command registry	`apps/web/components/chat/composer/slash/slash-commands.ts`
Menu / dispatch	`apps/web/components/chat/composer/slash/use-slash-menu.ts`
Selection	`apps/web/components/chat/composer/index.tsx` (`selectSlashCommand`)
Keyboard	`apps/web/components/chat/composer/input/use-composer-keydown.ts`
Enter submission	`apps/web/components/chat/composer/submit/use-chat-submit.ts`
Manual compaction over WS	`apps/server/openprogram_server/_webui/ws_actions/chat.py` → `handle_compact`
Compaction pipeline	`openprogram/context/engine.py`, `summarize.py`
Automatic compaction	`openprogram/agent/dispatcher/loop_runner.py`
Context-window overflow recovery	`openprogram/context/reactive.py`
Frontend event handling	`apps/web/lib/runtime-bridge/chat-handlers.ts`
Ring / panel	`context-badge.tsx`, `context-breakdown-panel.tsx`
Occupancy measurement	occupancy-status.html