Browse documentation

Technical documentation / Reference

Session summarizer

Mission Control can turn recent session activity into short progress entries in the Summary pane. It is disabled by default.

Start and stop

The pane shows Running during a provider request, Waiting while enabled between requests, and Stopped when disabled. Stopping keeps existing entries. The enabled state and progress are saved per session and restored when you resume it. This Mission Control instance observes only the selected session.

The header toggle controls visibility only. Hiding Summary does not stop generation or clear entries. By default it overlays the right side below the header without resizing the transcript or changing tabs. Alt-A pins all column modals beside Transcript instead. Pinned modals hide on Changes and return on Transcript. Prompt stays available in either mode.

Session files and scrolling

Session Files is a separate panel (Alt-2) and works while summarizing is stopped. It shows recorded file-tool activity and, when enabled, inferred Bash activity. It is not a full filesystem audit.

Filenames show successful read counts in parentheses, such as main.rs (3). Failed reads and inferred Bash reads do not increase these counts. Counts are restored from session history when you resume.

Scroll Summary with arrows, Page Up/Down, or the mouse wheel. See Mission Control for panel focus, pinning, and visibility controls.

Settings

Set agent.summarizer in global or project settings.json; see Configuration for an example and merge rules.

SettingBehavior
auto_startDefault false. Supplies the initial enabled state when a session has no saved summarizer state; a saved Start/Stop choice takes precedence.
provider, modelOmit both to follow the active agent. A model-only override uses the active provider. A different built-in provider can use its default model; a different custom provider requires an explicit model.
reasoningOmit to inherit the active agent's level, or choose default, low, medium, high, x_high, or max. Support depends on the provider/model.
promptReplaces the built-in system prompt, rather than adding instructions. Blank values are rejected; only the first 8,000 characters are used.

A custom prompt must still request a JSON array containing zero to six nonempty strings, each at most 600 characters. Return [] when there is no meaningful progress. Treat observed activity as data, not instructions. The built-in prompt is prompts/summarizer.md.

The built-in prompt asks for zero to two meaningful outcomes per update, grouping related steps and skipping routine reads, searches, plans, and repeated progress. Activity headers identify the primary session and each subagent. Child outcomes must use Subagent <source id>: so their work is not attributed to the primary agent. Custom prompts replace these instructions; the parser still accepts up to six entries.

Cost, storage, and limits

Edit this page on GitHub ↗