Conversations

Queueing and interrupts while a turn runs, composer suggestions, context compaction, and custom commands.

Queue and interrupts

Field	Description	Default
MaxSessionQueueSize	Maximum queued requests per session; 0 means unlimited	3
InstantInterruptEnabled	Restart the current model request with a mid-turn message that arrives before the request produces output or while it waits to retry	true
AgentInterruptMessageEnabled	Tell the model when a turn is intentionally interrupted, including an active-turn fork	true

Suggestions

Field	Description	Default
WelcomeSuggestions.Enabled	Generates welcome-screen suggestions from workspace memory and applied Dreams. Requires Memory.Enabled	true
PromptSuggestions.Enabled	Suggests one next message in an empty conversation composer after a completed turn. Independent of workspace memory	true

Both settings can be changed in Settings → Personalization. A prompt suggestion remains editable and is sent only when the user submits it.

Compaction

Field	Description	Default
Compaction.AutoCompactEnabled	Enables threshold-based auto compaction	true
Compaction.ReactiveCompactEnabled	Enables reactive compaction for prompt_too_long errors	true
Compaction.MaxContextWindow	Client context budget in tokens. -1 uses the model catalog window; a positive value caps it	-1
Compaction.SummaryReserveTokens	Tokens reserved for summary output	20000
Compaction.SummaryMaxOutputTokens	Maximum output tokens for a compaction summary request	12000
Compaction.AutoCompactBufferTokens	Token buffer below the hard limit that triggers auto compaction	13000
Compaction.PostTurnCompactThresholdPercent	Percentage of the effective context window that also triggers compaction at the end of a Turn; 0 disables it	0
Compaction.WarningBufferTokens	Token buffer before auto threshold that emits warning	20000
Compaction.ErrorBufferTokens	Token buffer before auto threshold that emits error	10000
Compaction.ManualCompactBufferTokens	Headroom below the effective context window used for the reported context-pressure limit	3000
Compaction.KeepRecentMinTokens	Minimum recent tail tokens after partial summary	10000
Compaction.KeepRecentMinGroups	Minimum recent API groups after partial summary	3
Compaction.KeepRecentMaxTokens	Maximum recent tail tokens after partial summary	40000
Compaction.MicrocompactEnabled	Enables micro-compaction	true
Compaction.MicrocompactKeepRecent	Recent tool results kept during micro-compaction	8
Compaction.MicrocompactGapMinutes	Also triggers after this many minutes since last assistant message; 0 disables it	20
Compaction.MaxConsecutiveFailures	Consecutive failures before circuit breaking compaction	3

Set the context budget in global or workspace config.json. For example, limit the client to 256K tokens:

{
  "Compaction": {
    "MaxContextWindow": 256000
  }
}

Omitting the setting uses the model catalog window. Workspace settings override global settings; set MaxContextWindow to -1 in the workspace to clear an inherited limit. The budget cannot increase model capacity. Summary reserves and safety buffers are deducted afterward, so this value is not the exact automatic-compaction trigger. Reasoning effort does not change the budget.

A model's context window comes from the model capability catalog.

Custom commands

Custom commands are Markdown files rather than a config.json field. DotCraft loads them from the global ~/.craft/commands/ directory and the workspace .craft/commands/ directory, and a workspace file wins a name collision. Each file becomes a /name command usable from the CLI, Desktop, and other entry points.

A command file may open with YAML frontmatter, which is stripped before the body is sent. Inside the body, $ARGUMENTS expands to the full argument string and $1 through $9 expand to positional arguments.