compact-boundary-session-distillation

SKILL: Compact Boundary Session Distillation

Domain: context-management
Trigger: When skillify captures a repeating process, use the latest session memory and the user messages after the most recent compact boundary so the prompt only contains the current instructions. Source Pattern: Distilled from reviewed context, compaction, and memory-governance patterns.

Core Method

• Call getSessionMemoryContent() to fill the <session_memory> block; the helper already returns null when no file exists so the prompt displays No session memory available. instead of throwing.
• Slice the transcript with getMessagesAfterCompactBoundary(context.messages) from runtime/utils/messages.ts, which locates the last compact boundary and optionally projects out snipped history via projectSnippedView when HISTORY_SNIP is on.
• Feed that clip into extractUserMessages, which keeps only type === 'user' entries, flattens mixed content blocks into strings, trims blank text, and yields the freshest user steering cues.
• Replace {{sessionMemory}}, {{userMessages}}, and the optional {{userDescriptionBlock}} inside SKILLIFY_PROMPT so the kernel prompt knows which context to expose to the skill authoring assistant.

Key Rules

• Do not let user text from before the compact boundary slip into the distilled user stanza; the getMessagesAfterCompactBoundary call enforces that slice.
• Always wrap the memory summary in <session_memory> and the user messages in <user_messages> so downstream tooling can parse them reproducibly.
• Respect the description the user optionally supplied, because it seeds the suggested skill name, goals, and trigger phrases in the AskUserQuestion rounds that follow.

Example Application

When the user redirects the conversation toward automation mid-session, this pattern yields a short prompt that highlights the current process and memory highlights without dragging the whole transcript along.

Anti-Patterns (What NOT to do)

• Do not include assistant or system messages—even if they appear in the slice—since only the user's intent should feed into the new skill definition.
• Do not drop the session memory block; even if it resolves to null, the placeholder text keeps the prompt structure stable for downstream parsers.