Distillation

Date: Sep 17, 2026

• 🔴 (08:36) User directed that local ownership be replaced with durable multi-node ownership when clustered.
• 🟡 (08:37) packages/core/src/session/compaction.ts implements automatic Session context compaction with defaults DEFAULT_BUFFER = 20_000, DEFAULT_KEEP_TOKENS = 8_000, TOOL_OUTPUT_MAX_CHARS = 2_000, and SUMMARY_OUTPUT_TOKENS = 4_096.
• 🟡 (08:37) SUMMARY_TEMPLATE requires exactly ordered Markdown sections: ## Objective; ## Important Details; ## Work State with ### Completed, ### Active, and ### Blocked; ## Next Move with 2 numbered actions; and ## Relevant Files. It requires every section even when empty, terse bullets, preservation of exact paths/symbols/commands/errors/URLs/identifiers, and no mention of compaction or the summary process.
• 🟡 (08:37) SUMMARY_UPDATE_INSTRUCTIONS tells the model to combine <prior-summary> with the newer <conversation>, preserve still-needed objectives, constraints, directives, decisions, and parallel workstreams, prefer newer conflicting facts, add new progress/context, move finished work from Active to Completed, update resolved blockers, and refresh Objective and Next Move.
• 🟡 (08:37) settings(documents) reads info.compaction from Config.Document entries and reduces them in order; each document may override auto, buffer, and keep.tokens, with defaults { auto: true, buffer: 20_000, tokens: 8_000 }.
• 🟡 (08:37) Compaction serialization in packages/core/src/session/compaction.ts labels message content as [User], [Assistant], [Assistant reasoning], [Assistant tool call], [Tool result], [Tool error], [System update], [Synthetic context], or [Shell]; user/file attachments are rendered as [Attached ${mime}: ${name ?? uri}], and tool attachments as [Attached ${mime}${name === undefined ? "" : \: ${name}`}]`.
• 🟡 (08:37) truncate(value) limits serialized tool and shell output to 2_000 characters and appends \n[truncated]; serializeToolContent(...) joins tool content items with newlines.
• 🟡 (08:37) select(entries, tokens) excludes messages whose type is "compaction", serializes the rest, discards empty serialization, and walks backward to retain the newest messages fitting within the token budget; it returns older serialized content as head and retained content as recent.
• 🟡 (08:37) buildPrompt(...) supports 2 paths: without a prior summary, it asks for a new anchored summary from <conversation>; with one, it includes both <conversation> and <prior-summary>, then applies SUMMARY_UPDATE_INSTRUCTIONS. Both paths append SUMMARY_TEMPLATE.
• 🟡 (08:37) SessionCompaction.compactIfNeeded returns false when automatic compaction is disabled, the model context limit is absent/nonpositive, or estimated request tokens fit within context - Math.max(output, config.buffer); otherwise it delegates to compactAfterOverflow.
• 🟡 (08:37) SessionCompaction.compactAfterOverflow returns false when the context limit is absent/nonpositive, no usable conversation is selected, there is neither older head content nor a prior compaction, or the summary prompt exceeds context - summaryOutput.
• 🟡 (08:37) compactAfterOverflow uses the first "compaction" entry as previousSummary; its prompt context combines that entry’s saved recent text with newly selected head, while the prior summary text is supplied separately to buildPrompt.
• 🟡 (08:37) Summary generation uses the same model and input.request.http, sends one Message.user(summaryPrompt), disables tools with tools: [], and sets generation.maxTokens to Math.min(configuredOutput || 4_096, 4_096).
• 🟡 (08:37) Before summary streaming, compactAfterOverflow publishes SessionEvent.Compaction.Started with a fresh SessionMessage.ID, current DateTime, and reason: "auto"; input.eventCommit is attached to this start event when provided, and optional input.beforeStream() runs after the start publication but before calling the LLM.
• 🟡 (08:37) Summary streaming concatenates LLMEvent.is.textDelta chunks and marks any LLMEvent.is.providerError as failure; an LLM.Error, provider error, unsuccessful stream, or blank summary returns false.
• 🟡 (08:37) Successful summary generation publishes SessionEvent.Compaction.Ended with the same messageID, current DateTime, reason: "auto", generated text, and recent: selected.recent; this end event uses { commit: input.commit }, after which compactAfterOverflow returns true.
• 🟡 (08:37) SessionCompaction.make(dependencies) exposes exactly compactIfNeeded and compactAfterOverflow; dependencies are EventV2.Interface, an LLM stream(request) function, and readonly Config.Entry[].