Distillation

Date: Sep 16, 2026

• 🔴 (11:15) User directed that local ownership be replaced with durable multi-node ownership when clustered.
• 🔴 (11:15) SessionRunner’s roadmap still has unchecked work to honor interruption and reject stale work after runtime attachment replacement.
• 🔴 (11:15) SessionRunner’s completed roadmap items include persisting assistant text and usage events incrementally as they arrive and reloading projected history to start the next explicit provider turn after local tool results.
• 🔴 (11:15) packages/core/src/session/compaction.ts defines compaction defaults DEFAULT_BUFFER = 20_000, DEFAULT_KEEP_TOKENS = 8_000, TOOL_OUTPUT_MAX_CHARS = 2_000, and SUMMARY_OUTPUT_TOKENS = 4_096.
• 🔴 (11:15) The compaction summary template requires sections in this exact order: ## Objective, ## Important Details, ## Work State with ### Completed, ### Active, and ### Blocked, then ## Next Move with numbered items 1 and 2, and ## Relevant Files.
• 🔴 (11:15) Compaction-summary rules require every section even when empty, terse bullets rather than prose paragraphs, preservation of exact file paths/symbols/commands/error strings/URLs/identifiers, and no mention that summarization or context compaction occurred.
• 🔴 (11:15) SUMMARY_UPDATE_INSTRUCTIONS says a new summary must combine <prior-summary> and <conversation> because the prior summary is discarded; carry forward unfinished objectives, constraints, directives, decisions, and parallel workstreams; prefer the newer conversation on conflicts; add new progress/context; move completed work to Completed; update resolved blockers while retaining continuation details; and refresh Objective and Next Move.
• 🔴 (11:15) packages/core/src/session/compaction.ts resolves configuration by reducing all document entries with info.compaction; later configured values override earlier ones through auto, buffer, and keep.tokens, with defaults { auto: true, buffer: 20_000, tokens: 8_000 }.
• 🔴 (11:15) serializeToolContent(...) renders text items directly and attachments as [Attached ${item.mime}] or [Attached ${item.mime}: ${item.name}]; serialized tool and shell output longer than 2_000 characters is sliced to 2_000 characters and suffixed with \n[truncated].
• 🔴 (11:15) Compaction serialization labels messages as [User], [Assistant], [Assistant reasoning], [Assistant tool call], [Tool result], [Tool error], [System update], [Synthetic context], and [Shell]; user files are represented as [Attached ${file.mime}: ${file.name ?? file.uri}].
• 🔴 (11:15) The compaction select(...) algorithm excludes messages with type === "compaction", serializes nonempty messages, then walks backward while the estimated token total remains within the keep-token limit; it returns older content as head and retained tail content as recent.
• 🔴 (11:15) buildPrompt(...) creates a new anchored summary from <conversation> when there is no prior summary. With a prior summary, it includes <conversation>, <prior-summary>, SUMMARY_UPDATE_INSTRUCTIONS, and SUMMARY_TEMPLATE in that order.
• 🔴 (11:15) SessionCompaction.compactIfNeeded(...) returns false when automatic compaction is disabled, the model context limit is absent/nonpositive, or estimated { system, messages, tools } input is at most context - Math.max(output, config.buffer); otherwise it delegates to compactAfterOverflow(...).
• 🔴 (11:15) SessionCompaction.compactAfterOverflow(...) uses input.request.generation?.maxTokens, falling back to the model output limit and then 0; chooses a summary output limit of Math.min(output || 4_096, 4_096); and refuses compaction if the summary prompt estimate exceeds context - summaryOutput.
• 🔴 (11:15) For repeated compaction, compactAfterOverflow(...) finds the first existing compaction message, supplies its summary as previousSummary, and combines its saved recent text with the newly selected older head; it returns false if there is neither selected head content nor an existing compaction summary.
• 🔴 (11:15) Before summary streaming, compaction creates a SessionMessage.ID, publishes SessionEvent.Compaction.Started with reason: "auto" and an optional eventCommit, then invokes optional beforeStream().
• 🔴 (11:15) The summarization LLM request reuses the selected model and original HTTP configuration, sends one Message.user(summaryPrompt), disables tools with tools: [], and sets generation.maxTokens to the computed summary output limit.
• 🔴 (11:15) Summary streaming concatenates all LLMEvent.is.textDelta(...) text, marks any LLMEvent.is.providerError(...) as failure, converts tagged LLM.Error failure to false, and declines to publish a completed compaction when streaming failed, a provider error occurred, or the resulting summary is blank.
• 🔴 (11:15) Successful automatic compaction publishes SessionEvent.Compaction.Ended with the same sessionID and messageID, a fresh timestamp, reason: "auto", generated text, and selected recent, using { commit: input.commit }; it then returns true.