@vellumai/assistant 0.6.4 → 0.6.6

package/src/context/window-manager.ts

@@ -1,3 +1,6 @@
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+
 import type { ContextWindowConfig } from "../config/types.js";
 import type {
   ContentBlock,
@@ -5,6 +8,7 @@ import type {
   Message,
   Provider,
 } from "../providers/types.js";
+import { resolveBundledDir } from "../util/bundled-asset.js";
 import { getLogger } from "../util/logger.js";
 import { safeStringSlice } from "../util/unicode.js";
 import {
@@ -26,20 +30,204 @@ const COMPACTION_TOOL_RESULT_MAX_CHARS = 6_000;
 const MIN_COMPACTABLE_PERSISTED_MESSAGES = 2;
 const INTERNAL_CONTEXT_SUMMARY_MESSAGES = new WeakSet<Message>();
 
-const SUMMARY_SYSTEM_PROMPT = [
-  "You compress long assistant conversations into durable working memory.",
-  "Focus on actionable state, not prose.",
-  "Preserve concrete facts: goals, constraints, decisions, pending questions, file paths, commands, errors, and TODOs.",
-  "Remove repetition and stale details that were superseded.",
-  "Return concise markdown using these section headers exactly:",
-  "## Goals",
-  "## Constraints",
-  "## Decisions",
-  "## Open Conversations",
-  "## Key Artifacts",
-  "## Recent Progress",
+/**
+ * When the existing summary is this fraction or more of the per-summary
+ * token budget, inject a "compress older content aggressively" instruction
+ * so incremental-update passes don't let the summary grow unboundedly.
+ */
+const SUMMARY_COMPRESSION_PRESSURE_RATIO = 0.6;
+
+/**
+ * Text-block prefixes that persist in live history (for prefix-caching
+ * stability and model grounding) but pollute the summarizer's view of the
+ * actual conversation. These blocks are system-metadata attached to user
+ * turns — memory injections, turn context, workspace hints, etc. They are
+ * stripped ONLY from the messages fed to the summarization LLM call. Live
+ * history is never mutated, so prefix caching is preserved.
+ *
+ * This list intentionally overlaps with `RUNTIME_INJECTION_PREFIXES` in
+ * `conversation-runtime-assembly.ts`. That list governs in-flight turn
+ * assembly via pure prefix matching; this one governs compaction input.
+ * Keep the two lists in sync when a new injection type is added.
+ *
+ * Compaction strip coverage is two-tier: this prefix list catches
+ * internal-vocabulary tags and any tag carrying the `__injected`
+ * attribute, while `COMPACTION_ONLY_WRAPPED_STRIP_TAGS` below matches
+ * ambiguous bare-tag blocks that are shaped like a runtime-emitted
+ * open/close wrap. A new ambiguous tag added upstream needs to be
+ * evaluated against both tiers — internal-vocabulary names go here,
+ * and names whose bare form collides with ordinary English
+ * (`<memory>`, `<workspace>`, `<knowledge_base>`, `<pkb>`,
+ * `<system_reminder>`) go in the wrapped-strip list so user prose
+ * mentioning the tag is preserved.
+ */
+const COMPACTION_ONLY_STRIP_PREFIXES = [
+  "<memory __injected>",
+  "<memory_image __injected>",
+  "</memory_image>",
+  "<memory_context __injected>",
+  "<turn_context>",
+  "<channel_turn_context>",
+  "<guardian_context>",
+  "<inbound_actor_context>",
+  "<interface_turn_context>",
+  "<workspace_top_level>",
+  "<now_scratchpad>",
+  "<NOW.md Always keep this up to date",
+  "<active_thread>",
+  "<active_subagents>",
+  "<active_workspace>",
+  "<active_dynamic_page>",
+  "<channel_capabilities>",
+  "<channel_command_context>",
+  "<voice_call_control>",
+  "<transport_hints>",
+  "<system_notice>",
+  "<non_interactive_context>",
+  "<temporal_context>",
+];
+
+/**
+ * Tags whose bare form (`<tag>`) is common English vocabulary or markup a
+ * user might legitimately type in prose. For these we only strip a text
+ * block if it is shaped exactly like a runtime injection: starts with
+ * `<tag>\n` and ends with `</tag>`. This bare-tag wrapped shape
+ * (e.g. `<memory>\n...\n</memory>`) appears in persisted history
+ * alongside the `__injected`-attributed variants, which the prefix list
+ * above already catches via `<memory __injected>`. A user who mentions
+ * `<memory>` in a sentence or inlines `<workspace>...</workspace>` within
+ * other prose will not match this shape.
+ */
+const COMPACTION_ONLY_WRAPPED_STRIP_TAGS = [
+  "memory",
+  "memory_context",
+  "workspace",
+  "knowledge_base",
+  "pkb",
+  "system_reminder",
+];
+
+function isCompactionInjectedBlock(text: string): boolean {
+  if (COMPACTION_ONLY_STRIP_PREFIXES.some((p) => text.startsWith(p))) {
+    return true;
+  }
+  return COMPACTION_ONLY_WRAPPED_STRIP_TAGS.some(
+    (tag) => text.startsWith(`<${tag}>\n`) && text.endsWith(`</${tag}>`),
+  );
+}
+
+/**
+ * Remove text blocks that look like runtime injections from user messages.
+ * Non-text blocks (images, tool_use, tool_result, etc.) are untouched.
+ * Empty messages (every block filtered out) are dropped from the output.
+ *
+ * Used only on the `compactableMessages` slice right before it is
+ * serialized for the summarization LLM — the caller's original message
+ * array is never mutated.
+ */
+export function stripCompactionOnlyInjections(messages: Message[]): Message[] {
+  return messages
+    .map((message) => {
+      if (message.role !== "user") return message;
+      const nextContent = message.content.filter((block) => {
+        if (block.type !== "text") return true;
+        return !isCompactionInjectedBlock(block.text);
+      });
+      if (nextContent.length === message.content.length) return message;
+      if (nextContent.length === 0) return null;
+      return { ...message, content: nextContent };
+    })
+    .filter(
+      (message): message is NonNullable<typeof message> => message != null,
+    );
+}
+
+/**
+ * Load the compaction summary system prompt from the bundled markdown asset.
+ *
+ * `resolveBundledDir` handles the compiled-binary case where the caller path
+ * points to `/$bunfs/` and the asset lives next to the executable (macOS app
+ * bundle `Contents/Resources/` or sibling dir). In source mode it falls back
+ * to the sibling `prompts/` directory.
+ */
+export function loadCompactPrompt(): string {
+  const callerDir = import.meta.dirname ?? __dirname;
+  const promptsDir = resolveBundledDir(callerDir, "prompts", "compact-prompts");
+  const promptPath = join(promptsDir, "compact.md");
+  const contents = readFileSync(promptPath, "utf-8");
+  if (contents.length === 0) {
+    throw new Error(
+      `compact.md at ${promptPath} is empty — compaction summary prompt missing`,
+    );
+  }
+  return contents;
+}
+
+/**
+ * Hardcoded fallback prompt used when the bundled `compact.md` asset is
+ * missing or unreadable, so the daemon can still compact conversations
+ * rather than failing module import at startup.
+ */
+const SUMMARY_PROMPT_FALLBACK = [
+  "You are summarizing a long conversation so that the assistant can keep working with it after older messages are dropped. Your summary will REPLACE those messages — the assistant's only access to what was said earlier will be what you write here.",
+  "",
+  "Be thorough. Capture what happened, why it mattered, what's unresolved, and what was felt. Do not compress away emotional tone, relationship context, or nuance. Keep specific details (names, numbers, file paths, commands, URLs, exact phrasings) when they might matter later.",
+  "",
+  "Target length: aim for 1500–4000 tokens. Use the upper end when the conversation is rich in decisions, relationships, emotional content, or threads that are still open. Use the lower end for short or simple task execution.",
+  "",
+  "Open with a 1–2 paragraph narrative describing what the conversation is about and where it currently stands. Then use `## ` section headers. Use these when they apply; skip sections that have nothing to say; add your own headers when something doesn't fit:",
+  "- `## What We're Working On`",
+  "- `## Decisions & Commitments`",
+  "- `## Facts Worth Remembering`",
+  "- `## Open Threads`",
+  "- `## Emotional Arc / Relationship Notes` (include when relevant)",
+  "- `## Artifacts & References`",
+  "",
+  "If an existing summary is provided, update it: merge new information in, prefer the most recent and explicit detail on conflicts, and preserve anything still unresolved or still true. Do not restart from scratch.",
+  "",
+  "Never include in the summary: content inside `<memory __injected>`, `<memory>`, `<turn_context>`, `<workspace>`, `<knowledge_base>`, `<system_reminder>`, `<now_scratchpad>`, `<NOW.md …>`, `<active_thread>`, `<channel_capabilities>`, `<transport_hints>`, `<system_notice>`, or any other angle-bracket-tagged system blocks. Tool-call boilerplate (retries, failed attempts the assistant recovered from, routine status updates) — summarize the outcome instead. Repetitive chit-chat that adds nothing.",
+  "",
+  'Thread anchors (Slack only): if the input includes a "Retained Thread References" section, each listed reply cites its parent via `→ Mxxxxxx`. If that parent appears in the Transcript, preserve its text verbatim. Omit when absent.',
+  "",
+  "Return only the summary itself in markdown — no preamble, no meta-commentary.",
 ].join("\n");
 
+/**
+ * Load the compact prompt with graceful fallback. If `loader` throws (missing
+ * or unreadable bundled asset, partial deployment, filesystem corruption),
+ * logs a warning and returns the hardcoded fallback string so module import
+ * never fails. The loader is injectable for testability.
+ */
+export function loadCompactPromptOrFallback(
+  loader: () => string = loadCompactPrompt,
+): string {
+  try {
+    return loader();
+  } catch (err) {
+    log.warn(
+      { err },
+      "Failed to load compact.md from bundle; using inline fallback prompt. The bundled asset may be missing or unreadable.",
+    );
+    return SUMMARY_PROMPT_FALLBACK;
+  }
+}
+
+const SUMMARY_SYSTEM_PROMPT = loadCompactPromptOrFallback();
+
+/**
+ * Pattern matching a Slack-style reply tag-line's parent-alias reference.
+ * The chronological renderer emits reply lines as
+ * `[MM/DD/YY HH:MM @sender → Mxxxxxx]: body`, or, for edited replies,
+ * `[MM/DD/YY HH:MM @sender → Mxxxxxx, edited MM/DD/YY HH:MM]: body`. The
+ * character after the 6-hex parent alias is therefore `]` for a plain reply
+ * or `,` for an edited one — the regex accepts either. `Mxxxxxx` is the
+ * first 6 hex chars of sha256(threadTs). A retained-tail text block that
+ * contains this pattern is carrying a live reference to a parent that may
+ * still live in the compactable region — the summarizer needs to know about
+ * it to act on the Thread-anchors clause of SUMMARY_SYSTEM_PROMPT.
+ */
+const THREAD_REPLY_REFERENCE_PATTERN = /→ M[0-9a-f]{6}[,\]]/;
+
 export interface ContextWindowResult {
   messages: Message[];
   compacted: boolean;
@@ -58,6 +246,13 @@ export interface ContextWindowResult {
   summaryRawResponses?: unknown[];
   summaryText: string;
   reason?: string;
+  /**
+   * True when the summary LLM call threw and the local fallback produced the
+   * summary. Callers use this to distinguish provider-side summary failures
+   * from successful compactions so they can apply circuit-breaker logic
+   * without losing the fallback-compacted messages.
+   */
+  summaryFailed?: boolean;
 }
 
 export interface ShouldCompactResult {
@@ -73,13 +268,22 @@ export interface ContextWindowCompactOptions {
    * Override the minimum number of recent user turns to preserve.
    * Set to `0` for emergency recovery that can compact the entire history
    * (except the summary message itself). When omitted, the default floor
-   * of 1 recent user turn is enforced.
+   * is `1` (or `8` when `conversationOriginChannel === "slack"`).
    */
   minKeepRecentUserTurns?: number;
+  /**
+   * Origin channel hint used when `minKeepRecentUserTurns` is omitted.
+   * Slack-originated conversations bump the default keep floor so multi-turn
+   * thread context (replies, quoted messages) is not summarized away too
+   * aggressively. Explicit `minKeepRecentUserTurns` overrides this hint.
+   */
+  conversationOriginChannel?: string;
   /**
    * Override the target input token budget used for keep-boundary
-   * projected-fit checks. Allows the caller to demand a stricter fit
-   * than the normal `config.targetInputTokens` during forced recovery.
+   * projected-fit checks. Clamped to no looser than `config.targetInputTokens`
+   * — i.e. the override may only demand a *stricter* fit. Passing a looser
+   * value has no effect. Intended for forced recovery paths that need a
+   * tighter target than the default.
    */
   targetInputTokensOverride?: number;
   /**
@@ -274,6 +478,8 @@ export class ContextWindowManager {
     const keepPlan = this.pickKeepBoundary(messages, userTurnStarts, {
       minKeepRecentUserTurns: options?.minKeepRecentUserTurns,
       targetInputTokensOverride: options?.targetInputTokensOverride,
+      conversationOriginChannel: options?.conversationOriginChannel,
+      force: options?.force,
     });
     if (keepPlan.keepFromIndex <= summaryOffset) {
       // All turns fit after truncation projection, but the real in-memory
@@ -447,13 +653,24 @@ export class ContextWindowManager {
       };
     }
 
+    const retainedThreadRefs = collectRetainedThreadReferences(
+      messages.slice(keepPlan.keepFromIndex),
+    );
+    // Strip runtime injections (memory, turn context, workspace hints, etc.)
+    // from the messages fed to the summarizer. These blocks are system
+    // metadata; leaving them in causes the summary to echo rotating memory
+    // content instead of the actual conversation. The caller's live message
+    // array is untouched so prefix caching stays intact.
+    const transcriptSource = stripCompactionOnlyInjections(compactableMessages);
     const transcriptBlocks = this.capTranscriptBlocksToTokenBudget(
-      serializeMessagesToContentBlocks(compactableMessages),
+      serializeMessagesToContentBlocks(transcriptSource),
       existingSummary ?? "No previous summary.",
+      retainedThreadRefs,
     );
     const summaryUpdate = await this.updateSummary(
       existingSummary ?? "No previous summary.",
       transcriptBlocks,
+      retainedThreadRefs,
       signal,
     );
     const summary = summaryUpdate.summary;
@@ -463,6 +680,7 @@ export class ContextWindowManager {
     const summaryCacheCreationInputTokens =
       summaryUpdate.cacheCreationInputTokens;
     const summaryCacheReadInputTokens = summaryUpdate.cacheReadInputTokens;
+    const summaryFailed = summaryUpdate.failed;
     const summaryRawResponses: unknown[] = [];
     if (Array.isArray(summaryUpdate.rawResponse)) {
       summaryRawResponses.push(...summaryUpdate.rawResponse);
@@ -499,7 +717,9 @@ export class ContextWindowManager {
     // the summary at index 0 as child-owned.
     this.nonPersistedPrefixCount = Math.max(
       0,
-      this.nonPersistedPrefixCount - injectedInCompactable - injectedSummaryOffset,
+      this.nonPersistedPrefixCount -
+        injectedInCompactable -
+        injectedSummaryOffset,
     );
     this.summaryIsInjected = false;
 
@@ -532,6 +752,7 @@ export class ContextWindowManager {
       summaryCacheReadInputTokens,
       summaryRawResponses,
       summaryText: summary,
+      summaryFailed,
     };
   }
 
@@ -548,14 +769,24 @@ export class ContextWindowManager {
     opts?: {
       minKeepRecentUserTurns?: number;
       targetInputTokensOverride?: number;
+      conversationOriginChannel?: string;
+      force?: boolean;
     },
   ): { keepFromIndex: number; keepTurns: number } {
+    // Slack-originated conversations rely on multi-turn thread context
+    // (reply chains, quoted messages, contextual references). Bump the
+    // default keep floor for them so compaction does not summarize away
+    // recent turns that the next reply may directly cite. Explicit
+    // `minKeepRecentUserTurns` (including emergency `0`) wins.
+    const defaultTurns = opts?.conversationOriginChannel === "slack" ? 8 : 1;
     const minFloor = Math.min(
-      Math.max(0, Math.floor(opts?.minKeepRecentUserTurns ?? 1)),
+      Math.max(0, Math.floor(opts?.minKeepRecentUserTurns ?? defaultTurns)),
       userTurnStarts.length,
     );
-    const targetTokens =
-      opts?.targetInputTokensOverride ?? this.targetInputTokens;
+    const targetTokens = Math.min(
+      opts?.targetInputTokensOverride ?? this.targetInputTokens,
+      this.targetInputTokens,
+    );
 
     // Binary search for the maximum keepTurns whose projected tokens fit
     // within the budget. Token count is monotonically non-decreasing with
@@ -597,6 +828,32 @@ export class ContextWindowManager {
       lo = hi;
     }
 
+    // Under forced compaction with only the implicit default floor in play,
+    // that floor stops being an absolute override when the kept region still
+    // exceeds the target. Walk keepTurns below the floor — down to 0 if
+    // needed — so /compact can always drive the conversation toward target,
+    // even when the floor turn itself is oversized (e.g. a huge paste in the
+    // last user message). Exceptions that still treat the floor as hard:
+    //   - Explicit `minKeepRecentUserTurns` (the caller opted in to that
+    //     floor; emergency recovery already passes 0 when it wants to go all
+    //     the way down).
+    //   - Slack origin (the bumped 8-turn floor protects thread reply chains
+    //     and quoted-message context that the next reply may directly cite).
+    // Automatic mid-loop compaction (force !== true) always honors the floor
+    // so the in-flight agent turn isn't summarized away.
+    const floorIsImplicitDefault =
+      opts?.minKeepRecentUserTurns === undefined &&
+      opts?.conversationOriginChannel !== "slack";
+    if (
+      opts?.force &&
+      floorIsImplicitDefault &&
+      projectedTokensForKeep(lo) > targetTokens
+    ) {
+      while (lo > 0 && projectedTokensForKeep(lo) > targetTokens) {
+        lo--;
+      }
+    }
+
     const keepTurns = lo;
     const rawKeepFromIndex =
       keepTurns === 0
@@ -628,10 +885,13 @@ export class ContextWindowManager {
   private capTranscriptBlocksToTokenBudget(
     blocks: ContentBlock[],
     currentSummary: string,
+    retainedThreadRefs: string[],
   ): ContentBlock[] {
+    const retainedRefsText = retainedThreadRefs.join("\n");
     const overheadTokens =
       estimateTextTokens(SUMMARY_SYSTEM_PROMPT) +
       estimateTextTokens(currentSummary) +
+      estimateTextTokens(retainedRefsText) +
       // Scaffolding text in buildSummaryContentBlocks ("Update the summary...",
       // section headers, etc.) — generous fixed estimate.
       200 +
@@ -643,7 +903,9 @@ export class ContextWindowManager {
     );
 
     const estimateBlockTokens = (b: ContentBlock): number =>
-      estimateContentBlockTokens(b, { providerName: this.estimationProviderName });
+      estimateContentBlockTokens(b, {
+        providerName: this.estimationProviderName,
+      });
 
     let totalTokens = 0;
     for (const block of blocks) {
@@ -656,7 +918,11 @@ export class ContextWindowManager {
     // images to drop. Images are high-cost and their text context (message
     // headers, surrounding tool_use/tool_result serializations) is preserved.
     const result = [...blocks];
-    for (let i = 0; i < result.length && totalTokens > maxTranscriptTokens; i++) {
+    for (
+      let i = 0;
+      i < result.length && totalTokens > maxTranscriptTokens;
+      i++
+    ) {
       if (result[i].type === "image") {
         totalTokens -= estimateBlockTokens(result[i]);
         const stub: ContentBlock = {
@@ -674,7 +940,11 @@ export class ContextWindowManager {
     // than dropping it entirely so the summarizer always has content to work with.
     let dropUntil = 0;
     let droppedTokens = 0;
-    for (let i = 0; i < result.length && totalTokens > maxTranscriptTokens; i++) {
+    for (
+      let i = 0;
+      i < result.length && totalTokens > maxTranscriptTokens;
+      i++
+    ) {
       const blockTokens = estimateBlockTokens(result[i]);
       const excess = totalTokens - maxTranscriptTokens;
       if (blockTokens > excess && result[i].type === "text") {
@@ -722,6 +992,7 @@ export class ContextWindowManager {
   private async updateSummary(
     currentSummary: string,
     transcriptBlocks: ContentBlock[],
+    retainedThreadRefs: string[],
     signal?: AbortSignal,
   ): Promise<{
     summary: string;
@@ -731,19 +1002,38 @@ export class ContextWindowManager {
     cacheCreationInputTokens: number;
     cacheReadInputTokens: number;
     rawResponse?: unknown;
+    /**
+     * True when the provider.sendMessage call threw and the local fallback
+     * was used. Callers (the agent loop) use this to drive circuit-breaker
+     * state without having to reimplement the fallback themselves.
+     */
+    failed: boolean;
   }> {
+    // When the existing summary is already consuming most of its budget,
+    // nudge the model to compress older durable content aggressively so
+    // incremental-update passes don't let the summary grow unboundedly.
+    const existingSummaryTokens = estimateTextTokens(currentSummary);
+    const compressionPressure =
+      existingSummaryTokens >=
+      this.summaryMaxTokens * SUMMARY_COMPRESSION_PRESSURE_RATIO;
     const contentBlocks = buildSummaryContentBlocks(
       currentSummary,
       transcriptBlocks,
+      retainedThreadRefs,
+      { compressionPressure },
     );
     const summaryMessage: Message = { role: "user", content: contentBlocks };
+    let failed = false;
     try {
       const response = await this.provider.sendMessage(
         [summaryMessage],
         undefined,
         SUMMARY_SYSTEM_PROMPT,
         {
-          config: { max_tokens: this.summaryMaxTokens },
+          config: {
+            callSite: "conversationSummarization" as const,
+            max_tokens: this.summaryMaxTokens,
+          },
           signal,
         },
       );
@@ -759,15 +1049,19 @@ export class ContextWindowManager {
             response.usage.cacheCreationInputTokens ?? 0,
           cacheReadInputTokens: response.usage.cacheReadInputTokens ?? 0,
           rawResponse: response.rawResponse,
+          failed: false,
         };
       }
     } catch (err) {
+      failed = true;
       log.warn({ err }, "Summary generation failed, using local fallback");
     }
 
     // Fallback: extract text-only transcript for local summary generation.
     const textTranscript = transcriptBlocks
-      .filter((b): b is Extract<ContentBlock, { type: "text" }> => b.type === "text")
+      .filter(
+        (b): b is Extract<ContentBlock, { type: "text" }> => b.type === "text",
+      )
       .map((b) => b.text)
       .join("\n\n");
 
@@ -778,6 +1072,7 @@ export class ContextWindowManager {
       model: "",
       cacheCreationInputTokens: 0,
       cacheReadInputTokens: 0,
+      failed,
     };
   }
 
@@ -785,10 +1080,38 @@ export class ContextWindowManager {
     // Budget in tokens → approximate char limit (4 chars ≈ 1 token).
     const maxChars = this.summaryMaxTokens * 4;
     if (summary.length <= maxChars) return summary;
-    return `${safeStringSlice(summary, 0, maxChars)}...`;
+    return clampSummaryAtSectionBoundary(summary, maxChars);
   }
 }
 
+/**
+ * Truncate a markdown summary that exceeds `maxChars`, preferring a
+ * section boundary (`\n## `) so we never cut a heading mid-text. Falls
+ * back to a hard character slice when no boundary exists in the safe
+ * region (first half of the budget).
+ */
+export function clampSummaryAtSectionBoundary(
+  summary: string,
+  maxChars: number,
+): string {
+  if (summary.length <= maxChars) return summary;
+  const ELLIPSIS = "...";
+  // Hard limit we must stay under, leaving room for the ellipsis suffix.
+  const cutoff = maxChars - ELLIPSIS.length;
+  if (cutoff <= 0) return ELLIPSIS;
+  const head = safeStringSlice(summary, 0, cutoff);
+  // Find the last `## ` heading at a line start. Require it to be past the
+  // midpoint of the allowed region so we don't drop most of the summary
+  // just to hit a boundary — better to cut mid-section late than to keep
+  // almost nothing.
+  const halfway = Math.floor(cutoff / 2);
+  const boundary = head.lastIndexOf("\n## ");
+  if (boundary >= halfway) {
+    return `${head.slice(0, boundary).trimEnd()}\n${ELLIPSIS}`;
+  }
+  return `${head}${ELLIPSIS}`;
+}
+
 function collectUserTurnStartIndexes(messages: Message[]): number[] {
   const starts: number[] = [];
   for (let i = 0; i < messages.length; i++) {
@@ -854,7 +1177,11 @@ function adjustForToolPairs(
     // Collect tool_use_ids referenced by tool_results in this user message
     const referencedIds = new Set<string>();
     for (const block of msg.content) {
-      if ((block.type === "tool_result" || block.type === "web_search_tool_result") && "tool_use_id" in block) {
+      if (
+        (block.type === "tool_result" ||
+          block.type === "web_search_tool_result") &&
+        "tool_use_id" in block
+      ) {
         referencedIds.add((block as { tool_use_id: string }).tool_use_id);
       }
     }
@@ -930,26 +1257,76 @@ export function createContextSummaryMessage(summary: string): Message {
 function buildSummaryContentBlocks(
   currentSummary: string,
   transcriptBlocks: ContentBlock[],
+  retainedThreadRefs: string[],
+  options: { compressionPressure: boolean } = { compressionPressure: false },
 ): ContentBlock[] {
+  const lines = [
+    "Update the summary with new transcript data.",
+    "If new information conflicts with older notes, keep the most recent and explicit detail.",
+    "Keep all unresolved asks and next steps.",
+    "For any images included below, describe their visual content in the summary so the information is preserved after compaction.",
+  ];
+  if (options.compressionPressure) {
+    lines.push(
+      "The existing summary is approaching its token budget. Compress older durable content aggressively (drop detail that is no longer load-bearing, merge bullets, tighten prose) while preserving the most recent turns' nuance.",
+    );
+  }
+  lines.push(
+    "",
+    "### Existing Summary",
+    currentSummary.trim().length > 0 ? currentSummary.trim() : "None.",
+    "",
+  );
+  if (retainedThreadRefs.length > 0) {
+    lines.push(
+      "### Retained Thread References",
+      "These reply tag lines remain in the live context after compaction. Each `→ Mxxxxxx` cites a parent message by alias; if that parent appears in the Transcript below, preserve its text verbatim.",
+      ...retainedThreadRefs.map((ref) => `- ${ref}`),
+      "",
+    );
+  }
+  lines.push("### Transcript");
   return [
     {
       type: "text",
-      text: [
-        "Update the summary with new transcript data.",
-        "If new information conflicts with older notes, keep the most recent and explicit detail.",
-        "Keep all unresolved asks and next steps.",
-        "For any images included below, describe their visual content in the summary so the information is preserved after compaction.",
-        "",
-        "### Existing Summary",
-        currentSummary.trim().length > 0 ? currentSummary.trim() : "None.",
-        "",
-        "### Transcript",
-      ].join("\n"),
+      text: lines.join("\n"),
     } as ContentBlock,
     ...transcriptBlocks,
   ];
 }
 
+/**
+ * Scan retained-tail messages for Slack-style reply tag lines that cite a
+ * thread parent via the `→ Mxxxxxx` alias convention. Returns the full tag
+ * line for each match (de-duplicated, order-preserved) so the summarizer
+ * has a concrete list of parents whose text must be preserved verbatim.
+ *
+ * Non-slack conversations and retained tails without any reply markers
+ * produce an empty list — in that case the summarizer is told explicitly
+ * that no verbatim preservation is required.
+ */
+function collectRetainedThreadReferences(
+  retainedMessages: Message[],
+): string[] {
+  const seen = new Set<string>();
+  const out: string[] = [];
+  for (const msg of retainedMessages) {
+    for (const block of msg.content) {
+      if (block.type !== "text") continue;
+      const text = (block as { text: string }).text;
+      for (const line of text.split("\n")) {
+        if (!THREAD_REPLY_REFERENCE_PATTERN.test(line)) continue;
+        const trimmed = line.trim();
+        if (trimmed.length === 0) continue;
+        if (seen.has(trimmed)) continue;
+        seen.add(trimmed);
+        out.push(trimmed);
+      }
+    }
+  }
+  return out;
+}
+
 /**
  * Serialize messages into a sequence of content blocks. Text-based content
  * (tool calls, tool results, thinking, etc.) is serialized into text blocks.
@@ -970,7 +1347,8 @@ function serializeMessagesToContentBlocks(messages: Message[]): ContentBlock[] {
           textLines.length = 0;
         }
         blocks.push(block);
-      } else if (block.type === "tool_result") { // guard:allow-tool-result-only — web_search_tool_result handled by serializeBlock via else branch
+      } else if (block.type === "tool_result") {
+        // guard:allow-tool-result-only — web_search_tool_result handled by serializeBlock via else branch
         // Extract images from tool_result contentBlocks before serializing.
         const collectedImages: ImageContent[] = [];
         textLines.push(serializeToolResultBlock(block, collectedImages));

package/src/credential-execution/approval-bridge.ts

@@ -87,7 +87,6 @@ function mapUserDecisionToCesDecision(
         userDecision: decision,
       };
     case "always_allow":
-    case "always_allow_high_risk":
       // Persistent grant — no expiry. CES stores it with expiresAt: null.
       return {
         grantDecision: "approved",