@loreai/core 0.10.1 → 0.11.0

package/dist/bun/index.js

@@ -122,9 +122,11 @@ var require_extend = __commonJS({
 // src/temporal.ts
 var temporal_exports = {};
 __export(temporal_exports, {
+  CHUNK_TERMINATOR: () => CHUNK_TERMINATOR,
   bySession: () => bySession,
   count: () => count,
   markDistilled: () => markDistilled,
+  partsToText: () => partsToText,
   prune: () => prune,
   search: () => search2,
   searchScored: () => searchScored,
@@ -419,6 +421,58 @@ var MIGRATIONS = [
     to_id TEXT NOT NULL REFERENCES knowledge(id) ON DELETE CASCADE,
     PRIMARY KEY (from_id, to_id)
   );
+  `,
+  `
+  -- Version 11: F3b -- unambiguous chunk terminator in temporal_messages.content.
+  --
+  -- Pre-F3b, partsToText joined chunks with a newline. Tool-output payloads
+  -- can contain newlines too, so the boundary between a tool envelope and a
+  -- following plain-text or [reasoning] chunk was structurally ambiguous.
+  -- This caused two known limitations in the F3 distill-input truncator:
+  -- trailing text could be swallowed into a tool payload, and embedded
+  -- literal envelope strings inside a payload (e.g. when reading AGENTS.md)
+  -- could fabricate fake boundaries.
+  --
+  -- F3b switches the chunk separator to newline plus ASCII Unit Separator
+  -- (char 31). The Unit Separator is non-word so FTS5's unicode61 tokenizer
+  -- ignores it (zero BM25 impact). New rows are written via the post-F3b
+  -- partsToText. Existing rows are rewritten in place by the UPDATE below,
+  -- which uses pure SQL replace() to inject the Unit Separator after every
+  -- legacy chunk-prefix sequence -- the same boundary patterns the legacy
+  -- F3 reader was already trying to recover.
+  --
+  -- Trade-off (acceptable): any embedded legacy chunk-prefix sequence
+  -- inside a tool payload becomes a structural boundary post-migration.
+  -- This matches what the legacy F3 reader did at read-time anyway, baked
+  -- into the row permanently. The migration runs once per machine.
+  --
+  -- Idempotent: a row that already contains the Unit Separator before a
+  -- chunk prefix no longer matches the search literal (the separator
+  -- interposes), so re-running the UPDATE is a no-op for migrated rows.
+  -- (Important: migrate() in db.ts runs each migration via database.exec()
+  -- with no explicit BEGIN/COMMIT around the whole loop. SQLite makes this
+  -- single UPDATE statement atomic per-statement, so partial progress on
+  -- crash is safe to retry thanks to the idempotency above.)
+  --
+  -- char(10) = newline, char(31) = Unit Separator. SQLite has no native
+  -- regex, but two nested replace() calls on the literal prefixes are
+  -- sufficient because both legacy chunk prefixes match at line-start.
+  --
+  -- Each row UPDATE fires the temporal_fts_update trigger once; because
+  -- the Unit Separator is a non-word character, the re-indexed content
+  -- tokenizes identically -- net no-op for FTS scoring.
+  UPDATE temporal_messages
+  SET content = replace(
+    replace(
+      content,
+      char(10) || '[tool:',
+      char(10) || char(31) || '[tool:'
+    ),
+    char(10) || '[reasoning] ',
+    char(10) || char(31) || '[reasoning] '
+  )
+  WHERE content LIKE '%' || char(10) || '[tool:%'
+     OR content LIKE '%' || char(10) || '[reasoning] %';
   `
 ];
 function dataDir() {
@@ -442,6 +496,7 @@ function db() {
   instance = new Database(path);
   instance.exec("PRAGMA journal_mode = WAL");
   instance.exec("PRAGMA foreign_keys = ON");
+  instance.exec("PRAGMA busy_timeout = 5000");
   instance.exec("PRAGMA auto_vacuum = INCREMENTAL");
   migrate(instance);
   return instance;
@@ -10748,12 +10803,27 @@ EXACT NUMBERS: When two segments report different numbers for what seems like th
 
 EARLY-SESSION CONTENT: Bug fixes, code changes, and decisions from the start of a session are just as important as later work. Never drop them just because the segment is short or old. If the first segment contains a specific bug fix with file paths and root cause, it MUST survive into the reflection.
 
+ANCHORED UPDATES: If the prompt includes a <previous-meta-summary> block, treat it as the current consolidated state. Update it using the NEW observation segments \u2014 preserve still-true details, remove stale details, and merge in new facts. Keep the same section headings. Do NOT re-derive unchanged sections verbatim unless the new segments contradict them.
+
 Output ONLY an <observations> block with the consolidated observations.`;
-function recursiveUser(distillations) {
+function recursiveUser(distillations, previousMeta) {
   const entries = distillations.map(
     (d, i) => `Segment ${i + 1}:
 ${d.observations}`
   );
+  if (previousMeta) {
+    return `Update the anchored meta-summary below using the NEW observation segments. Preserve still-true details, remove stale details, and merge in new facts. Keep the same section headings.
+
+<previous-meta-summary>
+${previousMeta}
+</previous-meta-summary>
+
+---
+
+New observation segments to merge (chronological order):
+
+${entries.join("\n\n---\n\n")}`;
+  }
   return `Observation segments to consolidate (chronological order):
 
 ${entries.join("\n\n---\n\n")}`;
@@ -10908,6 +10978,61 @@ function formatDistillations(distillations) {
   }
   return sections.join("\n\n");
 }
+var COMPACT_SUMMARY_TEMPLATE = `Output exactly this Markdown structure. Keep every section in this order, even when empty (use "(none)").
+
+---
+## Goal
+- [single-sentence task summary]
+
+## Constraints & Preferences
+- [user constraints, preferences, specs, or "(none)"]
+
+## Progress
+### Done
+- [completed work or "(none)"]
+
+### In Progress
+- [current work or "(none)"]
+
+### Blocked
+- [blockers or "(none)"]
+
+## Key Decisions
+- [decision and why, or "(none)"]
+
+## Next Steps
+- [ordered next actions or "(none)"]
+
+## Critical Context
+- [important technical facts, errors, open questions, or "(none)"]
+
+## Relevant Files
+- [file or directory path: why it matters, or "(none)"]
+---
+
+Rules:
+- Keep every section, even when empty.
+- Use terse bullets, not prose paragraphs.
+- Preserve exact file paths, commands, error strings, and identifiers when known.
+- Do not mention the summary process or that context was compacted.
+- End with "I'm ready to continue." on its own line after the closing "---".`;
+function buildCompactPrompt(input) {
+  const distillSection = input.hasDistillations ? "Lore has pre-computed chunked summaries of the session history (injected above as context). Use them as the authoritative source \u2014 do NOT re-read raw conversation messages that conflict with them.\n\n" : "";
+  const anchorBlock = input.previousSummary ? `A prior compacted summary exists for this session. Update it using the conversation history above: preserve still-true details, remove stale details, and merge in new facts. Keep every section in place.
+
+<previous-summary>
+${input.previousSummary}
+</previous-summary>
+
+` : "";
+  const knowledgeBlock = input.knowledge ? `
+${input.knowledge}
+` : "";
+  return `You are producing a compacted session summary for an AI coding agent. This summary will be the ONLY context available in the next part of the conversation.
+
+${distillSection}${anchorBlock}${COMPACT_SUMMARY_TEMPLATE}
+${knowledgeBlock}`;
+}
 function estimateTokens(text4) {
   return Math.ceil(text4.length / 3);
 }
@@ -11160,6 +11285,7 @@ function isToolPart(p3) {
 function estimate(text4) {
   return Math.ceil(text4.length / 3);
 }
+var CHUNK_TERMINATOR = "";
 function partsToText(parts) {
   const chunks = [];
   for (const part of parts) {
@@ -11169,7 +11295,7 @@ function partsToText(parts) {
     else if (isToolPart(part) && part.state.status === "completed")
       chunks.push(`[tool:${part.tool}] ${part.state.output}`);
   }
-  return sanitizeSurrogates(chunks.join("\n"));
+  return sanitizeSurrogates(chunks.join("\n" + CHUNK_TERMINATOR));
 }
 function messageMetadata(info2, parts) {
   const meta3 = {};
@@ -11248,11 +11374,11 @@ function search2(input) {
   const limit = input.limit ?? 20;
   const q = ftsQuery(input.query);
   if (q === EMPTY_QUERY) return [];
-  const ftsSQL = input.sessionID ? `SELECT m.* FROM temporal_messages m
-       JOIN temporal_fts f ON m.rowid = f.rowid
+  const ftsSQL = input.sessionID ? `SELECT m.* FROM temporal_fts f
+       CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ? AND m.session_id = ?
-       ORDER BY rank LIMIT ?` : `SELECT m.* FROM temporal_messages m
-       JOIN temporal_fts f ON m.rowid = f.rowid
+       ORDER BY rank LIMIT ?` : `SELECT m.* FROM temporal_fts f
+       CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ?
        ORDER BY rank LIMIT ?`;
   const params = input.sessionID ? [q, pid, input.sessionID, limit] : [q, pid, limit];
@@ -11277,11 +11403,11 @@ function searchScored(input) {
   const limit = input.limit ?? 20;
   const q = ftsQuery(input.query);
   if (q === EMPTY_QUERY) return [];
-  const ftsSQL = input.sessionID ? `SELECT m.*, rank FROM temporal_messages m
-       JOIN temporal_fts f ON m.rowid = f.rowid
+  const ftsSQL = input.sessionID ? `SELECT m.*, rank FROM temporal_fts f
+       CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ? AND m.session_id = ?
-       ORDER BY rank LIMIT ?` : `SELECT m.*, rank FROM temporal_messages m
-       JOIN temporal_fts f ON m.rowid = f.rowid
+       ORDER BY rank LIMIT ?` : `SELECT m.*, rank FROM temporal_fts f
+       CROSS JOIN temporal_messages m ON m.rowid = f.rowid
        WHERE f.content MATCH ? AND m.project_id = ?
        ORDER BY rank LIMIT ?`;
   const params = input.sessionID ? [q, pid, input.sessionID, limit] : [q, pid, limit];
@@ -25483,18 +25609,63 @@ var LoreConfig = external_exports.object({
     providerID: external_exports.string(),
     modelID: external_exports.string()
   }).optional(),
+  /** Explicit worker model override. When set, all background workers (distillation,
+   *  curation, query expansion) use this model instead of the session model or the
+   *  auto-selected worker model. Bypasses dynamic worker model selection entirely. */
+  workerModel: external_exports.object({
+    providerID: external_exports.string(),
+    modelID: external_exports.string()
+  }).optional(),
   budget: external_exports.object({
     distilled: external_exports.number().min(0.05).max(0.5).default(0.25),
     raw: external_exports.number().min(0.1).max(0.7).default(0.4),
     output: external_exports.number().min(0.1).max(0.5).default(0.25),
-    /** Max fraction of usable context reserved for LTM system-prompt injection. Default: 0.10 (10%). */
-    ltm: external_exports.number().min(0.02).max(0.3).default(0.1)
-  }).default({ distilled: 0.25, raw: 0.4, output: 0.25, ltm: 0.1 }),
+    /** Max fraction of usable context reserved for LTM system-prompt injection. Default: 0.05 (5%). */
+    ltm: external_exports.number().min(0.02).max(0.3).default(0.05),
+    /** Per-turn cache-read cost target in dollars. Controls when layer 0 (full
+     *  passthrough) escalates to layer 1 (compressed). The cap is derived as:
+     *  maxLayer0Tokens = max(target / model.cost.cache.read, 40K).
+     *  Lower = cheaper but earlier compression. Default: 0.10. Set to 0 to
+     *  disable cost-aware capping (use the model's full context). */
+    targetCacheReadCostPerTurn: external_exports.number().min(0).default(0.1),
+    /** Direct override for the layer-0 token cap. When set, bypasses the
+     *  cost-aware formula from targetCacheReadCostPerTurn. 0 = disabled
+     *  (no cap, use full context). Default: undefined (use cost-aware auto). */
+    maxLayer0Tokens: external_exports.number().min(0).optional()
+  }).default({ distilled: 0.25, raw: 0.4, output: 0.25, ltm: 0.05, targetCacheReadCostPerTurn: 0.1 }),
+  /**
+   * Cold-cache idle-resume handling.
+   *
+   * Anthropic's prompt cache evicts entries after ~5 min (default tier) /
+   * ~1 hour (extended tier). When a session resumes after the eviction window,
+   * Lore's byte-identity caches (distilled prefix, raw window pin, LTM block)
+   * are providing no value because the underlying provider cache is already
+   * cold. On detection, Lore refreshes those caches so the next turn can
+   * produce a better-fitting window without paying a cache cost it would
+   * otherwise be trying to preserve. Reasoning blocks are NOT touched —
+   * Anthropic's April 23 postmortem identified dropping reasoning blocks as
+   * the root cause of forgetfulness/repetition.
+   *
+   * `idleResumeMinutes` is the threshold in minutes. Default 60 — matches
+   * Anthropic's extended-cache eviction window, conservative across providers.
+   * Set to 0 to disable the feature.
+   */
+  idleResumeMinutes: external_exports.number().min(0).max(24 * 60).default(60),
   distillation: external_exports.object({
-    minMessages: external_exports.number().min(3).default(8),
-    maxSegment: external_exports.number().min(5).default(50),
-    metaThreshold: external_exports.number().min(3).default(10)
-  }).default({ minMessages: 8, maxSegment: 50, metaThreshold: 10 }),
+    minMessages: external_exports.number().min(3).default(5),
+    maxSegment: external_exports.number().min(5).default(30),
+    metaThreshold: external_exports.number().min(3).default(10),
+    /** Max chars per tool output when rendering temporal messages for distillation input.
+     *  Outputs longer than this are replaced with a compact annotation preserving line
+     *  count, error signals, and file paths. Default: 2000 (matches upstream OpenCode's
+     *  TOOL_OUTPUT_MAX_CHARS during compaction). Set to 0 to disable. */
+    toolOutputMaxChars: external_exports.number().min(0).default(2e3)
+  }).default({
+    minMessages: 5,
+    maxSegment: 30,
+    metaThreshold: 10,
+    toolOutputMaxChars: 2e3
+  }),
   knowledge: external_exports.object({
     /** Set to false to disable long-term knowledge storage and system-prompt injection.
      *  Conversation recall (temporal search, distillation search) and context management
@@ -25599,6 +25770,7 @@ __export(embedding_exports, {
   vectorSearch: () => vectorSearch,
   vectorSearchDistillations: () => vectorSearchDistillations
 });
+var EMBED_TIMEOUT_MS = 1e4;
 var VOYAGE_API_URL = "https://api.voyageai.com/v1/embeddings";
 var VoyageProvider = class {
   maxBatchSize = 128;
@@ -25622,7 +25794,8 @@ var VoyageProvider = class {
         model: this.model,
         input_type: inputType,
         output_dimension: this.dimensions
-      })
+      }),
+      signal: AbortSignal.timeout(EMBED_TIMEOUT_MS)
     });
     if (!res.ok) {
       const body = await res.text().catch(() => "");
@@ -25658,7 +25831,8 @@ var OpenAIProvider = class {
         "Content-Type": "application/json",
         Authorization: `Bearer ${this.apiKey}`
       },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: AbortSignal.timeout(EMBED_TIMEOUT_MS)
     });
     if (!res.ok) {
       const responseBody = await res.text().catch(() => "");
@@ -26041,8 +26215,8 @@ function searchScored2(input) {
   const ftsSQL = `SELECT s.id, s.project_id, s.file, s.heading, s.depth, s.content,
          s.content_hash, s.first_paragraph, s.updated_at,
          bm25(lat_sections_fts, 6.0, 2.0) as rank
-       FROM lat_sections s
-       JOIN lat_sections_fts f ON s.rowid = f.rowid
+       FROM lat_sections_fts f
+       CROSS JOIN lat_sections s ON s.rowid = f.rowid
        WHERE lat_sections_fts MATCH ?
        AND s.project_id = ?
        ORDER BY rank LIMIT ?`;
@@ -26068,8 +26242,8 @@ function scoreForSession(projectPath, sessionContext, maxTokens) {
       `SELECT s.id, s.project_id, s.file, s.heading, s.depth, s.content,
                 s.content_hash, s.first_paragraph, s.updated_at,
                 bm25(lat_sections_fts, 6.0, 2.0) as rank
-         FROM lat_sections s
-         JOIN lat_sections_fts f ON s.rowid = f.rowid
+         FROM lat_sections_fts f
+         CROSS JOIN lat_sections s ON s.rowid = f.rowid
          WHERE lat_sections_fts MATCH ?
          AND s.project_id = ?
          ORDER BY rank`
@@ -26198,10 +26372,10 @@ function scoreEntriesFTS(sessionContext) {
   try {
     const results = db().query(
       `SELECT k.id, bm25(knowledge_fts, ?, ?, ?) as rank
-         FROM knowledge k
-         JOIN knowledge_fts f ON k.rowid = f.rowid
-         WHERE knowledge_fts MATCH ?
-         AND k.confidence > 0.2`
+          FROM knowledge_fts f
+          CROSS JOIN knowledge k ON k.rowid = f.rowid
+          WHERE knowledge_fts MATCH ?
+          AND k.confidence > 0.2`
     ).all(title, content3, category, q);
     if (!results.length) return /* @__PURE__ */ new Map();
     const ranks = results.map((r) => r.rank);
@@ -26335,13 +26509,13 @@ function search3(input) {
   const q = ftsQuery(input.query);
   if (q === EMPTY_QUERY) return [];
   const pid = input.projectPath ? ensureProject(input.projectPath) : null;
-  const ftsSQL = pid ? `SELECT ${KNOWLEDGE_COLS_K} FROM knowledge k
-       JOIN knowledge_fts f ON k.rowid = f.rowid
+  const ftsSQL = pid ? `SELECT ${KNOWLEDGE_COLS_K} FROM knowledge_fts f
+       CROSS JOIN knowledge k ON k.rowid = f.rowid
        WHERE knowledge_fts MATCH ?
        AND (k.project_id = ? OR k.project_id IS NULL OR k.cross_project = 1)
        AND k.confidence > 0.2
-       ORDER BY bm25(knowledge_fts, ?, ?, ?) LIMIT ?` : `SELECT ${KNOWLEDGE_COLS_K} FROM knowledge k
-       JOIN knowledge_fts f ON k.rowid = f.rowid
+       ORDER BY bm25(knowledge_fts, ?, ?, ?) LIMIT ?` : `SELECT ${KNOWLEDGE_COLS_K} FROM knowledge_fts f
+       CROSS JOIN knowledge k ON k.rowid = f.rowid
        WHERE knowledge_fts MATCH ?
        AND k.confidence > 0.2
        ORDER BY bm25(knowledge_fts, ?, ?, ?) LIMIT ?`;
@@ -26368,13 +26542,13 @@ function searchScored3(input) {
   if (q === EMPTY_QUERY) return [];
   const pid = input.projectPath ? ensureProject(input.projectPath) : null;
   const { title, content: content3, category } = ftsWeights();
-  const ftsSQL = pid ? `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge k
-       JOIN knowledge_fts f ON k.rowid = f.rowid
+  const ftsSQL = pid ? `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge_fts f
+       CROSS JOIN knowledge k ON k.rowid = f.rowid
        WHERE knowledge_fts MATCH ?
        AND (k.project_id = ? OR k.project_id IS NULL OR k.cross_project = 1)
        AND k.confidence > 0.2
-       ORDER BY rank LIMIT ?` : `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge k
-       JOIN knowledge_fts f ON k.rowid = f.rowid
+       ORDER BY rank LIMIT ?` : `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge_fts f
+       CROSS JOIN knowledge k ON k.rowid = f.rowid
        WHERE knowledge_fts MATCH ?
        AND k.confidence > 0.2
        ORDER BY rank LIMIT ?`;
@@ -26396,8 +26570,8 @@ function searchScoredOtherProjects(input) {
   if (q === EMPTY_QUERY) return [];
   const excludePid = ensureProject(input.excludeProjectPath);
   const { title, content: content3, category } = ftsWeights();
-  const ftsSQL = `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge k
-     JOIN knowledge_fts f ON k.rowid = f.rowid
+  const ftsSQL = `SELECT ${KNOWLEDGE_COLS_K}, bm25(knowledge_fts, ?, ?, ?) as rank FROM knowledge_fts f
+     CROSS JOIN knowledge k ON k.rowid = f.rowid
      WHERE knowledge_fts MATCH ?
      AND k.project_id IS NOT NULL
      AND k.project_id != ?
@@ -26530,8 +26704,8 @@ function check2(projectPath) {
     try {
       const { title, content: content3, category } = config2().search.ftsWeights;
       const matches = db().query(
-        `SELECT k.id, k.title FROM knowledge k
-           JOIN knowledge_fts f ON k.rowid = f.rowid
+        `SELECT k.id, k.title FROM knowledge_fts f
+           CROSS JOIN knowledge k ON k.rowid = f.rowid
            WHERE knowledge_fts MATCH ?
            AND k.id != ?
            AND k.confidence > 0.2
@@ -26565,9 +26739,13 @@ function check2(projectPath) {
 // src/distillation.ts
 var distillation_exports = {};
 __export(distillation_exports, {
+  latestMetaObservations: () => latestMetaObservations,
   loadForSession: () => loadForSession,
+  messagesToText: () => messagesToText,
+  metaDistill: () => metaDistill,
   parseSourceIds: () => parseSourceIds,
   run: () => run,
+  truncateToolOutputsInContent: () => truncateToolOutputsInContent,
   workerSessionIDs: () => workerSessionIDs
 });
 
@@ -26592,6 +26770,8 @@ function estimateMessage(msg) {
 }
 var contextLimit = 2e5;
 var outputReserved = 32e3;
+var maxLayer0Tokens = 0;
+var MIN_LAYER0_FLOOR = 4e4;
 var FIRST_TURN_OVERHEAD = 15e3;
 var calibratedOverhead = null;
 function makeSessionState() {
@@ -26605,7 +26785,11 @@ function makeSessionState() {
     forceMinLayer: 0,
     lastTransformEstimate: 0,
     prefixCache: null,
-    rawWindowCache: null
+    rawWindowCache: null,
+    lastTurnAt: 0,
+    cameOutOfIdle: false,
+    consecutiveHighLayer: 0,
+    lastPrefixHash: ""
   };
 }
 var sessionStates = /* @__PURE__ */ new Map();
@@ -26618,11 +26802,36 @@ function getSessionState(sessionID) {
   }
   return state;
 }
+function onIdleResume(sessionID, thresholdMs, now = Date.now()) {
+  if (thresholdMs <= 0) return { triggered: false };
+  const state = getSessionState(sessionID);
+  if (state.lastTurnAt === 0) return { triggered: false };
+  const idleMs = now - state.lastTurnAt;
+  if (idleMs < thresholdMs) return { triggered: false };
+  state.prefixCache = null;
+  state.rawWindowCache = null;
+  state.cameOutOfIdle = true;
+  return { triggered: true, idleMs };
+}
+function consumeCameOutOfIdle(sessionID) {
+  const state = sessionStates.get(sessionID);
+  if (!state || !state.cameOutOfIdle) return false;
+  state.cameOutOfIdle = false;
+  return true;
+}
 var ltmTokens = 0;
 function setModelLimits(limits) {
   contextLimit = limits.context || 2e5;
   outputReserved = Math.min(limits.output || 32e3, 32e3);
 }
+function setMaxLayer0Tokens(tokens) {
+  maxLayer0Tokens = Math.max(0, Math.floor(tokens));
+}
+function computeLayer0Cap(targetCostPerTurn, cacheReadCostPerToken) {
+  if (targetCostPerTurn <= 0 || cacheReadCostPerToken <= 0) return 0;
+  const rawCap = Math.floor(targetCostPerTurn / cacheReadCostPerToken);
+  return Math.max(rawCap, MIN_LAYER0_FLOOR);
+}
 function setLtmTokens(tokens) {
   ltmTokens = tokens;
 }
@@ -26667,6 +26876,19 @@ function setForceMinLayer(layer, sessionID) {
     }
   }
 }
+function inspectSessionState(sessionID) {
+  const state = sessionStates.get(sessionID);
+  if (!state) return null;
+  return {
+    hasPrefixCache: state.prefixCache !== null,
+    hasRawWindowCache: state.rawWindowCache !== null,
+    cameOutOfIdle: state.cameOutOfIdle,
+    lastTurnAt: state.lastTurnAt
+  };
+}
+function setLastTurnAtForTest(sessionID, ms) {
+  getSessionState(sessionID).lastTurnAt = ms;
+}
 function loadDistillations(projectPath, sessionID) {
   const pid = ensureProject(projectPath);
   const query = sessionID ? "SELECT id, observations, generation, token_count, created_at, session_id FROM distillations WHERE project_id = ? AND session_id = ? AND archived = 0 ORDER BY created_at ASC" : "SELECT id, observations, generation, token_count, created_at, session_id FROM distillations WHERE project_id = ? AND archived = 0 ORDER BY created_at ASC";
@@ -26699,12 +26921,17 @@ function cleanParts(parts) {
   }
   return filtered.length > 0 ? filtered : parts;
 }
+var ANNOTATION_PATH_SCAN_LIMIT = 64 * 1024;
+var PATH_RE = /(?:[\w.-]+\/)+[\w.-]+\.\w{1,5}/g;
 function toolStripAnnotation(toolName, output) {
   const lines = output.split("\n").length;
-  const chars = output.length;
   const hasError = /\b(?:error|fail(?:ed|ure)?|exception|panic|traceback)\b/i.test(output);
-  const paths = output.match(/(?:[\w.-]+\/)+[\w.-]+\.\w{1,5}/g);
-  const uniquePaths = paths ? [...new Set(paths)].slice(0, 5) : [];
+  let uniquePaths = [];
+  if (output.indexOf("/") !== -1) {
+    const pathScan = output.length > ANNOTATION_PATH_SCAN_LIMIT ? output.slice(0, ANNOTATION_PATH_SCAN_LIMIT) : output;
+    const paths = pathScan.match(PATH_RE);
+    if (paths) uniquePaths = [...new Set(paths)].slice(0, 5);
+  }
   let annotation = `[output omitted \u2014 ${toolName}: ${lines} lines`;
   if (hasError) annotation += ", contained errors";
   if (uniquePaths.length > 0) annotation += `, paths: ${uniquePaths.join(", ")}`;
@@ -27135,7 +27362,11 @@ function transformInner(input) {
     expectedInput = messageTokens + overhead + ltmTokens;
   }
   const layer0Input = calibrated ? expectedInput : expectedInput * UNCALIBRATED_SAFETY;
-  if (effectiveMinLayer === 0 && layer0Input <= maxInput) {
+  let layer0Ceiling = maxLayer0Tokens > 0 ? Math.min(maxInput, maxLayer0Tokens) : maxInput;
+  if (!calibrated && layer0Ceiling < maxInput) {
+    layer0Ceiling = Math.floor(layer0Ceiling * 0.7);
+  }
+  if (effectiveMinLayer === 0 && layer0Input <= layer0Ceiling) {
     const messageTokens = calibrated ? expectedInput - (ltmTokens - sessState.lastKnownLtm) : expectedInput - overhead - ltmTokens;
     return {
       messages: input.messages,
@@ -27216,14 +27447,31 @@ function transformInner(input) {
     (sum, m) => sum + estimateMessage(m),
     0
   );
-  const nuclearRaw = input.messages.slice(-3).map((m) => ({
+  const tailBudget = Math.max(2e3, Math.min(8e3, Math.floor(usable * 0.25)));
+  const nuclearTurnStart = currentTurnStart(input.messages);
+  const currentTurn = input.messages.slice(nuclearTurnStart).map((m) => ({
     info: m.info,
     parts: cleanParts(m.parts)
   }));
-  const nuclearRawTokens = nuclearRaw.reduce(
+  const currentTurnTokens = currentTurn.reduce(
     (sum, m) => sum + estimateMessage(m),
     0
   );
+  const olderMessages = [];
+  let olderTokens = 0;
+  const remaining = Math.max(0, tailBudget - currentTurnTokens);
+  for (let i = nuclearTurnStart - 1; i >= 0 && olderTokens < remaining; i--) {
+    const msg = input.messages[i];
+    const est = estimateMessage(msg);
+    if (olderTokens + est > remaining) break;
+    olderMessages.unshift({
+      info: msg.info,
+      parts: cleanParts(msg.parts)
+    });
+    olderTokens += est;
+  }
+  const nuclearRaw = [...olderMessages, ...currentTurn];
+  const nuclearRawTokens = olderTokens + currentTurnTokens;
   return {
     messages: [...nuclearPrefix, ...nuclearRaw],
     layer: 4,
@@ -27245,6 +27493,28 @@ function transform2(input) {
     state.lastTransformEstimate = result.totalTokens;
     state.lastLayer = result.layer;
     state.lastWindowMessageIDs = new Set(result.messages.map((m) => m.info.id));
+    state.lastTurnAt = Date.now();
+    const prefixIds = result.messages.slice(0, 5).map((m) => m.info.id).join(",");
+    const prefixHash = `${result.layer}:${prefixIds}`;
+    if (state.lastPrefixHash && state.lastPrefixHash !== prefixHash) {
+      info(
+        `cache-bust detected: session=${sid} layer=${state.lastLayer}\u2192${result.layer} msgs=${state.lastTransformedCount}\u2192${result.messages.length} prefix=${state.lastPrefixHash.slice(0, 30)}\u2192${prefixHash.slice(0, 30)}`
+      );
+    }
+    state.lastPrefixHash = prefixHash;
+    if (result.layer >= 2) {
+      state.consecutiveHighLayer++;
+      if (state.consecutiveHighLayer === 3) {
+        info(
+          `session ${sid} has been at gradient layer ${result.layer}+ for 3 consecutive turns. Consider running /compact to reset the context window.`
+        );
+      }
+    } else {
+      state.consecutiveHighLayer = 0;
+    }
+    info(
+      `gradient: session=${sid} layer=${result.layer} tokens=${result.totalTokens} (distilled=${result.distilledTokens} raw=${result.rawTokens}) usable=${result.usable} cap=${maxLayer0Tokens || "off"}`
+    );
   }
   return result;
 }
@@ -27338,8 +27608,39 @@ function formatTime(ms) {
   const m = d.getMinutes().toString().padStart(2, "0");
   return `${h3}:${m}`;
 }
-function messagesToText(messages) {
-  return messages.map((m) => `[${m.role}] (${formatTime(m.created_at)}) ${m.content}`).join("\n\n");
+var CHUNK_SEPARATOR = "\n" + CHUNK_TERMINATOR;
+function truncateToolOutputsInContent(content3, maxChars) {
+  if (maxChars <= 0 || content3.length === 0) return content3;
+  if (content3.indexOf(CHUNK_TERMINATOR) === -1) {
+    return truncateSingleChunk(content3, maxChars);
+  }
+  const chunks = content3.split(CHUNK_SEPARATOR);
+  let anyToolChunk = false;
+  for (const c of chunks) {
+    if (c.startsWith("[tool:")) {
+      anyToolChunk = true;
+      break;
+    }
+  }
+  if (!anyToolChunk) return content3;
+  const out = chunks.map((chunk) => truncateSingleChunk(chunk, maxChars));
+  return out.join(CHUNK_SEPARATOR);
+}
+function truncateSingleChunk(chunk, maxChars) {
+  if (!chunk.startsWith("[tool:")) return chunk;
+  const closeBracket = chunk.indexOf("] ");
+  if (closeBracket < 0) return chunk;
+  const toolName = chunk.slice(6, closeBracket);
+  const payload = chunk.slice(closeBracket + 2);
+  if (payload.length <= maxChars) return chunk;
+  return `[tool:${toolName}] ${toolStripAnnotation(toolName, payload)}`;
+}
+function messagesToText(messages, toolOutputMaxChars) {
+  const cap = toolOutputMaxChars ?? config2().distillation.toolOutputMaxChars;
+  return messages.map((m) => {
+    const body = m.role === "user" ? m.content : truncateToolOutputsInContent(m.content, cap);
+    return `[${m.role}] (${formatTime(m.created_at)}) ${body}`;
+  }).join("\n\n");
 }
 function parseDistillationResult(text4) {
   const match = text4.match(/<observations>([\s\S]*?)<\/observations>/i);
@@ -27354,6 +27655,19 @@ function latestObservations(projectPath, sessionID) {
   ).get(pid, sessionID);
   return row?.observations || void 0;
 }
+function latestMetaObservations(projectPath, sessionID) {
+  return latestMeta(projectPath, sessionID)?.observations;
+}
+function latestMeta(projectPath, sessionID) {
+  const pid = ensureProject(projectPath);
+  const row = db().query(
+    `SELECT observations, generation FROM distillations
+       WHERE project_id = ? AND session_id = ? AND generation > 0
+       ORDER BY generation DESC, created_at DESC LIMIT 1`
+  ).get(pid, sessionID);
+  if (!row || !row.observations) return void 0;
+  return row;
+}
 function parseSourceIds(raw) {
   try {
     const parsed = JSON.parse(raw);
@@ -27363,11 +27677,10 @@ function parseSourceIds(raw) {
     return [];
   }
 }
-function loadForSession(projectPath, sessionID) {
+function loadForSession(projectPath, sessionID, includeArchived = false) {
   const pid = ensureProject(projectPath);
-  const rows = db().query(
-    "SELECT id, project_id, session_id, observations, source_ids, generation, token_count, created_at FROM distillations WHERE project_id = ? AND session_id = ? ORDER BY created_at ASC"
-  ).all(pid, sessionID);
+  const sql = includeArchived ? "SELECT id, project_id, session_id, observations, source_ids, generation, token_count, created_at FROM distillations WHERE project_id = ? AND session_id = ? ORDER BY created_at ASC" : "SELECT id, project_id, session_id, observations, source_ids, generation, token_count, created_at FROM distillations WHERE project_id = ? AND session_id = ? AND archived = 0 ORDER BY created_at ASC";
+  const rows = db().query(sql).all(pid, sessionID);
   return rows.map((r) => ({
     ...r,
     source_ids: parseSourceIds(r.source_ids)
@@ -27532,8 +27845,13 @@ async function distillSegment(input) {
 }
 async function metaDistill(input) {
   const existing = loadGen0(input.projectPath, input.sessionID);
-  if (existing.length < 3) return null;
-  const userContent = recursiveUser(existing);
+  const priorMeta = latestMeta(input.projectPath, input.sessionID);
+  if (priorMeta) {
+    if (existing.length === 0) return null;
+  } else {
+    if (existing.length < 3) return null;
+  }
+  const userContent = recursiveUser(existing, priorMeta?.observations);
   const model = input.model ?? config2().model;
   const responseText = await input.llm.prompt(
     RECURSIVE_SYSTEM,
@@ -27543,19 +27861,30 @@ async function metaDistill(input) {
   if (!responseText) return null;
   const result = parseDistillationResult(responseText);
   if (!result) return null;
-  const maxGen = Math.max(...existing.map((d) => d.generation));
+  const maxGen = Math.max(
+    ...existing.map((d) => d.generation),
+    priorMeta?.generation ?? 0
+  );
   const allSourceIDs = existing.flatMap((d) => d.source_ids);
-  const metaId = storeDistillation({
-    projectPath: input.projectPath,
-    sessionID: input.sessionID,
-    observations: result.observations,
-    sourceIDs: allSourceIDs,
-    generation: maxGen + 1
-  });
+  let metaId;
+  db().exec("BEGIN IMMEDIATE");
+  try {
+    metaId = storeDistillation({
+      projectPath: input.projectPath,
+      sessionID: input.sessionID,
+      observations: result.observations,
+      sourceIDs: allSourceIDs,
+      generation: maxGen + 1
+    });
+    archiveDistillations(existing.map((d) => d.id));
+    db().exec("COMMIT");
+  } catch (e) {
+    db().exec("ROLLBACK");
+    throw e;
+  }
   if (isAvailable()) {
     embedDistillation(metaId, result.observations);
   }
-  archiveDistillations(existing.map((d) => d.id));
   return result;
 }
 
@@ -27714,13 +28043,13 @@ function searchDistillationsScored(input) {
   const q = ftsQuery(input.query);
   if (q === EMPTY_QUERY) return [];
   const ftsSQL = input.sessionID ? `SELECT d.id, d.observations, d.generation, d.created_at, d.session_id, rank
-       FROM distillations d
-       JOIN distillation_fts f ON d.rowid = f.rowid
+       FROM distillation_fts f
+       CROSS JOIN distillations d ON d.rowid = f.rowid
        WHERE distillation_fts MATCH ?
        AND d.project_id = ? AND d.session_id = ?
        ORDER BY rank LIMIT ?` : `SELECT d.id, d.observations, d.generation, d.created_at, d.session_id, rank
-       FROM distillations d
-       JOIN distillation_fts f ON d.rowid = f.rowid
+       FROM distillation_fts f
+       CROSS JOIN distillations d ON d.rowid = f.rowid
        WHERE distillation_fts MATCH ?
        AND d.project_id = ?
        ORDER BY rank LIMIT ?`;
@@ -28163,7 +28492,130 @@ function importFromFile(input) {
     }
   }
 }
+
+// src/worker-model.ts
+var worker_model_exports = {};
+__export(worker_model_exports, {
+  WORKER_JUDGE_SYSTEM: () => WORKER_JUDGE_SYSTEM,
+  computeModelFingerprint: () => computeModelFingerprint,
+  getValidatedWorkerModel: () => getValidatedWorkerModel,
+  isValidationStale: () => isValidationStale,
+  parseJudgeScore: () => parseJudgeScore,
+  resolveWorkerModel: () => resolveWorkerModel,
+  selectWorkerCandidates: () => selectWorkerCandidates,
+  storeValidatedWorkerModel: () => storeValidatedWorkerModel,
+  structuralCheck: () => structuralCheck,
+  workerJudgeUser: () => workerJudgeUser
+});
+var KV_PREFIX = "lore:worker_model:";
+function selectWorkerCandidates(sessionModel, providerModels) {
+  const eligible = providerModels.filter(
+    (m) => m.providerID === sessionModel.providerID && m.status === "active" && m.capabilities.input.text
+  );
+  if (eligible.length === 0) return [];
+  const sorted = [...eligible].sort((a, b) => a.cost.input - b.cost.input);
+  const cheapest = sorted[0];
+  const belowSession = sorted.filter((m) => m.cost.input < sessionModel.cost.input).pop();
+  const candidates = /* @__PURE__ */ new Map();
+  candidates.set(cheapest.id, cheapest);
+  if (belowSession && belowSession.id !== cheapest.id) {
+    candidates.set(belowSession.id, belowSession);
+  }
+  if (cheapest.id === sessionModel.id || cheapest.cost.input >= sessionModel.cost.input) {
+    return [cheapest];
+  }
+  return [...candidates.values()];
+}
+function computeModelFingerprint(providerID, sessionModelID, activeModelIDs) {
+  const sorted = [...activeModelIDs].sort();
+  return sha256(
+    JSON.stringify({ providerID, sessionModelID, modelIDs: sorted })
+  );
+}
+function getValidatedWorkerModel(providerID) {
+  const row = db().query("SELECT value FROM kv_meta WHERE key = ?").get(`${KV_PREFIX}${providerID}`);
+  if (!row) return null;
+  try {
+    return JSON.parse(row.value);
+  } catch {
+    return null;
+  }
+}
+function storeValidatedWorkerModel(result) {
+  const key = `${KV_PREFIX}${result.providerID}`;
+  const value = JSON.stringify(result);
+  db().query(
+    "INSERT INTO kv_meta (key, value) VALUES (?, ?) ON CONFLICT(key) DO UPDATE SET value = ?"
+  ).run(key, value, value);
+}
+function isValidationStale(stored, currentFingerprint) {
+  if (!stored) return true;
+  return stored.fingerprint !== currentFingerprint;
+}
+function structuralCheck(candidateObservations, referenceObservations) {
+  if (candidateObservations == null || candidateObservations.length === 0) {
+    return { passed: false, observationCount: 0, tokenCount: 0, reason: candidateObservations === null ? "parse_failed" : "empty" };
+  }
+  const countObs = (text4) => text4.split("\n").filter((l) => l.trim().length > 0).length;
+  const refCount = countObs(referenceObservations);
+  const candCount = countObs(candidateObservations);
+  const candTokens = Math.ceil(candidateObservations.length / 3);
+  if (refCount > 0 && (candCount < refCount * 0.5 || candCount > refCount * 1.5)) {
+    return {
+      passed: false,
+      observationCount: candCount,
+      tokenCount: candTokens,
+      reason: `observation_count_${candCount}_vs_ref_${refCount}`
+    };
+  }
+  const refTokens = Math.ceil(referenceObservations.length / 3);
+  if (candTokens === 0) {
+    return { passed: false, observationCount: candCount, tokenCount: candTokens, reason: "empty" };
+  }
+  if (refTokens > 0 && candTokens > refTokens * 3) {
+    return {
+      passed: false,
+      observationCount: candCount,
+      tokenCount: candTokens,
+      reason: `token_count_${candTokens}_vs_ref_${refTokens}_3x`
+    };
+  }
+  return { passed: true, observationCount: candCount, tokenCount: candTokens };
+}
+var WORKER_JUDGE_SYSTEM = `You are evaluating distillation quality. You will be given a REFERENCE distillation (produced by a capable model) and a CANDIDATE distillation (produced by a cheaper model) of the same conversation segment.
+
+Rate the candidate on a scale of 1-5:
+5 = Captures all key facts and decisions, equivalent to reference
+4 = Captures most facts, minor omissions
+3 = Captures the essential facts, some detail loss acceptable
+2 = Missing important facts or technical details
+1 = Significantly incomplete or inaccurate
+
+Respond with ONLY a single digit (1-5).`;
+function workerJudgeUser(reference, candidate) {
+  return `<reference>
+${reference}
+</reference>
+
+<candidate>
+${candidate}
+</candidate>`;
+}
+function parseJudgeScore(response) {
+  const match = response.trim().match(/^([1-5])/);
+  if (!match) return null;
+  return parseInt(match[1], 10);
+}
+function resolveWorkerModel(providerID, configWorkerModel, configModel) {
+  if (configWorkerModel) return configWorkerModel;
+  const validated = getValidatedWorkerModel(providerID);
+  if (validated) {
+    return { providerID: validated.providerID, modelID: validated.modelID };
+  }
+  return configModel;
+}
 export {
+  COMPACT_SUMMARY_TEMPLATE,
   CONSOLIDATION_SYSTEM,
   CURATOR_SYSTEM,
   DISTILLATION_SYSTEM,
@@ -28172,10 +28624,14 @@ export {
   RECALL_PARAM_DESCRIPTIONS,
   RECALL_TOOL_DESCRIPTION,
   RECURSIVE_SYSTEM,
+  WORKER_JUDGE_SYSTEM,
+  buildCompactPrompt,
   calibrate,
   close,
+  computeLayer0Cap,
   config2 as config,
   consolidationUser,
+  consumeCameOutOfIdle,
   curator_exports as curator,
   curatorUser,
   db,
@@ -28197,6 +28653,7 @@ export {
   h,
   importFromFile,
   inline,
+  inspectSessionState,
   isFirstRun,
   isReasoningPart,
   isTextPart,
@@ -28211,6 +28668,7 @@ export {
   ltm_exports as ltm,
   needsUrgentDistillation,
   normalize,
+  onIdleResume,
   p,
   projectId,
   projectName,
@@ -28222,15 +28680,20 @@ export {
   saveForceMinLayer,
   serialize,
   setForceMinLayer,
+  setLastTurnAtForTest,
   setLtmTokens,
+  setMaxLayer0Tokens,
   setModelLimits,
   shouldImport,
   strong2 as strong,
   t,
   temporal_exports as temporal,
+  toolStripAnnotation,
   transform2 as transform,
   ul,
   unescapeMarkdown,
+  workerJudgeUser,
+  worker_model_exports as workerModel,
   workerSessionIDs
 };
 //# sourceMappingURL=index.js.map