@cortexkit/opencode-magic-context 0.9.1 → 0.10.1

package/dist/index.js

@@ -14070,7 +14070,7 @@ var init_agent_overrides = __esm(() => {
 });
 
 // src/config/schema/magic-context.ts
-var DEFAULT_NUDGE_INTERVAL_TOKENS = 1e4, DEFAULT_EXECUTE_THRESHOLD_PERCENTAGE = 65, DEFAULT_COMPARTMENT_TOKEN_BUDGET = 20000, DEFAULT_HISTORIAN_TIMEOUT_MS = 300000, DEFAULT_HISTORY_BUDGET_PERCENTAGE = 0.15, DEFAULT_LOCAL_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2", DREAMER_TASKS, DreamingTaskSchema, DEFAULT_DREAMER_TASKS, DreamerConfigSchema, SidekickConfigSchema, BaseEmbeddingConfigSchema, EmbeddingConfigSchema, MagicContextConfigSchema;
+var DEFAULT_NUDGE_INTERVAL_TOKENS = 1e4, DEFAULT_EXECUTE_THRESHOLD_PERCENTAGE = 65, DEFAULT_HISTORIAN_TIMEOUT_MS = 300000, DEFAULT_HISTORY_BUDGET_PERCENTAGE = 0.15, DEFAULT_LOCAL_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2", DREAMER_TASKS, DreamingTaskSchema, DEFAULT_DREAMER_TASKS, DreamerConfigSchema, SidekickConfigSchema, BaseEmbeddingConfigSchema, EmbeddingConfigSchema, MagicContextConfigSchema;
 var init_magic_context = __esm(() => {
   init_zod();
   init_agent_overrides();
@@ -14156,7 +14156,6 @@ var init_magic_context = __esm(() => {
     drop_tool_structure: exports_external.boolean().default(true),
     clear_reasoning_age: exports_external.number().min(10).default(50),
     iteration_nudge_threshold: exports_external.number().min(5).default(15),
-    compartment_token_budget: exports_external.number().min(1e4).default(DEFAULT_COMPARTMENT_TOKEN_BUDGET),
     history_budget_percentage: exports_external.number().min(0.05).max(0.5).default(DEFAULT_HISTORY_BUDGET_PERCENTAGE),
     historian_timeout_ms: exports_external.number().min(60000).default(DEFAULT_HISTORIAN_TIMEOUT_MS),
     commit_cluster_trigger: exports_external.object({
@@ -16166,7 +16165,57 @@ function cosineSimilarity(a, b) {
 
 // src/features/magic-context/memory/embedding-local.ts
 import { mkdirSync } from "fs";
+import { open, stat, unlink, writeFile } from "fs/promises";
 import { join as join9 } from "path";
+async function acquireModelLoadLock(lockPath) {
+  const waitStart = Date.now();
+  while (true) {
+    try {
+      const handle = await open(lockPath, "wx");
+      try {
+        await handle.writeFile(`pid=${process.pid} started=${Date.now()}
+`);
+      } catch {}
+      await handle.close();
+      return async () => {
+        try {
+          await unlink(lockPath);
+        } catch {}
+      };
+    } catch (error48) {
+      const code = error48.code;
+      if (code !== "EEXIST" && code !== "EPERM") {
+        throw error48;
+      }
+      try {
+        const info = await stat(lockPath);
+        if (Date.now() - info.mtimeMs > STALE_LOCK_MS) {
+          log(`[magic-context] embedding-load lock stale (>${STALE_LOCK_MS}ms), taking over`);
+          try {
+            await unlink(lockPath);
+          } catch {}
+          continue;
+        }
+      } catch {
+        continue;
+      }
+      if (Date.now() - waitStart > MAX_LOCK_WAIT_MS) {
+        log("[magic-context] embedding-load lock wait exceeded, proceeding without lock");
+        return async () => {};
+      }
+      await new Promise((resolve2) => setTimeout(resolve2, LOCK_POLL_MS));
+    }
+  }
+}
+function startLockHeartbeat(lockPath) {
+  const HEARTBEAT_MS = Math.floor(STALE_LOCK_MS / 3);
+  const timer = setInterval(() => {
+    writeFile(lockPath, `pid=${process.pid} alive=${Date.now()}
+`).catch(() => {});
+  }, HEARTBEAT_MS);
+  timer.unref?.();
+  return () => clearInterval(timer);
+}
 async function withQuietConsole(fn) {
   const origWarn = console.warn;
   const origError = console.error;
@@ -16259,30 +16308,37 @@ class LocalEmbeddingProvider {
           log("[magic-context] could not create model cache dir, using library default");
         }
         const createPipeline = transformersModule.pipeline;
-        const MAX_ATTEMPTS = 3;
-        let lastError;
-        for (let attempt = 1;attempt <= MAX_ATTEMPTS; attempt++) {
-          try {
-            this.pipeline = await withQuietConsole(() => createPipeline("feature-extraction", this.model, {
-              quantized: true,
-              dtype: "fp32"
-            }));
-            lastError = undefined;
-            break;
-          } catch (error48) {
-            lastError = error48;
-            if (!isTransientLoadError(error48) || attempt === MAX_ATTEMPTS) {
+        const lockPath = join9(modelCacheDir, ".load.lock");
+        const releaseLock = await acquireModelLoadLock(lockPath);
+        const stopHeartbeat = startLockHeartbeat(lockPath);
+        try {
+          const MAX_ATTEMPTS = 3;
+          let lastError;
+          for (let attempt = 1;attempt <= MAX_ATTEMPTS; attempt++) {
+            try {
+              this.pipeline = await withQuietConsole(() => createPipeline("feature-extraction", this.model, {
+                dtype: "fp32"
+              }));
+              lastError = undefined;
               break;
+            } catch (error48) {
+              lastError = error48;
+              if (!isTransientLoadError(error48) || attempt === MAX_ATTEMPTS) {
+                break;
+              }
+              const delayMs = 300 * attempt + Math.floor(Math.random() * 200);
+              log(`[magic-context] embedding model load attempt ${attempt}/${MAX_ATTEMPTS} failed transiently, retrying in ${delayMs}ms`);
+              await new Promise((resolve2) => setTimeout(resolve2, delayMs));
             }
-            const delayMs = 300 * attempt + Math.floor(Math.random() * 200);
-            log(`[magic-context] embedding model load attempt ${attempt}/${MAX_ATTEMPTS} failed transiently, retrying in ${delayMs}ms`);
-            await new Promise((resolve2) => setTimeout(resolve2, delayMs));
           }
-        }
-        if (this.pipeline) {
-          log(`[magic-context] embedding model loaded: ${this.model}`);
-        } else {
-          throw lastError ?? new Error("unknown embedding load failure");
+          if (this.pipeline) {
+            log(`[magic-context] embedding model loaded: ${this.model}`);
+          } else {
+            throw lastError ?? new Error("unknown embedding load failure");
+          }
+        } finally {
+          stopHeartbeat();
+          await releaseLock();
         }
       } catch (error48) {
         log("[magic-context] embedding model failed to load:", error48);
@@ -16357,10 +16413,13 @@ class LocalEmbeddingProvider {
     return this.pipeline !== null;
   }
 }
+var LOCK_POLL_MS = 150, STALE_LOCK_MS, MAX_LOCK_WAIT_MS;
 var init_embedding_local = __esm(() => {
   init_magic_context();
   init_data_path();
   init_logger();
+  STALE_LOCK_MS = 3 * 60000;
+  MAX_LOCK_WAIT_MS = 5 * 60000;
 });
 
 // src/features/magic-context/memory/embedding-openai.ts
@@ -18374,6 +18433,145 @@ var init_storage = __esm(() => {
   init_storage_tags();
 });
 
+// src/shared/models-dev-cache.ts
+import { createHash } from "crypto";
+import { existsSync as existsSync5, readFileSync as readFileSync4 } from "fs";
+import { homedir as homedir5, platform as platform2 } from "os";
+import { join as join11 } from "path";
+function hashFast(input) {
+  return createHash("sha1").update(input).digest("hex");
+}
+function getModelsJsonPath() {
+  const explicit = process.env.OPENCODE_MODELS_PATH?.trim();
+  if (explicit)
+    return explicit;
+  const xdgCache = process.env.XDG_CACHE_HOME;
+  const os3 = platform2();
+  let cacheBase;
+  if (xdgCache) {
+    cacheBase = xdgCache;
+  } else if (os3 === "win32") {
+    cacheBase = process.env.LOCALAPPDATA ?? join11(homedir5(), "AppData", "Local");
+  } else {
+    cacheBase = join11(homedir5(), ".cache");
+  }
+  const source = process.env.OPENCODE_MODELS_URL?.trim();
+  const filename = source && source !== "https://models.dev" ? `models-${hashFast(source)}.json` : "models.json";
+  return join11(cacheBase, "opencode", filename);
+}
+function getOpencodeConfigPath() {
+  const envDir = process.env.OPENCODE_CONFIG_DIR?.trim();
+  const configDir = envDir ? envDir : platform2() === "win32" ? join11(homedir5(), ".config", "opencode") : join11(process.env.XDG_CONFIG_HOME || join11(homedir5(), ".config"), "opencode");
+  const jsonc = join11(configDir, "opencode.jsonc");
+  if (existsSync5(jsonc))
+    return jsonc;
+  const json2 = join11(configDir, "opencode.json");
+  if (existsSync5(json2))
+    return json2;
+  return null;
+}
+function loadModelsDevLimitsFromFile() {
+  const limits = new Map;
+  const modelsJsonPath = getModelsJsonPath();
+  let fileFound = false;
+  try {
+    if (existsSync5(modelsJsonPath)) {
+      fileFound = true;
+      const raw = readFileSync4(modelsJsonPath, "utf-8");
+      const data = JSON.parse(raw);
+      for (const [providerId, provider2] of Object.entries(data)) {
+        if (!provider2?.models || typeof provider2.models !== "object")
+          continue;
+        for (const [modelId, model] of Object.entries(provider2.models)) {
+          const context = model?.limit?.context;
+          if (typeof context === "number" && context > 0) {
+            limits.set(`${providerId}/${modelId}`, context);
+            const modes = model?.experimental?.modes;
+            if (modes && typeof modes === "object") {
+              for (const mode of Object.keys(modes)) {
+                limits.set(`${providerId}/${modelId}-${mode}`, context);
+              }
+            }
+          }
+        }
+      }
+    }
+  } catch (error48) {
+    sessionLog("global", `models-dev-cache: failed to read models.json at ${modelsJsonPath}:`, error48 instanceof Error ? error48.message : String(error48));
+  }
+  try {
+    const configPath = getOpencodeConfigPath();
+    if (configPath && existsSync5(configPath)) {
+      let raw = readFileSync4(configPath, "utf-8");
+      raw = raw.replace(/"(?:[^"\\]|\\.)*"|\/\/.*$/gm, (match) => match.startsWith('"') ? match : "");
+      const config2 = JSON.parse(raw);
+      if (config2.provider && typeof config2.provider === "object") {
+        for (const [providerId, provider2] of Object.entries(config2.provider)) {
+          if (!provider2?.models || typeof provider2.models !== "object")
+            continue;
+          for (const [modelId, model] of Object.entries(provider2.models)) {
+            const context = model?.limit?.context;
+            if (typeof context === "number" && context > 0) {
+              limits.set(`${providerId}/${modelId}`, context);
+            }
+          }
+        }
+      }
+    }
+  } catch (error48) {
+    sessionLog("global", "models-dev-cache: failed to read opencode config for custom models:", error48 instanceof Error ? error48.message : String(error48));
+  }
+  sessionLog("global", `models-dev-cache: file-layer loaded ${limits.size} model limits (modelsJsonPath=${modelsJsonPath}, found=${fileFound})`);
+  return limits;
+}
+async function refreshModelLimitsFromApi(client) {
+  try {
+    const result = await client.config.providers();
+    const data = result.data;
+    const providers = data?.providers;
+    if (!Array.isArray(providers)) {
+      sessionLog("global", "models-dev-cache: API refresh returned no providers payload");
+      return;
+    }
+    const map2 = new Map;
+    for (const entry of providers) {
+      const p = entry;
+      if (!p?.id || !p.models || typeof p.models !== "object")
+        continue;
+      for (const [modelId, model] of Object.entries(p.models)) {
+        const context = model?.limit?.context;
+        if (typeof context === "number" && context > 0) {
+          map2.set(`${p.id}/${modelId}`, context);
+        }
+      }
+    }
+    apiCache = map2;
+    apiLoadedAt = Date.now();
+    sessionLog("global", `models-dev-cache: API layer loaded ${map2.size} model limits`);
+  } catch (error48) {
+    sessionLog("global", "models-dev-cache: API refresh failed:", error48 instanceof Error ? error48.message : String(error48));
+  }
+}
+function getModelsDevContextLimit(providerID, modelID) {
+  const key = `${providerID}/${modelID}`;
+  if (apiCache) {
+    const fromApi = apiCache.get(key);
+    if (typeof fromApi === "number")
+      return fromApi;
+  }
+  const now = Date.now();
+  if (!fileCache || now - fileLastAttempt > RELOAD_INTERVAL_MS) {
+    fileLastAttempt = now;
+    fileCache = loadModelsDevLimitsFromFile();
+  }
+  return fileCache.get(key);
+}
+var RELOAD_INTERVAL_MS, apiCache = null, apiLoadedAt = 0, fileCache = null, fileLastAttempt = 0;
+var init_models_dev_cache = __esm(() => {
+  init_logger();
+  RELOAD_INTERVAL_MS = 5 * 60 * 1000;
+});
+
 // src/features/magic-context/memory/project-identity.ts
 import { execSync } from "child_process";
 import path3 from "path";
@@ -18566,6 +18764,67 @@ var init_send_session_notification = __esm(() => {
   init_logger();
 });
 
+// src/hooks/magic-context/derive-budgets.ts
+var exports_derive_budgets = {};
+__export(exports_derive_budgets, {
+  resolveHistorianContextLimit: () => resolveHistorianContextLimit,
+  deriveTriggerBudget: () => deriveTriggerBudget,
+  deriveHistorianChunkTokens: () => deriveHistorianChunkTokens
+});
+function deriveTriggerBudget(mainContextLimit, executeThresholdPercentage) {
+  if (!Number.isFinite(mainContextLimit) || mainContextLimit <= 0) {
+    return TRIGGER_BUDGET_MIN;
+  }
+  const thresholdFraction = Math.max(0, executeThresholdPercentage) / 100;
+  const usable = mainContextLimit * thresholdFraction;
+  const derived = Math.round(usable * TRIGGER_BUDGET_PERCENTAGE);
+  return Math.max(TRIGGER_BUDGET_MIN, Math.min(TRIGGER_BUDGET_MAX, derived));
+}
+function deriveHistorianChunkTokens(historianContextLimit) {
+  if (!Number.isFinite(historianContextLimit) || historianContextLimit <= 0) {
+    return HISTORIAN_CHUNK_MIN;
+  }
+  const derived = Math.round(historianContextLimit * HISTORIAN_CHUNK_PERCENTAGE);
+  return Math.max(HISTORIAN_CHUNK_MIN, Math.min(HISTORIAN_CHUNK_MAX, derived));
+}
+function resolveHistorianContextLimit(historianModelOverride) {
+  if (typeof historianModelOverride === "string" && historianModelOverride.includes("/")) {
+    const [providerID, ...rest] = historianModelOverride.split("/");
+    const modelID = rest.join("/");
+    if (providerID && modelID) {
+      const limit = getModelsDevContextLimit(providerID, modelID);
+      if (typeof limit === "number" && limit > 0)
+        return limit;
+    }
+    return DEFAULT_HISTORIAN_CONTEXT_FALLBACK;
+  }
+  if (typeof historianModelOverride === "string" && historianModelOverride.trim() !== "") {
+    console.warn(`[magic-context] historian.model "${historianModelOverride}" lacks provider prefix ("provider/model-id"); using fallback chain for chunk-budget derivation.`);
+  }
+  const chain = AGENT_MODEL_REQUIREMENTS[HISTORIAN_AGENT]?.fallbackChain;
+  if (!chain || chain.length === 0)
+    return DEFAULT_HISTORIAN_CONTEXT_FALLBACK;
+  const expanded = expandFallbackChain(chain);
+  let minLimit;
+  for (const key of expanded) {
+    const [providerID, ...rest] = key.split("/");
+    const modelID = rest.join("/");
+    if (!providerID || !modelID)
+      continue;
+    const limit = getModelsDevContextLimit(providerID, modelID);
+    if (typeof limit !== "number" || limit <= 0)
+      continue;
+    if (minLimit === undefined || limit < minLimit)
+      minLimit = limit;
+  }
+  return minLimit ?? DEFAULT_HISTORIAN_CONTEXT_FALLBACK;
+}
+var TRIGGER_BUDGET_PERCENTAGE = 0.05, TRIGGER_BUDGET_MIN = 5000, TRIGGER_BUDGET_MAX = 50000, HISTORIAN_CHUNK_PERCENTAGE = 0.25, HISTORIAN_CHUNK_MIN = 8000, HISTORIAN_CHUNK_MAX = 50000, DEFAULT_HISTORIAN_CONTEXT_FALLBACK = 128000;
+var init_derive_budgets = __esm(() => {
+  init_model_requirements();
+  init_models_dev_cache();
+});
+
 // src/features/magic-context/compaction-marker.ts
 import { Database as Database4 } from "bun:sqlite";
 import { join as join12 } from "path";
@@ -19980,7 +20239,7 @@ async function runCompartmentAgent(deps) {
     client,
     db,
     sessionId,
-    tokenBudget,
+    historianChunkTokens,
     directory,
     historianTimeoutMs,
     getNotificationParams
@@ -20015,7 +20274,7 @@ No new compartments or facts were written. Rebuild or clear the broken compartme
     if (protectedTailStart <= offset) {
       return;
     }
-    const chunk = readSessionChunk(sessionId, tokenBudget, offset, protectedTailStart);
+    const chunk = readSessionChunk(sessionId, historianChunkTokens, offset, protectedTailStart);
     if (!chunk.text || chunk.messageCount === 0) {
       return;
     }
@@ -20158,7 +20417,7 @@ async function executeContextRecompInternal(deps) {
     client,
     db,
     sessionId,
-    tokenBudget,
+    historianChunkTokens,
     directory,
     historianTimeoutMs,
     getNotificationParams
@@ -20212,7 +20471,7 @@ No eligible raw history exists before the protected tail, so nothing was rebuilt
     let candidateFacts = existingStaging?.facts ?? [];
     let offset = existingStaging ? existingStaging.lastEndMessage + 1 : 1;
     let passCount = existingStaging?.passCount ?? 0;
-    let currentTokenBudget = tokenBudget;
+    let currentTokenBudget = historianChunkTokens;
     let passAttempt = 1;
     const resumed = existingStaging !== null;
     if (resumed) {
@@ -20310,7 +20569,7 @@ Nothing was written.`;
       ];
       candidateFacts = validatedPass.facts ?? [];
       passCount += 1;
-      currentTokenBudget = tokenBudget;
+      currentTokenBudget = historianChunkTokens;
       passAttempt = 1;
       saveRecompStagingPass(db, sessionId, passCount, candidateCompartments, candidateFacts);
       const nextOffset = (validatedPass.compartments?.[validatedPass.compartments.length - 1]?.endMessage ?? chunk.endIndex) + 1;
@@ -28510,7 +28769,7 @@ Check verifiable memories against actual repository state. Update stale wording,
 4. **Be conservative.** If you cannot find the referenced code but it might be in a location you haven't checked, do NOT archive. Move on.
 
 ### Verification examples
-- Memory: "compartment_token_budget defaults to 20000" \u2192 grep schema for \`compartment_token_budget\`, check \`.default(...)\`
+- Memory: "history_budget_percentage defaults to 0.15" \u2192 grep schema for \`history_budget_percentage\`, check \`.default(...)\`
 - Memory: "Durable state lives in ~/.local/share/opencode/storage/plugin/magic-context/context.db" \u2192 check storage-db.ts for the path construction
 - Memory: "ctx_search searches memories, facts, and history" \u2192 grep for ctx_search tool definition and unified search implementation
 
@@ -30239,110 +30498,13 @@ function createCompactionHandler() {
     }
   };
 }
-// src/shared/models-dev-cache.ts
-init_logger();
-import { existsSync as existsSync5, readFileSync as readFileSync4 } from "fs";
-import { homedir as homedir5, platform as platform2 } from "os";
-import { join as join11 } from "path";
-var cachedLimits = null;
-var lastLoadAttempt = 0;
-var RELOAD_INTERVAL_MS = 5 * 60 * 1000;
-function getModelsJsonPath() {
-  const xdgCache = process.env.XDG_CACHE_HOME;
-  const os3 = platform2();
-  let cacheBase;
-  if (xdgCache) {
-    cacheBase = xdgCache;
-  } else if (os3 === "win32") {
-    cacheBase = process.env.LOCALAPPDATA ?? join11(homedir5(), "AppData", "Local");
-  } else {
-    cacheBase = join11(homedir5(), ".cache");
-  }
-  return join11(cacheBase, "opencode", "models.json");
-}
-function getOpencodeConfigPath() {
-  const envDir = process.env.OPENCODE_CONFIG_DIR?.trim();
-  const configDir = envDir ? envDir : platform2() === "win32" ? join11(homedir5(), ".config", "opencode") : join11(process.env.XDG_CONFIG_HOME || join11(homedir5(), ".config"), "opencode");
-  const jsonc = join11(configDir, "opencode.jsonc");
-  if (existsSync5(jsonc))
-    return jsonc;
-  const json2 = join11(configDir, "opencode.json");
-  if (existsSync5(json2))
-    return json2;
-  return null;
-}
-function loadModelsDevLimits() {
-  const limits = new Map;
-  const modelsJsonPath = getModelsJsonPath();
-  try {
-    if (existsSync5(modelsJsonPath)) {
-      const raw = readFileSync4(modelsJsonPath, "utf-8");
-      const data = JSON.parse(raw);
-      for (const [providerId, provider2] of Object.entries(data)) {
-        if (!provider2?.models || typeof provider2.models !== "object")
-          continue;
-        for (const [modelId, model] of Object.entries(provider2.models)) {
-          const context = model?.limit?.context;
-          if (typeof context === "number" && context > 0) {
-            limits.set(`${providerId}/${modelId}`, context);
-            const modes = model?.experimental?.modes;
-            if (modes && typeof modes === "object") {
-              for (const mode of Object.keys(modes)) {
-                limits.set(`${providerId}/${modelId}-${mode}`, context);
-              }
-            }
-          }
-        }
-      }
-    }
-  } catch (error48) {
-    sessionLog("global", "models-dev-cache: failed to read models.json:", error48 instanceof Error ? error48.message : String(error48));
-  }
-  try {
-    const configPath = getOpencodeConfigPath();
-    if (configPath && existsSync5(configPath)) {
-      let raw = readFileSync4(configPath, "utf-8");
-      raw = raw.replace(/"(?:[^"\\]|\\.)*"|\/\/.*$/gm, (match) => match.startsWith('"') ? match : "");
-      const config2 = JSON.parse(raw);
-      if (config2.provider && typeof config2.provider === "object") {
-        for (const [providerId, provider2] of Object.entries(config2.provider)) {
-          if (!provider2?.models || typeof provider2.models !== "object")
-            continue;
-          for (const [modelId, model] of Object.entries(provider2.models)) {
-            const context = model?.limit?.context;
-            if (typeof context === "number" && context > 0) {
-              limits.set(`${providerId}/${modelId}`, context);
-            }
-          }
-        }
-      }
-    }
-  } catch (error48) {
-    sessionLog("global", "models-dev-cache: failed to read opencode config for custom models:", error48 instanceof Error ? error48.message : String(error48));
-  }
-  return limits;
-}
-function getModelsDevContextLimit(providerID, modelID) {
-  const now = Date.now();
-  if (!cachedLimits || now - lastLoadAttempt > RELOAD_INTERVAL_MS) {
-    lastLoadAttempt = now;
-    cachedLimits = loadModelsDevLimits();
-  }
-  return cachedLimits.get(`${providerID}/${modelID}`);
-}
-
 // src/hooks/magic-context/event-resolvers.ts
+init_models_dev_cache();
 var DEFAULT_CONTEXT_LIMIT = 128000;
-function resolveContextLimit(providerID, modelID, config2) {
+function resolveContextLimit(providerID, modelID) {
   if (!providerID) {
     return DEFAULT_CONTEXT_LIMIT;
   }
-  if (modelID) {
-    const modelSpecific = config2.modelContextLimitsCache?.get(`${providerID}/${modelID}`);
-    if (typeof modelSpecific === "number" && modelSpecific > 0) {
-      return modelSpecific;
-    }
-  }
   if (modelID) {
     const modelsDevLimit = getModelsDevContextLimit(providerID, modelID);
     if (modelsDevLimit !== undefined) {
@@ -30625,7 +30787,6 @@ init_storage_tags();
 init_logger();
 
 // src/hooks/magic-context/compartment-trigger.ts
-init_magic_context();
 init_compartment_storage();
 init_storage();
 init_logger();
@@ -30703,7 +30864,7 @@ var TAIL_INFO_DEFAULTS = {
   tokenEstimate: 0,
   commitClusterCount: 0
 };
-function getUnsummarizedTailInfo(db, sessionId, compartmentTokenBudget) {
+function getUnsummarizedTailInfo(db, sessionId, triggerBudget) {
   return withRawSessionMessageCache(() => {
     try {
       const lastCompartmentEnd = getLastCompartmentEndMessage(db, sessionId);
@@ -30714,7 +30875,7 @@ function getUnsummarizedTailInfo(db, sessionId, compartmentTokenBudget) {
       if (!hasEligibleHistory) {
         return { ...TAIL_INFO_DEFAULTS, nextStartOrdinal };
       }
-      const scanBudget = Math.max(MIN_PROACTIVE_TAIL_TOKEN_ESTIMATE, compartmentTokenBudget * TAIL_SIZE_TRIGGER_MULTIPLIER);
+      const scanBudget = Math.max(MIN_PROACTIVE_TAIL_TOKEN_ESTIMATE, triggerBudget * TAIL_SIZE_TRIGGER_MULTIPLIER);
       const chunk = readSessionChunk(sessionId, scanBudget, nextStartOrdinal, protectedTailStart);
       const isMeaningful = chunk.hasMore || chunk.tokenEstimate >= MIN_PROACTIVE_TAIL_TOKEN_ESTIMATE || chunk.messageCount >= MIN_PROACTIVE_TAIL_MESSAGE_COUNT;
       return {
@@ -30730,11 +30891,11 @@ function getUnsummarizedTailInfo(db, sessionId, compartmentTokenBudget) {
     }
   });
 }
-function checkCompartmentTrigger(db, sessionId, sessionMeta, usage, _previousPercentage, executeThresholdPercentage, compartmentTokenBudget = DEFAULT_COMPARTMENT_TOKEN_BUDGET, autoDropToolAge, protectedTagCount, clearReasoningAge, dropToolStructure = true, commitClusterTrigger) {
+function checkCompartmentTrigger(db, sessionId, sessionMeta, usage, _previousPercentage, executeThresholdPercentage, triggerBudget, autoDropToolAge, protectedTagCount, clearReasoningAge, dropToolStructure = true, commitClusterTrigger) {
   if (sessionMeta.compartmentInProgress) {
     return { shouldFire: false };
   }
-  const tailInfo = getUnsummarizedTailInfo(db, sessionId, compartmentTokenBudget);
+  const tailInfo = getUnsummarizedTailInfo(db, sessionId, triggerBudget);
   if (!tailInfo.hasNewRawHistory) {
     return { shouldFire: false };
   }
@@ -30750,12 +30911,12 @@ function checkCompartmentTrigger(db, sessionId, sessionMeta, usage, _previousPer
   }
   const clusterEnabled = commitClusterTrigger?.enabled ?? true;
   const minClusters = commitClusterTrigger?.min_clusters ?? DEFAULT_MIN_COMMIT_CLUSTERS_FOR_TRIGGER;
-  if (clusterEnabled && tailInfo.commitClusterCount >= minClusters && tailInfo.tokenEstimate >= compartmentTokenBudget) {
+  if (clusterEnabled && tailInfo.commitClusterCount >= minClusters && tailInfo.tokenEstimate >= triggerBudget) {
     sessionLog(sessionId, `compartment trigger: commit-cluster fire \u2014 ${tailInfo.commitClusterCount} clusters (min=${minClusters}), ~${tailInfo.tokenEstimate} tokens in eligible prefix`);
     return { shouldFire: true, reason: "commit_clusters" };
   }
-  if (tailInfo.tokenEstimate >= compartmentTokenBudget * TAIL_SIZE_TRIGGER_MULTIPLIER) {
-    sessionLog(sessionId, `compartment trigger: tail-size fire \u2014 ~${tailInfo.tokenEstimate} tokens exceeds ${compartmentTokenBudget * TAIL_SIZE_TRIGGER_MULTIPLIER} budget threshold`);
+  if (tailInfo.tokenEstimate >= triggerBudget * TAIL_SIZE_TRIGGER_MULTIPLIER) {
+    sessionLog(sessionId, `compartment trigger: tail-size fire \u2014 ~${tailInfo.tokenEstimate} tokens exceeds ${triggerBudget * TAIL_SIZE_TRIGGER_MULTIPLIER} budget threshold`);
     return { shouldFire: true, reason: "tail_size" };
   }
   const proactiveTriggerPercentage = getProactiveCompartmentTriggerPercentage(executeThresholdPercentage);
@@ -31099,11 +31260,15 @@ Historian recomp started. Rebuilding compartments and facts from raw session his
   };
 }
 
+// src/hooks/magic-context/hook.ts
+init_derive_budgets();
+
 // src/hooks/magic-context/event-handler.ts
 init_storage();
 init_storage_meta_persisted();
 init_logger();
 init_compaction_marker_manager();
+init_derive_budgets();
 
 // src/hooks/magic-context/event-payloads.ts
 function getSessionProperties(properties) {
@@ -31175,7 +31340,6 @@ function getMessageRemovedInfo(properties) {
 init_note_nudger();
 
 // src/hooks/magic-context/transform-compartment-phase.ts
-init_magic_context();
 init_compartment_storage();
 init_storage();
 init_logger();
@@ -31251,7 +31415,7 @@ async function runCompartmentPhase(args) {
         client: args.client,
         db: args.db,
         sessionId: args.sessionId,
-        tokenBudget: args.compartmentTokenBudget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+        historianChunkTokens: args.historianChunkTokens,
         historyBudgetTokens: args.historyBudgetTokens,
         historianTimeoutMs: args.historianTimeoutMs,
         directory: args.compartmentDirectory,
@@ -31272,7 +31436,7 @@ async function runCompartmentPhase(args) {
         client: args.client,
         db: args.db,
         sessionId: args.sessionId,
-        tokenBudget: args.compartmentTokenBudget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+        historianChunkTokens: args.historianChunkTokens,
         historyBudgetTokens: args.historyBudgetTokens,
         historianTimeoutMs: args.historianTimeoutMs,
         directory: args.compartmentDirectory,
@@ -31425,9 +31589,7 @@ function createEventHandler2(deps) {
         }
         if (hasUsageTokens) {
           const totalInputTokens = (info.tokens?.input ?? 0) + (info.tokens?.cache?.read ?? 0) + (info.tokens?.cache?.write ?? 0);
-          const contextLimit = resolveContextLimit(info.providerID, info.modelID, {
-            modelContextLimitsCache: deps.config.modelContextLimitsCache
-          });
+          const contextLimit = resolveContextLimit(info.providerID, info.modelID);
           const percentage = contextLimit > 0 ? totalInputTokens / contextLimit * 100 : 0;
           sessionLog(info.sessionID, `event message.updated: totalInputTokens=${totalInputTokens} contextLimit=${contextLimit} percentage=${percentage.toFixed(1)}%`);
           deps.contextUsageMap.set(info.sessionID, {
@@ -31447,7 +31609,9 @@ function createEventHandler2(deps) {
           const sessionMeta = getOrCreateSessionMeta(deps.db, info.sessionID);
           const previousPercentage = sessionMeta.lastContextPercentage;
           if (!sessionMeta.isSubagent) {
-            const triggerResult = checkCompartmentTrigger(deps.db, info.sessionID, sessionMeta, { percentage, inputTokens: totalInputTokens }, previousPercentage, resolveExecuteThreshold(deps.config.execute_threshold_percentage ?? 65, modelKey, 65), undefined, deps.config.auto_drop_tool_age ?? 100, deps.config.protected_tags, deps.config.clear_reasoning_age ?? 50, deps.config.drop_tool_structure ?? true, deps.config.commit_cluster_trigger);
+            const effectiveExecuteThreshold = resolveExecuteThreshold(deps.config.execute_threshold_percentage ?? 65, modelKey, 65);
+            const triggerBudget = deriveTriggerBudget(contextLimit, effectiveExecuteThreshold);
+            const triggerResult = checkCompartmentTrigger(deps.db, info.sessionID, sessionMeta, { percentage, inputTokens: totalInputTokens }, previousPercentage, effectiveExecuteThreshold, triggerBudget, deps.config.auto_drop_tool_age ?? 100, deps.config.protected_tags, deps.config.clear_reasoning_age ?? 50, deps.config.drop_tool_structure ?? true, deps.config.commit_cluster_trigger);
             if (triggerResult.shouldFire) {
               sessionLog(info.sessionID, `compartment trigger: firing (reason=${triggerResult.reason})`);
               updateSessionMeta(deps.db, info.sessionID, {
@@ -31725,7 +31889,6 @@ function createTextCompleteHandler() {
 }
 
 // src/hooks/magic-context/transform.ts
-init_magic_context();
 init_compartment_storage();
 init_project_identity();
 init_storage();
@@ -32806,6 +32969,14 @@ function isToolProtocolPart(part) {
 function hasToolProtocolParts(message) {
   return message.parts.some(isToolProtocolPart);
 }
+function hasThinkingBearingParts(message) {
+  return message.parts.some((part) => {
+    if (part === null || typeof part !== "object")
+      return false;
+    const p = part;
+    return p.type === "thinking" || p.type === "reasoning" || p.type === "redacted_thinking";
+  });
+}
 function isMessageDropped(message) {
   const textParts = message.parts.filter(isTextPart);
   if (textParts.length === 0)
@@ -32840,6 +33011,11 @@ function reinjectNudgeAtAnchor(messages, nudgeText, nudgePlacements, sessionId)
       nudgePlacements.clear(sessionId);
       return false;
     }
+    if (hasThinkingBearingParts(message)) {
+      sessionLog(sessionId, `nudge anchor abandoned: message ${message.info.id} now contains thinking/reasoning parts (signed, immutable)`);
+      nudgePlacements.clear(sessionId);
+      return false;
+    }
     for (let j = message.parts.length - 1;j >= 0; j--) {
       const part = message.parts[j];
       if (isTextPart(part)) {
@@ -32868,6 +33044,8 @@ function appendNudgeToAssistant(messages, nudge, nudgePlacements, sessionId) {
       continue;
     if (isMessageDropped(message))
       continue;
+    if (hasThinkingBearingParts(message))
+      continue;
     for (let j = message.parts.length - 1;j >= 0; j--) {
       const part = message.parts[j];
       if (isTextPart(part)) {
@@ -33151,12 +33329,13 @@ function runPostTransformPhase(args) {
   const forceMaterialization = args.fullFeatureMode && args.contextUsage.percentage >= args.forceMaterializationPercentage;
   const activeCompartmentRun = args.canRunCompartments ? getActiveCompartmentRun(args.sessionId) : undefined;
   const compartmentRunning = args.canRunCompartments && !args.awaitedCompartmentRun && activeCompartmentRun !== undefined;
-  const shouldReadPendingOps = isExplicitFlush || args.schedulerDecision === "execute" || compartmentRunning;
+  const emergencyBypassCompartmentGate = forceMaterialization;
+  const shouldReadPendingOps = isExplicitFlush || args.schedulerDecision === "execute" || forceMaterialization || compartmentRunning;
   const pendingOps = shouldReadPendingOps ? getPendingOps(args.db, args.sessionId) : [];
   const hasPendingUserOps = pendingOps.length > 0;
-  const shouldApplyPendingOps = (args.schedulerDecision === "execute" || isExplicitFlush) && !compartmentRunning;
+  const shouldApplyPendingOps = (args.schedulerDecision === "execute" || isExplicitFlush || forceMaterialization) && (!compartmentRunning || emergencyBypassCompartmentGate);
   const isCacheBustingPass = isExplicitFlush || shouldApplyPendingOps;
-  const shouldRunHeuristics = args.fullFeatureMode && !compartmentRunning && (isExplicitFlush || forceMaterialization || args.schedulerDecision === "execute" && !alreadyRanThisTurn);
+  const shouldRunHeuristics = args.fullFeatureMode && (!compartmentRunning || emergencyBypassCompartmentGate) && (isExplicitFlush || forceMaterialization || args.schedulerDecision === "execute" && !alreadyRanThisTurn);
   if (shouldRunHeuristics) {
     const reason = isExplicitFlush ? "explicit_flush" : forceMaterialization ? `force_materialization (${args.contextUsage.percentage.toFixed(1)}% >= ${args.forceMaterializationPercentage}%)` : `scheduler_execute (pendingOps=${pendingOps.length}, scheduler=${args.schedulerDecision})`;
     sessionLog(args.sessionId, `heuristics WILL RUN \u2014 reason=${reason}, context=${args.contextUsage.percentage.toFixed(1)}%, turn=${args.currentTurnId}`);
@@ -33165,7 +33344,11 @@ function runPostTransformPhase(args) {
     sessionLog(args.sessionId, `transform: skipping heuristics (already ran for turn ${args.currentTurnId})`);
   }
   if (compartmentRunning && hasPendingUserOps) {
-    sessionLog(args.sessionId, "transform: deferring pending ops \u2014 compartment agent in progress");
+    if (emergencyBypassCompartmentGate) {
+      sessionLog(args.sessionId, `transform: emergency bypass \u2014 applying ${pendingOps.length} pending ops while compartment agent runs (${args.contextUsage.percentage.toFixed(1)}%)`);
+    } else {
+      sessionLog(args.sessionId, "transform: deferring pending ops \u2014 compartment agent in progress");
+    }
   }
   try {
     if (shouldApplyPendingOps) {
@@ -33524,7 +33707,7 @@ function createTransform(deps) {
         client: deps.client,
         db,
         sessionId,
-        tokenBudget: deps.compartmentTokenBudget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+        historianChunkTokens: deps.getHistorianChunkTokens?.() ?? 20000,
         historyBudgetTokens,
         historianTimeoutMs: deps.historianTimeoutMs,
         directory: compartmentDirectory,
@@ -33645,7 +33828,7 @@ Historian previously failed ${historianFailureState.failureCount} time(s), so ma
       db,
       sessionId,
       resolvedSessionId,
-      compartmentTokenBudget: deps.compartmentTokenBudget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+      historianChunkTokens: deps.getHistorianChunkTokens?.() ?? 20000,
       historyBudgetTokens,
       historianTimeoutMs: deps.historianTimeoutMs,
       compartmentDirectory,
@@ -34352,6 +34535,7 @@ function createMagicContextHook(deps) {
   const projectPath = resolveProjectIdentity(deps.directory);
   registerDreamProjectDirectory(projectPath, deps.directory);
   let lastScheduleCheckMs = 0;
+  const getHistorianChunkTokens = () => deriveHistorianChunkTokens(resolveHistorianContextLimit(deps.config.historian?.model));
   const nudgePlacements = createNudgePlacementStore(db);
   const flushedSessions = new Set;
   const lastHeuristicsTurnId = new Map;
@@ -34389,7 +34573,7 @@ function createMagicContextHook(deps) {
       enabled: deps.config.memory.enabled,
       injectionBudgetTokens: deps.config.memory.injection_budget_tokens
     } : undefined,
-    compartmentTokenBudget: deps.config.compartment_token_budget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+    getHistorianChunkTokens,
     historyBudgetPercentage: deps.config.history_budget_percentage,
     executeThresholdPercentage: deps.config.execute_threshold_percentage,
     historianTimeoutMs: deps.config.historian_timeout_ms ?? DEFAULT_HISTORIAN_TIMEOUT_MS,
@@ -34470,7 +34654,7 @@ function createMagicContextHook(deps) {
       client: deps.client,
       db,
       sessionId,
-      tokenBudget: deps.config.compartment_token_budget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET,
+      historianChunkTokens: getHistorianChunkTokens(),
       historianTimeoutMs: deps.config.historian_timeout_ms ?? DEFAULT_HISTORIAN_TIMEOUT_MS,
       directory: deps.directory,
       fallbackModelId: (() => {
@@ -34596,7 +34780,7 @@ function createSessionHooks(args) {
         clear_reasoning_age: pluginConfig.clear_reasoning_age,
         iteration_nudge_threshold: pluginConfig.iteration_nudge_threshold,
         execute_threshold_percentage: pluginConfig.execute_threshold_percentage ?? DEFAULT_EXECUTE_THRESHOLD_PERCENTAGE,
-        compartment_token_budget: pluginConfig.compartment_token_budget,
+        historian: pluginConfig.historian,
         history_budget_percentage: pluginConfig.history_budget_percentage,
         historian_timeout_ms: pluginConfig.historian_timeout_ms,
         memory: pluginConfig.memory,
@@ -34921,17 +35105,18 @@ function registerRpcHandlers(rpcServer, args) {
       return { ok: false, error: "no session" };
     const { executeContextRecomp: executeContextRecomp2 } = await Promise.resolve().then(() => (init_compartment_runner(), exports_compartment_runner));
     const { sendIgnoredMessage: sendIgnoredMessage2 } = await Promise.resolve().then(() => (init_send_session_notification(), exports_send_session_notification));
+    const { deriveHistorianChunkTokens: deriveHistorianChunkTokens2, resolveHistorianContextLimit: resolveHistorianContextLimit2 } = await Promise.resolve().then(() => (init_derive_budgets(), exports_derive_budgets));
     const db = getDb();
     if (!db)
       return { ok: false, error: "db unavailable" };
-    const DEFAULT_COMPARTMENT_TOKEN_BUDGET2 = 20000;
     const DEFAULT_HISTORIAN_TIMEOUT_MS2 = 600000;
+    const historianChunkTokens = deriveHistorianChunkTokens2(resolveHistorianContextLimit2(config2.historian?.model));
     log(`[rpc] recomp requested for session ${sessionId}`);
     executeContextRecomp2({
       client: args.client,
       db,
       sessionId,
-      tokenBudget: config2.compartment_token_budget ?? DEFAULT_COMPARTMENT_TOKEN_BUDGET2,
+      historianChunkTokens,
       historianTimeoutMs: config2.historian_timeout_ms ?? DEFAULT_HISTORIAN_TIMEOUT_MS2,
       directory,
       getNotificationParams: () => getNotificationParams(sessionId)
@@ -36136,6 +36321,7 @@ init_conflict_detector();
 init_data_path();
 init_logger();
 init_model_requirements();
+init_models_dev_cache();
 
 // src/shared/rpc-server.ts
 init_logger();
@@ -36144,11 +36330,11 @@ import { createServer } from "http";
 import { dirname } from "path";
 
 // src/shared/rpc-utils.ts
-import { createHash } from "crypto";
+import { createHash as createHash2 } from "crypto";
 import { join as join15 } from "path";
 function projectHash(directory) {
   const normalized = directory.replace(/\/+$/, "");
-  return createHash("sha256").update(normalized).digest("hex").slice(0, 16);
+  return createHash2("sha256").update(normalized).digest("hex").slice(0, 16);
 }
 function rpcPortFilePath(storageDir, directory) {
   return join15(storageDir, "rpc", projectHash(directory), "port");
@@ -36333,6 +36519,10 @@ var plugin = async (ctx) => {
     rpcServer.start().catch((err) => {
       log(`[magic-context] RPC server failed to start: ${err}`);
     });
+    refreshModelLimitsFromApi(ctx.client);
+    setInterval(() => {
+      refreshModelLimitsFromApi(ctx.client);
+    }, 5 * 60 * 1000);
   }
   if (conflictResult?.hasConflict) {
     sendConflictWarning(ctx.client, ctx.directory, conflictResult);