@iowarp/clio-coder 0.3.3 → 0.3.6

package/dist/{chunk-LZSJBIVT.js → chunk-DJVECN66.js}

@@ -1,10 +1,14 @@
 import { createRequire as __clioCreateRequire } from "node:module"; const require = __clioCreateRequire(import.meta.url);
+import {
+  createSentinelStripper,
+  stripTokenizerSentinels
+} from "./chunk-CFGTUFWB.js";
 import {
   readSettings
-} from "./chunk-A3CYT5EX.js";
+} from "./chunk-AD2SYQYC.js";
 import {
   THINKING_LEVELS
-} from "./chunk-ZDOOVTXZ.js";
+} from "./chunk-4OC57DA6.js";
 import {
   authNotRequiredStatus,
   openAuthStorage,
@@ -13,10 +17,6 @@ import {
   resolveRuntimeAuthTarget,
   targetRequiresAuth
 } from "./chunk-4VP4KH3K.js";
-import {
-  cwdHash,
-  sessionPaths
-} from "./chunk-M6SHUN7Q.js";
 import {
   sleep
 } from "./chunk-FQ4SKYE4.js";
@@ -24,8 +24,10 @@ import {
   withStateFileLock
 } from "./chunk-IKCO5N3L.js";
 import {
-  require_dist
-} from "./chunk-APJ265NV.js";
+  ceilChars,
+  estimateAgentMessageTokens,
+  toolSchemaChars
+} from "./chunk-GEYXPTRF.js";
 import {
   activateExternalPluginApiBridge,
   calculateEngineCost,
@@ -34,15 +36,20 @@ import {
   getEngineSupportedThinkingLevels,
   registerEngineApiProvider
 } from "./chunk-FCSXB6T2.js";
+import {
+  require_dist
+} from "./chunk-APJ265NV.js";
 import {
   clioConfigDir,
   clioStatePath,
   resolveClioDirs
 } from "./chunk-BNAZZHFG.js";
 import {
-  BusChannels,
   getSharedBus
-} from "./chunk-OQ33BKR3.js";
+} from "./chunk-ZGVHUX3M.js";
+import {
+  BusChannels
+} from "./chunk-G7MUEIGA.js";
 import {
   __toESM,
   init_esm_shims
@@ -363,597 +370,15 @@ function credentialsPresent() {
   return present;
 }
 
-// src/domains/session/context-accounting.ts
-init_esm_shims();
-import { appendFileSync, existsSync, readFileSync } from "node:fs";
-import { dirname, join as join2 } from "node:path";
-
-// src/domains/session/compaction/auto.ts
-init_esm_shims();
-var DEFAULT_COMPACTION_THRESHOLD = 0.8;
-function shouldCompact(contextTokens, threshold, contextWindow) {
-  const base = {
-    shouldCompact: false,
-    pressure: null,
-    contextTokens,
-    contextWindow,
-    threshold
-  };
-  if (!Number.isFinite(threshold) || threshold <= 0 || threshold > 1) return base;
-  if (!Number.isFinite(contextTokens) || contextTokens <= 0) return base;
-  if (!Number.isFinite(contextWindow) || contextWindow <= 0) return base;
-  const pressure = contextTokens / contextWindow;
-  return {
-    shouldCompact: pressure >= threshold,
-    pressure,
-    contextTokens,
-    contextWindow,
-    threshold
-  };
-}
-var AutoCompactionTrigger = class {
-  inFlight = null;
-  /** True while a task is in flight. Callers can short-circuit on this. */
-  isBusy() {
-    return this.inFlight !== null;
-  }
-  /**
-   * Run `task` unless one is already in flight. Returns the Promise the
-   * task is running under. Callers always get the eventual result of the
-   * first task, never a second call. Rejections clear the slot so a later
-   * fire can start fresh.
-   */
-  fire(task) {
-    if (this.inFlight) return this.inFlight;
-    const run = task().finally(() => {
-      this.inFlight = null;
-    });
-    this.inFlight = run;
-    return run;
-  }
-};
-
-// src/domains/session/context-ledger.ts
-init_esm_shims();
-var SEGMENT_CATEGORY = {
-  identity: "system",
-  "operating-contract": "system",
-  delegation: "agents",
-  skills: "skills",
-  safety: "system",
-  runtime: "system",
-  "tool-contract": "system",
-  "retrieval-hints": "system",
-  "tool-catalog": "tools",
-  "tools-and-agents": "agents",
-  "agent-fleet-deltas": "agents",
-  "skills-catalog": "skills",
-  memory: "memory",
-  "project-context": "project",
-  "context.project-rules": "project",
-  "context.operator-profile": "memory",
-  "context.clio-repo-awareness": "system",
-  "history-summary": "messages"
-};
-var CONTEXT_CATEGORY_LABEL = {
-  system: "System prompt",
-  tools: "Tool definitions",
-  agents: "Agent fleet",
-  skills: "Skills",
-  memory: "Memory",
-  project: "Project context",
-  messages: "Messages",
-  pending: "Pending input",
-  reserve: "Autocompact reserve",
-  free: "Free space",
-  streaming: "Streaming output"
-};
-var CONTENT_ORDER = [
-  "system",
-  "tools",
-  "agents",
-  "skills",
-  "memory",
-  "project",
-  "messages",
-  "pending",
-  "streaming"
-];
-function categoryForSegment(id) {
-  return SEGMENT_CATEGORY[id] ?? "system";
-}
-function finiteNonNegative(value) {
-  return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
-}
-function percentOf(tokens, window) {
-  if (window <= 0) return null;
-  return Math.min(100, tokens / window * 100);
-}
-function buildContextLedger(input) {
-  const window = finiteNonNegative(input.contextWindow);
-  const raw = {
-    system: 0,
-    tools: 0,
-    agents: 0,
-    skills: 0,
-    memory: 0,
-    project: 0,
-    messages: 0,
-    pending: 0,
-    reserve: 0,
-    free: 0,
-    streaming: 0
-  };
-  const segments = input.promptSegments ?? [];
-  if (segments.length > 0) {
-    for (const segment of segments) {
-      raw[categoryForSegment(segment.id)] += finiteNonNegative(segment.tokenEstimate);
-    }
-  } else {
-    raw.system += finiteNonNegative(input.systemPromptTokens);
-  }
-  raw.tools += finiteNonNegative(input.toolSchemaTokens);
-  raw.messages += finiteNonNegative(input.messageTokens);
-  raw.pending += finiteNonNegative(input.pendingTokens);
-  raw.streaming += finiteNonNegative(input.streamingTokens);
-  raw.agents += finiteNonNegative(input.agentsTokens);
-  raw.skills += finiteNonNegative(input.skillsTokens);
-  raw.memory += finiteNonNegative(input.memoryTokens);
-  raw.project += finiteNonNegative(input.projectTokens);
-  const decomposed = CONTENT_ORDER.reduce((sum, category) => sum + raw[category], 0);
-  const liveTotal = finiteNonNegative(input.liveTotalTokens);
-  if (liveTotal > decomposed) {
-    raw.messages += liveTotal - decomposed;
-  }
-  const usedTokens = Math.max(decomposed, liveTotal);
-  const effectiveThreshold = typeof input.compactionThreshold === "number" && input.compactionThreshold > 0 && input.compactionThreshold < 1 ? input.compactionThreshold : input.compactionAuto ? DEFAULT_COMPACTION_THRESHOLD : null;
-  let reserveTokens = 0;
-  let freeTokens = 0;
-  if (window > 0) {
-    const remaining = Math.max(0, window - usedTokens);
-    reserveTokens = effectiveThreshold !== null ? Math.min(Math.round(window * (1 - effectiveThreshold)), remaining) : 0;
-    freeTokens = Math.max(0, window - usedTokens - reserveTokens);
-  }
-  const groups = [];
-  for (const category of CONTENT_ORDER) {
-    const tokens = raw[category];
-    if (tokens <= 0) continue;
-    groups.push({
-      category,
-      label: CONTEXT_CATEGORY_LABEL[category],
-      tokens,
-      percent: percentOf(tokens, window)
-    });
-  }
-  const meter = [...groups];
-  if (reserveTokens > 0) {
-    meter.push({
-      category: "reserve",
-      label: CONTEXT_CATEGORY_LABEL.reserve,
-      tokens: reserveTokens,
-      percent: percentOf(reserveTokens, window)
-    });
-  }
-  if (window > 0) {
-    meter.push({
-      category: "free",
-      label: CONTEXT_CATEGORY_LABEL.free,
-      tokens: freeTokens,
-      percent: percentOf(freeTokens, window)
-    });
-  }
-  return {
-    provider: input.provider ?? null,
-    model: input.model ?? null,
-    contextWindow: window,
-    contextWindowSource: input.contextWindowSource ?? null,
-    usedTokens,
-    reserveTokens,
-    freeTokens,
-    percent: percentOf(usedTokens, window),
-    measured: input.measured === true && liveTotal > 0,
-    compactionThreshold: effectiveThreshold,
-    compactionAuto: input.compactionAuto === true,
-    projectPreload: input.projectPreload ?? null,
-    projectHandbookFiles: input.projectHandbookFiles ?? null,
-    toolCount: Math.max(0, Math.floor(input.toolCount ?? 0)),
-    groups,
-    meter,
-    lastCompaction: input.lastCompaction ?? null,
-    promptCache: input.promptCache ?? null
-  };
-}
-
-// src/domains/session/context-accounting.ts
-var IMAGE_CHAR_ESTIMATE = 4800;
-var TOKEN_CHARS = 4;
-var MESSAGE_OVERHEAD_TOKENS = 16;
-function ceilChars(chars) {
-  return Math.ceil(Math.max(0, chars) / TOKEN_CHARS);
-}
-function isRecord(value) {
-  return !!value && typeof value === "object" && !Array.isArray(value);
-}
-function jsonLength(value) {
-  if (value === void 0 || value === null) return 0;
-  if (typeof value === "string") return value.length;
-  if (typeof value === "number" || typeof value === "boolean") return String(value).length;
-  try {
-    return JSON.stringify(value).length;
-  } catch {
-    return String(value).length;
-  }
-}
-function argsLength(value) {
-  if (value === void 0 || value === null) return 0;
-  return typeof value === "string" ? value.length : jsonLength(value);
-}
-function blockChars(block) {
-  if (typeof block === "string") return block.length;
-  if (!isRecord(block)) return jsonLength(block);
-  if (block.type === "text" && typeof block.text === "string") return block.text.length;
-  if (block.type === "thinking" && typeof block.thinking === "string") return block.thinking.length;
-  if (block.type === "image") return IMAGE_CHAR_ESTIMATE;
-  if (block.type === "toolCall") {
-    const name = typeof block.name === "string" ? block.name : "";
-    const id = typeof block.id === "string" ? block.id : "";
-    return id.length + name.length + argsLength(block.arguments ?? block.args ?? block.input);
-  }
-  if (Array.isArray(block.content)) return contentChars(block.content);
-  if (typeof block.text === "string") return block.text.length;
-  if (typeof block.thinking === "string") return block.thinking.length;
-  return jsonLength(block);
-}
-function contentChars(content) {
-  if (typeof content === "string") return content.length;
-  if (!Array.isArray(content)) return jsonLength(content);
-  return content.reduce((sum, block) => sum + blockChars(block), 0);
-}
-function messageChars(message) {
-  if (!isRecord(message)) return jsonLength(message);
-  let chars = 0;
-  chars += typeof message.role === "string" ? message.role.length : 0;
-  const payload = message.payload !== void 0 ? message.payload : message.content;
-  chars += contentChars(payload);
-  chars += typeof message.toolCallId === "string" ? message.toolCallId.length : 0;
-  chars += typeof message.toolName === "string" ? message.toolName.length : 0;
-  chars += typeof message.errorMessage === "string" ? message.errorMessage.length : 0;
-  return chars;
-}
-function toolSchemaChars(tool) {
-  if (!isRecord(tool)) return jsonLength(tool);
-  const name = typeof tool.name === "string" ? tool.name.length : 0;
-  const description = typeof tool.description === "string" ? tool.description.length : 0;
-  const parameters = jsonLength(tool.parameters);
-  return name + description + parameters;
-}
-function usageTotalTokens(usage) {
-  if (!isRecord(usage)) return null;
-  const total = usage.totalTokens;
-  if (typeof total === "number" && Number.isFinite(total) && total > 0) return total;
-  const input = typeof usage.input === "number" && Number.isFinite(usage.input) ? usage.input : 0;
-  const output = typeof usage.output === "number" && Number.isFinite(usage.output) ? usage.output : 0;
-  const cacheRead = typeof usage.cacheRead === "number" && Number.isFinite(usage.cacheRead) ? usage.cacheRead : 0;
-  const cacheWrite = typeof usage.cacheWrite === "number" && Number.isFinite(usage.cacheWrite) ? usage.cacheWrite : 0;
-  const sum = input + output + cacheRead + cacheWrite;
-  return sum > 0 ? sum : null;
-}
-function usageInvalidated(message) {
-  return message.contextUsageInvalidated === true;
-}
-function latestUsableAssistantUsage(messages) {
-  for (let i = messages.length - 1; i >= 0; i -= 1) {
-    const message = messages[i];
-    if (message?.role !== "assistant") continue;
-    if (usageInvalidated(message)) continue;
-    const stopReason = message.stopReason;
-    if (stopReason === "error" || stopReason === "aborted") continue;
-    const tokens = usageTotalTokens(message.usage);
-    if (tokens !== null) return { index: i, tokens };
-  }
-  return null;
-}
-function finiteNonNegative2(value) {
-  return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : null;
-}
-function nestedNumber(root, path) {
-  let current = root;
-  for (const key of path) {
-    if (!isRecord(current)) return null;
-    current = current[key];
-  }
-  return finiteNonNegative2(current);
-}
-function estimateReasoningTextTokens(text) {
-  if (typeof text !== "string" || text.trim().length === 0) return null;
-  return Math.max(1, Math.round(text.length / TOKEN_CHARS));
-}
-function extractReasoningTokens(usage) {
-  if (!isRecord(usage)) return null;
-  const direct = finiteNonNegative2(usage.reasoningTokens) ?? finiteNonNegative2(usage.reasoning_tokens) ?? finiteNonNegative2(usage.reasoning);
-  if (direct !== null) return direct;
-  const paths = [
-    ["outputDetails", "reasoningTokens"],
-    ["output_details", "reasoning_tokens"],
-    ["output_tokens_details", "reasoning_tokens"],
-    ["completion_tokens_details", "reasoning_tokens"],
-    ["completionTokensDetails", "reasoningTokens"],
-    ["details", "reasoningTokens"]
-  ];
-  for (const path of paths) {
-    const value = nestedNumber(usage, path);
-    if (value !== null) return value;
-  }
-  return null;
-}
-function estimateAgentMessageTokens(message) {
-  return ceilChars(messageChars(message)) + MESSAGE_OVERHEAD_TOKENS;
-}
-function estimateAgentContextBreakdown(input) {
-  return {
-    systemPromptTokens: input.systemPrompt ? ceilChars(input.systemPrompt.length) : 0,
-    messageTokens: input.messages.reduce((sum, message) => sum + estimateAgentMessageTokens(message), 0),
-    pendingUserTokens: input.pendingUserText ? ceilChars(input.pendingUserText.length) : 0,
-    toolSchemaTokens: (input.tools ?? []).reduce((sum, tool) => sum + ceilChars(toolSchemaChars(tool)), 0)
-  };
-}
-function estimateAgentContextTokens(input) {
-  const breakdown = estimateAgentContextBreakdown(input);
-  const projection = breakdown.systemPromptTokens + breakdown.messageTokens + breakdown.pendingUserTokens + breakdown.toolSchemaTokens;
-  const usage = latestUsableAssistantUsage(input.messages);
-  if (!usage) return projection;
-  const trailingTokens = input.messages.slice(usage.index + 1).reduce((sum, message) => sum + estimateAgentMessageTokens(message), 0);
-  const anchored = usage.tokens + trailingTokens + breakdown.pendingUserTokens + breakdown.toolSchemaTokens;
-  return Math.max(projection, anchored);
-}
-function contextUsageSnapshot(tokens, contextWindow, breakdown) {
-  const window = typeof contextWindow === "number" && Number.isFinite(contextWindow) ? Math.max(0, contextWindow) : 0;
-  const base = tokens === null || tokens <= 0 || window <= 0 ? { tokens, contextWindow: window, percent: null } : { tokens, contextWindow: window, percent: Math.min(100, tokens / window * 100) };
-  return breakdown ? { ...base, breakdown } : base;
-}
-function getSnapshotsFilePath(meta) {
-  const safeMeta = {
-    ...meta,
-    cwdHash: meta.cwdHash || cwdHash(meta.cwd || process.cwd())
-  };
-  const paths = sessionPaths(safeMeta);
-  return join2(dirname(paths.current), "context-snapshots.jsonl");
-}
-function persistableSnapshot(snapshot) {
-  const {
-    systemPrompt: _systemPrompt,
-    conversationMessages: _messages,
-    activeToolSchemas: _tools,
-    pendingUserInput: _pending,
-    images: _images,
-    ...slim
-  } = snapshot;
-  return slim;
-}
-function appendContextSnapshot(meta, snapshot) {
-  try {
-    const file = getSnapshotsFilePath(meta);
-    appendFileSync(file, `${JSON.stringify(persistableSnapshot(snapshot))}
-`, "utf8");
-  } catch {
-  }
-}
-function getContextSnapshots(meta) {
-  try {
-    const file = getSnapshotsFilePath(meta);
-    if (!existsSync(file)) return [];
-    const content = readFileSync(file, "utf8");
-    const snapshots = [];
-    for (const line of content.split("\n")) {
-      if (line.trim().length === 0) continue;
-      try {
-        snapshots.push(JSON.parse(line));
-      } catch {
-      }
-    }
-    return snapshots;
-  } catch {
-    return [];
-  }
-}
-function getLatestContextSnapshot(meta) {
-  const list = getContextSnapshots(meta);
-  return list.at(-1) ?? null;
-}
-function buildSnapshotCategories(inputs) {
-  const raw = {
-    system: 0,
-    tools: 0,
-    agents: 0,
-    skills: 0,
-    memory: 0,
-    project: 0,
-    messages: 0
-  };
-  if (inputs.promptSegments && inputs.promptSegments.length > 0) {
-    for (const segment of inputs.promptSegments) {
-      const cat = categoryForSegment(segment.id);
-      raw[cat] = (raw[cat] ?? 0) + segment.tokenEstimate;
-    }
-  } else if (inputs.systemPrompt) {
-    raw.system += ceilChars(inputs.systemPrompt.length);
-  }
-  if (inputs.tools) {
-    for (const tool of inputs.tools) {
-      raw.tools += ceilChars(toolSchemaChars(tool));
-    }
-  }
-  if (inputs.messages) {
-    for (const msg of inputs.messages) {
-      raw.messages += estimateAgentMessageTokens(msg);
-    }
-  }
-  const decomposed = Object.keys(raw).reduce((sum, k) => sum + (raw[k] ?? 0), 0);
-  const window = inputs.effectiveContextWindow;
-  const threshold = inputs.compactionThreshold ?? DEFAULT_COMPACTION_THRESHOLD;
-  let reserve = 0;
-  let free = 0;
-  if (window > 0) {
-    const remaining = Math.max(0, window - decomposed);
-    reserve = Math.min(Math.round(window * (1 - threshold)), remaining);
-    free = Math.max(0, window - decomposed - reserve);
-  }
-  return {
-    system: raw.system,
-    tools: raw.tools,
-    agents: raw.agents,
-    skills: raw.skills,
-    memory: raw.memory,
-    project: raw.project,
-    messages: raw.messages,
-    reserve,
-    free,
-    streaming: 0
-  };
-}
-var SNAPSHOT_SPLIT_KEYS = ["system", "tools", "agents", "skills", "memory", "project", "messages"];
-function estimatedSplitSources() {
-  const splits = {};
-  for (const key of SNAPSHOT_SPLIT_KEYS) splits[key] = "estimated";
-  return splits;
-}
-var snapshotCounter = 0;
-function nextSnapshotId() {
-  snapshotCounter += 1;
-  return `snap-${Date.now()}-${snapshotCounter.toString(36)}-${Math.random().toString(36).slice(2, 7)}`;
-}
-function captureContextSnapshot(input) {
-  const categories = buildSnapshotCategories({
-    systemPrompt: input.systemPrompt,
-    promptSegments: input.promptSegments,
-    tools: input.activeToolSchemas,
-    messages: input.conversationMessages,
-    effectiveContextWindow: input.effectiveContextWindow,
-    compactionThreshold: input.compactionThreshold
-  });
-  return {
-    snapshotId: nextSnapshotId(),
-    sessionId: input.sessionId,
-    turnId: input.turnId,
-    providerId: input.providerId,
-    runtimeId: input.runtimeId,
-    modelId: input.modelId,
-    systemPrompt: input.systemPrompt,
-    promptSegments: input.promptSegments ? [...input.promptSegments] : void 0,
-    conversationMessages: [...input.conversationMessages],
-    activeToolSchemas: [...input.activeToolSchemas],
-    pendingUserInput: input.pendingUserInput,
-    images: input.images ? [...input.images] : void 0,
-    desiredContextWindow: input.desiredContextWindow,
-    effectiveContextWindow: input.effectiveContextWindow,
-    contextWindowSource: input.contextWindowSource,
-    categories,
-    sources: { total: "estimated", splits: estimatedSplitSources() },
-    promptHash: input.promptHash,
-    toolSignature: input.toolSignature,
-    messageRange: input.conversationMessages.length > 0 ? { start: 0, end: input.conversationMessages.length - 1 } : void 0
-  };
-}
-function snapshotInputTokens(snapshot) {
-  return SNAPSHOT_SPLIT_KEYS.reduce((sum, key) => sum + (snapshot.categories[key] ?? 0), 0);
-}
-function reconcileSnapshot(snapshot, usage) {
-  const exactInput = (usage.input || 0) + (usage.cacheRead || 0) + (usage.cacheWrite || 0);
-  const exactOutput = usage.output || 0;
-  if (exactInput <= 0) {
-    const updatedCategories2 = {
-      ...snapshot.categories,
-      streaming: exactOutput
-    };
-    const window2 = snapshot.effectiveContextWindow;
-    const used2 = snapshotInputTokens(snapshot) + exactOutput;
-    let reserve2 = snapshot.categories.reserve;
-    let free2 = 0;
-    if (window2 > 0) {
-      const remaining = Math.max(0, window2 - used2);
-      reserve2 = Math.min(reserve2, remaining);
-      free2 = Math.max(0, window2 - used2 - reserve2);
-    }
-    updatedCategories2.reserve = reserve2;
-    updatedCategories2.free = free2;
-    return {
-      ...snapshot,
-      categories: updatedCategories2,
-      sources: {
-        total: "estimated",
-        splits: snapshot.sources.splits
-      }
-    };
-  }
-  const keys = SNAPSHOT_SPLIT_KEYS;
-  const currentCategories = { ...snapshot.categories };
-  const sum = keys.reduce((s, k) => s + (currentCategories[k] ?? 0), 0);
-  let normalized;
-  if (sum === 0) {
-    normalized = keys.reduce(
-      (acc, k) => {
-        acc[k] = k === "messages" ? exactInput : 0;
-        return acc;
-      },
-      {}
-    );
-  } else {
-    normalized = {};
-    let newSum = 0;
-    for (const key of keys) {
-      const val = currentCategories[key] ?? 0;
-      const norm = Math.round(val / sum * exactInput);
-      normalized[key] = norm;
-      newSum += norm;
-    }
-    const diff = exactInput - newSum;
-    if (diff !== 0) {
-      const largestKey = keys.reduce((a, b) => (normalized[a] ?? 0) >= (normalized[b] ?? 0) ? a : b);
-      normalized[largestKey] = (normalized[largestKey] ?? 0) + diff;
-    }
-  }
-  const updatedCategories = {
-    ...snapshot.categories,
-    ...normalized,
-    streaming: exactOutput
-  };
-  const window = snapshot.effectiveContextWindow;
-  const used = exactInput + exactOutput;
-  let reserve = snapshot.categories.reserve;
-  let free = 0;
-  if (window > 0) {
-    const remaining = Math.max(0, window - used);
-    reserve = Math.min(reserve, remaining);
-    free = Math.max(0, window - used - reserve);
-  }
-  updatedCategories.reserve = reserve;
-  updatedCategories.free = free;
-  const splitsSources = keys.reduce(
-    (acc, k) => {
-      acc[k] = "reconciled";
-      return acc;
-    },
-    {}
-  );
-  return {
-    ...snapshot,
-    categories: updatedCategories,
-    sources: {
-      total: "reconciled",
-      splits: splitsSources
-    }
-  };
-}
-
 // src/engine/apis/residency.ts
 init_esm_shims();
 
 // src/engine/apis/residency-lock.ts
 init_esm_shims();
-import { join as join3 } from "node:path";
+import { join as join2 } from "node:path";
 var LOCK_WAIT_MS = 12e4;
 function lockTargetFor(targetKey) {
-  return join3(clioStatePath(), "residency-locks", targetKey.replace(/[^a-zA-Z0-9._-]+/g, "_"));
+  return join2(clioStatePath(), "residency-locks", targetKey.replace(/[^a-zA-Z0-9._-]+/g, "_"));
 }
 async function withResidencyLock(targetKey, fn) {
   return withStateFileLock(lockTargetFor(targetKey), fn, {
@@ -1506,7 +931,7 @@ init_esm_shims();
 // src/domains/providers/types/local-model-quirks.ts
 init_esm_shims();
 var KV_CACHE_QUANTS = ["f32", "f16", "q8_0", "q4_0", "q4_1", "iq4_nl", "q5_0", "q5_1"];
-function isRecord2(value) {
+function isRecord(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
 function asKvCacheQuant(value) {
@@ -1525,7 +950,7 @@ function asPenalty(value) {
   return typeof value === "number" && Number.isFinite(value) ? value : void 0;
 }
 function extractKvCache(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   const k = asKvCacheQuant(raw.kQuant);
   if (k !== void 0) out.kQuant = k;
@@ -1535,7 +960,7 @@ function extractKvCache(raw) {
   return Object.keys(out).length > 0 ? out : void 0;
 }
 function extractSamplingProfile(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   const temperature = asPositive(raw.temperature);
   if (temperature !== void 0) out.temperature = temperature;
@@ -1556,7 +981,7 @@ function extractSamplingProfile(raw) {
   return Object.keys(out).length > 0 ? out : void 0;
 }
 function extractSampling(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   const thinking = extractSamplingProfile(raw.thinking);
   if (thinking) out.thinking = thinking;
@@ -1576,7 +1001,7 @@ function asThinkingMechanism(value) {
   return THINKING_MECHANISMS.includes(value) ? value : void 0;
 }
 function extractBudgetByLevel(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   const minimal = asInteger(raw.minimal);
   if (minimal !== void 0) out.minimal = minimal;
@@ -1591,7 +1016,7 @@ function extractBudgetByLevel(raw) {
   return Object.keys(out).length > 0 ? out : void 0;
 }
 function extractEffortByLevel(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   if (typeof raw.off === "string" && raw.off.length > 0) out.off = raw.off;
   if (typeof raw.minimal === "string" && raw.minimal.length > 0) out.minimal = raw.minimal;
@@ -1602,7 +1027,7 @@ function extractEffortByLevel(raw) {
   return Object.keys(out).length > 0 ? out : void 0;
 }
 function extractThinkingQuirks(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const mechanism = asThinkingMechanism(raw.mechanism);
   if (!mechanism) return void 0;
   const out = { mechanism };
@@ -1614,7 +1039,7 @@ function extractThinkingQuirks(raw) {
   return out;
 }
 function extractLocalModelQuirks(raw) {
-  if (!isRecord2(raw)) return void 0;
+  if (!isRecord(raw)) return void 0;
   const out = {};
   const kvCache = extractKvCache(raw.kvCache);
   if (kvCache) out.kvCache = kvCache;
@@ -2279,7 +1704,7 @@ function targetRootUrl(target) {
 
 // src/domains/providers/runtimes/common/lmstudio-http.ts
 var catalogsByHost = /* @__PURE__ */ new Map();
-function isRecord3(value) {
+function isRecord2(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
 function positiveNumber(value) {
@@ -2339,7 +1764,7 @@ async function requestLmStudioJson(url2, init, timeoutMs, signal) {
     };
     if (data !== void 0) result.data = data;
     if (!response.ok) {
-      const message = isRecord3(data) ? nonEmptyString(data.error) : void 0;
+      const message = isRecord2(data) ? nonEmptyString(data.error) : void 0;
       result.error = message ?? `HTTP ${response.status}`;
     }
     return result;
@@ -2358,28 +1783,28 @@ function loadedInstances(value) {
   if (!Array.isArray(value)) return [];
   const instances = [];
   for (const raw of value) {
-    if (!isRecord3(raw)) continue;
+    if (!isRecord2(raw)) continue;
     const id = nonEmptyString(raw.id);
     if (!id) continue;
-    instances.push({ id, config: isRecord3(raw.config) ? { ...raw.config } : {} });
+    instances.push({ id, config: isRecord2(raw.config) ? { ...raw.config } : {} });
   }
   return instances;
 }
 function reasoningOptions(capabilities2) {
-  if (!isRecord3(capabilities2) || !isRecord3(capabilities2.reasoning)) return void 0;
+  if (!isRecord2(capabilities2) || !isRecord2(capabilities2.reasoning)) return void 0;
   const allowed = capabilities2.reasoning.allowed_options;
   if (!Array.isArray(allowed)) return void 0;
   const options = allowed.filter((entry) => typeof entry === "string");
   return options.length > 0 ? options : void 0;
 }
 function parseLmStudioV1Models(data) {
-  if (!isRecord3(data) || !Array.isArray(data.models)) return null;
+  if (!isRecord2(data) || !Array.isArray(data.models)) return null;
   const models = [];
   for (const raw of data.models) {
-    if (!isRecord3(raw)) continue;
+    if (!isRecord2(raw)) continue;
     const key = nonEmptyString(raw.key);
     if (!key) continue;
-    const capabilities2 = isRecord3(raw.capabilities) ? raw.capabilities : void 0;
+    const capabilities2 = isRecord2(raw.capabilities) ? raw.capabilities : void 0;
     const info = {
       key,
       loadedInstances: loadedInstances(raw.loaded_instances),
@@ -2397,7 +1822,7 @@ function parseLmStudioV1Models(data) {
     const options = reasoningOptions(capabilities2);
     if (options) info.reasoningOptions = options;
     if (capabilities2 && "reasoning" in capabilities2) {
-      info.reasoning = capabilities2.reasoning === true || isRecord3(capabilities2.reasoning);
+      info.reasoning = capabilities2.reasoning === true || isRecord2(capabilities2.reasoning);
     }
     models.push(info);
   }
@@ -2495,10 +1920,10 @@ function lmStudioReasoningEffort(level, options) {
   return "low";
 }
 function v0Models(data) {
-  if (!isRecord3(data) || !Array.isArray(data.data)) return null;
+  if (!isRecord2(data) || !Array.isArray(data.data)) return null;
   const models = [];
   for (const raw of data.data) {
-    if (!isRecord3(raw)) continue;
+    if (!isRecord2(raw)) continue;
     const key = nonEmptyString(raw.id);
     if (!key) continue;
     const state = nonEmptyString(raw.state);
@@ -2522,10 +1947,10 @@ function v0Models(data) {
   return models;
 }
 function openAIModels(data) {
-  if (!isRecord3(data) || !Array.isArray(data.data)) return null;
+  if (!isRecord2(data) || !Array.isArray(data.data)) return null;
   const models = [];
   for (const raw of data.data) {
-    if (!isRecord3(raw)) continue;
+    if (!isRecord2(raw)) continue;
     const key = nonEmptyString(raw.id);
     if (key) models.push({ key, loadedInstances: [], metadata: { ...raw } });
   }
@@ -2585,7 +2010,7 @@ async function greetLmStudio(target, ctx) {
     ctx.httpTimeoutMs,
     ctx.signal
   );
-  if (result.ok && isRecord3(result.data) && result.data.lmstudio === true) {
+  if (result.ok && isRecord2(result.data) && result.data.lmstudio === true) {
     return { ok: true, latencyMs: result.latencyMs };
   }
   return { ok: false, latencyMs: result.latencyMs, error: result.error ?? "endpoint did not return {lmstudio:true}" };
@@ -2914,6 +2339,15 @@ async function probeOpenAICompatReasoning(opts) {
 
 // src/domains/providers/runtimes/common/probe-helpers.ts
 init_esm_shims();
+
+// src/domains/providers/types/context-window-slots.ts
+init_esm_shims();
+function formatContextWindowSlots(contextWindow, slots) {
+  const format = (n) => Math.round(n).toLocaleString("en-US");
+  return `${format(contextWindow)} (${format(slots.totalContextSize)} / ${slots.slots} slots)`;
+}
+
+// src/domains/providers/runtimes/common/probe-helpers.ts
 async function probeUrl(url2, ctx, method = "GET") {
   const base = { url: url2, timeoutMs: ctx.httpTimeoutMs, method };
   return ctx.signal ? probeHttp({ ...base, signal: ctx.signal }) : probeHttp(base);
@@ -2958,11 +2392,15 @@ async function probeOpenAIModelCatalog(base, ctx, modelsPath = "/v1/models") {
     const state = modelStateFromOpenAIModelEntry(row) ?? (detailRow ? modelStateFromOpenAIModelEntry(detailRow) : void 0) ?? // A reported loaded context is itself the residency answer: nothing
     // serves a window for a model it has not loaded.
     (loadedContext !== void 0 ? { state: "loaded" } : void 0);
-    if (state) modelStates[row.id] = loadedContext === void 0 ? state : { ...state, contextLength: loadedContext };
+    const contextSlots = contextSlotsFromEntry(row) ?? (detailRow ? contextSlotsFromEntry(detailRow) : void 0);
+    const withSlots = contextSlots ? { ...state ?? { state: "unknown" }, contextSlots } : state;
+    if (withSlots) {
+      modelStates[row.id] = loadedContext === void 0 ? withSlots : { ...withSlots, contextLength: loadedContext };
+    }
   }
   return { models, modelCapabilities, modelStates };
 }
-function isRecord4(value) {
+function isRecord3(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
 function positiveNumber2(value) {
@@ -2989,7 +2427,7 @@ function booleanFromAny(record, keys) {
 }
 function nestedRecord(record, key) {
   const value = record[key];
-  return isRecord4(value) ? value : null;
+  return isRecord3(value) ? value : null;
 }
 function firstString(record, keys) {
   if (!record) return void 0;
@@ -3012,7 +2450,9 @@ function normalizeModelState(raw) {
   if (!value) return void 0;
   if (value === "loaded" || value === "ready" || value === "running" || value === "active") return "loaded";
   if (value === "loading" || value === "pending" || value === "queued" || value === "starting") return "loading";
-  if (value === "unloaded" || value === "not-loaded" || value === "idle" || value === "stopped") return "unloaded";
+  if (value === "unloaded" || value === "not-loaded" || value === "idle" || value === "sleeping" || value === "stopped") {
+    return "unloaded";
+  }
   if (value === "failed" || value === "error" || value === "errored") return "failed";
   if (value === "unknown") return "unknown";
   return void 0;
@@ -3043,11 +2483,14 @@ function statusArgsFromEntry(row) {
   const status = nestedRecord(row, "status");
   return argsFromStatus(status);
 }
+function contextSlotsFromEntry(row) {
+  return llamaCppRequestContextWindow(parseLlamaCppServerFlags(statusArgsFromEntry(row)))?.slots;
+}
 function capabilitiesFromOpenAIModelEntry(row) {
   const caps = {};
   const meta = nestedRecord(row, "meta");
   const flags = parseLlamaCppServerFlags(statusArgsFromEntry(row));
-  const contextWindow = positiveNumber2(flags.contextSize) ?? firstPositiveNumber(row, [
+  const contextWindow = llamaCppRequestContextWindow(flags)?.contextWindow ?? firstPositiveNumber(row, [
     // What is actually loaded outranks what the model could support: a
     // model served at 8k out of a possible 262k has an 8k window today,
     // and the run has to be planned against the real one.
@@ -3100,27 +2543,45 @@ async function probeOpenAIModelEntries(base, ctx) {
   if (!result.ok) return [];
   return modelEntries(result.data);
 }
+function llamaCppRequestContextWindow(flags) {
+  const total = positiveNumber2(flags.contextSize);
+  if (total === void 0) return void 0;
+  const parallel = positiveNumber2(flags.parallel);
+  if (parallel === void 0 || parallel <= 1 || flags.kvUnified === true) return { contextWindow: Math.floor(total) };
+  const slots = Math.floor(parallel);
+  return {
+    contextWindow: Math.floor(total / slots),
+    slots: { totalContextSize: Math.floor(total), slots }
+  };
+}
 function argsFromStatus(status) {
-  if (!isRecord4(status)) return [];
+  if (!isRecord3(status)) return [];
   const args = status.args;
   if (Array.isArray(args)) return args.filter((entry) => typeof entry === "string");
   if (typeof args === "string") return args.trim().split(/\s+/).filter(Boolean);
   return [];
 }
-function valueAfter(args, flag) {
-  const index = args.indexOf(flag);
-  if (index < 0) return void 0;
-  const value = args[index + 1];
-  return value && !value.startsWith("--") ? value : void 0;
+function looksLikeFlag(token) {
+  return token.startsWith("-") && Number.isNaN(Number(token));
 }
-function numberFlag(args, flag) {
-  const value = valueAfter(args, flag);
+function valueAfter(args, ...flags) {
+  for (const flag of flags) {
+    const index = args.indexOf(flag);
+    if (index < 0) continue;
+    const value = args[index + 1];
+    return value && !looksLikeFlag(value) ? value : void 0;
+  }
+  return void 0;
+}
+function numberFlag(args, ...flags) {
+  const value = valueAfter(args, ...flags);
   if (value === void 0) return void 0;
   const parsed = Number(value);
   return Number.isFinite(parsed) ? parsed : void 0;
 }
-function booleanFlag(args, flag) {
-  if (!args.includes(flag)) return void 0;
+function booleanFlag(args, ...flags) {
+  const flag = flags.find((candidate) => args.includes(candidate));
+  if (flag === void 0) return void 0;
   const value = valueAfter(args, flag);
   if (value === void 0) return true;
   const normalized = value.toLowerCase();
@@ -3130,7 +2591,7 @@ function booleanFlag(args, flag) {
 }
 function parseLlamaCppServerFlags(args) {
   const flags = {};
-  const ctxSize = numberFlag(args, "--ctx-size");
+  const ctxSize = numberFlag(args, "--ctx-size", "-c");
   if (ctxSize !== void 0) flags.contextSize = ctxSize;
   const maxTokens = numberFlag(args, "--n-predict");
   if (maxTokens !== void 0) flags.maxTokens = maxTokens;
@@ -3150,8 +2611,12 @@ function parseLlamaCppServerFlags(args) {
   if (topK !== void 0) flags.topK = topK;
   const nGpuLayers = numberFlag(args, "--n-gpu-layers");
   if (nGpuLayers !== void 0) flags.nGpuLayers = nGpuLayers;
-  const parallel = numberFlag(args, "--parallel");
+  const parallel = numberFlag(args, "--parallel", "-np");
   if (parallel !== void 0) flags.parallel = parallel;
+  const kvUnifiedAt = Math.max(args.lastIndexOf("--kv-unified"), args.lastIndexOf("-kvu"));
+  const noKvUnifiedAt = args.lastIndexOf("--no-kv-unified");
+  if (noKvUnifiedAt > kvUnifiedAt) flags.kvUnified = false;
+  else if (kvUnifiedAt >= 0) flags.kvUnified = booleanFlag(args, "--kv-unified", "-kvu") ?? true;
   const cacheTypeK = valueAfter(args, "--cache-type-k");
   if (cacheTypeK) flags.cacheTypeK = cacheTypeK;
   const cacheTypeV = valueAfter(args, "--cache-type-v");
@@ -3168,7 +2633,7 @@ function selectedModelEntry(entries, target) {
   return entries[0] ?? null;
 }
 function statusNotes(id, status) {
-  if (!isRecord4(status)) return [];
+  if (!isRecord3(status)) return [];
   const notes = [];
   if (status.failed === true) notes.push(`llama.cpp router marks ${id} as failed`);
   const state = typeof status.state === "string" ? status.state : void 0;
@@ -3183,7 +2648,8 @@ async function probeLlamaCppModelStatus(base, target, ctx) {
   if (args.length === 0) return { notes: statusNotes(selected.id, selected.status) };
   const flags = parseLlamaCppServerFlags(args);
   const caps = {};
-  if (flags.contextSize !== void 0 && flags.contextSize > 0) caps.contextWindow = flags.contextSize;
+  const window = llamaCppRequestContextWindow(flags);
+  if (window !== void 0) caps.contextWindow = window.contextWindow;
   if (flags.maxTokens !== void 0 && flags.maxTokens > 0) caps.maxTokens = flags.maxTokens;
   if (flags.reasoning === true || flags.reasoningBudget !== void 0) caps.reasoning = true;
   if (flags.mmproj) caps.vision = true;
@@ -3191,6 +2657,11 @@ async function probeLlamaCppModelStatus(base, target, ctx) {
   const enrichment = { modelId: selected.id, serverFlags: flags };
   if (Object.keys(caps).length > 0) enrichment.discoveredCapabilities = caps;
   const notes = statusNotes(selected.id, selected.status);
+  if (window?.slots) {
+    notes.push(
+      `${selected.id} context window ${formatContextWindowSlots(window.contextWindow, window.slots)}: --ctx-size is split across --parallel slots without --kv-unified`
+    );
+  }
   if (notes.length > 0) enrichment.notes = notes;
   return enrichment;
 }
@@ -5148,6 +4619,9 @@ function loadedContextWindowForModel(status, modelId) {
   const reported = status?.discoveredModelStates?.[modelId]?.contextLength;
   return typeof reported === "number" && Number.isFinite(reported) && reported > 0 ? reported : null;
 }
+function contextSlotsForModel(status, modelId) {
+  return status?.discoveredModelStates?.[modelId]?.contextSlots ?? null;
+}
 function modelResidencyForStatus(status, modelId) {
   if (loadedContextWindowForModel(status, modelId) !== null) return "resident";
   switch (status?.discoveredModelStates?.[modelId]?.state) {
@@ -5414,7 +4888,9 @@ function resolveRuntimeTarget(providers, input) {
     wireModelId,
     providers.knowledgeBase,
     probedContextWindow,
-    loadedContextWindow
+    loadedContextWindow,
+    void 0,
+    contextSlotsForModel(status, wireModelId)
   );
   capabilities2.contextWindow = contextWindowDetails.effectiveContextWindow;
   if (contextWindowDetails.warning) {
@@ -5495,7 +4971,8 @@ function refineRuntimeTargetWithModelHints(target, model, knowledgeBase) {
     // one the backend has open. Re-resolving without the loaded number would
     // hand the planner the declared window back on the first refinement.
     target.contextWindowDetails.loadedContextWindow,
-    modelHintContextWindow
+    modelHintContextWindow,
+    target.contextWindowDetails.contextWindowSlots
   );
   capabilities2.contextWindow = contextWindowDetails.effectiveContextWindow;
   const modelRuntime = resolveModelRuntimeCapabilities({
@@ -5554,12 +5031,16 @@ function firstRuntimeResolutionError(diagnostics) {
 function runtimeResolutionWarnings(diagnostics) {
   return diagnostics.filter((entry) => entry.severity === "warning").map((entry) => entry.message);
 }
+function runtimeResolutionWarningsBesideThinkingNotice(diagnostics, thinkingNotice) {
+  if (thinkingNotice.trim().length === 0) return runtimeResolutionWarnings(diagnostics);
+  return runtimeResolutionWarnings(diagnostics.filter((entry) => !entry.code.startsWith("thinking-")));
+}
 var DESIRED_CONTEXT_WINDOW = CLIO_MIN_CONTEXT_WINDOW;
 var FALLBACK_CONTEXT_WINDOW = CLIO_MIN_CONTEXT_WINDOW;
 function positiveWindow(value) {
   return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : void 0;
 }
-function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase, probedContextWindow, loadedContextWindow = null, modelHintContextWindow) {
+function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase, probedContextWindow, loadedContextWindow = null, modelHintContextWindow, probedContextSlots = null) {
   const catalogModel = getCatalogModelForRuntime(runtime.id, wireModelId);
   const kbHit = knowledgeBase?.lookup(wireModelId) ?? null;
   let modelDeclared;
@@ -5614,6 +5095,7 @@ function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase
   if (source === "descriptor-default" || source === "unknown") {
     provenanceNotice = `Context window ${effective} is Clio's assumed minimum, not a figure this target reported. Run 'clio-coder targets --probe' to read the real one.`;
   }
+  const contextWindowSlots = source === "probe" && probedContextSlots !== null && Math.floor(probedContextSlots.totalContextSize / probedContextSlots.slots) === effective ? probedContextSlots : null;
   return {
     declaredContextWindow,
     probedContextWindow,
@@ -5621,6 +5103,7 @@ function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase
     desiredContextWindow: desired,
     effectiveContextWindow: effective,
     contextWindowSource: source,
+    contextWindowSlots,
     warning,
     provenanceNotice
   };
@@ -5837,63 +5320,6 @@ function filterGemmaChannelStream(source, enabled) {
   return filtered;
 }
 
-// src/engine/strip-tokenizer-sentinels.ts
-init_esm_shims();
-var KNOWN_SENTINELS = [
-  "<|endoftext|>",
-  "<|im_start|>",
-  "<|im_end|>",
-  "<|eot_id|>",
-  "<|end_of_turn|>",
-  "<|fim_prefix|>",
-  "<|fim_middle|>",
-  "<|fim_suffix|>",
-  "<|file_separator|>",
-  "<s>",
-  "</s>"
-];
-var SENTINEL_RE = new RegExp(KNOWN_SENTINELS.map((s) => s.replace(/[\\^$.*+?()[\]{}|]/g, "\\$&")).join("|"), "g");
-var MAX_SENTINEL_LEN = KNOWN_SENTINELS.reduce((max, s) => s.length > max ? s.length : max, 0);
-function stripTokenizerSentinels(text) {
-  if (text.length === 0) return text;
-  return text.replace(SENTINEL_RE, "");
-}
-function isSentinelPrefix(tail) {
-  if (tail.length === 0) return false;
-  for (const sentinel of KNOWN_SENTINELS) {
-    if (sentinel.length > tail.length && sentinel.startsWith(tail)) return true;
-  }
-  return false;
-}
-function createSentinelStripper() {
-  let buffer = "";
-  return {
-    push(chunk) {
-      if (chunk.length === 0) return "";
-      buffer += chunk;
-      const stripped = buffer.replace(SENTINEL_RE, "");
-      const holdMax = Math.min(stripped.length, MAX_SENTINEL_LEN - 1);
-      let hold = 0;
-      for (let take = holdMax; take > 0; take--) {
-        const tail = stripped.slice(stripped.length - take);
-        if (isSentinelPrefix(tail)) {
-          hold = take;
-          break;
-        }
-      }
-      const safe = hold === 0 ? stripped : stripped.slice(0, stripped.length - hold);
-      buffer = hold === 0 ? "" : stripped.slice(stripped.length - hold);
-      return safe;
-    },
-    flush() {
-      if (buffer.length === 0) return "";
-      const remainder = buffer.replace(SENTINEL_RE, "");
-      buffer = "";
-      return remainder;
-    }
-  };
-}
-
 // src/engine/apis/output-budget.ts
 init_esm_shims();
 var CONTEXT_BUDGET_SAFETY_TOKENS = 1024;
@@ -6340,6 +5766,7 @@ var ollamaNativeApiProvider = {
 
 // src/engine/apis/openai-completions.ts
 init_esm_shims();
+import { TextDecoder as TextDecoder2 } from "node:util";
 import {
   createAssistantMessageEventStream as createAssistantMessageEventStream3
 } from "@earendil-works/pi-ai";
@@ -6472,24 +5899,24 @@ var POLL_INTERVAL_MS = 500;
 function isResidentState(state) {
   return state === "loaded" || state === "loading" || state === "sleeping";
 }
-function isRecord5(value) {
+function isRecord4(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
 function routerModelState(entry) {
   const status = entry.status;
-  const state = isRecord5(status) ? status.value : status;
+  const state = isRecord4(status) ? status.value : status;
   if (state === "loaded" || state === "loading" || state === "sleeping" || state === "unloaded" || state === "failed") {
     return state;
   }
   return "unknown";
 }
 function parseRouterModels(payload) {
-  if (!isRecord5(payload)) return [];
+  if (!isRecord4(payload)) return [];
   const data = payload.data;
   if (!Array.isArray(data)) return [];
   const models = [];
   for (const entry of data) {
-    if (!isRecord5(entry) || typeof entry.id !== "string") continue;
+    if (!isRecord4(entry) || typeof entry.id !== "string") continue;
     const tags = Array.isArray(entry.tags) ? entry.tags.filter((tag) => typeof tag === "string") : [];
     models.push({ id: entry.id, state: routerModelState(entry), tags });
   }
@@ -6499,7 +5926,7 @@ function residentModel(model) {
   return isResidentState(model.state);
 }
 function parseLlamaCppRouterProps(payload) {
-  if (!isRecord5(payload)) return {};
+  if (!isRecord4(payload)) return {};
   const maxInstances = payload.max_instances;
   return typeof maxInstances === "number" && Number.isFinite(maxInstances) && maxInstances > 0 ? { maxInstances: Math.floor(maxInstances) } : {};
 }
@@ -6727,6 +6154,12 @@ async function loadOwnedInstance(target, targetKey, body, apiKey, signal) {
   if (instanceId) ownedInstances(targetKey).add(instanceId);
   return instanceId;
 }
+var announcedResidencyFacts = /* @__PURE__ */ new Set();
+function emitResidencyNoticeOnce(key, notice) {
+  if (announcedResidencyFacts.has(key)) return;
+  announcedResidencyFacts.add(key);
+  emitResidencyNotice(notice);
+}
 async function ensureLmStudioResidency(model, options = {}) {
   const target = targetForModel(model);
   if (!target) return model.id;
@@ -6749,7 +6182,7 @@ async function ensureLmStudioResidency(model, options = {}) {
     );
   }
   if (resolution.wireModelId !== model.id) {
-    emitResidencyNotice({
+    emitResidencyNoticeOnce(`resolved|${info.targetId}|${model.id}|${resolution.wireModelId}`, {
       kind: "co-resident",
       level: "info",
       targetId: info.targetId,
@@ -6760,14 +6193,16 @@ async function ensureLmStudioResidency(model, options = {}) {
     });
   }
   if (resolution.peerTargets.length > 0) {
-    emitResidencyNotice({
+    const peers = resolution.peerTargets.join(", ");
+    const requested = resolution.wireModelId === model.id ? "" : ` (requested '${model.id}')`;
+    emitResidencyNoticeOnce(`peer|${info.targetId}|${model.id}|${resolution.wireModelId}|${peers}`, {
       kind: "co-resident",
       level: "warning",
       targetId: info.targetId,
       runtimeId: "lmstudio",
       model: resolution.wireModelId,
-      message: `LM Studio instance '${resolution.wireModelId}' is also loaded on ${resolution.peerTargets.join(", ")}; this request may be served by that LM Link peer.`,
-      detail: { peerTargets: resolution.peerTargets.join(", ") }
+      message: `LM Studio instance '${resolution.wireModelId}'${requested} is also loaded on ${peers}; a request may be served by that LM Link peer, and the footer and usage ledger name the id that answered when it differs.`,
+      detail: { requestedModel: model.id, wireModel: resolution.wireModelId, peerTargets: peers }
     });
   }
   if (!load || Object.keys(load).length === 0 || resolution.instance) return resolution.wireModelId;
@@ -6853,6 +6288,98 @@ function pickSamplingProfile2(quirks, thinkingActive) {
 function isPlainRecord(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
+function observeResponseModelIdLine(line, capture) {
+  if (capture.done) return;
+  const normalized = line.endsWith("\r") ? line.slice(0, -1) : line;
+  if (!normalized.startsWith("data:")) return;
+  const data = normalized.slice("data:".length).trimStart();
+  if (data.length === 0 || data === "[DONE]") return;
+  try {
+    const payload = JSON.parse(data);
+    if (!isPlainRecord(payload)) return;
+    const model = payload.model;
+    if (typeof model === "string" && model.trim().length > 0) {
+      capture.reportedModelId = model.trim();
+      capture.done = true;
+    }
+  } catch {
+  }
+}
+function observeResponseModelIdBytes(chunk, capture, flush = false) {
+  if (capture.done || !capture.decoder) return;
+  capture.buffer += chunk ? capture.decoder.decode(chunk, { stream: !flush }) : capture.decoder.decode();
+  let newline = capture.buffer.indexOf("\n");
+  while (newline >= 0) {
+    const line = capture.buffer.slice(0, newline);
+    capture.buffer = capture.buffer.slice(newline + 1);
+    observeResponseModelIdLine(line, capture);
+    if (capture.done) {
+      capture.buffer = "";
+      capture.decoder = null;
+      return;
+    }
+    newline = capture.buffer.indexOf("\n");
+  }
+  if (flush && capture.buffer.length > 0) {
+    observeResponseModelIdLine(capture.buffer, capture);
+    capture.buffer = "";
+  }
+  if (capture.done || flush) capture.decoder = null;
+}
+function captureResponseModelId(response, capture) {
+  if (!response.body || !response.headers.get("content-type")?.toLowerCase().includes("text/event-stream")) {
+    return response;
+  }
+  const body = response.body.pipeThrough(
+    new TransformStream({
+      transform(chunk, controller) {
+        capture.observed = true;
+        if (!capture.done) observeResponseModelIdBytes(chunk, capture);
+        controller.enqueue(chunk);
+      },
+      flush() {
+        capture.observed = true;
+        if (!capture.done) observeResponseModelIdBytes(void 0, capture, true);
+      }
+    })
+  );
+  return new Response(body, {
+    status: response.status,
+    statusText: response.statusText,
+    headers: response.headers
+  });
+}
+function withResponseModelIdCapture(options, sourceFactory) {
+  const capture = {
+    reportedModelId: null,
+    observed: false,
+    done: false,
+    buffer: "",
+    decoder: new TextDecoder2()
+  };
+  const fetchImpl = options.fetch ?? ((input, init) => globalThis.fetch(input, init));
+  const capturedOptions = {
+    ...options,
+    fetch: async (input, init) => captureResponseModelId(await fetchImpl(input, init), capture)
+  };
+  const source = sourceFactory(capturedOptions);
+  const annotated = createAssistantMessageEventStream3();
+  (async () => {
+    try {
+      for await (const event of source) {
+        const observation = capture.observed ? capture.reportedModelId === null ? { state: "not-reported" } : { state: "reported", reportedModelId: capture.reportedModelId } : { state: "not-observed" };
+        if (event.type === "done") event.message.responseModelIdObservation = observation;
+        else if (event.type === "error") event.error.responseModelIdObservation = observation;
+        annotated.push(event);
+      }
+      annotated.end();
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      throw new Error(message);
+    }
+  })();
+  return annotated;
+}
 function samplingParamsFromProfile(profile) {
   return {
     ...profile.topP !== void 0 ? { top_p: profile.topP } : {},
@@ -7336,16 +6863,15 @@ var openAICompletionsApiProvider = {
         stripSentinelsFromStream(
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
-              withLocalResidency(
-                model,
-                {
-                  ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
-                  ...options?.signal !== void 0 ? { signal: options.signal } : {}
-                },
-                (requestModel) => piOpenAICompletions.stream(
-                  requestModel,
-                  effectiveContext,
-                  withRemainingContextBudget(model, effectiveContext, withSamplers)
+              withResponseModelIdCapture(
+                withRemainingContextBudget(model, effectiveContext, withSamplers),
+                (capturedOptions) => withLocalResidency(
+                  model,
+                  {
+                    ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
+                    ...options?.signal !== void 0 ? { signal: options.signal } : {}
+                  },
+                  (requestModel) => piOpenAICompletions.stream(requestModel, effectiveContext, capturedOptions)
                 )
               ),
               resolved.family === "gemma-4"
@@ -7368,16 +6894,15 @@ var openAICompletionsApiProvider = {
         stripSentinelsFromStream(
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
-              withLocalResidency(
-                model,
-                {
-                  ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
-                  ...options?.signal !== void 0 ? { signal: options.signal } : {}
-                },
-                (requestModel) => piOpenAICompletions.streamSimple(
-                  requestModel,
-                  effectiveContext,
-                  withRemainingContextBudget(model, effectiveContext, withSamplers)
+              withResponseModelIdCapture(
+                withRemainingContextBudget(model, effectiveContext, withSamplers),
+                (capturedOptions) => withLocalResidency(
+                  model,
+                  {
+                    ...options?.apiKey !== void 0 ? { apiKey: options.apiKey } : {},
+                    ...options?.signal !== void 0 ? { signal: options.signal } : {}
+                  },
+                  (requestModel) => piOpenAICompletions.streamSimple(requestModel, effectiveContext, capturedOptions)
                 )
               ),
               resolved.family === "gemma-4"
@@ -7404,8 +6929,8 @@ function registerClioApiProviders() {
 
 // src/domains/providers/knowledge-base-path.ts
 init_esm_shims();
-import { existsSync as existsSync2, statSync as statSync2 } from "node:fs";
-import { delimiter, dirname as dirname2, join as join4 } from "node:path";
+import { existsSync, statSync as statSync2 } from "node:fs";
+import { delimiter, dirname, join as join3 } from "node:path";
 import { fileURLToPath } from "node:url";
 var MODEL_CATALOG_OVERLAY_DIR = "model-catalog.d";
 var MODEL_CATALOG_DIRS_ENV = "CLIO_CODER_MODEL_CATALOG_DIRS";
@@ -7417,25 +6942,25 @@ function isDirectory(path) {
   }
 }
 function resolveProvidersModelsDir(importMetaUrl) {
-  const start = dirname2(fileURLToPath(importMetaUrl));
+  const start = dirname(fileURLToPath(importMetaUrl));
   const directCandidates = [
-    join4(start, "models"),
-    join4(start, "..", "domains", "providers", "models"),
-    join4(start, "..", "providers-models")
+    join3(start, "models"),
+    join3(start, "..", "domains", "providers", "models"),
+    join3(start, "..", "providers-models")
   ];
   for (const candidate of directCandidates) {
     if (isDirectory(candidate)) return candidate;
   }
   let cursor = start;
   for (let i = 0; i < 8; i++) {
-    const packageJson = join4(cursor, "package.json");
-    const sourceModels = join4(cursor, "src", "domains", "providers", "models");
-    const distModels = join4(cursor, "dist", "providers-models");
-    if (existsSync2(packageJson)) {
+    const packageJson = join3(cursor, "package.json");
+    const sourceModels = join3(cursor, "src", "domains", "providers", "models");
+    const distModels = join3(cursor, "dist", "providers-models");
+    if (existsSync(packageJson)) {
       if (isDirectory(sourceModels)) return sourceModels;
       if (isDirectory(distModels)) return distModels;
     }
-    const parent = dirname2(cursor);
+    const parent = dirname(cursor);
     if (parent === cursor) break;
     cursor = parent;
   }
@@ -7461,8 +6986,8 @@ function resolveProviderModelCatalogDirs(importMetaUrl, options = {}) {
   const bundled = resolveProvidersModelsDir(importMetaUrl);
   const cwd = options.cwd ?? process.cwd();
   const overlays = uniqueExistingDirs([
-    join4(resolveClioDirs().config, MODEL_CATALOG_OVERLAY_DIR),
-    join4(cwd, ".clio-coder", MODEL_CATALOG_OVERLAY_DIR),
+    join3(resolveClioDirs().config, MODEL_CATALOG_OVERLAY_DIR),
+    join3(cwd, ".clio-coder", MODEL_CATALOG_OVERLAY_DIR),
     ...envOverlayDirs()
   ]);
   return {
@@ -7481,7 +7006,7 @@ function resolveProviderKnowledgeBaseRoots(importMetaUrl, options = {}) {
 
 // src/domains/providers/plugins.ts
 init_esm_shims();
-import { join as join5 } from "node:path";
+import { join as join4 } from "node:path";
 function extractPluginPackages(settings) {
   if (!settings || typeof settings !== "object") return [];
   const raw = settings.runtimePlugins;
@@ -7490,7 +7015,7 @@ function extractPluginPackages(settings) {
 }
 async function loadPluginRuntimes(registry, settings) {
   const loaded = [];
-  const pluginDir = join5(clioConfigDir(), "runtimes");
+  const pluginDir = join4(clioConfigDir(), "runtimes");
   const packages = extractPluginPackages(settings);
   try {
     const ids = await registry.loadFromDir(pluginDir, activateExternalPluginApiBridge);
@@ -7518,8 +7043,8 @@ async function loadPluginRuntimes(registry, settings) {
 // src/domains/providers/types/knowledge-base.ts
 init_esm_shims();
 var import_yaml = __toESM(require_dist(), 1);
-import { readdirSync as readdirSync2, readFileSync as readFileSync2, statSync as statSync3 } from "node:fs";
-import { join as join6 } from "node:path";
+import { readdirSync as readdirSync2, readFileSync, statSync as statSync3 } from "node:fs";
+import { join as join5 } from "node:path";
 var FileKnowledgeBase = class {
   roots;
   loaded = [];
@@ -7533,7 +7058,7 @@ var FileKnowledgeBase = class {
       const files = collectYamlFiles(root.dir);
       for (const file of files) {
         const name = root.label ? `${root.label}:${file.name}` : file.name;
-        const raw = readFileSync2(file.path, "utf8");
+        const raw = readFileSync(file.path, "utf8");
         const parsed = (0, import_yaml.parse)(raw);
         if (!Array.isArray(parsed)) {
           throw new Error(`knowledge base file ${name} must be a YAML list of KnowledgeBaseEntry`);
@@ -7591,7 +7116,7 @@ function collectYamlFiles(dir, prefix = "") {
   const out = [];
   const entries = readdirSync2(dir, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name));
   for (const entry of entries) {
-    const path = join6(dir, entry.name);
+    const path = join5(dir, entry.name);
     const name = prefix ? `${prefix}/${entry.name}` : entry.name;
     if (entry.isDirectory()) {
       out.push(...collectYamlFiles(path, name));
@@ -8171,24 +7696,6 @@ export {
   getCatalogModelForRuntime,
   getRuntimeRegistry,
   credentialsPresent,
-  stripTokenizerSentinels,
-  DEFAULT_COMPACTION_THRESHOLD,
-  shouldCompact,
-  AutoCompactionTrigger,
-  buildContextLedger,
-  ceilChars,
-  contentChars,
-  estimateReasoningTextTokens,
-  extractReasoningTokens,
-  estimateAgentMessageTokens,
-  estimateAgentContextBreakdown,
-  estimateAgentContextTokens,
-  contextUsageSnapshot,
-  appendContextSnapshot,
-  getLatestContextSnapshot,
-  captureContextSnapshot,
-  snapshotInputTokens,
-  reconcileSnapshot,
   setGlobalDefaultMaxOutputTokens,
   resolveReservedOutputTokens,
   setResidencyNoticeSink,
@@ -8207,6 +7714,7 @@ export {
   registerClioApiProviders,
   resolveProviderKnowledgeBaseRoots,
   loadPluginRuntimes,
+  formatContextWindowSlots,
   registerBuiltinRuntimes,
   FileKnowledgeBase,
   isOrchestratorEligibleRuntime,
@@ -8228,7 +7736,8 @@ export {
   runtimeTargetSnapshot,
   firstRuntimeResolutionError,
   runtimeResolutionWarnings,
+  runtimeResolutionWarningsBesideThinkingNotice,
   normalizeCostProvenance,
   ProvidersDomainModule
 };
-//# sourceMappingURL=chunk-LZSJBIVT.js.map
+//# sourceMappingURL=chunk-DJVECN66.js.map