@iowarp/clio-coder 0.3.8 → 0.3.9

package/dist/{chunk-YS5VLNH5.js → chunk-LXPJXFM5.js}

@@ -5,19 +5,19 @@ import {
 import {
   probeWorkspace
 } from "./chunk-HVDIIIQW.js";
-import {
-  isSessionEntry,
-  isSessionHeader
-} from "./chunk-GOXNB3AO.js";
 import {
   protectArtifact
 } from "./chunk-MV3K5QF2.js";
+import {
+  isSessionEntry,
+  isSessionHeader
+} from "./chunk-RAPCMZL4.js";
 import {
   stripTokenizerSentinels
 } from "./chunk-CFGTUFWB.js";
 import {
   BusChannels
-} from "./chunk-TT36MB5S.js";
+} from "./chunk-RKRLDWD3.js";
 import {
   CURRENT_SESSION_FORMAT_VERSION,
   appendSessionFileEntry,
@@ -29,7 +29,7 @@ import {
   readSessionMeta,
   resumeSession,
   sessionPaths
-} from "./chunk-3BINW3FP.js";
+} from "./chunk-KV2AOLDF.js";
 import {
   clioStateDir,
   stateRootRemoved
@@ -1475,4 +1475,4 @@ export {
   ledgerUsageCalls,
   SessionDomainModule
 };
-//# sourceMappingURL=chunk-YS5VLNH5.js.map
+//# sourceMappingURL=chunk-LXPJXFM5.js.map

package/dist/{chunk-7RFXX52T.js → chunk-MIX5N5AC.js}

@@ -10,10 +10,10 @@ import {
   ceilChars,
   estimateAgentMessageTokens,
   toolSchemaChars
-} from "./chunk-RWSI4YD7.js";
+} from "./chunk-AEYBF3TB.js";
 import {
   readSettings
-} from "./chunk-IJ7RPIYJ.js";
+} from "./chunk-PNY46YEY.js";
 import {
   sleep
 } from "./chunk-FQ4SKYE4.js";
@@ -22,7 +22,7 @@ import {
 } from "./chunk-ZGVHUX3M.js";
 import {
   BusChannels
-} from "./chunk-TT36MB5S.js";
+} from "./chunk-RKRLDWD3.js";
 import {
   CLIO_CONTEXT_WINDOW_WARN_BELOW,
   CLIO_MIN_CONTEXT_WINDOW,
@@ -34,10 +34,10 @@ import {
   registerBuiltinRuntimes,
   requestLmStudioJson,
   resolveLmStudioInstance
-} from "./chunk-NMPKI6XL.js";
+} from "./chunk-JEQQR47K.js";
 import {
   listKnownModelsForRuntime
-} from "./chunk-IIZWH4XA.js";
+} from "./chunk-774ILSRL.js";
 import {
   authNotRequiredStatus,
   openAuthStorage,
@@ -65,18 +65,100 @@ import {
   getCatalogModelForRuntime,
   mergeCapabilities,
   resolveCostProvenance
-} from "./chunk-PT7HYKEM.js";
+} from "./chunk-VO2LKSTM.js";
 import {
   activateExternalPluginApiBridge,
   calculateEngineCost,
   ensurePiAiRegistered,
   registerEngineApiProvider
-} from "./chunk-FCSXB6T2.js";
+} from "./chunk-UOSL25KY.js";
 import {
   __toESM,
   init_esm_shims
 } from "./chunk-3R73A4XB.js";
 
+// src/domains/providers/endpoint-capacity.ts
+init_esm_shims();
+
+// src/core/endpoint-key.ts
+init_esm_shims();
+function canonicalEndpointUrl(raw) {
+  if (!raw?.trim()) return null;
+  try {
+    const url = new URL(raw.trim().replace(/^ws:/u, "http:").replace(/^wss:/u, "https:"));
+    url.protocol = url.protocol.toLowerCase();
+    if (url.protocol !== "http:" && url.protocol !== "https:") return null;
+    url.hostname = url.hostname.toLowerCase();
+    url.username = "";
+    url.password = "";
+    url.hash = "";
+    url.search = "";
+    let path = url.pathname.replace(/\/{2,}/gu, "/").replace(/\/$/u, "");
+    if (path.endsWith("/v1")) path = path.slice(0, -3);
+    url.pathname = path || "/";
+    return url.toString().replace(/\/$/u, "");
+  } catch {
+    return null;
+  }
+}
+
+// src/domains/providers/endpoint-capacity.ts
+var UNBOUNDED_LOCAL_NATIVE_RUNTIMES = /* @__PURE__ */ new Set(["vllm", "sglang"]);
+function canonicalEndpointKey(value) {
+  return canonicalEndpointUrl(typeof value === "string" ? value : value.url);
+}
+function endpointLabel(endpointKey) {
+  try {
+    const url = new URL(endpointKey);
+    return `${url.host}${url.pathname === "/" ? "" : url.pathname}`;
+  } catch {
+    return endpointKey;
+  }
+}
+function positiveInteger(value) {
+  return typeof value === "number" && Number.isInteger(value) && value > 0 ? value : void 0;
+}
+function endpointCapacityForStatus(status) {
+  const key = canonicalEndpointKey(status.target);
+  if (key === null) return null;
+  const override = positiveInteger(status.target.maxConcurrentRequests);
+  if (override !== void 0) return { key, label: endpointLabel(key), limit: override, source: "override" };
+  const discovered = positiveInteger(status.probeCapabilities?.parallelSlots);
+  if (discovered !== void 0) return { key, label: endpointLabel(key), limit: discovered, source: "discovered" };
+  if (status.runtime?.tier !== "local-native" || UNBOUNDED_LOCAL_NATIVE_RUNTIMES.has(status.runtime.id)) return null;
+  return { key, label: endpointLabel(key), limit: 1, source: "local-native-default" };
+}
+function sourceRank(source) {
+  return source === "override" ? 2 : source === "discovered" ? 1 : 0;
+}
+function endpointCapacitiesForStatuses(statuses) {
+  const capacities = {};
+  for (const status of statuses) {
+    const candidate = endpointCapacityForStatus(status);
+    if (candidate === null) continue;
+    const current = capacities[candidate.key];
+    if (current === void 0 || sourceRank(candidate.source) > sourceRank(current.source) || sourceRank(candidate.source) === sourceRank(current.source) && candidate.limit < current.limit) {
+      capacities[candidate.key] = candidate;
+    }
+  }
+  return capacities;
+}
+var foregroundStreams = /* @__PURE__ */ new Map();
+function registerForegroundStream(endpointKey) {
+  foregroundStreams.set(endpointKey, (foregroundStreams.get(endpointKey) ?? 0) + 1);
+  let held = true;
+  return () => {
+    if (!held) return;
+    held = false;
+    const next = (foregroundStreams.get(endpointKey) ?? 1) - 1;
+    if (next <= 0) foregroundStreams.delete(endpointKey);
+    else foregroundStreams.set(endpointKey, next);
+  };
+}
+function foregroundStreamUsage() {
+  return Object.fromEntries(foregroundStreams);
+}
+
 // src/engine/apis/residency.ts
 init_esm_shims();
 
@@ -85,7 +167,8 @@ init_esm_shims();
 import { join } from "node:path";
 var LOCK_WAIT_MS = 12e4;
 function lockTargetFor(targetKey) {
-  return join(clioStatePath(), "residency-locks", targetKey.replace(/[^a-zA-Z0-9._-]+/g, "_"));
+  const key = canonicalEndpointKey(targetKey) ?? targetKey;
+  return join(clioStatePath(), "residency-locks", key.replace(/[^a-zA-Z0-9._-]+/g, "_"));
 }
 async function withResidencyLock(targetKey, fn) {
   return withStateFileLock(lockTargetFor(targetKey), fn, {
@@ -114,6 +197,9 @@ function emitResidencyNotice(notice) {
   } catch {
   }
 }
+function emitResidencyMutation(mutation) {
+  getSharedBus().emit(BusChannels.ResidencyMutation, { ...mutation, at: Date.now() });
+}
 var protectedModelsProvider = null;
 function setProtectedModelsProvider(provider) {
   protectedModelsProvider = provider;
@@ -344,6 +430,13 @@ async function restoreEvictedModels(adapter, evicted, loadError) {
     try {
       await adapter.reloadEvicted(entry.modelId);
       restored.push(entry.modelId);
+      emitResidencyMutation({
+        targetKey: adapter.targetKey,
+        targetId: adapter.targetId,
+        runtimeId: adapter.runtimeId,
+        model: entry.modelId,
+        operation: "load"
+      });
     } catch {
       failed.push(entry.modelId);
     }
@@ -460,12 +553,28 @@ async function reconcileResidency(adapter) {
           await adapter.unload(entry.modelId);
           forgetClioLoaded(adapter.targetKey, entry.modelId);
           evicted.push(entry);
+          emitResidencyMutation({
+            targetKey: adapter.targetKey,
+            targetId: adapter.targetId,
+            runtimeId: adapter.runtimeId,
+            model: entry.modelId,
+            operation: "evict"
+          });
         } catch {
         }
       }
       if (!adapter.load) return;
       try {
         await adapter.load(adapter.keepModelId);
+        if (!plan.keepResident) {
+          emitResidencyMutation({
+            targetKey: adapter.targetKey,
+            targetId: adapter.targetId,
+            runtimeId: adapter.runtimeId,
+            model: adapter.keepModelId,
+            operation: "load"
+          });
+        }
       } catch (error) {
         await restoreEvictedModels(adapter, evicted, error);
         throw error;
@@ -1477,7 +1586,7 @@ function resolveRuntimeTarget(providers, input) {
   const capabilityResolution = modelCapabilitiesFor(providers, status, wireModelId);
   const capabilities = { ...capabilityResolution.capabilities };
   const probedContextWindow = probeCapabilitiesForModel(status, wireModelId)?.contextWindow ?? null;
-  const loadedContextWindow = loadedContextWindowForModel(status, wireModelId);
+  const loadedContextWindow = loadedContextWindowForModel(status, wireModelId) ?? input.knownLoadedContextWindow ?? null;
   const contextWindowDetails = resolveContextWindowDetails(
     target,
     runtime,
@@ -1705,6 +1814,12 @@ function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase
   };
 }
 
+// src/domains/providers/types/cost-provenance.ts
+init_esm_shims();
+function normalizeCostProvenance(value) {
+  return value ?? "unknown";
+}
+
 // src/domains/providers/index.ts
 init_esm_shims();
 
@@ -1722,6 +1837,18 @@ import {
   Ollama
 } from "ollama";
 
+// src/core/residency-target-key.ts
+init_esm_shims();
+var RESIDENCY_RUNTIME_IDS = /* @__PURE__ */ new Set([
+  "llamacpp",
+  "lmstudio",
+  "ollama-native"
+]);
+function residencyTargetKey(runtimeId, baseUrl) {
+  if (!RESIDENCY_RUNTIME_IDS.has(runtimeId) || typeof baseUrl !== "string") return null;
+  return canonicalEndpointUrl(baseUrl) ?? baseUrl;
+}
+
 // src/engine/gemma-channel-filter.ts
 init_esm_shims();
 import { createAssistantMessageEventStream } from "@earendil-works/pi-ai";
@@ -1991,7 +2118,7 @@ async function reconcileOllamaResidency(model, headers) {
   if (!baseUrl) return;
   const metadata2 = model.clio;
   const adapter = {
-    targetKey: `ollama-native|${baseUrl}`,
+    targetKey: residencyTargetKey("ollama-native", baseUrl),
     targetId: ollamaTargetId(model),
     runtimeId: "ollama-native",
     keepModelId: model.id,
@@ -2559,6 +2686,26 @@ async function fetchRouterModels(input, fetchImpl) {
   if (!response.ok) throw new Error(`HTTP ${response.status}`);
   return parseRouterModels(await response.json());
 }
+var RouterModelPostError = class extends Error {
+  constructor(url, status, responseBody) {
+    super(`llama.cpp router rejected ${url}: HTTP ${status}; response: ${responseBody}`);
+    this.url = url;
+    this.status = status;
+    this.responseBody = responseBody;
+    this.name = "RouterModelPostError";
+  }
+  url;
+  status;
+  responseBody;
+};
+async function routerErrorBody(response) {
+  try {
+    const body = (await response.text()).trim();
+    return body.length > 0 ? body : "<empty body>";
+  } catch (error) {
+    return `<unreadable body: ${error instanceof Error ? error.message : String(error)}>`;
+  }
+}
 async function postRouterModel(fetchImpl, url, modelId) {
   const response = await fetchImpl(url, {
     method: "POST",
@@ -2567,7 +2714,27 @@ async function postRouterModel(fetchImpl, url, modelId) {
     signal: AbortSignal.timeout(LOAD_TIMEOUT_MS)
   });
   if (response.ok) return;
-  throw new Error(`llama.cpp router rejected ${url}: HTTP ${response.status}`);
+  throw new RouterModelPostError(url, response.status, await routerErrorBody(response));
+}
+function alreadyRunningResponse(body) {
+  return body.toLowerCase().includes("already running");
+}
+async function postRouterLoad(input, fetchImpl, modelId) {
+  try {
+    await postRouterModel(fetchImpl, loadUrl(input.baseUrl), modelId);
+  } catch (error) {
+    if (!(error instanceof RouterModelPostError)) throw error;
+    let model;
+    try {
+      model = (await fetchRouterModels(input, fetchImpl)).find((entry) => entry.id === modelId);
+    } catch {
+      if (alreadyRunningResponse(error.responseBody)) return;
+      throw error;
+    }
+    if (alreadyRunningResponse(error.responseBody)) return;
+    if (model !== void 0 && model.state !== "unloaded" && model.state !== "failed") return;
+    throw error;
+  }
 }
 async function waitForLoaded(input, fetchImpl, modelId) {
   const started = performance.now();
@@ -2582,7 +2749,7 @@ async function waitForLoaded(input, fetchImpl, modelId) {
 }
 async function ensureModelLoaded(input, fetchImpl, modelId, knownState) {
   if (knownState === "loaded" || knownState === "sleeping") return;
-  if (knownState !== "loading") await postRouterModel(fetchImpl, loadUrl(input.baseUrl), modelId);
+  if (knownState !== "loading") await postRouterLoad(input, fetchImpl, modelId);
   await waitForLoaded(input, fetchImpl, modelId);
 }
 async function restoreDisplacedPinned(input, fetchImpl, before) {
@@ -2609,7 +2776,7 @@ async function ensureLlamaCppResidency(input) {
   const fetchImpl = input.fetchImpl ?? fetch;
   let snapshot = [];
   const adapter = {
-    targetKey: `llamacpp|${input.baseUrl}`,
+    targetKey: residencyTargetKey("llamacpp", input.baseUrl),
     targetId: input.targetId,
     runtimeId: input.runtimeId,
     keepModelId: input.keepModelId,
@@ -2809,7 +2976,7 @@ async function ensureLmStudioResidency(model, options = {}) {
     (entry) => entry.loadedInstances.map((instance) => ({ modelKey: entry.key, identifier: instance.id, instance }))
   );
   const managed = residencyManagedFor(info.lifecycle);
-  const targetKey = `lmstudio|${lmStudioRootUrl(target.url ?? model.baseUrl)}`;
+  const targetKey = residencyTargetKey("lmstudio", target.url ?? model.baseUrl);
   const contextLength = target.lmstudio?.load?.contextLength;
   const plan = await reconcileResidency({
     targetKey,
@@ -2852,17 +3019,36 @@ async function ensureLmStudioResidency(model, options = {}) {
       });
     }
   }
+  const loadAndReport = async () => {
+    const instanceId = await loadOwnedInstance(target, targetKey, body, options.apiKey, options.signal);
+    emitResidencyMutation({
+      targetKey,
+      targetId: info.targetId,
+      runtimeId: "lmstudio",
+      model: modelKey,
+      operation: "load"
+    });
+    return instanceId ?? model.id;
+  };
   try {
-    return await loadOwnedInstance(target, targetKey, body, options.apiKey, options.signal) ?? model.id;
+    return await loadAndReport();
   } catch (error) {
     if (plan.decision === "observe" || plan.fallbackEvict.length === 0) throw error;
     return withResidencyLock(targetKey, async () => {
       for (const candidate of plan.fallbackEvict) {
         for (const entry of instances.filter((resident) => resident.modelKey === candidate.modelId)) {
-          await unloadOwnedInstance(target, targetKey, entry.identifier, options.apiKey, options.signal);
+          if (await unloadOwnedInstance(target, targetKey, entry.identifier, options.apiKey, options.signal)) {
+            emitResidencyMutation({
+              targetKey,
+              targetId: info.targetId,
+              runtimeId: "lmstudio",
+              model: candidate.modelId,
+              operation: "evict"
+            });
+          }
         }
       }
-      return await loadOwnedInstance(target, targetKey, body, options.apiKey, options.signal) ?? model.id;
+      return loadAndReport();
     });
   }
 }
@@ -2884,8 +3070,36 @@ function pickSamplingProfile2(quirks, thinkingActive) {
 function isPlainRecord(value) {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
-function observeResponseModelIdLine(line, capture) {
-  if (capture.done) return;
+function nonnegativeFiniteNumber(value) {
+  return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : null;
+}
+function backendCompletionTimings(value, source) {
+  if (!isPlainRecord(value)) return null;
+  const promptN = nonnegativeFiniteNumber(value.prompt_n);
+  const predictedN = nonnegativeFiniteNumber(value.predicted_n);
+  const promptMs = nonnegativeFiniteNumber(value.prompt_ms);
+  const predictedMs = nonnegativeFiniteNumber(value.predicted_ms);
+  if (promptN === null || predictedN === null || promptMs === null || predictedMs === null) return null;
+  const cacheN = nonnegativeFiniteNumber(value.cache_n);
+  return {
+    promptTokens: promptN + (cacheN ?? 0),
+    cachedTokens: cacheN,
+    predictedTokens: predictedN,
+    promptMs,
+    predictedMs,
+    source
+  };
+}
+function backendTimingsSourceForModel(model) {
+  const metadata2 = model.clio;
+  if (model.provider === "llamacpp" && metadata2?.runtimeId === "llamacpp") return "llamacpp-timings";
+  if (model.provider === "lmstudio" && metadata2?.runtimeId === "lmstudio") return "lmstudio-timings";
+  return null;
+}
+function captureCanStopEarly(capture) {
+  return capture.modelIdDone && capture.backendTimingsSource === null;
+}
+function observeResponseMetadataLine(line, capture) {
   const normalized = line.endsWith("\r") ? line.slice(0, -1) : line;
   if (!normalized.startsWith("data:")) return;
   const data = normalized.slice("data:".length).trimStart();
@@ -2893,23 +3107,29 @@ function observeResponseModelIdLine(line, capture) {
   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;
+    if (!capture.modelIdDone) {
+      const model = payload.model;
+      if (typeof model === "string" && model.trim().length > 0) {
+        capture.reportedModelId = model.trim();
+        capture.modelIdDone = true;
+      }
+    }
+    if (capture.backendTimingsSource !== null) {
+      const timings = backendCompletionTimings(payload.timings, capture.backendTimingsSource);
+      if (timings !== null) capture.backendTimings = timings;
     }
   } catch {
   }
 }
 function observeResponseModelIdBytes(chunk, capture, flush = false) {
-  if (capture.done || !capture.decoder) return;
+  if (!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) {
+    observeResponseMetadataLine(line, capture);
+    if (captureCanStopEarly(capture)) {
       capture.buffer = "";
       capture.decoder = null;
       return;
@@ -2917,10 +3137,10 @@ function observeResponseModelIdBytes(chunk, capture, flush = false) {
     newline = capture.buffer.indexOf("\n");
   }
   if (flush && capture.buffer.length > 0) {
-    observeResponseModelIdLine(capture.buffer, capture);
+    observeResponseMetadataLine(capture.buffer, capture);
     capture.buffer = "";
   }
-  if (capture.done || flush) capture.decoder = null;
+  if (captureCanStopEarly(capture) || flush) capture.decoder = null;
 }
 function captureResponseModelId(response, capture) {
   if (!response.body || !response.headers.get("content-type")?.toLowerCase().includes("text/event-stream")) {
@@ -2930,12 +3150,12 @@ function captureResponseModelId(response, capture) {
     new TransformStream({
       transform(chunk, controller) {
         capture.observed = true;
-        if (!capture.done) observeResponseModelIdBytes(chunk, capture);
+        observeResponseModelIdBytes(chunk, capture);
         controller.enqueue(chunk);
       },
       flush() {
         capture.observed = true;
-        if (!capture.done) observeResponseModelIdBytes(void 0, capture, true);
+        observeResponseModelIdBytes(void 0, capture, true);
       }
     })
   );
@@ -2945,11 +3165,13 @@ function captureResponseModelId(response, capture) {
     headers: response.headers
   });
 }
-function withResponseModelIdCapture(options, sourceFactory) {
+function withResponseModelIdCapture(model, options, sourceFactory) {
   const capture = {
     reportedModelId: null,
     observed: false,
-    done: false,
+    modelIdDone: false,
+    backendTimings: null,
+    backendTimingsSource: backendTimingsSourceForModel(model),
     buffer: "",
     decoder: new TextDecoder()
   };
@@ -2964,8 +3186,13 @@ function withResponseModelIdCapture(options, sourceFactory) {
     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;
+        if (event.type === "done") {
+          event.message.responseModelIdObservation = observation;
+          if (capture.backendTimings !== null) event.message.backendTimings = capture.backendTimings;
+        } else if (event.type === "error") {
+          event.error.responseModelIdObservation = observation;
+          if (capture.backendTimings !== null) event.error.backendTimings = capture.backendTimings;
+        }
         annotated.push(event);
       }
       annotated.end();
@@ -3460,6 +3687,7 @@ var openAICompletionsApiProvider = {
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
               withResponseModelIdCapture(
+                model,
                 withRemainingContextBudget(model, effectiveContext, withSamplers),
                 (capturedOptions) => withLocalResidency(
                   model,
@@ -3491,6 +3719,7 @@ var openAICompletionsApiProvider = {
           stripNeverReasoningFromStream(
             filterGemmaChannelStream(
               withResponseModelIdCapture(
+                model,
                 withRemainingContextBudget(model, effectiveContext, withSamplers),
                 (capturedOptions) => withLocalResidency(
                   model,
@@ -4283,12 +4512,6 @@ function resolveModelReference(rawPattern, providers) {
   return buildResult(base, pickFirst(partial), thinkingLevel);
 }
 
-// src/domains/providers/types/cost-provenance.ts
-init_esm_shims();
-function normalizeCostProvenance(value) {
-  return value ?? "unknown";
-}
-
 // src/domains/providers/index.ts
 var ProvidersDomainModule = {
   manifest: ProvidersManifest,
@@ -4296,8 +4519,15 @@ var ProvidersDomainModule = {
 };
 
 export {
+  residencyTargetKey,
   setGlobalDefaultMaxOutputTokens,
   resolveReservedOutputTokens,
+  canonicalEndpointKey,
+  endpointLabel,
+  endpointCapacityForStatus,
+  endpointCapacitiesForStatuses,
+  registerForegroundStream,
+  foregroundStreamUsage,
   setResidencyNoticeSink,
   setProtectedModelsProvider,
   withRunOverrides,
@@ -4330,4 +4560,4 @@ export {
   normalizeCostProvenance,
   ProvidersDomainModule
 };
-//# sourceMappingURL=chunk-7RFXX52T.js.map
+//# sourceMappingURL=chunk-MIX5N5AC.js.map