@iowarp/clio-coder 0.3.4 → 0.3.6

package/dist/{chunk-QWU7ZBO7.js → chunk-DJVECN66.js}

@@ -27,7 +27,7 @@ import {
   ceilChars,
   estimateAgentMessageTokens,
   toolSchemaChars
-} from "./chunk-BPGS2WCQ.js";
+} from "./chunk-GEYXPTRF.js";
 import {
   activateExternalPluginApiBridge,
   calculateEngineCost,
@@ -49,7 +49,7 @@ import {
 } from "./chunk-ZGVHUX3M.js";
 import {
   BusChannels
-} from "./chunk-TTNYS3EA.js";
+} from "./chunk-G7MUEIGA.js";
 import {
   __toESM,
   init_esm_shims
@@ -2339,6 +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);
@@ -2383,7 +2392,11 @@ 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 };
 }
@@ -2437,7 +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;
@@ -2468,11 +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.
@@ -2525,6 +2543,17 @@ 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 (!isRecord3(status)) return [];
   const args = status.args;
@@ -2532,20 +2561,27 @@ function argsFromStatus(status) {
   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();
@@ -2555,7 +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;
@@ -2575,8 +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");
@@ -2608,7 +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;
@@ -2616,6 +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;
 }
@@ -4573,6 +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) {
@@ -4839,7 +4888,9 @@ function resolveRuntimeTarget(providers, input) {
     wireModelId,
     providers.knowledgeBase,
     probedContextWindow,
-    loadedContextWindow
+    loadedContextWindow,
+    void 0,
+    contextSlotsForModel(status, wireModelId)
   );
   capabilities2.contextWindow = contextWindowDetails.effectiveContextWindow;
   if (contextWindowDetails.warning) {
@@ -4920,7 +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({
@@ -4979,12 +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;
@@ -5039,6 +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,
@@ -5046,6 +5103,7 @@ function resolveContextWindowDetails(target, runtime, wireModelId, knowledgeBase
     desiredContextWindow: desired,
     effectiveContextWindow: effective,
     contextWindowSource: source,
+    contextWindowSlots,
     warning,
     provenanceNotice
   };
@@ -5708,6 +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";
@@ -6095,6 +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;
@@ -6117,7 +6182,7 @@ async function ensureLmStudioResidency(model, options = {}) {
     );
   }
   if (resolution.wireModelId !== model.id) {
-    emitResidencyNotice({
+    emitResidencyNoticeOnce(`resolved|${info.targetId}|${model.id}|${resolution.wireModelId}`, {
       kind: "co-resident",
       level: "info",
       targetId: info.targetId,
@@ -6128,14 +6193,16 @@ async function ensureLmStudioResidency(model, options = {}) {
     });
   }
   if (resolution.peerTargets.length > 0) {
-    emitResidencyNotice({
+    const peers = resolution.peerTargets.join(", ");
+    const requested = resolution.wireModelId === model.id ? "" : ` (requested '${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;
@@ -6221,6 +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 } : {},
@@ -6704,16 +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"
@@ -6736,16 +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"
@@ -7557,6 +7714,7 @@ export {
   registerClioApiProviders,
   resolveProviderKnowledgeBaseRoots,
   loadPluginRuntimes,
+  formatContextWindowSlots,
   registerBuiltinRuntimes,
   FileKnowledgeBase,
   isOrchestratorEligibleRuntime,
@@ -7578,7 +7736,8 @@ export {
   runtimeTargetSnapshot,
   firstRuntimeResolutionError,
   runtimeResolutionWarnings,
+  runtimeResolutionWarningsBesideThinkingNotice,
   normalizeCostProvenance,
   ProvidersDomainModule
 };
-//# sourceMappingURL=chunk-QWU7ZBO7.js.map
+//# sourceMappingURL=chunk-DJVECN66.js.map