muonroi-cli 1.8.4 → 1.8.5

package/dist/src/providers/vision-proxy.js

@@ -1,15 +1,13 @@
 /**
  * Vision proxy layer for text-only models (e.g. DeepSeek V4).
  *
- * When the active model does not support vision, image parts in the message
- * are extracted, sent to a cheap vision model on SiliconFlow
- * (Qwen/Qwen2.5-VL-32B-Instruct @ $0.27/M tokens), and replaced with
- * structured text descriptions before the message reaches the primary model.
+ * Image parts in messages are sent to the catalog-configured vision backend
+ * (default: Z.ai glm-4.6v-flash) and replaced with native-sight observations
+ * so the primary model reasons as if it saw the image directly.
  */
 import { getModelInfo } from "../models/registry.js";
-import { apiBaseFor } from "./endpoints.js";
-import { loadKeyForProvider } from "./keychain.js";
 import { UI_LAYOUT_SCHEMA_HINT } from "./mcp-vision-bridge.js";
+import { callVisionBackend, collectVisionUnavailableReasons, findNativeVisionFallback, formatNativeVisionObservation, formatNativeVisionUnavailable, isVisionBackendAvailable, looksLikeOcrIntent, resolveAvailableVisionChain, wrapAnalyzerInstructions, } from "./vision-backend.js";
 const DESIGN_INTENT_PATTERNS = [
     /\bredesign\b/i,
     /\bdesign\s*(system|token|spec|review)?\b/i,
@@ -26,24 +24,40 @@ function looksLikeDesignIntent(text) {
         return false;
     return DESIGN_INTENT_PATTERNS.some((re) => re.test(text));
 }
-const VISION_MODELS = [
-    "Qwen/Qwen3-VL-8B-Instruct",
-    "Qwen/Qwen3-VL-30B-A3B-Instruct",
-    "Qwen/Qwen3-VL-32B-Instruct",
-];
-const SILICONFLOW_BASE = apiBaseFor("siliconflow");
-const REQUEST_TIMEOUT_MS = 90_000;
-/**
- * Returns true when the model cannot handle image content natively.
- */
 export function needsVisionProxy(modelId) {
     const info = getModelInfo(modelId);
     return info?.supportsVision === false;
 }
 /**
- * Process messages through the vision proxy.
- * Only messages with image parts are modified; text-only messages pass through.
+ * Decide how to handle images when the active model is text-only:
+ * proxy backend, switch to a native vision model, or surface unavailable.
  */
+export async function planImageHandlingForTextOnlyModel(opts) {
+    if (!needsVisionProxy(opts.primaryModelId)) {
+        return { strategy: "proxy" };
+    }
+    const kind = opts.kind ?? "default";
+    if (await isVisionBackendAvailable(kind)) {
+        return { strategy: "proxy" };
+    }
+    const fallback = await findNativeVisionFallback({ excludeModelId: opts.primaryModelId });
+    if (fallback) {
+        return { strategy: "native_model", fallback };
+    }
+    const reasons = await collectVisionUnavailableReasons(kind);
+    return {
+        strategy: "unavailable",
+        notice: formatNativeVisionUnavailable(opts.imageCount, reasons),
+    };
+}
+/** True when proxy backend or a native vision model can handle images for this text-only model. */
+export async function canHandleImagesForTextOnlyModel(modelId) {
+    if (!needsVisionProxy(modelId))
+        return true;
+    if (await isVisionBackendAvailable())
+        return true;
+    return (await findNativeVisionFallback({ excludeModelId: modelId })) !== null;
+}
 export async function proxyVision(messages, modelId, signal) {
     if (!needsVisionProxy(modelId)) {
         return { messages, proxied: false, imageCount: 0 };
@@ -66,6 +80,7 @@ export async function proxyVision(messages, modelId, signal) {
         totalImages += imageParts.length;
         const userText = textParts.map((t) => t.text).join("\n");
         const designIntent = looksLikeDesignIntent(userText);
+        const ocrIntent = looksLikeOcrIntent(userText);
         const svgFast = trySvgFastPath(imageParts, designIntent);
         if (svgFast) {
             processed.push({
@@ -74,27 +89,22 @@ export async function proxyVision(messages, modelId, signal) {
             });
             continue;
         }
-        const descriptions = await describeImages(imageParts, textParts, signal, designIntent);
+        const kind = designIntent ? "design" : ocrIntent ? "ocr" : "default";
+        const descriptions = await describeImages(imageParts, textParts, signal, kind);
         const newContent = [...textParts, { type: "text", text: descriptions }];
         processed.push({ ...msg, content: newContent });
     }
     return { messages: processed, proxied: totalImages > 0, imageCount: totalImages };
 }
-async function describeImages(images, contextTexts, signal, designIntent = false) {
-    let apiKey;
-    try {
-        apiKey = await loadKeyForProvider("siliconflow");
-    }
-    catch {
-        return buildFallbackDescription(images, ["SILICONFLOW_API_KEY not configured"]);
-    }
+async function describeImages(images, contextTexts, signal, kind = "default") {
     const userContext = contextTexts.map((t) => t.text).join("\n");
+    const designIntent = kind === "design";
     const visionContent = [];
     visionContent.push({
         type: "text",
-        text: designIntent
+        text: wrapAnalyzerInstructions(designIntent
             ? buildDesignPrompt(userContext, images.length)
-            : buildAnalysisPrompt(userContext, images.length),
+            : buildAnalysisPrompt(userContext, images.length, kind === "ocr"), kind),
     });
     for (const img of images) {
         visionContent.push({
@@ -106,24 +116,12 @@ async function describeImages(images, contextTexts, signal, designIntent = false
         });
     }
     const responseFormat = designIntent ? { type: "json_object" } : undefined;
-    const failureReasons = [];
-    for (const model of VISION_MODELS) {
-        try {
-            const result = await callVisionModel(model, visionContent, apiKey, signal, responseFormat);
-            if (result.ok)
-                return formatVisionResult(result.text, images.length, model, designIntent);
-            failureReasons.push(`${model}: ${result.reason}`);
-            console.warn(`[vision-proxy] ${model} failed (${result.reason}), trying next...`);
-        }
-        catch (err) {
-            if (signal?.aborted)
-                throw err;
-            const msg = err instanceof Error ? err.message : String(err);
-            failureReasons.push(`${model}: ${msg}`);
-            console.warn(`[vision-proxy] ${model} threw (${msg}), trying next...`);
-        }
+    const chain = await resolveAvailableVisionChain(kind);
+    const result = await callVisionBackend(chain, visionContent, signal, responseFormat);
+    if (result.ok) {
+        return formatNativeVisionObservation(result.text, { imageCount: images.length });
     }
-    return buildFallbackDescription(images, failureReasons);
+    return formatNativeVisionUnavailable(images.length, [result.reason]);
 }
 function trySvgFastPath(images, designIntent) {
     if (images.length === 0)
@@ -142,64 +140,23 @@ function trySvgFastPath(images, designIntent) {
         return svg.length > 32_000 ? `${svg.slice(0, 32_000)}\n<!-- truncated -->` : svg;
     })
         .join("\n\n");
-    const header = `[Vision Proxy — ${images.length} SVG source(s) passed through (fast-path, no vision call)]`;
-    const guidance = designIntent
-        ? `\nThe raw SVG below IS the layout contract. Map nodes to:\n${UI_LAYOUT_SCHEMA_HINT}\n`
-        : "\nThe raw SVG below is vector text — read element attributes (x, y, width, fill, text content) directly.\n";
-    return `\n${header}${guidance}\n\`\`\`svg\n${decoded}\n\`\`\`\n[/Vision Proxy]\n`;
+    const observation = designIntent
+        ? `I see ${images.length} SVG source(s). The vector markup below is exact — use element attributes (x, y, width, fill, text) directly.\n\nMap to layout schema:\n${UI_LAYOUT_SCHEMA_HINT}\n\n\`\`\`svg\n${decoded}\n\`\`\``
+        : `I see ${images.length} SVG source(s). Vector markup (read attributes directly):\n\n\`\`\`svg\n${decoded}\n\`\`\``;
+    return formatNativeVisionObservation(observation, { imageCount: images.length });
 }
-async function callVisionModel(model, content, apiKey, signal, responseFormat) {
-    const controller = new AbortController();
-    let timedOut = false;
-    const timeout = setTimeout(() => {
-        timedOut = true;
-        controller.abort();
-    }, REQUEST_TIMEOUT_MS);
-    if (signal) {
-        signal.addEventListener("abort", () => controller.abort(), { once: true });
-    }
-    let res;
-    try {
-        res = await fetch(`${SILICONFLOW_BASE}/chat/completions`, {
-            method: "POST",
-            headers: {
-                "Content-Type": "application/json",
-                Authorization: `Bearer ${apiKey}`,
-            },
-            body: JSON.stringify({
-                model,
-                messages: [{ role: "user", content }],
-                max_tokens: 3072,
-                temperature: 0.1,
-                ...(responseFormat ? { response_format: responseFormat } : {}),
-            }),
-            signal: controller.signal,
-        });
-    }
-    catch (err) {
-        clearTimeout(timeout);
-        if (timedOut)
-            return { ok: false, reason: `timeout after ${REQUEST_TIMEOUT_MS}ms` };
-        if (signal?.aborted)
-            throw err;
-        return { ok: false, reason: `network error: ${err instanceof Error ? err.message : String(err)}` };
-    }
-    clearTimeout(timeout);
-    if (!res.ok) {
-        const errText = await res.text().catch(() => "unknown error");
-        return { ok: false, reason: `HTTP ${res.status} ${errText.slice(0, 200)}` };
-    }
-    const data = (await res.json().catch(() => null));
-    const text = data?.choices?.[0]?.message?.content;
-    if (!text)
-        return { ok: false, reason: "empty response body" };
-    return { ok: true, text };
-}
-function buildAnalysisPrompt(userContext, imageCount) {
+function buildAnalysisPrompt(userContext, imageCount, ocrFocus) {
     const plural = imageCount > 1 ? `${imageCount} images` : "the image";
+    const ocrBlock = ocrFocus
+        ? [
+            "OCR focus: transcribe ALL visible text exactly (preserve line breaks, labels, error codes).",
+            "Note any text that is blurry or partially cut off.",
+            "",
+        ]
+        : [];
     return [
-        `You are a vision model. Analyze ${plural} for a software developer working in a coding CLI tool.`,
-        "Use your full vision capability — do NOT refuse on the basis of being text-only.",
+        `Analyze ${plural} for a software developer working in a coding CLI.`,
+        ...ocrBlock,
         "Focus on:",
         "- UI layout and component structure (if screenshot)",
         "- All visible text, labels, buttons, and form elements",
@@ -212,45 +169,19 @@ function buildAnalysisPrompt(userContext, imageCount) {
         userContext ? `\nDeveloper's context: "${userContext}"` : "",
     ].join("\n");
 }
-function formatVisionResult(description, imageCount, model, designIntent = false) {
-    const usedModel = model ?? VISION_MODELS[0];
-    const tag = designIntent ? "design contract via" : "analyzed via";
-    const single = designIntent ? "design contract via" : "analyzed via";
-    const header = imageCount > 1
-        ? `[Vision Proxy — ${imageCount} images ${tag} ${usedModel}]`
-        : `[Vision Proxy — image ${single} ${usedModel}]`;
-    return `\n${header}\n${description}\n[/Vision Proxy]\n`;
-}
 function buildDesignPrompt(userContext, imageCount) {
     const plural = imageCount > 1 ? `${imageCount} images` : "the image";
     return [
-        `Analyze ${plural} as a UI/UX design contract for a software developer who CANNOT see the image.`,
-        "Output will be consumed by another AI agent to recreate or redesign this UI, so precision matters more than prose.",
-        "",
-        "Return ONLY a single JSON object matching this exact shape (no markdown, no commentary, no code fences):",
+        `Analyze ${plural} as a UI/UX design contract for a software developer.`,
+        "Return ONLY a single JSON object matching this exact shape (no markdown, no commentary):",
         UI_LAYOUT_SCHEMA_HINT,
         "",
         "Extraction rules:",
-        "- bbox: pixel coordinates relative to the visible image (origin top-left). Estimate when unsure but never omit.",
-        "- Colors: 6-digit lowercase hex (#rrggbb). Sample dominant pixel.",
-        "- Typography sizePx: estimate from cap-height, round to nearest 2.",
-        "- spacingScalePx / radiusScalePx: distinct observed values, sorted ascending.",
-        "- components[]: every visible interactive or content element. Stable IDs like 'btn_signup'.",
-        "- children[]: component IDs nested inside this component.",
-        "- hierarchy[]: visual reading order of headings — level 1 hero, 2 section, 3 sub.",
+        "- bbox: pixel coordinates relative to the visible image (origin top-left).",
+        "- Colors: 6-digit lowercase hex (#rrggbb).",
+        "- components[]: every visible interactive or content element.",
         "- text: EXACT visible string. null only when no text (icon-only).",
-        "- notes[]: things schema cannot capture — gradients, illustrations, motion.",
-        "",
-        "If a field is genuinely not determinable, use null. Never invent values.",
         userContext ? `\nDeveloper's context: "${userContext}"` : "",
     ].join("\n");
 }
-function buildFallbackDescription(images, reasons = []) {
-    const count = images.length;
-    const types = [...new Set(images.map((i) => i.mediaType))].join(", ");
-    const detail = reasons.length > 0
-        ? `Reason(s): ${reasons.join(" | ")}`
-        : "Reason: unknown — check SILICONFLOW_API_KEY and model availability.";
-    return `\n[Vision Proxy — unavailable, ${count} image(s) (${types}) could not be analyzed. ${detail}]\n`;
-}
 //# sourceMappingURL=vision-proxy.js.map

package/dist/src/providers/wire-debug.js

@@ -90,6 +90,95 @@ function summarizeMessage(m) {
     }
     return { role, contentKind: "string", textChars: 0 };
 }
+/**
+ * Reduce a request body (from AI SDK's `error.requestBodyValues`) to a
+ * PII-safe shape: top-level param key list + each assistant message's field
+ * key list. Values are NEVER persisted — only key names. This is the
+ * diagnostic evidence needed when a provider rejects "an invalid parameter"
+ * (Z.ai 1210, SiliconFlow 20015) and we must see which field was present or
+ * missing on each assistant turn in a multi-step tool loop.
+ */
+function summarizeRequestBodyShape(body) {
+    if (!body || typeof body !== "object")
+        return undefined;
+    const b = body;
+    const paramKeys = Object.keys(b).sort();
+    const configParamValues = extractConfigParamValues(b);
+    const messages = b.messages;
+    if (!Array.isArray(messages)) {
+        return {
+            paramKeys,
+            assistantFieldKeys: [],
+            assistantCount: 0,
+            assistantLens: [],
+            totalReasoningChars: 0,
+            responseFormat: b.response_format ?? undefined,
+            configParamValues,
+        };
+    }
+    const assistantMsgs = messages.filter((m) => m?.role === "assistant");
+    // De-duplicate the field-key signature across assistant messages so a long
+    // history doesn't bloat the log — we only care WHICH fields co-occur.
+    const signatures = new Set();
+    const assistantLens = [];
+    let totalReasoningChars = 0;
+    for (const m of assistantMsgs) {
+        const msg = m;
+        const keys = Object.keys(msg ?? {})
+            .sort()
+            .join(",");
+        signatures.add(keys);
+        const rc = msg?.reasoning_content;
+        const rcLen = typeof rc === "string" ? rc.length : null;
+        if (rcLen !== null)
+            totalReasoningChars += rcLen;
+        const contentLen = typeof msg?.content === "string"
+            ? msg.content.length
+            : Array.isArray(msg?.content)
+                ? msg.content.reduce((n, p) => n + (typeof p?.text === "string" ? p.text.length : 0), 0)
+                : 0;
+        const toolCallsLen = Array.isArray(msg?.tool_calls) ? msg.tool_calls.length : 0;
+        assistantLens.push({ rc: rcLen, content: contentLen, toolCalls: toolCallsLen });
+    }
+    return {
+        paramKeys,
+        assistantFieldKeys: [...signatures],
+        assistantCount: assistantMsgs.length,
+        assistantLens,
+        totalReasoningChars,
+        responseFormat: b.response_format ?? undefined,
+        configParamValues,
+    };
+}
+/**
+ * Scalar SDK-config params whose VALUES are PII-safe (generation config, not
+ * user prompt content) and which are the likely culprits when a provider
+ * rejects with a generic "Invalid API parameter" (Z.ai 1210, SiliconFlow
+ * 20015). `messages` and `tools` are deliberately excluded — their values may
+ * carry prompts/PII, and counts are already captured elsewhere.
+ */
+const CONFIG_PARAM_KEYS = [
+    "reasoning_effort",
+    "verbosity",
+    "tool_choice",
+    "response_format",
+    "max_tokens",
+    "temperature",
+    "top_p",
+    "frequency_penalty",
+    "presence_penalty",
+    "seed",
+    "stop",
+    "parallel_tool_calls",
+];
+function extractConfigParamValues(body) {
+    const out = {};
+    for (const k of CONFIG_PARAM_KEYS) {
+        if (k in body)
+            out[k] = body[k] ?? null;
+    }
+    return out;
+}
 export const wireDebug = {
     enabled: ENABLED,
     /** Log an outgoing streamText/streamObject call. */
@@ -129,6 +218,12 @@ export const wireDebug = {
             statusCode: typeof e?.statusCode === "number" ? e.statusCode : undefined,
             url: typeof e?.url === "string" ? e.url : undefined,
             responseBody: typeof e?.responseBody === "string" ? e.responseBody.slice(0, 4000) : undefined,
+            // Evidence for "Invalid API parameter" (e.g. Z.ai code 1210,
+            // SiliconFlow 20015): capture the SHAPE (key names only — no values,
+            // PII-safe) of every assistant message in the request body so we can
+            // see which field the provider rejected and whether reasoning_content
+            // was present/missing across the multi-step tool-loop history.
+            requestBodyShape: summarizeRequestBodyShape(e?.requestBodyValues),
             cause: e?.cause && typeof e.cause === "object"
                 ? {
                     message: typeof e.cause.message === "string"

package/dist/src/router/decide.d.ts

@@ -14,6 +14,14 @@ export interface DecideOpts {
     signal?: AbortSignal;
     defaultModel: string;
     defaultProvider: string;
+    /**
+     * Optional session id for audit logging. When provided, a `routing`
+     * interaction event is emitted whenever the turn's routed model differs from
+     * `defaultModel` (the stored session.model) — making per-turn overrides
+     * observable instead of silent. Fixes the "session.model lie": a user on
+     * flash could not see when the router silently promoted them to pro.
+     */
+    sessionId?: string;
     /** Override home directory for ledger (testing). */
     homeOverride?: string;
     /** PIL enrichment signals — forwarded to EE context. */
@@ -27,6 +35,11 @@ export interface DecideOpts {
         projectSize?: "small" | "medium" | "large" | null;
         filesTouched?: number;
         mode?: string | null;
+        turnIndex?: number;
+        messageCount?: number;
+        compactionCount?: number;
+        totalSavedTokens?: number;
+        compactionSummary?: string | null;
     };
 }
 export declare function clearRouteCache(): void;