@steipete/summarize 0.10.0 → 0.11.1

package/dist/esm/run/flows/url/summary.js

@@ -1,20 +1,21 @@
-import { isTwitterStatusUrl, isYouTubeUrl } from '@steipete/summarize-core/content/url';
-import { countTokens } from 'gpt-tokenizer';
-import { render as renderMarkdownAnsi } from 'markdansi';
-import { buildLanguageKey, buildLengthKey, buildPromptHash, buildSummaryCacheKey, hashString, normalizeContentForHash, } from '../../../cache.js';
-import { formatOutputLanguageForJson } from '../../../language.js';
-import { parseGatewayStyleModelId } from '../../../llm/model-id.js';
-import { buildAutoModelAttempts } from '../../../model-auto.js';
-import { buildLinkSummaryPrompt, SUMMARY_LENGTH_TARGET_CHARACTERS, SUMMARY_SYSTEM_PROMPT, } from '../../../prompts/index.js';
-import { parseCliUserModelId } from '../../env.js';
-import { buildExtractFinishLabel, buildLengthPartsForFinishLine, writeFinishLine, } from '../../finish-line.js';
-import { resolveTargetCharacters } from '../../format.js';
-import { writeVerbose } from '../../logging.js';
-import { prepareMarkdownForTerminal } from '../../markdown.js';
-import { runModelAttempts } from '../../model-attempts.js';
-import { buildOpenRouterNoAllowedProvidersMessage } from '../../openrouter.js';
-import { isRichTty, markdownRenderWidth, supportsColor } from '../../terminal.js';
-import { coerceSummaryWithSlides, interleaveSlidesIntoTranscript, normalizeSummarySlideHeadings, } from './slides-text.js';
+import { isTwitterStatusUrl, isYouTubeUrl } from "@steipete/summarize-core/content/url";
+import { countTokens } from "gpt-tokenizer";
+import { render as renderMarkdownAnsi } from "markdansi";
+import { buildLanguageKey, buildLengthKey, buildPromptHash, buildSummaryCacheKey, hashString, normalizeContentForHash, } from "../../../cache.js";
+import { formatOutputLanguageForJson } from "../../../language.js";
+import { parseGatewayStyleModelId } from "../../../llm/model-id.js";
+import { buildAutoModelAttempts } from "../../../model-auto.js";
+import { buildLinkSummaryPrompt, SUMMARY_LENGTH_TARGET_CHARACTERS, SUMMARY_SYSTEM_PROMPT, } from "../../../prompts/index.js";
+import { readLastSuccessfulCliProvider, writeLastSuccessfulCliProvider, } from "../../cli-fallback-state.js";
+import { parseCliUserModelId } from "../../env.js";
+import { buildExtractFinishLabel, buildLengthPartsForFinishLine, writeFinishLine, } from "../../finish-line.js";
+import { resolveTargetCharacters } from "../../format.js";
+import { writeVerbose } from "../../logging.js";
+import { prepareMarkdownForTerminal } from "../../markdown.js";
+import { runModelAttempts } from "../../model-attempts.js";
+import { buildOpenRouterNoAllowedProvidersMessage } from "../../openrouter.js";
+import { isRichTty, markdownRenderWidth, supportsColor } from "../../terminal.js";
+import { coerceSummaryWithSlides, interleaveSlidesIntoTranscript, normalizeSummarySlideHeadings, } from "./slides-text.js";
 const MAX_SLIDE_TRANSCRIPT_CHARS_BY_PRESET = {
     short: 2500,
     medium: 5000,
@@ -25,7 +26,7 @@ const MAX_SLIDE_TRANSCRIPT_CHARS_BY_PRESET = {
 const SLIDE_TRANSCRIPT_DEFAULT_EDGE_SECONDS = 30;
 const SLIDE_TRANSCRIPT_LEEWAY_SECONDS = 10;
 function parseTimestampSeconds(value) {
-    const parts = value.split(':').map((item) => Number(item));
+    const parts = value.split(":").map((item) => Number(item));
     if (parts.some((item) => !Number.isFinite(item)))
         return null;
     if (parts.length === 2) {
@@ -42,9 +43,9 @@ function parseTranscriptTimedText(input) {
     if (!input)
         return [];
     const segments = [];
-    for (const line of input.split('\n')) {
+    for (const line of input.split("\n")) {
         const trimmed = line.trim();
-        if (!trimmed.startsWith('['))
+        if (!trimmed.startsWith("["))
             continue;
         const match = trimmed.match(/^\[(\d{1,2}:\d{2}(?::\d{2})?)\]\s*(.*)$/);
         if (!match)
@@ -52,7 +53,7 @@ function parseTranscriptTimedText(input) {
         const seconds = parseTimestampSeconds(match[1]);
         if (seconds == null)
             continue;
-        const text = (match[2] ?? '').trim();
+        const text = (match[2] ?? "").trim();
         if (!text)
             continue;
         segments.push({ startSeconds: seconds, text });
@@ -65,20 +66,20 @@ function formatTimestamp(seconds) {
     const hours = Math.floor(clamped / 3600);
     const minutes = Math.floor((clamped % 3600) / 60);
     const secs = clamped % 60;
-    const mm = String(minutes).padStart(2, '0');
-    const ss = String(secs).padStart(2, '0');
+    const mm = String(minutes).padStart(2, "0");
+    const ss = String(secs).padStart(2, "0");
     if (hours <= 0)
         return `${minutes}:${ss}`;
-    const hh = String(hours).padStart(2, '0');
+    const hh = String(hours).padStart(2, "0");
     return `${hh}:${mm}:${ss}`;
 }
 function truncateTranscript(value, limit) {
     if (value.length <= limit)
         return value;
     const truncated = value.slice(0, limit).trimEnd();
-    const clean = truncated.replace(/\s+\S*$/, '').trim();
+    const clean = truncated.replace(/\s+\S*$/, "").trim();
     const result = clean.length > 0 ? clean : truncated.trim();
-    return result.length > 0 ? `${result}…` : '';
+    return result.length > 0 ? `${result}…` : "";
 }
 function buildSlidesPromptText({ slides, transcriptTimedText, preset, }) {
     if (!slides || slides.slides.length === 0)
@@ -114,19 +115,19 @@ function buildSlidesPromptText({ slides, transcriptTimedText, preset, }) {
                 break;
             excerptParts.push(segment.text);
         }
-        const excerptRaw = excerptParts.join(' ').trim().replace(/\s+/g, ' ');
+        const excerptRaw = excerptParts.join(" ").trim().replace(/\s+/g, " ");
         const excerptBudget = remaining > 0 ? Math.min(perSlideBudget, remaining) : 0;
-        const excerpt = excerptRaw && excerptBudget > 0 ? truncateTranscript(excerptRaw, excerptBudget) : '';
+        const excerpt = excerptRaw && excerptBudget > 0 ? truncateTranscript(excerptRaw, excerptBudget) : "";
         const label = `[slide:${slide.index}] [${formatTimestamp(start)}–${formatTimestamp(end)}]`;
         const block = excerpt ? `${label}\n${excerpt}` : label;
         blocks.push(block);
         remaining = Math.max(0, remaining - block.length);
     }
-    return blocks.length > 0 ? blocks.join('\n\n') : null;
+    return blocks.length > 0 ? blocks.join("\n\n") : null;
 }
 export function buildUrlPrompt({ extracted, outputLanguage, lengthArg, promptOverride, lengthInstruction, languageInstruction, slides, }) {
-    const isYouTube = extracted.siteName === 'YouTube';
-    const preset = lengthArg.kind === 'preset' ? lengthArg.preset : 'medium';
+    const isYouTube = extracted.siteName === "YouTube";
+    const preset = lengthArg.kind === "preset" ? lengthArg.preset : "medium";
     const slidesText = buildSlidesPromptText({
         slides,
         transcriptTimedText: extracted.transcriptTimedText,
@@ -140,10 +141,10 @@ export function buildUrlPrompt({ extracted, outputLanguage, lengthArg, promptOve
         content: extracted.content,
         truncated: extracted.truncated,
         hasTranscript: isYouTube ||
-            (extracted.transcriptSource !== null && extracted.transcriptSource !== 'unavailable'),
+            (extracted.transcriptSource !== null && extracted.transcriptSource !== "unavailable"),
         hasTranscriptTimestamps: Boolean(extracted.transcriptTimedText),
         slides: slidesText ? { count: slides?.slides.length ?? 0, text: slidesText } : null,
-        summaryLength: lengthArg.kind === 'preset' ? lengthArg.preset : { maxCharacters: lengthArg.maxCharacters },
+        summaryLength: lengthArg.kind === "preset" ? lengthArg.preset : { maxCharacters: lengthArg.maxCharacters },
         outputLanguage,
         shares: [],
         promptOverride: promptOverride ?? null,
@@ -161,7 +162,7 @@ function shouldBypassShortContentSummary({ extracted, lengthArg, forceSummary, m
         return false;
     if (extracted.content.length > targetCharacters)
         return false;
-    if (!json && typeof maxOutputTokensArg === 'number') {
+    if (!json && typeof maxOutputTokensArg === "number") {
         const tokenCount = countTokens(extracted.content);
         if (tokenCount > maxOutputTokensArg)
             return false;
@@ -176,7 +177,7 @@ async function outputSummaryFromExtractedContent({ ctx, url, extracted, extracti
         const finishReport = flags.shouldComputeReport ? await hooks.buildReport() : null;
         const payload = {
             input: {
-                kind: 'url',
+                kind: "url",
                 url,
                 timeoutMs: flags.timeoutMs,
                 youtube: flags.youtubeMode,
@@ -184,9 +185,9 @@ async function outputSummaryFromExtractedContent({ ctx, url, extracted, extracti
                 format: flags.format,
                 markdown: effectiveMarkdownMode,
                 timestamps: flags.transcriptTimestamps,
-                length: flags.lengthArg.kind === 'preset'
-                    ? { kind: 'preset', preset: flags.lengthArg.preset }
-                    : { kind: 'chars', maxCharacters: flags.lengthArg.maxCharacters },
+                length: flags.lengthArg.kind === "preset"
+                    ? { kind: "preset", preset: flags.lengthArg.preset }
+                    : { kind: "chars", maxCharacters: flags.lengthArg.maxCharacters },
                 maxOutputTokens: flags.maxOutputTokensArg,
                 model: model.requestedModelLabel,
                 language: formatOutputLanguageForJson(flags.outputLanguage),
@@ -258,17 +259,17 @@ const pickModelForFinishLine = (llmCalls, fallback) => {
         }
         return null;
     };
-    return (findLastModel('summary') ??
-        findLastModel('markdown') ??
+    return (findLastModel("summary") ??
+        findLastModel("markdown") ??
         (llmCalls.length > 0 ? (llmCalls[llmCalls.length - 1]?.model ?? null) : null) ??
         fallback);
 };
 const buildModelMetaFromAttempt = (attempt) => {
-    if (attempt.transport === 'cli') {
-        return { provider: 'cli', canonical: attempt.userModelId };
+    if (attempt.transport === "cli") {
+        return { provider: "cli", canonical: attempt.userModelId };
     }
     const parsed = parseGatewayStyleModelId(attempt.llmModelId ?? attempt.userModelId);
-    const canonical = attempt.userModelId.toLowerCase().startsWith('openrouter/')
+    const canonical = attempt.userModelId.toLowerCase().startsWith("openrouter/")
         ? attempt.userModelId
         : parsed.canonical;
     return { provider: parsed.provider, canonical };
@@ -280,14 +281,14 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
         extracted: { diagnostics: extracted.diagnostics },
         format: flags.format,
         markdownMode: effectiveMarkdownMode,
-        hasMarkdownLlmCall: model.llmCalls.some((call) => call.purpose === 'markdown'),
+        hasMarkdownLlmCall: model.llmCalls.some((call) => call.purpose === "markdown"),
     });
     const finishModel = pickModelForFinishLine(model.llmCalls, null);
     if (flags.json) {
         const finishReport = flags.shouldComputeReport ? await hooks.buildReport() : null;
         const payload = {
             input: {
-                kind: 'url',
+                kind: "url",
                 url,
                 timeoutMs: flags.timeoutMs,
                 youtube: flags.youtubeMode,
@@ -295,9 +296,9 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
                 format: flags.format,
                 markdown: effectiveMarkdownMode,
                 timestamps: flags.transcriptTimestamps,
-                length: flags.lengthArg.kind === 'preset'
-                    ? { kind: 'preset', preset: flags.lengthArg.preset }
-                    : { kind: 'chars', maxCharacters: flags.lengthArg.maxCharacters },
+                length: flags.lengthArg.kind === "preset"
+                    ? { kind: "preset", preset: flags.lengthArg.preset }
+                    : { kind: "chars", maxCharacters: flags.lengthArg.maxCharacters },
                 maxOutputTokens: flags.maxOutputTokensArg,
                 model: model.requestedModelLabel,
                 language: formatOutputLanguageForJson(flags.outputLanguage),
@@ -345,12 +346,12 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
     const extractCandidate = flags.transcriptTimestamps &&
         extracted.transcriptTimedText &&
         extracted.transcriptSource &&
-        extracted.content.toLowerCase().startsWith('transcript:')
+        extracted.content.toLowerCase().startsWith("transcript:")
         ? `Transcript:\n${extracted.transcriptTimedText}`
         : extracted.content;
     const slideTags = slides?.slides && slides.slides.length > 0
-        ? slides.slides.map((slide) => `[slide:${slide.index}]`).join('\n')
-        : '';
+        ? slides.slides.map((slide) => `[slide:${slide.index}]`).join("\n")
+        : "";
     if (slidesOutput && slides?.slides && slides.slides.length > 0) {
         const transcriptText = extracted.transcriptTimedText
             ? `Transcript:\n${extracted.transcriptTimedText}`
@@ -390,7 +391,7 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
         }
         return;
     }
-    const renderedExtract = flags.format === 'markdown' && !flags.plain && isRichTty(io.stdout)
+    const renderedExtract = flags.format === "markdown" && !flags.plain && isRichTty(io.stdout)
         ? renderMarkdownAnsi(prepareMarkdownForTerminal(extractCandidate), {
             width: markdownRenderWidth(io.stdout, io.env),
             wrap: true,
@@ -398,14 +399,14 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
             hyperlinks: true,
         })
         : extractCandidate;
-    if (flags.format === 'markdown' && !flags.plain && isRichTty(io.stdout)) {
-        io.stdout.write(`\n${renderedExtract.replace(/^\n+/, '')}`);
+    if (flags.format === "markdown" && !flags.plain && isRichTty(io.stdout)) {
+        io.stdout.write(`\n${renderedExtract.replace(/^\n+/, "")}`);
     }
     else {
         io.stdout.write(renderedExtract);
     }
-    if (!renderedExtract.endsWith('\n')) {
-        io.stdout.write('\n');
+    if (!renderedExtract.endsWith("\n")) {
+        io.stdout.write("\n");
     }
     hooks.restoreProgressAfterStdout?.();
     const slideFooter = slides ? [`slides ${slides.slides.length}`] : [];
@@ -434,9 +435,12 @@ export async function outputExtractedUrl({ ctx, url, extracted, extractionUi, pr
 }
 export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi, prompt, effectiveMarkdownMode, transcriptionCostLabel, onModelChosen, slides, slidesOutput, }) {
     const { io, flags, model, cache: cacheState, hooks } = ctx;
+    const lastSuccessfulCliProvider = model.isFallbackModel
+        ? await readLastSuccessfulCliProvider(io.envForRun)
+        : null;
     const promptPayload = { system: SUMMARY_SYSTEM_PROMPT, userText: prompt };
     const promptTokens = countTokens(promptPayload.userText);
-    const kindForAuto = extracted.siteName === 'YouTube' ? 'youtube' : 'website';
+    const kindForAuto = extracted.siteName === "YouTube" ? "youtube" : "website";
     const attempts = await (async () => {
         if (model.isFallbackModel) {
             const catalog = await model.getLiteLlmCatalog();
@@ -450,6 +454,9 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 catalog,
                 openrouterProvidersFromEnv: null,
                 cliAvailability: model.cliAvailability,
+                isImplicitAutoSelection: model.isImplicitAutoSelection,
+                allowAutoCliFallback: model.allowAutoCliFallback,
+                lastSuccessfulCliProvider,
             });
             if (flags.verbose) {
                 for (const attempt of list.slice(0, 8)) {
@@ -457,7 +464,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 }
             }
             return list.map((attempt) => {
-                if (attempt.transport !== 'cli')
+                if (attempt.transport !== "cli")
                     return model.summaryEngine.applyZaiOverrides(attempt);
                 const parsed = parseCliUserModelId(attempt.userModelId);
                 return { ...attempt, cliProvider: parsed.provider, cliModel: parsed.model };
@@ -465,12 +472,12 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
         }
         /* v8 ignore next */
         if (!model.fixedModelSpec) {
-            throw new Error('Internal error: missing fixed model spec');
+            throw new Error("Internal error: missing fixed model spec");
         }
-        if (model.fixedModelSpec.transport === 'cli') {
+        if (model.fixedModelSpec.transport === "cli") {
             return [
                 {
-                    transport: 'cli',
+                    transport: "cli",
                     userModelId: model.fixedModelSpec.userModelId,
                     llmModelId: null,
                     cliProvider: model.fixedModelSpec.cliProvider,
@@ -481,7 +488,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 },
             ];
         }
-        const openaiOverrides = model.fixedModelSpec.requiredEnv === 'Z_AI_API_KEY'
+        const openaiOverrides = model.fixedModelSpec.requiredEnv === "Z_AI_API_KEY"
             ? {
                 openaiApiKeyOverride: model.apiStatus.zaiApiKey,
                 openaiBaseUrlOverride: model.apiStatus.zaiBaseUrl,
@@ -490,7 +497,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             : {};
         return [
             {
-                transport: model.fixedModelSpec.transport === 'openrouter' ? 'openrouter' : 'native',
+                transport: model.fixedModelSpec.transport === "openrouter" ? "openrouter" : "native",
                 userModelId: model.fixedModelSpec.userModelId,
                 llmModelId: model.fixedModelSpec.llmModelId,
                 openrouterProviders: model.fixedModelSpec.openrouterProviders,
@@ -500,26 +507,29 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             },
         ];
     })();
-    const cacheStore = cacheState.mode === 'default' && !flags.summaryCacheBypass ? cacheState.store : null;
+    const cacheStore = cacheState.mode === "default" && !flags.summaryCacheBypass ? cacheState.store : null;
     const contentHash = cacheStore ? hashString(normalizeContentForHash(extracted.content)) : null;
     const promptHash = cacheStore ? buildPromptHash(prompt) : null;
     const lengthKey = buildLengthKey(flags.lengthArg);
     const languageKey = buildLanguageKey(flags.outputLanguage);
+    const autoSelectionCacheModel = model.isFallbackModel
+        ? `selection:${model.requestedModelInput.toLowerCase()}`
+        : null;
     let summaryResult = null;
     let usedAttempt = null;
     let summaryFromCache = false;
     let cacheChecked = false;
-    const isTweet = extracted.siteName?.toLowerCase() === 'x' || isTwitterStatusUrl(extracted.url);
-    const isYouTube = extracted.siteName === 'YouTube' || isYouTubeUrl(url);
+    const isTweet = extracted.siteName?.toLowerCase() === "x" || isTwitterStatusUrl(extracted.url);
+    const isYouTube = extracted.siteName === "YouTube" || isYouTubeUrl(url);
     const hasMedia = Boolean(extracted.video) ||
-        (extracted.transcriptSource != null && extracted.transcriptSource !== 'unavailable') ||
-        (typeof extracted.mediaDurationSeconds === 'number' && extracted.mediaDurationSeconds > 0) ||
+        (extracted.transcriptSource != null && extracted.transcriptSource !== "unavailable") ||
+        (typeof extracted.mediaDurationSeconds === "number" && extracted.mediaDurationSeconds > 0) ||
         extracted.isVideoOnly === true;
     const autoBypass = ctx.model.isFallbackModel && !ctx.model.isNamedModelSelection;
     const canBypassShortContent = (autoBypass || isTweet) &&
         !flags.slides &&
         !hasMedia &&
-        flags.streamMode !== 'on' &&
+        flags.streamMode !== "on" &&
         !isYouTube &&
         shouldBypassShortContentSummary({
             extracted,
@@ -538,41 +548,78 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             effectiveMarkdownMode,
             transcriptionCostLabel,
             slides,
-            footerLabel: 'short content',
-            verboseMessage: 'short content: skipping summary',
+            footerLabel: "short content",
+            verboseMessage: "short content: skipping summary",
         });
         return;
     }
     if (cacheStore && contentHash && promptHash) {
         cacheChecked = true;
-        for (const attempt of attempts) {
-            if (!model.summaryEngine.envHasKeyFor(attempt.requiredEnv))
-                continue;
+        if (autoSelectionCacheModel) {
             const key = buildSummaryCacheKey({
                 contentHash,
                 promptHash,
-                model: attempt.userModelId,
+                model: autoSelectionCacheModel,
                 lengthKey,
                 languageKey,
             });
-            const cached = cacheStore.getText('summary', key);
-            if (!cached)
-                continue;
-            writeVerbose(io.stderr, flags.verbose, 'cache hit summary', flags.verboseColor, io.envForRun);
-            onModelChosen?.(attempt.userModelId);
-            summaryResult = {
-                summary: cached,
-                summaryAlreadyPrinted: false,
-                modelMeta: buildModelMetaFromAttempt(attempt),
-                maxOutputTokensForCall: null,
-            };
-            usedAttempt = attempt;
-            summaryFromCache = true;
-            break;
+            const cached = cacheStore.getJson("summary", key);
+            const cachedSummary = cached && typeof cached.summary === "string" ? cached.summary.trim() : null;
+            const cachedModelId = cached && typeof cached.model === "string" ? cached.model.trim() : null;
+            if (cachedSummary) {
+                const cachedAttempt = cachedModelId
+                    ? (attempts.find((attempt) => attempt.userModelId === cachedModelId) ?? null)
+                    : null;
+                const fallbackAttempt = attempts.find((attempt) => model.summaryEngine.envHasKeyFor(attempt.requiredEnv)) ??
+                    attempts[0] ??
+                    null;
+                const matchedAttempt = cachedAttempt && model.summaryEngine.envHasKeyFor(cachedAttempt.requiredEnv)
+                    ? cachedAttempt
+                    : fallbackAttempt;
+                if (matchedAttempt) {
+                    writeVerbose(io.stderr, flags.verbose, "cache hit summary (auto selection)", flags.verboseColor, io.envForRun);
+                    onModelChosen?.(cachedModelId || matchedAttempt.userModelId);
+                    summaryResult = {
+                        summary: cachedSummary,
+                        summaryAlreadyPrinted: false,
+                        modelMeta: buildModelMetaFromAttempt(matchedAttempt),
+                        maxOutputTokensForCall: null,
+                    };
+                    usedAttempt = matchedAttempt;
+                    summaryFromCache = true;
+                }
+            }
+        }
+        if (!summaryFromCache) {
+            for (const attempt of attempts) {
+                if (!model.summaryEngine.envHasKeyFor(attempt.requiredEnv))
+                    continue;
+                const key = buildSummaryCacheKey({
+                    contentHash,
+                    promptHash,
+                    model: attempt.userModelId,
+                    lengthKey,
+                    languageKey,
+                });
+                const cached = cacheStore.getText("summary", key);
+                if (!cached)
+                    continue;
+                writeVerbose(io.stderr, flags.verbose, "cache hit summary", flags.verboseColor, io.envForRun);
+                onModelChosen?.(attempt.userModelId);
+                summaryResult = {
+                    summary: cached,
+                    summaryAlreadyPrinted: false,
+                    modelMeta: buildModelMetaFromAttempt(attempt),
+                    maxOutputTokensForCall: null,
+                };
+                usedAttempt = attempt;
+                summaryFromCache = true;
+                break;
+            }
         }
     }
     if (cacheChecked && !summaryFromCache) {
-        writeVerbose(io.stderr, flags.verbose, 'cache miss summary', flags.verboseColor, io.envForRun);
+        writeVerbose(io.stderr, flags.verbose, "cache miss summary", flags.verboseColor, io.envForRun);
     }
     ctx.hooks.onSummaryCached?.(summaryFromCache);
     let lastError = null;
@@ -618,7 +665,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
         };
         if (model.isNamedModelSelection) {
             if (lastError === null && missingRequiredEnvs.size > 0) {
-                throw new Error(withFreeTip(`Missing ${Array.from(missingRequiredEnvs).sort().join(', ')} for --model ${model.requestedModelInput}.`));
+                throw new Error(withFreeTip(`Missing ${Array.from(missingRequiredEnvs).sort().join(", ")} for --model ${model.requestedModelInput}.`));
             }
             if (lastError instanceof Error) {
                 if (sawOpenRouterNoAllowedProviders) {
@@ -642,21 +689,41 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             effectiveMarkdownMode,
             transcriptionCostLabel,
             slides,
-            footerLabel: 'no model',
+            footerLabel: "no model",
             verboseMessage: lastError instanceof Error ? `auto failed all models: ${lastError.message}` : null,
         });
         return;
     }
     if (!summaryFromCache && cacheStore && contentHash && promptHash) {
-        const key = buildSummaryCacheKey({
+        const perModelKey = buildSummaryCacheKey({
             contentHash,
             promptHash,
             model: usedAttempt.userModelId,
             lengthKey,
             languageKey,
         });
-        cacheStore.setText('summary', key, summaryResult.summary, cacheState.ttlMs);
-        writeVerbose(io.stderr, flags.verbose, 'cache write summary', flags.verboseColor, io.envForRun);
+        cacheStore.setText("summary", perModelKey, summaryResult.summary, cacheState.ttlMs);
+        writeVerbose(io.stderr, flags.verbose, "cache write summary", flags.verboseColor, io.envForRun);
+        if (autoSelectionCacheModel) {
+            const selectionKey = buildSummaryCacheKey({
+                contentHash,
+                promptHash,
+                model: autoSelectionCacheModel,
+                lengthKey,
+                languageKey,
+            });
+            cacheStore.setJson("summary", selectionKey, { summary: summaryResult.summary, model: usedAttempt.userModelId }, cacheState.ttlMs);
+            writeVerbose(io.stderr, flags.verbose, "cache write summary (auto selection)", flags.verboseColor, io.envForRun);
+        }
+    }
+    if (!summaryFromCache &&
+        model.isFallbackModel &&
+        usedAttempt.transport === "cli" &&
+        usedAttempt.cliProvider) {
+        await writeLastSuccessfulCliProvider({
+            env: io.envForRun,
+            provider: usedAttempt.cliProvider,
+        });
     }
     const { summary, summaryAlreadyPrinted, modelMeta, maxOutputTokensForCall } = summaryResult;
     const normalizedSummary = slides && slides.slides.length > 0 ? normalizeSummarySlideHeadings(summary) : summary;
@@ -664,7 +731,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
         const finishReport = flags.shouldComputeReport ? await hooks.buildReport() : null;
         const payload = {
             input: {
-                kind: 'url',
+                kind: "url",
                 url,
                 timeoutMs: flags.timeoutMs,
                 youtube: flags.youtubeMode,
@@ -672,9 +739,9 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 format: flags.format,
                 markdown: effectiveMarkdownMode,
                 timestamps: flags.transcriptTimestamps,
-                length: flags.lengthArg.kind === 'preset'
-                    ? { kind: 'preset', preset: flags.lengthArg.preset }
-                    : { kind: 'chars', maxCharacters: flags.lengthArg.maxCharacters },
+                length: flags.lengthArg.kind === "preset"
+                    ? { kind: "preset", preset: flags.lengthArg.preset }
+                    : { kind: "chars", maxCharacters: flags.lengthArg.maxCharacters },
                 maxOutputTokens: flags.maxOutputTokensArg,
                 model: model.requestedModelLabel,
                 language: formatOutputLanguageForJson(flags.outputLanguage),
@@ -695,7 +762,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 provider: modelMeta.provider,
                 model: usedAttempt.userModelId,
                 maxCompletionTokens: maxOutputTokensForCall,
-                strategy: 'single',
+                strategy: "single",
             },
             metrics: flags.metricsEnabled ? finishReport : null,
             summary: normalizedSummary,
@@ -707,7 +774,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
                 stderr: io.stderr,
                 env: io.envForRun,
                 elapsedMs: Date.now() - flags.runStartedAtMs,
-                elapsedLabel: summaryFromCache ? 'Cached' : null,
+                elapsedLabel: summaryFromCache ? "Cached" : null,
                 label: extractionUi.finishSourceLabel,
                 model: usedAttempt.userModelId,
                 report: finishReport,
@@ -750,15 +817,15 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             })
             : normalizedSummary;
         if (!flags.plain && isRichTty(io.stdout)) {
-            io.stdout.write(`\n${rendered.replace(/^\n+/, '')}`);
+            io.stdout.write(`\n${rendered.replace(/^\n+/, "")}`);
         }
         else {
             if (isRichTty(io.stdout))
-                io.stdout.write('\n');
-            io.stdout.write(rendered.replace(/^\n+/, ''));
+                io.stdout.write("\n");
+            io.stdout.write(rendered.replace(/^\n+/, ""));
         }
-        if (!rendered.endsWith('\n')) {
-            io.stdout.write('\n');
+        if (!rendered.endsWith("\n")) {
+            io.stdout.write("\n");
         }
         hooks.restoreProgressAfterStdout?.();
     }
@@ -769,7 +836,7 @@ export async function summarizeExtractedUrl({ ctx, url, extracted, extractionUi,
             stderr: io.stderr,
             env: io.envForRun,
             elapsedMs: Date.now() - flags.runStartedAtMs,
-            elapsedLabel: summaryFromCache ? 'Cached' : null,
+            elapsedLabel: summaryFromCache ? "Cached" : null,
             label: extractionUi.finishSourceLabel,
             model: modelMeta.canonical,
             report,