@trazum/cli 1.26.0 → 1.28.0

package/dist/index.js

@@ -2,7 +2,7 @@
 import { readdir, readFile, stat, writeFile } from 'node:fs/promises';
 import { join, resolve as resolvePath } from 'node:path';
 import { gunzipSync } from 'node:zlib';
-import { applyRewrites, BASELINE_FILENAME, BASELINE_VERSION, breaches, cacheableMinimum, analyzeCachePrefix, billLevers, cacheEconomics, cacheHitRate, comparePrompts, compareToBaseline, computeSavings, countTokensAnthropic, DEFAULT_USAGE, detectFromSource, driversBetween, estimateTokens, evaluate, extractPrompts, findExamples, formatBaseline, formatSignedUsd, formatUsd, getMessages, getModel, hasMarker, LOCALES, MAX_BASELINE_BYTES, moneyIsComparable, mostSpecificMatch, nearestName, optimize, parseBaseline, PHRASE_LANGUAGES, plannedCalls, profilePrompt, profileToCsv, profileUsage, promptId, providerFromEnv, pruneExamples, refineWithLlm, rejectionText, reorderForCache, repriceProfile, reviewAgeDays, reviewExamples, RULES, sharedPrefixes, sharesOf, SOURCE_EXTENSIONS, suggestRewrites, toOtlpMetrics, toPromptfoo, TTL_1H_MS, UNLABELLED, withExactTokenCounts, } from '@trazum/core';
+import { applyRewrites, BASELINE_FILENAME, BASELINE_VERSION, breaches, cacheableMinimum, analyzeCachePrefix, billLevers, cacheEconomics, cacheHitRate, contextPressure, comparePrompts, compareToBaseline, computeSavings, countTokensAnthropic, DEFAULT_USAGE, detectFromSource, driversBetween, estimateTokens, evaluate, extractPrompts, findExamples, formatBaseline, formatSignedUsd, formatUsd, getMessages, getModel, hasMarker, LOCALES, MAX_BASELINE_BYTES, moneyIsComparable, mostSpecificMatch, nearestName, optimize, parseBaseline, PHRASE_LANGUAGES, plannedCalls, profilePrompt, profileToCsv, profileUsage, promptId, providerFromEnv, pruneExamples, refineWithLlm, rejectionText, reorderForCache, repriceProfile, reviewAgeDays, reviewExamples, RULES, sharedPrefixes, sharesOf, SOURCE_EXTENSIONS, suggestRewrites, toOtlpMetrics, toPromptfoo, TTL_1H_MS, UNLABELLED, withExactTokenCounts, } from '@trazum/core';
 import { cacheDir, cacheStats, cachingProvider, clearCache } from './suggest-cache.js';
 import { dayOf, formatGap, median, spanDays } from './time.js';
 // Everything that reads the filesystem, on its own entry point so the web
@@ -1660,6 +1660,12 @@ async function commandProfile(args, config, pricing, t) {
      * got a report with no answer in it, and has no way to tell a typo from a
      * model this comparison had nothing to say about.
      */
+    /**
+     * The largest call against each model's window, computed once here so the
+     * terminal, the JSON and the markdown state the same ratio — the
+     * denominator lives in the (possibly overlaid) catalogue.
+     */
+    const pressures = contextPressure(report, pricing);
     const whatIfModel = stringFlag(args, 'what-if');
     const whatIf = whatIfModel !== undefined ? repriceProfile(report, whatIfModel, pricing) : null;
     if (whatIfModel !== undefined && whatIf === null) {
@@ -1832,8 +1838,12 @@ async function commandProfile(args, config, pricing, t) {
          * sound in both directions, the pass is a floor, and the pass message
          * says so when the span does not start and end on a day boundary.
          */
-        if (typeof args.flags.get('max-day-usd') === 'string') {
-            const maxDay = numberFlag(args, 'max-day-usd', 0, t);
+        if (typeof args.flags.get('max-day-usd') === 'string' || config.spend?.maxDayUsd !== undefined) {
+            // The flag beats the config, like every gate here: the config is the
+            // repository's standing policy, the flag is this invocation's word.
+            const maxDay = typeof args.flags.get('max-day-usd') === 'string'
+                ? numberFlag(args, 'max-day-usd', 0, t)
+                : config.spend.maxDayUsd;
             if (report.spendByDay.length === 0) {
                 console.error(c.red(t.profile.maxDayNoClock()));
                 process.exitCode = 1;
@@ -1899,6 +1909,7 @@ async function commandProfile(args, config, pricing, t) {
                 // handed over, so the summary in a pull request cannot disagree
                 // with the terminal about what a move would cost.
                 ...(whatIf !== null ? { whatIf } : {}),
+                pressure: pressures,
                 // The comparison, when there was one — the same figures and the same
                 // drivers the terminal printed, never re-derived here.
                 ...(previous !== null
@@ -1937,7 +1948,7 @@ async function commandProfile(args, config, pricing, t) {
              * somebody sums wrong.
              */
             const shape = stringFlag(args, 'csv-shape') ?? 'slice';
-            if (shape !== 'slice' && shape !== 'day' && shape !== 'hour') {
+            if (shape !== 'slice' && shape !== 'day' && shape !== 'hour' && shape !== 'model-day') {
                 throw new Error(t.profile.badCsvShape(shape));
             }
             await writeFile(csvOut, profileToCsv(report, { unlabelled: t.profile.unlabelled(), shape }), 'utf8');
@@ -1994,6 +2005,10 @@ async function commandProfile(args, config, pricing, t) {
                     },
                 }
                 : {}),
+            // How close each slice's largest call is to its model's window —
+            // derived like `cache` and `levers`, so a dashboard need not
+            // re-derive a ratio whose denominator lives in the catalogue.
+            contextPressure: pressures,
             // Present only when --what-if was passed. `sameTokensAssumed` rides
             // along inside it so a consumer cannot print the dollar figure
             // without the caveat being in the same object.
@@ -2697,6 +2712,52 @@ async function commandProfile(args, config, pricing, t) {
             }
         }
     }
+    /**
+     * How close the largest call is to the model's ceiling.
+     *
+     * The failure a bill cannot show: input grows turn by turn or document by
+     * document, costs nothing extra to grow, and then one call crosses the
+     * context window and the API refuses it. Loud from 85% — close enough
+     * that the next retrieval bump or long conversation plausibly crosses —
+     * and quiet above half, so the reader sees it coming either way. No
+     * prediction of *when*: a straight line through two points is a guess
+     * wearing arithmetic's clothes.
+     */
+    {
+        if (pressures.length > 0) {
+            console.log();
+            console.log(c.bold(t.profile.pressureHeading()));
+            for (const row of pressures.slice(0, 3)) {
+                const label = row.label === UNLABELLED ? t.profile.unlabelled() : row.label;
+                const line = t.profile.pressureLine(label, row.modelName, n(row.maxCallInputTokens), n(row.contextWindow), pct(row.share));
+                console.log();
+                if (row.share >= 0.85) {
+                    console.log(`  ${c.yellow('!')} ${c.bold(wrap(line, 74, '    '))}`);
+                    console.log(`  ${c.dim(wrap(t.profile.pressureAdvice(), 74, '  '))}`);
+                }
+                else {
+                    console.log(`  ${c.dim(wrap(line, 74, '  '))}`);
+                }
+            }
+        }
+    }
+    /**
+     * The mix moving inside one log — the drift `--against` needs a second log
+     * to see. Spoken only past fifteen points, a presentation threshold stated
+     * in the copy; the data states exact shares and the JSON carries them all.
+     */
+    if (report.modelMixDrift !== null) {
+        const moved = report.modelMixDrift.models.filter((m) => Math.abs(m.lastShare - m.firstShare) >= 0.15);
+        if (moved.length > 0) {
+            console.log();
+            console.log(c.bold(t.profile.mixDriftHeading()));
+            for (const m of moved.slice(0, 3)) {
+                console.log();
+                console.log(`  ${c.yellow('!')} ${c.bold(wrap(t.profile.mixDriftLine(m.model, pct(m.firstShare), pct(m.lastShare), n(report.modelMixDrift.firstDays), n(report.modelMixDrift.lastDays), formatUsd(m.lastUsd)), 74, '    '))}`);
+            }
+            console.log(`  ${c.dim(wrap(t.profile.mixDriftNote(), 74, '  '))}`);
+        }
+    }
     /**
      * The same request, sent again a moment later.
      *
@@ -2761,6 +2822,20 @@ async function commandProfile(args, config, pricing, t) {
         if (ceiling !== undefined) {
             console.log(`    ${c.dim(wrap(t.profile.truncatedCeiling(n(ceiling.p95WithinTokens)), 74, '    '))}`);
         }
+        /**
+         * The "billed again" half, measured. The sentence above has always said
+         * truncated answers are frequently retried; this is the count and the
+         * money, when the log can carry it — a pattern, never a certainty, and
+         * attributed to the truncated call's slice, where the ceiling that
+         * caused it lives.
+         */
+        for (const row of report.truncationRetries.slice(0, 3)) {
+            const name = row.label === UNLABELLED ? t.profile.unlabelled() : row.label;
+            console.log(`  ${c.yellow('!')} ${c.bold(wrap(t.profile.truncationRetryLine(name, row.modelName, n(row.retried), n(row.truncatedCalls), n(Math.round(row.withinMs / 1000)), formatUsd(row.wastedUsd), formatUsd(row.retryUsd)), 74, '    '))}`);
+        }
+        if (report.truncationRetries.length > 0) {
+            console.log(`  ${c.dim(wrap(t.profile.truncationRetryNote(), 74, '  '))}`);
+        }
     }
     else if (report.total.stopReasonCalls === 0) {
         console.log();