@bli-cockpit/cli 0.2.89 → 0.2.91

package/dist/commands/local-args-tower-models.js

@@ -4,6 +4,7 @@
  *   cockpit models                       # every card: documented vs measured
  *   cockpit models list --json
  *   cockpit models show openai:gpt-5.6-terra
+ *   cockpit models compare openai:gpt-5.6-luna openai:gpt-5.6-terra --highlight
  *
  * Two nouns that look alike and are not: `cockpit model` is the ORG SETTING for
  * which model a slot runs on (`model show`, `model set`); `cockpit models` is
@@ -11,20 +12,47 @@
  * quirks, probe verdicts and stale marks. The help text for each says so.
  */
 import { optionalNonEmpty, optionalUrl, parseNamedArgs } from "./local-arg-values.js";
+/** Two is not enough to compare against, five stops being readable (BLI-3919). */
+const COMPARE_MIN = 2;
+const COMPARE_MAX = 4;
 export function parseModelsArgs(args) {
     const values = parseNamedArgs(args, {
-        allowedFlags: ["--home", "--dashboard-url", "--json"],
+        allowedFlags: ["--home", "--dashboard-url", "--json", "--highlight"],
         valueFlags: ["--home", "--dashboard-url"],
     });
-    if (values.positionals.length > 2) {
-        throw new Error("models accepts at most an action (list|show) and one model id.");
-    }
-    const [rawAction, rawId] = values.positionals;
     const base = {
         homeDir: optionalNonEmpty(values.flags.get("--home")),
         dashboardUrl: optionalUrl(values.flags.get("--dashboard-url")),
         json: values.booleans.has("--json"),
     };
+    // `compare` is the one verb that takes several ids, so it is parsed before
+    // the at-most-two rule the other two live under (BLI-3919).
+    if (values.positionals[0] === "compare") {
+        const modelIds = values.positionals.slice(1).map((id) => id.trim()).filter((id) => id.length > 0);
+        const unique = [...new Set(modelIds)];
+        if (unique.length < COMPARE_MIN) {
+            throw new Error(`models compare needs at least ${COMPARE_MIN} model ids, e.g. ` +
+                "`models compare openai:gpt-5.6-luna openai:gpt-5.6-terra`.");
+        }
+        if (unique.length > COMPARE_MAX) {
+            throw new Error(`models compare holds at most ${COMPARE_MAX} models; ${unique.length} named. ` +
+                "A fifth column stops being readable.");
+        }
+        return {
+            kind: "models",
+            action: "compare",
+            modelIds: unique,
+            highlight: values.booleans.has("--highlight"),
+            ...base,
+        };
+    }
+    if (values.booleans.has("--highlight")) {
+        throw new Error("--highlight only means something for `models compare`.");
+    }
+    if (values.positionals.length > 2) {
+        throw new Error("models accepts at most an action (list|show) and one model id.");
+    }
+    const [rawAction, rawId] = values.positionals;
     if (!rawAction || rawAction === "list") {
         if (rawAction === "list" && rawId) {
             throw new Error("models list takes no model id — use `models show <provider:model>`.");

package/dist/commands/local-help-commands.js

@@ -276,7 +276,7 @@ export function localSubcommandHelp(command) {
         [
             "models",
             [
-                "Usage: cockpit models [list|show <provider:model>] [--dashboard-url <url>] [--json]",
+                "Usage: cockpit models [list|show <provider:model>|compare <id> <id> [...]] [--highlight] [--dashboard-url <url>] [--json]",
                 "",
                 "The model-card shelf: one card per model, with what the PROVIDER documents (context",
                 "window, max output, price per million, reasoning modes, tool calling, structured output,",
@@ -287,6 +287,15 @@ export function localSubcommandHelp(command) {
                 "STALE:<field> means a measurement contradicted the documentation — reasoning-off refused",
                 "where the card says it can be turned off, the belt refused where the card says tool",
                 "calling. The line under it is the command that re-reads the provider.",
+                "compare puts two to four cards side by side: the same rows the page at /settings/models/",
+                "compare renders, including `$ per 100 JARVIS turns` (our own measured tokens a turn times",
+                "the provider's price, with the window it was measured over beside it — `not measured`",
+                "where no profile exists, never a per-million price relabelled) and the bench row, which",
+                "appears only when every model was judged in the SAME run against the SAME baseline;",
+                "otherwise it says `benched separately — not a head-to-head` rather than subtracting two",
+                "win rates. --highlight marks the winner in each scored row (lowest wins on price and",
+                "latency, highest on tokens/s, most wins on the bench); a row where any column is `—` or",
+                "`not measured` is not scored at all, and says why.",
                 "`cockpit models` is the SHELF; `cockpit model` is the org setting saying which model a",
                 "slot runs on. Different nouns, one letter apart, on purpose.",
                 "--json writes one object to stdout; operational lines stay on stderr.",

package/dist/commands/local-help.js

@@ -71,7 +71,7 @@ export function localCommandHelp(command) {
         "  cockpit analyze [--workspace <path>] [--dashboard-url <url>] [--max-depth <n>] [--max-repos <n>] [--json]",
         "  cockpit jarvis [question] [--prompt <question>] [--as <person>] [--date <YYYY-MM-DD>] [--thread <name>] [--model <key>] [--image <path>|--file <path>] [--no-stream] [--threads|--history [--limit <n>]] [--trace <id|last>] [--dashboard-url <url>] [--json [--show-approval-code]]",
         "  cockpit model [show|set <provider:model>] [--json]",
-        "  cockpit models [list|show <provider:model>] [--dashboard-url <url>] [--json]",
+        "  cockpit models [list|show <provider:model>|compare <provider:model> <provider:model> [...]] [--highlight] [--dashboard-url <url>] [--json]",
         "  cockpit scout [start|dismiss|undo <experiment-id>] [--days <n>] [--dashboard-url <url>] [--json]",
         "  cockpit ops [status [--job <id>] [--skips] [--memory] [--models] | recompile --person <email|name|id> [--dry-run]] [--dashboard-url <url>] [--json]",
         "  cockpit slack [coverage [--workspace bli|blue_pearl] [--stale-only] | read [--person <p>] [--channel <c>] [--query <text>] [--since <YYYY-MM-DD>] [--until <YYYY-MM-DD>] [--limit <n>]] [--json]",

package/dist/commands/memory-hook-counts.js

@@ -27,7 +27,7 @@
  * which is the exact false green this ticket exists to remove.
  */
 import fs from "node:fs";
-import { memoryHookStatsFilePath, parseMemoryHookStats, summariseMemoryHookWindow, } from "@bli-cockpit/telemetry-core";
+import { memoryHookCharsPercentile, memoryHookStatsFilePath, parseMemoryHookStats, summariseMemoryHookWindow, } from "@bli-cockpit/telemetry-core";
 export function readMemoryHookCounts(options) {
     const readText = options.readText ?? defaultReadText;
     const file = memoryHookStatsFilePath(options.homeDir);
@@ -55,6 +55,17 @@ export function readMemoryHookCounts(options) {
                 : {}),
             hook_skipped_trivial_24h: window.skippedTrivial,
             hook_billing_exhausted_24h: window.billingExhausted,
+            // Ticket 3934. Reported as the bin's LOWER bound — the smaller, true
+            // claim — and only when something was measured. A machine that recorded
+            // no histogram carries no key at all, which the board reads as
+            // "not recorded"; a zero here would say its recalls were empty.
+            ...(window.charsMeasured && window.charsSamples > 0
+                ? {
+                    hook_chars_p50: memoryHookCharsPercentile(window.chars, 50)?.lower ?? 0,
+                    hook_chars_p95: memoryHookCharsPercentile(window.chars, 95)?.lower ?? 0,
+                    hook_chars_samples: window.charsSamples,
+                }
+                : {}),
         },
         reason: "ok",
     };

package/dist/commands/models.js

@@ -6,11 +6,24 @@
  * tokens/s and last probe verdict, and a `STALE:<field>` mark wherever the two
  * disagree — with the command that fixes it on the next line.
  *
+ * `models show <id>` asks the same door for the SIX SECTIONS (BLI-3918) —
+ * what's going on, what's wrong, what's right, what should we do, double down,
+ * forget about it — so a terminal reads the model page a browser reads, in the
+ * same order, including today's spend and the `stale:<field>` markers.
+ *
+ * `models compare <id> <id> [...]` asks it for TWO TO FOUR cards side by side
+ * (BLI-3919): the same rows, the same `$ per 100 JARVIS turns` and the same
+ * bench head-to-head the page at `/settings/models/compare` renders, computed
+ * by one pure builder on the server. `--highlight` marks the best cell in each
+ * scored row; it is off by default, exactly as the page's toggle is.
+ *
  * **Every line comes from the server** (`/api/models/cards`, rendered by
- * `lib/models/model-lines.ts`). This file decides an exit code and nothing
- * else. A terminal that formatted a card itself would eventually disagree with
- * the board about what "stale" means, and the one that is wrong is always the
- * one somebody is reading.
+ * `lib/models/model-lines.ts` and `card-section-lines.ts`). This file decides an
+ * exit code and nothing else. A terminal that formatted a card itself would
+ * eventually disagree with the board about what "stale" means, and the one that
+ * is wrong is always the one somebody is reading. A server that has not shipped
+ * the sections yet sends no `section_lines`, and the whole-card `lines` are
+ * printed instead — never an empty screen.
  *
  * **Exit codes.** Zero when the shelf was read, whatever it says — a stale card
  * is a fact to act on, not a broken command. One when the read failed or when a
@@ -21,13 +34,29 @@ import { writeLine } from "./cli-io.js";
 import { loadPairedSession, towerJsonRequest, } from "../tower-client.js";
 /** A card read is checked-in data on the server side; it should never hang a shell. */
 const READ_DEADLINE_MS = 30_000;
+/**
+ * Which question this invocation asks the door.
+ *
+ * `compare` sends `--highlight` along, because the winner is decided ON THE
+ * SERVER and each surface only decides whether to SHOW it — a terminal that
+ * scored the rows itself would eventually disagree with the page about which
+ * model is cheaper.
+ */
+function requestPath(command) {
+    if (command.action === "compare") {
+        const ids = (command.modelIds ?? []).join(",");
+        return `/api/models/cards?compare=${encodeURIComponent(ids)}${command.highlight ? "&highlight=1" : ""}`;
+    }
+    if (command.action === "show" && command.modelId) {
+        return `/api/models/cards?id=${encodeURIComponent(command.modelId)}&sections=1`;
+    }
+    return "/api/models/cards";
+}
 export async function runModels(command, io) {
     const session = await loadPairedSession("models", command.homeDir);
     const dashboardUrl = command.dashboardUrl ?? session.dashboard_url;
     const log = (line) => writeLine(io.stderr, line);
-    const path = command.action === "show" && command.modelId
-        ? `/api/models/cards?id=${encodeURIComponent(command.modelId)}`
-        : "/api/models/cards";
+    const path = requestPath(command);
     const result = await towerJsonRequest({
         dashboardUrl,
         path,
@@ -47,7 +76,14 @@ export async function runModels(command, io) {
         writeLine(io.stdout, JSON.stringify({ ok: true, ...payload }));
     }
     else {
-        for (const line of payload.lines ?? [])
+        // A server that predates a branch sends no lines for it; the next-best
+        // rendering is printed rather than an empty screen.
+        const lines = payload.compare_lines?.length
+            ? payload.compare_lines
+            : payload.section_lines?.length
+                ? payload.section_lines
+                : (payload.lines ?? []);
+        for (const line of lines)
             writeLine(io.stdout, line);
     }
     // Never silent about a card the server could not read: a shelf quietly one
@@ -58,7 +94,10 @@ export async function runModels(command, io) {
     writeLine(io.stderr, `[models cli] read ${JSON.stringify({
         action: command.action,
         model: command.modelId ?? null,
+        models: command.modelIds?.length ?? 0,
         cards: Array.isArray(payload.cards) ? payload.cards.length : payload.card ? 1 : 0,
+        sections: payload.section_lines?.length ? "server" : "absent",
+        compare: payload.compare_lines?.length ? "server" : "absent",
         unreadable: payload.unreadable?.length ?? 0,
     })}`);
     return 0;
@@ -82,5 +121,6 @@ function writeFailure(command, io, failure) {
         reason: failure.reason,
         http_status: failure.httpStatus ?? null,
         model: command.modelId ?? null,
+        models: command.modelIds?.length ?? 0,
     })}`);
 }

package/dist/commands/ops-render-memory.js

@@ -110,6 +110,11 @@ export function renderMemoryHooks(hooks, dim) {
         // printed verbatim, like every other sentence in this section.
         if (device.viaLine)
             lines.push(dim(`    ${device.viaLine}`));
+        // Ticket 3934: how much of a person's context the recall took. Written on
+        // the server (`lib/ops/memory-hook-misses.ts`) and printed verbatim, like
+        // every other sentence in this section.
+        if (device.charsLine)
+            lines.push(dim(`    ${device.charsLine}`));
     }
     return lines;
 }

package/dist/commands/ops-render-spend.js

@@ -29,8 +29,21 @@ function usdColumn(usd) {
  */
 export function renderSpend(spend, dim) {
     const lines = ["", `SPEND  ${spend.summary ?? "(no summary)"}`];
+    // The subscription question, on its own line and in the server's own words
+    // (ticket 3934). Printed BEFORE the per-model rows because it is the number
+    // somebody acts on: the rows say where the money went, this says whether it
+    // had to be spent at all.
+    if (spend.offsettableLine)
+        lines.push(dim(`  ${spend.offsettableLine}`));
     for (const row of spend.models ?? []) {
-        lines.push(dim(`  ${usdColumn(row.usd).padStart(9)}  ${row.model ?? "?"} · ${row.calls ?? 0} call(s) · ${(row.bases ?? []).join("+") || "?"} · ${(row.sources ?? []).join(",") || "?"}`));
+        // Two dollar columns: what the provider LISTS and what the account was
+        // BILLED. They are equal on Fireworks and OpenRouter and differ by the
+        // credit divisor on OpenAI, so printing one alone was three times the bill
+        // for every OpenAI row and said nothing about which figure it was.
+        const real = typeof row.realUsd === "number" && row.realUsd !== row.usd
+            ? ` (${usdColumn(row.realUsd)} real)`
+            : "";
+        lines.push(dim(`  ${usdColumn(row.usd).padStart(9)}${real}  ${row.model ?? "?"} · ${row.calls ?? 0} call(s) · ${row.purpose ?? "?"} · ${(row.bases ?? []).join("+") || "?"} · ${(row.sources ?? []).join(",") || "?"}`));
     }
     return lines;
 }

package/dist/commands/public-root.js

@@ -15,7 +15,7 @@ export async function runCockpitCli(argv, io) {
   }
 
   if (command === "--version" || command === "-V" || command === "version") {
-    writeLine(io?.stdout ?? process.stdout, "0.2.89");
+    writeLine(io?.stdout ?? process.stdout, "0.2.91");
     return 0;
   }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bli-cockpit/cli",
-  "version": "0.2.89",
+  "version": "0.2.91",
   "private": false,
   "type": "module",
   "bin": {
@@ -27,8 +27,8 @@
     "test": "node dist/cli.js --help && node ../../scripts/assert-public-cli-routing.mjs && node ../../scripts/assert-public-cli-verb-help.mjs && node ../../scripts/assert-public-cli-runtime-files.mjs && node ../../scripts/assert-public-cli-no-fleet-posts.mjs && node ../../scripts/assert-public-package-pack.mjs --workspace=@bli-cockpit/cli"
   },
   "dependencies": {
-    "@bli-cockpit/memory-mcp": "0.1.20",
-    "@bli-cockpit/mcp": "0.1.22",
-    "@bli-cockpit/telemetry-core": "0.1.37"
+    "@bli-cockpit/memory-mcp": "0.1.21",
+    "@bli-cockpit/mcp": "0.1.24",
+    "@bli-cockpit/telemetry-core": "0.1.38"
   }
 }