ollanet 0.5.0 → 0.6.1

package/README.md

@@ -32,11 +32,12 @@ They share the same Ollama API; Finetuna shapes the models, ollanet finds and ch
 - **Discover** Ollama hosts from localhost, `config.hosts`, `OLLANET_HOSTS`, optional Tailscale, and optional LAN scan (`--lan`, off by default)
 - **Pull** a library model onto a named host (`ollanet pull studio gemma3:12b`) — the server downloads; re-pull updates
 - **Show / rm / ps** — inspect a Modelfile, delete a model (`--yes`), see what’s in VRAM. Scan marks Finetuna-style `[tuned]` names
-- **Prompt** any hostname/IP/model with streaming replies
+- **Prompt** any hostname/IP/model with streaming replies (`--file` for `.txt` / `.md`)
+- **Compare** 2–5 models on one host with the same prompt; summary table + `compares/<id>.md`
 - **Persist** chats as `responses/<hash>.json` with topic, machine, model, timestamps
 - **Continue** any thread with `--chat <hash>`
 - **Configure** per-machine defaults (model, temperature, context, …)
-- **MCP** — `ollanet mcp` exposes scan/prompt/pull/show/rm/ps/chats as Model Context Protocol tools for agents
+- **MCP** — `ollanet mcp` exposes scan/prompt/compare/pull/show/rm/ps/chats as Model Context Protocol tools for agents
 - **Library** — `import { scanNetwork } from "ollanet"` for apps (Node 20+, not the browser)
 
 ## Use cases
@@ -125,6 +126,8 @@ ollanet ps studio
 
 # Talk to a machine by name, MagicDNS name, or IP
 ollanet prompt localhost "What is MagicDNS?"
+ollanet prompt localhost --file ./notes.md
+ollanet compare studio gemma3:12b llama3.2:3b --prompt "Explain MagicDNS"
 ollanet prompt 192.168.1.50 gemma4:12b "Hello from the LAN"
 
 # Continue later
@@ -154,6 +157,7 @@ for (const s of servers) {
 |---|---|---|
 | `scanNetwork()` | localhost + config + `OLLANET_HOSTS` + Tailscale | Button / first “find servers” |
 | `scanNetwork({ lanScan: true })` | plus TCP sweep of local `/24`s on port 11434 | Checkbox; never the default scan, never on every health check |
+| `lastScan()` | last saved snapshot, no network | UI “show last inventory” |
 | `discoverHosts()` then probe yourself | candidates only, no `/api/tags` | If the app already has its own Ollama client |
 
 Pass `config` to skip the file (useful when the app already has its own config):
@@ -172,7 +176,7 @@ await scanNetwork({
 
 `pullModel({ machine, model })` asks that host to download (or update) a library model. The bits never transit ollanet. `showModel`, `removeModel({ yes: true })`, and `listLoaded` are the inspect / delete / VRAM trio.
 
-Requires **ollanet ≥ 0.4.0** (library). **Model management (`pull` / `show` / `rm` / `ps`) needs ≥ 0.5.0.** Bundlers (Vite / SvelteKit SSR):
+Requires **ollanet ≥ 0.4.0** (library). **Model management (`pull` / `show` / `rm` / `ps`) needs ≥ 0.5.0.** **Compare, `lastScan`, and bench `--hot` need ≥ 0.6.0.** **`resolveTarget` / `listTargets` need ≥ 0.6.1.** Bundlers (Vite / SvelteKit SSR):
 
 ```ts
 // vite.config.ts
@@ -189,7 +193,8 @@ optimizeDeps: { exclude: ["ollanet"] },
 | `ollanet show <machine> <model>` | Inspect Modelfile / params / capabilities (`[tuned]` when it looks like Finetuna) |
 | `ollanet rm <machine> <model> --yes` | Delete a model from that host’s disk |
 | `ollanet ps [machine]` | Models loaded in VRAM (omit machine = every discovered host) |
-| `ollanet prompt …` | Send a prompt / continue a chat |
+| `ollanet prompt …` | Send a prompt / continue a chat (`--file` for `.txt` / `.md`) |
+| `ollanet compare …` | Same prompt on 2–5 models; summary + `compares/<id>.md` |
 | `ollanet chats` | List or inspect saved transcripts |
 | `ollanet bench …` | Benchmark models for tok/s + lightweight quality checks |
 | `ollanet mcp` | Stdio [MCP](https://modelcontextprotocol.io) server for agents |
@@ -199,6 +204,7 @@ optimizeDeps: { exclude: ["ollanet"] },
 - `--json` — machine-readable output
 - `--all` — also probe offline Tailscale peers
 - `--lan` — scan local LAN CIDRs for open Ollama ports
+- `--last` / `--cached` — print the last saved scan (no network). Written to `~/.ollanet/last-scan.json` when installed, or `./last-scan.json` in a checkout (`OLLANET_LAST_SCAN` overrides).
 
 ### `prompt` options
 
@@ -219,12 +225,36 @@ ollanet prompt --chat <hash> <prompt...>
 | `--keep-alive <value>` | Keep model loaded (`5m`, `0`, `-1`, …) |
 | `--format json\|<schema>` | JSON mode / schema |
 | `--think` / `--no-think` | Enable/disable model thinking (default **off** so replies aren’t empty on qwen3-style models) |
+| `--file <path>` | Prompt from a `.txt` or `.md` file (joins argv / stdin) |
 | `--no-stream` | Buffer the full reply |
 | `--no-save` | Do not write a transcript |
 | `--json` | Emit JSON (includes chat id when saved) |
 
 Machine can be a discovered name, hostname, FQDN, or IP (`192.168.1.50`, `host:11434`). Direct addresses work even if they were never scanned.
 
+### `compare` options
+
+```text
+ollanet compare <machine> <model> <model> [model...]
+ollanet compare <machine> <model> <model> [model...] --prompt <text>
+ollanet compare <machine> <model> <model> [model...] --file <path.txt|.md>
+```
+
+Same prompt on **2–5 models** on one host. Prints a tok/s table and writes `compares/<id>.md` (readable) plus `.json`. This is not `bench` (no fixed suite / repeats). Omit `--prompt` / `--file` to use the built-in mesh-host tasting prompt (printed on stderr when used).
+
+| Flag | Meaning |
+|---|---|
+| `--prompt` / `-p` | Prompt text (default: built-in) |
+| `--file` / `-f` | Prompt from `.txt` or `.md` |
+| `--system` | System prompt |
+| `--temperature` / `--num-predict` / `--num-ctx` / `--keep-alive` | Same as `prompt` |
+| `--think` / `--no-think` | Thinking (default off) |
+| `--unload` | Unload each model before the next (fairer tok/s) |
+| `--no-save` | Do not write `compares/*` |
+| `--json` | Full record on stdout |
+
+Override dir with `OLLANET_COMPARES_DIR`.
+
 ### `pull` options
 
 ```text
@@ -274,8 +304,9 @@ Runs a **stdio MCP server** (no extra npm deps). Cursor / Claude Desktop / any M
 
 | Tool | What it does |
 |---|---|
-| `ollanet_scan` | Discover hosts + models (`lan`, `all` optional) |
-| `ollanet_prompt` | Prompt a host or continue with `chat_id` |
+| `ollanet_scan` | Discover hosts + models (`lan`, `all`, `last` optional) |
+| `ollanet_prompt` | Prompt a host or continue with `chat_id` (`file` optional) |
+| `ollanet_compare` | Same prompt on 2–5 models (`machine`, `models`; `prompt` / `file` optional) |
 | `ollanet_pull` | Pull / update a model on a host (`machine`, `model`) |
 | `ollanet_show` | Inspect a model (`machine`, `model`) |
 | `ollanet_rm` | Delete a model (`machine`, `model`, `confirm: true`) |
@@ -304,11 +335,14 @@ Or, from a checkout: `"command": "pnpm"`, `"args": ["ollanet", "--", "mcp"]` (wi
 ollanet bench <machine|ip> [model...]
 ollanet bench <machine|ip> --all
 ollanet bench localhost gemma4:12b --runs 5 --cold-load --json
+ollanet bench localhost gemma4:12b --hot --runs 5
 ```
 
 Runs a built-in suite (`quick` default, or `--suite full`) against one or more **completion**
 models: instruction checks once, throughput case repeated (`--runs`, default 3) with
-pinned `num_predict` / `seed` / `temperature`. Reports median tok/s + spread. Under `--all`,
+pinned `num_predict` / `seed` / `temperature`. Reports median tok/s + spread.
+`--hot` discards the first throughput run and leaves models loaded so the counted
+repeats are already warm (Finetuna steady-state). Under `--all`,
 skips non-completion models (embeddings); multimodal `completion`+`vision` models stay in
 (early-stop / pass_rate handle text-suite failures). Use `--exclude-vision` to drop them.
 Saves `benchmarks/<id>.json` (or `~/.ollanet/benchmarks/` when installed).

package/dist/argv.d.ts

@@ -0,0 +1,17 @@
+/**
+ * Shared CLI flag helpers. Commands keep their own usage() and unique flags.
+ */
+import type { GenerateSettings } from "./config.js";
+export type UsageFn = () => never;
+export declare function takeValue(args: string[], flag: string, usage: UsageFn): string;
+/** `--name value` or `--name=value`. Returns undefined if `arg` is not this flag. */
+export declare function takeFlag(arg: string, name: string, args: string[], usage: UsageFn): string | undefined;
+export declare function parseNumberFlag(raw: string, flag: string, usage: UsageFn): number;
+export declare function parseKeepAlive(raw: string): string | number;
+export interface SettingsFlagOpts {
+    temperature?: boolean;
+    numPredict?: boolean;
+    system?: boolean;
+}
+/** Shared generate-settings flags. Returns true when `arg` was consumed. */
+export declare function consumeSettingsFlag(arg: string, args: string[], settings: GenerateSettings, usage: UsageFn, opts?: SettingsFlagOpts): boolean;

package/dist/argv.js

@@ -0,0 +1,73 @@
+export function takeValue(args, flag, usage) {
+    const value = args.shift();
+    if (!value) {
+        console.error(`${flag} requires a value`);
+        usage();
+    }
+    return value;
+}
+/** `--name value` or `--name=value`. Returns undefined if `arg` is not this flag. */
+export function takeFlag(arg, name, args, usage) {
+    if (arg === name)
+        return takeValue(args, name, usage);
+    if (arg.startsWith(`${name}=`))
+        return arg.slice(name.length + 1);
+    return undefined;
+}
+export function parseNumberFlag(raw, flag, usage) {
+    const n = Number(raw);
+    if (!Number.isFinite(n)) {
+        console.error(`${flag} must be a number (got "${raw}")`);
+        usage();
+    }
+    return n;
+}
+export function parseKeepAlive(raw) {
+    const asNum = Number(raw);
+    return raw.trim() !== "" && Number.isFinite(asNum) && String(asNum) === raw ? asNum : raw;
+}
+/** Shared generate-settings flags. Returns true when `arg` was consumed. */
+export function consumeSettingsFlag(arg, args, settings, usage, opts = {}) {
+    if (arg === "--think") {
+        settings.think = true;
+        return true;
+    }
+    if (arg === "--no-think") {
+        settings.think = false;
+        return true;
+    }
+    if (opts.system) {
+        const system = takeFlag(arg, "--system", args, usage);
+        if (system !== undefined) {
+            settings.system = system;
+            return true;
+        }
+    }
+    if (opts.temperature) {
+        if (arg === "-t" || arg === "--temperature" || arg.startsWith("--temperature=")) {
+            const raw = arg === "-t" || arg === "--temperature"
+                ? takeValue(args, "--temperature", usage)
+                : arg.slice("--temperature=".length);
+            settings.temperature = parseNumberFlag(raw, "--temperature", usage);
+            return true;
+        }
+    }
+    if (opts.numPredict) {
+        const raw = takeFlag(arg, "--num-predict", args, usage);
+        if (raw !== undefined) {
+            settings.num_predict = Math.trunc(parseNumberFlag(raw, "--num-predict", usage));
+            return true;
+        }
+    }
+    const numCtx = takeFlag(arg, "--num-ctx", args, usage);
+    if (numCtx !== undefined) {
+        settings.num_ctx = Math.trunc(parseNumberFlag(numCtx, "--num-ctx", usage));
+        return true;
+    }
+    const keepAlive = takeFlag(arg, "--keep-alive", args, usage);
+    if (keepAlive !== undefined) {
+        settings.keep_alive = parseKeepAlive(keepAlive);
+        return true;
+    }
+    return false;
+}

package/dist/bench-suite.d.ts

@@ -21,5 +21,7 @@ export declare function comparabilityKey(opts: {
     temperature: number;
     think: boolean;
     numCtx: number | null;
+    /** Steady-state tok/s: first throughput run discarded, no inter-model unload. */
+    hot?: boolean;
 }): string;
 export declare function median(values: number[]): number | undefined;

package/dist/bench-suite.js

@@ -134,6 +134,7 @@ export function comparabilityKey(opts) {
         `temperature=${opts.temperature}`,
         `think=${opts.think ? 1 : 0}`,
         `num_ctx=${opts.numCtx ?? ""}`,
+        ...(opts.hot ? ["hot=1"] : []),
     ].join("|");
     return createHash("sha256").update(payload).digest("hex").slice(0, 16);
 }

package/dist/bench.js

@@ -5,7 +5,9 @@ import { defaultModelForHost, loadConfig, machineSettingsForHost, mergeSettings,
 import { comparabilityKey, getSuiteCases, median, suiteRevision, } from "./bench-suite.js";
 import { benchmarksDir, newBenchId, saveBenchmark } from "./bench-store.js";
 import { readFile } from "node:fs/promises";
-import { discoverHosts, envInt, ollamaBaseUrl, resolveHost, shortName, } from "./hosts.js";
+import { envInt, ollamaBaseUrl, shortName } from "./hosts.js";
+import { consumeSettingsFlag, parseNumberFlag, takeFlag } from "./argv.js";
+import { resolveTarget } from "./target.js";
 import { projectPath } from "./paths.js";
 import { contextLengthForModel, isCompletionCapable, isVisionCapable, ollamaChat, ollamaShow, ollamaTags, ollamaUnload, ollamaVersion, ollamaPs, psHasModel, shouldWarnNoThinking, vramForModel, waitUntilUnloaded, } from "./ollama-chat.js";
 const MODEL_COL = 34;
@@ -16,11 +18,16 @@ function usage() {
   ollanet bench <machine|ip> [model...] [options]
   ollanet bench <machine|ip> --all [options]
 
+Examples:
+  ollanet bench studio gemma3:12b --runs 5
+  ollanet bench studio gemma3:12b --hot --runs 5
+
 Options:
   --all                 Every completion-capable model from /api/tags
   --exclude-vision      With --all, skip models that advertise vision
   --suite quick|full    Prompt suite (default quick)
   --runs <n>            Throughput repeats (default 3)
+  --hot                 Discard first throughput run; keep models loaded
   --warmup / --no-warmup
   --cold-load           Measure cold load via unload + /api/ps
   --num-predict <n>     Pin throughput length (default 256)
@@ -36,22 +43,6 @@ Options:
 See docs/bench-spec.md for measurement rules.`);
     process.exit(1);
 }
-function takeValue(args, flag) {
-    const value = args.shift();
-    if (!value) {
-        console.error(`${flag} requires a value`);
-        usage();
-    }
-    return value;
-}
-function parseNumberFlag(raw, flag) {
-    const n = Number(raw);
-    if (!Number.isFinite(n)) {
-        console.error(`${flag} must be a number (got "${raw}")`);
-        usage();
-    }
-    return n;
-}
 function isLoopback(host) {
     return (host.source === "localhost" ||
         host.ip === "127.0.0.1" ||
@@ -85,6 +76,8 @@ function parseArgs(argv) {
     let all = false;
     let excludeVision = false;
     let warmup = true;
+    let warmupSet = false;
+    let hot = false;
     let coldLoad = false;
     let json = false;
     let save = true;
@@ -126,10 +119,16 @@ function parseArgs(argv) {
         }
         if (arg === "--warmup") {
             warmup = true;
+            warmupSet = true;
             continue;
         }
         if (arg === "--no-warmup") {
             warmup = false;
+            warmupSet = true;
+            continue;
+        }
+        if (arg === "--hot") {
+            hot = true;
             continue;
         }
         if (arg === "--cold-load") {
@@ -144,57 +143,35 @@ function parseArgs(argv) {
             failOnError = true;
             continue;
         }
-        if (arg === "--think") {
-            settings.think = true;
-            continue;
-        }
-        if (arg === "--no-think") {
-            settings.think = false;
-            continue;
-        }
         if (arg === "--judge") {
             judge = true;
             continue;
         }
-        if (arg === "--suite" || arg.startsWith("--suite=")) {
-            const raw = arg.includes("=") ? arg.slice("--suite=".length) : takeValue(args, "--suite");
-            if (raw !== "quick" && raw !== "full") {
-                console.error(`Unknown suite "${raw}" (use quick|full)`);
-                usage();
-            }
-            suite = raw;
-            continue;
-        }
-        if (arg === "--runs" || arg.startsWith("--runs=")) {
-            const raw = arg.includes("=") ? arg.slice("--runs=".length) : takeValue(args, "--runs");
-            runs = Math.max(1, Math.trunc(parseNumberFlag(raw, "--runs")));
+        if (consumeSettingsFlag(arg, args, settings, usage)) {
             continue;
         }
-        if (arg === "--num-predict" || arg.startsWith("--num-predict=")) {
-            const raw = arg.includes("=")
-                ? arg.slice("--num-predict=".length)
-                : takeValue(args, "--num-predict");
-            throughputNumPredict = Math.trunc(parseNumberFlag(raw, "--num-predict"));
+        const suiteRaw = takeFlag(arg, "--suite", args, usage);
+        if (suiteRaw !== undefined) {
+            if (suiteRaw !== "quick" && suiteRaw !== "full") {
+                console.error(`Unknown suite "${suiteRaw}" (use quick|full)`);
+                usage();
+            }
+            suite = suiteRaw;
             continue;
         }
-        if (arg === "--num-ctx" || arg.startsWith("--num-ctx=")) {
-            const raw = arg.includes("=") ? arg.slice("--num-ctx=".length) : takeValue(args, "--num-ctx");
-            settings.num_ctx = Math.trunc(parseNumberFlag(raw, "--num-ctx"));
+        const runsRaw = takeFlag(arg, "--runs", args, usage);
+        if (runsRaw !== undefined) {
+            runs = Math.max(1, Math.trunc(parseNumberFlag(runsRaw, "--runs", usage)));
             continue;
         }
-        if (arg === "--keep-alive" || arg.startsWith("--keep-alive=")) {
-            const raw = arg.includes("=")
-                ? arg.slice("--keep-alive=".length)
-                : takeValue(args, "--keep-alive");
-            const asNum = Number(raw);
-            settings.keep_alive =
-                raw.trim() !== "" && Number.isFinite(asNum) && String(asNum) === raw ? asNum : raw;
+        const predictRaw = takeFlag(arg, "--num-predict", args, usage);
+        if (predictRaw !== undefined) {
+            throughputNumPredict = Math.trunc(parseNumberFlag(predictRaw, "--num-predict", usage));
             continue;
         }
-        if (arg === "--judge-model" || arg.startsWith("--judge-model=")) {
-            judgeModel = arg.includes("=")
-                ? arg.slice("--judge-model=".length)
-                : takeValue(args, "--judge-model");
+        const judgeRaw = takeFlag(arg, "--judge-model", args, usage);
+        if (judgeRaw !== undefined) {
+            judgeModel = judgeRaw;
             continue;
         }
         if (arg.startsWith("-")) {
@@ -221,7 +198,8 @@ function parseArgs(argv) {
         excludeVision,
         suite,
         runs,
-        warmup,
+        warmup: hot && !warmupSet ? false : warmup,
+        hot,
         coldLoad,
         json,
         save,
@@ -382,12 +360,15 @@ async function runCaseAttempt(opts) {
         };
     }
 }
+function countedAttempts(attempts) {
+    return attempts.filter((a) => !a.discarded);
+}
 function summarizeModel(cases, cold, selfJudge) {
     const throughput = cases.find((c) => c.id === "throughput");
-    const fullLen = throughput?.attempts.filter((a) => !a.error && !a.early_stop && a.tok_s != null) ?? [];
+    const counted = countedAttempts(throughput?.attempts ?? []);
+    const fullLen = counted.filter((a) => !a.error && !a.early_stop && a.tok_s != null);
     const toks = fullLen.map((a) => a.tok_s);
-    const earlyStopCount = throughput?.attempts.filter((a) => a.early_stop || (a.done_reason && a.done_reason !== "length"))
-        .length ?? 0;
+    const earlyStopCount = counted.filter((a) => a.early_stop || (a.done_reason && a.done_reason !== "length")).length;
     const checkAttempts = cases
         .filter((c) => c.role === "check")
         .flatMap((c) => c.attempts)
@@ -404,12 +385,15 @@ function summarizeModel(cases, cold, selfJudge) {
         self_judge: selfJudge,
     };
 }
-function printTable(machine, suite, runs, revision, compKey, models, skipped, coldLoad) {
+function printTable(machine, suite, runs, revision, compKey, models, skipped, coldLoad, hot) {
     const lines = [];
-    lines.push(`Bench: ${models.length} completion model(s) × suite=${suite} × runs=${runs}`);
+    lines.push(`Bench: ${models.length} completion model(s) × suite=${suite} × runs=${runs}` +
+        (hot ? " (hot: first discarded)" : ""));
     lines.push(`comparability_key=${compKey}   suite_revision=${revision}   timeout: ${BENCH_TIMEOUT_MS}ms/case`);
     if (coldLoad)
         lines.push("Cold-load: on");
+    if (hot)
+        lines.push("Hot: first throughput run discarded; models stay loaded");
     lines.push("");
     const header = coldLoad
         ? pad("Model", MODEL_COL) +
@@ -462,6 +446,8 @@ function printTable(machine, suite, runs, revision, compKey, models, skipped, co
             notes.push(m.error.slice(0, 60));
         if (m.summary.early_stop_count > 0)
             notes.push(`${m.summary.early_stop_count}× early-stop`);
+        if (hot)
+            notes.push("hot");
         if (m.summary.self_judge)
             notes.push("self-judge");
         const load = m.summary.load_ms != null && m.summary.load_ms > 0
@@ -505,11 +491,7 @@ function pad(s, n) {
 export async function main() {
     const parsed = parseArgs(process.argv.slice(2));
     const config = await loadConfig();
-    const { hosts: targets } = await discoverHosts({
-        hosts: config.hosts,
-        discovery: config.discovery,
-    });
-    const host = resolveHost(targets, parsed.machine);
+    const host = await resolveTarget(parsed.machine, config);
     const baseUrl = ollamaBaseUrl(host);
     const machineLabel = shortName(host);
     const tags = await ollamaTags(baseUrl, Math.min(BENCH_TIMEOUT_MS, 15_000));
@@ -590,11 +572,13 @@ export async function main() {
         temperature: 0,
         think: baseSettings.think === true,
         numCtx: baseSettings.num_ctx ?? null,
+        hot: parsed.hot,
     });
     const suiteCases = getSuiteCases(parsed.suite);
+    const throughputReps = parsed.runs + (parsed.hot ? 1 : 0);
     const caseCount = selected.length *
         (suiteCases.filter((c) => c.role !== "throughput").length +
-            suiteCases.filter((c) => c.role === "throughput").length * parsed.runs);
+            suiteCases.filter((c) => c.role === "throughput").length * throughputReps);
     if (!parsed.json) {
         const skipBits = [
             skipped.filter((s) => s.reason === "non-completion").length
@@ -606,16 +590,22 @@ export async function main() {
         ].filter(Boolean);
         console.error(`Bench: ${selected.length} completion model(s)` +
             (skipBits.length ? ` (skipped ${skipBits.join(", ")})` : "") +
-            ` × suite=${parsed.suite} × runs=${parsed.runs}`);
+            ` × suite=${parsed.suite} × runs=${parsed.runs}` +
+            (parsed.hot ? " (hot: +1 discarded)" : ""));
         console.error(`Cases: ~${caseCount} chat calls   timeout: ${BENCH_TIMEOUT_MS}ms/case` +
             `   worst-case ceiling: ~${Math.ceil((caseCount * BENCH_TIMEOUT_MS) / 60000)}m if every call times out`);
-        if (!isLoopback(host)) {
-            console.error(`Note: benchmarking unloads models on ${machineLabel} between runs` +
-                (parsed.coldLoad ? " (and for --cold-load)" : "") +
-                ".");
+        if (!parsed.hot && !isLoopback(host) && (selected.length >= 2 || parsed.coldLoad)) {
+            const bits = [
+                selected.length >= 2 ? "unloads each model before the next" : "",
+                parsed.coldLoad ? "unloads for --cold-load" : "",
+            ].filter(Boolean);
+            console.error(`Note: benchmarking ${bits.join(" and ")} on ${machineLabel}.`);
         }
         console.error("Proceeding…");
     }
+    if (parsed.hot) {
+        console.error("Hot: discarding the first throughput run; models stay loaded (no unload between models).");
+    }
     if (parsed.settings.think === true) {
         for (const name of selected) {
             const caps = await fetchCapabilities(baseUrl, name);
@@ -665,18 +655,22 @@ export async function main() {
         let modelError;
         for (const caseDef of suiteCases) {
             const attempts = [];
-            const reps = caseDef.role === "throughput" ? parsed.runs : 1;
-            for (let r = 1; r <= reps; r += 1) {
+            const discardFirst = parsed.hot && caseDef.role === "throughput";
+            const reps = caseDef.role === "throughput" ? parsed.runs + (discardFirst ? 1 : 0) : 1;
+            for (let i = 0; i < reps; i += 1) {
+                const discarded = discardFirst && i === 0;
                 const attempt = await runCaseAttempt({
                     baseUrl,
                     model,
                     caseDef,
-                    run: r,
+                    run: discarded ? 0 : discardFirst ? i : i + 1,
                     settings: baseSettings,
                     throughputNumPredict: parsed.throughputNumPredict,
                     judgeModel,
                     selfJudge,
                 });
+                if (discarded)
+                    attempt.discarded = true;
                 attempts.push(attempt);
                 if (attempt.error) {
                     anyError = true;
@@ -717,8 +711,9 @@ export async function main() {
         });
         if (parsed.failFast && modelError)
             break;
-        // Unload before next model when ≥2 models remain after this one / multi-model run
-        if (selected.length >= 2 && mi < selected.length - 1) {
+        // Unload before next model when ≥2 models remain after this one / multi-model run.
+        // --hot keeps weights resident so the discarded first shot warms the rest.
+        if (!parsed.hot && selected.length >= 2 && mi < selected.length - 1) {
             try {
                 await ollamaUnload(baseUrl, model, BENCH_TIMEOUT_MS);
                 await waitUntilUnloaded(baseUrl, model, UNLOAD_WAIT_MS);
@@ -755,6 +750,7 @@ export async function main() {
             throughput_num_predict: parsed.throughputNumPredict,
             num_ctx: baseSettings.num_ctx ?? null,
             warmup: parsed.warmup,
+            hot: parsed.hot,
             cold_load: parsed.coldLoad,
             think: baseSettings.think === true,
             bench_timeout_ms: BENCH_TIMEOUT_MS,
@@ -778,7 +774,7 @@ export async function main() {
     }
     else {
         process.stdout.write("\n");
-        printTable(machineLabel, parsed.suite, parsed.runs, revision, compKey, modelResults, skipped, parsed.coldLoad);
+        printTable(machineLabel, parsed.suite, parsed.runs, revision, compKey, modelResults, skipped, parsed.coldLoad, parsed.hot);
         if (parsed.save)
             console.error(`benchmarks dir: ${benchmarksDir()}`);
     }

package/dist/chats.js

@@ -7,6 +7,7 @@
  *   ollanet chats --id a1b2c3d4e5f6
  */
 import { loadChat, listChats, responsesDir, } from "./chat-store.js";
+import { takeFlag } from "./argv.js";
 function usage() {
     console.error(`Usage: ollanet chats [--json] [--id <hash>]
 
@@ -31,12 +32,9 @@ function parseArgs(argv) {
             json = true;
             continue;
         }
-        if (arg === "--id" || arg.startsWith("--id=")) {
-            id = arg.includes("=") ? arg.slice("--id=".length) : args.shift();
-            if (!id) {
-                console.error("--id requires a value");
-                usage();
-            }
+        const idFlag = takeFlag(arg, "--id", args, usage);
+        if (idFlag !== undefined) {
+            id = idFlag;
             continue;
         }
         console.error(`Unknown flag: ${arg}`);

package/dist/cli.js

@@ -16,9 +16,11 @@ Discover Ollama hosts on localhost, your LAN, Tailscale, or any IP/hostname
 you can reach — then chat with hash-addressed transcripts.
 
 Usage:
-  ollanet scan [--json] [--all] [--lan]
+  ollanet scan [--json] [--all] [--lan] [--last]
   ollanet prompt <machine|ip> [model] <prompt...>
   ollanet prompt --chat <hash> <prompt...>
+  ollanet prompt <machine|ip> --file <path.txt|.md>
+  ollanet compare <machine> <model> <model> [model...] [--prompt <text> | --file <path>]
   ollanet pull <machine|ip> <model>
   ollanet show <machine|ip> <model>
   ollanet rm <machine|ip> <model> --yes
@@ -30,8 +32,11 @@ Usage:
 Examples:
   ollanet scan
   ollanet scan --lan
+  ollanet scan --last            # print the last saved scan (no network)
   ollanet prompt localhost "What is MagicDNS?"
-  ollanet prompt 192.168.1.50 gemma4:12b "Hello"
+  ollanet prompt localhost --file ./notes.md
+  ollanet compare studio gemma3:12b llama3.2:3b
+  ollanet compare studio gemma3:12b llama3.2:3b --prompt "Explain MagicDNS"
   ollanet prompt --chat a1b2c3d4e5f6 "Tell me more"
   ollanet pull studio gemma3:12b
   ollanet show studio gemma4-ctx32k
@@ -40,12 +45,14 @@ Examples:
   ollanet chats
   ollanet bench localhost --all
   ollanet bench localhost llama3.2:1b --runs 3
+  ollanet bench localhost gemma3:12b --hot --runs 5
   ollanet mcp                  # stdio MCP server for agents
 
 Config: ~/.ollanet/config.json when installed, ./config.json in a checkout
         (override with OLLANET_CONFIG)
 Hosts:  config.hosts, OLLANET_HOSTS, optional Tailscale + --lan
 Chats:  ~/.ollanet/responses or ./responses (override with OLLANET_RESPONSES_DIR)
+Scan:   ~/.ollanet/last-scan.json or ./last-scan.json (override with OLLANET_LAST_SCAN)
 `;
 async function main() {
     // pnpm/npm often forward a bare `--`; ignore it.
@@ -109,6 +116,13 @@ async function main() {
             await run();
             return;
         }
+        case "compare":
+        case "vs":
+        case "h2h": {
+            const { main: run } = await import("./compare.js");
+            await run();
+            return;
+        }
         case "bench":
         case "benchmark": {
             const { main: run } = await import("./bench.js");

package/dist/compare-store.d.ts

@@ -0,0 +1,8 @@
+import type { CompareRecord } from "./compare.js";
+export declare function comparesDir(): string;
+export declare function newCompareId(): string;
+export declare function renderCompareMarkdown(record: CompareRecord): string;
+export declare function saveCompare(record: CompareRecord): Promise<{
+    json: string;
+    md: string;
+}>;