pennyrouter 0.3.37 → 0.3.39

package/README.md

@@ -87,12 +87,12 @@ of `npx pennyrouter install`; to install or refresh it separately, run:
 npx pennyrouter compare
 ```
 
-On macOS/Linux, the installer puts `penny-compare` and `penny-session` on PATH alongside `penny`.
-These commands require `tmux`, Python 3.10+, and the harnesses being compared. On Windows,
-`penny-compare` starts a local browser terminal with native Claude and `penny claude` side by
-side. Its live token and cost readouts use Claude transcripts, with no tmux or Python install.
-Open the local URL printed by the command. Sync is Ctrl+Shift+S; Exit and save report is
-Ctrl+Shift+X. The Windows path currently supports Claude Code comparisons.
+Run `penny compare` on macOS, Linux, or Windows to open the browser-based setup. It mirrors the
+classic launcher's Claude/Codex pane, directory, label, and model choices. The browser version
+uses Python session metering on macOS/Linux and Node transcript metering on Windows, so it does
+not need tmux. The classic terminal interface remains available as `penny-compare` on macOS/Linux
+and requires tmux and Python 3.10+. Sync is Ctrl+Shift+S; Exit saves the report and shows the
+final summary.
 
 On macOS and Linux, a normal install adds a small `penny` launcher under `~/.local/bin`.
 On Windows it installs `penny.cmd` under `%LOCALAPPDATA%\\PennyRouter\\bin` and adds that

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pennyrouter",
-  "version": "0.3.37",
+  "version": "0.3.39",
   "description": "Install and manage PennyRouter local coding-agent integrations.",
   "homepage": "https://pennyrouter.com",
   "bugs": {

package/src/cli.js

@@ -138,9 +138,8 @@ export async function main(argv = process.argv.slice(2)) {
       throw new Error("Penny Compare requires macOS, Linux, or Windows.");
     }
     console.log(`Penny Compare installed at ${result.path}`);
-    console.log(process.platform === "win32"
-      ? "Run `penny-compare`, then open the local URL it prints."
-      : "Run `penny-compare` or `penny-session` to use it.");
+    console.log("Run `penny compare` to open the browser-based comparison setup.");
+    if (process.platform !== "win32") console.log("Run `penny-compare` for the classic tmux interface.");
     return;
   }
 
@@ -1010,9 +1009,9 @@ export async function install(flags = {}) {
     }
   }
   if (runtimeResult && !runtimeResult.available) {
-    console.log("Add the PennyRouter launcher directory to PATH before using the `penny` command.");
+    console.log(color(color("Add the PennyRouter launcher directory to PATH before using the `penny` command.", "bold"), "yellow"));
   } else if (runtimeResult?.changed) {
-    console.log("Open a new terminal (or reload your shell) before using the `penny` command.");
+    console.log(color(color("Open a new terminal (or reload your shell) before using the `penny` command.", "bold"), "yellow"));
   }
 
   // Registered after routing, which is the primary job: a naming conflict here reports itself
@@ -1460,9 +1459,9 @@ async function update(flags) {
 
   console.log("");
   if (runtimeResult && !runtimeResult.available) {
-    console.log("Add the PennyRouter launcher directory to PATH before using the `penny` command.");
+    console.log(color(color("Add the PennyRouter launcher directory to PATH before using the `penny` command.", "bold"), "yellow"));
   } else if (runtimeResult?.changed) {
-    console.log("Open a new terminal (or reload your shell) before using the `penny` command.");
+    console.log(color(color("Open a new terminal (or reload your shell) before using the `penny` command.", "bold"), "yellow"));
   }
   console.log("Restart any running Claude Code or Codex sessions to pick up the update.");
   reportFailures(failures);
@@ -2391,6 +2390,7 @@ function color(text, style) {
     bold: ["\x1b[1m", "\x1b[22m"],
     cyan: ["\x1b[36m", "\x1b[39m"],
     green: ["\x1b[32m", "\x1b[39m"],
+    yellow: ["\x1b[33m", "\x1b[39m"],
   };
   const pair = codes[style];
   return pair ? `${pair[0]}${text}${pair[1]}` : text;

package/src/launch.js

@@ -314,13 +314,13 @@ export async function installPennyCompare() {
   }
   await mkdir(join(destination, "reports"), { recursive: true, mode: 0o700 });
   await mkdir(pennyBinDir(), { recursive: true, mode: 0o755 });
+  const modules = join(destination, "web-demo", "node_modules");
+  for (const name of ["@xterm/addon-fit", "@xterm/xterm", "node-pty", "ws"]) {
+    const target = join(modules, name);
+    await mkdir(dirname(target), { recursive: true });
+    await cp(await findInstalledPackage(name), target, { recursive: true });
+  }
   if (process.platform === "win32") {
-    const modules = join(destination, "web-demo", "node_modules");
-    for (const name of ["@xterm/addon-fit", "@xterm/xterm", "node-pty", "ws"]) {
-      const target = join(modules, name);
-      await mkdir(dirname(target), { recursive: true });
-      await cp(await findInstalledPackage(name), target, { recursive: true });
-    }
     const link = join(pennyBinDir(), "penny-compare.cmd");
     const marker = "@REM PennyRouter-managed compare launcher";
     let previous = "";
@@ -917,6 +917,9 @@ export async function writeClaudeLaunchSettings({
             // An empty value is treated as unset by Claude Code, and prevents a
             // stale API key in the shared settings from creating dual auth.
             ANTHROPIC_API_KEY: "",
+            // Claude Code turns tool search off behind a custom base URL unless told
+            // otherwise; Penny relays deferred tools and their tool_reference blocks.
+            ENABLE_TOOL_SEARCH: "true",
             ANTHROPIC_DEFAULT_HAIKU_MODEL: "",
             ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME: "",
             ANTHROPIC_DEFAULT_HAIKU_MODEL_DESCRIPTION: "",
@@ -1413,6 +1416,7 @@ export function buildPennyInvocation(harness, args, config, models = null) {
     // Ask Claude Code to discover the gateway's model catalog at startup. The gateway
     // advertises Penny-facing ids in /v1/models; this is separate from inference routing.
     env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1";
+    env.ENABLE_TOOL_SEARCH = "true";
     delete env.ANTHROPIC_API_KEY;
     for (const slot of ["haiku", "sonnet", "opus", "fable"]) {
       const name = `ANTHROPIC_DEFAULT_${slot.toUpperCase()}_MODEL`;
@@ -1501,16 +1505,36 @@ export async function runPennyClaudeStats(args, profile) {
   return result.code ?? 1;
 }
 
+/** Launch the separately installed Penny Compare app from the `penny` command. */
+async function runPennyCompare(args) {
+  const compareDir = pennyCompareDir();
+  const executable = join(compareDir, "web-demo", "server.mjs");
+  try {
+    await stat(executable);
+  } catch {
+    throw new Error("Penny Compare is not installed. Run `npx pennyrouter compare` first.");
+  }
+  const child = spawn(process.execPath, [executable, "--compare", ...args], {
+    cwd: process.cwd(), env: process.env, stdio: "inherit",
+  });
+  const result = await waitForChild(child);
+  if (result.signal) return 128 + (osConstants.signals[result.signal] || 0);
+  return result.code ?? 1;
+}
+
 export async function runPenny(argv = process.argv.slice(2)) {
   const harness = argv[0];
   if (!harness || harness === "help" || harness === "--help" || harness === "-h") {
-    console.log("Usage: penny <claude|codex|opencode> [--penny-profile <name>] [native arguments...]");
+    console.log("Usage: penny <claude|codex|opencode|compare> [arguments...]");
+    console.log("       penny compare launches Penny Compare; install it with `npx pennyrouter compare`.");
+    console.log("       penny <claude|codex|opencode> [--penny-profile <name>] [native arguments...]");
     console.log("       --private / --cloud select the like-named profile.");
     console.log("       penny claude --stats adds the live single-session token readout.");
     return 0;
   }
+  if (harness === "compare") return runPennyCompare(argv.slice(1));
   if (!["claude", "codex", "opencode"].includes(harness)) {
-    throw new Error(`Unknown harness "${harness}". Try penny claude, penny codex, or penny opencode.`);
+    throw new Error(`Unknown Penny command "${harness}". Try penny claude, penny codex, penny opencode, or penny compare.`);
   }
   const statsSelection = extractStatsArgs(argv.slice(1));
   if (statsSelection.stats && harness !== "claude") {

package/src/launch.test.js

@@ -139,6 +139,7 @@ async function main() {
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_BASE_URL, "https://gateway.example/");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_AUTH_TOKEN, "pr-super-secret");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_API_KEY, "");
+    assert.equal(claudeLaunchSettings.env.ENABLE_TOOL_SEARCH, "true");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_DEFAULT_OPUS_MODEL, "");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_DEFAULT_OPUS_MODEL_NAME, "");
     await writeClaudeLaunchSettings({
@@ -264,6 +265,7 @@ async function main() {
     assert.equal(claude.env.ANTHROPIC_AUTH_TOKEN, "pr-super-secret");
     assert.equal(claude.env.ANTHROPIC_BASE_URL, "https://gateway.example/");
     assert.equal(claude.env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY, "1");
+    assert.equal(claude.env.ENABLE_TOOL_SEARCH, "true");
     // 1M-capable slots carry the `[1m]` suffix; Haiku has no 1M variant and stays bare.
     // Per-slot suffixes, not the global CLAUDE_CODE_MAX_CONTEXT_TOKENS override, which
     // would wrongly hand Haiku a 1M window too.

package/standalone/penny-compare

@@ -62,9 +62,11 @@ label_right=""
 wire=""
 wire_both=""
 wire_https=""
+pricing_profile=""
 while [ $# -gt 0 ]; do
   case "$1" in
     --model) model="$2"; shift 2 ;;
+    --pricing-profile) pricing_profile="$2"; shift 2 ;;
     --dir-left) dir_left="$2"; shift 2 ;;
     --dir-right) dir_right="$2"; shift 2 ;;
     --native-proxy) native_proxy="--native-proxy"; shift ;;
@@ -175,6 +177,17 @@ else
   fi
 fi
 
+# Pricing is an estimate-only display setting, so choose it explicitly instead of silently
+# coupling it to the native transport proxy. This also keeps direct (two-argument) launches
+# consistent with menu launches.
+if [ -z "$pricing_profile" ]; then
+  pricing_profile="$(choose 'pricing profile' standard standard govcloud)"
+fi
+case "$pricing_profile" in
+  standard|govcloud) ;;
+  *) echo "unknown pricing profile: $pricing_profile (try standard or govcloud)" >&2; exit 1 ;;
+esac
+
 # A directory that does not exist would otherwise surface as a confusing tmux failure.
 for candidate in "$dir_left" "$dir_right"; do
   if [ -n "$candidate" ] && [ ! -d "${candidate/#\~/$HOME}" ]; then
@@ -243,9 +256,7 @@ workdir_right_arg=" --cwd $(printf '%q' "$workdir_right")"
 # a penny-* pane already owns its own proxy routing and ignores the flag either way.
 native_proxy_left=""
 native_proxy_right=""
-pricing_profile="standard"
 if [ -n "$native_proxy" ]; then
-  pricing_profile="govcloud"
   case "$left" in claude|codex) native_proxy_left="$native_proxy" ;; esac
   case "$right" in claude|codex) native_proxy_right="$native_proxy" ;; esac
 fi

package/standalone/pricing.py

@@ -7,15 +7,27 @@ that uses 1-hour caching, which the gateway does by default.
 
 from __future__ import annotations
 
-# input, cache_read, write_5m, write_1h, output
-RATES = {
-    "claude-opus-5":     (5.00e-06, 5.00e-07, 6.25e-06, 1.00e-05, 2.50e-05),
-    "claude-sonnet-5":   (3.00e-06, 3.00e-07, 3.75e-06, 6.00e-06, 1.50e-05),
-    "claude-haiku-4-5":  (1.00e-06, 1.00e-07, 1.25e-06, 2.00e-06, 5.00e-06),
-    # OpenAI does not offer a 1-hour cache tier; its write rate applies uniformly.
-    "gpt-5-codex":       (1.25e-06, 1.25e-07, 1.25e-06, 1.25e-06, 1.00e-05),
-    "gpt-5.1-codex":     (1.25e-06, 1.25e-07, 1.25e-06, 1.25e-06, 1.00e-05),
-}
+import json
+from pathlib import Path
+
+
+def _load_standard_rates():
+    """Standard per-token rates from rates.json (scripts/generate_compare_rates.py), the same
+    file the web-demo meters read. Returns (flat rates, tiered rates) keyed by model id; a tier
+    entry is (inclusive threshold, <= threshold rates, > threshold rates)."""
+    document = json.loads((Path(__file__).with_name("rates.json")).read_text())
+    flat, tiered = {}, {}
+    for name, model in document["models"].items():
+        flat[name] = tuple(model["rates"])
+        if "above" in model:
+            above = model["above"]
+            tiered[name] = (above["threshold"], tuple(model["rates"]), tuple(above["rates"]))
+    return flat, tiered
+
+
+# input, cache_read, write_5m, write_1h, output. OpenAI does not offer a 1-hour cache tier;
+# its write rate applies uniformly.
+RATES, STANDARD_TIERED_RATES = _load_standard_rates()
 
 # --native-proxy means the comparison is measuring the GovCloud route. Pricing remains the
 # same on both panes, but both use this schedule instead of the standard one.
@@ -26,9 +38,7 @@ GOVCLOUD_RATES = {
     "claude-haiku-4-5":  (1.10e-06, 1.10e-07, 1.375e-06, 2.20e-06, 5.50e-06),
 }
 
-# GPT GovCloud pricing changes when an individual request's input context exceeds 272K.
-# Each entry is (inclusive threshold, <= threshold rates, > threshold rates), with rate tuples
-# in the same input/read/write-5m/write-1h/output order as RATES.
+# The GovCloud schedule for the same tiered GPT models.
 GOVCLOUD_TIERED_RATES = {
     "gpt-5.6-luna": (
         272_000,
@@ -85,11 +95,11 @@ def _matching(name: str, choices) -> str | None:
 
 def rates_for(model: str, profile: str = "standard", context_tokens: int = 0):
     name = _normalize(model)
-    if profile == "govcloud":
-        tiered = _matching(name, GOVCLOUD_TIERED_RATES)
-        if tiered:
-            threshold, low, high = GOVCLOUD_TIERED_RATES[tiered]
-            return high if context_tokens > threshold else low
+    tiers = GOVCLOUD_TIERED_RATES if profile == "govcloud" else STANDARD_TIERED_RATES
+    tiered = _matching(name, tiers)
+    if tiered:
+        threshold, low, high = tiers[tiered]
+        return high if context_tokens > threshold else low
     rates = GOVCLOUD_RATES if profile == "govcloud" else RATES
     match = _matching(name, rates)
     if match: