troxy-cli 1.29.38 → 1.29.40

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.29.38",
+  "version": "1.29.40",
   "description": "A secure control layer for AI agents: policies across payments, messages, logins, destructive actions, model usage, and secrets, all enforceable from the CLI",
   "homepage": "https://troxy.io",
   "bugs": {

package/src/mods/troxy-guard/.claude-plugin/plugin.json

@@ -2,5 +2,8 @@
   "name": "troxy-guard",
   "version": "0.1.0",
   "description": "Troxy keeps your Claude usage bar going longer. Type /troxy off to pause it.",
-  "author": { "name": "Troxy" }
+  "author": {
+    "name": "Troxy"
+  },
+  "types": "./types/index.d.ts"
 }

package/src/mods/troxy-guard/hooks/register.js

@@ -50,7 +50,7 @@ async function failClosed($, e, next, err) {
 //
 // Ported from the proxy's scorer (troxy-tf-live model-proxy/app.py; bands match
 // the production ROUTE_SCORER_BANDS=0.2:haiku,0.6:sonnet,0.8:opus,
-// never-upward rule, 50k context floor for Haiku). The mod scores each human
+// never-upward rule; the 50k context floor here covers every downward route, not just Haiku). The mod scores each human
 // prompt once with the user's own session via $.model.complete, so there is no
 // Troxy round trip and no credential to carry. One decision per prompt, applied
 // to every step of its turn, so the cache prefix stays stable within a turn.
@@ -62,7 +62,7 @@ const MODELS = {
 };
 const RANK = { haiku: 0, sonnet: 1, opus: 2 };
 const BANDS = [[0.2, 'haiku'], [0.6, 'sonnet'], [0.8, 'opus']];
-const HAIKU_CONTEXT_FLOOR_TOKENS = 50000;
+const ROUTE_CONTEXT_FLOOR_TOKENS = 50000;
 const SCORE_TIMEOUT_MS = 3000;
 const SCORER_PROMPT_CHARS = 4000;
 const SCORER_SYSTEM = 'You are a routing classifier for a coding-assistant proxy. Read the request below '
@@ -98,9 +98,9 @@ function pickModel(score, requested, contextTokens) {
   for (const [limit, name] of BANDS) {
     if (score <= limit) { if (RANK[name] < RANK[rq]) tier = name; break; }
   }
-  if (tier === 'haiku' && rq !== 'haiku' && contextTokens > HAIKU_CONTEXT_FLOOR_TOKENS) {
-    tier = RANK.sonnet < RANK[rq] ? 'sonnet' : rq;
-  }
+  // The cache is per model: the first prompt on a new model re-writes the whole
+  // conversation, so past the floor a switch costs more than one cheap turn saves.
+  if (tier !== rq && contextTokens > ROUTE_CONTEXT_FLOOR_TOKENS) tier = rq;
   return tier === rq ? requested : MODELS[tier];
 }
 
@@ -119,7 +119,16 @@ function troxyCommand(args, paused) {
 // ponytail: turn-weighted and ignores cache writes after a model switch, so it
 // is an estimate; weight by r.usage tokens if the number must be exact.
 const MODEL_COST = { haiku: 0.3, sonnet: 0.5, opus: 1.0 };
-const stats = { prompts: 0, routed: 0, cost: 0 }; // cost: actual spend in "requested-model prompts"
+// Kept in $.state (per session) so a `troxy update` hot reload keeps the counts.
+// cost: actual spend in "requested-model prompts".
+const STATS = { plugin: 'troxy-guard', key: 'stats' };
+
+async function addStats($, prompts, routed, cost) {
+  const { value: s = { prompts: 0, routed: 0, cost: 0 } } = await $.state.get(STATS);
+  const next = { prompts: s.prompts + prompts, routed: s.routed + routed, cost: s.cost + cost };
+  await $.state.set(STATS, next);
+  $.ui.status(usageLine(next));
+}
 
 function costRatio(fromModel, toModel) {
   const f = MODEL_COST[tierOf(fromModel)], t = MODEL_COST[tierOf(toModel)];
@@ -162,9 +171,7 @@ async function scorePrompt($, e, next) {
   );
   const score = r.isAnswered ? parseScore(r.text) : null;
   $.ui.log(`troxy scorer: ${r.isAnswered ? JSON.stringify(r.text) : r.reason} -> ${score}`, { to: 'debug' });
-  stats.prompts += 1;
-  stats.cost += 1; // until routeStep says otherwise
-  $.ui.status(usageLine(stats));
+  await addStats($, 1, 0, 1); // full cost until routeStep says otherwise
   if (score !== null) {
     const u = await $.session.usage();
     decision = { score, tokens: estimateTokens(u.context && u.context.tokens, e.text) };
@@ -179,9 +186,7 @@ async function* routeStep($, e, next) {
   if (model !== e.model) {
     if (!decision.counted) {
       decision.counted = true;
-      stats.routed += 1;
-      stats.cost += costRatio(e.model, model) - 1;
-      $.ui.status(usageLine(stats));
+      await addStats($, 0, 1, costRatio(e.model, model) - 1);
     }
     const sid = await $.session.id();
     const file = modRouteFile($.plugin.root, sid);
@@ -198,6 +203,9 @@ export function register(on) {
   on('tool.call', { tool: 'Bash' }, guardBash).catch(failClosed);
   on('session.start', async ($, e, next) => {
     await $.command.register({ name: 'troxy', description: 'Pause or resume Troxy model routing', argumentHint: '[on|off|status]' });
+    // session.start fires again on a hot reload: put the bar back from the kept counts
+    const { value } = await $.state.get(STATS);
+    if (value && value.prompts && !(await $.store.get('paused'))) $.ui.status(usageLine(value));
     return next(e);
   });
   on('command.run', { command: 'troxy' }, async ($, e) => {

package/src/mods/troxy-guard/types/index.d.ts

@@ -0,0 +1,7 @@
+export type UsageStats = { prompts: number; routed: number; cost: number }
+
+declare module 'claude-code' {
+  interface PluginState {
+    'troxy-guard': { stats: UsageStats }
+  }
+}

package/src/tests/mods-route.test.js

@@ -31,10 +31,12 @@ test('pickModel: never routes upward', () => {
   assert.equal(pickModel(0.4, MODELS.haiku, 1000), MODELS.haiku);
 });
 
-test('pickModel: no haiku once the context passes the 50k floor, sonnet instead', () => {
-  assert.equal(pickModel(0.05, MODELS.opus, 60000), MODELS.sonnet);
+test('pickModel: past the 50k floor nothing routes down, a cold cache would cost more than it saves', () => {
+  assert.equal(pickModel(0.05, MODELS.opus, 60000), MODELS.opus);
+  assert.equal(pickModel(0.4, MODELS.opus, 60000), MODELS.opus);
   assert.equal(pickModel(0.05, MODELS.sonnet, 60000), MODELS.sonnet);
   assert.equal(pickModel(0.05, MODELS.haiku, 60000), MODELS.haiku);
+  assert.equal(pickModel(0.4, MODELS.opus, 50000), MODELS.sonnet);
 });
 
 test('pickModel: a model it does not recognise is left alone', () => {
@@ -71,6 +73,6 @@ test('estimateTokens: counts the prompt being submitted, not just the context so
   assert.equal(estimateTokens(30000, 'x'.repeat(40000)), 40000);
   assert.equal(estimateTokens(undefined, 'hi'), 1);
   assert.equal(estimateTokens(5, ''), 5);
-  // a first-prompt of ~60k tokens must trip the Haiku floor
-  assert.equal(pickModel(0.05, MODELS.opus, estimateTokens(0, 'x'.repeat(240000))), MODELS.sonnet);
+  // a first-prompt of ~60k tokens must trip the floor
+  assert.equal(pickModel(0.05, MODELS.opus, estimateTokens(0, 'x'.repeat(240000))), MODELS.opus);
 });