troxy-cli 1.29.37 → 1.29.39

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.29.37",
+  "version": "1.29.39",
   "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/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];
 }
 
@@ -113,28 +113,24 @@ function troxyCommand(args, paused) {
 }
 
 
-// Session savings tracking: cumulative % savings from routing
-let sessionSavingsPercent = 0;
-let routingCount = 0;
-
-// Cost ratio of models (relative to Opus=1.0)
-const MODEL_COST = {
-  'claude-haiku-4-5-20251001': 0.3,
-  'claude-sonnet-5-5': 0.6,
-  'claude-opus-5-5': 1.0,
-};
+// Session usage tracking, one count per human prompt (not per step).
+// Weights relative to Opus: list prices put Haiku at 0.25 and Sonnet at 0.5; a
+// Pro-plan meter test (30 prompts each) put Haiku at ~0.33. Haiku sits between.
+// 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"
 
-function calculateSavings(fromModel, toModel) {
-  const fromCost = MODEL_COST[fromModel] || 1.0;
-  const toCost = MODEL_COST[toModel] || 1.0;
-  if (toCost >= fromCost) return 0; // no savings if routing upward
-  return Math.round(((fromCost - toCost) / fromCost) * 100);
+function costRatio(fromModel, toModel) {
+  const f = MODEL_COST[tierOf(fromModel)], t = MODEL_COST[tierOf(toModel)];
+  return f && t ? t / f : 1;
 }
 
-// Status-line text for the running session savings, e.g. [████░░░░░░░░░░░░░░░░] Troxy: -20% saved
-function savingsBar() {
-  const filled = Math.round((sessionSavingsPercent / 100) * 20);
-  return `[${'█'.repeat(filled)}${'░'.repeat(20 - filled)}] Troxy: -${sessionSavingsPercent}% saved`;
+// e.g. [████████░░░░░░░░░░░░] Troxy extended your usage by 87% · 2 of 3 prompts on a cheaper model
+function usageLine({ prompts, routed, cost }) {
+  const extended = cost > 0 ? Math.round((prompts / cost - 1) * 100) : 0;
+  const filled = prompts ? Math.round((routed / prompts) * 20) : 0;
+  return `[${'█'.repeat(filled)}${'░'.repeat(20 - filled)}] Troxy extended your usage by ${extended}% · ${routed} of ${prompts} prompts on a cheaper model`;
 }
 
 // Where the local daemon looks for this session's routing note (see interceptor.js).
@@ -165,6 +161,10 @@ async function scorePrompt($, e, next) {
     { model: 'haiku', system: SCORER_SYSTEM, prompt: e.text.slice(0, SCORER_PROMPT_CHARS), maxTokens: 8, effort: 'low', timeoutMs: SCORE_TIMEOUT_MS },
   );
   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));
   if (score !== null) {
     const u = await $.session.usage();
     decision = { score, tokens: estimateTokens(u.context && u.context.tokens, e.text) };
@@ -177,11 +177,11 @@ async function* routeStep($, e, next) {
   if (!decision || e.agentId) return yield* next(e);
   const model = pickModel(decision.score, e.model, decision.tokens);
   if (model !== e.model) {
-    const savings = calculateSavings(e.model, model);
-    if (savings > 0) {
-      routingCount += 1;
-      sessionSavingsPercent = Math.round((sessionSavingsPercent * (routingCount - 1) + savings) / routingCount);
-      $.ui.status(savingsBar());
+    if (!decision.counted) {
+      decision.counted = true;
+      stats.routed += 1;
+      stats.cost += costRatio(e.model, model) - 1;
+      $.ui.status(usageLine(stats));
     }
     const sid = await $.session.id();
     const file = modRouteFile($.plugin.root, sid);
@@ -213,4 +213,4 @@ export function register(on) {
 }
 
 // Exported for tests only - not part of the mod's runtime contract.
-export const _internal = { toDomainAndFields, guardBash, failClosed, parseScore, pickModel, troxyCommand, modRouteFile, estimateTokens, MODELS };
+export const _internal = { toDomainAndFields, guardBash, failClosed, parseScore, pickModel, troxyCommand, modRouteFile, estimateTokens, MODELS, usageLine, costRatio };

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);
 });

package/src/tests/mods-troxy-guard.test.js

@@ -62,3 +62,13 @@ test('real subprocess: a real Bash command still runs normally through the mod (
     fs.rmSync(workdir, { recursive: true, force: true });
   }
 });
+
+test('usageLine: counts every prompt, extension reflects routed share', () => {
+  const { usageLine, costRatio, MODELS } = _internal;
+  assert.equal(costRatio(MODELS.opus, MODELS.haiku), 0.3);
+  assert.equal(costRatio('some-new-model', MODELS.haiku), 1);
+  // 3 prompts, 2 routed opus->haiku: cost 1 + 0.3 + 0.3 = 1.6, 3/1.6 = +88%
+  assert.match(usageLine({ prompts: 3, routed: 2, cost: 1.6 }), /extended your usage by 88% · 2 of 3 prompts/);
+  assert.match(usageLine({ prompts: 3, routed: 0, cost: 3 }), /by 0% · 0 of 3/);
+  assert.match(usageLine({ prompts: 0, routed: 0, cost: 0 }), /by 0% · 0 of 0/);
+});