troxy-cli 1.29.38 → 1.29.39

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.29.38",
+  "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];
 }
 

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);
 });