troxy-cli 1.29.59 → 1.29.61

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.29.59",
+  "version": "1.29.61",
   "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

@@ -188,9 +188,9 @@ function sessionLine(t) {
 // by the server. undefined until something was saved.
 function savingsText(s) {
   const c = (s && s.compressed_tokens) || 0, t = (s && s.trimmed_tokens) || 0;
-  const k = n => n >= 1000 ? `${Math.round(n / 1000)}k` : String(n);
+  const k = n => n >= 1e6 ? `${Math.round(n / 1e5) / 10}M`.replace('.0M', 'M') : n >= 1000 ? `${Math.round(n / 1000)}k` : String(n);
   // Always show, even if 0, so the band is transparent: 0 means no trimming/compression on this turn
-  return `${k(c + t)} tokens saved by compression and trimming`;
+  return `${k(c + t)} tokens not re-sent, thanks to compression and trimming`;
 }
 
 // This session's savings from the server, or null when there is no key, no session id or
@@ -440,13 +440,17 @@ const held = new Map();
 
 // Which model a step runs on. At a cold moment (first step, 55+ min idle, after a
 // compaction) any model is free to pick. While warm, hold the model, except at the
-// first step of a new prompt once it has held HOLD_PROMPTS prompts: then re-pick by
-// score, down or up. The user's own /model always wins and becomes the held model.
+// first step of a new prompt: up at once, down once it has held HOLD_PROMPTS prompts,
+// re-picking by score. The user's own /model always wins and becomes the held model.
 function stepModel({ arm, requested, switched, cold, score, warm, newPrompt, heldPrompts, tokens }) {
   if (arm !== 'treated' || switched) return requested;
   if (cold) return pickModel(score, requested, 0, null);
-  if (newPrompt && score !== null && heldPrompts >= HOLD_PROMPTS && tokens <= WARM_SWITCH_MAX_TOKENS) {
-    return pickModel(score, requested, 0, null);
+  if (newPrompt && score !== null && tokens <= WARM_SWITCH_MAX_TOKENS) {
+    const pick = pickModel(score, requested, 0, null);
+    // Up is immediate: a cheap model chosen for an easy first prompt must not keep
+    // running a harder one (the 2026-10-15 A/B: Haiku held a whole coding session and
+    // made 2-4x the calls). Down still waits HOLD_PROMPTS, since it re-writes the cache.
+    if (RANK[tierOf(pick)] > RANK[tierOf(warm)] || heldPrompts >= HOLD_PROMPTS) return pick;
   }
   const w = tierOf(warm), rq = tierOf(requested);
   return w !== null && rq !== null && RANK[w] <= RANK[rq] ? warm : requested;

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

@@ -134,10 +134,12 @@ test('compactPlan: big idle sessions compact, small ones and a nearly full bar d
 
 test('savingsText: shows compression + trimming together, 0 when nothing was saved', () => {
   const { savingsText } = _internal;
-  assert.equal(savingsText({ compressed_tokens: 5200, trimmed_tokens: 40100 }), '45k tokens saved by compression and trimming');
-  assert.equal(savingsText({ compressed_tokens: 300, trimmed_tokens: 0 }), '300 tokens saved by compression and trimming');
-  assert.equal(savingsText({ compressed_tokens: 0, trimmed_tokens: 0 }), '0 tokens saved by compression and trimming');
-  assert.equal(savingsText(null), '0 tokens saved by compression and trimming');
+  assert.equal(savingsText({ compressed_tokens: 5200, trimmed_tokens: 40100 }), '45k tokens not re-sent, thanks to compression and trimming');
+  assert.equal(savingsText({ compressed_tokens: 300, trimmed_tokens: 0 }), '300 tokens not re-sent, thanks to compression and trimming');
+  assert.equal(savingsText({ compressed_tokens: 0, trimmed_tokens: 0 }), '0 tokens not re-sent, thanks to compression and trimming');
+  assert.equal(savingsText({ compressed_tokens: 123277, trimmed_tokens: 154512861 }), '154.6M tokens not re-sent, thanks to compression and trimming');
+  assert.equal(savingsText({ compressed_tokens: 0, trimmed_tokens: 2000000 }), '2M tokens not re-sent, thanks to compression and trimming');
+  assert.equal(savingsText(null), '0 tokens not re-sent, thanks to compression and trimming');
 });
 
 test('fetchSavings: sends the saved key and the session id; null without a key or on an error', async () => {
@@ -213,6 +215,18 @@ test('stepModel: warm switch at a new prompt only after the hold, and under the
   assert.equal(stepModel({ ...base, arm: 'holdout', newPrompt: true, heldPrompts: HOLD_PROMPTS }), M.opus);
 });
 
+test('stepModel: a harder prompt upgrades a cheap warm session at once, a lighter one waits', () => {
+  const { stepModel, MODELS: M, HOLD_PROMPTS } = _internal;
+  const base = { arm: 'treated', requested: M.opus, switched: false, cold: false, warm: M.haiku, newPrompt: true, tokens: 5000 };
+  assert.equal(stepModel({ ...base, score: 0.9, heldPrompts: 0 }), M.opus);   // hard prompt on haiku: up now
+  assert.equal(stepModel({ ...base, score: 0.5, heldPrompts: 0 }), M.sonnet);
+  assert.equal(stepModel({ ...base, score: 0.1, heldPrompts: 0 }), M.haiku);  // still easy: stay
+  assert.equal(stepModel({ ...base, warm: M.opus, score: 0.1, heldPrompts: 0 }), M.opus);               // down waits
+  assert.equal(stepModel({ ...base, warm: M.opus, score: 0.1, heldPrompts: HOLD_PROMPTS }), M.haiku);   // ...for the hold
+  assert.equal(stepModel({ ...base, score: 0.9, heldPrompts: 0, tokens: 300000 }), M.haiku);            // too big to re-write
+  assert.equal(stepModel({ ...base, score: 0.9, heldPrompts: 0, newPrompt: false }), M.haiku);          // mid-turn: hold
+});
+
 test('pickModel: past the floor, stay on the warm model unless the prompt needs a stronger one', () => {
   const { pickModel, MODELS } = _internal;
   const big = 80000;