troxy-cli 1.30.9 → 1.30.11

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.30.9",
+  "version": "1.30.11",
   "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

@@ -56,7 +56,7 @@ async function failClosed($, e, next, err) {
 // to every step of its turn, so the cache prefix stays stable within a turn.
 
 const MODELS = {
-  haiku: 'claude-haiku-4-5-20251001',
+  haiku: 'claude-haiku-5-5',
   sonnet: 'claude-sonnet-5-5',
   opus: 'claude-opus-5-5',
 };
@@ -67,6 +67,15 @@ const RANK = { haiku: 0, sonnet: 1, opus: 2 };
 // pick flipped between Sonnet and Opus. Sonnet is the default; only a clearly hard
 // prompt (above 0.85) keeps the requested model.
 const BANDS = [[0.85, 'sonnet']];
+// Haiku band, off unless switched on (config haikuBand, or TROXY_HAIKU_BAND=1 for an A/B run):
+// an easy prompt from a Sonnet session goes to Haiku 5.5. The 10-15 rejection was Haiku 4.5.
+// Sonnet sessions only; Opus keeps the bands above.
+const HAIKU_BANDS = [[0.2, 'haiku']];
+// A Haiku session moves up to Sonnet, once and for good, when a later prompt scores above
+// this or the context passes the Haiku price break (Haiku 5.5 is 5x dearer past 100k tokens).
+// Never back down: every switch re-writes the cache.
+const HAIKU_ESCALATE_ABOVE = 0.6;
+const HAIKU_MAX_TOKENS = 100000;
 const ROUTE_CONTEXT_FLOOR_TOKENS = 50000;
 const SCORE_TIMEOUT_MS = 3000;
 const SCORER_PROMPT_CHARS = 4000;
@@ -104,7 +113,7 @@ function asRequest(text) {
   return `<request>\n${String(text || '').replace(/<\/request>/gi, '</ request>')}\n</request>`;
 }
 
-function pickModel(score, requested, contextTokens, warm) {
+function pickModel(score, requested, contextTokens, warm, haiku = false) {
   const rq = tierOf(requested);
   if (rq === null) return requested; // unknown family (e.g. a newer tier): leave it alone
   // No score (scorer failed or echoed the prompt): stay on the warm model at any
@@ -114,7 +123,7 @@ function pickModel(score, requested, contextTokens, warm) {
     return w !== null && RANK[w] <= RANK[rq] ? warm : requested;
   }
   let tier = rq;
-  for (const [limit, name] of BANDS) {
+  for (const [limit, name] of (haiku && rq === 'sonnet' ? HAIKU_BANDS : BANDS)) {
     if (score <= limit) { if (RANK[name] < RANK[rq]) tier = name; break; }
   }
   // The cache is per model: the first prompt on a new model re-writes the whole
@@ -144,7 +153,8 @@ function troxyCommand(args, paused) {
 // List $/MTok in (output keeps the same ratios) times usage-bar points per list dollar,
 // measured from the bar's own readings (usage-audit bar_weights.py, 2026-10-13). The bar
 // charges Haiku more per dollar than list prices imply; same weights as the dashboard.
-const PRICE = { haiku: 1 * 4.45, sonnet: 2 * 3.19, opus: 4 * 2.80 };
+// Haiku 5.5 ($0.10 in): 5.2 points per list dollar puts it 12x under Sonnet per token, as measured 2026-10-22.
+const PRICE = { haiku: 0.1 * 5.2, sonnet: 2 * 3.19, opus: 4 * 2.80 };
 const WINDOW_LABEL = { five_hour: '5h', seven_day: '7d' };
 // Drawn in Troxy's teal: the status line can't be colored and shows as a yellow warning.
 const LINE = { plugin: 'troxy-guard', key: 'line' };
@@ -522,10 +532,12 @@ async function rememberAsked($, sid, want, ran) {
 // compaction) the cache is gone anyway, so pick by score. Otherwise hold the model the
 // cache is warm on: every switch re-writes the whole conversation (2026-10-16 A/B: 4-6
 // switches per session cost more bar than routing saved). /model always wins.
-function stepModel({ arm, requested, switched, cold, score, warm }) {
+function stepModel({ arm, requested, switched, cold, score, warm, haiku = false, tokens = 0 }) {
   if (arm !== 'treated' || switched) return requested;
-  if (cold) return pickModel(score, requested, 0, null);
+  if (cold) return pickModel(score, requested, 0, null, haiku && tokens <= HAIKU_MAX_TOKENS);
   const w = tierOf(warm), rq = tierOf(requested);
+  // ponytail: tokens is read at the prompt, so a turn that grows past 100k moves up at the next prompt.
+  if (w === 'haiku' && rq === 'sonnet' && (!haiku || score > HAIKU_ESCALATE_ABOVE || tokens > HAIKU_MAX_TOKENS)) return requested;
   return w !== null && rq !== null && RANK[w] <= RANK[rq] ? warm : requested;
 }
 
@@ -547,6 +559,11 @@ async function effortSwapOn($) {
   return Boolean((await troxyConfig($)).effortSwap);
 }
 
+async function haikuBandOn($) {
+  if ((await $.env.get('TROXY_HAIKU_BAND')) === '1') return true;
+  return Boolean((await troxyConfig($)).haikuBand);
+}
+
 async function* routeStep($, e, next) {
   // Main loop only: a subagent keeps whatever model it was given.
   if (e.agentId) {
@@ -575,6 +592,8 @@ async function* routeStep($, e, next) {
     arm, requested: want, switched,
     cold: isCold,
     score: decision ? decision.score : null, warm: ls && ls.model,
+    haiku: await haikuBandOn($).catch(() => false),
+    tokens: decision ? decision.tokens : 0,
   });
   coldNext.delete(sid);
   lastSent.set(sid, { model, at: now });

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

@@ -2,7 +2,7 @@ import { test } from 'node:test';
 import assert from 'node:assert/strict';
 import { _internal } from '../mods/troxy-guard/hooks/register.js';
 
-const { parseScore, pickModel, troxyCommand, modRouteFile, estimateTokens, MODELS } = _internal;
+const { parseScore, pickModel, troxyCommand, modRouteFile, estimateTokens, stepModel, MODELS } = _internal;
 
 test('parseScore: reads a bare number and clamps to 0..1', () => {
   assert.equal(parseScore('0.35'), 0.35);
@@ -90,3 +90,26 @@ test('estimateTokens: counts the prompt being submitted, not just the context so
   // a first-prompt of ~60k tokens must trip the floor
   assert.equal(pickModel(0.05, MODELS.opus, estimateTokens(0, 'x'.repeat(240000))), MODELS.opus);
 });
+
+test('pickModel: haiku band (flag) sends an easy Sonnet prompt to Haiku 5.5, nothing else changes', () => {
+  assert.equal(MODELS.haiku, 'claude-haiku-5-5');
+  assert.equal(pickModel(0.1, MODELS.sonnet, 1000, null, true), MODELS.haiku);
+  assert.equal(pickModel(0.2, MODELS.sonnet, 1000, null, true), MODELS.haiku);
+  assert.equal(pickModel(0.3, MODELS.sonnet, 1000, null, true), MODELS.sonnet);
+  assert.equal(pickModel(0.1, MODELS.sonnet, 1000, null, false), MODELS.sonnet); // off by default
+  assert.equal(pickModel(0.1, MODELS.opus, 1000, null, true), MODELS.sonnet);    // Opus keeps its bands
+  assert.equal(pickModel(0.1, MODELS.sonnet, 60000, null, true), MODELS.sonnet); // 50k floor still holds
+});
+
+test('stepModel: a Haiku session moves up to Sonnet on a hard prompt or a long context, never back down', () => {
+  const base = { arm: 'treated', requested: MODELS.sonnet, switched: false, haiku: true };
+  assert.equal(stepModel({ ...base, cold: true, score: 0.1, warm: null }), MODELS.haiku);
+  assert.equal(stepModel({ ...base, cold: true, score: 0.1, warm: null, tokens: 150000 }), MODELS.sonnet); // too long to start on Haiku
+  assert.equal(stepModel({ ...base, cold: false, score: 0.1, warm: MODELS.haiku }), MODELS.haiku);        // easy prompt: hold
+  assert.equal(stepModel({ ...base, cold: false, score: null, warm: MODELS.haiku }), MODELS.haiku);       // no score: hold
+  assert.equal(stepModel({ ...base, cold: false, score: 0.7, warm: MODELS.haiku }), MODELS.sonnet);       // hard prompt: up
+  assert.equal(stepModel({ ...base, cold: false, score: 0.1, warm: MODELS.haiku, tokens: 120000 }), MODELS.sonnet); // long: up
+  assert.equal(stepModel({ ...base, cold: false, score: 0.1, warm: MODELS.sonnet }), MODELS.sonnet);      // Sonnet never goes down
+  assert.equal(stepModel({ ...base, haiku: false, cold: false, score: 0.1, warm: MODELS.haiku }), MODELS.sonnet); // flag off: up
+  assert.equal(stepModel({ ...base, arm: 'holdout', cold: true, score: 0.1, warm: null }), MODELS.sonnet); // holdout untouched
+});