troxy-cli 1.30.8 → 1.30.10

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.30.8",
+  "version": "1.30.10",
   "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 +67,10 @@ 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']];
 const ROUTE_CONTEXT_FLOOR_TOKENS = 50000;
 const SCORE_TIMEOUT_MS = 3000;
 const SCORER_PROMPT_CHARS = 4000;
@@ -104,7 +108,7 @@ 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 +118,7 @@ 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
@@ -522,9 +526,9 @@ async function rememberAsked($, sid, want, ran) {
 // compaction) the cache is gone anyway, so pick by score. Otherwise hold the model the
 // cache is warm on: every switch re-writes the whole conversation (2026-10-16 A/B: 4-6
 // switches per session cost more bar than routing saved). /model always wins.
-function stepModel({ arm, requested, switched, cold, score, warm }) {
+function stepModel({ arm, requested, switched, cold, score, warm, haiku = false }) {
   if (arm !== 'treated' || switched) return requested;
-  if (cold) return pickModel(score, requested, 0, null);
+  if (cold) return pickModel(score, requested, 0, null, haiku);
   const w = tierOf(warm), rq = tierOf(requested);
   return w !== null && rq !== null && RANK[w] <= RANK[rq] ? warm : requested;
 }
@@ -547,6 +551,11 @@ async function effortSwapOn($) {
   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($, e, next) {
   // Main loop only: a subagent keeps whatever model it was given.
   if (e.agentId) {
@@ -575,6 +584,7 @@ async function* routeStep($, e, next) {
     arm, requested: want, switched,
     cold: isCold,
     score: decision ? decision.score : null, warm: ls && ls.model,
+    haiku: await haikuBandOn($).catch(() => false),
   });
   coldNext.delete(sid);
   lastSent.set(sid, { model, at: now });

package/src/sandbox/posttooluse-hook.js

@@ -171,6 +171,9 @@ async function readStdin(timeoutMs) {
 // (`troxy posttooluse-hook`, registered via patchClaudeCodeHooks in
 // init.js) and, below, by this file's own direct-execution guard.
 export async function runPostToolUseHook() {
+  // Off by default: the 2026-10-22 split A/B showed output compression alone made the model
+  // re-run and re-read (about 4.5x the tool calls) and cost more bar. TROXY_COMPRESS=1 re-enables it.
+  if (process.env.TROXY_COMPRESS !== '1') process.exit(0);
   try {
     const raw = await readStdin(STDIN_TIMEOUT_MS);
     const input = JSON.parse(raw);

package/src/tests/e2e-sandbox-chain.test.js

@@ -114,6 +114,7 @@ test('end-to-end: real init -> real hook subprocess -> real recall, and real fai
       ...process.env,
       HOME: tmpHome,
       PATH: '/usr/bin:/bin',
+      TROXY_COMPRESS: '1',
     };
 
     let stdout = '';

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

@@ -90,3 +90,13 @@ 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
+});

package/src/tests/posttooluse-hook.test.js

@@ -255,3 +255,19 @@ test('the hook records the tokens fetched back with troxy_recall, per session, a
   await buildHookResponse({ tool_name: 'mcp__troxy__troxy_recall', tool_response: [{ type: 'text', text: 'z'.repeat(800) }] }, { storeDir: dir });
   assert.equal(fs.readdirSync(path.join(dir, 'recalled')).length, 1);
 });
+
+test('posttooluse-hook is a no-op unless TROXY_COMPRESS=1', () => {
+  const tmpHome = fs.mkdtempSync(path.join(os.tmpdir(), 'troxy-home-'));
+  try {
+    const run = (extra) => execFileSync(process.execPath, [troxyBin, 'posttooluse-hook'], {
+      input: JSON.stringify({ session_id: 's1', tool_name: 'Bash', tool_response: bashResult('x'.repeat(5000)) }),
+      env: { ...process.env, HOME: tmpHome, TROXY_COMPRESS: '', ...extra },
+      encoding: 'utf8',
+    });
+    assert.equal(run({}), '');
+    // Enabled: either compressed (treated arm) or left alone (holdout arm), never an error.
+    assert.doesNotThrow(() => run({ TROXY_COMPRESS: '1' }));
+  } finally {
+    fs.rmSync(tmpHome, { recursive: true, force: true });
+  }
+});