troxy-cli 1.29.45 → 1.29.46

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.29.45",
+  "version": "1.29.46",
   "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

@@ -91,16 +91,27 @@ function parseScore(text) {
   return Number.isFinite(n) ? Math.min(1, Math.max(0, n)) : null;
 }
 
-function pickModel(score, requested, contextTokens) {
+function pickModel(score, requested, contextTokens, warm) {
   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
+  // size, since a switch would re-write the cache for no reason.
+  if (score === null) {
+    const w = tierOf(warm);
+    return w !== null && RANK[w] <= RANK[rq] ? warm : requested;
+  }
   let tier = rq;
   for (const [limit, name] of 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
-  // conversation, so past the floor a switch costs more than one cheap turn saves.
-  if (tier !== rq && contextTokens > ROUTE_CONTEXT_FLOOR_TOKENS) tier = rq;
+  // conversation, so past the floor stay on the model the cache is warm on
+  // (`warm`, what this session last ran), unless the prompt needs a stronger one.
+  if (contextTokens > ROUTE_CONTEXT_FLOOR_TOKENS) {
+    const w = tierOf(warm);
+    if (w === null || RANK[w] > RANK[rq]) tier = rq;
+    else if (RANK[tier] <= RANK[w]) tier = w;
+  }
   return tier === rq ? requested : MODELS[tier];
 }
 
@@ -198,7 +209,7 @@ function statusText({ version, paused, totals, rateLimits = [], compactOff = fal
     : 'This session: nothing moved to a cheaper model yet');
   const meter = rateLimits.filter(w => WINDOW_LABEL[w.kind]).map(w => `${WINDOW_LABEL[w.kind]} ${Math.round(w.percentUsed)}%`);
   if (meter.length) lines.push(`Plan usage now: ${meter.join(' · ')}`);
-  lines.push(`Prompts with over ${ROUTE_CONTEXT_FLOOR_TOKENS / 1000}k tokens of context stay on your model (switching would rewrite the cache).`);
+  lines.push(`Past ${ROUTE_CONTEXT_FLOOR_TOKENS / 1000}k tokens of context a session stays on the model it is on (switching would rewrite the cache), unless a prompt needs a stronger one.`);
   lines.push(compactOff ? 'Idle compaction is off (/troxy compact on).'
     : `Idle compaction is on: sessions over ${COMPACT_MIN_TOKENS / 1000}k tokens are compacted after ${COMPACT_IDLE_MS / 60000} min idle (/troxy compact off).`);
   lines.push('All sessions: dash.troxy.io');
@@ -302,26 +313,31 @@ async function scorePrompt($, e, next) {
   );
   const score = r.isAnswered ? parseScore(r.text) : null;
   $.ui.log(`troxy scorer: ${r.isAnswered ? JSON.stringify(r.text) : r.reason} -> ${score}`, { to: 'debug' });
-  if (score !== null) {
-    const u = await $.session.usage();
-    decision = { score, tokens: estimateTokens(u.context && u.context.tokens, e.text) };
-  }
+  const u = await $.session.usage();
+  decision = { score, tokens: estimateTokens(u.context && u.context.tokens, e.text) };
   return next(e);
 }
 
 // The model the session last asked for, per session, so a /model switch is seen.
 const lastAsked = new Map();
+// The model each session last ran on and when: its cache is warm for about an hour.
+const lastSent = new Map();
+const WARM_MS = 55 * 60 * 1000;
 
 async function* routeStep($, e, next) {
   // Main loop only: a subagent keeps whatever model it was given.
-  if (!decision || e.agentId) return yield* next(e);
+  if (e.agentId) return yield* next(e);
+  const sid = await $.session.id();
+  const now = await $.clock.now();
   // The user picked a different model since the last turn: their choice wins
   // for this turn (the proxy does the same, route_reason 'user_switch').
-  const sid = await $.session.id();
   const prev = lastAsked.get(sid);
   lastAsked.set(sid, e.model);
-  if (prev && prev !== e.model) { decision = null; return yield* next(e); }
-  const model = pickModel(decision.score, e.model, decision.tokens);
+  if (prev && prev !== e.model) decision = null;
+  if (!decision) { lastSent.set(sid, { model: e.model, at: now }); return yield* next(e); }
+  const ls = lastSent.get(sid);
+  const model = pickModel(decision.score, e.model, decision.tokens, ls && now - ls.at < WARM_MS ? ls.model : null);
+  lastSent.set(sid, { model, at: now });
   if (model !== e.model) {
     const file = modRouteFile($.plugin.root, sid);
     const key = `${sid}:${e.model}>${model}`;

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

@@ -98,7 +98,7 @@ test('statusText: version, routing, this session, plan usage, where the account
     'Troxy 1.29.42 · routing is on\n'
     + 'This session: your bar lasts 4.0x longer · $0.09 saved of $0.12 at list price · estimate\n'
     + 'Plan usage now: 5h 23% · 7d 77%\n'
-    + 'Prompts with over 50k tokens of context stay on your model (switching would rewrite the cache).\n'
+    + 'Past 50k tokens of context a session stays on the model it is on (switching would rewrite the cache), unless a prompt needs a stronger one.\n'
     + 'Idle compaction is on: sessions over 100k tokens are compacted after 55 min idle (/troxy compact off).\n'
     + 'All sessions: dash.troxy.io\n'
     + 'Type /troxy off to pause routing.');
@@ -147,10 +147,11 @@ test('fetchSavings: sends the saved key and the session id; null without a key o
 // (score 0.1) and the context is small, so the mod would route down to haiku.
 function routingApi() {
   const kv = new Map();
+  const sid = 'sess-' + Math.random().toString(36).slice(2, 10);
   return {
     state: { get: async (k) => ({ value: kv.get(k.key) }), set: async (k, v) => kv.set(k.key, v) },
     store: { get: async () => undefined },
-    session: { id: async () => 'sess-switch-test', usage: async () => ({ context: { tokens: 1000 }, rateLimits: [] }) },
+    session: { id: async () => sid, usage: async () => ({ context: { tokens: 1000 }, rateLimits: [] }) },
     plugin: { root: '/tmp/not-installed' },
     model: { complete: async () => ({ isAnswered: true, text: '0.1' }) },
     ui: { log: () => {} },
@@ -175,3 +176,28 @@ test('routeStep: a /model switch is honored for that turn, then routing resumes'
   assert.equal(await runTurn($, sonnet), sonnet); // user picked sonnet: kept
   assert.equal(await runTurn($, sonnet), haiku);  // same choice again: routing resumes
 });
+
+test('pickModel: past the floor, stay on the warm model unless the prompt needs a stronger one', () => {
+  const { pickModel, MODELS } = _internal;
+  const big = 80000;
+  assert.equal(pickModel(0.1, MODELS.opus, big, MODELS.haiku), MODELS.haiku); // easy: stay warm on haiku
+  assert.equal(pickModel(0.5, MODELS.opus, big, MODELS.haiku), MODELS.sonnet); // needs sonnet: step up
+  assert.equal(pickModel(0.1, MODELS.opus, big, MODELS.opus), MODELS.opus);   // warm on opus: no switch down
+  assert.equal(pickModel(0.1, MODELS.opus, big, null), MODELS.opus);          // cold or unknown: as before
+  assert.equal(pickModel(0.1, MODELS.opus, 1000, MODELS.opus), MODELS.haiku); // under the floor: route freely
+});
+
+test('routeStep: the session that crossed 50k on haiku stays on haiku (no cache rewrite)', async () => {
+  const $ = routingApi();
+  const { opus, haiku } = _internal.MODELS;
+  assert.equal(await runTurn($, opus), haiku);
+  $.session.usage = async () => ({ context: { tokens: 60000 }, rateLimits: [] });
+  assert.equal(await runTurn($, opus), haiku);
+});
+
+test('pickModel: no score never causes a switch away from the warm model past the floor', () => {
+  const { pickModel, MODELS } = _internal;
+  assert.equal(pickModel(null, MODELS.opus, 80000, MODELS.haiku), MODELS.haiku);
+  assert.equal(pickModel(null, MODELS.opus, 1000, MODELS.haiku), MODELS.haiku); // small context too
+  assert.equal(pickModel(null, MODELS.opus, 1000, null), MODELS.opus);           // nothing warm: as asked
+});