troxy-cli 1.30.0 → 1.30.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.30.0",
+  "version": "1.30.2",
   "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

@@ -181,13 +181,19 @@ function sessionLine(t) {
     Object.entries(t.meter).map(([kind, percentUsed]) => ({ kind, percentUsed })));
 }
 
-// Tokens Troxy kept out of the requests of this session (compression + trimming), counted
-// by the server. undefined until something was saved.
-function savingsText(s) {
-  const c = (s && s.compressed_tokens) || 0, t = (s && s.trimmed_tokens) || 0;
+// This session's Troxy work, from the server's count of its calls: how many ran on a
+// lighter model than asked, plus tokens kept out of requests (compression + trimming,
+// API-key setups through the proxy only). A holdout session says so: it is never routed.
+// undefined until there is something to say.
+function savingsText(s, arm) {
+  const calls = (s && s.calls) || 0, routed = (s && s.routed_calls) || 0;
+  const kept = ((s && s.compressed_tokens) || 0) + ((s && s.trimmed_tokens) || 0);
   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 `${k(c + t)} tokens not re-sent, thanks to compression and trimming`;
+  if (arm === 'holdout') return calls ? 'comparison session: not routed, so Troxy can measure the difference' : undefined;
+  const parts = [];
+  if (calls) parts.push(`${routed} of ${calls} calls on a lighter model`);
+  if (kept > 0) parts.push(`${k(kept)} tokens not re-sent`);
+  return parts.length ? parts.join(' · ') : undefined;
 }
 
 // This session's savings from the server, or null when there is no key, no session id or
@@ -229,7 +235,7 @@ const REPORT_KEEP_ROWS = 1000; // kept for the next try while Troxy is unreachab
 let pending = [];
 let compacting = false;
 
-function reportRow({ sid, usage, requested, arm, cold, score, compact, rateLimits }) {
+function reportRow({ sid, usage, requested, arm, cold, score, compact, rateLimits, ts, effort, effortAsked }) {
   const rl = {};
   for (const w of rateLimits || []) {
     if (!WINDOW_LABEL[w.kind]) continue;
@@ -241,7 +247,7 @@ function reportRow({ sid, usage, requested, arm, cold, score, compact, rateLimit
     arm: arm || null, switch: arm ? (cold ? 'cold' : 'warm') : null, score: score == null ? null : score,
     compact: Boolean(compact), input_tokens: usage.input_tokens, output_tokens: usage.output_tokens,
     cache_read_tokens: usage.cache_read_input_tokens, cache_creation_tokens: usage.cache_creation_input_tokens,
-    rate_limits: rl,
+    rate_limits: rl, ts, effort: effort || null, effort_requested: effortAsked || null,
   };
 }
 
@@ -250,7 +256,7 @@ async function noteStep($, result, meta) {
   if (!usage || !usage.model) return;
   let rateLimits = [];
   try { rateLimits = (await $.session.usage()).rateLimits; } catch {}
-  pending.push(reportRow({ ...meta, sid: await $.session.id(), usage, compact: compacting, rateLimits }));
+  pending.push(reportRow({ ...meta, sid: await $.session.id(), usage, compact: compacting, rateLimits, ts: await $.clock.now() }));
   if (pending.length > REPORT_KEEP_ROWS) pending = pending.slice(-REPORT_KEEP_ROWS);
 }
 
@@ -504,6 +510,24 @@ function stepModel({ arm, requested, switched, cold, score, warm }) {
   return w !== null && rq !== null && RANK[w] <= RANK[rq] ? warm : requested;
 }
 
+// Effort swap: an easy prompt (scorer) runs at lower effort, so the model thinks and
+// writes less and makes fewer tool calls. Never raised. Claude Code sends the change
+// per message, which keeps the prompt cache (tested 2026-10-20: about 65 cache tokens
+// written on a changed turn, same as an unchanged one). Off unless switched on
+// (config effortSwap, or TROXY_EFFORT_SWAP=1 for an A/B run) until it is measured.
+const EFFORT_RANK = { low: 0, medium: 1, high: 2, xhigh: 3, max: 4 };
+const EFFORT_LOW_BELOW = 0.4;
+const EFFORT_MEDIUM_BELOW = 0.85;
+function stepEffort({ arm, requested, score, on }) {
+  if (!on || arm !== 'treated' || score == null || !(requested in EFFORT_RANK)) return requested;
+  const cap = score < EFFORT_LOW_BELOW ? 'low' : score < EFFORT_MEDIUM_BELOW ? 'medium' : null;
+  return cap && EFFORT_RANK[cap] < EFFORT_RANK[requested] ? cap : requested;
+}
+async function effortSwapOn($) {
+  if ((await $.env.get('TROXY_EFFORT_SWAP')) === '1') return true;
+  return Boolean((await troxyConfig($)).effortSwap);
+}
+
 async function* routeStep($, e, next) {
   // Main loop only: a subagent keeps whatever model it was given.
   if (e.agentId) {
@@ -546,8 +570,13 @@ async function* routeStep($, e, next) {
   }
   try { await rememberAsked($, sid, want, model); } catch {} // a missed write only loses this on resume
   if (model !== want && turn) turn.ratio = PRICE[tierOf(want)] / PRICE[tierOf(model)];
-  const result = yield* next(model === e.model ? e : { ...e, model });
-  try { await noteStep($, result, { requested: want, arm, cold: isCold, score: decision ? decision.score : null }); } catch {} // a missed row only thins the dashboard
+  let effort = e.effort;
+  if (!(await $.store.get('paused'))) {
+    try { effort = stepEffort({ arm, requested: e.effort, score: decision ? decision.score : null, on: await effortSwapOn($) }); } catch {}
+  }
+  const sent = model === e.model && effort === e.effort ? e : { ...e, model, ...(effort === undefined ? {} : { effort }) };
+  const result = yield* next(sent);
+  try { await noteStep($, result, { requested: want, arm, cold: isCold, score: decision ? decision.score : null, effort: sent.effort, effortAsked: e.effort }); } catch {} // a missed row only thins the dashboard
   return result;
 }
 
@@ -601,7 +630,7 @@ export function register(on) {
     let savings = null;
     try {
       savings = await fetchSavings($);
-      if (savings) await $.state.set(SAVINGS, savingsText(savings));
+      if (savings) await $.state.set(SAVINGS, savingsText(savings, (await $.state.get(ARM)).value) || null);
     } catch {} // a missed fetch only leaves the band as it was
     if (turn) {
       const { start, ratio } = turn;
@@ -623,4 +652,4 @@ export function register(on) {
 }
 
 // Exported for tests only - not part of the mod's runtime contract.
-export const _internal = { reportRow, noteStep, flushReport, pendingRows: () => pending, troxyConfig, askedFor, rememberAsked, asked, armCompact, cancelCompact, toDomainAndFields, guardBash, failClosed, parseScore, asRequest, pickModel, troxyCommand, modRouteFile, estimateTokens, MODELS, savedLine, sessionLine, addTurn, withKept, statusText, compactPlan, savingsText, fetchSavings, scorePrompt, routeStep, stepModel, lastSent };
+export const _internal = { stepEffort, reportRow, noteStep, flushReport, pendingRows: () => pending, troxyConfig, askedFor, rememberAsked, asked, armCompact, cancelCompact, toDomainAndFields, guardBash, failClosed, parseScore, asRequest, pickModel, troxyCommand, modRouteFile, estimateTokens, MODELS, savedLine, sessionLine, addTurn, withKept, statusText, compactPlan, savingsText, fetchSavings, scorePrompt, routeStep, stepModel, lastSent };

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

@@ -128,14 +128,14 @@ test('compactPlan: big idle sessions compact, small ones and a nearly full bar d
   assert.deepEqual(compactPlan({ tokens: 210000, now, rateLimits: [{ kind: 'seven_day', percentUsed: 99 }] }), { action: 'skip' });
 });
 
-test('savingsText: shows compression + trimming together, 0 when nothing was saved', () => {
+test('savingsText: calls on a lighter model, kept tokens only when there are some, holdout says so', () => {
   const { savingsText } = _internal;
-  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');
+  assert.equal(savingsText({ calls: 5, routed_calls: 5, compressed_tokens: 0, trimmed_tokens: 0 }), '5 of 5 calls on a lighter model');
+  assert.equal(savingsText({ calls: 4, routed_calls: 1, compressed_tokens: 5200, trimmed_tokens: 40100 }), '1 of 4 calls on a lighter model · 45k tokens not re-sent');
+  assert.equal(savingsText({ calls: 9, routed_calls: 0, compressed_tokens: 123277, trimmed_tokens: 154512861 }), '0 of 9 calls on a lighter model · 154.6M tokens not re-sent');
+  assert.equal(savingsText({ calls: 3, routed_calls: 0 }, 'holdout'), 'comparison session: not routed, so Troxy can measure the difference');
+  assert.equal(savingsText({ calls: 0 }), undefined);
+  assert.equal(savingsText(null), undefined);
 });
 
 test('fetchSavings: sends the saved key and the session id; null without a key or on an error', async () => {
@@ -364,6 +364,7 @@ test('usage report: each request is queued with model, tokens, bar and arm, then
   const row = pendingRows().at(-1);
   assert.equal(row.model, M.sonnet); assert.equal(row.model_requested, M.opus);
   assert.equal(row.arm, 'treated'); assert.equal(row.switch, 'cold'); assert.equal(row.cache_read_tokens, 900);
+  assert.equal(typeof row.ts, 'number');                  // when the request ran, not when the turn ended
   assert.deepEqual(row.rate_limits, { five_hour: { percent_used: 42, resets_at: '2026-10-20T10:00:00Z' } });
   assert.equal(JSON.stringify(row).includes('reply only ok'), false); // never prompt text
   await flushReport($);
@@ -402,3 +403,32 @@ test('usage report: a server error keeps the rows for the next turn', async () =
   await flushReport($);
   assert.equal(pendingRows().length, 0);
 });
+
+test('stepEffort: easy prompts drop to low, middle ones to medium, never raised, off unless on', () => {
+  const { stepEffort } = _internal;
+  const on = { arm: 'treated', on: true };
+  assert.equal(stepEffort({ ...on, requested: 'high', score: 0.1 }), 'low');
+  assert.equal(stepEffort({ ...on, requested: 'xhigh', score: 0.6 }), 'medium');
+  assert.equal(stepEffort({ ...on, requested: 'high', score: 0.9 }), 'high');     // hard: kept
+  assert.equal(stepEffort({ ...on, requested: 'low', score: 0.6 }), 'low');       // never raised
+  assert.equal(stepEffort({ ...on, requested: 'medium', score: 0.6 }), 'medium');
+  assert.equal(stepEffort({ ...on, requested: 'high', score: null }), 'high');    // no score: no change
+  assert.equal(stepEffort({ ...on, requested: undefined, score: 0.1 }), undefined); // model without effort
+  assert.equal(stepEffort({ ...on, requested: 16000, score: 0.1 }), 16000);       // a numeric budget: left alone
+  assert.equal(stepEffort({ arm: 'holdout', on: true, requested: 'high', score: 0.1 }), 'high');
+  assert.equal(stepEffort({ arm: 'treated', on: false, requested: 'high', score: 0.1 }), 'high');
+});
+
+test('routeStep: with the effort swap on, an easy prompt goes out at low effort and the row records it', async () => {
+  const { MODELS: M, routeStep, scorePrompt, pendingRows } = _internal;
+  const $ = { ...routingApi(),
+    env: { get: async (k) => (k === 'TROXY_EFFORT_SWAP' ? '1' : undefined) },
+    fs: { write: async () => {}, read: async () => { throw new Error('no config'); } } };
+  await scorePrompt($, { text: 'reply only ok', origin: { kind: 'composer' } }, async (e) => e);
+  let sent;
+  const gen = routeStep($, { model: M.opus, effort: 'high' }, async function* (e) { sent = e; return { usage: { model: e.model, output_tokens: 1 } }; });
+  for await (const _ of gen) { /* drain */ }
+  assert.equal(sent.effort, 'low');
+  assert.equal(pendingRows().at(-1).effort, 'low');
+  assert.equal(pendingRows().at(-1).effort_requested, 'high');
+});