troxy-cli 1.30.3 → 1.30.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.30.3",
+  "version": "1.30.5",
   "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

@@ -161,7 +161,7 @@ function bar(fraction) {
 }
 
 // One line, most important first, e.g.
-// This session: savings under measurement · 5h: 19% used · 7d: 12% used
+// 5h: 19% used · 7d: 12% used
 function savedLine(money, rateLimits = []) {
   const paid = money && money.cost_paid_usd;
   if (!(paid > 0)) return undefined;
@@ -170,8 +170,7 @@ function savedLine(money, rateLimits = []) {
   // measured replacement exists.
   const moved = rateLimits.filter(w => WINDOW_LABEL[w.kind] && Math.round(w.percentUsed) >= 1);
   if (!(money.dollars_saved_usd > 0) && !moved.length) return undefined; // nothing routed or cleared yet
-  return ['This session: savings under measurement',
-    ...moved.map(w => `${WINDOW_LABEL[w.kind]}: ${Math.round(w.percentUsed)}% used`)].join(' · ');
+  return moved.map(w => `${WINDOW_LABEL[w.kind]}: ${Math.round(w.percentUsed)}% used`).join(' · ') || undefined;
 }
 
 // t.kept: what the content Troxy cleared or compressed would have cost (server's
@@ -181,19 +180,15 @@ function sessionLine(t) {
     Object.entries(t.meter).map(([kind, percentUsed]) => ({ kind, percentUsed })));
 }
 
-// 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.
+// What Troxy did in this session, from the server's count of its calls: the share that
+// ran on a lighter model instead of the model asked for. A holdout session says so: it is never
+// routed. undefined until there is something to say. (Tokens cleared by the old proxy are
+// not shown: subscription sessions no longer go through it.)
 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);
-  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;
+  if (!calls) return undefined;
+  if (arm === 'holdout') return 'comparison session: not routed, so Troxy can measure the difference';
+  return routed ? `${Math.max(1, Math.round(routed / calls * 100))}% of this session ran on a lighter model` : 'no calls moved to a lighter model yet';
 }
 
 // This session's savings from the server, or null when there is no key, no session id or
@@ -235,7 +230,7 @@ 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, ts, effort, effortAsked, surfaces }) {
+function reportRow({ sid, usage, requested, arm, cold, score, compact, rateLimits, ts, effort, effortAsked, surfaces, compressed }) {
   const rl = {};
   for (const w of rateLimits || []) {
     if (!WINDOW_LABEL[w.kind]) continue;
@@ -248,10 +243,26 @@ function reportRow({ sid, usage, requested, arm, cold, score, compact, rateLimit
     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, ts, effort: effort || null, effort_requested: effortAsked || null,
+    compressed_tokens: compressed > 0 ? compressed : null,
     surface: (surfaces || []).includes('desktop') ? 'desktop' : (surfaces || []).includes('terminal') ? 'terminal' : null,
   };
 }
 
+// Tokens the compression hook kept out of this session since the last report (it appends one
+// line per compressed tool output, see sandbox/posttooluse-hook.js). Each count is reported once.
+const savedSeen = new Map();
+async function compressedSince($, sid) {
+  try {
+    if (!/^[A-Za-z0-9-]{8,128}$/.test(String(sid))) return 0;
+    const home = await $.env.get('HOME');
+    const text = await $.fs.read(`${home}/.troxy/sandbox/saved/${sid}.log`);
+    const total = String(text).split('\n').reduce((n, l) => n + (Number(l) || 0), 0);
+    const delta = Math.max(0, total - (savedSeen.get(sid) || 0));
+    savedSeen.set(sid, total);
+    return delta;
+  } catch { return 0; }
+}
+
 async function noteStep($, result, meta) {
   const usage = result && result.usage;
   if (!usage || !usage.model) return;
@@ -259,7 +270,8 @@ async function noteStep($, result, meta) {
   try { rateLimits = (await $.session.usage()).rateLimits; } catch {}
   let surfaces = [];
   try { surfaces = await $.session.surfaces(); } catch {}
-  pending.push(reportRow({ ...meta, sid: await $.session.id(), usage, compact: compacting, rateLimits, surfaces, ts: await $.clock.now() }));
+  const sid = await $.session.id();
+  pending.push(reportRow({ ...meta, sid, usage, compact: compacting, rateLimits, surfaces, compressed: await compressedSince($, sid), ts: await $.clock.now() }));
   if (pending.length > REPORT_KEEP_ROWS) pending = pending.slice(-REPORT_KEEP_ROWS);
 }
 
@@ -278,10 +290,9 @@ async function flushReport($) {
 }
 
 // /troxy status: everything about this session's Troxy in one place.
-function statusText({ version, paused, totals, rateLimits = [], compactOff = false }) {
+function statusText({ version, paused, savings, rateLimits = [], compactOff = false }) {
   const lines = [`Troxy ${version || ''}`.trim() + ` · routing is ${paused ? 'off' : 'on'}`];
-  lines.push(sessionLine(totals) ? `This session: ${sessionLine(totals).replace(/^.*?This session: /, '')}`
-    : 'This session: nothing moved to a cheaper model yet');
+  lines.push(`This session: ${savings || 'nothing moved to a lighter model yet'}`);
   const meter = rateLimits.filter(w => WINDOW_LABEL[w.kind]).map(w => `${WINDOW_LABEL[w.kind]} ${Math.round(w.percentUsed)}%`);
   if (meter.length) lines.push(`Plan usage now: ${meter.join(' · ')}`);
   lines.push('The model is picked at the start, after an hour idle or after a compaction, then kept (switching would rewrite the cache). /model always wins.');
@@ -606,8 +617,7 @@ export function register(on) {
     await $.store.set('paused', r.paused);
     if (/^\s*(on|off)\s*$/i.test(String(e.args || ''))) return { text: r.text };
     const manifest = JSON.parse(await $.fs.read(`${$.plugin.root}/.claude-plugin/plugin.json`).catch(() => '{}'));
-    const { value: totals } = await $.state.get(SAVED);
-    return { text: statusText({ version: manifest.version, paused: r.paused, totals, rateLimits: (await $.session.usage()).rateLimits, compactOff: Boolean(await $.store.get('compactOff')) }) };
+    return { text: statusText({ version: manifest.version, paused: r.paused, savings: (await $.state.get(SAVINGS)).value, rateLimits: (await $.session.usage()).rateLimits, compactOff: Boolean(await $.store.get('compactOff')) }) };
   });
   // Routing fails open: a scorer or hook error must never break the session,
   // it just means the requested model is used as-is.
@@ -624,7 +634,7 @@ export function register(on) {
     const { value: kept } = await $.state.get(SAVINGS);
     if ((!line && !kept && !note) || e.props.hasSurvey) return next(e);
     const { Text } = $.ui.resolve(e);
-    return h(Text, { color: TROXY_TEAL }, `Troxy ${[line, kept, note && note.text].filter(Boolean).join(' · ')}`);
+    return h(Text, { color: TROXY_TEAL }, `Troxy · ${[kept, line, note && note.text].filter(Boolean).join(' · ')}`);
   });
   // After the answer, so counting never delays a prompt.
   on('turn.complete', async ($, e, next) => {
@@ -655,4 +665,4 @@ export function register(on) {
 }
 
 // Exported for tests only - not part of the mod's runtime contract.
-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 };
+export const _internal = { compressedSince, 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/sandbox/posttooluse-hook.js

@@ -38,6 +38,18 @@ function sessionArm(dir, sessionId, holdoutFraction, randomFn) {
   return arm;
 }
 
+// One line per compression: the tokens it kept out of the conversation (about 4 characters
+// each). The troxy-guard mod reads this per session and reports it, since in the mod-only setup
+// no proxy sees the traffic. Append-only so parallel hook processes never overwrite each other.
+export function recordSaved(dir, sessionId, tokens) {
+  if (typeof sessionId !== 'string' || !SESSION_ID_RE.test(sessionId) || !(tokens > 0)) return;
+  try {
+    const file = path.join(dir, 'saved', `${sessionId}.log`);
+    fs.mkdirSync(path.dirname(file), { recursive: true });
+    fs.appendFileSync(file, `${Math.round(tokens)}\n`);
+  } catch { /* fails open: a missed count never blocks the tool output */ }
+}
+
 function writeState(dir, state) {
   try {
     fs.mkdirSync(dir, { recursive: true });
@@ -117,6 +129,7 @@ export async function buildHookResponse(toolUseBlock, {
     const { handle } = await put(output, { summary, tags: [] }, storeDir);
     writeState(storeDir, 'treated');
     const updatedText = `${summary}\n\n(troxy_recall("${handle}") to fetch the full output)`;
+    recordSaved(storeDir, toolUseBlock?.session_id, (output.length - updatedText.length) / 4);
     return {
       hookSpecificOutput: {
         hookEventName: 'PostToolUse',

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

@@ -67,10 +67,10 @@ test('savedLine: no savings claim until measured, then the meter', () => {
   const { savedLine } = _internal;
   const meter = [{ kind: 'five_hour', percentUsed: 19 }, { kind: 'seven_day', percentUsed: 12.4 }];
   assert.equal(savedLine({ cost_paid_usd: 19, dollars_saved_usd: 13 }, meter),
-    'This session: savings under measurement · 5h: 19% used · 7d: 12% used');
+    '5h: 19% used · 7d: 12% used');
   assert.equal(savedLine({ cost_paid_usd: 92, dollars_saved_usd: 8 }, [{ kind: 'five_hour', percentUsed: 12 }, { kind: 'seven_day', percentUsed: 0 }]),
-    'This session: savings under measurement · 5h: 12% used');
-  assert.equal(savedLine({ cost_paid_usd: 1.64, dollars_saved_usd: 4.53 }, []), 'This session: savings under measurement');
+    '5h: 12% used');
+  assert.equal(savedLine({ cost_paid_usd: 1.64, dollars_saved_usd: 4.53 }, []), undefined);
   assert.equal(savedLine(null, meter), undefined);
   assert.doesNotMatch(savedLine({ cost_paid_usd: 10, dollars_saved_usd: 50 }, meter), /longer|instead of/);
 });
@@ -83,11 +83,11 @@ test('addTurn: a session total from its own turns, a routed one counted at the a
   // Not routed; the 5h window reset mid-turn, so its move counts as 0, not negative.
   t = addTurn(t, { cost: 1.5, meter: { five_hour: 13, seven_day: 5 } }, { cost: 2, meter: { five_hour: 1, seven_day: 6 } }, 1);
   assert.deepEqual(t, { paid: 1, full: 1.5, meter: { five_hour: 3, seven_day: 2 } });
-  assert.equal(sessionLine(t), 'This session: savings under measurement · 5h: 3% used · 7d: 2% used');
+  assert.equal(sessionLine(t), '5h: 3% used · 7d: 2% used');
   // Cleared/compressed content counts too, and survives the next turn: $1 paid, $0.5 swap + $1.5 kept = 3x.
   t = addTurn({ ...t, kept: 1.5 }, { cost: 2, meter: {} }, { cost: 2, meter: {} }, 1);
   assert.equal(t.kept, 1.5);
-  assert.match(sessionLine(t), /savings under measurement · 5h: 3% used/);
+  assert.match(sessionLine(t), /5h: 3% used/);
   // Server total of a session already 28 dollars deep: only growth after the first reading counts.
   const { withKept } = _internal;
   let k = { paid: 1, full: 1, meter: {} };
@@ -102,16 +102,15 @@ test('addTurn: a session total from its own turns, a routed one counted at the a
 
 test('statusText: version, routing, this session, plan usage, where the account view is', () => {
   const { statusText } = _internal;
-  const totals = { paid: 0.03, full: 0.12, meter: {} };
-  assert.equal(statusText({ version: '1.29.42', paused: false, totals, rateLimits: [{ kind: 'five_hour', percentUsed: 23.4 }, { kind: 'seven_day', percentUsed: 77 }] }),
+    assert.equal(statusText({ version: '1.29.42', paused: false, savings: '55% of this session ran on a lighter model', rateLimits: [{ kind: 'five_hour', percentUsed: 23.4 }, { kind: 'seven_day', percentUsed: 77 }] }),
     'Troxy 1.29.42 · routing is on\n'
-    + 'This session: savings under measurement\n'
+    + 'This session: 55% of this session ran on a lighter model\n'
     + 'Plan usage now: 5h 23% · 7d 77%\n'
     + 'The model is picked at the start, after an hour idle or after a compaction, then kept (switching would rewrite the cache). /model always wins.\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.');
-  assert.match(statusText({ paused: true }), /^Troxy · routing is off\nThis session: nothing moved to a cheaper model yet\n/);
+  assert.match(statusText({ paused: true }), /^Troxy · routing is off\nThis session: nothing moved to a lighter model yet\n/);
 });
 
 test('compactPlan: big idle sessions compact, small ones and a nearly full bar do not', () => {
@@ -128,11 +127,11 @@ 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: calls on a lighter model, kept tokens only when there are some, holdout says so', () => {
+test('savingsText: share of calls on a lighter model, holdout says so, nothing before the first call', () => {
   const { savingsText } = _internal;
-  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: 5, routed_calls: 5 }), '100% of this session ran on a lighter model');
+  assert.equal(savingsText({ calls: 1121, routed_calls: 622, trimmed_tokens: 35100000 }), '55% of this session ran on a lighter model');
+  assert.equal(savingsText({ calls: 9, routed_calls: 0 }), 'no calls moved to a lighter model yet');
   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);
@@ -440,3 +439,24 @@ test('reportRow: surface is desktop when the app draws the session, terminal oth
   assert.equal(reportRow({ ...base, surfaces: ['terminal'] }).surface, 'terminal');
   assert.equal(reportRow({ ...base, surfaces: [] }).surface, null);
 });
+
+test('reportRow carries the tokens compression kept out, null when none', () => {
+  const { reportRow } = _internal;
+  const usage = { model: 'claude-sonnet-5-5', input_tokens: 1, output_tokens: 1 };
+  assert.equal(reportRow({ sid: 's1', usage, compressed: 4200 }).compressed_tokens, 4200);
+  assert.equal(reportRow({ sid: 's1', usage, compressed: 0 }).compressed_tokens, null);
+  assert.equal(reportRow({ sid: 's1', usage }).compressed_tokens, null);
+});
+
+test('compressedSince reports each compression once, per session', async () => {
+  const { compressedSince } = _internal;
+  const sid = 'cccccccc-1111-2222-3333-444444444444';
+  let log = '1000\n500\n';
+  const $ = { env: { get: async () => '/home/u' }, fs: { read: async (p) => { assert.ok(p.endsWith(`/.troxy/sandbox/saved/${sid}.log`)); return log; } } };
+  assert.equal(await compressedSince($, sid), 1500);
+  assert.equal(await compressedSince($, sid), 0);
+  log += '250\n';
+  assert.equal(await compressedSince($, sid), 250);
+  assert.equal(await compressedSince({ ...$, fs: { read: async () => { throw new Error('no file'); } } }, 'dddddddd-1111-2222'), 0);
+  assert.equal(await compressedSince($, '../etc/passwd'), 0);
+});

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

@@ -223,3 +223,23 @@ test('session_id with path characters is ignored (no file outside the store)', a
     { holdoutFraction: 0.2, randomFn: () => 0, storeDir: dir });
   assert.equal(fs.existsSync(path.join(dir, 'sessions')), false);
 });
+
+test('a compression records the tokens it kept out, per session, one line each', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'troxy-sandbox-'));
+  const sid = 'aaaaaaaa-1111-2222-3333-444444444444';
+  const opts = { sandboxFraction: 1, holdoutFraction: 0, randomFn: () => 0.5, storeDir: dir };
+  const big = 'x'.repeat(8000);
+  await buildHookResponse({ session_id: sid, tool_response: bashResult(big) }, opts);
+  await buildHookResponse({ session_id: sid, tool_response: bashResult(big) }, opts);
+  const lines = fs.readFileSync(path.join(dir, 'saved', `${sid}.log`), 'utf8').trim().split('\n').map(Number);
+  assert.equal(lines.length, 2);
+  assert.ok(lines.every(n => n > 1000 && n <= 2000)); // about 8000 chars / 4, minus the excerpt
+});
+
+test('no session id or a small output records nothing', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'troxy-sandbox-'));
+  const opts = { sandboxFraction: 1, holdoutFraction: 0, randomFn: () => 0.5, storeDir: dir };
+  await buildHookResponse({ tool_response: bashResult('x'.repeat(8000)) }, opts);
+  await buildHookResponse({ session_id: 'bbbbbbbb-1111-2222-3333-444444444444', tool_response: bashResult('short') }, opts);
+  assert.equal(fs.existsSync(path.join(dir, 'saved')), false);
+});