troxy-cli 1.30.11 → 1.30.13

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "troxy-cli",
-  "version": "1.30.11",
+  "version": "1.30.13",
   "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

@@ -74,7 +74,7 @@ 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_ESCALATE_ABOVE = 0.3; // job B's bug and fix prompts scored 0.4 to 0.5 and Sonnet won them 8 to 2; at 0.6 they stayed on Haiku
 const HAIKU_MAX_TOKENS = 100000;
 const ROUTE_CONTEXT_FLOOR_TOKENS = 50000;
 const SCORE_TIMEOUT_MS = 3000;

package/src/sandbox/posttooluse-hook.js

@@ -107,6 +107,23 @@ function rebuildWithText(toolResponse, text) {
   return { ...toolResponse, content: text };
 }
 
+// Only outputs Claude rarely fetches back get shortened. Measured 2026-10-23 on 1,435 shortened
+// outputs: Grep 0% recalled; wc, ls, head, tail, test runs and logs 0-15%; Read of .js/.html
+// 17-20%. Targeted reads (sed, cat, Bash grep, Read of .py/.md/.txt) came back 30-86%, so
+// shortening them only adds a recall step.
+const SHORTEN_COMMANDS = new Set(['wc', 'ls', 'head', 'tail', 'ps', 'gh', 'aws', 'pytest']);
+const SHORTEN_READ_EXT = /\.(js|html)$/i;
+
+export function isShortenable(toolName, toolInput) {
+  if (toolName === undefined) return true; // ponytail: payloads without tool_name (old tests); Claude Code always sends it
+  if (toolName === 'Grep') return true;
+  if (toolName === 'Read') return SHORTEN_READ_EXT.test(String(toolInput?.file_path || ''));
+  if (toolName !== 'Bash') return false;
+  const cmd = String(toolInput?.command || '').trim().replace(/^(cd\s+\S+\s*(&&|;)\s*)+/, '');
+  if (/^python3?\s+-m\s+pytest\b/.test(cmd)) return true;
+  return SHORTEN_COMMANDS.has(path.basename(cmd.split(/\s+/)[0] || ''));
+}
+
 export async function buildHookResponse(toolUseBlock, {
   sandboxFraction = 1,
   holdoutFraction = 0.2,
@@ -124,7 +141,7 @@ export async function buildHookResponse(toolUseBlock, {
     return undefined;
   }
   const output = extractText(toolResponse);
-  if (output === null || !shouldSandbox(output)) {
+  if (output === null || !shouldSandbox(output) || !isShortenable(toolUseBlock?.tool_name, toolUseBlock?.tool_input)) {
     writeState(storeDir, 'skipped');
     return undefined;
   }

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

@@ -107,7 +107,8 @@ test('stepModel: a Haiku session moves up to Sonnet on a hard prompt or a long c
   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: 0.4, warm: MODELS.haiku }), MODELS.sonnet);       // harder prompt: up
+  assert.equal(stepModel({ ...base, cold: false, score: 0.3, warm: MODELS.haiku }), MODELS.haiku);       // 0.3 or lower: hold
   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

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

@@ -271,3 +271,25 @@ test('posttooluse-hook is a no-op unless TROXY_COMPRESS=1', () => {
     fs.rmSync(tmpHome, { recursive: true, force: true });
   }
 });
+
+test('isShortenable: only output kinds Claude rarely fetches back', async () => {
+  const { isShortenable } = await import('../sandbox/posttooluse-hook.js');
+  assert.equal(isShortenable('Grep', {}), true);
+  assert.equal(isShortenable('Read', { file_path: '/a/b/app.js' }), true);
+  assert.equal(isShortenable('Read', { file_path: '/a/b/notes.md' }), false);
+  assert.equal(isShortenable('Read', { file_path: '/a/b/x.py' }), false);
+  assert.equal(isShortenable('Bash', { command: 'cd /repo && ls -la src' }), true);
+  assert.equal(isShortenable('Bash', { command: 'python3 -m pytest -q' }), true);
+  assert.equal(isShortenable('Bash', { command: 'sed -n 1,200p file.js' }), false);
+  assert.equal(isShortenable('Bash', { command: 'cat big.log' }), false);
+  assert.equal(isShortenable('mcp__claude_ai_Gmail__get_thread', {}), false);
+});
+
+test('a large sed output passes through unshortened', async () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'troxy-sandbox-'));
+  const response = await buildHookResponse(
+    { tool_name: 'Bash', tool_input: { command: 'sed -n 1,400p a.js' }, tool_response: bashResult('x'.repeat(5000)) },
+    { sandboxFraction: 1, holdoutFraction: 0, randomFn: () => 0.5, storeDir: dir });
+  assert.equal(response, undefined);
+  assert.equal(readState(dir).state, 'skipped');
+});