lynkr 9.14.16 → 9.14.18

package/config/routing.example.json

@@ -0,0 +1,186 @@
+{
+  "version": 1,
+  "mode": "observe",
+  "signals": {
+    "anchor": {
+      "type": "anchor_score"
+    },
+    "judge": {
+      "type": "jev",
+      "min_confidence": 0.5
+    },
+    "harness": {
+      "type": "harness"
+    },
+    "structured": {
+      "type": "request_field",
+      "any_of": [
+        "output_format",
+        "response_format"
+      ]
+    },
+    "phase": {
+      "type": "session_phase"
+    },
+    "prev": {
+      "type": "prev_outcome"
+    },
+    "forensics": {
+      "type": "keyword",
+      "terms": [
+        "recover",
+        "deleted",
+        "disk image",
+        "forensic",
+        "carve"
+      ],
+      "min_hits": 1
+    },
+    "complexity": {
+      "type": "complexity",
+      "threshold": 0.6,
+      "scale": 0.15
+    }
+  },
+  "decisions": [
+    {
+      "name": "harness_regressing_escalate",
+      "priority": 300,
+      "rules": {
+        "operator": "AND",
+        "conditions": [
+          {
+            "signal": "harness"
+          },
+          {
+            "signal": "prev",
+            "in": [
+              "regression",
+              "no_progress"
+            ]
+          },
+          {
+            "signal": "prev",
+            "min": 0
+          }
+        ]
+      },
+      "tier": "COMPLEX",
+      "effort": "medium"
+    },
+    {
+      "name": "harness_hard_task",
+      "priority": 200,
+      "rules": {
+        "operator": "AND",
+        "conditions": [
+          {
+            "signal": "harness"
+          },
+          {
+            "operator": "OR",
+            "conditions": [
+              {
+                "signal": "judge",
+                "tier_in": [
+                  "COMPLEX",
+                  "REASONING"
+                ]
+              },
+              {
+                "signal": "anchor",
+                "band_in": [
+                  "COMPLEX",
+                  "REASONING"
+                ]
+              },
+              {
+                "signal": "forensics"
+              },
+              {
+                "signal": "complexity",
+                "min": 0.6
+              }
+            ]
+          }
+        ]
+      },
+      "tier": "COMPLEX",
+      "effort": "medium",
+      "hosts": [
+        "DeepSeek"
+      ],
+      "plugins": {
+        "grounding": {
+          "mode": "observe",
+          "contradiction_threshold": 0.6
+        }
+      }
+    },
+    {
+      "name": "harness_easy_task",
+      "priority": 100,
+      "rules": {
+        "operator": "AND",
+        "conditions": [
+          {
+            "signal": "harness"
+          }
+        ]
+      },
+      "tier": "MEDIUM",
+      "effort": "low",
+      "hosts": [
+        "Friendli",
+        "Parasail"
+      ],
+      "plugins": {
+        "response_cache": false,
+        "format_guard": false,
+        "grounding": {
+          "mode": "observe",
+          "contradiction_threshold": 0.6
+        }
+      }
+    },
+    {
+      "name": "structured_output_no_mutation",
+      "priority": 50,
+      "rules": {
+        "operator": "AND",
+        "conditions": [
+          {
+            "signal": "structured"
+          }
+        ]
+      },
+      "tier": "from:legacy",
+      "plugins": {
+        "response_cache": false,
+        "format_guard": false
+      }
+    },
+    {
+      "name": "legacy",
+      "priority": 0,
+      "rules": {
+        "operator": "AND",
+        "conditions": []
+      },
+      "tier": "from:legacy"
+    }
+  ],
+  "switch_gate": {
+    "mode": "observe",
+    "escalate_after_regressions": 2,
+    "downgrade_after_recoveries": 3,
+    "min_turns_before_switch": 2,
+    "max_switches_per_session": 2,
+    "cooldown_turns": 2
+  },
+  "cascade": {
+    "mode": "observe",
+    "trigger_margin": 0.15,
+    "escalate_to": "COMPLEX"
+  }
+}

package/config/routing.json

@@ -0,0 +1,93 @@
+{
+  "version": 1,
+  "mode": "observe",
+  "anchor_bands": {
+    "SIMPLE": [
+      0,
+      20
+    ],
+    "MEDIUM": [
+      20,
+      51
+    ],
+    "COMPLEX": [
+      51,
+      76
+    ],
+    "REASONING": [
+      76,
+      101
+    ]
+  },
+  "harness": {
+    "patterns": [
+      {
+        "name": "terminus",
+        "preamble": "You are an AI assistant tasked with solving command-line tasks",
+        "instruction": "(?:^|\\n)Instruction:\\s*\\n([\\s\\S]*?)\\n\\s*\\n(?:Your response must be|Your response|Respond)"
+      }
+    ]
+  },
+  "signals": {
+    "anchor": {
+      "type": "anchor_score"
+    },
+    "judge": {
+      "type": "jev"
+    },
+    "harness": {
+      "type": "harness"
+    },
+    "structured": {
+      "type": "request_field",
+      "any_of": [
+        "output_format",
+        "response_format"
+      ]
+    },
+    "tool_count": {
+      "type": "tool_count"
+    },
+    "turn": {
+      "type": "session_turn"
+    },
+    "phase": {
+      "type": "session_phase"
+    },
+    "risk": {
+      "type": "risk"
+    },
+    "prev": {
+      "type": "prev_outcome"
+    },
+    "complexity": {
+      "type": "complexity",
+      "threshold": 0.6,
+      "scale": 0.15
+    }
+  },
+  "decisions": [
+    {
+      "name": "legacy",
+      "priority": 0,
+      "rules": {
+        "operator": "AND",
+        "conditions": []
+      },
+      "tier": "from:legacy"
+    }
+  ],
+  "switch_gate": {
+    "mode": "observe",
+    "escalate_after_regressions": 2,
+    "downgrade_after_recoveries": 3,
+    "min_turns_before_switch": 2,
+    "max_switches_per_session": 2,
+    "cooldown_turns": 2
+  },
+  "cascade": {
+    "mode": "observe",
+    "trigger_margin": 0.15,
+    "escalate_to": "COMPLEX"
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lynkr",
-  "version": "9.14.16",
+  "version": "9.14.18",
   "files": [
     "index.js",
     "install.sh",
@@ -38,7 +38,7 @@
     "start:supervised": "while true; do node index.js 2>&1 | npx pino-pretty --sync; code=$?; echo \"[supervisor] lynkr exited ($code) — restarting in 3s\"; sleep 3; done",
     "lint": "eslint src index.js",
     "test": "npm run test:unit && npm run test:performance",
-    "test:unit": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com LOG_FILE_ENABLED=false node --test test/routing.test.js test/hybrid-routing-integration.test.js test/retry-logic.test.js test/sse-transformer.test.js test/passthrough-stream.test.js test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js test/azure-openai-config.test.js test/azure-openai-format-conversion.test.js test/azure-openai-routing.test.js test/azure-openai-streaming.test.js test/azure-openai-error-resilience.test.js test/azure-openai-integration.test.js test/openai-integration.test.js test/atlas-integration.test.js test/toon-compression.test.js test/gcf-compression.test.js test/llamacpp-integration.test.js test/resilience.test.js test/telemetry-routing.test.js test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js test/distill.test.js test/large-payload.test.js test/prompt-cache-injection.test.js test/risk-analyzer.test.js test/interaction-block.test.js test/preflight.test.js test/token-reduction.test.js test/session-affinity.test.js test/cache-state.test.js test/cache-switch-cost.test.js test/lens-recommendations.test.js test/model-registry-cost.test.js test/output-format-guard.test.js test/tier-fallback.test.js test/wrap.test.js test/init.test.js test/tool-call-response-metadata.test.js test/degradation.test.js test/routing-telemetry-columns.test.js test/sticky-routing.test.js test/knn-ambiguous-escalate.test.js test/deescalator.test.js test/client-profiles.test.js test/strip-internal-fields.test.js test/complexity-tool-subtraction.test.js test/bandit.test.js test/routing-propensity.test.js test/reward-pipeline.test.js test/knn-cold-start.test.js test/calibration.test.js test/feedback-loop.test.js test/session-fingerprint.test.js test/side-request-guards.test.js test/verifier.test.js test/intent-score.test.js test/difficulty-classifier.test.js test/classifier-setup.test.js test/usage-stats.test.js test/loop-guard.test.js test/moonshot-model-mapping.test.js test/baidu-model-mapping.test.js test/tenant-policy-ingress-parity.test.js test/decide.test.js test/embeddings-degradation.test.js test/health-probe.test.js test/stuck-detector.test.js test/onnx-embedder.test.js test/ope.test.js test/hierarchical-budget.test.js test/token-rate-limit.test.js test/otel-export.test.js test/mcp-broker.test.js test/compression-budget.test.js test/gpt-utils.test.js test/dedup-observe-only.test.js test/context-window-header.test.js test/token-budget-auto.test.js test/opencode-setup.test.js test/auth-mode-first-party.test.js test/harness-envelope.test.js test/task-ledger.test.js test/jev-router.test.js test/jev-routing.test.js test/force-patterns.test.js test/upstream-fidelity.test.js test/tool-schema-compression.test.js test/passthrough-route.test.js test/quota-ledger.test.js",
+    "test:unit": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com LOG_FILE_ENABLED=false node --test test/routing.test.js test/hybrid-routing-integration.test.js test/retry-logic.test.js test/sse-transformer.test.js test/passthrough-stream.test.js test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js test/azure-openai-config.test.js test/azure-openai-format-conversion.test.js test/azure-openai-routing.test.js test/azure-openai-streaming.test.js test/azure-openai-error-resilience.test.js test/azure-openai-integration.test.js test/openai-integration.test.js test/atlas-integration.test.js test/toon-compression.test.js test/gcf-compression.test.js test/llamacpp-integration.test.js test/resilience.test.js test/telemetry-routing.test.js test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js test/distill.test.js test/large-payload.test.js test/prompt-cache-injection.test.js test/risk-analyzer.test.js test/interaction-block.test.js test/preflight.test.js test/token-reduction.test.js test/session-affinity.test.js test/cache-state.test.js test/cache-switch-cost.test.js test/lens-recommendations.test.js test/model-registry-cost.test.js test/output-format-guard.test.js test/tier-fallback.test.js test/wrap.test.js test/init.test.js test/tool-call-response-metadata.test.js test/degradation.test.js test/routing-telemetry-columns.test.js test/sticky-routing.test.js test/knn-ambiguous-escalate.test.js test/deescalator.test.js test/client-profiles.test.js test/strip-internal-fields.test.js test/complexity-tool-subtraction.test.js test/bandit.test.js test/routing-propensity.test.js test/reward-pipeline.test.js test/knn-cold-start.test.js test/calibration.test.js test/feedback-loop.test.js test/session-fingerprint.test.js test/side-request-guards.test.js test/verifier.test.js test/intent-score.test.js test/difficulty-classifier.test.js test/classifier-setup.test.js test/usage-stats.test.js test/loop-guard.test.js test/moonshot-model-mapping.test.js test/baidu-model-mapping.test.js test/tenant-policy-ingress-parity.test.js test/decide.test.js test/embeddings-degradation.test.js test/health-probe.test.js test/stuck-detector.test.js test/onnx-embedder.test.js test/ope.test.js test/hierarchical-budget.test.js test/token-rate-limit.test.js test/otel-export.test.js test/mcp-broker.test.js test/compression-budget.test.js test/gpt-utils.test.js test/dedup-observe-only.test.js test/context-window-header.test.js test/token-budget-auto.test.js test/opencode-setup.test.js test/auth-mode-first-party.test.js test/harness-envelope.test.js test/task-ledger.test.js test/jev-router.test.js test/jev-routing.test.js test/force-patterns.test.js test/upstream-fidelity.test.js test/tool-schema-compression.test.js test/passthrough-route.test.js test/quota-ledger.test.js test/shortfall-semantic.test.js test/routing-decisions.test.js test/routing-corpus.test.js test/switch-gate-grounding.test.js",
     "test:memory": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/memory/store.test.js test/memory/surprise.test.js test/memory/extractor.test.js test/memory/search.test.js test/memory/retriever.test.js test/memory/distiller.test.js test/memory/distiller-freeze.test.js test/memory/wiki.test.js test/memory/skills-cache.test.js test/memory/tencentdb-launcher.test.js",
     "test:new-features": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node --test test/passthrough-mode.test.js test/openrouter-error-resilience.test.js test/format-conversion.test.js",
     "test:performance": "LYNKR_KNN_DIR=/tmp/lynkr-test-knn DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/hybrid-routing-performance.test.js && DATABRICKS_API_KEY=test-key DATABRICKS_API_BASE=http://test.com node test/performance-tests.js",

package/scripts/build-complexity-exemplars.js

@@ -0,0 +1,84 @@
+#!/usr/bin/env node
+/**
+ * Build config/complexity-exemplars.json from measured task outcomes.
+ *
+ * "Hard" exemplars are task instructions the STRONG model failed in every
+ * solo run supplied; "easy" exemplars are instructions the CHEAP model passed.
+ * Everything else is left out (ambiguous). The complexity signal
+ * (signals.js, type: complexity) embeds both sets once and scores a request
+ * as sim(hard) − sim(easy), normalised to [0,1].
+ *
+ * Usage:
+ *   node scripts/build-complexity-exemplars.js \
+ *     --strong ~/tb-runs-auto-v2 [--strong ~/tb-runs-auto-full] \
+ *     --cheap ~/tb-runs-lynkr-v14 --cheap-model glm-5.3-flash \
+ *     [--out config/complexity-exemplars.json]
+ */
+'use strict';
+
+const fs = require('fs');
+const os = require('os');
+const path = require('path');
+
+function expand(p) { return p.startsWith('~') ? path.join(os.homedir(), p.slice(1)) : path.resolve(p); }
+function findRunRoot(dir) {
+  if (fs.existsSync(path.join(dir, 'results.json'))) return dir;
+  const subs = fs.readdirSync(dir).filter((d) => fs.existsSync(path.join(dir, d, 'results.json'))).sort().reverse();
+  return subs.length ? path.join(dir, subs[0]) : null;
+}
+function loadRun(root) {
+  const res = JSON.parse(fs.readFileSync(path.join(root, 'results.json'), 'utf8')).results || [];
+  const out = {};
+  for (const r of res) out[r.task_id] = { resolved: !!r.is_resolved, instruction: r.instruction || null };
+  return out;
+}
+function servedBy(root, task, needle) {
+  // share of episodes whose debug record names the model
+  const tdir = path.join(root, task); let trial;
+  try { trial = fs.readdirSync(tdir).find((x) => fs.existsSync(path.join(tdir, x, 'agent-logs'))); } catch { return 0; }
+  if (!trial) return 0;
+  const eps = fs.readdirSync(path.join(tdir, trial, 'agent-logs')).filter((e) => e.startsWith('episode-'));
+  let hit = 0, n = 0;
+  for (const e of eps) {
+    try { const s = JSON.parse(fs.readFileSync(path.join(tdir, trial, 'agent-logs', e, 'debug.json'), 'utf8')).original_response || ''; n++; if (s.toLowerCase().includes(needle.toLowerCase())) hit++; } catch { /* skip */ }
+  }
+  return n ? hit / n : 0;
+}
+
+function main() {
+  const argv = process.argv.slice(2);
+  const strong = [], cheap = []; let cheapModel = null, out = path.join(__dirname, '..', 'config', 'complexity-exemplars.json');
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === '--strong') strong.push(expand(argv[++i]));
+    else if (argv[i] === '--cheap') cheap.push(expand(argv[++i]));
+    else if (argv[i] === '--cheap-model') cheapModel = argv[++i];
+    else if (argv[i] === '--out') out = expand(argv[++i]);
+  }
+  if (!strong.length) { console.error('need at least one --strong run dir'); process.exit(2); }
+  const strongRuns = strong.map(findRunRoot).filter(Boolean).map(loadRun);
+  const cheapRoots = cheap.map(findRunRoot).filter(Boolean);
+  const cheapRuns = cheapRoots.map(loadRun);
+  const tasks = new Set(); strongRuns.forEach((r) => Object.keys(r).forEach((t) => tasks.add(t)));
+  const hard = [], easy = [], skipped = [];
+  for (const t of tasks) {
+    const instr = strongRuns.map((r) => r[t]?.instruction).find(Boolean); if (!instr) continue;
+    const strongFailedAll = strongRuns.every((r) => r[t] && !r[t].resolved);
+    const strongPassedAny = strongRuns.some((r) => r[t] && r[t].resolved);
+    let cheapPassed = false;
+    cheapRuns.forEach((r, i) => { if (r[t]?.resolved && (!cheapModel || servedBy(cheapRoots[i], t, cheapModel) >= 0.8)) cheapPassed = true; });
+    if (strongFailedAll) hard.push({ task: t, text: instr.slice(0, 1200) });
+    else if (cheapPassed) easy.push({ task: t, text: instr.slice(0, 1200) });
+    else if (strongPassedAny) skipped.push(t);
+  }
+  const heads = ['reasoning', 'codegen', 'debugging', 'tool_use'];
+  const doc = {
+    generated: new Date().toISOString(),
+    note: 'hard = strong model failed every supplied solo run; easy = cheap model passed. Edit freely; the signal embeds these at load.',
+    // One shared pool per side; per-head pools can be curated later.
+    heads: Object.fromEntries(heads.map((h) => [h, { hard: hard.map((x) => x.text), easy: easy.map((x) => x.text) }])),
+    provenance: { hard: hard.map((x) => x.task), easy: easy.map((x) => x.task), ambiguous: skipped },
+  };
+  fs.writeFileSync(out, JSON.stringify(doc, null, 1));
+  console.log(`hard ${hard.length}, easy ${easy.length}, ambiguous ${skipped.length} → ${out}`);
+}
+main();

package/scripts/build-routing-corpus.js

@@ -0,0 +1,55 @@
+#!/usr/bin/env node
+/**
+ * Build / refresh the routing corpus fixtures used by test/routing-corpus.test.js.
+ *
+ * For every recorded request (a directory tree of Terminal-Bench runs, or any
+ * directory of *.json request bodies), run the full routing decision under
+ * the given config and store the signal snapshot plus the decision reached.
+ *
+ * Usage:
+ *   node scripts/build-routing-corpus.js --from <run-dir|requests-dir> \
+ *     --config config/routing.example.json --out test/fixtures/routing-corpus/<name>.json
+ *
+ * Needs the operator .env (embeddings, judge key) like the server does.
+ */
+'use strict';
+
+const fs = require('fs');
+const os = require('os');
+const path = require('path');
+
+function arg(name, dflt) { const i = process.argv.indexOf(name); return i >= 0 ? process.argv[i + 1] : dflt; }
+
+async function main() {
+  const from = arg('--from'); const cfgPath = arg('--config', 'config/routing.example.json'); const out = arg('--out');
+  if (!from || !out) { console.error('usage: --from <dir> --config <routing.json> --out <fixture.json>'); process.exit(2); }
+  const envPath = fs.existsSync(path.join(process.cwd(), '.env')) ? path.join(process.cwd(), '.env') : path.join(os.homedir(), '.env');
+  require('dotenv').config({ path: envPath });
+  process.env.LOG_LEVEL = 'silent'; process.env.LOG_FILE_ENABLED = 'false';
+  process.env.LYNKR_ROUTING_CONFIG = path.resolve(cfgPath);
+  const { determineProviderSmart } = require('../src/routing');
+
+  // Collect requests: Terminal-Bench layout (<task>/<trial>/agent-logs/episode-0/prompt.txt) or *.json bodies.
+  const requests = [];
+  for (const entry of fs.readdirSync(from)) {
+    const p = path.join(from, entry);
+    if (entry.endsWith('.json') && fs.statSync(p).isFile()) { requests.push({ name: entry.replace(/\.json$/, ''), payload: JSON.parse(fs.readFileSync(p, 'utf8')) }); continue; }
+    if (!fs.statSync(p).isDirectory()) continue;
+    const trial = fs.readdirSync(p).find((x) => fs.existsSync(path.join(p, x, 'agent-logs', 'episode-0', 'prompt.txt')));
+    if (trial) requests.push({ name: entry, payload: { messages: [{ role: 'user', content: fs.readFileSync(path.join(p, trial, 'agent-logs', 'episode-0', 'prompt.txt'), 'utf8') }], tools: [] } });
+  }
+  const fixtures = []; const dist = {};
+  for (const r of requests) {
+    const d = await determineProviderSmart(r.payload, {});
+    const e = d.engine; if (!e) continue;
+    const signals = {};
+    for (const [k, v] of Object.entries(e.signals)) signals[k] = { matched: v.matched, value: (v.value && typeof v.value === 'object') ? null : v.value, confidence: v.confidence, band: v.band, tier: v.tier };
+    fixtures.push({ task: r.name, legacyTier: d.tier, signals, expected: { decision: e.decision, tier: e.tier, effort: e.effort } });
+    dist[e.decision] = (dist[e.decision] || 0) + 1;
+  }
+  fs.mkdirSync(path.dirname(out), { recursive: true });
+  fs.writeFileSync(out, JSON.stringify({ generated: new Date().toISOString(), config: path.relative(process.cwd(), path.resolve(cfgPath)), fixtures }, null, 1));
+  console.log(`wrote ${fixtures.length} fixtures → ${out}; decisions: ${JSON.stringify(dist)}`);
+}
+
+main().catch((e) => { console.error(e); process.exit(1); });

package/scripts/cache-probe.js

@@ -0,0 +1,66 @@
+#!/usr/bin/env node
+/**
+ * Prompt-cache probe: does each configured OpenRouter host actually serve
+ * cache hits, and what does warm vs cold cost in time?
+ *
+ * Sends the same ~3k-token prefix twice to every (model, host) pair from
+ * OPENROUTER_PROVIDER_ORDER_MAP (or --model/--host), reports cached_tokens,
+ * cache price, TTFT-ish latency for both calls, and the provider-billed cost.
+ *
+ * Usage: node scripts/cache-probe.js [--model <id> --host <name>]... [--prefix-tokens 3000]
+ */
+'use strict';
+
+const fs = require('fs');
+const path = require('path');
+const os = require('os');
+
+async function main() {
+  const envPath = fs.existsSync(path.join(process.cwd(), '.env')) ? path.join(process.cwd(), '.env') : path.join(os.homedir(), '.env');
+  require('dotenv').config({ path: envPath });
+  const KEY = process.env.OPENROUTER_API_KEY;
+  if (!KEY) { console.error('OPENROUTER_API_KEY not set'); process.exit(2); }
+  const argv = process.argv.slice(2);
+  const pairs = [];
+  for (let i = 0; i < argv.length; i++) if (argv[i] === '--model') pairs.push({ model: argv[++i], host: argv[i + 1] === '--host' ? argv[(i += 2)] : null });
+  if (!pairs.length) {
+    const map = (process.env.OPENROUTER_PROVIDER_ORDER_MAP || '').replace(/^"|"$/g, '');
+    for (const entry of map.split(';')) {
+      const eq = entry.indexOf('='); if (eq < 0) continue;
+      const k = entry.slice(0, eq).trim(); const hosts = entry.slice(eq + 1).split(',').map((x) => x.trim()).filter(Boolean);
+      // Resolve the short model key to a configured TIER_* model id.
+      const tiers = ['TIER_SIMPLE', 'TIER_MEDIUM', 'TIER_COMPLEX', 'TIER_REASONING'].map((t) => process.env[t] || '').filter((v) => v.startsWith('openrouter:') && v.includes(k));
+      const model = tiers.length ? tiers[0].slice('openrouter:'.length).split(/\s+#/)[0].trim() : k;
+      for (const h of hosts) pairs.push({ model, host: h });
+    }
+  }
+  const pi = argv.indexOf('--prefix-tokens'); const prefixTokens = pi >= 0 ? Number(argv[pi + 1]) || 3000 : 3000;
+  const filler = Array.from({ length: Math.round(prefixTokens / 12) }, (_, i) => `-rw-r--r-- 1 root root ${100000 + i * 37} Oct  2 05:00 file_${i}.log`).join('\n');
+  const nonce = Date.now();
+  const prefix = `[cache-probe ${nonce}]\nYou are an AI assistant tasked with solving command-line tasks in a Linux environment.\n\nInstruction:\nSummarise the directory listing below in one sentence.\n\n${filler}\n\n`;
+
+  const call = async (model, host, suffix) => {
+    const body = { model, messages: [{ role: 'user', content: prefix + suffix }], max_tokens: 64, temperature: 0, usage: { include: true } };
+    if (host) body.provider = { order: [host], allow_fallbacks: false };
+    const t0 = Date.now();
+    const res = await fetch('https://openrouter.ai/api/v1/chat/completions', { method: 'POST', headers: { Authorization: `Bearer ${KEY}`, 'Content-Type': 'application/json' }, body: JSON.stringify(body) });
+    const ms = Date.now() - t0; const j = await res.json();
+    if (j.error) return { error: j.error.message?.slice(0, 80), ms };
+    const u = j.usage || {};
+    return { ms, provider: j.provider, prompt: u.prompt_tokens, cached: u.prompt_tokens_details?.cached_tokens ?? 0, cost: u.cost ?? null };
+  };
+
+  console.log(`prefix ≈ ${prefixTokens} tokens, two calls per host\n`);
+  console.log(`${'model'.padEnd(34)} ${'host'.padEnd(12)} ${'cold ms'.padStart(8)} ${'warm ms'.padStart(8)} ${'cached'.padStart(7)} ${'hit%'.padStart(5)} ${'cold $'.padStart(9)} ${'warm $'.padStart(9)}  note`);
+  for (const { model, host } of pairs) {
+    const a = await call(model, host, 'Reply with one word: ready.');
+    await new Promise((r) => setTimeout(r, 1500));
+    const b = await call(model, host, 'Reply with one word: again.');
+    if (a.error || b.error) { console.log(`${model.padEnd(34)} ${String(host).padEnd(12)} ERROR ${a.error || b.error}`); continue; }
+    const hit = b.prompt ? Math.round(100 * b.cached / b.prompt) : 0;
+    const note = b.cached === 0 ? 'NO CACHE HITS REPORTED' : (b.ms < a.ms * 0.8 ? 'warm faster' : 'no latency gain');
+    console.log(`${model.padEnd(34)} ${String(host).padEnd(12)} ${String(a.ms).padStart(8)} ${String(b.ms).padStart(8)} ${String(b.cached).padStart(7)} ${String(hit).padStart(5)} ${a.cost != null ? a.cost.toFixed(6).padStart(9) : '        -'} ${b.cost != null ? b.cost.toFixed(6).padStart(9) : '        -'}  ${note}`);
+  }
+}
+
+main().catch((e) => { console.error(e); process.exit(1); });