@nexus-cortex/server 4.124.64 → 4.124.66

package/.cortex/skills/cortex-bench/SKILL.md

@@ -85,7 +85,7 @@ pkill -9 -f "node dist/index.js" 2>/dev/null; sleep 2
 ps -eo pid,args | grep "[d]ist/index.js" && echo "ZOMBIE — kill -9" || echo "clean"
 
 # Start stateless (CRITICAL — persistent mode leaks context across probes)
-cd /home/runner/workspace/omniclaude-v4/packages/server && \
+cd "$CORTEX_ROOT/packages/server" && \   # CORTEX_ROOT = the nexus-cortex checkout
   DEBUG=true \
   MENTORSHIP_ENABLED=false \
   ENABLE_SERVER_SIDE_TOOLS=true \
@@ -358,7 +358,7 @@ Most benchmarks compare JSON from `/v1/messages` — but that **cannot see what
 **Use the native `TmuxSession` tool** (proven method — same one used during the earlier active-improvement rounds; it wraps `TmuxCapture.ts`/`TmuxManager.ts`, stores metadata in `.cortex/tmux-sessions/`, binary via `TMUX_BIN`). The tool exposes **create → send commands → capture output → list → kill**. The agentic flow:
 
 1. `TmuxSession create` — a persistent session sized like a real terminal (e.g. 200×50; also test 80×24 — different widths expose different wrapping/overflow bugs).
-2. `TmuxSession send` → `cd /home/runner/workspace/omniclaude-v4 && neoncortex` (Ink UI; or `cortex-cli`). Wait for boot.
+2. `TmuxSession send` → `cd "$CORTEX_ROOT" && neoncortex` (Ink UI; or `cortex-cli`). Wait for boot.
 3. `TmuxSession send` → the prompt that exercises the rendering under test (e.g. *"Show a markdown table of the 5 cheapest models with a fenced code example"*). Wait for the stream to finish.
 4. `TmuxSession capture` — **this captured pane IS the metric.** It's the human-visible truth; the raw model text is NOT (a model can emit perfect markdown the TUI then mangles — only the capture shows it).
 5. Evaluate the capture for **frontend-design deficiencies**, treating the render as a real UI surface (hierarchy, alignment, contrast, density — not just "did text appear"):
@@ -390,10 +390,10 @@ Then compare all 4 answers side-by-side and report discrepancies.
 
 ## Location Reference
 
-- Server: `/home/runner/workspace/omniclaude-v4/packages/server/dist/index.js`
+- Server: `$CORTEX_ROOT/packages/server/dist/index.js`
 - Server log: `/tmp/omniclaude-server.log`
-- Bench results / deficiency ledgers: `/home/runner/workspace/omniclaude-v4/.cortex/bench/`
-- Routing matrix (the metric store): `/home/runner/workspace/omniclaude-v4/.cortex/router-matrix.jsonl`
+- Bench results / deficiency ledgers: `$CORTEX_ROOT/.cortex/bench/`
+- Routing matrix (the metric store): `$CORTEX_ROOT/.cortex/router-matrix.jsonl`
 - **Cortex skill (load it — full tool/subsystem reference incl. `TmuxSession`, `WorkspaceManager`):** `.agents/skills/cortex/SKILL.md`
 - **Worktree/team proven pattern:** `WorkspaceManagerTool` (core: `tools/definitions/`, executor: `executors/.../execution/`) + memory `agent-team-workspace`
 - **TUI capture proven pattern:** `TmuxSession` tool, `TmuxCapture.ts` / `TmuxManager.ts`, sessions in `.cortex/tmux-sessions/`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/server",
-  "version": "4.124.64",
+  "version": "4.124.66",
   "description": "Thin Express server wrapper for Nexus Cortex core library",
   "type": "module",
   "main": "dist/index.js",
@@ -20,8 +20,8 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.124.64",
-    "@nexus-cortex/executors": "4.124.64",
+    "@nexus-cortex/core": "4.124.66",
+    "@nexus-cortex/executors": "4.124.66",
     "chalk": "^5.3.0",
     "cors": "^2.8.5",
     "dotenv": "^16.4.5",

package/.cortex/bench/tasks/p3-anchor-holdout.json

@@ -1,14 +0,0 @@
-[
-  {
-    "id": "p3h-impl-intervals",
-    "taskType": "T2",
-    "prompt": "In /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p3-anchor, create src/intervals.js exporting mergeIntervals(list) for [start,end] integer pairs: sort by start and merge OVERLAPPING intervals only — intervals that merely TOUCH (end === next start) must NOT be merged. Throw a TypeError if any pair has end < start. Create src/intervals.test.js (plain Node) printing 'PASS <label>'/'FAIL <label>' per check, covering: [[1,5],[4,8]] -> [[1,8]]; [[1,5],[5,8]] -> [[1,5],[5,8]] (touching stays split); [[3,4],[1,2],[2,3]] -> [[1,2],[2,3],[3,4]]; [[1,0]] throws TypeError. Run it and reply with the full test output.",
-    "verifier": { "type": "regex", "pattern": "PASS[\\s\\S]*PASS[\\s\\S]*PASS[\\s\\S]*PASS" }
-  },
-  {
-    "id": "p3h-trace-canontools",
-    "taskType": "T3",
-    "prompt": "In this repository (omniclaude-v4), for the canon cross-harness tool-mapping layer: (1) which exported constant maps canonical tool concepts to per-harness tool names, (2) which exported type enumerates the harness names, and (3) what is the EXACT grok-build tool name mapped to the edit_file concept? Cite exact identifiers.",
-    "verifier": { "type": "contains", "all": ["TOOL_CONCEPTS", "HarnessName", "search_replace"], "caseInsensitive": false }
-  }
-]

package/.cortex/bench/tasks/p3-anchor-train.json

@@ -1,32 +0,0 @@
-[
-  {
-    "id": "p3-fix-ledger",
-    "taskType": "T2",
-    "prompt": "The Node project at /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p3-anchor has MULTIPLE independent bugs. First run `git -C /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p3-anchor checkout -- .` to reset it. Read README.md for the business rules, then find and fix ALL bugs in src/ledger.js so `node src/run.js` prints the correct summary for the data in src/run.js. Do not modify src/run.js or the data. Actually edit the file, run it to verify, and reply with the exact final JSON line it prints.",
-    "verifier": { "type": "contains", "all": ["\"first\":\"bolts\"", "\"latest\":\"12\"", "\"page1Count\":3", "19.15"], "caseInsensitive": false }
-  },
-  {
-    "id": "p3-impl-duration",
-    "taskType": "T2",
-    "prompt": "In /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p3-anchor, create src/duration.js exporting parseDuration(str) -> total seconds. Rules: str is 1-3 components in strict h,m,s ORDER (e.g. '2h45m10s', '90m', '1h5s'); integers only — reject decimals like '1.5h'; reject unknown units, wrong order (e.g. '10s2h'), empty string, and repeated units, all by throwing a TypeError. Then create src/duration.test.js using plain Node asserts that prints one line per check, 'PASS <label>' or 'FAIL <label>', covering: parseDuration('2h45m10s')===9910, parseDuration('90m')===5400, parseDuration('1h5s')===3605, '1.5h' throws TypeError, '10s2h' throws TypeError. Run it and reply with the full test output.",
-    "verifier": { "type": "regex", "pattern": "PASS[\\s\\S]*PASS[\\s\\S]*PASS[\\s\\S]*PASS[\\s\\S]*PASS" }
-  },
-  {
-    "id": "p3-trace-gemini-thinking",
-    "taskType": "T3",
-    "prompt": "In this repository (omniclaude-v4), trace how tunable Gemini thinking is plumbed to the wire. Report: (1) the exact helper method used to write the nested request param, (2) the exact dotted param path written into the request, (3) the exact model-card field whose value rides that path. Cite exact identifiers from the code.",
-    "verifier": { "type": "contains", "all": ["setNestedParam", "generationConfig.thinkingConfig.thinkingLevel", "defaultEffort"], "caseInsensitive": false }
-  },
-  {
-    "id": "p3-trace-searchtools-widening",
-    "taskType": "T3",
-    "prompt": "In this repository (omniclaude-v4), for the CLI harness deferred-tool path: after the model calls SearchTools, (1) which ClientSideToolFilter method records the discovered tool names, (2) which filter method does the orchestrator re-run before the continuation request so the discovered schemas enter the tools array, and (3) if the model calls a standard-tier tool COLD (never discovered via SearchTools), is the call executed or refused at dispatch? Answer (3) with exactly one word: 'executed' or 'refused'. Cite exact identifiers for (1) and (2).",
-    "verifier": { "type": "contains", "all": ["recordDiscoveredTools", "getFilteredTools", "executed"], "caseInsensitive": true }
-  },
-  {
-    "id": "p3-count-essential",
-    "taskType": "T3",
-    "prompt": "In this repository, count EXACTLY how many tool registrations in packages/core/src/tools/registries/BaseToolRegistry.ts carry discoveryTier: 'essential'. Be exhaustive — do not estimate. Reply with a line of the form 'COUNT: <n>'.",
-    "verifier": { "type": "contains", "all": ["COUNT: 14"], "caseInsensitive": true }
-  }
-]

package/.cortex/bench/tasks/tool-profile-p2-holdout.json

@@ -1,14 +0,0 @@
-[
-  {
-    "id": "p2h-refactor-discount",
-    "taskType": "T2",
-    "prompt": "In the project at /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p2-tool-profile (run `git -C /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p2-tool-profile checkout -- .` first, then re-apply the known fixes: loop bound < and pct/100), refactor src/inventory.js to add a new exported function netTotal(items, pct) that composes totalValue and applyDiscount, update src/report.js to ALSO print {\"net\": netTotal(items, 20)} on a second line, run it, and reply with both output lines.",
-    "verifier": { "type": "contains", "all": ["\"discounted\":40", "\"net\":40"], "caseInsensitive": true }
-  },
-  {
-    "id": "p2h-harness-usage-trace",
-    "taskType": "T3",
-    "prompt": "In this repository, packages/core/src/adapters/GatewayTranslationLayer.ts extracts token usage. Report: (1) the exact interface field that carries tokens read from cache, (2) which provider convention EXCLUDES cached tokens from its reported input_tokens (name the provider), (3) the field name added 2026-08 that carries reasoning token counts. Cite exact identifiers.",
-    "verifier": { "type": "contains", "all": ["cacheReadTokens", "anthropic", "reasoningTokens"], "caseInsensitive": true }
-  }
-]

package/.cortex/bench/tasks/tool-profile-p2-train.json

@@ -1,26 +0,0 @@
-[
-  {
-    "id": "p2-debug-fix-inventory",
-    "taskType": "T2",
-    "prompt": "The Node project at /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p2-tool-profile has a bug: `node src/report.js` crashes. First run `git -C /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p2-tool-profile checkout -- .` to reset the fixture. Then find and FIX all bugs in src/inventory.js so report.js prints the mathematically correct result for the data in src/report.js (a 20% discount means the customer pays 80%). Actually edit the file, run report.js to verify, and reply with the final JSON line it prints.",
-    "verifier": { "type": "contains", "all": ["\"total\":50", "\"discounted\":40"], "caseInsensitive": true }
-  },
-  {
-    "id": "p2-multifile-trace",
-    "taskType": "T3",
-    "prompt": "In this repository, trace the CORTEX_TOOL_PROFILE feature end to end and report: (1) the exact function name that filters a tool list by profile, (2) the file where the filter is applied inside getAllTools, (3) the two tool names that are ALWAYS retained regardless of profile, (4) the field name stamped into decision records for the active profile. Cite exact identifiers.",
-    "verifier": { "type": "contains", "all": ["applyToolProfile", "ToolFactory", "EndTurn", "AskUserQuestion", "toolProfile"], "caseInsensitive": false }
-  },
-  {
-    "id": "p2-write-tested-util",
-    "taskType": "T2",
-    "prompt": "In the project at /home/runner/workspace/omniclaude-v4/.cortex/bench/fixtures/p2-tool-profile, create a new file src/median.js exporting a function median(nums) that returns the median of a numeric array (average of two middles for even length; throw a TypeError on empty input). Then create src/median.test.js that requires it and asserts median([1,3,2])===2, median([1,2,3,4])===2.5, and that median([]) throws — printing PASS for each assertion that holds and FAIL otherwise, using plain Node (no test framework). Run it with node and reply with the full test output.",
-    "verifier": { "type": "regex", "pattern": "PASS[\\s\\S]*PASS[\\s\\S]*PASS" }
-  },
-  {
-    "id": "p2-cross-module-count",
-    "taskType": "T3",
-    "prompt": "Across packages/core/src/training in this repository, determine: (1) which file computes the FWER-adjusted alpha threshold and the exact function name, (2) which file auto-stamps experimentTag from CORTEX_EXPERIMENT_TAG into benchmark records, (3) the name of the verdict field that reports distinct task families. Cite exact identifiers.",
-    "verifier": { "type": "contains", "all": ["mcFwerThreshold", "AutoResearchStats", "ModelRouterMatrix", "familiesCovered"], "caseInsensitive": false }
-  }
-]