@nexus-cortex/cli 4.124.11 → 4.124.13

package/.env.defaults

@@ -1135,6 +1135,32 @@ CORTEX_ENDTURN_GATE=true
 # [LEDGER 2026-09-20 @4.121.0] BUILT + unit-tested. FIELD: pending (cell g2).
 #CORTEX_JUDGE_SPEC_REPEAT_MAX=
 #CORTEX_JUDGE_SPEC_TESTS_AT=
+# CORTEX_JUDGE_REQ_LEDGER / _MAX / _HOLD_MAX — R187 HB-REQUIREMENT-LEDGER (4.124.12, 2026-09-23; DARK, default off). The consistency
+# lever the top-7 read named (tb4-consistency-top7-2026-09-23.md): five of the seven tasks flash passes ≥50% of the time fail on a
+# requirement the TASK TEXT STATES, and the failing run finishes EARLIER with more budget left, the resolver accepting — the writer's own
+# checks test the path it chose. At lift a mentor call (endturn-resolver wire, pro @ max, task text + env report ONLY) extracts each stated
+# requirement as `REQ n | kind | text | CHECK: cmd|NONE` (training/requirementLedger.ts); the writer sees the ledger with its plan of attack
+# ("each line must be EXERCISED before you finish"); at every finish the harness runs the checks — FAILED = R166 veto evidence (with the
+# spec/named checks), OPEN (never passed nor failed) holds a finish that would otherwise stand ONCE (_HOLD_MAX), naming the open lines and
+# their checks; UNVERIFIABLE (CHECK: NONE) is shown to the judge only. Never below CORTEX_JUDGE_VETO_MIN_REMAINING. Events:
+# requirement_ledger (lines), requirement_ledger_delivered, endturn_resolver += reqLines/reqOpen/reqExercised/reqFailed/reqUnverifiable/reqHeld.
+# Differs from R174 spec tests (commands only, no writer-facing lines, no hold on "unexercised"). Cell: the seven tasks (n≈18 each) first.
+# [LEDGER 2026-09-23 @4.124.13] LOCAL SMOKE PASSED (tools mode, g1c config + ledger on, toy CSV task with 4 stated requirements): 6 lines
+# authored in 13 s (1 non-read-only check refused → unverifiable), DELIVERED to the writer with the plan of attack (1919 chars), at the
+# finish 5 lines exercised by harness checks + 1 closed by Jev (0.3 s), finish accepted with 0 open; answer correct. 4.124.12 had the
+# ordering bug (ledger authored AFTER the plan delivery → never delivered to the writer; finish-time half worked) — fixed in 4.124.13.
+#CORTEX_JUDGE_REQ_LEDGER=
+#CORTEX_JUDGE_REQ_LEDGER_MAX=
+#CORTEX_JUDGE_REQ_LEDGER_HOLD_MAX=
+# CORTEX_JUDGE_REQ_LEDGER_JEV / _JEV_MIN — R187b (operator: "use jev in additional one-shot calls to verify the end turn meets the
+# requirements of the task"): E0 measured Jev as a WHOLE-task doneness judge at AUC 0.64 (weak; leg D, the Southbridge phase/forecast
+# read, no pass signal) but as a typed gate on concrete evidence at 0.98. The ledger makes doneness typed: after the shell checks, ONE Jev
+# call asks `met_<id>` per OPEN/UNVERIFIABLE line — "exercised AND met, judging only from the writer's attestation, the work-product tail and
+# the harness result" — and closes lines ≥ _JEV_MIN (0.7) with provenance `jev:`; the rest stay open (→ the hold). Jev never sees the judge's
+# verdict (gate independence). Default on when the ledger is on; needs TYPESAFE_API_KEY (fail-open to shell-only). Banked: reqJevAsked /
+# reqJevClosed / reqJevLatencyMs.
+#CORTEX_JUDGE_REQ_LEDGER_JEV=
+#CORTEX_JUDGE_REQ_LEDGER_JEV_MIN=
 # CORTEX_JUDGE_INDEPENDENT_DERIVATION / _TOL — R176 HB-INDEPENDENT-DERIVATION (4.122.0, 2026-09-21; design docs/R176_INDEPENDENT_DERIVATION_DESIGN.md).
 # The largest never-passed TB4.0 class (~14 tasks: foodstuff, roy-polymorph, glycan, gsea, hof, vigenere, sound-change, cargo, ctr,
 # shadow-relay, satb, takens…) is a computed value / exact artifact graded against a golden answer with NO ground truth in the container:
@@ -1224,6 +1250,9 @@ CORTEX_ENDTURN_GATE=true
 # options is handed back unexecuted unless the action carries `single_reason` (one sentence why only one action makes sense); bounded 5
 # re-asks/session; the author still widens afterwards. CORTEX_FRAME_AUTHOR_MODEL (unset = HELPER_MODEL_ID = flash): set deepseek-v4-pro for
 # pro-authored alternatives (thinking off). Banked: singleReason on frame_turn. E3 re-pinned 4.124.11.
+# [LEDGER 2026-09-23 @4.124.11] LOCAL SMOKE PASSED (E3 env, adapter-shaped request, toy CSV-sum): the writer gave 2 distinct options on every
+# turn under the new contract (no re-asks, no single_reason needed), the author added a 3rd, Jev's questions were pick_<id> only (5 keys), it
+# chose the AUTHOR's option on turn 3 (0.81, ran rc 0) and FINISH on turn 4 (0.62); sum correct in 4 turns.
 #CORTEX_FRAME_KEY_GUARD=
 #CORTEX_FRAME_MIN_CANDIDATES=
 #CORTEX_FRAME_AUTHOR=

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.11",
+  "version": "4.124.13",
   "description": "Nexus Cortex CLI - Terminal interface for multi-provider LLM orchestration",
   "main": "dist/index.js",
   "type": "module",
@@ -19,7 +19,7 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.124.11",
+    "@nexus-cortex/core": "4.124.13",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",