@nexus-cortex/cli 4.124.10 → 4.124.12

package/.env.defaults

@@ -1135,6 +1135,28 @@ 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.
+#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:
@@ -1211,9 +1233,26 @@ CORTEX_ENDTURN_GATE=true
 # guard and named-template resolution like writer candidates, and Jev picks among writer + author + templates; the escape stays the writer's c1.
 # With an author present the R184 re-ask does not fire. Banked per frame_turn: authored, authorMs, author; meta.pickSource. Cost ≈ one helper
 # call per turn where the writer gave one candidate (89% of E2 turns): ~$0.0005 and ~3 s each. Cell E3 = this + chooser on, pin 4.124.10.
+# [LEDGER 2026-09-22 @4.124.10] LOCAL SMOKE PASSED (E3 cell env, adapter-shaped /v1/messages, toy CSV-sum task): 6 frame turns, author fired on
+# every turn (2 alternatives, 1.0–1.6 s each), Jev chose among 3 real candidates each turn (c1 at 0.62–0.82, then FINISH at 0.87), sum correct,
+# ~3.5–6 s per step. Smoke trap: a stale server held :4187 and the fresh server moved to :4188 — four "no FrameAction" runs were the stale box;
+# read the port from the server log. The orchestrator now logs `[Frame] tools offered (n): …` in frame mode (engagement evidence).
+# [LEDGER 2026-09-23 @4.124.11] R186 CHOOSER BACK TO ITS ROLE (operator: "the chooser is drifting from the core concept of jev in this role").
+# REPLAY of the E2 rows (3652 Jev-answered turns): the gate questions changed the action 56× (1.5%) — repeat 55 (a fact code already writes
+# on the option), unsafe 1, unaddressed_error 4 of 382 fires. REMOVED from the chooser: repeat / unsafe / questionable / unaddressed_error /
+# intended_<id> nouls and the restrict/refuse branches. Jev now answers ONE question per option ("is this the one to press"); code facts
+# (already run n×, DESTRUCTIVE: …) are written onto the option text; safety = the keystroke guard in code; a soft-destructive option runs only
+# with the writer's stated `why` in BOTH chooser modes. The WRITER OWES THE MENU (operator): fewer than CORTEX_FRAME_MIN_CANDIDATES distinct
+# 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=
+#CORTEX_FRAME_AUTHOR_MODEL=
 # (no lever) R169 HB-NEARDUP-SCRATCH-STEM — REFUTED before shipping (2026-09-17, pilot-12 trajectory replay). The distiller's 16
 # "repeated-identical-retry" candidates were re-read: the two stuck sessions (coq-block-bound base arm: `cat > /tmp/tN.v <<EOF` with a
 # DIFFERENT proof body each attempt; rs-archive-clone: 191 distinct `python3 -c` bodies) are unconverged EXPLORATION, not retries —

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.10",
+  "version": "4.124.12",
   "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": "4.124.10",
+    "@nexus-cortex/core": "4.124.12",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",