@nexus-cortex/cli 4.124.39 → 4.124.41

package/.env.defaults

@@ -993,6 +993,44 @@ CORTEX_ENDTURN_GATE=true
 # server + adapter sibling survive a matching pattern, killall spares the sibling, pgrep hides the harness, the c21ba command shape through
 # the real Bash tool); ShellTool/BashOutput/KillShell suites unchanged (145 tests).
 #CORTEX_KILL_GUARD=
+# CORTEX_PID1_GUARD — R209 HB-CONTAINER-PID1 (DARK, 2026-10-01). c23gb r1 exit 137: the agent SIGKILLed `supervisord` with os.kill inside
+# a Python script (no shell kill guard can see that); PID 1 was /app/entrypoint.sh doing `wait $supervisord_pid`, so PID 1 exited, the
+# container stopped and the harness was SIGKILLed. on: at the first tool_result boundary (the boot observation, normally orient's output;
+# independent of the tool anchor) the harness reads /proc/1/cmdline and the direct children of PID 1 (/proc/*/stat ppid == 1, the harness
+# and its ancestors excluded, max 3) and, when PID 1 is not the harness, appends ONE system-reminder: "Container init: PID 1 is `<cmd>`; it
+# supervises `<child>` (pid N). Killing pid N (or PID 1) stops the container and ends your session — restart services through their
+# supervisor (e.g. `supervisorctl restart <name>`) instead." Linux-only, one /proc pass, fail-safe (any read error → no note). One-shot
+# per orchestrator. Event pid1_guard {noted, pid1Cmd, children, chars} when the note is added. off: nothing read, nothing added.
+# [LEDGER 2026-10-01 @4.124.40] DARK — BUILT + unit-tested (pid1Guard.test.ts: fake /proc builder cases). NO bench arm yet.
+#CORTEX_PID1_GUARD=
+# CORTEX_PLATEAU_STOP / _REL / _WINDOW / _MIN_POINTS — R210 HB-PLATEAU-STOP (DARK, 2026-10-02). v39a layout-config-recreation2: the
+# agent cannot see the grader's pass threshold on an optimization task, so it polished a self-computed score for 5.3 h (483 iterations,
+# 840,636 → 851,547 with ever-finer searches); budget steering reports time left, never "enough". Operator design = a PLATEAU VARIATION
+# TARGET. on: at the first tool_result boundary (same carrier as the R209 PID 1 note) ONE system-reminder asks the model, when optimizing
+# a measurable objective, to print `PLATEAU_METRIC <name>=<number> dir=<max|min>` after every evaluation and keep the deliverable at its
+# best result. Every tool round, the tool_result texts are scanned for that line (last occurrence per result wins); a per-session series
+# is kept per metric name (orchestrator/plateauStop.ts). Plateau = n >= MIN_POINTS and the best of the last WINDOW reports improves on
+# the best before them by < REL, relative (max: (bestRecent-bestBefore)/|bestBefore|; min inverted; a series that got worse counts).
+# The first plateau steers ONE reminder on the budget-steering channel (appended to the round's first tool_result, steering_injected
+# kind 'plateau'): "PLATEAU: your <name> improved only <x>% over the last <window> evaluations (target ≥ <rel>%). Stop optimizing.
+# Make sure the deliverable holds the best result (<best>), run the task's own verification/tests once, then finish." At most 2 per
+# session; the second only after >= WINDOW more reports with still no real gain. Events: plateau_metric_seen {name, dir, value} (once
+# per metric name), plateau_stop {name, dir, best, improvementPct, points, firing}. STATUS segment (with CORTEX_TURN_STATUS on): once a
+# metric was reported, the per-turn STATUS line gains ONE segment for the most recently reported metric — `score <name>: best <v> (<n>
+# evals); +<x>% over last <window> (target ≥<rel>%)`, plus ` — PLATEAU` while the detector says plateau (improvement part only once the
+# series is comparable). No metric seen or lever off → STATUS byte-identical. off: no instruction, results not scanned, no segment.
+# REL = relative-improvement target (default 0.002 = 0.2%); WINDOW = most recent reports compared (default 6); MIN_POINTS = reports
+# needed before any stop can fire (default 8). Bad values fall back to the defaults.
+# [LEDGER 2026-10-02 @4.124.41] DARK — BUILT + unit-tested (plateauStop.test.ts: detector rising/flat/worse/min/min-points/window edges,
+# noisy-line parsing, firing cap, lever-off path, STATUS segment; plateauStop.integration.test.ts: both loops, off = no instruction/
+# reminder/segment, on = instruction + 2 reminders + PLATEAU segment). NO bench arm yet. evidence: R210 v39a layout-config-recreation2 (483 iterations, 5.3 h).
+#CORTEX_PLATEAU_STOP=
+# [LEDGER 2026-10-02 @4.124.41] DARK — default 0.002 (0.2%) relative; untested on a bench. evidence: R210 v39a layout (840,636 → 851,547 over 5.3 h; per-window gains not yet measured).
+#CORTEX_PLATEAU_REL=
+# [LEDGER 2026-10-02 @4.124.41] DARK — default 6 reports; untested on a bench. evidence: R210 v39a layout.
+#CORTEX_PLATEAU_WINDOW=
+# [LEDGER 2026-10-02 @4.124.41] DARK — default 8 reports; untested on a bench. evidence: R210 v39a layout.
+#CORTEX_PLATEAU_MIN_POINTS=
 # (no lever) R154 HB-IMAGE-REJECTION-HEAL (4.109.0, 2026-09-16): retro-console-soc generated a 128x128 8-bit GRAYSCALE PNG, called
 # ReadImage, and DeepSeek answered the next continuation with `400 .messages[290].image[0]: You have uploaded an unsupported image`;
 # the exception escaped after the tool batch, the already-executed guard left the loop, and the R29a synthesis failed on the same block —
@@ -1087,6 +1125,16 @@ CORTEX_ENDTURN_GATE=true
 # hold a finish. Event field namedInconclusive.
 #CORTEX_JUDGE_VETO=
 #CORTEX_JUDGE_EVIDENCE_MAX_VETOES=
+# CORTEX_VERIFY_SCALE — R208 HB-VERIFY-AT-SCALE (DARK, 2026-10-01). c23 payments-pipeline-fix 0/6: the agent applied a plausible fix,
+# verified it with its own small run (latency 0.2–1.9 s) and finished; the hidden tests ran ~920k records against a 5 s SLA and measured
+# 8–22 s. The EndTurn judge accepted the small-scale self-report. on: the judge persona (resolverSystemPrompt) carries ONE extra rubric item
+# (RESOLVER_SCALE_CLAUSE): does the verification reproduce the scale, volume, concurrency and timing the task's tests will use (data size,
+# records/requests, parallel clients, restarts, time limits) rather than a toy run? Evidence that covers only a smaller run is a GAP, and
+# the judge names a CHECK at the stated scale — under VETO=evidence that check is what can hold the finish (45 s check cap applies).
+# off: persona byte-identical. Events: verify_scale_item {surface, clauseChars, roundsUsed, meets, parsed, confidence} per adjudication;
+# endturn_resolver gains {verifyScale}. Needs CORTEX_ENDTURN_RESOLVER=true.
+# [LEDGER 2026-10-01 @4.124.40] DARK — BUILT + unit-tested (endTurnResolver.test.ts R208 cases). NO bench arm yet.
+#CORTEX_VERIFY_SCALE=
 # CORTEX_TURN_CONTRACT — HB-TURN-CONTRACT (4.115.0, 2026-09-17; Terminus-2 disparity "suppress vs channel"). The boot-minimal door SUPPRESSES
 # deliberation ("Prefer acting over deliberating"; tuned at effort low). At effort high the model fought it: tb4-flash-v3 had 58 turns in
 # 23 sessions reason to the 65536 cap and emit nothing (R153). Terminus 2 CHANNELS deliberation: think freely, then every response must

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.39",
+  "version": "4.124.41",
   "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": "4.124.39",
+    "@nexus-cortex/core": "4.124.41",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",