@nexus-cortex/cli 4.124.33 → 4.124.35

package/.env.defaults

@@ -872,6 +872,37 @@ CORTEX_ENDTURN_GATE=true
 # [LEDGER 2026-09-15 @4.108.21] BUILT + unit-tested (timeBudget.test.ts bands/line/levers; TurnEndGuards.test.ts open-items shapes).
 #CORTEX_BUDGET_VISIBILITY=
 #CORTEX_BUDGET_CONTINUE_MIN_REMAINING=
+# CORTEX_TURN_STATUS — HB-TURN-STATUS (DARK, 2026-09-29). TB4.0 early self-finish root-cause (r-tb4-early-finish-rootcause-2026-09-29):
+# 141/162 fails finished with >50% of the budget left; the model's sense of elapsed time ran 6-10x fast and it gave up "out of time"
+# at <1 h of 8 h — the only clock it saw was the R151 WALL BUDGET line, once per 10% band. off (default; unset = byte-identical requests)
+# | auto (on when CORTEX_TURN_DEADLINE_MS > 0 — the bench always sets it) | on (always; elapsed only without a deadline). On EVERY
+# tool-result round (sendMessage + streamMessage, every handleToolCalls path) one line is appended to the TAIL of the round's newest
+# tool_result: <system-reminder>STATUS: elapsed 0h12m of 7h12m (3%), ~7h00m left; context ~84K of 1M tokens (8%); bg: <id> "<cmd>"
+# 4m12s running</system-reminder> (context = last API prompt_tokens + request growth, R132 anchor; bg = the 3 most recent entries of
+# executors' BackgroundProcessRegistry — the Bash run_in_background / BashOutput / KillShell registry). CACHE: persisted in the in-memory
+# history like the band line, so every earlier request is a byte-identical prefix; only the new tail varies. The band line stays.
+# Event: one turn_status row per turn {lines, mode, streaming}.
+# [LEDGER 2026-09-29 @4.124.33] DARK — BUILT + unit-tested (turnStatus.test.ts line/levers/bg/placement; turnStatus.integration.test.ts
+# both loops: STATUS tail on every continuation, prior request prefixes byte-identical, unset/auto-without-deadline = no STATUS). NO bench arm yet.
+#CORTEX_TURN_STATUS=
+# CORTEX_EFFORT_RAMP — HB-EFFORT-RAMP (DARK, 2026-09-30). TB4.0 t4cw traces: the action model hit the 65,536-token output wall right
+# after reading the task (photonic r2 call 3, protein r1 call 6), trying to solve the routing plan / hidden rule in its head before writing
+# code; it only started acting once forced to low. The first N MAIN action calls of a turn (0 = initial request; retry/synthesis calls
+# don't count) run at the ramp level, then the configured effort. off (default) | <level>[:<calls>] (low:8) | on (= low:8). DeepSeek
+# levels are low | high | max (medium = high). Precedence: R153 backoff > effort pulse > ramp > request param > card. Events: effort_ramp
+# {phase start|end}. CAUTION: GLOBAL low was harmful on TB4.0 (e6: 3/29 vs high 10/29) — this lever only covers the opening calls.
+# [LEDGER 2026-09-30 @4.124.34] DARK — BUILT + tested (effortLevers.test.ts; effortLevers.integration.test.ts both loops: calls 0..N-1 at
+# the ramp level, then the configured effort; unset = untouched). NO bench arm yet.
+#CORTEX_EFFORT_RAMP=
+# CORTEX_COMPUTE_NUDGE / _TOKENS / _COOLDOWN / _MAX — HB-COMPUTE-NUDGE (DARK, 2026-09-30). Same traces: steps that DID act often carried
+# 35K-52K reasoning tokens first; photonic r2's productive stretch began when it ran the task's checker. After a round whose response
+# reasoned past _TOKENS (default 16000), one line is appended to the tail of the newest tool_result (the HB-TURN-STATUS carrier — earlier
+# request prefixes stay byte-identical): "COMPUTE, DON'T DELIBERATE: your last step spent ~NK reasoning tokens ... write and run a short
+# script or the task's own checker ...". Cooldown _COOLDOWN rounds (3), cap _MAX per turn (6). off (default) | on. Event compute_nudge.
+# [LEDGER 2026-09-30 @4.124.34] DARK — BUILT + tested (unit + both-loop integration: fires on a heavy step, cooldown respected; unset =
+# never). NO bench arm yet.
+#CORTEX_COMPUTE_NUDGE=
+#CORTEX_COMPUTE_NUDGE_TOKENS=
 # (no lever) R152 HB-EMPTY-COMPLETION-RETRY (4.108.22, 2026-09-15): during a 27-min DeepSeek incident (1:30-1:57 AM PT) chat/completions
 # returned bodies that parsed to null; the SDK handed them back, the translation layer dereferenced `.choices`, and the tool loop ended the turn
 # on the TypeError — 30 sessions across tb21-a21 A/B + tb4-flash-v3 (largest-eigenval died at iteration 1). R150's ladder did not apply (a null
@@ -910,6 +941,12 @@ CORTEX_ENDTURN_GATE=true
 # from it) and the counter covers the retry call plus TURNS continuations.
 #CORTEX_REASONING_EXHAUST_BACKOFF=
 #CORTEX_REASONING_EXHAUST_BACKOFF_TURNS=
+# CORTEX_REASONING_EXHAUST_LADDER — R153c (2026-09-30, DEFAULT CHANGE). The R153 backoff stepped high/max → 'medium', which DeepSeek
+# runs as HIGH (reasoning_effort ladder low | high | max) — the first backoff was a NO-OP on DeepSeek (t4cw photonic r2: wall at high,
+# "medium" wall again, only then low). provider (default): step down the provider's real ladder (DeepSeek high → low, max → high; others
+# unchanged low | medium | high) | legacy: the pre-R153c steps. reasoning_exhaustion events now carry {ladder, provider}.
+# [LEDGER 2026-09-30 @4.124.34] ON (bug fix) — unit-tested (effortLevers.test.ts). NO bench arm.
+#CORTEX_REASONING_EXHAUST_LADDER=
 # (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 —

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.33",
+  "version": "4.124.35",
   "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": "4.124.33",
+    "@nexus-cortex/core": "4.124.35",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",